content
stringlengths 1
15.9M
|
|---|
\section{INTRODUCTION}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{figures/final9.png}
\vspace{-0.3cm}
\caption{An overview of our system. During training, images from the center camera are embedded into a low-dimensional representation, $Z_c$, which is deterministically transformed into the predicted embeddings of the left and right viewpoints respectively, via equivariant mappings $M$. Both the actual steering command (solid red arrow) and the predicted steering commands (dashed red arrows) are included in the training set for the policy. During testing, the policy is executed on embeddings from the center camera. The training set for the mappings $M$ can be different than the test set for $\pi$.}
\label{fig:first_page}
\vspace{-0.5cm}
\end{figure}
From autonomous driving to flying and manipulation, vision-based policies learned through imitation are starting to be deployed on robot systems of critical importance. To ensure safety during deployment, we would ideally like to have formal guarantees about the generalization and robustness properties of these policies. While desirable, however, formal guarantees for out-of-distribution generalization have been attained on perturbation models that are theoretically convenient, but limited in practice~\cite{pmlr-v97-cohen19c}, as they often do not capture the full range of plausible unseen states. Due to this reason, data augmentation during training has high practical significance. Yet, existing augmentation schemes focus cropping, rotating, blurring, perturbing by Gaussian noise, and other operations that do not capture the spatial extent of demonstrations.
In this paper we address this problem by proposing a dataset augmentation scheme for training policies whose inputs are image embeddings, as shown in Fig.~\ref{fig:first_page}.
Given an observed image from the center camera of a vehicle, our method allows us to learn a map that predicts the embeddings of corresponding nearby viewpoints, as well as the actions that would have needed to be taken if the vehicle's camera was at those viewpoints. Our method relies upon predicting embeddings of nearby viewpoints using \emph{equivariant mappings}~\cite{Jayaraman_2015_ICCV, lenc2015understanding} that are trained to transform the embedding of the center camera to the embeddings of nearby viewpoints, both for ground vehicles and for flying vehicles, as we will show in the evaluation section. Equivariance has emerged as an important structural bias for neural network design in the last few years. Here we use it as an auxiliary task to the main behavioral cloning loss. Our contributions are the following:
\begin{itemize}
\item We show that augmenting the training dataset of the policy with predicted embeddings of nearby viewpoints by learned equivariant maps increases the robustness of the vision-based policy and leads to fewer interventions from
the operator and lower cross-track error on average.
\item We demonstrate in simulation that this increased robustness occurs both for ground vehicles and for flying vehicles in the context of UAV racing.
\item We also show that it is the case for real-world visual navigation with a terrestrial mobile robot.
\end{itemize}
\section{RELATED WORK}
\textbf{Equivariant representations:} Equivariance for image representations captures relationships between encodings of two related images that are determined by the transformation of those images~\cite{lenc2015understanding}. If the encodings are the same, then they are invariant to that transformation. In our work, we are dealing with image pairs that look at the same scene from different, but nearby viewpoints, for instance the center and the left camera mounted on a self-driving car. We want the representation of the left camera to be predictable by the center camera, so we want a high degree of equivariance. Feature equivariance and invariance in terms of rotation transformations was studied in~\cite{equivariant_rotations, Zhao2019QuaternionEC, lifd_survey}, as well as in the context of probabilistic models~\cite{dbn_equivariant, Bilovs2020EquivariantNF, Rezende2019EquivariantHF}. The utility of equivariant representations for multiple downstream tasks, learned by egomotion was recognized in~\cite{Jayaraman_2015_ICCV}. Our work here is related the most to this paper.
Equivariant representations have also been useful for 3D volumetric data~\cite{Weiler20183DSC}, even in the fully unsupervised setting~\cite{Spezialetti2019LearningAE}, addressing equivariance and invariance to both rotation and translation for 3D data~\cite{Fuchs2020SE3Transformers3R, Wang2020EquivariantMF}.
Additionally, instead of imposing equivariance through a loss function, as in~\cite{Jayaraman_2015_ICCV}, many recent works build equivariance in the structure of the network~\cite{cubenet, worrall2017harmonic, cohen2019gauge, Tai2019EquivariantTN, Cohen2019AGT, steerable_filters_equivariant_cnn}, for example in the multi-view setting~\cite{emvn, multiviews, gvcnn, qi2016volumetric, you2018pvnet}. Finally, equivariance has also been used to improve state representations for planning in MDPs~\cite{Pol2020PlannableAT}.
\textbf{Automated data augmentations:} Typical dataset augmentations for supervised deep learning have included addition of noise, rotations, crops, blurring, and others. While this is usually a set of manually hand-crafted augmentation schemes, there has been increased interest in automatically computing data augmentations~\cite{ratner2017learning, zhang2019adversarial, dao2019kernel, raileanu2020automatic} to increase the robustness of classifiers, regressors, and RL policies. Augmentations have also been used in monocular depth estimation~\cite{ravi_unsupervised_depth}.
\textbf{Robustness in behavioral cloning:}
Behavioral cloning techniques that learn policies through pure supervised learning from a given dataset of state-action pairs usually suffer from covariate shift~\cite{Loquercio_2018,lecun_obstacle_avoidance}. Back in the early 1990s Dean Pomerleau's PhD thesis~\cite{pomerleau_thesis} and pioneering work on ALVINN~\cite{alvinn} provided one of the first visually guided self-driving cars. His thesis details a data augmentation scheme that injects synthetic images of the road ahead from viewpoints nearby the camera, to the policy training scheme. Our paper provides an alternative that is based on neural network view synthesis, which in the last years has seen significant progress. Pomerleau's thesis also identified issues of covariate shift, which methods such as DAgger~\cite{DAgger_ross} address by iteratively labeling the newly visited states after each policy evaluation. The question of dataset augmentation in one round of behavioral cloning, however, remains. Thus, collecting large driving datasets, in a way that no manual annotation is required from the user, has proved to be very popular among self-driving car and visual navigation research circles. The first instance of this was shown in UAV navigation in forest trails~\cite{trail_following}.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/train_smath.png}
\vspace{-0.4cm}
\caption{Representation learning phase: we use an encoder-decoder network to learn image embeddings that are equivariant to input camera viewpoint. Mapping network $M_{i \rightarrow j}$ is learned to map these embeddings from viewpoint $i$ to viewpoint $j$. The encoder-decoder and the mapping networks are trained collectively by optimizing a combined loss function shown in Eqn. \ref{eqn:repr_loss}.}
\label{fig:eqSetup}
\end{figure*}
The images of the center camera are automatically annotated with the forward steering command, the images of the left camera with the right, and the images of the right camera are annotated with the left steering command. A similar hardware setup and automated data collection was used by the NVIDIA self-driving car team~\cite{nvidia_driving}, as well as more recent work on behavior cloning conditioned on human commands (e.g. turn left at the next intersection)~\cite{codevilla_2017}.
The crux behind the idea of the 3-camera setup is that behavioral cloning can become more robust if we obtain expert actions from states that are nearby the current one. In cases where we cannot address this problem using multiple sensor measurements from nearby viewpoints, one way to proceed is to explicitly generate nearby states that the expert will have to annotate. One simple way to do this is shown in DART~\cite{dart_laskey}, where Gaussian noise is added to the expert's demonstrations in order to explicitly include nearby labeled states in the dataset.
\section{METHODOLOGY}
Our goal is to learn an encoding method that respects equivariance constraints with respect to camera poses or viewpoints nearby the original camera. This allows us to predict the embeddings of nearby viewpoints from the observations of the center camera.
{\color{blue} Here, nearby viewpoints refer to horizontal and vertical translation of the camera over short distances capturing the same scene, our case involves translations of 0.25 to 0.5 meters.}
Given an image $I_i \in \mathbf{X}$ of a scene taken from viewpoint $i$, we compute its corresponding embedding in the latent space $Z_i \in \mathbf{Z}$, and use an equivariant map $M_{i \rightarrow j}$ to transform it to the embedding corresponding to
another desired nearby viewpoint $j$. This predicted embedding $Z\sp{\prime}_j \in \mathbf{Z}$ can then be used as
augmented training data for policies trained on such embeddings and actions.
We train an image encoder network on a dataset of image pairs, where each pair consists of images of the same scene from neighboring viewpoints. In our case, the images are obtained from synchronized videos recorded by a set of cameras mounted on a vehicle visually navigating through an environment. The relative poses of the cameras are assumed to be known and looking at the same scene. With each image pair we simultaneously also record the corresponding expert control command used to navigate the vehicle at that time.
We then use the embeddings obtained from the images to train imitation learning policies used to control both ground and flying vehicles, and compare their performance to those trained without augmented data.
\subsection{Equivariant Feature Learning} \label{eqLearning}
Given an image set $\textbf{I} = \langle I_1, I_2, ..., I_n \rangle$, our model learns an encoder $z_\theta (\cdot) : I_i \to \mathbb{R}^D$, parametrized by $\theta$, that maps an element of \textbf{I} to an embedding $Z$. We want the embeddings to exhibit equivariance, i.e. change predictably with respect to transformations applied to the viewpoint $i$ of the input image $I_i$. Let $g$ be a transformation applied to the input image $I$ in the pixel space, and let $M_g$ define a corresponding transformation in the latent space. Then:
\begin{equation}
z_\theta (g(I)) \approx M_g (z_\theta(I))
\end{equation}
For example, suppose the images $I_c$ and $I_l$ correspond to the center and left camera images in Fig.~\ref{fig:first_page}, respectively. If $g$ is an image transformation such that \mbox{$g(I_c) = I_l$}, then $M_{g}=M_{c \rightarrow l}$ defines a transformation in the latent space such that $ M_{g}(Z_c) = Z\sp{\prime}_{{ l}} \approx Z_l$, i.e. the embedding $Z_c$ derived from $I_c$, when transformed using $M_{g}$, approximates the embedding $Z_l$ derived from $I_l$. In latent space, $Z_c$ and $Z_l$ maintain a relation directly corresponding to the one between $I_c$ and $I_l$ in the pixel space.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/control_smath.png}
\vspace{-0.4cm}
\caption{Policy training phase: the center image is passed through the encoder network, and the resulting embedding is passed through the FCN to get a control output for the vehicle.}
\label{fig:controlsetup}
\vspace{-0.5cm}
\end{figure}
\subsection{Equivariant Feature Learning as an Auxiliary Task} \label{eqObjective}
Now, we design an objective function that encourages the learned feature space $Z$ to exhibit equivariance with respect to camera transformations, or camera viewpoints in our case. Given image set $\textbf{I} = \langle I_1, I_2, ..., I_n\rangle$ where $I_i$ is an image taken from the $i^{th}$ viewpoint, and the corresponding embeddings $\textbf{Z} = \langle Z_1, Z_2, ..., Z_n \rangle$, we setup an objective function to enforce equivariance between any two image viewpoints.
For any pair of viewpoints $(i,j)$, we can learn a map $ M_{i \rightarrow j}$, and by minimizing the $L_2$ norm between original and mapped latent representations, the function forces the embedding map $ M_{i \rightarrow j}$ from the $i^{th}$ to the $j^{th}$ viewpoint to preserve equivariant relations. The loss function is as follows:
\begin{equation}
L_{eq}(M) = \textstyle \sum\limits_{i} \sum\limits_{j} ||Z_j - M_{i \rightarrow j}(Z_i)||^2
\label{eqn:leq}
\end{equation}
\subsection{Image Reconstruction as an Auxiliary Task} \label{mappingFun}
In addition to the equivariance loss among all available viewpoints we make use of image reconstruction as an auxiliary task, so that the embeddings passed into the loss in Eqn.~\ref{eqn:leq} above capture the information needed to reconstruct images from latent space. The function $z_\theta (\cdot)$ encodes images into embeddings and the mapping function $ M_{i \rightarrow j}$, implemented in our case as a multilayer perceptron, transforms these embeddings from one viewpoint into another. We make use of a Siamese encoder-decoder architecture and share the convolutional encoding and decoding weights among all pairs of viewpoints.
The encoder, decoder, as well as the equivariant mapping $M_{i \rightarrow j}$ are optimized together in a shared objective.
The latent representations $Z_i$ are decoded back to images by a decoder $p_\varphi (\cdot)$. The output of the decoder is then used to compute and minimize the reconstruction loss $L_{rc} $:
\begin{equation}
L_{rc}(\theta, \varphi) = \textstyle \sum\limits_{i} ||I_i - p_{\varphi}(z_{\theta}(I_i))||^2
\label{reconLoss}
\end{equation}
\noindent To further reinforce the effectiveness of the equivariant maps $ M_{i \rightarrow j}$, we enforce an additional reconstruction penalty between each image and its reconstructions from other viewpoints:
\begin{equation}
L_{eq\textrm{-}rc}(\theta, \varphi, M) = \textstyle \sum\limits_{i} \sum\limits_{j} ||I_j - p_{\varphi}(M_{i \rightarrow j}(z_{\theta}(I_i))) ||^2
\label{reconLoss}
\end{equation}
The overall representation learning loss is a combination of the reconstruction, equivariant reconstruction, as well as the equivariant loss as follows:
\begin{equation}
L(\theta, \varphi, M) = \lambda_1 L_{rc}(\theta, \varphi) + \lambda_2 L_{eq}(M) + \lambda_3 L_{eq\textrm{-}rc}(\theta, \varphi, M)
\label{eqn:repr_loss}
\end{equation}
We backprop through the combined loss in Eqn.~\ref{eqn:repr_loss} to optimize $\theta, \varphi$, and $M$. The hyperparameters are chosen as $\lambda_1=1$, $\lambda_2=10$, and $\lambda_3=1$ to ensure all loss terms have relatively same scale.
\subsection{Imitation Learning with learned Features} \label{bcTraining}
We do representation learning separately from policy learning. Once we are done optimizing the loss in Eqn.~\ref{eqn:repr_loss}, we train a policy through imitation in order to control vehicles based on the learned encoder, as shown in Eqn.~\ref{eqn:policy_loss}. We use a multi-layer perceptron to represent the policy $\pi$ and we train it on embedding-action pairs both from the center camera (observed embeddings and actions) and from the nearby viewpoints (predicted embeddings from equivariance relations and predicted actions). The objective function used to train the policy is the following:
\begin{equation}
L(\pi) = \textstyle \sum\limits_{i} ||a_i - \pi(Z_i)||^2
\label{eqn:policy_loss}
\end{equation}
At test time and deployment, the encoder is used to convert the images to embeddings which are then used as input to the learned policy.
\section{Experiments}
We validate our approach in two different experimental settings, an aerial vehicle and a terrestrial vehicle. The experiments involve the following steps:
\begin{enumerate}
\item \textbf{Collecting Data:}
Each data sample $d = \langle \textit{\textbf{I}}, \textit{\textbf{A}} \rangle$ consists of an image set $\textit{\textbf{I}}$ that comprises of images from different camera viewpoints and an action set $\textit{\textbf{A}}$ that contains the corresponding control actions. We collect data both in simulation and in the real world, for the respective experiments.
\item\textbf{Learning Equivariant Features:} We learn the feature mapping function $z_\theta(\cdot)$ and equivariant mappings $M_{i \rightarrow j}$ using the collected image sets $\textit{\textbf{I}}$ in the dataset.
\item \textbf{Learning Control via Imitation Learning:} We learn to control the vehicle through an imitation learning policy that is trained using both image sets \textit{\textbf{I}} and action sets \textit{\textbf{A}}.
\end{enumerate}
In order to assess the effectiveness of our learned embeddings and equivariant mappings we evaluate and compare three imitation learning policies that each learn to map images to control actions:
\begin{enumerate}
\item The first policy is trained end-to-end on image-action pairs obtained from a single camera, in our case the center camera mounted on the vehicle.
\item The second policy is trained end-to-end on image-action pairs obtained from all the cameras mounted on the vehicle. In our case it's three for the car, five for the flying vehicle. The recorded expert control is modified to account for the shifted camera viewpoints for the off center cameras.
\item The third policy, the one that uses our equivariant augmentations, is trained on image-action pairs obtained from only a single camera, but the generated embeddings are then mapped to the viewpoints of the other cameras in the latent space. This corresponds to the training scheme described in Sec.~\ref{mappingFun}.
\end{enumerate}
The end-to-end imitation learning policy is trained using the same methods and network architecture used by \cite{nvidia_driving}. The network contains a set of convolutional layers followed by a set of fully connected layers. For consistency in comparisons, we ensure that the convolutional part of the imitation learning policy has the same architecture as the image encoder $z_\theta(\cdot)$. All the networks are optimized using ADAM optimizer \cite{adam} with a learning rate of $10^{-4}$ and a batch size of $64$. The policy is evaluated by how well it can navigate a vehicle along a specified path using a single center camera image as input as done in \cite{nvidia_driving}. We setup DAgger \cite{DAgger} to investigate the performance of the control policy with respect to DAgger iterations. We hypothesize that the policy trained using equivariant augmentations will perform better than a policy trained without them using only the data from a single camera, but not as well as a policy trained directly using the data captured from all the cameras.
As the DAgger runs increase, the difference in performance between the policies becomes less pronounced as the data aggregation process itself makes all three policies more robust. This provides us with one measure of how effective our embedding method is for data augmentation, particularly in the beginning of imitation. To compare the three policies, we record the number of interventions and average cross track error (deviation from the specified trajectory) to use as performance metrics. The experiment setup details and results are discussed in Sec. \ref{droneExp} and Sec. \ref{carExp}.
\subsection{Quadrotor Simulation Experiments} \label{droneExp}
\subsubsection{Experimental Setup}
For this experiment, we use a flying vehicle in the AirSim simulator \cite{airsim2017fsr}. The vehicle is set up with five front facing cameras.
For the dataset used in these experiments, each data sample comprises of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r, I_t, I_b \rangle$ corresponding to timestamped images from the left, center, right, top, and bottom cameras. These cameras are located at a distance of $50 cm$ each from the center camera in the corresponding directions.
The action set $\textit{\textbf{A}} = \langle \delta \textrm{yaw}, \delta z \rangle$, is the expert control used to navigate the flying vehicle, applied in terms of the relative change in yaw and $z$ at each timestep. The augmented actions for non-center cameras are computed as shown in Table \ref{tab:flying-corrections}. Note that clockwise rotation is positive $\delta yaw$ and downwards direction is positive $\delta z$.
\begin{table}[!t]
\caption{Augmentation label calculations (flying vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Camera} & \textbf{\textit{$\delta yaw$ $(rad)$}}& \textbf{\textit{$\delta z$ $(m)$}} \\
\hline
Left Camera & $\delta yaw_{cen}+0.03$ $rad$ & $\delta z_{cen}$ \\
\hline
Right Camera & $\delta yaw_{cen}-0.03$ $rad$ & $\delta z_{cen}$ \\
\hline
Top Camera & $\delta yaw_{cen}$ & $\delta z_{cen} + 0.5$$m$ \\
\hline
Bottom Camera & $\delta yaw_{cen}$ & $\delta z_{cen} - 0.5m$ \\
\hline
\end{tabular}
\label{tab:flying-corrections}
\end{center}
\vspace*{-0.4cm}
\end{table}
\subsubsection{Network}
We use a CNN as the encoding network $z_\theta(\cdot)$ and a separate FCN for each of the equivariant\mbox{ maps $\langle M_{c \rightarrow l}, M_{c \rightarrow r}, M_{c \rightarrow t}, M_{c \rightarrow b} \rangle$. }
\subsubsection{Results}
\begin{figure*}[!t]
\hspace{0.8cm}\includegraphics[width=7in, trim={4.5cm 0.0cm 0.0cm 0.5cm},clip]{figures/cte_both.png}
\vspace{-0.7cm}
\caption{Cross track error (m) vs DAgger iterations for flying vehicle (left) and ground vehicle (right). It can be seen that network trained with equivariant augmentation lies in between the network trained using center camera and all cameras in terms of performance.}
\label{fig:cte_drone}
\end{figure*}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/refvsflowntraj-paper.png}
\caption{Reference vs Flown Trajectory. The all-cameras model (right) tracks the trajectory the best, followed by the equivariant-augmentation model(center), followed by the center-camera model(left). Note that missing sections in the flown trajectory are due to an intervention where expert control was used to navigate the vehicle back to the reference trajectory.}
\label{fig:trajs}
\end{figure}
Figure \ref{fig:cte_drone} (left) shows the cross track error vs DAgger iterations for three different test scenarios. As hypothesized we see that using the augmentations generated via the equivariant maps for training improves the imitation performance. Figure \ref{fig:trajs} overlays the trajectory flown on top of the reference trajectory. It can be seen that the policy trained with equivariant augmentations is able to track better than a policy trained using only the center-camera data.
\begin{table}[!t]
\caption{out-of-domain dataset experiment results (flying vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Setup} & \textbf{\textit{Interventions}}& \textbf{\textit{Autonomy (\%)}}& \makecell{\textbf{\textit{Cross Track}} \\ \textbf{\textit{Error (m)}}} \\
\hline
Center Camera & 3 & 84.23 & 2.12 \\
\hline
\makecell{Equiv. Augment. \\Disjoint Training} & 2 & 89.46 & 1.39 \\
\hline
\makecell{Equiv. Augment. \\ Same Training} & 2 & 89.51 & 1.26 \\
\hline
All Cameras & 0 & 100 & 0.17 \\
\hline
\end{tabular}
\label{tab:eqfrozen-flying}
\end{center}
\vspace*{-0.6cm}
\end{table}
We also perform another experiment to gauge how well our method generalizes to new environments. We train the equivariant map on a dataset captured in a separate area of the simulation map with from where we test the model.
As per the results in \mbox{Table \ref{tab:eqfrozen-flying}}, the policy trained using the generated augmentations shows a significant improvement in performance over the policy trained directly on the center-camera data. Furthermore, it performs only marginally worse than the policy which uses the equivariant map trained on the same area of the simulation map as used for testing. These preliminary results suggest that the features captured by the equivariant map are robust and generalize within environments with similar features.
\subsection{Driving Simulation Experiments} \label{carExp}
\subsubsection{Experimental Setup}
For this experiment, we use the CARLA simulator \cite{carla}. The driving vehicle is set up with three front facing cameras mounted behind the windshield, each 25cm apart. Each data sample of the dataset consists of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r\rangle$ of timestamped images from the left, center, and right cameras respectively, paired with their corresponding expert control, in this case, the steering angle for the car collected from the CARLA driving autopilot. The augmentations for non-center cameras are computed as shown in Table \ref{tab:driving-corrections}. Note that clockwise direction is the direction of positive steering in carla setup.
\subsubsection{Network}
We use a very similar architecture as the flying vehile experiment. A CNN as the encoding network $z_\theta(\cdot)$ and a separate FCN for each of the equivariant maps $\langle M_{c \rightarrow l}, M_{c \rightarrow r}\rangle$.
\subsubsection{Results}
\begin{table}[!t]
\caption{Augmentation label calculations (ground vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Camera} & \textbf{\textit{Steering Angle $(normalized$ $[-1,1])$}} \\
\hline
Left Camera & $steering_{cen}+0.05$ \\
\hline
Right Camera & $steering_{cen}-0.05$ \\
\hline
\end{tabular}
\label{tab:driving-corrections}
\end{center}
\vspace*{-0.4cm}
\end{table}
Figure \ref{fig:cte_drone} (right) shows the average cross track error vs DAgger iterations for three different test scenarios. Again we can see that the driving policy trained using equivariant augmentations performs better than policy trained with just the center camera directly. The performance increase subsides as we add more varied data to the training dataset with each DAgger iteration, as expected.
\begin{table}[!t]
\label{table1}
\caption{out-of-domain dataset experiment results (ground vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Setup} & \textbf{\textit{Interventions}}& \textbf{\textit{Autonomy (\%)}}& \makecell{\textbf{\textit{Cross Track}} \\ \textbf{\textit{Error (m)}}} \\
\hline
Center Camera & 8 & 80.24 & 4.00 \\
\hline
\makecell{Equiv. Augment. \\Disjoint Training} & 6 & 87.21 & 2.66 \\
\hline
\makecell{Equiv. Augment. \\Same Training} & 5 & 88.01 & 2.61 \\
\hline
All Cameras & 4 & 89.97 & 2.25 \\
\hline
\end{tabular}
\label{tab:eqfrozen-ground}
\end{center}
\vspace*{-0.6cm}
\end{table}
As per the flying vehicle experiments, we also performed an additional experiment to test generalization of the learned features. The aforementioned experiments used an urban area of Town10 in the CARLA simulator \cite{carla} with lots of buildings and other urban structures. For this experiment however, the equivariance map was trained using a dataset collected in a different area of the map, not used to train the control policy. As per the results in Table \ref{tab:eqfrozen-ground}, we again observe that the policy trained using the generated augmentations shows a significant performance improvement over the policy trained without them, despite the fact that the equivariance map used was trained elsewhere on a different area of the map.
We also run experiments adding noise augmentations to center camera images and training a policy with this augmented dataset. The noise augmentations consist of zero mean gaussian noise with varying standard deviations \mbox{$\sigma = \{0.01, 0.05, 0.1, 0.2, 0.3\}$.} Fig.~\ref{fig:gaussnoise} shows the average cross track error along with the error bars for different noise $\sigma$ settings. Adding noise augmentations improves the performance of networks trained using the center camera data. However, on average our method still performs better in most cases than network trained using single camera.
\begin{figure}[h]
\centering
\hspace*{-0.6cm}\includegraphics[width=3.6in]{figures/sigma_avg_errbars.png}
\vspace*{-0.9cm}
\caption{Average cross track error vs DAgger iterations with center camera network trained with gaussian noise augmentations on the input images.}
\label{fig:gaussnoise}
\vspace{-0.4cm}
\end{figure}
\subsection{Terrestrial Robot Experiments}
\subsubsection{Setup}
These experiments were performed using a Clearpath Husky robot equipped with three front facing cameras. The Husky is setup similar to the driving vehicle in CARLA simulation as described in Sec.~\ref{carExp}. The cameras are mounted 25cm apart on the Husky. Fig.~\ref{fig:huskysetup} shows the image of the Husky with the three ZED cameras used for these experiments. Note that the cameras are inherently stereo, but were used as monocular cameras. Also, no other sensors (Lidar and GPS) mounted on the Husky were used for our experiments.
\begin{table}[!t]
\caption{Augmentation label calculations (husky)}
\begin{center}
\begin{tabular}{|c|c|}
\hline
\textbf{Training Camera} & \textbf{\textit{Yaw Rate} $(rad/s)$} \\
\hline
Left Camera & $\dot{yaw}_{cen}-0.1$ $rad/s$ \\
\hline
Right Camera & $\dot{yaw}_{cen}+0.1$ $rad/s$\\
\hline
\end{tabular}
\label{tab:husky-corrections}
\end{center}
\vspace*{-0.6cm}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figures/husky_setup.png}
\caption{Clearpath Husky used for ground robot experiments with three front facing cameras mounted in the front and a Jetson Xavier on the base plate.}
\label{fig:huskysetup}
\end{figure}
Various other sensors mounted on the Husky include an IMU, Lidar, GPS etc., though they were not used for these experiments. All of the computation involved is performed by an NVIDIA Jetson Xavier, which can be seen in Fig. \ref{fig:huskysetup} mounted on the back-end of the Husky's payload platform.
Each data sample consists of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r\rangle$ of timestamped images from left, center, and right cameras respectively, paired with their corresponding expert control---in this case the yaw rate of the robot, which is recorded from the joystick controller command issued to drive the robot around. The control correction for the left and right camera images are calculated as depicted in table \ref{tab:husky-corrections}.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/path2.png}
\caption{A bird's eye view of the testing path for the Husky experiment. Shows the interventions needed while navigating for each policy.}
\label{fig:huskypath}
\vspace{-0.6cm}
\end{figure}
Figure \ref{fig:huskypath} shows the overhead view of the test path used for the experiment. The path is highlighted in red. Three different models are trained to drive the robot along the path - a model trained using data from a single camera, a model trained using data from a single camera plus equivariant features, and a model trained using data from all three cameras. Once trained, each model's performance navigating the path is then evaluated.
\subsubsection{Network}
We use the same setup as described previously for the Carla simulation experiment. A CNN as the encoding network $z_\theta(\cdot)$ and a separate FCN for each of the equivariant maps $\langle M_{c \rightarrow l}, M_{c \rightarrow r}\rangle$.
\subsubsection{Results}
As hypothesized, the model trained using only center camera data performs worse than the others, steering off the path at several points. The model trained on data generated using equivariant maps as well as the model trained using all three cameras both perform significantly better. Figure \ref{fig:huskypath} shows a snapshot of the entire test run for three setups respectively. The locations where an intervention occurred are denoted by a yellow cross. Both the equivariant feature model and the all camera model fail once at roughly the same location at a turn while the center camera model fails at 9 different locations. The video from this experiment can be found here: \href{https://youtu.be/5g4Kg3-YWvA}{https://youtu.be/5g4Kg3-YWvA}.
\subsection{Drone Racing}
\subsubsection{Experimental Setup}
In this experiment, we showcase the effectiveness of equivariant features in a flying vehicle in the Drone-Racing environment developed by Microsoft \cite{droneracing}. The environment used is an outdoor environment where the drone navigates through a trail with rectangular gates. The aim is to navigate the path while passing through the center of the gates.
The network architecture is the same as described in Section \ref{droneExp}.
\subsubsection{Results}
Again, we train three separate models, using only the center camera data, center camera data plus equivariant augmentations, and all camera data, each of them trained using same amount of expert runs and compare their performance through number of expert interventions required when completing the path. We run DAgger iterations and observe the performance improvements with respect to DAgger iterations.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\columnwidth]{figures/drone-racing-interventions.png}
\caption{Number of interventions vs DAgger iterations for the flying vehicle in drone-racing environment. Model trained using equivariant augmentations lies in between the model trained using a single cameras and the model trained using all five cameras.}
\label{fig:droneracingDAgger}
\vspace{-0.6cm}
\end{figure}
The model trained using data from all 5 cameras completes the path without needing any interventions. Figure \ref{fig:droneracingDAgger} shows the results for this experiment. It is observed that the number of interventions required for the model trained using equivariant augmentations lies between the other two models. This again showcases the effectiveness of augmentations produced using equivariant features in improving imitation learning performance.
\subsection{Learnt vs Non-Learnt Equivariance Maps}
These experiments examine the significance of learned equivariance maps as opposed to other types of maps toward improving imitation learning performance. In each of the experiments thus far, we modeled the equivariance relation using a neural network that is optimized using expert data. We investigate the effectiveness of this learned equivariance mapping by comparing it to two different types of maps - first, a completely random map that does not enforce any equivariance relation. And second, a deterministic map that tries to enforce a deterministic constraint in the embedding space relative to input image viewpoint. The following sections describe these setups in detail. Our results indicate that both alternative techniques fall short of achieving the performance improvements gained using an equivariance map.
\subsubsection{Random Equivariance Maps} \label{randomEqMap}
We use a randomly initialized fully connected network for the equivariance map $M_{i \rightarrow j}$. This random equivariance map is then used to train an imitation learning policy for driving the car in Carla Simulation as described in section \ref{carExp}. Fig. ~\ref{fig:detmap_cte} shows the imitation learning performance with respect to DAgger iterations using this setup. As expected, a learned equivariance map performs much better compared to a random map. This illustrates the role of equivariance towards improving representation for imitation learning.
\subsubsection{Deterministic Equivariance Maps} \label{detEqMap}
This experiment uses a pre-defined equivariance map, essentially a linear translation in the embedding space. This map is used to train an imitation learning policy as before, and compare its performance with learned equivariance map. The experiment setup here is analogous to the CARLA simulation setup used previously. The latent space transformation used is defined in table \ref{tab:detMapDef}. Fig.~\ref{fig:detmap_cte} shows the imitation learning performance with respect to DAgger iterations using this setup. The learned equivariant mapping significantly outperforms this deterministic mapping.
\begin{table}[!t]
\caption{Deterministic mapping used in the embedding space between different input viewpoints.}
\begin{center}
\begin{tabular}{|c|c|}
\hline
\textbf{Mapping} & \textbf{Transformation} \\
\hline
$M_{c \rightarrow l}$ & $z_{cen}+0.25$ \\
\hline
$M_{c \rightarrow r}$ & $z_{cen}-0.25$ \\
\hline
\end{tabular}
\label{tab:detMapDef}
\end{center}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/detEqCTEshrunk.png}
\caption{Cross track error(m) for driving policy trained using a deterministic map and a learned map enforcing equivariant features.}
\label{fig:detmap_cte}
\vspace{-0.6cm}
\end{figure}
\section{CONCLUSION}
In this paper, we investigate the use of equivariant maps for visual navigation by mobile robots. By training a neural network model that learns image features which are equivariant, we can predict the latent representations of images from viewpoints nearby to the one observed. Our results indicate that by augmenting the training dataset with these representations, one can significantly improve navigation performance in a variety of settings as demonstrated by our simulation experiments involving flying and ground vehicles.
Additionally, through our ground robot experiments over a 500m path, we showed that our method also transfers over to the real world.
\section{INTRODUCTION}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{figures/final9.png}
\vspace{-0.3cm}
\caption{An overview of our system. During training, images from the center camera are embedded into a low-dimensional representation, $Z_c$, which is deterministically transformed into the predicted embeddings of the left and right viewpoints respectively, via equivariant mappings $M$. Both the actual steering command (solid red arrow) and the predicted steering commands (dashed red arrows) are included in the training set for the policy. During testing, the policy is executed on embeddings from the center camera. The training set for the mappings $M$ can be different than the test set for $\pi$.}
\label{fig:first_page}
\vspace{-0.5cm}
\end{figure}
From autonomous driving to flying and manipulation, vision-based policies learned through imitation are starting to be deployed on robot systems of critical importance. To ensure safety during deployment, we would ideally like to have formal guarantees about the generalization and robustness properties of these policies. While desirable, however, formal guarantees for out-of-distribution generalization have been attained on perturbation models that are theoretically convenient, but limited in practice~\cite{pmlr-v97-cohen19c}, as they often do not capture the full range of plausible unseen states. Due to this reason, data augmentation during training has high practical significance. Yet, existing augmentation schemes focus cropping, rotating, blurring, perturbing by Gaussian noise, and other operations that do not capture the spatial extent of demonstrations.
In this paper we address this problem by proposing a dataset augmentation scheme for training policies whose inputs are image embeddings, as shown in Fig.~\ref{fig:first_page}.
Given an observed image from the center camera of a vehicle, our method allows us to learn a map that predicts the embeddings of corresponding nearby viewpoints, as well as the actions that would have needed to be taken if the vehicle's camera was at those viewpoints. Our method relies upon predicting embeddings of nearby viewpoints using \emph{equivariant mappings}~\cite{Jayaraman_2015_ICCV, lenc2015understanding} that are trained to transform the embedding of the center camera to the embeddings of nearby viewpoints, both for ground vehicles and for flying vehicles, as we will show in the evaluation section. Equivariance has emerged as an important structural bias for neural network design in the last few years. Here we use it as an auxiliary task to the main behavioral cloning loss. Our contributions are the following:
\begin{itemize}
\item We show that augmenting the training dataset of the policy with predicted embeddings of nearby viewpoints by learned equivariant maps significantly increases the robustness of the vision-based policy resulting in a lower cross track error, and fewer human interventions needed for navigation tasks.
\item Our simulation experiment results indicate that this increased robustness occurs both for ground vehicle and aerial vehicle navigation tasks.
\item We demonstrate that our method applies to real world navigation scenarios by deploying it on a terrestrial mobile robot on campus, resulting in significantly improved navigation performance.
\end{itemize}
\section{RELATED WORK}
\textbf{Equivariant representations:} Equivariance for image representations captures relationships between encodings of two related images that are determined by the transformation of those images~\cite{lenc2015understanding}. If the encodings are the same, then they are invariant to that transformation. In our work, we are dealing with image pairs that look at the same scene from different, but nearby viewpoints, for instance the center and the left camera mounted on a self-drivingvehicle. We want the representation of the left camera to be predictable by the center camera, so we want a high degree of equivariance. Feature equivariance and invariance in terms of rotation transformations was studied in~\cite{equivariant_rotations, Zhao2019QuaternionEC, lifd_survey}, as well as in the context of probabilistic models~\cite{dbn_equivariant, Bilovs2020EquivariantNF, Rezende2019EquivariantHF}. The utility of equivariant representations for multiple downstream tasks, learned by egomotion was recognized in~\cite{Jayaraman_2015_ICCV}. Our work here is related the most to this paper.
Equivariant representations have also been useful for 3D volumetric data~\cite{Weiler20183DSC}, even in the fully unsupervised setting~\cite{Spezialetti2019LearningAE}, addressing equivariance and invariance to both rotation and translation for 3D data~\cite{Fuchs2020SE3Transformers3R, Wang2020EquivariantMF}.
Additionally, instead of imposing equivariance through a loss function, as in~\cite{Jayaraman_2015_ICCV}, many recent works build equivariance in the structure of the network~\cite{cubenet, worrall2017harmonic, cohen2019gauge, Tai2019EquivariantTN, Cohen2019AGT, steerable_filters_equivariant_cnn}, for example in the multi-view setting~\cite{emvn, multiviews, gvcnn, qi2016volumetric, you2018pvnet}. Finally, equivariance has also been used to improve state representations for planning in MDPs~\cite{Pol2020PlannableAT}.
\textbf{Automated data augmentations:} Typical dataset augmentations for supervised deep learning have included addition of noise, rotations, crops, blurring, and others. While this is usually a set of manually hand-crafted augmentation schemes, there has been increased interest in automatically computing data augmentations~\cite{ratner2017learning, zhang2019adversarial, dao2019kernel, raileanu2020automatic} to increase the robustness of classifiers, regressors, and RL policies. Augmentations have also been used in monocular depth estimation~\cite{ravi_unsupervised_depth}.
\textbf{Robustness in behavioral cloning:}
Behavioral cloning techniques that learn policies through pure supervised learning from a given dataset of state-action pairs usually suffer from covariate shift~\cite{Loquercio_2018,lecun_obstacle_avoidance}. Back in the early 1990s Dean Pomerleau's PhD thesis~\cite{pomerleau_thesis} and pioneering work on ALVINN~\cite{alvinn} provided one of the first visually guided self-driving cars. His thesis details a data augmentation scheme that injects synthetic images of the road ahead from viewpoints nearby the camera, to the policy training scheme. Our paper provides an alternative that is based on neural network view synthesis, which in the last years has seen significant progress. Pomerleau's thesis also identified issues of covariate shift, which methods such as DAgger~\cite{DAgger_ross} address by iteratively labeling the newly visited states after each policy evaluation. The question of dataset augmentation in one round of behavioral cloning, however, remains. Thus, collecting large driving datasets, in a way that no manual annotation is required from the user, has proved to be very popular among self-driving car and visual navigation research circles. The first instance of this was shown in UAV navigation in forest trails~\cite{trail_following}.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/train_smath.png}
\vspace{-0.4cm}
\caption{Representation Learning Phase: we use an encoder-decoder network to learn image embeddings that are equivariant to input camera viewpoint. Mapping network $M_{i \rightarrow j}$ is learned to map these embeddings from viewpoint $i$ to viewpoint $j$. The encoder-decoder and the mapping networks are trained collectively by optimizing a combined loss function shown in Eqn. \ref{eqn:repr_loss}. See~\ref{methods} for details.}
\label{fig:eqSetup}
\end{figure*}
The images of the center camera are automatically annotated with the forward steering command, the images of the left camera with the right, and the images of the right camera are annotated with the left steering command. A similar hardware setup and automated data collection was used by the NVIDIA self-driving car team~\cite{nvidia_driving}, as well as more recent work on behavior cloning conditioned on human commands (e.g. turn left at the next intersection)~\cite{codevilla_2017}.
The crux behind the idea of the 3-camera setup is that behavioral cloning can become more robust if we obtain expert actions from states that are nearby the current one.
\color{blue}{Our method builds upon this approach by possibly eliminating the need for multiple sensors once you have a trained policy that can simply predict the latent state of those nearby viewpoints. } \color{black} { }In cases where we cannot address this problem using multiple sensor measurements from nearby viewpoints, one way to proceed is to explicitly generate nearby states that the expert will have to annotate. One simple way to do this is shown in DART~\cite{dart_laskey}, where Gaussian noise is added to the expert's demonstrations in order to explicitly include nearby labeled states in the dataset.
\section{METHODOLOGY} \label{methods}
Our goal is to learn an encoding method that satisfies equivariance constraints with respect to camera poses or viewpoints nearby the original camera. This allows us to predict the embeddings of nearby viewpoints from the observations of the center camera.
Here, nearby viewpoints refer to horizontal and vertical translation of the camera over short distances capturing the same scene. Our case involves translations of 0.25 to 0.5 meters.
Given an image $I_i \in \mathbf{X}$ of a scene taken from viewpoint $i$, we compute its corresponding embedding in the latent space $Z_i \in \mathbf{Z}$, and use an equivariant map $M_{i \rightarrow j}$ to transform it to the embedding corresponding to
another desired nearby viewpoint $j$. This predicted embedding $Z\sp{\prime}_j \in \mathbf{Z}$ can then be used as
augmented training data for policies trained on such embeddings and actions.
We train an image encoder network on a dataset of image pairs, where each pair consists of images of the same scene from neighboring viewpoints. In our case, the images are obtained from synchronized videos recorded by a set of cameras mounted on a vehicle visually navigating through an environment. The relative poses of the cameras are assumed to be known and looking at the same scene. With each image pair we simultaneously also record the corresponding expert control command used to navigate the vehicle at that time.
We then use the learned embeddings to train policies using imitation learning.
\subsection{Equivariant Feature Learning} \label{eqLearning}
Given an image set $\textbf{I} = \langle I_1, I_2, ..., I_n \rangle$, our model learns an encoder $z_\theta (\cdot) : I_i \to \mathbb{R}^D$, parametrized by $\theta$, that maps an element of \textbf{I} to an embedding $Z$. We want the embeddings to exhibit equivariance, i.e. change predictably with respect to transformations applied to the viewpoint $i$ of the input image $I_i$. Let $g$ be a transformation applied to the input image $I$ in the pixel space, and let $M_g$ define a corresponding transformation in the latent space. Then:
\begin{equation}
z_\theta (g(I)) \approx M_g (z_\theta(I))
\end{equation}
For example, suppose the images $I_c$ and $I_l$ correspond to the center and left camera images in Fig.~\ref{fig:first_page}, respectively. If $g$ is an image transformation such that \mbox{$g(I_c) = I_l$}, then $M_{g}=M_{c \rightarrow l}$ defines a transformation in the latent space such that $ M_{g}(Z_c) = Z\sp{\prime}_{{ l}} \approx Z_l$, i.e. the embedding $Z_c$ derived from $I_c$, when transformed using $M_{g}$, approximates the embedding $Z_l$ derived from $I_l$. In latent space, $Z_c$ and $Z_l$ maintain a relation directly corresponding to the one between $I_c$ and $I_l$ in the pixel space.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/control_smath.png}
\vspace{-0.4cm}
\caption{Policy Training Phase: The center image is passed to an encoder, then the resulting embedding is mapped to nearby viewpoints (for experiments using an equivariant map), and finally the embeddings are passed to a FCN model to produce steering output. Note that the encoder and FCN models are trained \textit{separately}. See~\ref{bcTraining} for details.}
\label{fig:controlsetup}
\vspace{-0.5cm}
\end{figure}
\subsection{Equivariant Feature Learning as an Auxiliary Task} \label{eqObjective}
Now, we design an objective function that encourages the learned feature space $Z$ to exhibit equivariance with respect to camera transformations, or camera viewpoints in our case. Given image set $\textbf{I} = \langle I_1, I_2, ..., I_n\rangle$ where $I_i$ is an image taken from the $i^{th}$ viewpoint, and the corresponding embeddings $\textbf{Z} = \langle Z_1, Z_2, ..., Z_n \rangle$, we setup an objective function to enforce equivariance between any two image viewpoints.
For any pair of viewpoints $(i,j)$, we can learn a map $ M_{i \rightarrow j}$, and by minimizing the $L_2$ norm between original and mapped latent representations, the function forces the embedding map $ M_{i \rightarrow j}$, from the $i^{th}$ to the $j^{th}$ viewpoint to preserve equivariant relations. The loss function is as follows:
\begin{equation}
L_{eq}(M) = \textstyle \sum\limits_{i} \sum\limits_{j} ||Z_j - M_{i \rightarrow j}(Z_i)||^2
\label{eqn:leq}
\end{equation}
\subsection{Image Reconstruction as an Auxiliary Task} \label{mappingFun}
In addition to the equivariance loss among all available viewpoints we make use of image reconstruction as an auxiliary task, so that the embeddings passed into the loss in Eqn.~\ref{eqn:leq} above capture the information needed to reconstruct images from latent space. The function $z_\theta (\cdot)$ encodes images into embeddings and the mapping function $ M_{i \rightarrow j}$, implemented in our case as a multilayer perceptron, transforms these embeddings from one viewpoint into another. We jointly train an encoder-decoder model for each viewpoint. This is shown in Fig.~\ref{fig:eqSetup}.
The encoder, decoder, as well as the equivariant mapping $M_{i \rightarrow j}$ are optimized together in a shared objective.
The latent representations $Z_i$ are decoded back to images by a decoder $p_\varphi (\cdot)$. The output of the decoder is then used to compute and minimize the reconstruction loss $L_{rc} $:
\begin{equation}
L_{rc}(\theta, \varphi) = \textstyle \sum\limits_{i} ||I_i - p_{\varphi}(z_{\theta}(I_i))||^2
\label{reconLoss}
\end{equation}
\noindent To further reinforce the effectiveness of the equivariant maps $ M_{i \rightarrow j}$, we enforce an additional reconstruction penalty between each image and its reconstructions from other viewpoints:
\begin{equation}
L_{eq\textrm{-}rc}(\theta, \varphi, M) = \textstyle \sum\limits_{i} \sum\limits_{j} ||I_j - p_{\varphi}(M_{i \rightarrow j}(z_{\theta}(I_i))) ||^2
\label{eqReconLoss}
\end{equation}
The overall representation learning loss is a combination of the reconstruction, equivariant reconstruction, as well as the equivariant loss as follows:
\begin{equation}
L(\theta, \varphi, M) = \lambda_1 L_{rc}(\theta, \varphi) + \lambda_2 L_{eq}(M) + \lambda_3 L_{eq\textrm{-}rc}(\theta, \varphi, M)
\label{eqn:repr_loss}
\end{equation}
We backprop through the combined loss in Eqn.~\ref{eqn:repr_loss} to optimize $\theta, \varphi$, and $M$. The hyperparameters are chosen as $\lambda_1=1$, $\lambda_2=10$, and $\lambda_3=1$ to ensure all loss terms have relatively same scale.
\subsection{Imitation Learning with learned Features} \label{bcTraining}
We train policies in a two-stage process where we first learn scene representations using a learning module, i.e. our CNN image encoder representing $z_\theta (\cdot)$, and then use the learned representations to learn a control policy, i.e. a multi-layer perceptron representing $\pi (\cdot)$, via imitation learning.
The control policy $\pi (\cdot)$ is trained on embedding-action pairs both from the center camera (observed embeddings and actions) and from the nearby viewpoints (predicted embeddings from equivariance relations and predicted actions). The objective function used is the following:
\begin{equation}
L(\pi) = \textstyle \sum\limits_{i} ||a_i - \pi(Z_i)||^2
\label{eqn:policy_loss}
\end{equation}
At test time and deployment, the encoder is used to convert the images to embeddings which are then used as input to the learned policy. See Fig.~\ref{fig:eqSetup}.
\section{Experiments}
We validate our approach in two different experimental settings, an aerial vehicle and a terrestrial vehicle. The experiments involve the following steps:
\begin{enumerate}
\item \textbf{Collecting Data:}
Each data sample $d = \langle \textit{\textbf{I}}, \textit{\textbf{A}} \rangle$ consists of an image set $\textit{\textbf{I}}$ that comprises of images from different camera viewpoints and an action set $\textit{\textbf{A}}$ that contains the corresponding control actions. We collect data both in simulation and in the real world, for the respective experiments.
\item\textbf{Learning Equivariant Features:} We learn the feature mapping function $z_\theta(\cdot)$ and equivariant mappings $M_{i \rightarrow j}$ using the collected image sets $\textit{\textbf{I}}$ in the dataset.
\item \textbf{Learning Control via Imitation Learning:} We learn to control the vehicle through an imitation learning policy that is trained using both image sets \textit{\textbf{I}} and action sets \textit{\textbf{A}}.
\end{enumerate}
In order to assess the effectiveness of our learned embeddings and equivariant mappings we evaluate and compare three imitation learning policies that each learn to map images to control actions:
\begin{enumerate}
\item The first policy is trained end-to-end on image-action pairs obtained from a single camera, in our case the center camera mounted on the vehicle.
\item The second policy is trained end-to-end on image-action pairs obtained from all the cameras mounted on the vehicle. In our case it's three for the car, five for the flying vehicle. The recorded expert control is modified to account for the shifted camera viewpoints for the off center cameras.
\item The third policy, the one that uses our equivariant augmentations, is trained on image-action pairs obtained from only a single camera, but the generated embeddings are then mapped to the viewpoints of the other cameras in the latent space. This corresponds to the training scheme described in Sec.~\ref{mappingFun}.
\end{enumerate}
The end-to-end imitation learning policy is trained using the same methods and network architecture used by \cite{nvidia_driving}. The network contains a set of convolutional layers followed by a set of fully connected layers. For consistency in comparisons, we ensure that the convolutional part of the imitation learning policy has the same architecture as the image encoder $z_\theta(\cdot)$. All the networks are optimized using ADAM optimizer \cite{adam} with a learning rate of $10^{-4}$ and a batch size of $64$. The policy is evaluated by how well it can navigate a vehicle along a specified path using a single center camera image as input as done in \cite{nvidia_driving}. We setup DAgger \cite{DAgger} to investigate the performance of the control policy with respect to DAgger iterations. We hypothesize that the policy trained using equivariant augmentations will perform better than a policy trained without them using only the data from a single camera, but not as well as a policy trained directly using the data captured from all the cameras.
As the DAgger runs increase, the difference in performance between the policies becomes less pronounced as the data aggregation process itself makes all three policies more robust. This provides us with one measure of how effective our embedding method is for data augmentation, particularly in the beginning of imitation. To compare the three policies, we record the number of interventions and average cross track error (deviation from the specified trajectory) to use as performance metrics. The experiment setup details and results are discussed in Sec. \ref{droneExp} and Sec. \ref{carExp}.
\subsection{Quadrotor Simulation Experiments} \label{droneExp}
\subsubsection{Experimental Setup}
For this experiment, we use a flying vehicle in the AirSim simulator \cite{airsim2017fsr}. The vehicle is set up with five front facing cameras.
For the dataset used in these experiments, each data sample comprises of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r, I_t, I_b \rangle$ corresponding to timestamped images from the left, center, right, top, and bottom cameras. These cameras are located at a distance of $50 cm$ each from the center camera in the corresponding directions.
The action set $\textit{\textbf{A}} = \langle \delta \textrm{yaw}, \delta z \rangle$, is the expert control used to navigate the flying vehicle, applied in terms of the relative change in yaw and $z$ at each timestep. The augmented actions for non-center cameras are computed as shown in Table \ref{tab:flying-corrections}. Note that clockwise rotation is positive $\delta yaw$ and downwards direction is positive $\delta z$.
\begin{table}[!t]
\caption{out-of-domain dataset experiment results (flying vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Setup} & \textbf{\textit{Interventions}}& \textbf{\textit{Autonomy (\%)}}& \makecell{\textbf{\textit{Cross Track}} \\ \textbf{\textit{Error (m)}}} \\
\hline
Center Camera & 3 & 84.23 & 2.12 \\
\hline
\makecell{Equiv. Augment. \\Disjoint Training} & 2 & 89.46 & 1.39 \\
\hline
\makecell{Equiv. Augment. \\ Same Training} & 2 & 89.51 & 1.26 \\
\hline
All Cameras & 0 & 100 & 0.17 \\
\hline
\end{tabular}
\label{tab:eqfrozen-flying}
\end{center}
\vspace*{-0.5cm}
\end{table}
\begin{table}[!t]
\caption{Augmentation label calculations (flying vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Camera} & \textbf{\textit{$\delta yaw$ $(rad)$}}& \textbf{\textit{$\delta z$ $(m)$}} \\
\hline
Left Camera & $\delta yaw_{cen}+0.03$ $rad$ & $\delta z_{cen}$ \\
\hline
Right Camera & $\delta yaw_{cen}-0.03$ $rad$ & $\delta z_{cen}$ \\
\hline
Top Camera & $\delta yaw_{cen}$ & $\delta z_{cen} + 0.5$$m$ \\
\hline
Bottom Camera & $\delta yaw_{cen}$ & $\delta z_{cen} - 0.5m$ \\
\hline
\end{tabular}
\label{tab:flying-corrections}
\end{center}
\vspace*{-0.4cm}
\end{table}
\subsubsection{Network}
We use a CNN as the encoding network $z_\theta(\cdot)$ and a separate FCN for each of the equivariant\mbox{ maps $\langle M_{c \rightarrow l}, M_{c \rightarrow r}, M_{c \rightarrow t}, M_{c \rightarrow b} \rangle$. }
\subsubsection{Results}
\begin{figure*}[!t]
\hspace{0.8cm}\includegraphics[width=7in, trim={4.5cm 0.0cm 0.0cm 0.5cm},clip]{figures/cte_both.png}
\vspace{-0.7cm}
\caption{Cross track error (m) vs DAgger iterations for flying vehicle (left) and ground vehicle (right). It can be seen that network trained with equivariant augmentation lies in between the network trained using center camera and all cameras in terms of performance. See~\ref{droneExp} and ~\ref{carExp} for details.}
\label{fig:cte_drone}
\end{figure*}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/refvsflowntraj-paper.png}
\caption{Reference vs Flown Trajectory. The all-cameras model (right) tracks the trajectory the best, followed by the equivariant-augmentation model(center), followed by the center-camera model(left). Note that missing sections in the flown trajectory are due to an intervention where expert control was used to navigate the vehicle back to the reference trajectory. See~\ref{droneExp} for details.}
\label{fig:trajs}
\end{figure}
Fig. \ref{fig:cte_drone} (left) shows the cross track error vs DAgger iterations for three different test scenarios. As hypothesized we see that using the augmentations generated via the equivariant maps for training improves the imitation performance. Fig. \ref{fig:trajs} overlays the trajectory flown on top of the reference trajectory. It can be seen that the policy trained with equivariant augmentations is able to track better than a policy trained using only the center-camera data.
\subsubsection{Out-Of-Domain Experiments} {
To gauge how well our method generalizes to new environments, we perform another set of experiments where we train the equivariant map on a dataset captured in a separate area of the simulation map from where we test the model.
As per the results in \mbox{Table \ref{tab:eqfrozen-flying}}, the policy trained using the generated augmentations shows a significant improvement in performance over the policy trained directly on the center-camera data. \color{blue} { Autonomy here is calculated as in~\cite{nvidia_driving}, i.e. the percentage of time the car was not being driven by the learned policy due to an intervention. More specifically, for an experiment with $n$ interventions:
\begin{equation}
\textrm{autonomy} = ( 1 - \frac{ n \cdot (6 \textrm{ seconds })}{ \textrm{ elapsed time in seconds} } ) \cdot 100
\label{eqn:policy_loss}
\end{equation} } \color{black}{}
Furthermore, it performs only marginally worse than the policy which uses the equivariant map trained on the same area of the simulation map as used for testing. These preliminary results suggest that the features captured by the equivariant map are robust and generalize within environments with similar features.
helllo
}
\subsection{Driving Simulation Experiments} \label{carExp}
\subsubsection{Experimental Setup}
For this experiment, we use the CARLA simulator \cite{carla}. The driving vehicle is set up with three front facing cameras mounted behind the windshield, each 25cm apart. Each data sample of the dataset consists of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r\rangle$ of timestamped images from the left, center, and right cameras respectively, paired with their corresponding expert control, in this case, the steering angle for the car collected from the CARLA driving autopilot. The augmentations for non-center cameras are computed as shown in Table \ref{tab:driving-corrections}. Note that clockwise direction is the direction of positive steering in Carla setup.
\subsubsection{Network}
We use a very similar architecture as the flying vehile experiment. A CNN as the encoding network $z_\theta(\cdot)$ and a separate FCN for each of the equivariant maps $\langle M_{c \rightarrow l}, M_{c \rightarrow r}\rangle$.
\subsubsection{Results}
\begin{table}[!t]
\label{table1}
\caption{out-of-domain dataset experiment results (ground vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Setup} & \textbf{\textit{Interventions}}& \textbf{\textit{Autonomy (\%)}}& \makecell{\textbf{\textit{Cross Track}} \\ \textbf{\textit{Error (m)}}} \\
\hline
Center Camera & 8 & 80.24 & 4.00 \\
\hline
\makecell{Equiv. Augment. \\Disjoint Training} & 6 & 87.21 & 2.66 \\
\hline
\makecell{Equiv. Augment. \\Same Training} & 5 & 88.01 & 2.61 \\
\hline
All Cameras & 4 & 89.97 & 2.25 \\
\hline
\end{tabular}
\label{tab:eqfrozen-ground}
\end{center}
\vspace*{-0.4cm}
\end{table}
\begin{table}[!t]
\caption{Augmentation label calculations (ground vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Camera} & \textbf{\textit{Steering Angle $(normalized$ $[-1,1])$}} \\
\hline
Left Camera & $steering_{cen}+0.05$ \\
\hline
Right Camera & $steering_{cen}-0.05$ \\
\hline
\end{tabular}
\label{tab:driving-corrections}
\end{center}
\vspace*{-0.4cm}
\end{table}
Fig. \ref{fig:cte_drone} (right) shows the average cross track error vs DAgger iterations for three different test scenarios. Again we can see that the driving policy trained using equivariant augmentations performs better than policy trained with just the center camera directly. The performance increase subsides as we add more varied data to the training dataset with each DAgger iteration, as expected.
\subsubsection{Out-Of-Domain Experiments} {
As per the flying vehicle experiments, we also performed an additional experiment to test generalization of the learned features. The aforementioned experiments used an urban area of 'Town10' in the CARLA simulator \cite{carla} with lots of buildings and other urban structures. For this experiment however, the equivariance map was trained using a dataset collected in a different area of the map, not used to train the control policy. As per the results in Table \ref{tab:eqfrozen-ground}, we again observe that the policy trained using the generated augmentations shows a significant performance improvement over the policy trained without them, despite the fact that the equivariance map used was trained elsewhere on a different area of the map.
}
\subsubsection{Out-Of-Domain Experiments} \label{noisexp} {
We also run experiments adding noise augmentations to center camera images and training a policy with this augmented dataset. The noise augmentations consist of zero mean gaussian noise with varying standard deviations \mbox{$\sigma = \{0.01, 0.05, 0.1, 0.2, 0.3\}$.} Fig.~\ref{fig:gaussnoise} shows the average cross track error along with the error bars for different noise $\sigma$ settings. Adding noise augmentations improves the performance of networks trained using the center camera data. However, on average our method still performs better in most cases than network trained using single camera.
}
\begin{figure}[h]
\centering
\hspace*{-0.5cm}\includegraphics[width=3.5in]{figures/sigma_avg_errbars.png}
\vspace*{-0.6cm}
\caption{Average cross track error vs DAgger iterations with center camera network trained with gaussian noise augmentations on the input images. See~\ref{noisexp} for details.}
\label{fig:gaussnoise}
\vspace{-0.3cm}
\end{figure}
\subsection{Terrestrial Robot Experiments} \label{huskyexp}
\subsubsection{Setup} \label{husksetup}
These experiments were performed using a Clearpath Husky robot equipped with three front facing cameras. The Husky is setup similar to the driving vehicle in CARLA simulation as described in Sec.~\ref{carExp}. The cameras are mounted 25cm apart on the Husky. Fig.~\ref{fig:huskysetup} shows the image of the Husky with the three ZED cameras used for these experiments. Note that the cameras are inherently stereo, but were used as monocular cameras. Also, no other sensors (Lidar and GPS) mounted on the Husky were used for our experiments.
\begin{table}[!t]
\caption{Augmentation label calculations (husky)}
\begin{center}
\begin{tabular}{|c|c|}
\hline
\textbf{Training Camera} & \textbf{\textit{Yaw Rate} $(rad/s)$} \\
\hline
Left Camera & $\dot{yaw}_{cen}-0.1$ $rad/s$ \\
\hline
Right Camera & $\dot{yaw}_{cen}+0.1$ $rad/s$\\
\hline
\end{tabular}
\label{tab:husky-corrections}
\end{center}
\vspace*{-0.6cm}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figures/husky_setup.png}
\caption{Clearpath Husky used for ground robot experiments with three front facing cameras mounted in the front and a Jetson Xavier on the base plate. See~\ref{husksetup} for details. }
\label{fig:huskysetup}
\end{figure}
Various other sensors mounted on the Husky include an IMU, Lidar, GPS etc., though they were not used for these experiments. All of the computation involved is performed by an NVIDIA Jetson Xavier, which can be seen in Fig. \ref{fig:huskysetup} mounted on the back-end of the Husky's payload platform.
Each data sample consists of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r\rangle$ of timestamped images from left, center, and right cameras respectively, paired with their corresponding expert control---in this case the yaw rate of the robot, which is recorded from the joystick controller command issued to drive the robot around. The control correction for the left and right camera images are calculated as depicted in Table \ref{tab:husky-corrections}.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/path2.png}
\caption{A bird's eye view of the testing path for the Husky experiment. Shows the interventions needed while navigating for each policy. See~\ref{huskres} for details.}
\label{fig:huskypath}
\vspace{-0.6cm}
\end{figure}
Fig. \ref{fig:huskypath} shows the overhead view of the test path used for the experiment. The path is highlighted in red. Three different models are trained to drive the robot along the path - a model trained using data from a single camera, a model trained using data from a single camera plus equivariant features, and a model trained using data from all three cameras. Once trained, each model's performance navigating the path is then evaluated.
\subsubsection{Network}
We use the same setup as described previously for the Carla simulation experiment. A CNN as the encoding network $z_\theta(\cdot)$ and a separate FCN for each of the equivariant maps $\langle M_{c \rightarrow l}, M_{c \rightarrow r}\rangle$.
\subsubsection{Results} \label{huskres}
As hypothesized, the model trained using only center camera data performs worse than the others, steering off the path at several points. The model trained on data generated using equivariant maps as well as the model trained using all three cameras both perform significantly better. Fig. \ref{fig:huskypath} shows a snapshot of the entire test run for three setups respectively. The locations where an intervention occurred are denoted by a yellow cross. Both the equivariant feature model and the all camera model fail once at roughly the same location at a turn while the center camera model fails at 9 different locations. The video from this experiment can be found here: \href{https://youtu.be/5g4Kg3-YWvA}{https://youtu.be/5g4Kg3-YWvA}.
\subsection{Drone Racing}
\subsubsection{Experimental Setup}
In this experiment, we showcase the effectiveness of equivariant features in a flying vehicle in the Drone-Racing environment developed by Microsoft \cite{droneracing}. The environment used is an outdoor environment where the drone navigates through a trail with rectangular gates. The aim is to navigate the path while passing through the center of the gates.
The network architecture is the same as described in Section \ref{droneExp}.
\subsubsection{Results} \label{droneres}
Again, we train three separate models, using only the center camera data, center camera data plus equivariant augmentations, and all camera data, each of them trained using same amount of expert runs and compare their performance through number of expert interventions required when completing the path. We run DAgger iterations and observe the performance improvements with respect to DAgger iterations.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\columnwidth]{figures/drone-racing-interventions.png}
\caption{Number of interventions vs DAgger iterations for the flying vehicle in drone-racing environment. Model trained using equivariant augmentations lies in between the model trained using a single cameras and the model trained using all five cameras. See ~\ref{droneres} for details.}
\label{fig:droneracingDAgger}
\vspace{-0.6cm}
\end{figure}
The model trained using data from all 5 cameras completes the path without needing any interventions. Fig. \ref{fig:droneracingDAgger} shows the results for this experiment. It is observed that the number of interventions required for the model trained using equivariant augmentations lies between the other two models. This again showcases the effectiveness of augmentations produced using equivariant features in improving imitation learning performance.
\subsection{Learnt vs Non-Learnt Equivariance Maps}
These experiments examine the significance of learned equivariance maps as opposed to other types of maps toward improving imitation learning performance. In each of the experiments thus far, we modeled the equivariance relation using a neural network that is optimized using expert data. We investigate the effectiveness of this learned equivariance mapping by comparing it to two different types of maps - first, a completely random map that does not enforce any equivariance relation. And second, a deterministic map that tries to enforce a deterministic constraint in the embedding space relative to input image viewpoint. The following sections describe these setups in detail. Our results indicate that both alternative techniques fall short of achieving the performance improvements gained using an equivariance map.
\subsubsection{Random Equivariance Maps} \label{randomEqMap}
We use a randomly initialized fully connected network for the equivariance map $M_{i \rightarrow j}$. This random equivariance map is then used to train an imitation learning policy for driving the car in Carla Simulation as described in section \ref{carExp}. Fig. ~\ref{fig:detmap_cte} shows the imitation learning performance with respect to DAgger iterations using this setup. As expected, a learned equivariance map performs much better compared to a random map. This illustrates the role of equivariance towards improving representation for imitation learning.
\subsubsection{Deterministic Equivariance Maps} \label{detEqMap}
This experiment uses a pre-defined equivariance map, essentially a linear translation in the embedding space. This map is used to train an imitation learning policy as before, and compare its performance with learned equivariance map. The experiment setup here is analogous to the CARLA simulation setup used previously. The latent space transformation used is defined in Table \ref{tab:detMapDef}. Fig.~\ref{fig:detmap_cte} shows the imitation learning performance with respect to DAgger iterations using this setup. The learned equivariant mapping significantly outperforms this deterministic mapping.
\begin{table}[!t]
\caption{Deterministic mapping used in the embedding space between different input viewpoints.}
\begin{center}
\begin{tabular}{|c|c|}
\hline
\textbf{Mapping} & \textbf{Transformation} \\
\hline
$M_{c \rightarrow l}$ & $z_{cen}+0.25$ \\
\hline
$M_{c \rightarrow r}$ & $z_{cen}-0.25$ \\
\hline
\end{tabular}
\label{tab:detMapDef}
\end{center}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/detEqCTEshrunk.png}
\caption{Cross track error(m) for driving policy trained using a deterministic map and a learned map enforcing equivariant features.}
\label{fig:detmap_cte}
\vspace{-0.6cm}
\end{figure}
\section{CONCLUSION}
In this paper, we investigate the use of equivariant maps for visual navigation by mobile robots. By training a neural network model that learns image features which are equivariant, we can predict the latent representations of images from viewpoints nearby to the one observed. Our results indicate that by augmenting the training dataset with these representations, one can significantly improve navigation performance in a variety of settings as demonstrated by our simulation experiments involving flying and ground vehicles.
Additionally, through our ground robot experiments over a 500m path, we showed that our method also transfers over to the real world.
\section{Introduction}
This document is a model and instructions for \LaTeX.
Please observe the conference page limits.
\section{Related Works}
- Imitation learning works \\
- Equivariance related works \\
- Equivariant representation learning - Dinesh Jayaraman's work\\
\section{Proposed Approach}
Our goal is to learn an image representation or ``embedding" that is equivariant with respect to relative camera pose. Let $x_i$ $\epsilon$ $X$ be an image in the pixel space, and let $y_i$ $\epsilon$ $Y$ be the associated relative camera pose. The relative camera pose could be a change in position and/or orientation of the given camera with respect to a reference camera.
As input to our representation learning algorithm, we have training set $U$ of $N_u$ image pairs and their associated camera poses, $U = \{\langle(x_i, x_j), (y_i, y_j)\rangle\}_{(i,j)=1}^{N_u}$. The image pairs originate from video sequences and are adjacent frames in time.
The learned representations are then used to augment the training of an imitation learning policy to control a ground vehicle as well as a flying vehicle. In the following, we first explain in Sec. \ref{cameraImgPairs} how to obtain the camera image pairs. Then, Sec. \ref{eqLearning} presents the equivariant feature learning procedure and Sec. \ref{eqObjective} introduces the corresponding objective function. In Sec. \ref{mappingFun} we showcase how a multilayer perceptron is used as the mapping function to achieve equivariance in the learned latent space or ``feature space". Finally, Sec. \ref{bcTraining} discusses training an imitation learning policy to learn control actions.
\subsection{Generating Camera Image Pairs} \label{cameraImgPairs}
First, we obtain temporally coherent image frame pairs generated from different camera poses. These image pairs are obtained by setting up cameras at varying poses capturing distinct overlapping viewpoints.
In this paper, an image set refers to a set of images of the same scene obtained from different camera viewpoints.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/aerialVehicleCapture.png}
\caption{Image set obtained from an aerial vehicle using five different cameras located 50 cm apart from the center camera in the corresponding directions.}
\label{fig:imageAerial}
\end{figure}
Fig. \ref{fig:imageAerial}. shows a sample set obtained for an aerial vehicle using five distinct cameras with a fixed orientation but distinct relative positions in simulation. We obtain a dataset in simulation for two separate settings - a ground vehicle and an aerial vehicle, by setting up cameras at different relative positions.
\begin{figure*}[h]
\centering
\includegraphics[width=\textwidth]{figures/train_wmath.png}
\caption{Training Setup: Siamese encoder decoder network with shared weights for all inputs in the set. A mapping network $M_{i:j}$ exists for each input pair that enforces the equivariance relation in the latent state.}
\label{fig:eqSetup}
\end{figure*}
\begin{figure*}[h]
\centering
\includesvg[width=\textwidth]{figures/train_wmath.svg}
\caption{Training Setup: Siamese encoder decoder network with shared weights for all inputs in the set. A mapping network $M_{i:j}$ exists for each input pair that enforces the equivariance relation in the latent state.}
\label{fig:eqSetup}
\end{figure*}
\subsection{Equivariant Feature Learning} \label{eqLearning}
Given an image set $\textbf{I} = \langle I_1, I_2, ..., I_n \rangle$, our model learns a function $z_\theta(\cdot)$, parametrized by $\theta$, that maps each element of \textbf{I} to a $D$ dimensional feature space $Z$, i.e., $z_\theta (\cdot) : I_i \to \mathbb{R}^D$.
Additionally, the feature space $Z$ exhibits equivariance, i.e. the image embeddings in the latent space behave symmetrically and predictably with respect to transformations applied to the input image $I_i$:
\begin{equation}
\forall{I} \in \textbf{I} : z_\theta (g(I)) \approx M_g (z_\theta(I))
\end{equation}
Here $g$ is a transformation applied to the input image $I$ in the pixel space; and $M_g$ defines a corresponding transformation in the latent space. For example, suppose the images $I_c$ and $I_l$ correspond to the center and left camera images in Fig. \ref{fig:imageAerial}, respectively. If $g_{c:l}$ is a transformation such that \mbox{$g_{c:l}(I_c) = I_l$}, then $M_{g_{c:l}}$ defines a transformation in the latent space such that $ M_{g_{c:l}}(Z_c) = Z\sp{\prime}_{{c:l}} \approx Z_l$, i.e. the embedding $Z_c$ derived from $I_c$, when transformed using $M_{g_{c:l}}$, approximates the embedding $Z_l$ derived from $I_l$. In the latent space, $Z_c$ and $Z_l$ maintain a relation directly corresponding to the one between $I_c$ and $I_l$ in the pixel space.
\subsection{Equivariant Feature Learning Objective} \label{eqObjective}
Now, we design an objective function that encourages the learned feature space $Z$ to exhibit equivariance with respect to camera transformations, or camera viewpoints in our case.
Given image set $\textbf{I} = \langle I_1, I_2, ..., I_n\rangle$ where $I_i$ is an image taken from the $i^{th}$ viewpoint, and the corresponding embeddings $\textbf{Z} = \langle Z_1, Z_2, ..., Z_n \rangle$, we setup an objective function to enforce equivariance between any two image viewpoints, as follows:
\begin{equation}
L_{eq_{i:j}} = ||Z_j - M_{i:j}(Z_i)||^2
\label{eqloss}
\end{equation}
By minimizing the L2 norm between learned and mapped latent representations, the function forces the mapping $M_{j:i}$ from the $j^{th}$ to the $i^{th}$ embedding to preserve equivariant relations.
The overall equivariance learning objective uses the average loss over a large batch of embeddings in the latent space:
\begin{equation}
L_{eq} = \textstyle \sum\limits_{i} \sum\limits_{j} ||Z_j - M_{i:j}(Z_i)||^2
\end{equation}
\subsection{Feature Mapping Function} \label{mappingFun}
In addition to the parameterized function $z_\theta (\cdot)$ that maps images to embeddings in the feature space; we also use a mapping function $M_{i:j}$ that maps embeddings in the feature space from one viewpoint to another.
Fig. \ref{fig:eqSetup}. shows our ``Siamese Encoder-Decoder" \cite{NIPS1993_769} network setup and our mapping network.
The first network is a convolutional neural network with weights $\theta$. that models $z_\theta (\cdot)$ and maps or `encodes' the input image $I_i$ to a one-dimensional latent representation $Z_i$.
Then, we have an equivariant mapping $M_{i:j}$ for each viewpoint transformation $i$:$j$ which is modeled by a corresponding fully connected network $M_{\phi_{i:j}}$, with weights $\phi_{i:j}$. Given an image embedding in the latent space $Z_i$ from the $i^{th}$ viewpoint, $M_{\phi_{i:j}}$ attempts to map it to the $j^{th}$ viewpoint, \mbox{i.e. $M_{i:j} (Z_i) = Z\sp{\prime}_{i:j} \approx Z_{j} $.}
The weights $\theta$ as well as $\phi_{i:j}$ are optimized together as a shared objective so our model learns both the feature and equivariant maps, respectively.
The latent representations $Z_i$ are then ``decoded'' back to images $ \hat{I}_i $ by another convolutional network that attempts to model $z_{\theta}^{\text{-}1} (\cdot)$. It is defined as $p_\varphi (\cdot)$ with weights $\varphi$, and it reconstructs the image from the latent space, \mbox{ i.e. $ p_\varphi(Z_i) = \hat{I}_i \approx I_i $.} The output from the decoder is then used to compute and minimize the reconstruction loss $L_{rc} $:
\begin{equation}
L_{rc} = \textstyle \sum\limits_{i} ||I_i - \hat{I_i}||^2
\label{reconLoss}
\end{equation}
To further reinforce the effectiveness of the equivariance maps $M_{i:j}$, we enforce an additional reconstruction penalty between the image ${\widehat{I\sp\prime_{i:j}}}$ reconstructed from the ``mapped" embedding i.e. $\widehat{I\sp\prime_{i:j}} = p_\varphi(M_{i:j} (z_\theta (I_i)))$ and the original image $I_j$ from that viewpoint:
\begin{equation}
L_{eq\textrm{-}rc} = \textstyle \sum\limits_{i} \sum\limits_{j} ||I_j - \widehat{I\sp\prime_{i:j}}||^2
\label{reconLoss}
\end{equation}
The overall loss function is combination of the reconstruction, equivavriant reconstruction, as well as the equivariant loss as follows:
\begin{equation}
L = L_{rc} + L_{eq} + L_{eq\textrm{-}rc}
\label{loss}
\end{equation}
\subsection{Imitation Learning with learned Features} \label{bcTraining}
The learned feature map $Z_\theta(\cdot)$ and equivariance map $M_{i:j}$ are used to train an imitation learning policy $\pi_\Theta (a|I,\Theta)$. As showcased in \cite{e2eselfdriving}, an imitation learning policy can be enhanced by adding additional images action pairs $\langle I_j, a_j \rangle$ obtained by perturbing given image action pair $\langle I_i, a_i \rangle$.
The feature map $Z_\theta(\cdot)$ and equivariance map $M_{i:j}$ are used to obtain the approximate representation $Z_j$ from given image $I_i$. These additional representations provide augmentations for the training policy. The objective function used to train the policy is mean squared error.
\begin{equation}
L_{bc} = \frac{1}{n} \sum_{i=1}^n ||a_i - \pi (I_i|\Theta)||^2
\label{loss}
\end{equation}
\section{Experiments}
We validate our approach in two different experimental settings - an aerial vehicle and a ground vehicle. The experiments involve the following:
\begin{enumerate}
\item Collecting Data: The dataset is collected in simulation. Each data sample $D = \langle \textit{\textbf{I}}, \textit{\textbf{A}} \rangle$ consists of image set $\textit{\textbf{I}}$ that comprises of images from different camera viewpoints and an action set $\textit{\textbf{A}}$ that contains the corresponding control actions.
\item Learning Equivariant Features: We learn the feature mapping function $z_\theta(\cdot)$ and equivariant mappings $M_{i:j}$ using image sets $\textit{\textbf{I}}$ in the dataset.
\item Learning Control via Imitation Learning: We learn to control the vehicle through an imitation learning policy that is trained using both image sets \textit{\textbf{I}} and action sets \textit{\textbf{A}}.
\end{enumerate}
We perform the following three tests to assess the effectiveness of our learned equivariant representations and equivariant mappings:
\begin{enumerate}
\item Train and evaluate a behaviour cloning policy with data from a single camera. This policy is trained end to end to learn control from image.
\item Train and evaluate a behaviour cloning policy with data from multiple cameras providing additional augmentation. This policy is also trained end to end to learn control from image.
\item Train and evaluate a behaviour cloning policy from data from single camera along with augmentations applied in feature space using learned equivariant feature mappings. This policy uses the locked feature mapping function $z_\theta(\cdot)$ as the feature extractor and optimizes the fully connected layers to learn the control from equivariant representations.
\end{enumerate}
The end to end imitation learning policy is trained in the same way as \cite{e2eselfdriving}. The network contains a set of convolutional layers followed by a set of fully connected layers. For consistency in comparisons, we ensure that the convolutional part of the imitation learning policy has the same architecture as the feature mapping function $z_\theta(\cdot)$. All networks are optmized using ADAM optimizer \cite{adam} with a learning rate of $10^{-4}$ and a batch size of $64$. The evaluation of the policy is done by accessing the control of vehicle using a single camera input as done in \cite{e2eselfdriving}. We setup dagger \cite{dagger} to investigate the performance of the control policy with respect to dagger iterations. We hypothesize that policy trianed with multiple cameras should perform better than policy trained with single camera + equivariant augmentations which should perform better than the policy trained with a single camera. As we do more dagger iterations, the different will become less and less pronounced since dagger helps in obtaining necessary data to make the policy more robust. We use number of interventions and cross track error (deviation from the expert path) as the metric to compare the performance. The experiment setup details and results are discussed in Sec. \ref{carExp} and Sec. \ref{droneExp}.
\subsection{Quadrotor Simulation Experiments} \label{droneExp}
\subsubsection{Experimental Setup}
We use a flying vehicle in the airsim simulator \cite{airsim2017fsr}. The vehicle is setup with five front facing cameras as seen in figure \ref{fig:imageAerial}. The dataset used in these experiments comprises of image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r, I_t, T_b \rangle$ corresponding to images from left, center, right, top, and bottom cameras. These cameras are located at a distance of $50 cm$ each from the center camera in the corresponding directions.
The action set $\textit{\textbf{A}} = \langle \delta yaw, \delta z \rangle$, comprises of the relative change in yaw and $z$ between consecutive frames. It is worth noting that the flying vehicle in this setup is non-holonomic capable of moving in forward direction only while changing the yaw and $z$.
\subsubsection{Network}
The equivariant feature learning is setup with feature mapping network $z_\theta(\cdot)$ as a CNN with architecture: 24 conv(7x7)-stride(2)-ReLU $\to$ 36 conv(7x7)-stride(2)-ReLU $\to$ 48 conv(7x7)-stride(2)-ReLU $\to$ 64 conv(5x5)-stride(1)-ReLU $\to$ $D$ = 2816 full feature units. The parenthesis after conv indicate the size of convolutional filters and those after stride represent stride length.
The equivariance maps $\langle M_{c:l}, M_{c:l}, M_{c:l}, M_{c:l} \rangle$ are fully connected networks with architecture: linear(2816,128)-ReLU $\to$ linear(128,2816).
The imitation learning network fully connected part has the architecture: linear(2816,128)-ReLU $\to$ linear(128,8)-ReLU $\to$ linear(8,2).
\subsubsection{Results}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/cte_bar-drone.png}
\caption{Cross track error (m) vs dagger iterations for flying vehicle. It can be seen that network trained with equivariant augmentation lies in between the network trained using center camera and all cameras.}
\label{fig:cte_drone}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/refvsflowntraj-paper.png}
\caption{Reference vs Flown Trajectory. All cameras model (right) tracks the trajectory the best, followed by equivariant augmentation model(center), followed by center camera model(left).}
\label{fig:trajs}
\end{figure}
Figure \ref{fig:cte_drone} shows the cross track error vs dagger iterations for three different test scenarios. As hypothesized we see that using augmentation produced via equivariant representations help to improve the imitation performance. Figure \ref{fig:trajs} overlays the trajectory flown on top on reference trajectory. It can be seen that policy trained with equivariant augmentations is able to track better than a policy trained by center camera only. Note that missing parts in flown trajectory signify an intervention where the expert had to take over.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/frozenEqNet-paper.png}
\caption{Autonomy, Number of Interventions, and Cross Track error comparison between policies trained with data from center camera, equivariant augmentations, and all cameras. Equivariant augmentations in this case are produced by a network trained on a data that is disjoint from the data used to train imitation policy showcasing the ability to generalize.}
\label{fig:frozenEq}
\end{figure}
We also perform a test by learning equivariant representation and mapping networks using a dataset that is disjoint from that used to train the control policy. We observe that equivariant augmentations still help by improving the control performance showcasing the ability to generalize. Figure \ref{fig:frozenEq} displays the result of this experiment.
\subsection{Driving Simulation Experiments} \label{carExp}
\subsubsection{Experimental Setup}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/groundVehicle.png}
\caption{Image set obtained from ground vehicle in carla simulator. The left and right cameras are 25 cm apart from the center camera.}
\label{fig:carlasetup}
\end{figure}
We use the Carla simulator \cite{carla} for our experiments. The vehicle is setup with three front facing cameras producing images as seen in figure \ref{fig:carlasetup}. The dataset for this comprises of image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r\rangle$ corresponding to images from left, center, and right cameras respectively. The cameras are located at a distance of $25 cm$ from the center camera in the corresponding directions. The action set $\textit{\textbf{A}} = \langle \alpha \rangle$ consists of expert steering angles collected using carla autopilot.
\subsubsection{Network}
For equivariant feature learning, feature mapping network $z_\theta(\cdot)$ has architecture: 24 conv(5x5)-stride(2)-ELU $\to$ 36 conv(5x5)-stride(2)-ELU $\to$ 48 conv(5x5)-stride(2)-ELU $\to$ 64 conv(3x3)-stride(1)-ELU-dropout(p=0.5) $\to$ $D = 3840$ full feature units. The equivariance maps $\langle M_{c:l}, M_{c:r} \rangle$ are fully connected networks with architecture: linear(3840, 128)-ELU $\to$ linear(128, 3840).
The imitation learning network's conv part is same architecture as $z_\theta$. The fully connected part has architecture: linear(3840, 100)-ReLU-dropout(p=0.5) $\to$ linear(100,10)-ReLU $\to$ linear(10,1).
\subsubsection{Results}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/cte_bar-car.png}
\caption{Cross track error(m) vs dagger iterations for ground vehicle. I can be seen that network trained with equivariant augmentation lies mostly in between the network trained using center and all cameras.}
\label{fig:cte_car}
\end{figure}
Figure \ref{fig:cte_car} shows the cross track error vs dagger iterations for three different test scenarios. Again we can see that driving policy trained using equivariant augmentations performs better than policy trained with just center camera. With increasing dagger iterations, the effect is seen to subside as expected.
\section{Conclusion}
\section*{Acknowledgment}
\bibliographystyle{./bibliography/IEEEtran}
\section{INTRODUCTION}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{figures/final9.jpg}
\vspace{-0.3cm}
\caption{An overview of our system. Images from the center camera are embedded into a low-dimensional representation, $Z_c$, which is deterministically transformed into the predicted embeddings of the left and right viewpoints respectively, via equivariant mappings $M$. Both the actual steering command (solid red arrow) and the predicted steering commands (dashed red arrows) are included in the augmented training set on which the policy is trained. At test time, the policy is executed on embeddings from the center camera. The maps $M$ are trained separately from $\pi$.}
\label{fig:first_page}
\vspace{-0.5cm}
\end{figure}
From autonomous driving to flying and manipulation, vision-based policies learned through imitation are starting to be deployed on robot systems of critical importance. To ensure safety during deployment, we would ideally like to have formal guarantees about the generalization and robustness properties of these policies. While desirable, however, formal guarantees for out-of-distribution generalization have been attained on perturbation models that are theoretically convenient, but limited in practice~\cite{pmlr-v97-cohen19c}, as they often do not capture the full range of plausible unseen states. Data augmentation during training is significant in practice.
In this paper we address this problem by proposing a dataset augmentation scheme for training policies whose inputs are image embeddings, as shown in Fig.~\ref{fig:first_page}.
Given an observed image from the center camera of a vehicle, our method allows us to learn a map that predicts the embeddings of corresponding nearby viewpoints, as well as the actions that would have needed to be taken if the vehicle's camera was at those viewpoints. Our method relies on predicting embeddings of nearby viewpoints using \emph{equivariant mappings}~\cite{Jayaraman_2015_ICCV, lenc2015understanding} that are trained to transform the embedding of the center camera to the embeddings of nearby viewpoints, both for ground vehicles and for flying vehicles, as we will show in the evaluation section. Equivariance has emerged as an important structural bias for neural network design in the last few years. Here we use it as an auxiliary loss to the main behavioral cloning loss.
Our main contribution is showing that augmenting the training dataset with predicted embeddings of nearby viewpoints, using learned equivariant maps, increases the robustness of the vision-based policy. This results in lower cross track error and fewer human interventions needed for navigation tasks, both for ground vehicle and aerial vehicle navigation tasks in 2D and 3D respectively. We demonstrate that our method applies to real world visual navigation scenarios by deploying it on a terrestrial mobile robot, resulting in significantly improved navigation performance.
\section{RELATED WORK}
\textbf{Equivariant representations:} Equivariance for image representations captures relationships between encodings of two related images that are determined by the transformation of those images~\cite{lenc2015understanding}. If the encodings are the same, then they are invariant to that transformation. In our work, we want the representation of the left camera to be predictable by the center camera, so we want a high degree of equivariance. Feature equivariance and invariance in terms of rotation transformations was studied in~\cite{equivariant_rotations, Zhao2019QuaternionEC, lifd_survey}, as well as in the context of probabilistic models~\cite{dbn_equivariant, Bilovs2020EquivariantNF, Rezende2019EquivariantHF}. The utility of equivariant representations for multiple downstream tasks, learned by egomotion was recognized in~\cite{Jayaraman_2015_ICCV}. Our work here is related the most to this paper.
Equivariant representations have also been useful for 3D volumetric data~\cite{Weiler20183DSC}, even in the fully unsupervised setting~\cite{Spezialetti2019LearningAE}, addressing equivariance and invariance to both rotation and translation for 3D data~\cite{Fuchs2020SE3Transformers3R, Wang2020EquivariantMF}.
Additionally, instead of imposing equivariance through a loss function, as in~\cite{Jayaraman_2015_ICCV}, many recent works build equivariance in the structure of the network~\cite{cubenet, worrall2017harmonic, cohen2019gauge, Tai2019EquivariantTN, Cohen2019AGT, steerable_filters_equivariant_cnn}, for example in the multi-view setting~\cite{emvn, multiviews, gvcnn, qi2016volumetric, you2018pvnet}. Finally, equivariance has also been used to improve state representations for planning in MDPs~\cite{Pol2020PlannableAT}.
\textbf{Automated data augmentations:} Typical dataset augmentations for supervised deep learning have included addition of noise, rotations, crops, blurring, and others. While this is usually a set of manually hand-crafted augmentation schemes, there has been increased interest in automatically computing data augmentations~\cite{ratner2017learning, zhang2019adversarial, dao2019kernel, raileanu2020automatic} to increase the robustness of classifiers, regressors, and RL policies. Augmentations have also been used in monocular depth estimation~\cite{ravi_unsupervised_depth}.
\textbf{Robustness in behavioral cloning:}
Behavioral cloning techniques that learn policies through pure supervised learning from a given dataset of state-action pairs usually suffer from covariate shift~\cite{Loquercio_2018,lecun_obstacle_avoidance}. Dean Pomerleau's PhD thesis~\cite{pomerleau_thesis} and pioneering work on ALVINN~\cite{alvinn} details a data augmentation scheme that injects synthetic images of the road ahead from viewpoints nearby the camera to the training set. Our paper provides an alternative that is based on neural network view synthesis, which in the last years has seen significant progress. Pomerleau's thesis also identified issues of covariate shift, which methods such as DAgger~\cite{DAgger_ross} address by iteratively labeling the newly visited states after each policy evaluation. The question of dataset augmentation in one round of behavioral cloning, however, remains. Collecting large driving datasets, in a way that no manual annotation is required from the user, is critical for autonomous visual navigation.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/train_smath.jpg}
\vspace{-0.4cm}
\caption{Representation Learning Phase: we use an encoder-decoder network to learn image embeddings that are equivariant to input camera viewpoint. Mapping network $M_{i \rightarrow j}$ is learned to map these embeddings from viewpoint $i$ to viewpoint $j$. The encoder-decoder and the mapping networks are trained collectively by optimizing a combined loss function shown in Eqn. \ref{eqn:repr_loss}. See~\ref{methods} for details.}
\label{fig:eqSetup}
\end{figure*}
One example of this was in UAV navigation in forest trails~\cite{trail_following}. The images of the center camera are automatically annotated with the forward steering command, the images of the left camera with the right, and the images of the right camera are annotated with the left steering command. A similar hardware setup and automated data collection was used by the NVIDIA self-driving car team~\cite{nvidia_driving}, as well as more recent work on behavior cloning conditioned on human commands (e.g. turn left at the next intersection)~\cite{codevilla_2017}.
The crux behind the idea of the 3-camera setup is that behavioral cloning can become more robust if we obtain expert actions from states that are nearby the forward-facing one.
Our method builds upon this approach by replacing the need for additional sensors with a model that predicts their embeddings, whose corresponding actions are automatically annotated. We note that in other similar approaches, such as DART~\cite{dart_laskey} an expert demonstrator will have to manually annotate the additional states. In DART, Gaussian noise is added to the original training dataset in order to explicitly include nearby labeled states in the augmented dataset, together with manual annotations.
\section{METHODOLOGY} \label{methods}
Our goal is to learn an encoding method that satisfies equivariance constraints with respect to camera poses or viewpoints nearby the original camera. This allows us to predict the embeddings of nearby viewpoints from the observations of the center camera.
Here, nearby viewpoints refer to horizontal and vertical translation of the camera over short distances capturing the same scene. Our case involves translations of 0.25 to 0.5 meters.
Given an image $I_i$ of a scene taken from viewpoint $i$, we compute its corresponding embedding in the latent space $Z_i$, and use an equivariant map $M_{i \rightarrow j}$ to transform it to the embedding corresponding to
another desired nearby viewpoint $j$. This predicted embedding $Z\sp{\prime}_j$ can then be used as
augmented training data for policies trained on such embeddings and actions.
We train an image encoder network on a dataset of image pairs, where each pair consists of images of the same scene from neighboring viewpoints. In our case, the images are obtained from synchronized videos recorded by a set of cameras mounted on a vehicle visually navigating through an environment. The relative poses of the cameras are assumed to be known and looking at the same scene. With each image pair we simultaneously also record the corresponding expert control command used to navigate the vehicle at that time.
We then use the learned embeddings to train policies using imitation learning.
\subsection{Equivariant Feature Learning} \label{eqLearning}
Given a tuple of images $\textbf{I} = \langle I_1, I_2, ..., I_n \rangle$ that capture the scene at the same timestep, our model learns an encoder $z_\theta (\cdot): I_i \to \mathbb{R}^D$, parametrized by $\theta$, that maps an image to an embedding $Z_i$. We want the embeddings to exhibit equivariance, i.e. change predictably with respect to transformations applied to the viewpoint $i$ of the input image $I_i$. Let $g$ be a transformation applied to the input image $I$ in the pixel space, and let $M_g$ define a corresponding transformation in the latent space. Then:
\begin{equation}
z_\theta (g(I)) \approx M_g (z_\theta(I))
\end{equation}
For example, suppose the images $I_c$ and $I_l$ correspond to the center and left camera images in Fig.~\ref{fig:first_page}, respectively. If $g$ is an image transformation such that \mbox{$g(I_c) = I_l$}, then $M_{g}=M_{c \rightarrow l}$ defines a transformation in the latent space such that $ M_{c \rightarrow l}(Z_c) = Z\sp{\prime}_{{ l}} \approx Z_l$. In latent space, we want $Z_c$ and $Z_l$ to maintain a geometric relation directly corresponding to the one between images $I_c$ and $I_l$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/control_smath.jpeg}
\vspace{-0.3cm}
\caption{Policy Training Phase: The center image is passed to an encoder, then the resulting embedding is mapped to nearby viewpoints (for experiments using an equivariant map), and finally the embeddings are passed to a FCN model to produce steering output. Note that the encoder and FCN models are trained \textit{separately}. See~\ref{bcTraining} for details.}
\label{fig:controlsetup}
\vspace{-0.5cm}
\end{figure}
\subsection{Equivariant Map Learning as an Auxiliary Task} \label{eqObjective}
Now, we design an objective function that encourages the learned embeddings to exhibit equivariance with respect to camera transformations, or camera viewpoints in our case. Given a tuple of images $\textbf{I} = \langle I_1, I_2, ..., I_n\rangle$ captured at the same time, where image $I_i$ is taken from the $i^{th}$ viewpoint, and the corresponding embeddings $z_{\theta}(I_i)$, we set up the following objective function to enforce equivariance between any two image viewpoints.
\begin{equation}
L_{eq}(\theta, M) = \textstyle \sum\limits_{i} \sum\limits_{j} ||z_{\theta}(I_j) - M_{i \rightarrow j}(z_{\theta}(I_i))||^2
\label{eqn:leq}
\end{equation}
\subsection{Image Reconstruction as an Auxiliary Task} \label{mappingFun}
We make use of image reconstruction as an auxiliary task, so that the embeddings capture the information needed to reconstruct images across all different viewpoints. We jointly train an encoder-decoder model for each viewpoint, as shown in Fig.~\ref{fig:eqSetup}. The encoder, decoder, as well as the equivariant mapping $M_{i \rightarrow j}$ are optimized together in a shared objective.
The latent representations $Z_i$ are decoded back to images by a decoder $p_\varphi (\cdot)$. The output of the decoder is then used to compute and minimize the autoencoding loss $L_{rc}$ from the same viewpoint:
\begin{equation}
L_{rc}(\theta, \varphi) = \textstyle \sum\limits_{i} ||I_i - p_{\varphi}(z_{\theta}(I_i))||^2
\label{reconLoss}
\end{equation}
\noindent To further reinforce the effectiveness of the equivariant maps $ M_{i \rightarrow j}$, we enforce an additional reconstruction penalty between pairs of viewpoints:
\begin{equation}
L_{eq\textrm{-}rc}(\theta, \varphi, M) = \textstyle \sum\limits_{i} \sum\limits_{j} ||I_j - p_{\varphi}(M_{i \rightarrow j}(z_{\theta}(I_i))) ||^2
\label{eqReconLoss}
\end{equation}
The overall representation learning loss is a combination of the reconstruction, equivariant reconstruction, as well as the equivariant loss:
\begin{equation}
\begin{multlined}
L(\theta, \varphi, M) = \\ \lambda_1 L_{rc}(\theta, \varphi) + \lambda_2 L_{eq}(\theta, M) + \lambda_3 L_{eq\textrm{-}rc}(\theta, \varphi, M)
\label{eqn:repr_loss}
\end{multlined}
\end{equation}
\noindent We backprop through the combined loss in Eqn.~\ref{eqn:repr_loss} to optimize $\theta, \varphi$, and $M$. The hyperparameters are chosen as $\lambda_1=1$, $\lambda_2=10$, and $\lambda_3=1$ to ensure all loss terms have relatively same scale.
\subsection{Imitation Learning with learned Features} \label{bcTraining}
After we learn $z_\theta (\cdot)$ using the representation learning loss in Eqn.~\ref{eqn:repr_loss}, we fix the encoder weights and learn a control policy, i.e. a multi-layer perceptron representing $\pi (Z)$, via imitation learning. The control policy $\pi (\cdot)$ is trained on embedding-action pairs both from the center camera (observed embeddings and actions) and from the nearby viewpoints (predicted embeddings from equivariance relations and predicted actions). The objective function used is the following:
\begin{equation}
L(\pi) = \textstyle \sum\limits_{i} ||a_i - \pi(Z_i)||^2
\label{eqn:policy_loss}
\end{equation}
At test time and deployment, the encoder is used to convert the images to embeddings which are then used as input to the learned policy, as shown in Fig.~\ref{fig:eqSetup}.
\section{Experiments}
We validate our approach in two different experimental settings, an aerial vehicle and a terrestrial vehicle. The experiments involve the following steps:
\textbf{Data Collection.} Each data sample $\langle \textit{\textbf{I}}, \textit{\textbf{A}} \rangle$ consists of an image tuple $\textit{\textbf{I}}$ that consists of images from different camera viewpoints and an action set $\textit{\textbf{A}}$ that contains the corresponding control actions. We collect data both in simulation and in the real world, for the respective experiments.
\textbf{Learning Equivariant Features.} We learn the feature mapping function $z_\theta(\cdot)$ and equivariant mappings $M_{i \rightarrow j}$ using the collected image sets $\textit{\textbf{I}}$ in the original training dataset.
\textbf{Policy Evaluation.} In order to assess the effectiveness of our learned embeddings and equivariant mappings we evaluate and compare three imitation learning policies that each learn to map images to control actions:
\begin{enumerate}
\item The first policy is trained on image-action pairs obtained from a single camera, in our case the center camera mounted on the vehicle.
\item The second policy is trained on image-action pairs obtained from all the camera sensors mounted on the vehicle. In our case it is 3 viewpoints for the car and 5 for the flying vehicle. The recorded expert control is modified to account for the shifted camera viewpoints for the off center cameras.
\item The third policy uses our equivariant augmentations and is trained on image-action pairs obtained from the center camera and predicted nearby embeddings. The center camere embedding is mapped to the embeddings of the other viewpoints in the latent space. This corresponds to the training scheme described in Sec.~\ref{mappingFun}.
\end{enumerate}
\textbf{Evaluation Metrics.}
To compare the three policies, we record the number of interventions (which occur when the ground vehicle goes off of the current lane or when the flying vehicle deviates more than 2.5 meters away from the goal trajectory), average cross track error (deviation from the specified trajectory in meters), and autonomy.
Autonomy here is calculated as in~\cite{nvidia_driving}, i.e. the percentage of time the car was not being driven by the learned policy due to an intervention. More specifically, for an experiment with $n$ interventions:
\begin{equation}
\textrm{autonomy} = ( 1 - \frac{ n \cdot (6 \textrm{ seconds })}{ \textrm{ elapsed time in seconds} } ) \cdot 100
\label{eqn:policy_loss2}
\end{equation}
The experiment setup details and results are discussed in Sec. \ref{droneExp} and Sec. \ref{carExp}.
\textbf{Policy and Encoder Architectures.} The network architecture used for vehicle navigation was similar to the one used by NVIDIA in \cite{nvidia_driving}, it consists of a CNN encoder network followed by a FCN regressor network. An important distinction between their work and ours is that the entire model in \cite{nvidia_driving} is trained end-to-end, whereas we train the encoder separately from the MLP, as mentioned in Sec~\ref{bcTraining}. For consistent comparisons, we ensure that the encoder $z_{\theta}(I)$ has the same architecture as the image encoder in \cite{nvidia_driving}.
\textbf{Policy Training.} All the networks are optimized using ADAM optimizer~\cite{adam} with a learning rate of $10^{-4}$ and a batch size of $64$. We use DAgger \cite{DAgger} as our imitation learning algorithm and we investigate the performance of the control policy with respect to DAgger iterations. As the DAgger runs increase, the difference in performance between the policies becomes less pronounced as the data aggregation process itself makes all three policies more robust. This provides us with one measure of how effective our embedding method is for data augmentation, particularly in the beginning of imitation.
\subsection{Quadrotor Simulation Experiments} \label{droneExp}
\textbf{Experimental Setup.} For this set of experiments, we use a flying vehicle in the AirSim simulator \cite{airsim2017fsr}, and also a more complex Drone-Racing environment developed by Microsoft \cite{droneracing}, consisting of outdoor three dimensional racing trail marked by rectangular gates. The vehicle is set up with five front facing cameras.
For the dataset used in these experiments, each data sample comprises of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r, I_t, I_b \rangle$ corresponding to timestamped images from the left, center, right, top, and bottom cameras.
The action tuple $\textit{\textbf{A}} = \langle \delta \textrm{yaw}, \delta z \rangle$, is the expert control used to navigate the flying vehicle, applied in terms of the relative change in yaw and $z$ at each timestep. The augmented actions for non-center cameras are computed as shown in Table \ref{tab:flying-corrections}. Note that clockwise rotation is positive $\delta yaw$ and downwards direction is positive $\delta z$.
\begin{table}[!t]
\caption{out-of-distribution experiment results (flying vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Setup} & \textbf{Interventions}& \textbf{Autonomy(\%)}& \makecell{\textbf{Cross Track} \\ \textbf{Error (m)}} \\
\hline
Center Camera & 3 & 84.23 & 2.12 \\
\hline
\makecell{Equiv. Augment. \\Out-Of-Dist \\ Training} & 2 & 89.46 & 1.39 \\
\hline
\makecell{Equiv. Augment. \\ In-Dist \\ Training} & 2 & 89.51 & 1.26 \\
\hline
All Cameras & 0 & 100 & 0.17 \\
\hline
\end{tabular}
\label{tab:eqfrozen-flying}
\end{center}
\vspace*{-0.4cm}
\end{table}
\begin{table}[!t]
\caption{Augmentation label calculations (flying vehicle)}
\begin{center}
\vspace*{-0.2cm}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Center Camera} & \textbf{\textit{$\delta yaw$ $(rad)$}}& \textbf{\textit{$\delta z$ $(m)$}} \\
\hline
Left Camera & $\delta yaw+0.03$ $rad$ & $\delta z$ \\
\hline
Right Camera & $\delta yaw-0.03$ $rad$ & $\delta z$ \\
\hline
Top Camera & $\delta yaw$ & $\delta z + 0.5$$m$ \\
\hline
Bottom Camera & $\delta yaw$ & $\delta z - 0.5m$ \\
\hline
\end{tabular}
\label{tab:flying-corrections}
\end{center}
\vspace*{-0.7cm}
\end{table}
\begin{figure*}[!t]
\hspace{0.8cm}\includegraphics[width=7in, trim={4.5cm 0.0cm 0.0cm 0.5cm},clip]{figures/cte_both.png}
\vspace{-0.7cm}
\caption{Cross track error (m) vs DAgger iterations for flying vehicle (left) and ground vehicle (right). It can be seen that network trained with equivariant augmentation lies in between the network trained using center camera and all cameras in terms of performance.}
\label{fig:cte_drone}
\end{figure*}
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/refvsflowntraj-paper.jpeg}
\caption{Reference trajectory (orange) vs flown trajectory (blue). The all-cameras model (right) tracks the trajectory the best, followed by the equivariant-augmentation model (center), followed by the center-camera model (left). Note that missing sections in the flown trajectory are due to an intervention where expert control was used to navigate the vehicle back to the reference trajectory. See~\ref{droneExp} for details.}
\label{fig:trajs}
\end{figure}
\textbf{Results.} Fig. \ref{fig:cte_drone} (left) shows the cross track error vs DAgger iterations for three different test scenarios. We see that using the augmentations generated via the equivariant maps for training improves the imitation performance. Fig. \ref{fig:trajs} overlays the trajectory flown on top of the reference trajectory. The policy trained with equivariant augmentations is able to track better than a policy trained using only the center-camera data. Additionally, as shown in Fig. \ref{fig:droneracingDAgger}, we observe that the policy trained on our augmentations requires on average 2 fewer interventions when compared to the policy trained without them, for the Drone Racing environment.
\textbf{Out-Of-Distribution Experiments.} We perform another set of experiments where we train the equivariant map on a dataset captured in a separate area of the simulation map, a dense urban environment with buildings and roads, compared to where we test the model, a park with no city structures except a single paved path.
As shown in \mbox{Table \ref{tab:eqfrozen-flying}}, the policy trained using the generated augmentations has a lower cross track error by 0.73 meters when compared to the policy trained directly on the center-camera data.
Furthermore, its cross track error is only 0.13 meters higher than the policy which uses the equivariant map trained on the same area of the simulation map as used for testing.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\columnwidth]{figures/drone-racing-interventions.png}
\caption{Number of interventions vs DAgger iterations for the flying vehicle in drone-racing environment. Model trained using equivariant augmentations lies in between the model trained using a single cameras and the model trained using all five cameras.}
\label{fig:droneracingDAgger}
\end{figure}
\subsection{Driving Simulation Experiments} \label{carExp}
\textbf{Experimental Setup.} For this experiment, we use the Carla simulator \cite{carla}. The driving vehicle is set up with three front facing cameras. Each data sample of the dataset consists of an image set $\textit{\textbf{I}} = \langle I_l, I_c, I_r\rangle$ of timestamped images from the left, center, and right cameras respectively, paired with their corresponding expert control, in this case, the steering angle for the car collected from the Carla driving autopilot. The augmentations for non-center cameras are computed as shown in Table \ref{tab:driving-corrections}.
\begin{table}[!t]
\label{table1}
\caption{out-of-distribution experiment results (ground vehicle)}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Setup} & \textbf{Interventions}& \textbf{Autonomy (\%)}& \makecell{\textbf{Cross Track} \\ \textbf{Error (m)}} \\
\hline
Center Camera & 8 & 80.24 & 4.00 \\
\hline
\makecell{Equiv. Augment. \\Out-Of-Dist\\Training} & 6 & 87.21 & 2.66 \\
\hline
\makecell{Equiv. Augment. \\In-Dist\\Training} & 5 & 88.01 & 2.61 \\
\hline
All Cameras & 4 & 89.97 & 2.25 \\
\hline
\end{tabular}
\label{tab:eqfrozen-ground}
\end{center}
\vspace*{-0.3cm}
\end{table}
\begin{table}[!t]
\caption{Augmentation label calculations (ground vehicle)}
\begin{center}
\vspace*{-0.1cm}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Training Camera} & \textbf{Steering Angle} (normalized $[-1,1]$) \\
\hline
Left Camera & $steering_{cen}+0.05$ \\
\hline
Right Camera & $steering_{cen}-0.05$ \\
\hline
\end{tabular}
\label{tab:driving-corrections}
\end{center}
\vspace*{-0.6cm}
\end{table}
\textbf{Results.} Again we can see in Fig. \ref{fig:cte_drone} (right) that the driving policy trained using equivariant augmentations outperforms the policy trained with just the center camera directly, and that this benefit slowly subsides with more DAgger iterations.
\textbf{Out-Of-Distribution Experiments.}
As per the flying vehicle experiments, we perform an additional experiment to test generalization of the learned features to novel environments. The experiments so far used a specific path in 'Town10' from the Carla simulator \cite{carla} which resembles a densely packed urban environment with lots of city structures. For this experiment however, the equivariance map was trained using a dataset collected on a highway surrounded by green hills and no other city structures. Yet our results as shown in Table \ref{tab:eqfrozen-ground} illustrate the significant performance improvement gained by our generated augmentations over the policy trained without them.
\textbf{Gaussian Noise Augmentation Experiments.}
\label{noisexp}
We also compare our method to simply adding noise augmentations to the camera images. The noise augmentations consist of zero mean gaussian noise with varying standard deviations \mbox{$\sigma = \{0.01, 0.05, 0.1, 0.2, 0.3\}$.} While noise augmentations do improve navigation performance, out method still performs better in most cases as shown in Fig.~\ref{fig:gaussnoise}.
\begin{figure}[ht]
\centering
\hspace*{-0.5cm}\includegraphics[width=3.5in]{figures/sigma_avg_errbars.png}
\vspace*{-0.6cm}
\caption{Average cross track error vs DAgger iterations with center camera network trained with gaussian noise augmentations on the input images.}
\label{fig:gaussnoise}
\vspace{-0.3cm}
\end{figure}
\subsection{Terrestrial Robot Experiments} \label{huskyexp}
\textbf{Experimental Setup.}
\label{husksetup}
These experiments were performed using a Clearpath Husky robot equipped with three front facing cameras. The Husky setup is similar to the driving vehicle in Carla simulation as described in Sec.~\ref{carExp}. Three ZED cameras (used as monocular cameras) are mounted on the front, each 25cm apart. All of the computation was performed on the NVIDIA Jetson Xavier mounted on the bottom. The robot can be seen in Fig.~\ref{fig:huskysetup}.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figures/husky_setup.jpeg}
\vspace{-0.3cm}
\caption{Clearpath Husky used for ground robot experiments with three front facing cameras mounted in the front.}
\label{fig:huskysetup}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=3.3in]{figures/path2.jpg}
\caption{A bird's eye view of the testing path for the Husky experiment. Shows the interventions needed while navigating for each policy.}
\label{fig:huskypath}
\end{figure}
\noindent Fig. \ref{fig:huskypath} shows the overhead view of the test path used for the experiment. The path is highlighted in red. Three different models are trained to drive the robot along the path---a model trained using data from a single camera, a model trained using data from a single camera with equivariant augmentations, and a model trained using data from all three cameras.
\textbf{Results.} The model trained using only center camera data performs worse than the others, steering off the path 9 times. The model trained on data generated using equivariant maps as well as the model trained using all three cameras both perform significantly better, steering off the path only once. Fig. \ref{fig:huskypath} shows a snapshot of the entire test run for three setups respectively. The locations where an intervention occurred are denoted by a yellow cross.\footnote{The video from this experiment can be found here: \href{https://youtu.be/5g4Kg3-YWvA}{https://youtu.be/5g4Kg3-YWvA}}
\subsection{Learnt vs Non-Learnt Equivariance Maps}
In each of the experiments thus far, we modeled the equivariance maps using a neural network that is optimized using expert data. We investigate the necessity of learning these maps by comparing them to a random map that does not enforce any equivariance relation, and a fixed deterministic map (a linear translation in embedding space). Our results as shown in Fig. \ref{fig:detmap_cte} indicate that both alternative techniques fall short of achieving the performance improvements gained using an equivariance map.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{figures/detEqCTEshrunk.jpeg}
\caption{Cross track error(m) for driving policy trained using a deterministic map and a learned map enforcing equivariant features.}
\label{fig:detmap_cte}
\vspace{-0.6cm}
\end{figure}
\section{CONCLUSION}
In this paper, we investigate the use of equivariant maps for visual navigation by mobile robots. By training a neural network model that learns image features which are equivariant, we can predict the latent representations of images from viewpoints nearby to the one observed. Our results indicate that by augmenting the training dataset with these representations, one can significantly improve navigation performance in a variety of settings as demonstrated by our simulation experiments involving flying and ground vehicles.
Additionally, through our ground robot experiments over a 500m path, we showed that the benefits of our method also transfer over to real world settings.
|
\section{Introduction}
Cohen, Lenstra, and Martinet \cite{Cohen1984, Cohen1990} have given heuristics on the distribution of class groups of number fields that lead to the following conjecture. For a number field $k$, a transitive permutation group $G$, and a prime $p\nmid |G|$, there conjecturally exists a constant $c_{k,G,p}>0$ such that
\begin{equation}\label{E:conj}
\lim_{X\ra\infty} \frac{1}{|E_k(G,X)|}\sum_{K\in E_k(G,X) } |\Cl_K[p]| \stackrel{\textrm{Conj.}}{=}c_{k,G,p},
\end{equation}
where $E_k(G,X)$ is the set of extensions $K/k$ with Galois closure group $G$ (that is, the Galois group of the normal closure; see Section \ref{sec:notations})
and $\Disc(K)\leq X$.
The limiting average in \eqref{E:conj} was previously known for only two cases of $G$ and $p$:
for $G=S_2$ and $p=3$
by
Davenport and Heilbronn \cite{DH71} when $k=\ensuremath{{\mathbb{Q}}}$ and Datskovsky and Wright \cite{DW88} for general $k$, and for $G=S_3$ and $p=2$
by Bhargava \cite{Bha05} when $k=\ensuremath{{\mathbb{Q}}}$ (see also the work of Bhargava, Shankar, and Wang \cite{BSW15} for the case of general $k$).
This conjecture is particularly notable in light of the fact that the best general upper bound on $|\Cl(K)[p]|$ is $\Disc(K)^{1/2+\epsilon}$, and there has been a lot of recent breakthrough work (e.g. \cite{Heath-Brown2017, PTBW})
even to show that the average in \eqref{E:conj} is bounded by $X^{1/2-\delta}$ for some $\delta>0$, for certain $G$, despite the fact that it is conjectured to have constant limit.
Our first main result is that we find the average and prove the conjecture in \eqref{E:conj} when $G$ is any transitive $2$-group containing a transposition and $p=3$.
\begin{theorem}\label{T:G}
For any $m$, let $G\sub S_{2^m}$ be a transitive permutation $2$-group containing a transposition.
For any number field $k$, there exists a constant $c_{k,G,3}$ (given explicitly in Section~\ref{sec:main-theorem}) such that
$$
\lim_{X\ra\infty} \frac{1}{|E_k(G,X)|}\sum_{K\in E_k(G,X) } |\Cl_K[3]| =c_{k,G,3}.
$$
\end{theorem}
The simplest new example that Theorem \ref{T:G} provides is that the average size of $3$-torsion in the class groups of $D_4$-quartic fields over $k$ tends to a constant; when $k=\mathbb{Q}$, this constant works out to be around $1.42$.
It applies also to three permutation groups $G$ of degree $8$, to $26$ groups of degree $16$, and to at least as many groups in degree $2^m$ as in degree $2^{m-1}$.
The restriction that $G$ has a transposition is necessary for our proof, and is a natural condition in this setting. In particular, it is expected (for example, according to Malle's conjecture \cite{Mal04}) that the groups $G$ arising with positive density as Galois groups when fields of degree $n$ are ordered by discriminant are precisely those with a transposition in their degree $n$ permutation representation. This is known for $n=4$ where there are $\sim c_1 X$ quartic $D_4$ fields \cite{CDyDO02} and $\sim c_2X$ quartic $S_4$ fields \cite{Bha05} of absolute discriminant bounded by $X$ for certain constants $c_1,c_2 > 0$, and $O_\epsilon(X^{1/2+\epsilon})$ quartic fields with any other Galois closure group.
Thus, the $2$-groups to which Theorem \ref{T:G} applies---or, indeed, the groups susceptible to our method as a whole---are those expected to have positive density within the set of all fields of a given degree, which you might think of as ``generic'' Galois groups.
When $G$ is a transitive permutation $2$-group with a transposition and $K\in E_k(G,X)$, then $K$ has a unique index two subfield $F$, and we can also ask about the relative class group of $K/F$. Indeed for $G=D_4$ Cohen and Martinet \cite{Cohen1987} have said this part of $\Cl_K$ is of particular interest.
Our next main result gives the averages of $|\Cl_{K/F}[3]|=|\Cl_K[3]|/|\Cl_F[3]|$.
\begin{theorem}\label{T:rel}
For any $m$, let $G\sub S_{2^m}$ be a transitive permutation $2$-group containing a transposition.
Let $k$ be a number field, $u$ be an integer, and $E^u_k(G,X)$ be those extensions $K\in E_k(G,X)$ such that $\rk \O_K^*-\rk \O_{F_K}^*=u$, where $F_K$ is the unique extension of $k$ such that $[K:F_K]=2$. Then, if $E^u_k(G,\infty)$ is non-empty,
$$
\lim_{X\ra\infty} \frac{1}{|E^u_k(G,X)|}\sum_{K\in E^u_k(G,X) } |\Cl_{K/F_K}[3]| =1+3^{-u},
$$
and this particular average value is predicted by the Cohen--Lenstra--Martinet heuristics.
\end{theorem}
Cohen, Lenstra, and Martinet actually gave a conjecture that predicts the average of $f(\Cl_K)$ over $G$-extensions $K/k$ for any function $f$ not depending on certain ``bad'' Sylow subgroups of the class group.
The only non-trivial cases of these conjectures previously proven were for averages of the two $f$ of the form $f(\Cl_K)=|\Cl_K[p]|$ mentioned above.
Theorem \ref{T:rel} thus gives the first proof the Cohen-Lenstra-Martinet conjecture for a function not of the form $|\Cl_K[p]|$ (or trivial variations), and does this for infinitely many groups $G$.
(When $G=C_4$ and $f$ is bounded and only depends on the class group of the quadratic subfield of the cyclic quartic,
Bartel and Lenstra \cite{Bartel2020} showed that the average is computable in finite time to arbitrary
precision, though the example they compute does not agree with the conjecture---see Section~\ref{sec:CM} for further discussion.
There have
also been some averages proven on the ``bad'' part of the class group including work of Fouvry and Kl\"uners \cite{Fouvry2006a} and Klys \cite{Klys2016a}, and the recent groundbreaking work of Smith \cite{Smith2017} determining the entire distribution of $\Cl_K[2^\infty]$ when $G=S_2$. Other work in this direction includes that of Gerth \cite{Gerth1984, GerthIII1987} and Koymans and Pagano \cite{KP18}.)
\subsection{Methods}
The condition that $G$ contains a transposition implies that $G=C_2\wr H$ for some $H$.
So, for the extensions $K/k$ we consider, $K$ has a unique index two subfield $F$, and the overarching strategy of the paper is to
average and then sum over each $F$.
Accordingly, the work of Datskovsky and Wright on the average of $|\Cl_K[3]|$ over quadratic extensions of a fixed $F$ is a key input into our work.
If we could use just the main term of the average in place of the average, the argument would be essentially straightforward.
However, summing the error terms over $F$ in a straightforward way leads to error that is larger than the main term.
The major obstacle we overcome in this paper is proving a sufficiently good tail bound so that we can sum over $F$. In particular, the challenge is to bound the contribution when the average
is taken over a relatively small interval, given $\Disc(F)$.
Our first tool to do this is a heavily optimized upper bound on the average $3$-class number of quadratic extensions of a number field $F$, and the closely related
number of $S_3$ extensions of $F$, which we develop in Section~\ref{sec:uniform-cubic}.
For very small or very large intervals, compared to $\Disc(F)$, we use class field theory and the theory of the Shintani zeta function counting cubic orders due to Datskovsky and Wright \cite{DW88}, respectively.
While cubic extensions of a general number field are counted in \cite{DW88}, in order to sum over different base fields, we must determine the explicit dependence of the count on the base field, which has not been done in previous work.
For intermediate intervals, however --- summing $|\Cl(K)[3]|$ for roughly $\Disc (K)\in[\Disc(F)^{3.1}, \Disc(F)^6]$ ---
we require a new technique, which we call propagation of orders.
In this range, the bound on cubic extensions given by the Shintani zeta function is too large. However, the Shintani zeta function we use counts non-maximal orders as well as maximal orders. Given a maximal order, one can prove a lower bound on the number of non-maximal orders inside it, up to a certain discriminant bound.
Then, to count maximal orders of discriminant up to $X$, we count all orders of discriminant up to $Z$ for some $Z>X$, and then account for the known overcounting.
As $Z$ gets larger the total count goes up but the known overcounting also improves, and it turns out in the intermediate range these trade-offs work out to give us an improved and sufficiently good bound.
With these optimized bounds on $S_3$ extensions of a single $F$, especially in shorter intervals, we begin to take the sum over $F$, and we can bound the sum as required except in one problematic range, which we call the critical range. In this range, $\Disc(K)\approx \Disc(F)^3$. If $\Disc(K)$ is a power smaller or larger than this, then the methods above are sufficient. In the critical range, the best individual bounds we can prove on $S_3$ extensions of $F$
use class field theory and a trivial bound for $|\Cl_{K/F}[3]|$, and are too large when summed.
Ellenberg and Venkatesh \cite{EV} have shown that one can improve the trivial bound when $K$ has enough small split primes.
Ellenberg, Pierce, and the third author \cite{Ellen16} introduced the idea that in many situations, one can prove that most fields in a family have enough
small split primes to apply \cite{EV}. We use that approach, though there is one significant new aspect.
As we sum $|\Cl_{K}[3]|=|\Cl_{F}[3]||\Cl_{K/F}[3]|$ over $F$, we require a significant saving over the trivial bound.
We structure our proof as an induction, which allows us to assume an optimal (constant average) bound for the $|\Cl_{F}[3]|$ factor.
We then further need a non-trivial bound on the relative class number
$|\Cl_{K/F}[3]|= |\Cl_{K}[3]|/|\Cl_{F}[3]|.$
For this, in Section~\ref{sec:arakelov} we develop a relative version of the widely-used lemma of Ellenberg and Venkatesh \cite{EV}, using the relative Arakelov class group to show that one obtains a non-trivial bound on $|\Cl_{K/F}[p]|$ when there are sufficiently many small split primes in $K/F$.
To find that most fields have small split primes, in Section~\ref{sec:zero-density} we use a uniform zero density estimate for Hecke $L$-functions as in \cite[Proposition A.2]{Pasten} or \cite[Theorem 1.2]{ThornerZaman-LargeSieve} to show that most of the Dedekind zeta functions do not have zeros in the relevant region.
In Section~\ref{sec:tail}, we put these results together to prove the desired tail bound, and in Section~\ref{sec:main-theorem} we prove the main theorems.
Our methods apply more broadly than just to $2$-groups. We determine in Section \ref{sec:other-groups} the average size of $3$-torsion in the class group of $(C_2 \wr G)$-extensions for many general classes of group $G$, where $C_2 \wr G$ denotes the wreath product.
\subsection{Previous Work}
There is a large body of previous work in arithmetic statistics and analytic number theory on the questions that arise in the this paper. There have been many remarkable achievements on counting cubic extensions, and relatedly averaging three torsion in class groups over quadratic extensions, after the work of Davenport and Heilbronn \cite{DH71} and Datskovsky and Wright \cite{DW88} discussed above. Taniguchi and Thorne \cite{TT13} and Bhargava, Shankar, and Tsimerman \cite{BST} improved the error term in Davenport and Heilbronn's count so far as to uncover a secondary term, and recent work by Bhargava, Taniguchi, and Thorne \cite{Bhargava2021} has improved the error term even further. Bhargava, Shankar, and Wang \cite{BSW15} have given a geometry of numbers proof of the counting results of Datskovsky and Wright, opening the door to further applications. See also Thorne's exposition \cite{Th12} of the different approaches that have been used to count cubic fields.
The problem of proving non-trivial bounds for $|\Cl_K[p]|$ has also been the topic of many important works. Pierce, Helfgott, and Venkatesh \cite{Pie05,HV06,Pie06} proved the first non-trivial bounds (at a prime $p$ not dividing the order of the Galois group), breaking the trivial bound in the case that $K$ is quadratic field and $p=3$. This bound was improved by Ellenberg and Venkatesh \cite{EV} to $|\Cl_K[3]|=O_\epsilon(\Disc K^{1/3+\epsilon})$ for quadratic fields $K$.
Bhargava, Shankar, Taniguchi, Thorne, Tsimerman, and Zhao \cite{BSTTTZ} gave a non-trivial bound on $|\Cl_K[2]|$ for all $K$ using geometry of numbers.
Ellenberg and Venkatesh \cite{EV} also proved a non-trivial bound on $|\Cl_K[3]|$ whenever $[K:\ensuremath{{\mathbb{Q}}}]\leq 4$.
The second author gave a non-trivial bound on $|\Cl_K[p]|$ when $K$ is Galois with the Galois group a non-cyclic and non-quaternion $\ell$-group or a nilpotent group where every Sylow-$\ell$ subgroup is non-cyclic and non-quaternion \cite{Wang2021a,Wang2020}. In the setting when $p$ divides the order of the Galois closure group of $K$, Kl\"{u}ners and the second author \cite{KluWan} proved $|\Cl_K[p]|=O_\epsilon(\Disc K^{\epsilon})$ when $K$ has Galois closure group a $p$-group, generalizing the classical result from Gauss's genus theory for $K$ quadratic and $p=2$.
Soundararajan \cite[pg.690]{Sound-Divisibility} was the first to observe that non-trivial bounds for $|\Cl_K[p]|$ could be proved for almost all imaginary quadratic $K$, or equivalently, on average over $K$.
Heath-Brown and Pierce \cite{Heath-Brown2017} reinvigorated the study of such average bounds for $|\Cl_K[p]|$, as well as bounds on higher moments, by using the large sieve.
As mentioned above, Ellenberg, Pierce, and the third author \cite{Ellen16} proved non-trivial bounds for most fields in low degree families using
a non-trivial bound of Ellenberg and Venkatesh \cite{EV} conditional on the existence of small split primes, and precise field counting results with local conditions to prove such existence.
Pierce, Turnage-Butterbaugh, and the third author \cite{PTBW} found ways to prove non-trivial bounds for many more families using
zero density estimates for $L$-functions to prove the required existence of small primes, which is a strategy we use in this paper.
This strategy has been improved and extended to further cases by An \cite{An2020}, Frei and Widmer \cite{Frei2018,Frei2021}, Widmer \cite{Widmer2018}, Thorner and Zaman \cite{Thorner2019}, and the first author, Thorner and Zaman \cite{Oliver2021}.
\subsection{Further questions}
We expect the methods in this paper can be used in other settings. We have forthcoming work with Alberts which uses a similar inductive approach to prove many new cases of Malle's conjecture \cite{Mal04} on the asymptotics of $|E_k(G,X)|$.
It would be natural to apply our approach to the $2$-part of class groups of fields that can be obtained as cubic extensions of other extensions,
using \cite{Bha05} as a base case.
It was noted by Bhargava and Varma \cite{BhargavaVarma} that the average $3$-torsion in class group of quadratic extensions does not change when the fields are subject to finitely many local conditions at finite primes (see also \cite{Wood2018}). We expect that the average in Theorem \ref{T:rel} behaves similarly and is insensitive to local conditions, while the average in Theorem \ref{T:G} behaves differently and is quite sensitive to local conditions, and it would be interesting to investigate whether this is the case. A subtlety in this question is in defining a notion of local conditions that works well in towers of fields.
The contrast between the constants in Theorems~\ref{T:G} and \ref{T:rel} naturally leads to the question: when ordering fields by a particular invariant, which averages does one expect to be as predicted by Cohen--Lenstra--Martinet? For example, does the analogue of Theorem \ref{T:rel} hold when the fields $K$ are instead ordered by the
Artin conductor of the representation of $\Gal(\bar{k}/k)$ induced from the sign representation of $\Gal(K/F_K)$? (This is an irreducible representation, as we show in the proof of Theorem~\ref{P:asCM}.) In the case of quartic $D_4$ extensions of $k=\mathbb{Q}$, the fields themselves were recently counted by this Artin conductor in work of Alt\v{u}g, Shankar, Varma, and Wilson \cite{D4VS}.
It would be interesting to extend our methods to as wide a range of groups $G$ as possible (e.g., by incorporating ideas from Sections \ref{sec:arakelov} and \ref{sec:zero-density} into an argument as in Section \ref{sec:other-groups}). For example, does an analogue of Theorem \ref{T:G} hold for $(C_2 \wr G)$-extensions for arbitrary abelian groups $G$? Arbitrary $p$-groups?
Finally, as noted above, Malle's conjecture predicts asymptotics for $|E_k(G,X)|$. These predicted asymptotics are of the form $|E_k(G,X)| \sim c_{G,k} X^{1/a(G)} (\log X)^{b(k,G)-1}$ for specified integers $a(G)$ and $b(k,G)$, and an unspecified positive constant $c_{G,k}$, as $X$ tends to infinity. Our work raises the question, what is the least exponent $\alpha = \alpha(G)$ such that $|E_k(G,X)| \ll_{[k:\mathbb{Q}],\epsilon} \Disc(k)^{\alpha+\epsilon} X^{1/a(G)} (\log X)^{b(k,G)-1}$ for all $k$, all $\epsilon>0$, and all $X \geq 1$? Does such an exponent even exist? When $G$ is abelian, this problem is closely related to bounding torsion subgroups of class groups. Some of the key technical results of this paper (see Section \ref{sec:uniform-cubic}) provide the first such result for a non-abelian group $G$, namely $G=S_3$. What can be said for other non-abelian groups $G$?
This question when $G=S_3$ could naturally be approached by proving a discriminant-aspect subconvexity bound for the relevant Shintani zeta functions, which the methods of Section \ref{sec:uniform-cubic} fall a little short of proving. Can these or other methods be adapted to prove such a subconvexity result? We mention recent work of Hough and Lee \cite{HoughLee} that establishes a subconvexity bound in $t$-aspect for the Shintani zeta functions over $\mathbb{Q}$, though we expect these methods do not easily apply to discriminant-aspect.
\subsection{Technical remarks}
The conjectures of Cohen, Lenstra, and Martinet in fact fix some archimedean data and only consider families of fields with this fixed data. In particular, for Galois $\Gamma$-extensions $L/k$, they fix the $\Gamma$-module structure of $\O_L^* \tensor_\Z \ensuremath{{\mathbb{Q}}}$, which is equivalent \cite[Th\'{e}or\`{e}me 6.7]{Cohen1990} to fixing the representation $\oplus_{v} \Ind_{\Gamma_v}^{\Gamma} \ensuremath{{\mathbb{Q}}}$, where the sum is over infinite places $v$ of $k$ and $\Gamma_v$ is the decomposition group of $L/k$ at $v$, and we call this the \emph{enriched signature}. The papers of Cohen, Lenstra, and Martinet are agnostic on the relative frequency of the various enriched signatures (and even whether such limiting frequencies exist). So, to be precise, to obtain \eqref{E:conj} from their conjectures one must also assume such frequencies exist (which is natural to conjecture given all known counting results, e.g. \cite{ DH71, Wright, CDyDO02, Bha05}, and is conjectured by Malle \cite{Mal04}), and the conjectural value would be given as a weighted average of particular values for each enriched signature as conjectured by Cohen, Lenstra, and Martinet, where the weights are these frequencies.
When one fixes an enriched signature, one can use the results of \cite{Wang2021} to find the average value predicted by Cohen, Lenstra, and Martinet (see Section~\ref{sec:CM}).
We prove a version of Theorem~\ref{T:G} with a specified enriched signature (in fact with slightly more specification at infinity) in Theorem~\ref{thm:main}. However, even when we do this, we do not apparently obtain the values conjectured by Cohen and Martinet in \cite{Cohen1990}. We show this computationally is the case in Section~\ref{SS:wrongpred}. This is because in this setting each $F$ appears as the index $2$ subfield for a positive proportion of $K$, and so the $3$-ranks of the particular $F$ that appear with significant frequency bias the statistics, as in the work of Bartel and Lenstra \cite{Bartel2020}. When we only average $\Cl_{K/F}$ as in Theorem~\ref{T:rel}, this bias is removed. See Section~\ref{sec:CM} for further discussion of this phenomenon.
\subsection{List of notations}\label{sec:notations}
\noindent
$k$: a number field, for us a subfield of a fixed algebraic closure $\bar{\ensuremath{{\mathbb{Q}}}}$\\
$\O_k$: the ring of algebraic integers in $k$\\
$\disc(F/k)$: relative discriminant ideal in $\O_k$, for an extension $F/k$\\
$\Nm I$: norm $|\O_k/I|$ of an ideal $I$ in $\O_k$ and extend to any fractional ideals in $\O_k$ multiplicatively \\
$\Nm_{F/k} I$: an ideal in $\O_k$ that is $(I\cap \O_k)^{f(I| I\cap \O_k)}$ when $I$ is a prime ideal in $\O_F$ and $f$ is the inertia degree of $I$ over $I\cap \O_k$, and defined for any fractional ideals in $\O_F$ multiplicatively \\
$\Disc(F/k)$: $\Nm(\disc(F/k))$ (When $k$ omitted, $k=\ensuremath{{\mathbb{Q}}}$) \\
$\Gal(F/k)$: the \emph{Galois closure group} of $F/k$, defined as the Galois group of $\tilde{F}/k$ as a \emph{permutation group}, where $\tilde{F}$ is the Galois closure of $F$ over $k$ (in $\bar{\ensuremath{{\mathbb{Q}}}}$) acting on the embeddings of $F\ra \tilde{F}$ that fix $k$ pointwise\\
$G$-extension: for a permutation group $G$, an extension $L/k$ of number fields (in $\bar{\ensuremath{{\mathbb{Q}}}}$), \emph{and a choice of isomorphism} $\phi:\Gal(L/k)\isom G$ \emph{as permutation groups}
(see Definition~\ref{D:permgroup})
\\
$2$-extension: a $G$-extension for some $2$-group $G$\\
$E_k(G,X)$: the set of $G$-extensions $F/k$ with $\Disc(F)\le X$
\\
$N_k(G,X)$: the number of $G$-extensions $F/k$ with $\Disc(F/k)\le X$ (note relative discriminant, versus absolute above)\\
$\Cl_F$: class group of $F$ \\
$\Cl_{F/k}$: relative class group of $F/k$, when $k=\mathbb{Q}$ it is the usual class group of $F$ (see Section~\ref{sec:arakelov})
\\
$A[\ell]$: $\{ [\alpha]\in A \mid \ell [\alpha] = 0\in A \}$ for an abelian group $A$\\
$h(F/k)$, $h(F)$: the size of $\Cl_{F/k}$, $\Cl_{F}$\\
$h_{\ell}(F/k)$, $h_{\ell}(F)$: the size of $\Cl_{F/k}[\ell]$, $\Cl_{F}[\ell]$\\
$\ACl_L$: Arakelov class group\\
$\pi_k(X)$: the number of prime ideals $\mathfrak{p}$ in $k$ with $\Nm(\fp)<X$\\
$\pi_k(Y; F, \mathcal{C})$: the number of unramified prime ideals $\mathfrak{p}$ in $k$ with $\Nm_{k/\ensuremath{{\mathbb{Q}}}}(\fp)<Y$ and with Frobenius at $\fp$ in the conjugacy class of $\mathcal{C}$ in $\Gal(F/k)$\\
$r_1(k)$: the number of real embeddings of $k$\\
$r_2(k)$: the number of pairs of complex embeddings of $k$\\
$\rk \O_K^*$: the unit rank, i.e. $\dim_{\ensuremath{{\mathbb{Q}}}} (\O_k^*\tensor \ensuremath{{\mathbb{Q}}}$)\\
$\zeta_k(s)$: Dedekind zeta function of $k$\\
$\kappa_k$: the residue $\mathrm{Res}_{s=1}\zeta_k(s)$\\
$\Rg_k$: the regulator of $k$\\
$f(X) \sim g(X)$: $\lim_{X\to \infty} f(X)/g(X) = 1$ or $\lim_{X\ra\infty} f(X)=\lim_{X\ra\infty}g(X)=0$\\
$f(X) =O_{p_1,\dots}(g(X))$: there exists a constant $C$ depending only on $p_1,\dots$ such that for all values of all variables $|f(X)|\leq Cg(X)$\\
$f(X) \ll_{p_1,\dots}g(X)$: $f(X) =O_{p_1,\dots}(g(X))$\\
$f(X) \gg_{p_1,\dots}g(X)$: only used when both sides non-negative, and then means $g(X) \ll_{p_1,\dots}f(X)$
\section*{Acknowledgements}
We thank Arul Shankar, Frank Thorne, and Jesse Thorner for helpful conversations related to this project. We also thank Brandon Alberts, Alex Bartel, Jordan Ellenberg,
J\"{u}rgen Kl\"{u}ners,
Lillian Pierce, Frank Thorne, Jesse Thorner, and Asif Zaman for comments on an earlier version of this paper.
RJLO was partially supported by National Science Foundation grant DMS-1601398. JW was partially supported by a Foerster-Bernstein Fellowship at Duke University. MMW was partially supported by a Packard Fellowship for Science and Engineering, National Science Foundation grant DMS-1652116, and an NSF Waterman award.
\section{Nontrivial Bound for the Relative Class Group}
\label{sec:arakelov}
In this section, our main goal is to establish Lemma \ref{lem:relative-E-V}, which is a \textit{relative} version of a widely used lemma of Ellenberg--Venkatesh \cite[Lemma 2.3]{EV}.
Let $L/K$ be an extension of number fields. The relative class group $\Cl_{L/K}\subset \Cl_L$ is defined to be the kernel of the map $\Nm_{L/K}: \Cl_L\to \Cl_K$ induced from the usual norm on fractional ideals of $L$.
\begin{lemma}[Relative Ellenberg--Venkatesh]\label{lem:relative-E-V}
Let $L/K$ be an extension of number fields with $[L:K]=d$, let $\ell\ge 1$ be an integer, and let $\theta$ be a real number such that $0< \theta< \frac{1}{4\ell(d-1)}$. Suppose there exist $M$ pairs of distinct primes $(\wp_i, \bar{\wp}_i)$ in $\O_L$ and coprime integers $a_i,b_i$ such that
\begin{enumerate}
\item
For each $i$, we have $\wp_i \cap \O_K = \bar{\wp}_i\cap \O_K$ and $\Nm_{L/K}(\wp_i)^{a_i} = \Nm_{L/K}(\bar{\wp}_i)^{b_i}$;
\item
For each $i$, the norms satisfy $ \Nm(\wp_i)^{a_i}=\Nm(\bar{\wp_i})^{b_i}<\Disc(L/K)^{\theta}$; and
\item
For each $i$, there is no intermediate field $H$ with $K\subset H \subsetneq L$ such that $\wp_i$ and $\bar{\wp_i}$ are both extended from $H$. (A prime $\wp\subset \O_L $ is \emph{extended} from $H$ if there exists a prime $\fp\subset \O_H$ such that $\wp = \fp \O_L$.)
\end{enumerate}
Then
\begin{equation}\label{eqn:EV-relative}
|\Cl_{L/K}[\ell]|= O_{[L:\ensuremath{{\mathbb{Q}}}],\theta,\epsilon}\Big(\frac{\Disc(L/K)^{1/2+\epsilon} \Disc(K)^{(d-1)/2+\epsilon}}{M}\Big).
\end{equation}
\end{lemma}
\begin{remark}
In our application, we will only use prime ideals that are split completely in $L/K$, therefore the first and third conditions will be automatically satisfied, and the second condition will be equivalent to $\Nm(\wp\cap \O_K)<\Disc(L/K)^{\theta}$.
\end{remark}
The original lemma \cite{EV} gives a non-trivial bound
\begin{equation}\label{eqn:EV-lemma}
|\Cl_{L}[\ell]|= O_{[L:\ensuremath{{\mathbb{Q}}}],\theta,\epsilon}\Big(\frac{\Disc(L)^{1/2+\epsilon}}{M}\Big),
\end{equation}
for the absolute class group $\Cl_L[\ell]$ under the same assumptions. Our improvement is in proving a saving from the trivial bound on the size of the \emph{relative} class group $\Cl_{L/K}[\ell]$ instead of the \emph{absolute} class group. Such a refinement is crucial for our application since we will treat $\Cl_K$ and $\Cl_{L/K}$ separately.
When $\ell$ is relatively prime to $[L:K]$, there is a short exact sequence
\begin{center}\label{eqn:relative-class-group-size}
\begin{tikzcd}
0\arrow{r} & \Cl_{L/K}[\ell] \arrow{r} & \Cl_L[\ell] \arrow[r,"\Nm_{L/K}"] & \Cl_K[\ell] \arrow{r} & 0.
\end{tikzcd}
\end{center}
Let $\iota$ be the natural map from $\Cl_K$ to $\Cl_L$ by extension of ideals. The composition $\Nm_{L/K}\circ \iota: \Cl_K[\ell] \to \Cl_K[\ell]$ is equal to multiplication by $[L:K]$ and thus is an isomorphism since $(\ell, [L:K]) =1$, so $\iota$ induces a splitting section of the above sequence. So we have
$$ \Cl_{L}[\ell] \simeq \Cl_{L/K}[\ell] \times \Cl_{K}[\ell], \quad\quad h_{\ell}(L) = h_{\ell}(K)\cdot h_{\ell} (L/K).$$
This leads to a strategy we use in this paper to bound $h_{\ell}(L)$, that is, to bound $h_{\ell}(K)$ and $h_{\ell}(L/K)$ separately.
The organization of this section is as follows. In Section \ref{ssec:relative-arakelov-set-up}, we give a definition of a \emph{relative Arakelov class group}, compute the volume of it and derive the ``trivial'' bound for the relative class group in Lemma \ref{lem:relative-trivial}. In Section \ref{ssec:relative-EV}, we prove Lemma \ref{lem:relative-E-V}.
\subsection{The relative Arakelov class group}\label{ssec:relative-arakelov-set-up}
In this section we will define a relative version of the Arakelov class group. First we recall the usual
\emph{absolute} notion of the Arakelov class group. A nice reference for the theory of the absolute Arakelov class group is \cite{Schoof}.
For any number field $L$ with degree $n$ and $r=r_1+r_2$ infinite places, we let $I_L$ be the group of fractional ideals and $L_\infty := \prod_{v|\infty} \R^{\times}_{>0}$. We define the degree $0$ divisors to be $\Div^0(L) := \{(x,\mathfrak{a}) \in L^{\times}_\infty \times I_L: |x|_{\infty} =\Nm(\mathfrak{a})\}$ where $\Nm(\mathfrak{a}):=|\O_L/\mathfrak{a}|$ and $|x|_{\infty} = \prod_{v|\infty} |x_v|$. There is a natural embedding $\pi: L^{\times} \to \Div^0(L)$ by sending $\alpha\in L^{\times}$ to $ ((|\alpha|_{v}), (\alpha)) \in \Div^0(L)$ from the product formula (here $|\cdot|_v$ is the usual norm in $\R$ at the $r_1$ real embeddings and the square of the usual norm in $\R$ at the $r_2$ complex embeddings).
The absolute Arakelov class group $\ACl_L$ of $L$ is then defined as $\ACl_L:= \Div^0(L)/\pi(L^{\times})$, and it fits into the following short exact sequence
\begin{center}\label{fml:Arakelov-SES}
\begin{tikzcd}
0\arrow{r} & (\oplus \R)^{(0)}/\mathcal{L}(L) \arrow{r} & \ACl_L \arrow{r} & \Cl_L \arrow{r} & 0,
\end{tikzcd}
\end{center}
where $(\oplus_v \R)^{(0)}$ is isomorphic to $L_\infty^{(1)}$ (the elements of $L^{\times}_\infty$ of norm $1$) via the Minkowski embedding, i.e. the logarithm map $\log=\oplus_v \log$, and $\mathcal{L}(L) =\log( \pi(\O_L^{\times}))$ is a full rank lattice by Dirichlet's unit theorem.
The Arakelov class group is not only an abelian group but also a Lie group.
We fix a left-invariant Riemannian metric on $\widetilde{\Cl}_{L}$
that on $(\oplus_v \R)^{(0)}$ agrees with the restriction of the standard Riemannian metric from $\oplus_v \R$.
We will use this Riemannian metric on $\widetilde{\Cl}_{L}$, and the induced metrics on various sub-Lie groups, to compute all volumes in this section.
We compute $\Vol(\ACl_L) = h(L) \cdot \Rg(L)\cdot \sqrt{r}$.
Though the language of a Riemannian metric is convenient, we remark that in our case this metric is very concrete
(in particular, because unlike \cite{EV}, we are using the \emph{non-oriented Arakelov class group} as in \cite[Section $5$]{Schoof}).
All of our groups
are a disjoint union of finitely many copies of quotients of subspaces of $\R^n$, and the volumes we compute are always the ones that come from the usual metric on $\R^n$.
The absolute Arakelov class group associated to $L$ and $K$ can be related by a norm map $\Nm_{L/K}: \ACl_L\ra \ACl_K$
where we use the relative norm map $\Nm_{L/K}:I_L\to I_K$ for fractional ideals and $\Nm_{\infty}: (\oplus_v \R)^{(0)} \ra (\oplus_w \R)^{(0)}$ that sends $(x_v)_{v\mid\infty}\mapsto
(\sum_{v\mid w} x_v)_{w\mid\infty}$ for archimedean embeddings. This norm map $\Nm_{L/K}$ fits into the following commutative diagram:
\begin{equation}\label{diag:Arakelov-relative}
\begin{tikzcd}
0\arrow{r} & (\oplus_v \R)^{(0)}/\mathcal{L}(L) \arrow{r}\arrow[d,"\mathrm{Nm}_\infty"] & \ACl_L \arrow{r}\arrow[d, "\mathrm{Nm}_{L/K}"] & \Cl_L \arrow{r}\arrow[d, "\mathrm{Nm}_{L/K}"] & 0 \\
0\arrow{r} & (\oplus_w \R)^{(0)}/\mathcal{L}(K) \arrow{r} & \ACl_K \arrow{r} & \Cl_K \arrow{r} & 0.
\end{tikzcd}
\end{equation}
We now can define the relative Arakelov class group in a manner analogous to the relative class group, as a subgroup of the absolute Arakelov class group.
\begin{definition}[Relative Arakelov Class Group]
Given an extension $L/K$ of number fields,
we define the \emph{relative Arakelov class group} $\widetilde{\mathrm{Cl}}_{L/K}$
to be the kernel of $\Nm_{L/K}: \ACl_L\to \ACl_K.$
\end{definition}
We obtain a
short exact sequence for the relative Arakelov class group as an abelian group, similar to the sequence \ref{fml:Arakelov-SES}:
\begin{equation}\label{diag: Arakelov-relative}
\begin{tikzcd}[node distance = 1.5cm]
0\rar & \mathrm{Ker}(\mathrm{Nm}_\infty) \arrow{r} & \ACl_{L/K} \arrow{r} & \Cl_{L/K} \arrow{r} & 0,
\end{tikzcd}
\end{equation}
from an application of the snake lemma on \eqref{diag:Arakelov-relative}, since $\Nm_{\infty}$ is surjective.
Our next goal is to show that $\Vol(\Ker(\Nm_{\infty})) = C\Rg_L/\Rg_K$ where $C$ is a constant only depending on the signatures of $L$ and $K$. Let $v_{i,j}$ be the standard coordinates in $\oplus_v \R$ and $w_i$ in $\oplus_w \R$, where $\Nm_{\infty}$ maps $w_i = \sum_{1\le j\le i_k} v_{i,j}$ and $i_k$ is the number of infinite places above $w_i$. We then choose
a new coordinate system for $\oplus_v \R$ to be $\tilde{v}_{i,j}$ where $\tilde{v}_{i,j} = v_{i,j}$ for $j>1$ and $\tilde{v}_{i,1} = \sum_{j} v_{i,j} = w_i$. Then we can compute the covolume of $(\oplus_v \R)^{(0)}/\mathcal{L}(L)$ using the coordinate system $v_{i,j}$ or $\tilde{v}_{i,j}$ (without $v_{1,1}$ and $\tilde{v}_{1,1}$ respectively):
$$\Rg_L =\int_{\tilde{v} \in \mathcal{F}_L} \,\mathrm{d} \tilde{v} = \int_{w \in \mathcal{F}_K} \,\mathrm{d} w \int_{\tilde{v}' \in \Nm_{\infty}^{-1}(w)} \,\mathrm{d} \tilde{v}'= \Rg_K \cdot \int_{\tilde{v}' \in \Nm_{\infty}^{-1}(0)} \,\mathrm{d} \tilde{v}'.$$
Here ${\mathcal{F}}_L$ is the projection of a fundamental domain for $(\oplus_v \R)^{(0)}/\mathcal{L}(L)$ onto the $v_{i,j}((i,j)\neq (1,1))$-coordinate plane, and we define $\mathcal{F}_K$ analogously. The first equality comes from the definition of regulator. The second equality is an iterated integral where we use $\tilde{v}'$ to denote the variables in $\tilde{v}$ aside from $\tilde{v}_{i,1}$, and $\Nm_{\infty}^{-1}(w)= \{ v\in \mathcal{{F}}_L\mid \Nm_{\infty}(v)\equiv w \mod \mathcal{L}(K)\}$. In the last equality $\int_{\tilde{v}' \in \Nm_{\infty}^{-1}(0)} \,\mathrm{d} \tilde{v}' = \int_{\tilde{v}' \in \Nm_{\infty}^{-1}(w)} \,\mathrm{d} \tilde{v}'$ for any $w\in \mathcal{F}_K$ since the two domains for $\tilde{v}'$ only differ by a translation.
We have $ \Vol(\Ker(\Nm_{\infty})) = \prod_{i} \sqrt{i_k} \cdot \int_{\tilde{v}' \in \Nm_{\infty}^{-1}(0)} \,\mathrm{d} \tilde{v}' = \prod_{i} \sqrt{i_k} \frac{\Rg_L}{\Rg_K}$.
In the following lemma we give the description of the cokernel of the map $\mathrm{Nm}_{L/K}: \Cl_L \to \Cl_K$.
\begin{lemma}\label{lem:relative-Coker}
Given an extension $L/K$ of number fields,
the cokernel of $\mathrm{Nm}_{L/K}: \Cl_L \to \Cl_K$ is isomorphic to $\Gal(M/K)$,
where $M = H_K\cap L$ and $H_K$ is the Hilbert class field of $K$.
\end{lemma}
\begin{proof}
By class field theory, the map $\Nm_{L/K}: \Cl_L \to \Cl_K$ agrees with the restriction map on the Galois groups $\Gal(H_L/L) \to \Gal(H_K/K)$ after identifying class groups and Galois groups with the reciprocity map. Since $H_K/K$ is Galois, we have $\Gal(H_L/L) \twoheadrightarrow \Gal(H_KL/L)\simeq \Gal(H_K/M) = \mathrm{Im}(\Nm_{L/K}: \Cl_L \to \Cl_K)$, and the lemma follows.
\end{proof}
As a consequence, we have an upper bound $|\mathrm{Coker}(\mathrm{Nm}_{L/K}: \Cl_L \to \Cl_K )|\leq [L:K]$, and thus we can obtain the following upper bound on $h(L/K)$, which we consider as the ``trivial bound'' on the $\ell$-torsion in the relative class group.
\begin{lemma}\label{lem:relative-trivial}
Given a relative extension $L/K$ with $[L:K]=d$ and an arbitrary integer $\ell>0$, we have
$$h_{\ell}(L/K) \ll_{[L:\ensuremath{{\mathbb{Q}}}],\epsilon}\mathrm{Disc}(L/K)^{1/2+\epsilon} \cdot \mathrm{Disc}(K)^{(d-1)/2+\epsilon}.$$
\end{lemma}
\begin{proof}
We have
\begin{equation}
h_{\ell}(L/K) \le h(L/K) \le [L:K] \cdot \frac{h(L)}{h(K)} \ll_{[L:\ensuremath{{\mathbb{Q}}}], \epsilon} \mathrm{Disc}(L/K)^{1/2+\epsilon} \mathrm{Disc}(K)^{(d-1)/2+\epsilon},
\end{equation}
where the first inequality is trivial, and the second inequality follows from Lemma \ref{lem:relative-Coker}. The third inequality comes from the theorem of Brauer-Siegel, an absolute lower bound $\frac{\Rg(L)}{\Rg(K)} \gg_{[L:\ensuremath{{\mathbb{Q}}}]} 1$ on the ratio of regulator by \cite{FrSko}, and the expression of the relative discriminant in terms of absolute discriminants.
\end{proof}
Combining the discussion of $\Vol(\Ker(\Nm_{\infty}))$ and $\Cl_{L/K}$, we obtain an estimate of $\Vol(\ACl_{L/K})$.
\begin{lemma}\label{lem:relative-volume}
Given an extension $L/K$ of number fields with $[L:K]=d$, with the measure above, we have
$$ \Disc(L/K)^{1/2-\epsilon} \Disc(K)^{(d-1)/2-\epsilon}\ll_{[L:\ensuremath{{\mathbb{Q}}}],\epsilon} \Vol(\ACl_{L/K}) \ll_{[L:\ensuremath{{\mathbb{Q}}}],\epsilon} \Disc(L/K)^{1/2+\epsilon} \Disc(K)^{(d-1)/2+\epsilon}.$$
\end{lemma}
\begin{remark}
{\em
The upper bounds in Lemmas \ref{lem:relative-trivial} and \ref{lem:relative-volume} are both effective, despite the invocation of Brauer--Siegel to estimate the quotient of residues $\mathrm{Res}_{s=1} \zeta_L(s) / \mathrm{Res}_{s=1} \zeta_K(s)$, and the lower bound of Lemma \ref{lem:relative-volume} is effective if $L$ does not contain a nontrivial quadratic extension of the form $K(\sqrt{D})$ for $D \in \mathbb{Q}$.
In particular, there is always an effective upper bound on the residue $\mathrm{Res}_{s=1} \zeta_L(s)$ due to Landau (see also \cite[Lemma 3]{Brauer}), and there is an effective lower bound on $\mathrm{Res}_{s=1} \zeta_K(s)$ whenever $\zeta_K(s)$ does not have an exceptional Landau--Siegel zero due to Stark \cite{Stark}. Thus, if $\zeta_K(s)$ does not have an exceptional zero, we simply combine these two results to get an effective upper bound. When $\zeta_K(s)$ does have an exceptional zero, the quotient $\zeta_{\widetilde{L}}(s)/\zeta_K(s)$ is entire by the Aramata--Brauer theorem \cite[Theorem 1]{Brauer}, where $\widetilde{L}$ denotes the normal closure of $L$ over $K$, and does not have an exceptional zero by Stark \cite[Theorem 3]{Stark}. Moreover, by \cite[Lemma on p.244]{Brauer}, this quotient may be expressed as a product over all $L$-functions associated to nontrivial $1$-dimensional characters of subextensions of $\widetilde{L}/K$, where each such $L$-function appears with a strictly positive rational exponent. These $L$-functions thus also do not have an exceptional zero, and are subject to effective lower bounds \cite[Lemma 2]{Stark} in addition to the effective upper bounds that hold for any entire $L$-function \cite[Ch. 5]{IwaniecKowalski}. The quotient $\zeta_L(s)/\zeta_K(s)$ may be expressed as a product and quotient of these $L$-functions, and is thus subject to an effective upper bound in the case that $\zeta_K(s)$ has an exceptional zero as well. For the lower bound on $\mathrm{Res}_{s=1} \zeta_L(s) / \mathrm{Res}_{s=1} \zeta_K(s)$, if $\zeta_L(s)$ does not have an exceptional zero, we simply use the effective lower bound on its residue and the effective upper bound on that of $\zeta_K(s)$. If $\zeta_L(s)$ has an exceptional zero, it is inherited either from $\zeta_K(s)$ or from the Dedekind zeta function of a subextension of $L/K$ of the form $K(\sqrt{D})$ with $D \in \mathbb{Q}$ by \cite[Theorem 3]{Stark}. In the former case, we proceed as in the proof of the effective upper bound to also obtain an effective lower bound; in the latter, there does not appear to be a way to make the lower bound effective.
}
\end{remark}
\subsection{Relative Ellenberg--Venkatesh Lemma}\label{ssec:relative-EV}
Now we are ready to prove Lemma \ref{lem:relative-E-V}. The idea of the proof is similar to the proof of Lemma $2.3$ in \cite{EV}, where the major difference is we use a different set $T$ of ideals that cut out the non-trivial saving in the proof. As a result, we can show a saving that is completely coming from the relative class group.
\begin{proof}[Proof of Lemma \ref{lem:relative-E-V}]
We let $\psi: \Div^0(L) \to \ACl_L$
to be natural quotient map, and denote the following subgroups of $\Div^0(L)$
$$G = \psi^{-1}(\ACl_{L/K}),\quad P = \pi(L^{\times}), \quad P_{\ell} = \ell G + P.$$
We can pull-back the Riemannian metric from $\ACl_L$ to a left-invariant Riemannian metric on $G$.
The short exact sequence (\ref{diag: Arakelov-relative}) induces an isomorphism $\Cl_{L/K}/\ell\Cl_{L/K} \simeq \ACl_{L/K}/\ell\ACl_{L/K}$.
Therefore we obtain an upper bound on $|\Cl_{L/K}[\ell]| = |\Cl_{L/K}/\ell \Cl_{L/K}|$ by $\Vol( \ACl_{L/K})/ \Vol(\ell \ACl_{L/K})$ where $\ACl_{L/K} = G/P$ and $\ell \ACl_{L/K} = P_{\ell}/P$.
We will give a lower bound on $\Vol(P_{\ell}/P)$ and derive an upper bound for $\Vol(G/P_{\ell})$ using the trivial upper bound for $\Vol(G/P) = \Vol(\ACl_{L/K})$ in Lemma \ref{lem:relative-volume}.
Now we let $\Sigma_L$ be the set of infinite places of $L$, and for a constant $C>1$ we define
$$T:=\{(x, J)\in G \mid J= \wp_i ^{a_i}\bar{\wp}^{-b_i}_i\text{ for some $i$, and } \forall v_1, v_2\in \Sigma_L, C^{-1/\ell}< |x|_{v_1}/|x|_{v_2} < C^{1/\ell} \} \subseteq \Div^0(L),$$
where $(\wp_i, \bar{\wp_i})$ for $i = 1, \cdots, M$ are the pairs of prime ideals in $\O_L$ that are given in the statement of the lemma. We will show that
$$\Vol (P_{\ell}/P) \ge \Vol((\ell T +P)/P) \gg_{[L:\ensuremath{{\mathbb{Q}}}],C} M ,$$
from which it follows that $|\ACl_{L/K}[\ell]| \ll_{[L:\ensuremath{{\mathbb{Q}}}],C} \frac{\Vol(\ACl_{L/K})}{M}.$
The first inequality is clear since there is a natural embedding $\ell T+P \hookrightarrow P_{\ell} = \ell G+P$.
It suffices to prove the second inequality. For each $i$, we let $T_i\subset T$ be the subset of elements with $J=\wp_i ^{a_i} \bar{\wp}_i^{-b_i}$
and let $D_i$ be the corresponding disk $D_i := (\ell T_i+P)/P \subseteq \ACl_{L/K}$.
We now show that for each $i$, the volume of $D_i$ satisfies $\Vol(D_i)\gg_{[L:\ensuremath{{\mathbb{Q}}}],C} 1$.
We can compute that $\Vol(\ell T_i)$ in $G$ is a constant in terms of $C$ and $[L:\ensuremath{{\mathbb{Q}}}]$. If $t_1^{\ell}, t_2^{\ell}\in \ell T_i$ correspond to the same point in $D_i$, then let $u:=t_1^{\ell}/t_2^{\ell}$. We see that $u\in \O_L^{\times}$ since $u$ has no non-archimedean valuations, and moreover it satisfies $C^{-2}< |u|_{v_1}/|u|_{v_2}< C^2$ for any archimedean valuations $v_1, v_2\in \Sigma_L$. Now all archimedean valuations of $u$ are of roughly the same size and their product is $1$, therefore the number of such $u$ can be uniformly bounded in terms of $[L:\ensuremath{{\mathbb{Q}}}]$ and $C$, since the defining polynomial of $u$ has bounded integral coefficients and bounded degree. This then implies that the disk $D_i = (\ell T_i+P)/P$ has $\Vol(D_i) \gg_{[L:\ensuremath{{\mathbb{Q}}}],C} 1$.
On the other hand, we now show that for any $i\neq j$, the disks $D_i$ and $D_j$ do not overlap. Suppose $t_i\in T_i, t_j\in T_j$ with $i\neq j$ satisfy $u:=t_i^{\ell} t_j^{-\ell}\in L^{\times}$.
Then recall that for $u\in \Div^0_L$, we define the height of $u$ as in \cite{EV} to be
$$H(u):= \prod_{v|\infty} \max\{ u_v,1 \} \prod_{\wp} \max\{ |\wp|^{-\text{val}_{\wp}(u)}, 1 \}.$$
Then for the given $u$, we have $H(u) \le C^{2[L:\ensuremath{{\mathbb{Q}}}]} \Nm(\wp_j)^{a_j\ell} \Nm(\bar{\wp}_i)^{b_i\ell } $, which is bounded by $C^{2[L:\ensuremath{{\mathbb{Q}}}]}\Disc(L/K)^{2\theta \ell}$ from our assumption on $(\wp, \bar{\wp})$. By \cite[Lemma 2.2]{EV}, there is a constant $D_0 = D_0(\theta, [L:\ensuremath{{\mathbb{Q}}}])$ such that when $\Disc(L/K)$ is larger than $D_0$, the height of $u$ is not large enough to generate $L$. Therefore $u$ generates a strict subfield $K(u)\subsetneq L$. This is a contradiction, since the non-archimedean part of $u$ has valuations from $\wp_i$ and $\bar{\wp}_i$, which are not extended from any strict intermediate subfield by assumption. Thus, when $\Disc(L/K)$ is sufficiently large, we have shown that $D_i\cap D_j = \emptyset$ for any $i\neq j$.
Therefore by choosing an arbitrary absolute constant $C>1$, we get for $\Disc(L/K)$ sufficiently larger $D_0 = D_0(\theta, [L:\ensuremath{{\mathbb{Q}}}])$ that
$$\Vol((\ell T+P)/P) \ge \sum_i \Vol(D_i) \gg_{[L:\ensuremath{{\mathbb{Q}}}]} M.$$
\end{proof}
\section{Uniform bounds on cubic extensions and the average $3$-part of quadratic extensions}
\label{sec:uniform-cubic}
In this section we are going to bound the sum of $h_3(F/k)$ for quadratic extensions $F/k$ of a general number field $k$, and the closely related
number of $S_3$ cubic extensions of $k$. This proof will employ different techniques and ideas for different regimes of $X$ versus $\Disc(k)$.
Let $N_k(G,X)$ be the number of isomorphism classes of $G$-extensions $L/k$ with $\Disc(L/k)\le X$.
The main theorem is the following, which we expect will be useful for other applications as well.
\begin{theorem}\label{thm:cubic-bound}
Let $k$ be a number field, let $h = h_2(k)$ denote the size of the $2$-torsion subgroup of the class group of $k$, and let $D_k = \Disc(k)$.
For any $X \geq 1$
we have
\[
\sum_{\substack{[F:k]=2\\\Disc(F/k) \leq X}} h_3(F/k)
\ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon}D_k^\epsilon \cdot
\begin{cases}
h X^{3/2} D_k^{1/2}, & \text{if } X \leq D_k h^{-2/3}, \\
h^{1/3} X^{1/2} D_k^{3/2}, & \text{if } D_k h^{-2/3} \leq X \leq D_k^2 h^{2/3}, \\
X D_k^{1/2}, & \text{if } D_k^2 h^{2/3} \leq X \leq D_k^{5/2} h^{2/3}, \\
h^{2/3} D_k^3, & \text{if } D_k^{5/2}h^{2/3} \leq X \leq D_k^3 h^{2/3}, \text{and} \\
X, & \text{if } X \geq D_k^3 h^{2/3}.
\end{cases}
\]
The same upper bound holds for $N_k(S_3, X)$. (The sum is over $F\sub \bar{\ensuremath{{\mathbb{Q}}}}$.)
\end{theorem}
We obtain the following convenient corollary for general number fields $k$.
\begin{corollary}\label{cor:general-bound}
For any number field $k$ and any $X \geq 1$, we have
$$
\sum_{\substack{[F:k]=2\\\Disc(F/k) \leq X}} h_3(F/k)
= O_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon}(\Disc(k)^{1+\epsilon} h_2(k)^{2/3} X).
$$
The same upper bound holds for $N_k(S_3,X)$ under the same hypotheses.
\end{corollary}
Using the trivial bound on $h_2(k)$ in Corollary \ref{cor:general-bound}, the right-hand side becomes $O_{[k:\mathbb{Q}],\epsilon}(\Disc(k)^{4/3+\epsilon} X)$ for any $k$, and when $k$ is a $2$-extension, it follows from Lemma \ref{lem:two-part} that the right-hand side is $O_{[k:\mathbb{Q}],\epsilon}(\Disc(k)^{1+\epsilon} X)$.
\begin{remark}
Theorem \ref{thm:cubic-bound}
relies on an application of the Brauer--Siegel theorem to give a lower bound on the residue $\mathrm{Res}_{s=1} \zeta_k(s)$ of the Dedekind zeta function of $k$. In particular, the implied constant
depends on $\epsilon$ ineffectively.
However,
the key results of Sections \ref{sec:small-cft}--\ref{sec:mid-order} from which Theorem \ref{thm:cubic-bound} is obtained, namely Propositions \ref{prop:trivial-cubic-bound}, \ref{prop:effective-shintani}, and \ref{prop:cubic-bound-order-general}, have effective constants and an explicit dependence on the residue, and could be used to prove an effective analogue of
Theorem \ref{thm:cubic-bound} with dependence on the residue.
\end{remark}
In order to prove Theorem \ref{thm:cubic-bound}, we will use different approaches for three regimes of $X$ compared to $\Disc(k)$: the small range, the large range, and the intermediate range. In Section \ref{sec:small-cft}, for $X$ in the small range, we will use class field theory to give what we regard as a weak bound
on the number of cubic extensions with a fixed discriminant. This will give a correspondingly weak bound on the number of cubic extensions with bounded discriminant. In Section \ref{sec:large-Shintani}, for $X$ in the large range, by using the functional equation of the {Shintani zeta function} for cubic rings, we can use the bound on the coefficients we obtained from class field theory and get improved bounds on the number of cubic fields with discriminants quite large compared to $\Disc(k)$. In Section \ref{sec:mid-order}, we will take advantage of the fact that the bound from the Shintani zeta function is actually also an upper bound for counting all cubic rings. Since the number of cubic orders associated to each cubic field is large, we show it is impossible to get too many cubic fields with $X$ in the intermediate range. Finally, in Section \ref{sec:effective}, we prove Theorem \ref{thm:cubic-bound} by combining the previous propositions for all ranges of $X$.
\subsection{Small range: class field theory}\label{sec:small-cft}
In this section we will use class field theory to give the following upper bound on the number of small degree extensions of $k$.
\begin{proposition}\label{prop:trivial-cubic-bound}
For any number field $k$ and any $X \geq 1$, we have
$$N_k(S_3, X)+ N_k(C_3, X)+ N_k(S_2, X) = O_{[k:\mathbb{Q}],\epsilon}(X^{3/2+\epsilon}\mathrm{Disc}(k)^{1/2+\epsilon} h_2(k)).$$
\end{proposition}
To prove this proposition, we apply class field theory and the trivial bound on torsion in relative class groups
to give the following pointwise bound on the number of relative $S_3$ cubic extensions.
\begin{lemma}\label{lem:pointwise-bound}
For any positive integer $n$ and any number field $k$, the number of extensions $K/k$ in $\bar{\ensuremath{{\mathbb{Q}}}}$ with $\mathrm{Disc}(K/k)=n$ and $[K:k]\leq 3$
is
$O_{[k:\mathbb{Q}],\epsilon}(n^{1/2+\epsilon} h_2(k) \mathrm{Disc}(k)^{1/2+\epsilon}).$
\end{lemma}
\begin{proof}
If $K/k$ is an $S_3$ cubic extension with $\mathrm{Disc}(K/k)=n$, then its associated quadratic resolvent $F/k$ has $\Disc(F/k) = m$ and $m|\Disc(K/k)$. By class field theory, quadratic extensions of $k$ are in bijection to surjective homomorphisms from the id\`ele class group $C_k = \mathbb{A}^{\times}_k/k^{\times}$ to $C_2$, where $\mathbb{A}^{\times}_k$ is the id\`ele group. We have the following exact sequence of $C_k$,
$$1\rightarrow O_k^{\times} \rightarrow \prod_v O_{v}^{\times}\rightarrow C_k \rightarrow \Cl_k\rightarrow 1.$$
Since $\Hom(\cdot, A)$ is left exact, we get
$$1\rightarrow \Hom(\Cl_k, C_2 ) \rightarrow \Hom(C_k, C_2) \rightarrow \Hom( \prod_v O_{v}^{\times}, C_2).$$
Here $|\Hom(\Cl_k, C_2)| = h_2(k)$. Given a $\rho\in \Hom(C_k, C_2)$, its image in $\Hom( \prod_v O_{v}^{\times}, C_2)$ determines the relative discriminant ideal of the quadratic field corresponding to $\rho$. Therefore for each fixed $m$, the number of possible $F/k$ with $\Disc(F/k)=m$ is bounded by $h_2(k)$.
For each fixed $n$, there are at most $O_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon}(n^{\epsilon})$ possible $m$, therefore there are at most $O_{[k:\mathbb{Q}],\epsilon}(n^\epsilon h_2(k))$ quadratic extensions $F/k$ with $\mathrm{Disc}(F/k)|n$.
Similarly, for each fixed quadratic extension $F/k$, any associated $S_3$ cubic extension $K/k$ has Galois closure $\tilde{K}/k$ such that $\tilde{K}/F$ has $\Gal(\tilde{K}/F) = C_3$. Thus, we can use class field theory again to determine that, for each fixed $F/k$, the number of associated $S_3$ cubic extensions with $\Disc(K/k)|n$
is
$$\ll_{[k:\mathbb{Q}],\epsilon} n^{\epsilon} h_3(F/k) \ll_{[k:\mathbb{Q}],\epsilon} n^{\epsilon}\mathrm{Disc}(F/k)^{1/2+\epsilon}\mathrm{Disc}(k)^{1/2+\epsilon} \ll_{[k:\mathbb{Q}],\epsilon} n^{1/2+\epsilon} \mathrm{Disc}(k)^{1/2+\epsilon},$$
by the trivial bound on $h_3(F/k)$ in Lemma \ref{lem:relative-trivial}. Altogether, the total number of $S_3$-extensions is now $O_{[k:\mathbb{Q]},\epsilon}(n^{1/2+\epsilon} h_2(k) \mathrm{Disc}(k)^{1/2+\epsilon})$, as claimed. As above, we can bound the number of $C_3$-extensions $F/k$ by $O_{[k:\mathbb{Q}],\epsilon}(n^\epsilon h_3(k))$
and the lemma follows.
\end{proof}
Summing Lemma \ref{lem:pointwise-bound} over integers $n$ immediately yields Proposition \ref{prop:trivial-cubic-bound}.
\subsection{Large range: the Shintani zeta function}\label{sec:large-Shintani}
In this section, we are going to give a better bound for $N_k(S_3, X)$ when $X$ is very large compared to $\Disc(k)$. The main idea is to use the bound
in Lemma \ref{lem:pointwise-bound} to get a better estimate on $N_k(S_3, X)$ using \emph{Shintani zeta functions} over a general number field $k$. Moreover, the upper bound we obtain in this section is also an upper bound on the number of cubic rings over $k$, not just cubic fields. This will be important in Section \ref{sec:mid-order}.
Given a number field $k$, Shintani zeta functions over $k$ count \emph{cubic rings} over $k$ with a specified signature. More specifically, given a cubic ring $R$ over $k$ (that is, a ring that is a locally free rank three $\mathcal{O}_k$-module), let $\disc(R/\mathcal{O}_k)$ denote the relative discriminant ideal of $R$ and $\mathrm{Disc}(R/\mathcal{O}_k) = \Nm_{k/\ensuremath{{\mathbb{Q}}}}(\disc(R/\mathcal{O}_k))$.
For each archimedean place $v$ of $k$, let $k_v$ be the completion of $k$ at $v$, and let $R_v:= R\tensor_{\O_k} k_v$.
Each $R_v$ is a cubic \'etale algebra over $k_v$. If $k_v \simeq \mathbb{C}$, then there is only one such cubic \'etale algebra $\mathbb{C}^3$, while if $k_v \simeq \mathbb{R}$, then either $R_v \simeq \mathbb{R}^3$ or $R_v \simeq \mathbb{R}\times \mathbb{C}$. The signature of $R$ is the tuple $\mathrm{sgn}(R) = (R_v)_{v\mid \infty}$ that records the isomorphism class of each $R_v$.
We will use $N^3_k(X)$ to denote the number of isomorphism classes of cubic rings $K/k$ over a number field $k$ with $0<\Disc(K/k)\le X$, and $N^3_{k, \alpha}(X)$ to denote the number of isomorphism classes of cubic rings $K/k$ over a number field $k$ with $0<\Disc(K/k)\le X$ with signature $\alpha$.
Then the main proposition we are going to prove for this subsection is as follows.
\begin{proposition}\label{prop:effective-shintani}
For any number field $k$
and $X \geq \mathrm{Disc}(k)^3 h_2(k)^{2/3}$, we have
$$N^3_k(X) \ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^\epsilon X.$$
\end{proposition}
We start by introducing properties of Shintani zeta functions over a general number field, including all useful notations. For a fixed signature $\alpha$, let
\[
\xi_{k,\alpha}(s) := \sum_{\begin{subarray}{c} R/\mathcal{O}_k: \\ \,\mathrm{Disc}(R/\mathcal{O}_k) \neq 0, \\ \mathrm{sgn}(R) = \alpha \end{subarray}} \frac{|\mathrm{Aut}_{\O_k}(R)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s},
\]
where the sum runs over isomorphism classes of cubic rings $R$ over $k$.
Let $\zeta_k(s)$ be the Dekekind zeta function of $k$.
For convenience in what is to come, we define
\[
\mathfrak{A}_k := \frac{\zeta_k(2) \mathrm{Res}_{s=1}\zeta_k(s)}{2^{r_1(k)+r_2(k)+1}} \text{ and } \mathfrak{B}_k := \frac{3^{r_1(k)+r_2(k)/2}\zeta_k(1/3) \mathrm{Res}_{s=1} \zeta_k(s)}{6\cdot 2^{r_1(k)+r_2(k)} \mathrm{Disc}(k)^{1/2}} \left(\frac{\Gamma(1/3)^3}{2\pi}\right)^{[k:\mathbb{Q}]}.
\]
Additionally, for each pair of signatures $\alpha,\beta$ and each infinite place $v \mid \infty$ of $k$, we define a factor $c_{v,\alpha\beta}(s)$ as follows. If $k_v \simeq \mathbb{C}$, set $c_{v,\alpha\beta}(s) = \sin^2(\pi s)\sin(\pi s - \pi/6) \sin(\pi s + \pi/6)$. If $k_v \simeq \mathbb{R}$ and $\alpha_v \simeq \beta_v$, set $c_{v,\alpha\beta}(s) = \sin(2\pi s)/2$. In the remaining cases, set
\begin{equation}
\begin{aligned}
c_{v,\alpha\beta}(s) = \frac{1}{2} \left\{\begin{array}{ll} 3\sin(\pi s), & \text{if } \alpha_v \simeq \mathbb{R}^3, \beta_v \simeq \mathbb{R}\times\mathbb{C}, \\ \sin(\pi s), & \text{if } \alpha_v \simeq \mathbb{R}\times\mathbb{C}, \beta_v \simeq \mathbb{R}^3, \end{array}\right.
\end{aligned}
\end{equation}
and define $c_{\alpha\beta}(s) := \prod_{v\mid \infty} c_{v,\alpha\beta}(s)$.
The basic properties of $\xi_{k,\alpha}(s)$ are recorded in the following proposition that collects results due to Shintani \cite{Shintani}, Wright \cite{WrightThesis,WrightAdelic}, and Datskovsky and Wright \cite{DW86}. We also refer the interested reader to the work of Taniguchi \cite{Taniguchi} for more information.
\begin{proposition} \label{prop:shintani}
We have the following properties for Shintani zeta functions over a general number field $k$:
\begin{enumerate}
\item For each number field $k$ and each signature $\alpha$, the function $\xi_{k,\alpha}(s)$ converges absolutely in the region $\Re(s)>1$.
\item The function $\xi_{k,\alpha}(s)$ has a meromorphic continuation to $\mathbb{C}$ with poles only at $s=1$ and $s=5/6$. Each of these poles is simple, with residues
\[
\mathfrak{A}_k\cdot(1+3^{-r(\alpha)-r_2(k)}) \text{ and } \mathfrak{B}_k\cdot 3^{-r(\alpha)/2}
\]
respectively, where $r(\alpha) = \#\{ v \mid \infty : \alpha_v \simeq \mathbb{R}^3\}$.
\item The function $\xi_{k,\alpha}(s)$ satisfies a functional equation,
\[
\xi_{k,\alpha}(1-s) = \left(\frac{3^{(6s-2)}}{\pi^{4s}}\Gamma(s)^{2}\Gamma(s-1/6)\Gamma(s+1/6)\right)^{[k:\mathbb{Q}]} \mathrm{Disc}(k)^{4s-2} \sum_{\beta} c_{\alpha\beta}(s) \hat\xi_{k,\beta}(s),
\]
where the sum runs over possible signatures $\beta$ and where $\hat\xi_{k,\beta}(s)$ is the {dual Shintani zeta function} defined via the Dirichlet series
\[
\hat\xi_{k,\beta}(s) = \sum_{\substack{R/\mathcal{O}_k: \\ \mathrm{Disc}(R/\mathcal{O}_k) \neq 0 \\ \mathrm{sgn}(R)=\beta \\ 3\mid \mathrm{tr}(t)\, \forall t \in R}} \frac{|\mathrm{Aut}_{\O_k}(R)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s},
\]
where the sum runs over isomorphism classes of cubic rings $R$ over $k$.
\item
For each $\alpha$, the function $\hat\xi_{k,\alpha}(s)$ has a meromorphic continuation to all of $\mathbb{C}$ with poles at $s=1$ and $s=5/6$ (both simple), and satisfies the inequality $\hat\xi_{k,\alpha}(s) < \xi_{k,\alpha}(s)$ for real $s>1$.
\item
For each $\alpha$, the functions $(s-1)(s-5/6) \xi_{k,\alpha}(s)$ and $(s-1)(s-5/6)\hat\xi_{k,\alpha}(s)$ are entire of order $1$.
\end{enumerate}
\end{proposition}
\begin{proof}
We provide references for these claims without regard to necessarily providing the original reference. The first claim follows from \cite[Theorem 4.1]{WrightAdelic}. The second through fourth primarily follow from \cite[Theorem 6.2]{DW86}, with the claim that $\hat\xi_{k,\alpha}(s) < \xi_{k,\alpha}(s)$ for real $s>1$ following from the absolute convergence of $\xi_{k,\alpha}(s)$ in this region. The fifth follows from a generalization of \cite[Theorem 2.1.iii]{Shintani} as is explained on \cite[pg. 96]{WrightThesis}.
\end{proof}
We next record one more useful property of $\xi_{k,\alpha}(s)$.
\begin{lemma}\label{lem:shintani-vanishing}
If $k$ has degree at least $2$, then $\xi_{k,\alpha}(0)=0$ for each signature $\alpha$.
\end{lemma}
\begin{proof}
From the formulas given, it follows that each $c_{\alpha\beta}(s)$ has a zero of order $[k:\mathbb{Q}]$ at $s=1$. The claim then follows from the functional equation for $\xi_{k,\alpha}(s)$.
\end{proof}
While Proposition \ref{prop:shintani} establishes that $\xi_{k,\alpha}(s)$ converges absolutely in the region $\Re(s)>1$, it does not guarantee that this convergence is uniform in $k$. To control this uniformity, we must control the number of cubic rings over $k$ with small discriminant. This will essentially be afforded to us by Proposition \ref{prop:trivial-cubic-bound} on the number of small discriminant cubic algebras over $k$, except that $\xi_{k,\alpha}(s)$ counts all orders in cubic \'etale algebras over $k$, not just the maximal orders. We will later exploit this overcounting in the proof of Proposition \ref{prop:order-bound-sf}, but for now, we simply record the following result of Datskovsky and Wright \cite[Theorem 6.1]{DW86} in order to track the number of orders in a cubic \'etale algebra $A/k$.
\begin{lemma}[Datskovsky--Wright]\label{lem:datskovsky-wright}
Let $A$ be a cubic \'etale algebra over $k$ with maximal order $\mathcal{O}_A$. Thus, either $A\simeq k^3$, $A \simeq F \times k$ where $F/k$ is a quadratic extension, or $A \simeq K$ where $K/k$ is a relative cubic field extension. For each $n$, let $a_n$ denote the number of orders $\mathcal{O} \subset A$ for which $[\mathcal{O}_A:\mathcal{O}]^2 = n$, equivalently, the number of orders $\mathcal{O}\subset A$ for which $\Disc(\mathcal{O}/\mathcal{O}_k) = \Disc(\mathcal{O}_A/\mathcal{O}_k) \cdot n$.
Then the Dirichlet series $f_A(s):=\sum_{n} a_n n^{-s}$ satisfies
\[
f_A(s)=\zeta_k(4s) \zeta_k(6s-1) \frac{\zeta_A(2s)}{\zeta_A(4s)},
\]
where
\[
\zeta_A(s) =
\begin{cases}
\zeta_k(s)^3, & \text{if } A \simeq k^3, \\
\zeta_k(s) \zeta_F(s), & \text{if } A \simeq F \times k,\\
\zeta_K(s), & \text{if } A \simeq K,
\end{cases}
\]
with notation as above, and where for a number field $L$, $\zeta_L(s)$ denotes the Dedekind zeta function of $L$.
\end{lemma}
Combining Lemma \ref{lem:datskovsky-wright} and Lemma \ref{lem:pointwise-bound}, we derive the following lemma on the values of Shintani zeta function in a right half-plane.
\begin{lemma}\label{lem:shintani-bound}
For any real $\sigma>3/2$ and any $t \in \mathbb{R}$, we have $\xi_{k,\alpha}(\sigma+it) \ll_{[k:\mathbb{Q}],\sigma,\epsilon} \mathrm{Disc}(k)^{1/2+\epsilon}h_2(k)$.
\end{lemma}
\begin{proof}
By the triangle inequality, it suffices to prove the lemma when $t=0$. Assume $\sigma>3/2$ is given.
With the notation of Lemma \ref{lem:datskovsky-wright}, for any cubic \'etale algebra $A$, we have $\zeta_A(\sigma) \leq \zeta_k(\sigma)^3$ whenever $\sigma > 1$. It follows that
\begin{equation}
\begin{aligned}
\xi_{k,\alpha}(\sigma) & \leq \sum_{[L:k]\leq 3} \frac{1}{\mathrm{Disc}(L/k)^\sigma} \cdot \zeta_k(4\sigma) \zeta_k(6\sigma-1) \frac{\zeta_A(2\sigma)}{\zeta_A(4\sigma)} \\
& \leq \zeta_k(2\sigma)^3\zeta_k(4\sigma)\zeta_k(6\sigma-1) \sum_{[L:k]\leq 3} \frac{1}{\mathrm{Disc}(L/k)^\sigma},
\end{aligned}
\end{equation}
where the sum runs over isomorphism classes of field extensions $L/k$ of degree at most $3$. The second inequality comes from $\zeta_A(4\sigma)\ge 1$, valid for $\sigma>1$. The factors of $\zeta_k(\sigma)$ may be bounded by suitable powers of the Riemann zeta function $\zeta(\sigma)$,
depending only on the degree of $k/\mathbb{Q}$.
So it suffices to understand the Dirichlet series of field discriminants.
By Lemma \ref{lem:pointwise-bound}, we then find that for any $\epsilon < \sigma - 3/2$,
\[
\sum_{L/k} \frac{1}{\mathrm{Disc}(L/k)^\sigma} \ll_{[k:\mathbb{Q}],\epsilon} h_2(k)\mathrm{Disc}(k)^{1/2+\epsilon} \sum_{n}\frac{n^{1/2+\epsilon}}{n^\sigma} \ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon} \mathrm{Disc}(k)^{1/2+\epsilon}h_2(k) \zeta(\sigma-1/2-\epsilon),
\]
yielding the lemma.
\end{proof}
Next, we apply the functional equation to deduce a discriminant-aspect ``convexity bound'' for the Shintani zeta function.
\begin{lemma}\label{lem:shintani-convexity}
For any $-1/2 \leq \sigma \leq 3/2$ and any $t \in \mathbb{R}$, the Shintani zeta function satisfies
\[
\xi_{k,\alpha}(\sigma + it)
=O_{[k:\mathbb{Q}],\epsilon}( h_2(k) \mathrm{Disc}(k)^{\frac{7}{2}-2\sigma+\epsilon} (1+|t|)^{2[k:\mathbb{Q}](\frac{3}{2}-\sigma)+\epsilon}).
\]
\end{lemma}
\begin{proof}
Let $\delta > 0$ be small. For $\sigma = 3/2+\delta$, it follows from Lemma \ref{lem:shintani-bound} that
\[
\xi_k(\sigma+it)
\ll_{[k:\mathbb{Q}],\delta,\epsilon} \Disc(k)^{1/2+\epsilon} h_2(k).
\]
For $s = \sigma + it$ with $\sigma = 3/2 + \delta$, we apply the functional equation and Stirling's formula to extract cancellation between the gamma factors and the
$c_{\alpha\beta}(s)$ in order to find also
\begin{align*}
\xi_{k,\alpha}(1-s)
&= \left(\frac{3^{(6s-2)}}{\pi^{4s}}\Gamma(s)^{2}\Gamma(s-1/6)\Gamma(s+1/6)\right)^{[k:\mathbb{Q}]} \mathrm{Disc}(k)^{4s-2} \sum_{\beta} c_{\alpha\beta}(s) \hat\xi_{k,\beta}(s) \\
&\ll_{[k:\mathbb{Q}],\delta,\epsilon} (1+|t|)^{(4+4\delta)[k:\mathbb{Q}]+\epsilon} \mathrm{Disc}(k)^{9/2+4\delta+\epsilon} h_2(k).
\end{align*}
The general result follows upon using the Phragmen--Lindel\"of principle \cite[Theorem 5.53]{IwaniecKowalski} applied to the function $\frac{(s-1)(s-5/6)}{(s+1)^2} \xi_{k,\alpha}(s)$ and letting $\delta$ be sufficiently small. Specifically, we find for any $-1/2- \delta < \sigma < 3/2 + \delta$ that
\[
\xi_{k,\alpha}(\sigma+it)
\ll_{[k:\mathbb{Q}],\delta,\epsilon} (1+|t|)^{2[k:\mathbb{Q}](\frac{3}{2}+\delta-\sigma)+\epsilon} \Disc(k)^{\frac{7}{2}+2\delta-2\sigma+\epsilon}.
\]
Upon choosing $\delta = \epsilon$ and then replacing $\epsilon$ in the equation above by $\epsilon/(1+2[k:\mathbb{Q}])$, we obtain the required statement.
\end{proof}
Now we apply the functional equation in Proposition \ref{prop:shintani} and the above lemmas to bound the number of cubic rings $N^3_{k, \alpha}(X)$.
\begin{proof}[Proof of Proposition \ref{prop:effective-shintani}]
Let $\phi(x)$ be a positive smooth function satisfying $\phi(x)\geq 1$ whenever $0\leq x \leq 1$ and whose Mellin transform exhibits rapid decay away from the real axis; for example, we may take $\phi(x)=e^{1-x}$. Letting $R$ run over cubic rings over $k$, we have
\[
N^3_{k,\alpha}(X)
= \sum_{\substack{ 0<\mathrm{Disc}(R/\mathcal{O}_k) \leq X \\ \mathrm{sgn}(R) = \alpha}} 1
\ll_\phi \sum_{\substack{0<\mathrm{Disc}(R/\mathcal{O}_k) \\ \mathrm{sgn}(R) = \alpha}} \frac{\phi(\mathrm{Disc}(R/\mathcal{O}_k)/X)}{|\mathrm{Aut}_{\O_k}(R)|}.
\]
This latter term may be evaluated via Perron's formula and the full Shintani zeta function $\xi_{k,\alpha}(s)$,
\[
\sum_{\substack{0<\mathrm{Disc}(R/\mathcal{O}_k) \\ \mathrm{sgn}(R) = \alpha}} \frac{\phi(\mathrm{Disc}(R/\mathcal{O}_k)/X)}{|\mathrm{Aut}_{\O_k}(R)|} = \frac{1}{2\pi i} \int_{2-i\infty}^{2+i\infty} \xi_{k,\alpha}(s) \Phi(s) X^s \,ds,
\]
where $\Phi(s)$ is the Mellin transform of $\phi(x)$. By Lemma \ref{lem:shintani-convexity}, the exponential decay of $\Phi(s)$ off the real axis allows us to shift the contour to the line $\Re(s)=-1/2-\epsilon$ for any $\epsilon>0$. Doing so, we must account for the poles of $\xi_{k,\alpha}(s)$ at $s=1$ and $s=5/6$ and of $\Phi(s)$ at $s=0$. From Proposition \ref{prop:shintani}, it follows that the contribution from the pole at $s=1$ subsumes that of the pole at $s=5/6$. Moreover, Lemma \ref{lem:shintani-vanishing} shows that in fact the integrand has no pole at $s=0$ owing to the zero of $\xi_{k,\alpha}(s)$ at $s=0$. Altogether, the contribution from the poles is seen to be $O_{[k:\mathbb{Q}]}(X\cdot \mathrm{Res}_{s=1}\zeta_k(s))$.
Finally, appealing to the functional equation and Lemma \ref{lem:shintani-bound}, we find
\begin{align*}
&\frac{1}{2\pi i} \int_{-1/2-\epsilon-i\infty}^{-1/2-\epsilon+i\infty} \xi_{k,\alpha}(s) \Phi(s) X^s \,ds \\
\quad
&= \frac{1}{2\pi i} \int_{3/2+\epsilon-i\infty}^{3/2+\epsilon+i\infty} \left(\frac{3^{(6s-2)}}{\pi^{4s}}\Gamma(s)^{2}\Gamma(s-1/6)\Gamma(s+1/6)\right)^{[k:\mathbb{Q}]} \mathrm{Disc}(k)^{4s-2} \hat\xi(s) \Phi(1-s) X^{1-s} \,ds \\
&\ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^{9/2+\epsilon} h_2(k) X^{-1/2-\epsilon},
\end{align*}
where $\hat\xi(s) = \sum_{\beta}c_{\alpha\beta}(s)\hat\xi_{K,\beta}(s)$.
We conclude that for any number field $k$ and any $X \geq 1$, we have
\begin{equation}\label{eqn:prop:shintani-cubic-bound}
N^3_{k,\alpha}(X) \ll_{[k:\mathbb{Q}],\epsilon} X \cdot \mathrm{Res}_{s=1} \zeta_k(s) + X^{-1/2-\epsilon} h_2(k) \mathrm{Disc}(k)^{9/2+\epsilon}.
\end{equation}
Using the well-known upper bound $\mathrm{Res}_{s=1} \zeta_k(s) \ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^\epsilon$
due to Landau (see also \cite[Lemma 3]{Brauer}),
we conclude the proposition.
\end{proof}
\subsection{Intermediate range: propagation of orders}\label{sec:mid-order}
In this section, our main proposition is the following bound on $N_k(S_3, X)$ that is better than both Proposition \ref{prop:trivial-cubic-bound} and \ref{prop:effective-shintani} when $X$ is a bit smaller than the region allowed by Proposition \ref{prop:effective-shintani}.
Let $\kappa_k:=\mathrm{Res}_{s=1}\zeta_k(s)$.
\begin{proposition}\label{prop:cubic-bound-order-general}
Let $k$ a number field and let $X \geq 1$. We have
\[
N_k(S_3, X)
\ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^{1/2+\epsilon} \kappa_k^{-2} X^{1/2} ( {\mathrm{Disc}(k) h_2(k)^{1/3}}+ {X^{1/2} }).
\]
\end{proposition}
The key idea is that Proposition \ref{prop:effective-shintani} actually gives an upper bound on the number of cubic rings over $k$ instead of cubic fields, and if there are too many cubic fields in this intermediate range, then the orders inside these fields would overrun these bounds. We will first apply this idea to bound the average $3$-class number of quadratic extensions of $k$,
and then we apply class field theory to bound the number $N_k(S_3, X)$ of general $S_3$ cubic fields.
We begin with the following lemma.
\begin{lemma}\label{lem:gen-series-order-F}
Let $F/k$ be a quadratic extension of number fields, and let $\mathcal{R}(F)$ denote the set of isomorphism classes of
cubic rings $R$ over $k$
whose underlying cubic \'etale algebra $A:=R\tensor_\ensuremath{{\mathbb{Q}}} k$
has discriminant $\disc(F/k)$ and is either of the form $A \simeq k \times F$ or $A \simeq K$ for a non-Galois cubic extension $K/k$ with quadratic resolvent $F/k$.
Then we have
\[
\sum_{R \in \mathcal{R}(F)} \frac{|\mathrm{Aut}_{k}(A)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s}
= \frac{h_3(F/k)}{2 \mathrm{Disc}(F/k)^{s}} \zeta_k(2s) \zeta_k(6s-1) \sum_{\substack{\mathfrak{a} \subseteq \mathcal{O}_F \,\mathrm{sq. free} \\ [\mathfrak{a}] \in 3\mathrm{Cl}_F+\Cl_k}} \frac{1}{\Nm(\mathfrak{a})^{2s}},
\]
where the summation on the right-hand side runs over those squarefree ideals of $\mathcal{O}_F$ whose classes are in the subgroup $3\Cl_F+\Cl_k \subseteq \Cl_F$.
\end{lemma}
\begin{proof}
If $R \in \mathcal{R}(F)$, then $\mathrm{Disc}(R/\mathcal{O}_k) = \mathrm{Disc}(F/k) [\mathcal{O}_A:R]^2$. Thus, with the notation of Lemma \ref{lem:datskovsky-wright}, we have
\[
\sum_{R \in \mathcal{R}(F)} \frac{|\mathrm{Aut}_{k}(A)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s}
= \mathrm{Disc}(F/k)^{-s} \cdot\left(\frac{1}{2}f_{k\times F}(s) + \sum_{\substack{[K:k] =3 \\ \disc(K/k) = \disc(F/k) \\ F \subseteq \widetilde{K}/k}} f_K(s) \right).
\]
Note that the summation on the right is restricted to those cubic extensions $K/k$ with quadratic resolvent $F$ and which satisfy $\disc(K/k) =\disc(F/k)$. For each such $K/k$, by Lemma \ref{lem:datskovsky-wright} we have
\begin{equation}
f_K(s)
=\zeta_k(4s) \zeta_k(6s-1) \frac{\zeta_K(2s)}{\zeta_K{(4s)}}
= \zeta_k(2s) \zeta_k(6s-1)\frac{L(2s,\chi_{\widetilde{K}/F})}{L(4s,\chi_{\widetilde{K}/F})},
\end{equation}
where $\widetilde{K}$ is the Galois closure of $K$ over $k$, and $\chi_{\widetilde{K}/F}$ is one of the non-trivial cyclic cubic characters of the class group $\mathcal{C}= \mathrm{Cl}_F/\mathrm{Cl}_k$ that cuts out $\widetilde{K}/F$.
The second equality comes from the relation $\zeta_K(s)/\zeta_k(s)= L(s, \chi_{\widetilde{K}/F})$. In fact, we have $L(s,\chi_{\widetilde{K}/F}) = L(s,\chi_{\widetilde{K}/F}^2) = L(s, \overline{\chi}_{\widetilde{K}/F})$, since all are equal to the irreducible degree $2$ Artin $L$-function of $K/k$. Thus, we obtain also
\[
f_K(s)
= \zeta_k(2s) \zeta_k(6s-1)\frac{L(2s,\chi_{\widetilde{K}/F})}{L(4s,\chi_{\widetilde{K}/F}^2)}.
\]
Meanwhile, the Dirichlet series for orders inside $F\times k$ is
\begin{equation}
f_{F\times k}(s) = \zeta_k(2s) \zeta_k(6s-1)\frac{\zeta_F(2s)}{\zeta_F(4s)}.
\end{equation}
Notice that $\zeta_F(s)$ is the $L$-function attached to the trivial character of $\mathcal{C}$. Therefore,
\begin{align*}
\sum_{R \in \mathcal{R}(F)} \frac{|\mathrm{Aut}_{k}(A)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s}
& = \frac{ \zeta_k(2s) \zeta_k(6s-1)}{2\mathrm{Disc}(F/k)^{s}} \sum_{\chi \in \widehat{\mathcal{C}}[3]} \frac{L(2s,\chi_{\widetilde{K}/F})}{L(4s,\chi_{\widetilde{K}/F}^2)} \\
&= \frac{ \zeta_k(2s) \zeta_k(6s-1)}{2\mathrm{Disc}(F/k)^{s}} \sum_{\mathfrak{a}\subseteq\mathcal{O}_F\,\text{sq. free}} \frac{1}{\Nm(\mathfrak{a})^{2s}} \sum_{\chi \in \widehat{\mathcal{C}}[3]}\chi(\mathfrak{a})\\
& =\frac{ h_3(F/k)\zeta_k(2s) \zeta_k(6s-1)}{2\mathrm{Disc}(F/k)^{s}} \sum_{\substack{\mathfrak{a}\subseteq \mathcal{O}_F\,\text{sq. free}\\ \mathfrak{a}\in 3\Cl_F + \Cl_k }} \frac{1}{\Nm(\mathfrak{a})^{2s}},
\end{align*}
by orthogonality of characters and the relation $|\widehat{\mathcal{C}}[3]| = h_3(F/k)$.
\end{proof}
We next have the following simple lemma.
\begin{lemma}\label{lem:ideal-lower-bound}
Let $k$ be a number field and let $\delta>0$. There is an effectively computable constant $C$, depending at most on $\delta$ and $[k:\mathbb{Q}]$, such that for any $X \geq C \mathrm{Disc}(k)^{1/2+\delta} \kappa_k^{-1}$,
we have
\[
\#\{\mathfrak{a} \subseteq \mathcal{O}_K \text{ ideal}: \Nm(\mathfrak{a}) \leq X\}
\gg_{[k:\mathbb{Q}],\delta} \kappa_k X .
\]
\end{lemma}
\begin{proof}
Let $m \geq \frac{[k:\mathbb{Q}]}{2}+2$ be an integer. Then
\begin{align*}
\#\{\mathfrak{a} \subseteq \mathcal{O}_K \text{ ideal}: \Nm(\mathfrak{a}) \leq X\}
\geq \sum_{\Nm(\mathfrak{a}) \leq X} \left(1 - \frac{\Nm(\mathfrak{a})}{X}\right)^m
= \frac{m!}{2\pi i} \int_{2-i\infty}^{2+i\infty} \zeta_k(s) X^s \frac{ds}{s(s+1)\dots(s+m)}.
\end{align*}
The convexity bound for $\zeta_k(s)$ (see \cite[Equation (5.20)]{IwaniecKowalski}, for example) implies that when $\sigma:=\Re(s)$ lies in the critical strip $0 \leq \sigma \leq 1$, we have
\begin{equation}\label{eqn:convexity-dedekind}
(s-1) \zeta_k(s) \ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^{(1-\sigma)/2+\epsilon} (1+|s|)^{\frac{[k:\mathbb{Q}](1-\sigma)}{2}+1+\epsilon}.
\end{equation}
Thus, the integral above conveges absolutely when the line of integration is moved anywhere inside the critical strip. Taking it arbitrarily close to the line $\Re(s)=0$, we conclude that
\[
\#\{\mathfrak{a} \subseteq \mathcal{O}_K \text{ ideal}: \Nm(\mathfrak{a}) \leq X\}
\gg_{[k:\mathbb{Q}],\epsilon} X \cdot\mathrm{Res}_{s=1} \zeta_k(s) + O(X^\epsilon \mathrm{Disc}(k)^{\frac{1}{2}+\epsilon}).
\]
The result follows.
\end{proof}
We can now apply our count of orders to get an upper bound on the average $3$-class number of quadratic extensions.
\begin{proposition}\label{prop:order-bound-sf}
Let $k$ a number field and let $X \geq 1$. We have
\[
\sum_{\substack{[F:k]=2 \\ \mathrm{Disc}(F/k) \leq X}} h_3(F/k)
\ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^{1/2+\epsilon} \kappa_k^{-2} X^{1/2} ( {\mathrm{Disc}(k) h_2(k)^{1/3}}+ {X^{1/2} }).
\]
\end{proposition}
\begin{proof}
Let $Y$ be given by
\begin{equation}\label{eqn:Y-value}
Y := \max\left\{ { \mathrm{Disc}(k)^3 h_2(k)^{2/3}},\kappa_k^{-2} {X C^2\mathrm{Disc}(k)^{1+\epsilon}} \right\},
\end{equation}
where $C$ is the constant from Lemma \ref{lem:ideal-lower-bound}.
For each quadratic extension $F/k$ with discriminant at most $X$, we consider the contribution to $N^3_k(Y)$ from the orders in the family $\mathcal{R}(F)$ from Lemma \ref{lem:gen-series-order-F}.
For any $R \in \mathcal{R}(F)$, any ring automorphism extends to an automorphism of the \'etale algebra $A = R \otimes_\mathbb{Q} k$. Thus, $|\mathrm{Aut}_{\mathcal{O}_k} (R)| \leq |\mathrm{Aut}_k(A)|$, and we obtain an inequality of generating functions,
\[
\sum_{R \in \mathcal{R}(F)} \frac{|\mathrm{Aut}_{\mathcal{O}_k} (R)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s}
\geq \sum_{R \in \mathcal{R}(F)} \frac{|\mathrm{Aut}_{k}(A)|^{-1}}{\mathrm{Disc}(R/\mathcal{O}_k)^s},
\]
which is the Dirichlet series from Lemma \ref{lem:gen-series-order-F}. Appealing to Lemma \ref{lem:gen-series-order-F}, this series is in turn bounded below by
\[
\frac{h_3(F/k)}{2 \mathrm{Disc}(F/k)^{s}} \zeta_k(2s) \zeta_k(6s-1)
\geq \frac{h_3(F/k)}{2 \mathrm{Disc}(F/k)^s} \zeta_k(2s),
\]
since the summation over ideals $\mathfrak{a}$ has non-negative coefficients. It then follows from our assumption
$Y\geq \kappa_k^{-2} {X C^2\mathrm{Disc}(k)^{1+\epsilon}} $
that we may appeal to Lemma \ref{lem:ideal-lower-bound} (with $X = Y^{1/2}/\mathrm{Disc}(F/k)^{1/2}$) to find that the contribution to $N^3_k(Y)$ of orders arising in $\mathcal{R}(F)$ is
\[
\gg_{[k:\mathbb{Q}],\epsilon} \frac{h_3(F/k) Y^{1/2}
\kappa_k
}{\mathrm{Disc}(F/k)^{1/2}}.
\]
Thus,
\[
N^3_k(Y)
\gg_{[k:\mathbb{Q}],\epsilon} \sum_{\substack{[F:k]=2 \\ X/2 \leq \mathrm{Disc}(F/k) \leq X}} \frac{h_3(F/k) Y^{1/2}
\kappa_k
}{\mathrm{Disc}(F/k)^{1/2}},
\]
and we conclude that
\[
\sum_{\substack{[F:k]=2 \\ X/2 \le \mathrm{Disc}(F/k) \leq X}} h_3(F/k)
\ll_{[k:\mathbb{Q}],\epsilon} \frac{N^3_k(Y) X^{1/2} }{Y^{1/2}
\kappa_k}.
\]
Finally, by Proposition \ref{prop:effective-shintani}, we have
$N^3_k(Y) \ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^{\epsilon/2} Y$
whenever
$
Y\geq\mathrm{Disc}(k)^3h_2(k)^{2/3},$
and the result follows
using dyadic summation and the fact that $\kappa_k \ll_{[k:\mathbb{Q}],\epsilon} \mathrm{Disc}(k)^{\epsilon/2}$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:cubic-bound-order-general}]
Following the proof of Lemma $6.2$ in \cite{DW88}, for a square-free ideal $\mathfrak{q}$ of $k$ of norm $\Nm \mathfrak{q} \leq X^{1/2}$, the number of $S_3$ cubic fields $K/k$ with quadratic resolvent $F/k$ and $\disc(K/k) = \mathfrak{q}^2 \disc(F/k)$ is bounded by $O(4^{\omega(\mathfrak{q})} h_3(F/k))$, where $\omega(\mathfrak{q})$ denotes the number of prime ideals dividing $\mathfrak{q}$.
Thus, for fixed $\mathfrak{q}$, the number of $S_3$-extensions $K/k$ with $\disc(K/k)\leq X$ whose discriminant is $\mathfrak{q}^2$ times that of their quadratic resolvent, may be bounded by a uniform constant times
\[
\sum_{\substack{[F:k] = 2 \\ \Disc(F/k) \leq X/\Nm(\mathfrak{q})^2}} 4^{\omega(\mathfrak{q})} h_3(F/k)
\ll_{[k:\mathbb{Q}],\epsilon} 4^{\omega(\mathfrak{q})} \mathrm{Disc}(k)^{1/2+\epsilon} \kappa_k^{-2}
\left( \frac{\mathrm{Disc}(k) h_2(k)^{1/3}X^{1/2}}{\Nm(\mathfrak{q})} + \frac{X}{\Nm(\mathfrak{q})^{2}} \right),
\]
the latter inequality provided by Proposition \ref{prop:order-bound-sf}.
When $X \geq \Disc(k)^2 h_2(k)^{2/3}$, the expression is parentheses above is bounded by twice the second term, and
adding the above inequality across $\mathfrak{q}$ with norm at most $X^{1/2}$, we have
\begin{equation*}
N_k(S_3, X) \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \mathrm{Disc}(k)^{1/2+\epsilon} \kappa_k^{-2} X.
\end{equation*}
Otherwise, $X \leq \Disc(k)^3$, and adding over $\mathfrak{q}$ have have
\begin{equation*}
N_k(S_3, X) \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \mathrm{Disc}(k)^{3/2+\epsilon} \kappa_k^{-2} h_2(k)^{1/3} X^{1/2+\epsilon} \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \mathrm{Disc}(k)^{3/2+4\epsilon} \kappa_k^{-2} h_2(k)^{1/3} X^{1/2},
\end{equation*}
and the proposition follows.
\end{proof}
\subsection{Proof of Theorem \ref{thm:cubic-bound}}\label{sec:effective}
We can now combine our results in the various ranges to prove Theorem \ref{thm:cubic-bound}.
By the bijection between $3$-torsion elements in relative class groups of quadratic extensions and $S_3$ relative cubic extensions with square-free discriminant ideal \cite[Section 5]{DW88}, we have
\begin{equation}\label{eq:h3-bijection}
\sum_{\substack{[F:k]=2\\\Disc(F/k)\leq X}} h_3(F_2/k) \leq 2N_k(S_3,X) + N_{k}(C_2,X).
\end{equation}
Using this, the first case follows directly from Proposition \ref{prop:trivial-cubic-bound}.
The second and third cases follow from Propositions \ref{prop:order-bound-sf} and \ref{prop:cubic-bound-order-general}
and the fact that $\mathrm{Res}_{s=1} \zeta_k(s) \gg_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \mathrm{Disc}(k)^{-\epsilon}$
(though, this is only known to hold with an effectively computable implied constant when $\zeta_k(s)$ does not have an exceptional Landau--Siegel zero; see \cite{Stark}).
The fifth case follows from \eqref{eq:h3-bijection} and Proposition \ref{prop:effective-shintani}, and the fourth from the observation that $N_{k}(S_3, X) \leq N_k(S_3, D_k^3h^{2/3})$ for any $X \leq D_k^3h^{2/3}$ (and analogously for class numbers).
\section{Zero-density estimates for ray class $L$-functions}\label{sec:zero-density}
In this section, we collect the results we need on zero-density estimates for ray class $L$-functions and use them to show that most quadratic extensions will have sufficiently many small split primes. This will allow us to apply our relative Ellenberg--Venkatesh Lemma~\ref{lem:relative-E-V} and get a non-trivial bound on $h_3(F/k)$ for most quadratic $F/k$.
Let $\chi$ be a nontrivial ray class character of a number field $k$ with conductor $\mathfrak{f}_\chi$. The associated $L$-function $L(s,\chi)$ is defined by
\[
L(s,\chi) = \sum_{\mathfrak{a}} \frac{\chi(\mathfrak{a})}{\Nm(\mathfrak{a})^s},
\]
the series running over the integral ideals of $k$. Let $r_1^+(\chi)$ denote the number of real places of $k$ that are split in the cyclic extension of $k$ cut out by $\chi$, let $r_1^-(\chi)$ denote the number of real places that are ramified in the extension, and let $r_2$ denote the number of complex places of $k$. If we define
\[
\gamma_\chi(s) := \pi^{-[k:\mathbb{Q}]s/2} \Gamma\left(\frac{s}{2}\right)^{r_1^+(\chi)+r_2} \Gamma\left(\frac{s+1}{2}\right)^{r_1^-(\chi)+r_2},
\]
then the completed $L$-function $\Lambda(s,\chi) := (\Nm(\mathfrak{f}_\chi) \mathrm{Disc}(k))^{s/2} \gamma_\chi(s) L(s,\chi)$
is entire and satisfies a functional equation, $\Lambda(1-s,\chi) = \varepsilon(\chi) \Lambda(s,\bar\chi)$. The poles of $\gamma_\chi(s)$ thus lend $L(s,\chi)$ its trivial zeros; these are the non-positive even integers with multiplicity $r_1^+(\chi)+r_2$, and the negative odd integers with multiplicity $r_1^-(\chi)+r_2$.
The following lemma summarizes the analytic properties we shall need.
\begin{lemma}\label{lem:analytic-ray}
Assume notation as above and set $q_\chi = \Nm(\mathfrak{f}_\chi) \mathrm{Disc}(k)$. Then:
\begin{enumerate}
\item For any $t\geq 0$, the number of non-trivial zeros $\rho=\beta+i\gamma$ of $L(s,\chi)$ with $|\gamma-t| \leq 1$ is $O_{[k:\mathbb{Q}]}(\log((t+1) q_\chi))$.
\item For any $s$ with $-3/2 \leq \Re(s) \leq 2$, we have
\[
-\frac{L^\prime}{L}(s,\chi)
= -\frac{r_1^+(\chi)+r_2}{s} - \frac{r_1^-(\chi)+r_2}{s+1}-\sum_{\substack{ \rho: \\ |s-\rho|<1}} \frac{1}{s-\rho} + O_{[k:\mathbb{Q}]}(\log((|s|+1) q_\chi)),
\]
where the sum runs over the nontrivial zeros $\rho$ of $L(s,\chi)$.
\end{enumerate}
\end{lemma}
\begin{proof}
These are standard properties of $L$-functions. For example, see Proposition 5.7 in Iwaniec and Kowalski \cite{IwaniecKowalski}. The second conclusion is stated there only for $-1/2 \leq \Re(s) \leq 2$, but in the region $-3/2 \leq \Re(s) \leq -1/2$, the functional equation relates $-\frac{L^\prime}{L}(s,\chi)$ to $-\frac{L^\prime}{L}(1-s,\bar\chi)$, which is absolutely convergent. Thus, the conclusion is easier in this case, with the factors arising from the functional equation being dominant.
\end{proof}
We next recall a consequence of Proposition A.2 in the appendix by Lemke Oliver and Thorner to work of Pasten \cite{Pasten}, which follows upon taking $\pi$ to be the trivial representation of $k$. See also \cite[Theorem 1.2]{ThornerZaman-LargeSieve} for a somewhat stronger statement, which would allow also $\epsilon = 0$ below with an explicit value of $c$.
\begin{theorem}\label{thm:zero-density}
Let $k$ be a number field, and for any ray class character $\chi$ of $k$, let
\[
N_\chi(\sigma,T) := \#\{ \rho : L(\rho,\chi) = 0, \Re(\rho) \in (\sigma,1), |\Im(\rho)| \leq T\}.
\]
Then there is a constant $c>0$, depending only on the degree $[k:\mathbb{Q}]$, such that for any $Q,T>1$, any $1/2 \leq \sigma <1$, and any $\epsilon > 0$, we have
\[
\sum_{\Nm(\mathfrak{q}) \leq Q} \,\sideset{}{^\prime}\sum_{\chi \pmod{\mathfrak{q}}} N_\chi(\sigma,T)
\ll_{[k:\mathbb{Q}],\epsilon} (\mathrm{Disc}(k) QT)^{c(1-\sigma)+\epsilon}.
\]
Here, the outer summation runs over ideals $\mathfrak{q}$ of $\mathcal{O}_k$ with bounded norm and the inner summation runs over primitive ray class characters of conductor $\mathfrak{q}$.
\end{theorem}
Theorem \ref{thm:zero-density} implies that over a general number field $k$, most ray class $L$-functions do not have zeros near $s=1$. Therefore we will utilize Theorem \ref{thm:zero-density} to show that for most quadratic extensions over $k$ we can find enough prime ideals with a specified splitting behavior. Recall that $\pi_k(Y)$ denotes the prime ideal counting function of $k$, and for an extension $F/k$, $\pi_k(Y;F,e)$ denotes the number of primes in $k$ of norm at most $Y$ that are split in $F$.
\begin{lemma}\label{lem:effective-Chebo-over-k-average}
Let $k$ be a number field and let $\mathcal{F}_2(X)$ be the set of quadratic extensions of $k$ with $\Disc(F/k)\le X$. Let $c = c_{[k:\ensuremath{{\mathbb{Q}}}]}$ be the constant as given in Theorem \ref{thm:zero-density}. For any $\epsilon_1>0$
and $X\geq 2$,
there exists a set $\mathcal{E} = \mathcal{E}(k, X,
\epsilon_1)\subset \mathcal{F}_2(X)$ of exceptional quadratic extensions $F/k$ such that:
\begin{enumerate}
\item for all $F \in \mathcal{F}_2(X)\setminus \mathcal{E}$, for all $4 \leq Y \le X$
there exists a constant $C_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1}$ depending only on $[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1$ such that
$$
\pi_k(Y; F, e) \geq \frac{1}{8} \pi_k(Y/2) - C_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1}Y^{\sigma_1} \log^2( X \Disc(k)),
$$
where $\sigma_1 = \max(1- \frac{\epsilon_1}{4c},1/2)$; and
\item $|\mathcal{E}| \ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon_1} \Disc(k)^{\epsilon_1}X^{\epsilon_1}$.
\end{enumerate}
\end{lemma}
\begin{proof}
For each quadratic extension $F/k$, there is a unique quadratic ray class character $\chi_{F/k}$ with conductor $\disc(F/k)$ such that $\frac{\zeta_F(s)}{\zeta_k(s)} = L(s, \chi_{F/k})$.
As in the statement of the lemma, we set $\sigma_1 = 1-\epsilon_1/4c$
and define
\[
\mathcal{E}: = \{ F \in \mathcal{F}_2(X) : \exists \rho \text{ s.t. } L(\rho, \chi_{F/k}) =0, \Re(\rho) \in (\sigma_1, 1), \text{ and } |\Im(\rho)| \leq X^{1/2}
\}.
\]
Then Theorem \ref{thm:zero-density} implies $|\mathcal{E}| \ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon_1/4} (\Disc(k) X^{3/2}
)^{c(1-\sigma_1)+\epsilon_1/4}\ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1} \Disc(k)^{\epsilon_1}X^{\epsilon_1}$.
Suppose $F \notin \mathcal{E}$. In order to bound from below the number of primes of $k$ with $\chi_{F/k}(\mathfrak{p}) = 1$, we first observe that, away from the $O(\log \Disc(F))$ ramified primes, the function $(1+\chi_{F/k}(\mathfrak{p}))/2$ is precisely $1$ or $0$ according to whether $\mathfrak{p}$ is split or inert. As is standard, we first consider a version weighted by the von Mangoldt function $\Lambda_k$ of $k$, and we observe
\begin{align*}
\sum_{\Nm(\mathfrak{a}) \leq Y} \Lambda_k(\mathfrak{a})\, \frac{1+\chi_{F/k}(\mathfrak{a})}{2}
&\geq \sum_{\Nm(\mathfrak{a}) \leq Y} \Lambda_k(\mathfrak{a})\, \frac{1+\chi_{F/k}(\mathfrak{a})}{2} \left(1 - \frac{\Nm(\mathfrak{a})}{Y}\right) \\
&= \frac{1}{2} \sum_{\Nm(\mathfrak{a}) \leq Y} \Lambda_k(\mathfrak{a})\left(1 - \frac{\Nm(\mathfrak{a})}{Y}\right) + \frac{1}{2} \sum_{\Nm(\mathfrak{a}) \leq Y} \chi_{F/k}(\mathfrak{a})\Lambda_k(\mathfrak{a})\left(1 - \frac{\Nm(\mathfrak{a})}{Y}\right).
\end{align*}
The first term is related to $\pi_k(Y)$ by means of partial summation, so our goal is to show that the second is small.
Via Perron's formula, we find
\begin{align*}
\sum_{|\mathfrak{a}| \leq Y} \chi_{F/k}(\mathfrak{a})\Lambda_k(\mathfrak{a}) \left( 1-\frac{\Nm(\mathfrak{a})}{Y} \right)
&= \frac{-1}{2\pi i} \int_{2-i\infty}^{2+i\infty} \frac{L^\prime}{L}(s,\chi_{F/k}) \cdot \frac{Y^s}{s(s+1)}\,ds.
\end{align*}
It follows from Lemma \ref{lem:analytic-ray} and the functional equation that the integral converges absolutely over any vertical line. Thus, we may shift the contour all the way to the left, finding
\[
\sum_{\Nm(\mathfrak{a}) \leq Y} \chi_{F/k}(\mathfrak{a}) \Lambda_k(\mathfrak{a})\left( 1 - \frac{\Nm(\mathfrak{a})}{Y}\right)
=- \sum_{\substack{ \rho: \\ L(\rho,\chi)=0 \\ \rho\neq 0,-1}} \frac{Y^\rho}{\rho(\rho+1)} - \mathrm{Res}_0 - \mathrm{Res}_{-1},
\]
where the summation runs over all zeros of $L(s,\chi)$, nontrivial or trivial, and where $\mathrm{Res}_0$ and $\mathrm{Res}_{-1}$ denote the residues of the integrand at $s=0$ and $s=-1$, respectively.
The trivial zeros of $L(s,\chi_{F/k})$ occur at negative integers, and each has multiplicity at most $[k:\mathbb{Q}]$, so their total contribution to the sum may be bounded unifomly by $O_{[k:\mathbb{Q}]}(Y^{-2})$. Using Lemma \ref{lem:analytic-ray}, the residue $\mathrm{Res}_{-1}$ is bounded by $O_{[k:\mathbb{Q}]}((\log Y)(\log X\mathrm{Disc}(k))Y^{-1})$, which is sufficient. Again using Lemma \ref{lem:analytic-ray}, we see that
\begin{align*}
\mathrm{Res}_0
&= -(r_1^+(\chi)+r_2) \log Y + \sum_{\substack{\rho: \\ |\rho| \leq 1}} \frac{1}{\rho} + O_{[k:\mathbb{Q}]}(\log (X \mathrm{Disc}(k))) \\
&\ll_{[k:\mathbb{Q}]} \frac{\log (X \mathrm{Disc}(k))}{1-\sigma_1},
\end{align*}
since $F \not\in \mathcal{E}$ and $Y \leq X
$.
By Lemma \ref{lem:analytic-ray}, the summation over zeros of height greater than $T = X^{1/2}
$ is bounded
\[
\sum_{\substack{ \rho: \\ L(\rho, \chi)=0 \\ |\Im(\rho)| \geq T}} \frac{Y^\rho}{\rho(\rho+1)}
\ll_{[k:\mathbb{Q}]} \frac{Y \log(X\Disc(k)T)}{T} \ll_{[k:\ensuremath{{\mathbb{Q}}}]} Y^{1/2} \log(X\Disc(k)).
\]
Notice that by definition of $\mathcal{E}$, for $F \notin \mathcal{E}$ there are no zeros for $L(s, \chi_{F/k})$ with $\Re(\rho)>\sigma_1$ and $|\Im(\rho)| \leq T = X^{1/2}
$. Using the functional equation, there are also no zeros with $\Re(\rho)\in (0,1-\sigma_1)$ and $|\Im(\rho)| \leq T$. The summation over low lying zeroes of $L(s, \chi_{F/k})$ may therefore be bounded by
\[
\sum_{\substack{ \rho: \\ L(\rho,\chi)=0 \\ |\Im(\rho)| < T}} \frac{Y^\rho}{\rho(\rho+1)}
\ll_{[k:\mathbb{Q}]} \frac{Y^{\sigma_1} \log (X\Disc(k)T)}{1-\sigma_1} \ll_{[k:\ensuremath{{\mathbb{Q}}}],
\epsilon_1} Y^{\sigma_1}\log (X\Disc(k)).
\]
Finally, by the standard translation from the von Mangoldt function to the prime counting function, we can conclude the lemma.
\end{proof}
In order to apply Lemma \ref{lem:effective-Chebo-over-k-average} efficiently, we need the following effective lower bound on the number of prime ideals in a general number field $k$.
\begin{lemma}[\cite{ZamThesis}]\label{lem:prime-number-theorem-k-dependence}
There exist absolute, effective constants $\gamma$, $\beta$, and $D_0 >0$ such that if $k$ is a number field with $\Disc(k)\ge D_0$, then for $Y\ge \Disc(k)^{\beta}$, we have
$$\pi_k(Y) \gg_{[k:\ensuremath{{\mathbb{Q}}}]} \frac{1}{\Disc(k)^{\gamma}}\cdot \frac{Y}{\log Y}.$$
\end{lemma}
An effective lower bound, weaker than that above, can be derived from \cite{LMO}, but Zaman \cite{ZamThesis} was the first to prove a bound of this strength; he did so with $\beta = 35$ and $\gamma = 19$. It follows from \cite{ThornerZaman} that $\beta = 694$ and $\gamma = 5$ are also admissible, and from \cite{ZTChebo} that $\gamma = 1/[k:\ensuremath{{\mathbb{Q}}}]$ is admissible with an inexplicit value of $\beta$.
\section{A Uniform Tail Estimate}\label{sec:tail}
A \emph{$2$-extension} is a $G$-extension for some $2$-group $G$.
In this section, our main goal is to prove the following theorem on a tail estimate for the $3$-torsion in class groups of $2$-extensions with a uniform dependence on the base field $k$.
It will be the critical input for the proof of our main theorems in Section \ref{sec:main-theorem}.
\begin{theorem}\label{thm:tail}
Let $k$ be a number field. For each $m\ge 2$, there exist constants $\delta_m>0$ and $\alpha_m$ depending at most on $m$ and the degree $[k:\ensuremath{{\mathbb{Q}}}]$ such that for any
$X,Y>0$,
we have
\begin{equation}
\sum_{\substack{F_m/F_{m-1}/\cdots/ F_1/k \\ \Disc(F_{m-1})\ge Y \\ X/2\le \Disc(F_m)\le X}} h_3(F_m/k) = O_{\epsilon, [k:\ensuremath{{\mathbb{Q}}}],m}\left(\Big(\frac{X}{Y^{1-\epsilon}} + X^{1-\delta_m} \Big) \cdot \Disc(k)^{\alpha_m} \right),
\end{equation}
where the summation is over all towers $F_m/F_{m-1}/ \cdots F_1/F_0=k$
in $\bar{\ensuremath{{\mathbb{Q}}}}$
with $F_{i+1}/F_{i}$ being quadratic extensions for each $ 0\le i\le m-1$.
\end{theorem}
We emphasize the point that Theorem \ref{thm:tail} is stated with respect to the absolute discriminant, while the results of Section \ref{sec:uniform-cubic} are with respect to the relative discriminant.
\begin{remark}
The statement of Theorem \ref{thm:tail} for $m=1$ is empty except when $Y \le \Disc(k)$.
If $Y \leq \Disc(k)$, the statement for $m=1$ is non-trivial, but
follows from Corollary \ref{cor:general-bound}, and noting that it is stated with respect to \emph{relative} discriminant, we may take $\alpha_1$ to be any constant greater than $1/3$.
\end{remark}
\begin{remark}
For our purposes, the actual values of $\delta_m$ and $\alpha_m$ are irrelevant; it suffices to know only that the constants $\delta_m$ and $\alpha_m$ exist. However, in the course of the proof, we provide explicit but non-optimal values. This will show that the value of $\delta_2$ may be chosen without dependence on $[k:\mathbb{Q}]$.
We choose specific non-optimal values of many related exponents in the course of the proof, because we wish to avoid tracking further dependencies of implied constants that would come with avoiding such choices.
\end{remark}
Note that the tail estimate we prove in Theorem \ref{thm:tail} handles all $2$-extensions due to the following easy lemma.
\begin{lemma}[\cite{KluWan}, Lemma 2.3]\label{lem:tower}
Let $n=\ell^r$ be a prime power and $G\subset S_n$ be an $\ell$-group and $E/F$ be an extension of number fields with $\Gal(E/F)\cong G$. Then there exists a tower of fields
\begin{equation}
F=F_0 \subset F_1\subset \ldots \subset F_{r-1} \subset F_r=E
\end{equation}
such that $\Gal(F_{i+1}/F_i)=C_\ell$ for all $0\leq i \leq r-1$.
\end{lemma}
We will prove Theorem~\ref{thm:tail} first for $m = 2$. Then treating $m=2$ as the base case, we will apply an inductive proof to prove Theorem~\ref{thm:tail} for general $m$. For $m=2$, the theorem requires us to sum over all towers $F_2/F_1/F_0=k$ of relative quadratic extensions. We will separate the discussion depending on how large $\Disc(F_1)$ is. In Section \ref{ssec:m-2-non-critical}, we will consider the summation when $\Disc(F_1)$ is away from $X^{1/3}$ using results established in Section \ref{sec:uniform-cubic}. In Section \ref{ssec:initial-critical}, we will consider the summation when $\Disc(F_1)$ is near $X^{1/3}$ using results established in Section \ref{sec:arakelov} and \ref{sec:zero-density}.
Taking $Y=1$ in Theorem \ref{thm:tail}, we get the following immediate corollary, essentially an analogue of Corollary \ref{cor:general-bound}, which will be useful in the induction argument to come.
\begin{corollary} \label{cor:general-bound-m}
Let $k$ be a number field. For each $m \geq 2$, let $\alpha_m$ be the constant from Lemma \ref{thm:tail}. Then for any $X \geq 1$, we have
\[
\sum_{\substack{F_m/F_{m-1}/\cdots/ F_1/k \\ X/2\le \Disc(F_m)\le X}} h_3(F_m/k) = O_{ [k:\ensuremath{{\mathbb{Q}}}],m}\Big ( X \Disc(k)^{\alpha_m} \Big ).
\]
\end{corollary}
\subsection{Base Case: Non-critical Range}\label{ssec:m-2-non-critical}
In this subsection, we will consider the following summation:
\begin{equation} \label{eqn:non-critical-def}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{N}(X,Y) \\ X/2\le \Disc(F_2)\le X}} h_3(F_2/k),
\end{equation}
where
$\mathfrak{N}(X,Y)
:= [Y, X^{1/3-\delta_0}) \cup (X^{1/3+\delta_0}, X^{1/2}]$
for an arbitrary small number $\delta_0$ satisfying $0<\delta_0<1/6$. We regard this range for $F_1$ as non-critical because
the results of Section~\ref{sec:uniform-cubic}
together with the following lemma on the $2$-class number of
relative $2$-extensions will give the desired bound (Lemma~\ref{lem:m-2-non-critical}) in this range.
\begin{lemma}[\cite{KluWan}, Theorem $2.7$]\label{lem:two-part}
Let $k$ be a number field and let $F/k$ be a $2$-extension. Then for any $\epsilon>0$, we have $h_2(F) \ll_{[F:\mathbb{Q}],\epsilon} \mathrm{Disc}(F)^\epsilon h_2(k)^{[F:k]}$.
\end{lemma}
\begin{lemma}\label{lem:m-2-non-critical}
Let $k$ be a number field and let $\delta_0 \in (0, 1/6)$. There exists $\delta>0$, only depending on $\delta_0$, and an absolute constant $\alpha>0$ such that for all
$X,Y>0$,
we have
\begin{equation}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{N}(X,Y) \\ X/2\le \Disc(F_2)\le X}} h_3(F_2/k) = O_{\epsilon, [k:\ensuremath{{\mathbb{Q}}}],\delta_0}\Big (\frac{X}{Y^{1-\epsilon}} + X^{1-\delta} \Big) \cdot \Disc(k)^{\alpha},
\end{equation}
where the summation is over all towers $F_2/F_1/F_0=k$ with $F_2/F_1$ and $F_1/F_0$ being relative quadratic extensions and where $\mathfrak{N}(X,Y) = [Y, X^{1/3-\delta_0}) \cup (X^{1/3+\delta_0}, X^{1/2}]$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:m-2-non-critical}]
Writing $D := \Disc(F_1)$ and $h:=h_2(F_1)$, from Section~\ref{sec:uniform-cubic} (and the translation from relative to absolute discriminant), we have
\begin{equation}\label{eqn:3-torsion-summation}
\sum_{\substack{F_2/ F_1\\ X/2\le \Disc(F_2)\le X}} h_3(F_2/F_1) \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} h \cdot D^{\epsilon}
\begin{cases}
X^{3/2} D^{-5/2}, & \text{for all $D^2\leq X \leq D^3$}, \\
X^{1/2} D^{1/2} +X D^{-3/2}, & \text{for all $D^2\leq X$}, \\
X D^{-2}, & \text{if } X \geq D^5 h^{2/3}.
\end{cases}
\end{equation}
Specifically, the first line follows from Proposition \ref{prop:trivial-cubic-bound} and \eqref{eq:h3-bijection},
the second from Proposition \ref{prop:order-bound-sf},
and the third from the final case of Theorem~\ref{thm:cubic-bound}.
We will use the first line of \eqref{eqn:3-torsion-summation} for $\Disc(F_1) \in (X^{1/3+\delta_0}, X^{1/2}]$, the second line for $\Disc(F_1) \in [Y_0, X^{1/3-\delta_0})$,
and the third line for $\Disc(F_1) \in [Y, Y_0)$, where we choose a value for $Y_0$ so that we can apply the third line to the range $[Y, Y_0)$ for every $F_1/k$.
In particular, if $C_{[K:\ensuremath{{\mathbb{Q}}}]}$ is such that we always have $h(K)^{2/3}\leq C_{[K:\ensuremath{{\mathbb{Q}}}]} \Disc(K)$, we can choose $Y_0=C_{2[k:\ensuremath{{\mathbb{Q}}}]}^{-1/6}X^{1/6}$.
Applying \eqref{eqn:3-torsion-summation} as described and applying Lemma \ref{lem:two-part} to bound $h$, we find
\begin{equation}\label{eqn:non-critical-summation-m-2}
\begin{aligned}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{N}(X,Y)\\ X/2\le \Disc(F_2)\le X}} h_3(F_2/k)
& \ll_{\epsilon',\epsilon, [k:\ensuremath{{\mathbb{Q}}}]} h_2(k)^{2} \cdot X^{3/2} \sum_{\substack{F_1/k \\ \Disc(F_1) \in (X^{1/3+\delta_0}, X^{1/2})}} h_3(F_1/k) \Disc(F_1)^{-5/2+\epsilon'}\\
& + h_2(k)^{2} \cdot X^{1/2} \sum_{\substack{F_1/k \\ \Disc(F_1) \in [Y_0,X^{1/3-\delta_0})}} h_3(F_1/k) \Disc(F_1)^{1/2+\epsilon'}\\
& + h_2(k)^{2} \cdot X \sum_{\substack{F_1/k \\ \Disc(F_1) \in [Y_0,X^{1/3-\delta_0})}} h_3(F_1/k) \Disc(F_1)^{-3/2+\epsilon'}\\
& + h_2(k)^{2} \cdot X \sum_{\substack{F_1/k \\ \Disc(F_1) \in (Y,Y_0)}} h_3(F_1/k) \Disc(F_1)^{-2+\epsilon}\\
& \ll_{\epsilon',\epsilon, [k:\ensuremath{{\mathbb{Q}}}]} \Disc(k)^{1/3+\epsilon} \cdot (X^{1-3\delta_0/2+\epsilon'} + X^{11/12+\epsilon'} + \frac{X}{Y^{1-\epsilon}}),
\end{aligned}
\end{equation}
where for the last inequality we have applied partial summation and Corollary~\ref{cor:general-bound}, as well as the trivial bound for $h_2(k)$.
Then we can choose, for example, $\delta=\min\{ \delta_0,1/15\}$, and $\alpha=2/5$.
Above if we take $\epsilon'=\min\{\delta_0/2,1/60\}$, then the lemma follows for $0<\epsilon\leq \alpha- 1/3$, and hence for all $\epsilon> 0.$
\end{proof}
\subsection{Base Case: Critical Range}\label{ssec:initial-critical}
In this section, we will consider the following summation:
\begin{equation}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X,Y) \\ X/2\le \Disc(F_2)\le X}} h_3(F_2/k),
\end{equation}
where $\mathfrak{C}(X,Y) = \mathfrak{C}(X)= [X^{1/3-\delta_0}, X^{1/3+\delta_0}]$, which we regard as the critical range of $\Disc(F_1)$.
In this range, summing the best bound for each $F_1$ is not sufficient, and we need to extract an additional saving from the sum over $F_1$.
We will apply the results developed in Sections \ref{sec:arakelov} and \ref{sec:zero-density} to obtain this saving.
It will be convenient to use the following simple bound on the number of relative quadratic extensions.
\begin{lemma}\label{lem:quadratic-bound}
Let $k$ be a number field. Then
$$N_k(C_2, X) = O_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon}(h_2(k)\Disc(k)^{\epsilon} X).$$
\end{lemma}
\begin{proof}
By class field theory, the number
$N_k(C_2, X)$ is bounded by the product of $h_2(k)$ and the number of integral ideals with bounded norm.
The latter can be counted by integrating $\zeta_k(s)$.
If $a_n$ denotes the number of integral ideals of norm $n$, then by Perron's formula we have
\begin{equation}
\sum_{n \leq X} a_n \leq \sum_{n=1}^\infty a_n e^{1-\frac{n}{X}} = \frac{e}{2\pi i} \int_{1+\epsilon-i\infty}^{1+\epsilon+i\infty} \zeta_k(s)\cdot \Gamma(s)\cdot X^s\,ds.
\end{equation}
Shifting the contour integral to $\Re(s)=1-\epsilon$ and using the convexity bound \eqref{eqn:convexity-dedekind} on $\zeta_k(s)$,
we get the upper bound
\begin{equation}
\sum_{n \leq X} a_n = \Res(\zeta_k(s) \Gamma(s) X^s)_{s = 1} + O_{[k:\mathbb{Q}],\epsilon}(\Disc(k)^{\epsilon/2} X^{1-\epsilon}) = O_{[k:\mathbb{Q}],\epsilon}(\Disc(k)^{\epsilon} X),
\end{equation}
where the upper bound on the residue of the Dedekind zeta function comes from Landau.
\end{proof}
\begin{lemma}\label{lem:m-2-critical}
Let $k$ be a number field and let $\delta_0 \in (0, 1/100]$. There exists an absolute positive constant $\delta$, and a value $\alpha$, depending only on $[k:\ensuremath{{\mathbb{Q}}}]$, such that for all $X\ge 1$ we have
\begin{equation}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X) \\ X/2\le \Disc(F_2)\le X}} h_3(F_2/k) = O_{[k:\ensuremath{{\mathbb{Q}}}] }\Big (X^{1-\delta} \Big) \cdot \Disc(k)^{\alpha},
\end{equation}
where the summation is over all towers $F_2/F_1/F_0=k$ with $F_2/F_1$ and $F_1/F_0$ being relative quadratic extensions and $\mathfrak{C}(X) = [X^{1/3-\delta_0}, X^{1/3+\delta_0}]$.
\end{lemma}
\begin{proof}
We first give the idea of the proof in broad terms to motivate the notation that is to follow.
Exploiting the factorization $h_3(F_2/k) = h_3(F_2/F_1) h_3(F_1/k)$, Corollary \ref{cor:general-bound} provides strong control on the average of $h_3(F_1/k)$, while Lemma \ref{lem:relative-E-V} provides a non-trivial bound on $h_3(F_2/F_1)$ if there are ``enough'' small primes that split in $F_2/F_1$. Lemma \ref{lem:prime-number-theorem-k-dependence} ensures there are enough small primes in $k$. Applying Lemma \ref{lem:effective-Chebo-over-k-average} twice, we see first that ``most'' extensions $F_1/k$ have roughly the same number of small primes as $k$, and then second that in ``most'' extensions $F_2/F_1$ for such $F_1$, many of those small primes are split. This lets us apply Lemma \ref{lem:relative-E-V}, and yields the lemma.
To make this precise, and in preparation for applying Lemma \ref{lem:effective-Chebo-over-k-average}, we let $\delta_1=1/20$ and $\epsilon_1=1/30$.
We set $\sigma_1 = \max(1 - \epsilon_1/4c,1/2)$, where $c = c_{[k:\ensuremath{{\mathbb{Q}}}]}$ is the constant given in Theorem \ref{thm:zero-density}.
Throughout the proof, we will make assumptions on $X$, including that $X$ is at least any absolute constant necessary to apply the results we use, and then we will handle the remaining values of $X$ at the end of the proof.
Applying Lemma \ref{lem:effective-Chebo-over-k-average} the first time, we obtain an exceptional set $\mathcal{E}_0 =\mathcal{E}(k, X^{1/3+\delta_0}, \epsilon_1)$.
Then for each $F_1/k\notin \mathcal{E}_0$, we apply Lemma \ref{lem:effective-Chebo-over-k-average} a second time to obtain an exceptional set $\mathcal{E}(F_1) = \mathcal{E}(F_1, X/\Disc(F_1)^2, \epsilon_1)$. By Lemma \ref{lem:effective-Chebo-over-k-average}, both exceptional sets contain few elements:
\begin{equation} \label{eqn:exceptional-sparse}
|\mathcal{E}_0|\ll_{[k:\ensuremath{{\mathbb{Q}}}]} \Disc(k)^{\epsilon_1} X^{\epsilon_1} \quad \text{and} \quad |\mathcal{E}(F_1)|\ll_{[k:\ensuremath{{\mathbb{Q}}}]} \Disc(k)^{\epsilon_1} X^{\epsilon_1}.
\end{equation}
We now consider the consequences of Lemma \ref{lem:effective-Chebo-over-k-average} for fields outside the exceptional sets. For $F_1/k \notin \mathcal{E}_0$ with $\Disc(F_1) \in \mathfrak{C}(X)$ and $x =(X/(2\Disc(F_1)^2))^{\delta_1} \leq X^{1/3+\delta_0}$, by Lemma \ref{lem:effective-Chebo-over-k-average} we have, for $X$ sufficiently large,
\begin{equation}\label{eqn:getprimes}
\pi_{F_1}(x/2) \ge \pi_k(x/2;F_1, e) \geq \frac{1}{8}\pi_k(x/4)-C_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1}(x/2)^{\sigma_1} \log^2( X^{1/3+\delta_0}\Disc(k)).
\end{equation}
We next assume that $X\ge 4^{\frac{1}{(1/3-2/100)\delta_1}} \Disc(k)^A$
where $A =\frac{ \max \{ \beta, 8c(\gamma+2)/\epsilon_1 \} }{(1/3-2/100) \delta_1} = A([k:\ensuremath{{\mathbb{Q}}}])$, and $\gamma$ and $\beta$
are absolute constants for which Lemma \ref{lem:prime-number-theorem-k-dependence} holds; let also $D_0$ be an absolute constant such that Lemma \ref{lem:prime-number-theorem-k-dependence} holds for those $\gamma$ and $\beta$.
If $\Disc(k) \geq D_0$, we use Lemma \ref{lem:prime-number-theorem-k-dependence},
and for $X\ge \Disc(k)^A$ and $X$ sufficiently large given $[k:\ensuremath{{\mathbb{Q}}}]$, we obtain
\begin{equation} \label{eqn:X0-requirement}
\frac{1}{8}\pi_k(x/4)-C_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1}(x/2)^{\sigma_1} \log^2( X^{1/3+\delta_0}\Disc(k)) \gg_{[k:\ensuremath{{\mathbb{Q}}}]} \frac{x}{\Disc(k)^{\gamma} \log x}.
\end{equation}
We require that $X$ is large enough
so that \eqref{eqn:X0-requirement} holds also for the finitely many fields $k$ with $\Disc(k) \leq D_0$. Under these assumptions, we conclude for $F_1 \notin \mathcal{E}_0$ that
\begin{equation}\label{eqn:halfway}
\pi_{F_1}(x/2) \gg_{[k:\ensuremath{{\mathbb{Q}}}]} \frac{x}{\Disc(k)^{\gamma} \log x}.
\end{equation}
Next, for $F_2/F_1\notin \mathcal{E}(F_1)$, by Lemma \ref{lem:effective-Chebo-over-k-average} and \eqref{eqn:halfway}, and the assumptions above on $X$, we have
\begin{equation}\label{eqn:non-exceptional-primes}
\pi_{F_1}(x; F_2, e) \gg_{[k:\ensuremath{{\mathbb{Q}}}]} \frac{x}{ \Disc(k)^{\gamma} \log x}.
\end{equation}
We now consider the contribution to the summation from the various kinds of exceptional and non-exceptional fields. First, we consider the summation $F_2/F_1/k$ where $F_1/k\in \mathcal{E}_0$. Using the trivial bound on the class group
and the sparsity
of exceptional fields along with Lemmas~\ref{lem:quadratic-bound} and \ref{lem:two-part}, we find
\begin{equation}
\begin{aligned}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)\\ X/2\le \Disc(F_2)\le X, F_1/k\in \mathcal{E}_0}} h_3(F_2/k)
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} X^{1/2+\epsilon} \sum_{\substack{F_1/k \in \mathcal{E}_0 \\ \Disc(F_1) \in \mathfrak{C}(X)}} \sum_{\substack{F_2/ F_1 \\ X/2\le \Disc(F_2)\le X}} 1 \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} h_2(k)^2 \Disc(k)^{\epsilon_1} \cdot X^{5/6+2\delta_0+\epsilon_1+\epsilon} \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \Disc(k)^{1+\epsilon} \cdot X^{5/6+2\delta_0+\epsilon_1+\epsilon}.
\end{aligned}
\end{equation}
Second, we consider the contribution from those $F_2/F_1/k$ where $F_1 \notin \mathcal{E}_0$ but $F_2/F_1\in \mathcal{E}(F_1)$. Using the trivial bound on relative class group $h_3(F_2/F_1)$ and an average bound on $h_3(F_1/k)$,
we find
\begin{equation}
\begin{aligned}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)\\ X/2\le \Disc(F_2)\le X\\ F_2/F_1\in \mathcal{E}(F_1)}} h_3(F_2/k) & \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \sum_{\substack{F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)}} h_3(F_1/k) \sum_{\substack{F_2/ F_1 \\ X/2 \leq \Disc(F_2) \leq X \\ F_2/F_1\in \mathcal{E}(F_1)}} \Disc(F_2/F_1)^{1/2+\epsilon} \Disc(F_1)^{1/2+\epsilon} \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \Disc(k)^{\epsilon_1} X^{1/2+\epsilon_1+\epsilon} \sum_{\substack{F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)}} h_3(F_1/k) \Disc(F_1)^{-1/2+\epsilon}\\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \Disc(k)^{4/3+\epsilon_1+\epsilon} \cdot X^{2/3+\delta_0/2+\epsilon_1+\epsilon}. \\
\end{aligned}
\end{equation}
Precisely, the first inequality follows from Lemma \ref{lem:relative-trivial}. The second inequality follows from
the estimate $|\mathcal{E}(F_1)|\ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon_1} \Disc(k)^{\epsilon_1} X^{\epsilon_1}$
from \eqref{eqn:exceptional-sparse}.
The third inequality follows from Corollary \ref{cor:general-bound} and partial summation.
We now consider the contribution from those $F_2/F_1/k$ where neither $F_1/k \notin \mathcal{E}_0$ nor $F_2/F_1 \notin \mathcal{E}(F_1)$. Using the relative Ellenberg--Venkatesh method developed in Section \ref{sec:arakelov} and existence of small split primes for non-exceptional fields developed in Section \ref{sec:zero-density} and made explicit in \eqref{eqn:non-exceptional-primes}, we find
\begin{equation}
\begin{aligned}
&\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)\\ X/2\le \Disc(F_2)\le X \\ F_1 \notin \mathcal{E}_0, F_2/F_1\notin \mathcal{E}(F_1)}} h_3(F_2/k) \\
&\quad\quad \ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon} \Disc(k)^{\gamma} \cdot \sum_{\substack{F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)}} h_3(F_1/k) \left(\sum_{\substack{F_2/ F_1 \\ F_2\notin \mathcal{E}(F_1)\\ X/2\le \Disc(F_2)\le X }} \Disc(F_2/F_1)^{1/2-\delta_1+\epsilon} \Disc(F_1)^{1/2+\epsilon}\right) \\
&\quad\quad \ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon} \Disc(k)^{\gamma} \cdot h_2(k)^2 \cdot
\left(\sum_{\substack{F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X)}} h_3(F_1/k) \cdot (\frac{X}{\Disc(F_1)^2})^{3/2-\delta_1+\epsilon}\cdot \Disc(F_1)^{1/2+\epsilon}\right) \\
& \quad\quad \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \Disc(k)^{\gamma+7/3+\epsilon} X^{1 - \frac{\delta_1}{3}+\delta_0(\frac{3}{2}-2\delta_1)+\epsilon}.\\
\end{aligned}
\end{equation}
Precisely, the first inequality comes from Lemma \ref{lem:relative-E-V} and \eqref{eqn:non-exceptional-primes}.
The second inequality follows from Lemma \ref{lem:quadratic-bound} and Lemma~\ref{lem:two-part}.
We apply Corollary \ref{cor:general-bound} and partial summation for the last inequality.
Finally, in the complementary case that $X$ is not sufficiently large, we may assume that $X \ll_{[k:\mathbb{Q}]} \Disc(k)^A$, so that
by applying the trivial bound $h_3(F_2/k) \ll_{[k:\mathbb{Q}],\epsilon} X^{1/2+\epsilon}$, we find
\begin{equation}
\begin{aligned}
\sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X) \\ X/2\le \Disc(F_2)\le X}} h_3(F_2/k)
& \ll_{[k:\mathbb{Q}],\epsilon} X^{1/2+\epsilon} \cdot \sum_{\substack{F_2/ F_1/k \\ \Disc(F_1) \in \mathfrak{C}(X) \\ X/2\le \Disc(F_2)\le X}} 1 \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon} X^{4/3+\delta_0 + \epsilon} h_2(k)^3 \Disc(k)^{-2+\epsilon} \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \Disc(k)^{(4/3+\delta_0+\epsilon)A-1/2+\epsilon}.
\end{aligned}
\end{equation}
Here
the second inequality follows from Lemma \ref{lem:quadratic-bound} twice, and the last inequality follows from the bound on $X$.
So to deduce the statement of the lemma, we note we have four upper bounds of the form $\Disc(k)^{a_i} X^{b_i+\epsilon}$. We can take, e.g. $\epsilon=1/1000$, and note that each of the $b_i$ is less than $998/1000$. Then we can take $\delta=1/1000$, and $\alpha$ can be taken to be $\max_i{a_i}$, which depends on $[k:\ensuremath{{\mathbb{Q}}}]$.
\end{proof}
\subsection{Induction}
Finally, in this section, we will prove Theorem \ref{thm:tail} in general. We will use an induction argument with the initial case $m=2$.
\begin{proof}[Proof of Theorem \ref{thm:tail}]
The statement for $m = 2$ follows from the combination of Lemma \ref{lem:m-2-non-critical} and \ref{lem:m-2-critical}.
Now assume for some $i$ that we can prove the lemma for every $m\le i$. Our goal is to prove the lemma for $m = i+1$.
Firstly, we treat the summation when $\Disc(F_i) \in \mathfrak{N}(X,Y)=[Y, X^{1/3-\delta_0}) \cup (X^{1/3+\delta_0}, X^{1/2}]$, the non-critical range. We use an argument similar with that in Lemma \ref{lem:m-2-non-critical} where non-critical range is treated for $m=2$. For any $0< \delta_0 < 1/6$ and $Y \geq 1$, using the results on counting cubic fields in Section \ref{sec:uniform-cubic}, as collected in \eqref{eqn:3-torsion-summation}, and the induction hypothesis in the form of
Corollary \ref{cor:general-bound-m}, we find
\begin{equation}\label{eqn:induction-noncritical}
\begin{aligned}
\sum_{\substack{F_{i+1}/F_i/\cdots/F_1/k \\ \Disc(F_i) \in \mathfrak{N}(X,Y) \\ X/2\le \Disc(F_{i+1})\le X}} h_3(F_{i+1}/k)
&= \sum_{\substack{F_i/\cdots/F_1/k \\ \Disc(F_i) \in \mathfrak{N}(X,Y)}} h_3(F_i/k) \sum_{\substack{F_{i+1}/F_i \\ X/2\le \Disc(F_{i+1})\le X }} h_3(F_{i+1}/F_i) \\
&\ll_{[k:\ensuremath{{\mathbb{Q}}}],i,\epsilon',\epsilon} \Disc(k)^{\alpha_i-2^{i-1}+\epsilon} \Big(X^{1-\frac{3\delta_0}{2}+\epsilon'} + X^{11/12+\epsilon'}+\frac{X}{Y^{1-\epsilon}}\Big) .\\
\end{aligned}
\end{equation}
The inequality above follows as in \eqref{eqn:non-critical-summation-m-2}, with $F_i$ in place of $F_1$,
using the cases of \eqref{eqn:3-torsion-summation} in three ranges of $\Disc(F_i)$ followed by partial summation with Corollary \ref{cor:general-bound-m}
(in place of Corollary~\ref{cor:general-bound}),
and the $h_2(k)^2$ factor from bounding $h_2(F_1)$ is replaced by $h_2(k)^{2^i}$ in the bound for $h_2(F_i)$.
We now treat the contribution from towers with $\Disc(F_i)\in \mathfrak{C}(X) = [X^{1/3-\delta_0},X^{1/3+\delta_0}]$. We separate the discussion into two cases depending on the size of $\Disc(F_{i-1})$. Fix $1 \leq Y_2 \leq X^{\frac{1}{6}-\frac{\delta_0}{2}}$. Then,
we find
\begin{equation}\label{eqn:induction-1}
\begin{aligned}
\sum_{\substack{F_{i+1}/F_i/\cdots/F_1/k \\ \Disc(F_{i-1})\ge Y_2 \\\Disc(F_i) \in \mathfrak{C}(X) \\ X/2\le \Disc(F_{i+1})\le X}} h_3(F_{i+1}/k)
& = \sum_{\substack{F_i/\cdots/F_1/k \\ \Disc(F_{i-1})\ge Y_2 \\ \Disc(F_i) \in \mathfrak{C}(X,Y)}} h_3(F_{i}/k) \sum_{\substack{F_{i+1}/F_i\\ X/2\le \Disc(F_{i+1})\le X}} h_3(F_{i+1}/F_i)\\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],i, \epsilon} h_2(k)^{2^i} \sum_{\substack{F_i/\cdots/F_1/k \\ \Disc(F_{i-1})\ge Y_2 \\ \Disc(F_i) \in \mathfrak{C}(X)}} h_3(F_{i}/k) \cdot \Disc(F_i)^{1/2+\epsilon} \cdot (\frac{X}{\Disc(F_{i})^2})^{3/2+\epsilon}\\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}], i,\epsilon} h_2(k)^{2^i} \Disc(k)^{\alpha_i}\cdot X^{3/2+\epsilon} \cdot \Big( \frac{ X^{-\frac{1}{2}+\frac{3\delta_0}{2}} }{Y_2^{1-\epsilon}} + X^{-(1/3-\delta_0)(3/2+\delta_i)}\Big)\\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}], i,\epsilon} \Disc(k)^{\alpha_i+2^{i-1}+\epsilon} \cdot \Big( \frac{ X^{1+{3\delta_0}/{2}+\epsilon} }{Y_2^{1-\epsilon}} + X^{ 1+3\delta_0/2-\delta_i/3+\delta_0\delta_i+\epsilon }\Big).
\end{aligned}
\end{equation}
Precisely, the first inequality follows from
\eqref{eq:h3-bijection},
Proposition \ref{prop:trivial-cubic-bound}, Lemma~\ref{lem:quadratic-bound},and Lemma~\ref{lem:two-part}. The second inequality follows from partial summation, and the induction hypothesis with $m=i$ applied to $F_i/\cdots/F_1/k$ with $Y_2$ being the lower bound for $\Disc(F_{i-1})$.
The third inequality follows from the trivial bound on $h_2(k)$.
Finally we consider those towers with $\Disc(F_{i-1}) \le Y_2$. Let $\delta'_2$ and $\alpha'_2$ be the constants associated to $[F_{i-1}:\ensuremath{{\mathbb{Q}}}]$ and $m =2$.
We have
\begin{equation}\label{eqn:induction-2}
\begin{aligned}
\sum_{\substack{F_{i+1}/F_i/\cdots/F_1/k \\ \Disc(F_{i-1})\le Y_2 \\\Disc(F_i) \in \mathfrak{C}(X) \\ X/2\le \Disc(F_{i+1})\le X}} h_3(F_{i+1}/k)
&= \sum_{\substack{F_{i-1}/\cdots/F_1/k \\ \Disc(F_{i-1})\le Y_2}} h_3(F_{i-1}/k) \sum_{\substack{F_{i+1}/F_i/F_{i-1}\\ \Disc(F_i) \in \mathfrak{C}(X)\\ X/2\le \Disc(F_{i+1})\le X}} h_3(F_{i+1}/F_{i-1})\\
& \ll_{ [k:\ensuremath{{\mathbb{Q}}}],i, \epsilon} \sum_{\substack{F_{i-1}/\cdots/F_1/k \\ \Disc(F_{i-1})\le Y_2}} h_3(F_{i-1}/k) \Big( X^{2/3+\delta_0+\epsilon} + X^{1-\delta'_2} \Big) \cdot \Disc(F_{i-1})^{\alpha'_2} \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],i} \Disc(k)^{\alpha_{i-1}} X^{\max\{{3/4+\delta_0, 1-\delta'_2}\}} Y_2^{1+\alpha'_2}.\\
\end{aligned}
\end{equation}
Here precisely, the first inequality follows from the induction hypothesis with $m=2$ and base field $k = F_{i-1}$. The second inequality follows from the induction hypothesis with $m=i-1$.
If we let $Y_2=X^b$, then to obtain the desired bound, \eqref{eqn:induction-1} requires us to choose $\delta_0$ small, given $b$ and $\delta_i$,
and \eqref{eqn:induction-2} requires us to take $b$ small given $\delta_2'$ and $\alpha_2'$ (as long as we have taken $\delta_0$ absolutely sufficiently small).
Since we can do this by choosing $b$ and then $\delta_0$, we will be able to obtain the theorem.
In particular, making the convenient but non-optimal choices $b=\min\{\delta'_2/(2(1+\alpha'_2)),1/(8(1+\alpha'_2))\}$ and $Y_2 = X^{b}$ and $\delta_0 = \min\{\delta_i/9, b/2,1/16\}$, we conclude that the statement of the theorem holds for $m=i+1$ with
\[
\alpha_{i+1} = \max\{\alpha_i+2^{i}, \alpha_{i-1}\} \quad \text{and} \quad \delta_{i+1} = \min\{\frac{\delta'_2}{2}, \frac{b}{8},\frac{\delta_i}{36},\frac{1}{32}\}.
\]
\end{proof}
\section{Average of $3$-torsion in Class Groups of $2$-extensions}\label{sec:main-theorem}
In this section, we will prove a refined version of Theorem~\ref{T:G}, in Theorem~\ref{thm:main} below, using crucially the tail estimates from Section \ref{sec:tail}.
We use the same approach to prove Theorem~\ref{T:rel}.
The
\emph{group signature} $\Sigma$ of $K/k$ is the ordered tuple $(\Sigma_v)_v$, where $v$ ranges over the real places of $k$, and $\Sigma_v$ is the conjugacy class of complex conjugation in $\Gal(K/k)$ over $v$. This further refines the enriched signature of Cohen and Martinet.
Let $E_k(m,X)$
be the set of degree $2^m$ $2$-extensions $K/k$ (in $\bar{\ensuremath{{\mathbb{Q}}}}$) with $\Disc K\leq X$.
Let $E_k(G,X)$
be the set of $G$-extensions $K/k$ (in $\bar{\ensuremath{{\mathbb{Q}}}}$) with $\Disc K\leq X$ and let $E_k^\Sigma(G,X)$ be those with group signature $\Sigma$.
\begin{theorem}\label{thm:main}
Let $k$ be a number field and $m$ a positive integer.
Then there exists $C_m>0$
such that
\begin{equation}\label{eqn:m-average}
\lim_{X\to\infty}\frac{1}{|E_k(m,X)|} {\sum_{K\in E_k(m,X)}\quad h_3(K)}=C_m.
\end{equation}
Moreover, let $G\subset S_{2^m}$ be a transitive permutation $2$-group containing a transposition and $\Sigma$ a group signature
that occurs for some $G$-extension of $k$.
Then there exists $C_{G,\Sigma},C_G>0$ such that
\begin{equation} \label{eqn:G-average}
\lim_{X\to\infty} \frac{1}{|E_k^\Sigma(G,X)|} {\sum_{K\in E_k^\Sigma(G,X)}\quad h_3(K)}= C_{G,\Sigma}
\quad \quad \textrm{and} \quad \quad
\lim_{X\to\infty} \frac{1}{|E_k(G,X)|} {\sum_{K\in E_k(G,X)}\quad h_3(K)} = C_{G}.
\end{equation}
\end{theorem}
The proof of Theorem \ref{thm:main} gives the constants explicitly.
We let $r_1(F)$ denote the number of real places of $F$, and $r_2(F)$ the number of pairs of complex places of $F$.
We have
$$
C_m= \left( \sum_{F \in E_k(m-1,\infty)}
\frac{h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{r_2(F)}\zeta_{F}(2)\Disc(F)^2 } \cdot \left(1 + \frac{2^{r_1(F)}}{3^{r_1(F)+r_2(F)}}\right) \right) \left(
\sum_{\substack{F\in E_k(m-1,\infty)}} \frac{\mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{r_2(F)} \zeta_{F}(2)\Disc(F)^2
}
\right)^{-1}.
$$
When the group $G$ has a transposition, Lemma \ref{lem:2-grp-with-transposition-is-wreath-product} below shows that $G=C_2\wr H$ for some transitive subgroup $H \subseteq S_{2^{m-1}}$.
The image of a group signature $\Sigma$ for $G$ gives a group signature $\bar{\Sigma}$ for $H$.
So a $G$-extension $K/k$ of signature $\Sigma$ has an index $2$ subfield $F/k$ which is an $H$-extension of signature $\bar{\Sigma}$, and
there is a number $u(\Sigma)$, depending only on $\Sigma$, which gives the number of infinite places of $F$ that are split in $K$, or equivalently, $\rk\O_K^*-\rk \O_F^*$, \emph{the relative unit rank of $K/F$}. Then we have
$$
C_{G,\Sigma}=\left(
\sum_{F \in E_k^{\bar{\Sigma}}(H,\infty)}
\frac{h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{\zeta_{F}(2)\Disc(F)^2 } \cdot \left(1 +3^{-u(\Sigma)}\right)
\right)\left(\sum_{F \in E_k^{\bar{\Sigma}}(H,\infty)}
\frac{ \mathrm{Res}_{s=1} \zeta_{F}(s) }{\zeta_{F}(2)\Disc(F)^2 } \right)^{-1},
$$
$$
C_G =\left( \sum_{F \in E_k(H,\infty)} \frac{h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{\zeta_{F}(2)\Disc(F)^2 } \cdot
\left(1 + \frac{2^{r_1(F)}}{3^{r_1(F)+r_2(F)}}\right) \right)\left( \sum_{F \in E_k(H,\infty)}
\frac{ \mathrm{Res}_{s=1} \zeta_{F}(s) }{\zeta_{F}(2)\Disc(F)^2 } \right)^{-1}.
$$
Before proving Theorem \ref{thm:main}, we first give a result on the asymptotic number of $2$-extensions, i.e. the denominator in Theorem \ref{thm:main}.
\begin{theorem} \label{thm:counting-2-ext}
Let $k$ be a number field and let $m$ be a positive integer.
There exists $D_m>0$ such that
\begin{equation}
|E_k(m,X)| \sim D_mX.
\end{equation}
Moreover, let $G\subset S_{2^m}$ be a transitive permutation $2$-group containing a transposition and $\Sigma$ a group signature.
Then there exists $D_{G,\Sigma}\geq 0$ and $D_G>0$ such that
\begin{equation}
|E_k^\Sigma(G,X)| \sim D_{G,\Sigma} X \quad \textrm{and}\quad |E_k(G,X)| \sim D_{G}X.
\end{equation}
If there exists an $H$-extension of $k$ with group signature $\bar{\Sigma}$, then $D_{G,\Sigma}>0$.
Further, if $G\subset S_{2^m}$ is a transitive permutation $2$-group not containing a transposition, then $|E_k(G,X)| = O_{k,G,\epsilon}(X^{1/2+\epsilon})$.
\end{theorem}
\begin{proof}
Malle \cite{Mal02,Mal04} conjectured that for $2$-groups $|E_k(G,X)|$ has linear growth if and only if $G$ contains a transposition.
When $G$ has no transpositions, it is proved by \cite[Corollary 7.3]{KlunersMalle} that $|E_k(G,X)| = O_{k,G,\epsilon}(X^{1/2+\epsilon})$. When $G$ contains a transposition, by \cite[Lemma $5.5$]{Klu12} $G$ must be isomorphic to $C_2\wr H$ for some permutation $2$-group $H$. It is proved in \cite[Theorem $5.8$]{Klu12} that when there exists a $H$-extension and $|E_k(H,X)|$ is not growing too fast, Malle's conjecture $|E_k(G,X)| \sim D_G X$ holds for $G = C_2 \wr H$. The existence of $2$-group extensions is shown by a celebrated theorem of Shafarevich \cite{Shafa}, and the upper bound for $|E_k(H,X)|$ is known for all permutation $2$-groups $H$, again by \cite[Corollary 7.3]{KlunersMalle}.
This gives the first statement of the theorem and a version of the second statement without signature conditions.
Following Kl\"{u}ners proof, or a simpler version of the proof of Theorem~\ref{thm:main}, one can
obtain a version with the signature condition and also with the precise constants
$$
D_m = \sum_{\substack{F\in E_k(m-1,\infty)}} \frac{\mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{r_2(F)} \zeta_{F}(2)\Disc(F)^2
},
$$
$$
D_{G,\Sigma} = \sum_{F \in E_k^{\bar{\Sigma}}(H,\infty)}
\frac{ p_\Sigma 2^{2^{m-1}-r_2(F)} \mathrm{Res}_{s=1} \zeta_{F}(s) }{\zeta_{F}(2)\Disc(F)^2 },
\quad \quad \textrm{and}\quad\quad
D_G=
\sum_{F \in E_k(H,\infty)}
\frac{2^{2^{m-1}-r_2(F)} \mathrm{Res}_{s=1} \zeta_{F}(s) }{\zeta_{F}(2)\Disc(F)^2 }
,
$$
where $\bar{\Sigma}$ is the group signature on $H$ given from $\Sigma$ via the quotient map $G\ra H$, and $p_\Sigma$ is the proportion of the
$2^{r_1(F)}$ possible behaviors at infinity for a quadratic extension $K/F$ such that when $K$ is a $G$-extension with such behavior then it has group signature $\Sigma$. (We note that $p_\Sigma$ is determined group theoretically from $\Sigma$ and is non-zero---see the proof of Theorem~\ref{thm:main}.)
\end{proof}
Theorem \ref{thm:counting-2-ext} shows that, among $2$-extensions of a given degree, the fields with Galois groups without transpositions are \emph{thin} families. Thus, the restriction in Theorem \ref{thm:main} to groups $G$ with transpositions is natural. In fact, while Theorem \ref{thm:main} does not give class number averages for the thin families of extensions with Galois group $G$ for $G$ without transpositions, a key step in its proof is to show that class group elements arising from such extensions
give a negligible contribution to the average in \eqref{eqn:m-average}.
This is a result of independent interest that complements Theorem \ref{thm:main}. Thus, we begin by considering thin families in Section \ref{ssec:thin-2-group}, which culminates in Theorem \ref{thm:3-torsion-thin-2-ext} that makes this discussion precise. Then, in Section \ref{ssec:proof-main-theorem}, we turn to the proof of Theorem~\ref{thm:main}.
\subsection{Thin families of $2$-extensions}\label{ssec:thin-2-group}
In this section, we focus on studying $3$-torsion in class groups of $2$-extensions with a permutation Galois group $G$ without a transposition.
We begin by proving an upper bound on the count of these extensions, but with explicit base field discriminant dependence. We first give the following lemma, which is a uniform version of Theorem $1.6$ in \cite{KluWan} for $\ell=2$.
\begin{lemma}\label{lem:disc-multi-general}
Let $k$ be a number field. Then there exists $\alpha_m>0$ depending at most on $[k:\ensuremath{{\mathbb{Q}}}]$ and $m$ such that the number of degree $2^m$ $2$-extension $K/k$ with $\Disc(K/k) = D$ is bounded by $O_{[k:\ensuremath{{\mathbb{Q}}}], m, \epsilon}(D^{\epsilon} \Disc(k)^{\alpha_m})$.
\end{lemma}
\begin{proof}
We proceed by induction over $m$. For $m = 1$, we know from class field theory (see the proof of Lemma~\ref{lem:pointwise-bound}) that the number of relative quadratic extension $K/k$ with $\Disc(K/k) = D$ is at most $O_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon}(h_2(k)\cdot D^{\epsilon}) = O_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon}( D^{\epsilon} \Disc(k)^{1/2+\epsilon})$. Now assuming the statement is true for $m =i$, we will prove that it also holds for $m = i+1$. By Lemma \ref{lem:tower}, it suffices to count quadratic extensions of degree $2^i$ $2$-extensions $F/k$ with $\Disc(F/k)^2| D$. The number of such $F$ is bounded by $O_{[k:\ensuremath{{\mathbb{Q}}}], \epsilon, i}( D^{\epsilon} \Disc(k)^{\alpha_i})$ by the induction hypothesis. For each such $F$, the number of quadratic relative extension $K/F$ with $\Disc(K/F) = D/\Disc(F/k)^2$ is bounded by $O_{[F:\ensuremath{{\mathbb{Q}}}], \epsilon}(h_2(F) \cdot D^{\epsilon})$ from class field theory. Applying Lemma \ref{lem:two-part} and summation over all divisors of $D$, we obtain the upper bound $O_{[k:\ensuremath{{\mathbb{Q}}}], i+1, \epsilon}( D^{\epsilon} \Disc(k)^{\alpha_{i+1}})$, where $\alpha_{i+1}$ can be taken as $\alpha_{i}+ 2^{i-1}$.
\end{proof}
\begin{theorem}\label{thm:counting-thin-2-extensions}
Let $k$ be a number field $k$ and let $G \subseteq S_{2^m}$ be a transitive permutation $2$-group without a transposition.
Then there exists $\alpha>0$ depending at most on $[k:\ensuremath{{\mathbb{Q}}}]$ and $m$ such that
$$|E_k(G, X)| = O_{[k:\ensuremath{{\mathbb{Q}}}],m,\epsilon}(X^{1/2+\epsilon} \Disc(k)^{\alpha}).$$
\end{theorem}
\begin{proof}
Notice that when $G$ does not contain a transposition, the discriminant ideal $\disc(K/k)$ must have exponent at least $2$ at every ramified prime, i.e., it is powerful. Thus, its norm, $\Disc(K/k)$, must be a powerful integer, and there are $O(X^{1/2})$ powerful integers below $X$. The result now follows from Lemma \ref{lem:disc-multi-general}.
\end{proof}
The $\ell$-torsion conjecture then would imply that the summation of $|\Cl_K[3]|$ for $G$-extensions $K$ without a transposition should be also thin, i.e. $O(X^{1/2+\epsilon})$.
By applying the trivial bound for class group $h_3(K)= O_{[K:\ensuremath{{\mathbb{Q}}}], \epsilon}(\Disc(K)^{1/2+\epsilon})$, we can conclude immediately that $\sum_{K\in E_k(G,X)} h_3(K)= O_{k, \epsilon}(X^{1+\epsilon})$. However, it is crucial for our main theorem that we prove something slightly better than this.
\begin{theorem}\label{thm:3-torsion-thin-2-ext}
Let $k$ be a number field and let $G\subset S_{2^m}$ be a transitive permutation $2$-group without a transposition.
Then there exist $\delta>0$ and $\alpha$ depending on $G$ and $[k:\ensuremath{{\mathbb{Q}}}]$ such that
$$\sum_{K\in E_k(G,X)} h_3(K/k)= O_{[k:\ensuremath{{\mathbb{Q}}}],G}(X^{1-\delta} \cdot \Disc(k)^{\alpha}).$$
\end{theorem}
\begin{proof}
A $G$-extension $K$ can be constructed as a relative quadratic extension over a degree $2^{m-1}$ $2$-extension $F$ with $\Gal(F/k) =: H\subset S_{2^{m-1}}$ by Lemma \ref{lem:tower}. We fix $\delta_1=1/24$
and $\epsilon_1=1/4$.
Let $\gamma$, $\beta$ and $D_0$ be absolute constants allowable in Lemma \ref{lem:prime-number-theorem-k-dependence}. Let $Y = X^{\delta'}$ for
some $\delta'$ that we will choose sufficiently small in terms of $m$ and $[k:\ensuremath{{\mathbb{Q}}}]$.
Then by Theorem \ref{thm:tail} and dyadic summation, taking $\delta'\leq \delta_m$, we have
$$
\sum_{\substack{ K/F/k\\ \Disc(K)\le X \\ \Disc(F)\ge Y}} h_3(K/k) \ll_{[k:\ensuremath{{\mathbb{Q}}}], m,\epsilon} \Disc(k)^{\alpha_m} X^{1-\delta'+\epsilon}.
$$
Therefore it suffices to consider the summation over $G$-extensions where the associated $H$-extension $F/k$ has $\Disc(F)\le Y = X^{\delta'}$.
We will apply an argument similar to Section \ref{ssec:initial-critical} to complete the proof.
For each $F$, denote the set $\mathcal{E}(F) = \mathcal{E}(F, X/\Disc(F)^2, \epsilon_1)$ of exceptional quadratic extensions as given by Lemma \ref{lem:effective-Chebo-over-k-average}, for $X \gg 1$. The set has size $|\mathcal{E}(F)| \ll_{[F:\ensuremath{{\mathbb{Q}}}],\epsilon_1 } X^{\epsilon_1}$. Let $x = (X/(2Y^2))^{\delta_1}$. We assume $X$ is sufficiently large in terms of $[k:\mathbb{Q}]$
and $\delta'$ sufficiently small in terms of our absolute constants,
so that for $\sigma_1 = \max\{1- \epsilon_1/4c,1/2\}$ with $c = c_{[F:\ensuremath{{\mathbb{Q}}}]}$ in Theorem \ref{thm:zero-density} and $F$ with $\Disc(F) \geq D_0$, we may apply Lemmas \ref{lem:effective-Chebo-over-k-average} and \ref{lem:prime-number-theorem-k-dependence} in concert to conclude for $K\not\in \mathcal{E}(F)$,
\begin{equation}\label{eqn:thin-primes}
\pi_F(x;K, e) \geq \frac{1}{8}\pi_F(x/2)- C_{[F:\mathbb{Q}],\epsilon_1} x^{\sigma_1} (\log (X\Disc(F)))^2 \gg_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \frac{x}{\Disc(F)^{\gamma}\log x}.
\end{equation}
We also assume that $X$ is sufficiently large
that this holds for the finite number of fields $F$ with $\Disc(F) \leq D_0$.
Using the trivial bound for $h_3(K/k)$, the summation over $K/F/k$ with $K/F\in \mathcal{E}(F)$ is
\begin{equation}\label{eqn:thin-2-extension-exceptional}
\begin{aligned}
\sum_{\substack{K/F/k \\ \Disc(F)\le Y\\\Disc(K)\le X \\K/F\in \mathcal{E}(F)}} h_3(K/k) & \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon} \sum_{\Disc(F)\le Y} \sum_{\substack{ K/F\in \mathcal{E}(F)}} \Disc(K)^{1/2+\epsilon} \\
& \ll_{[k:\ensuremath{{\mathbb{Q}}}],\epsilon_1,m, \epsilon} \Disc(k)^{\alpha_{m-1}} \cdot X^{1/2+\epsilon_1+ \epsilon} Y,
\end{aligned}
\end{equation}
where the last inequality follows from combining the bound on $|\mathcal{E}(F)|$ with an upper bound on the number of such $F$, for example as is provided by Corollary \ref{cor:general-bound-m} and dyadic summation.
For fields $K/F \notin \mathcal{E}(F)$, we use Lemma \ref{lem:relative-E-V}, \eqref{eqn:thin-primes}, and the trivial bound on $h_3(F/k)$ to obtain that $h_3(K/k) \ll_{[k:\mathbb{Q}],m,\epsilon} X^{1/2-\delta_1+\epsilon}Y^{2\delta_1+\gamma}$. Then we find
\begin{align*}
\sum_{\substack{K/F/k \\ X/2\le \Disc(K)\le X \\ \Disc(F)\le Y\\K/F\notin \mathcal{E}(F)\\ \Gal(K/k) = G}} h_3(K/k)
&\ll_{[k:\ensuremath{{\mathbb{Q}}}],m,\epsilon} \sum_{\substack{K/k \\ X/2 \leq \Disc(K) \leq X \\ \Disc(F)\le Y\\ \mathrm{Gal}(K/k) = G}} X^{1/2-\delta_1+\epsilon}Y^{2\delta_1+\gamma} \\
&\ll_{[k:\ensuremath{{\mathbb{Q}}}],m,\epsilon} \Disc(k)^{\alpha} X^{1-\delta_1+\epsilon}Y^{2\delta_1+\gamma},
\end{align*}
by Theorem \ref{thm:counting-thin-2-extensions}, where $\alpha$ is associated to $[k:\ensuremath{{\mathbb{Q}}}]$ in Theorem \ref{thm:counting-thin-2-extensions}.
We can use dyadic summation to remove the lower bound on $\Disc(K)$ and obtain the same bound (with a different implied constant).
Finally, for $X\ll_{[k:\ensuremath{{\mathbb{Q}}}],\delta'} 1$, we have
\begin{align*}
\sum_{\substack{K/F/k \\ X/2\le \Disc(K)\le X \\ \Disc(F)\le Y\\ \Gal(K/k) = G}} h_3(K/k)
&\ll_{[k:\ensuremath{{\mathbb{Q}}}],m,\delta'} \Disc(F)^{\alpha_m} \leq X^{\delta' \alpha_m}
\end{align*}
by Corollary \ref{cor:general-bound-m}. Choosing $\delta'$ and sufficiently small in terms of $[k:\ensuremath{{\mathbb{Q}}}]$ and $m$, we conclude the theorem.
\end{proof}
\subsection{Proof of the Main Theorem}\label{ssec:proof-main-theorem}
Now we are ready to prove Theorem \ref{thm:main}. We start with the following lemmas that allow us to move from summing over $G$ extensions to summing over $H$-extensions, where $G\simeq C_2\wr H$.
\begin{lemma}\label{lem:2-grp-with-transposition-is-wreath-product}
A transitive permutation $2$-group $G\subset S_{2d}$ contains a transposition if and only if $G\simeq C_2\wr H$ for some $H\subset S_{d}$.
For such a $G$,
if $K/k$ is a $G$-extension, then there exists a unique subfield $F/k$ with $[K:F] = 2$. Moreover, $F/k$ is an $H$-extension.
For such a $G$, there is only one permutation group $H$ up to isomorphism such that $G\simeq C_2\wr H$.
\end{lemma}
\begin{proof}
The first statement is Lemma $5.5$ in \cite{Klu12}. For the second statement, we claim
if $M$ is a subgroup of $G$ and contains $\Stab_G(1)$ as a proper subgroup, then $C_2^d\rtimes \Stab_H(1)\subset M$.
To show the claim, we write $\Stab_G(1) = C_2^{d-1}\rtimes \Stab_H(1)$ where $C_2^{d-1}$ is the subspace of $C_2^d$ with first entry $0$. Suppose $g=v\rtimes h\in M$ and $g\notin \Stab_G(1)$. If $h\in \Stab_H(1)$, then $v_1\neq 0$, and multiplication by all $u\rtimes h^{-1}\in \Stab_G(1)$ shows that $C_2^d \subset M$.
If $h\notin \Stab_H(1)$, then for $u\in C_2^{d-1}$ with a $1$ in position $h^{-1}(1)$ and $0$'s elsewhere, we have that
$(v\rtimes h) u (v\rtimes h)^{-1}\in C_2^d$ with a $1$ in position $1$ and $0$'s elsewhere, so we conclude the claim.
The second and third statements of lemma then follow by Galois theory.
If $G \isom C_2\wr H$, then $H$ acts on $2$-element blocks of the elements that $G$ acts on, and all elements of $G$ preserve the block structure. From this, it follows that two elements are together in a block if and only if $G$ contains a transposition that interchanges them. From this it follows that $H$ is determined uniquely as a permutation group.
\end{proof}
We now want to be precise about exactly what we mean by permutation group isomorphisms.
\begin{definition}\label{D:permgroup}
A permutation group $G$ is a group $G$, a set $B_G$, and a faithful action of $G$ on $B_G$.
The degree of $G$ is $|B_G|$.
An isomorphism of permutation groups $G\ra H$ is a bijection $\sigma:B_G \ra B_H$
so that the induced map $\sigma_*: G\ra H$ is a a group isomorphism (where $\sigma_*(g)(b)= \sigma g\sigma^{-1}(b)$).
We write $\Aut_{\operatorname{perm}}(G)$ for the group of permutation automorphisms of $G$, which is also the normalizer $N_{S_{B_H}}(H)$ of $H$ in the symmetric group $S_{B_H}$.
\end{definition}
In particular, a $G$-extension $K/k$ includes the data of a bijection between the embeddings $K\ra \bar{k}$ and $B_G$. Note that $B_{C_2\wr H}=\{1,2\}\times B_H$.
\begin{lemma}\label{lem:KFbij}
Let $G$ be a transitive permuation $2$-group and $G=C_2\wr H$, with $H$ a permutation group of degree $d$.
Given an $H$-extension $F\in E_k(H,\infty)$, and a quadratic extension $K/F$ (in $\bar{\ensuremath{{\mathbb{Q}}}}$), then either (1)$\Gal(K/k)$ does not contain a transposition or
(2)$K/k$ can be realized as a $G$-extension with $2^d$ choices of
$\Gal(K/k)\isom G$
(i.e. choices of $\{K \ra\tilde{K} \}\ra B_G$)
that are compatible with the permutation group isomorphism $\Gal(F/k)\isom H$ (i.e. choice of $\{F \ra\tilde{F} \}\ra B_H$)
in the quotient.
Moreover, any $G$-extension $K\in E_k(G,\infty)$ arises in this construction from a unique $F$, one quadratic extension $K/F$, and one of the
$2^d$ choices of $\Gal(K/k)\isom G$.
\end{lemma}
\begin{proof}
Given $F\in E_k(H,\infty)$ and a quadratic extension $K/F$, we have $2^d$ choices of identification of the embeddings $K\ra \tilde{K}$ with $\{1,2\}\times B_H$ that are compatible with
the map from the embeddings of $F$ to $B_H$. Any of these gives an injection of groups $\Gal(K/k)\sub C_2\wr H$ compatible with the permutation actions.
Any transposition in $C_2\wr H$ is an element of $C_2^d$ that is non-trivial in exactly one coordinate. Since $H$ is transitive, this implies that any subgroup of
$C_2\wr H$ containing a transposition and with image in $H$ all of $H$
must contain all of $C_2^d$ and thus be all of $C_2\wr H$.
So if $\Gal(K/k)$ contains a transposition, we see that $\Gal(K/k)\isom G$, and there are $2^d$ ways of choosing such a permutation group isomorphism that are compatible with the map
from the embeddings of $F$ to $B_H$.
Given a $G$-extension $K\in E_k(G,\infty)$, from Lemma~\ref{lem:2-grp-with-transposition-is-wreath-product} we see that it
arises in this way from a unique $F$, and the lemma follows.
\end{proof}
Next we recall the following important result by Datskovsky--Wright \cite{DW88}.
\begin{theorem}[Datskovsky--Wright]\label{thm:cubic-count}
Let $F$ be a number field. For a group signature $\Sigma$ of a quadratic extension $K/F$, let $u(\Sigma)$ be the number of infinite places
of $F$ where $\Sigma$ is trivial (which is also the relative unit rank of $K/F$).
Then
\[
\sum_{ \substack{ [K:F]=2\\
\textrm{signature $\Sigma$} \\ |\mathrm{Disc}(K/F)| \leq X }} h_3(K/F)
\sim X \cdot
\frac{\mathrm{Res}_{s=1}\zeta_F(s)}{2^{r_1(F)+r_2(F)}\zeta_F(2)}(1+3^{-u(\Sigma)})
\quad \textrm{and} \sum_{ \substack{ [K:F]=2\\
\textrm{signature $\Sigma$} \\ |\mathrm{Disc}(K/F)| \leq X }} 1
\sim X \cdot
\frac{\mathrm{Res}_{s=1}\zeta_F(s)}{2^{r_1(F)+r_2(F)}\zeta_F(2)}.
\]
\end{theorem}
\begin{proof}
This follows from combining Theorems 4.2 and 5.1 in \cite{DW88}.
\end{proof}
We now turn to the proof of our main theorem.
\begin{proof}[Proof of Theorem \ref{thm:main}]
Suppose $G \subseteq S_{2^m}$ is a $2$-group with a transposition, so by Lemma \ref{lem:2-grp-with-transposition-is-wreath-product}, $G = C_2 \wr H$ for a unique $H \subseteq S_{2^{m-1}}$.
Let $\bar{\Sigma}$ be the group signature for $H$ that is the image of $\Sigma$.
So a $G$-extension $K/k$ with group signature $\Sigma$ has an index two subfield $F/k$ with the group signature $\bar{\Sigma}$.
We next wish to show that the group signature of $K/k$ is determined by $F$ and the group signature of $K/F$.
We first consider the structure of order $2$ elements of $G$, to understand the possible group signatures.
For $g\in G=C_2 \wr H$, let $\bar{g}$ denote the image of $g$ in $H$.
The group $H$ acts on a set of $2^{m-1}$ elements $B_H$. The element $g\in G$ has order $2$ if and only if $\bar{g}$ has order $2$ and the $C_2$ coordinates of $g$ (of which there is one for each element of $B_H$) are constant on $\bar{g}$ orbits of $B_H$.
Moreover, if $\bar{g}$ transposes two elements $a,b$ of $B_H$, and $g'$ is obtained from $g$ by changing the $C_2$ coordinates at positions $a,b$, then $g'$ is conjugate to $g$. In particular, a conjugacy class $[g]$ of an element $g\in G$ is determined by the image of the conjugacy class in $H$, and for any element $h$ of that conjugacy class in $H$, the $C_2$ coordinates of the $h$ fixed points of $B_H$.
Next, we need to relate the infinite places of $F$ above a place $v$ of $k$ to the permutation group $H$. To do this, at each real place $v$ of $k$ we choose an embedding $i_v: \bar{\ensuremath{{\mathbb{Q}}}}\ra \C$ that restricts to $v$ on $k$.
If $F$ is an $H$-extension, the elements $B_H$ correspond to embeddings $\tau: F\ra \tilde{F}$ (as part of the data of the $H$-extension), and via $i_v:\bar{\ensuremath{{\mathbb{Q}}}}\ra \C$ these correspond to embeddings $i_v\circ \tau: F\ra \C$. So for each real place $v$ of $k$, we now have a correspondence between $B_H$ and the complex embeddings of $F$ that restrict to $i_v$ on $k$.
With these choices of $i_v$, we can now specify an actual element of $H$ corresponding to complex conjugation over $v$ (as opposed to a conjugacy class). Precisely,
we define $\sigma_v(F)$ to be the element of $H$ that acts on the embeddings $\tau: F\ra \tilde{F}$ so that $\sigma_v(F)(\tau)=i_v^{-1} \circ \overline{i_v\circ\tau}$, where the bar denotes complex conjugation, i.e. $\sigma_v(F)$
is the pullback of complex conjugation via $i_v$.
Similarly, for a $G$-extension $K/k$ of group signature $\Sigma$, we have
$\sigma_v(K) \in {\Sigma}_v$.
If $F/k$ is the index two subfield of $K$, then
at the fixed points of $\sigma_v(F)$, our chosen complex conjugation, on $B_H$, the $C_2$ coordinate of $\sigma_v(K)$ is trivial when $K/F$ is split at the corresponding place of $F$ and non-trivial when $K/F$ is ramified at that place.
Using the conclusion of the previous paragraph, we conclude that if $K/k$ is a $G$-extension with index two subfield $F/k$,
the conjugacy class of $\sigma_v(K)$, and hence the group signature of $K/k$, is determined by $F$ and the group signature of $K/F$.
Next, we wish to determine the number of group signatures of the quadratic extension $K/F$ that will lead to group signature $\Sigma$ for $K/k$.
Let $F/k$ be an $H$-extension with group signature $\bar{\Sigma}.$
Let $M_F$ be the set of group signatures $\Sigma_2$ for quadratic extensions of $F$ such that
$K/k$ has group signature $\Sigma$ when $K/F$ has group signature $\Sigma_2$.
For a $G$-extension $K/k$ of group signature $\Sigma$, we let $u(\Sigma)$ be the difference in unit ranks between $K$ and $F$, which only depends on $\Sigma$.
Note that for any $\Sigma_2\in M_F$, we have that $u(\Sigma)$ is the number of split places in $\Sigma_2$.
Also let $M_\Sigma:=|M_F|$, and note that it only depends on $\Sigma$, since if we choose $\sigma_v\in \Sigma_v$ for all real places $v$ of $k$, then
$M_\Sigma$ is the product over $v$ of the number of elements $u\in C_2^{(B_H^{\sigma_v})}$ such that $u\sigma_v$ is conjugate to $\sigma_v$
(where $B_H^{\sigma_v}$ denotes the elements of $B_H$ fixed by $\sigma_v$).
Given an $H$-extension $F$, let $\mathcal{Q}_F^{\Sigma}$ be the set of quadratic extensions $K/F$ (in $\bar{\ensuremath{{\mathbb{Q}}}}$) of group signature in $M_F$.
Let $Y \geq 1$, and suppose $X_0 = X_0(Y)$ is sufficiently large so that: 1) $X_0\geq Y^2$; 2) and
for any $X\geq X_0$ and each $F \in E_k(H,Y)$, we have
\begin{equation} \label{eqn:X-hypothesis}
\left|\sum_{ \substack{ K\in \mathcal{Q}_F^{\Sigma} \\ \mathrm{Disc}(K/F) \leq \frac{X}{\Disc(F)^2 }
}} h_3(K/F)
- \frac{X}{\Disc(F)^2} \cdot \frac{\mathrm{Res}_{s=1} \zeta_F(s)}{2^{r_1(F)+r_2(F)}\zeta_F(2)} \cdot \left(1 + 3^{-u(\Sigma)}\right) \right| \leq \frac{X}{2^{2^{m-1}} h_3(F) Y |E_k(H,Y)|}.
\end{equation}
Since there are only finitely many fields $F \in E_k(H,Y)$, such an $X_0$ formally exists by Theorem \ref{thm:cubic-count}. By Theorem \ref{thm:tail} and dyadic summation, we find for any $X \geq X_0$
\begin{align*}
\sum_{K \in E_k^{\Sigma}(G,X)} h_3(K)
&= \sum_{\substack{ K \in E_k^{\Sigma}(G,X) \\ \Disc(F) \leq Y }} h_3(K) + O_{k,m,\epsilon}(X/Y^{1-\epsilon} + X^{1-\delta_m}),
\end{align*}
where $F \in E_k^{\bar{\Sigma}}(H,Y)$ is the unique index 2 subfield of $K$ containing $k$.
Then by
Lemma~\ref{lem:KFbij} and
Theorem \ref{thm:3-torsion-thin-2-ext}, we find, for $X\geq X_0$,
\begin{align*}
\sum_{\substack{ K \in E_k^{\Sigma}(G,X) \\ \Disc(F) \leq Y }} h_3(K)
&=
\sum_{\substack{F \in E_k^{\bar{\Sigma}}(H,Y) }}
h_3(F)
\sum_{\Sigma_2\in M_F} 2^{2^{m-1}}
\sum_{ \substack{ [K:F]=2\\\textrm{signature $\Sigma_2$} \\ \mathrm{Disc}(K/F) \leq X/\Disc(F)^2
}} h_3(K/F) + O_k(X^{1 - \delta}).
\end{align*}
Then by \eqref{eqn:X-hypothesis}, for $X\geq X_0$,
\begin{align*}
\sum_{\substack{ K \in E_k^{\Sigma}(G,X) \\ \Disc(F) \leq Y }} h_3(K)
&= X \sum_{\substack{F \in E_k^{\bar{\Sigma}}(H,Y) }}
\frac{M_\Sigma 2^{2^{m-1}}h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{
r_1(F)+
r_2(F)} \zeta_{F}(2)\Disc(F)^2
} \cdot \left(1 + 3^{-u(\Sigma)} \right) + O_k(X^{1-\delta}) + O(X/Y),
\end{align*}
and thus also
\[
\frac{1}{X}\sum_{K \in E_k^{\Sigma}(G,X)} h_3(K)
= \sum_{F \in E_k^{\bar{\Sigma}}(H,Y)}
\frac{ M_\Sigma 2^{2^{m-1}} h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{
r_1(F)+
r_2(F)} \zeta_{F}(2)\Disc(F)^2
} \cdot \left(1 +3^{-u(\Sigma)}\right) + O_{k,m,\epsilon}(Y^{-1+\epsilon} + Y^{-\delta_m} + Y^{-\delta}),
\]
since we assumed $X \geq X_0 \geq Y^2$. Letting $Y \to \infty$, the sum over $F \in E_k^{\bar{\Sigma}}(H,Y)$ converges by virtue of Theorem \ref{thm:counting-2-ext}, the trivial bounds $h_3(F) \ll_{[F:\mathbb{Q}],\epsilon} \Disc(F)^{1/2+\epsilon}$ and $\Res_{s=1}\zeta_F(s) \ll_{[F:\mathbb{Q}],\epsilon} \Disc(F)^{\epsilon}$, and partial summation.
The error term on the right-hand side also converges, so the left-hand side must converge as well. It follows that
\[
\lim_{X \to \infty} \frac{1}{X} \sum_{K \in E_k^{\Sigma}(G,X)} h_3(K)
= \sum_{F \in E_k^{\bar{\Sigma}}(H,\infty)}
\frac{M_\Sigma 2^{2^{m-1}} h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{
r_1(F)+
r_2(F)}\zeta_{F}(2)\Disc(F)^2 } \cdot \left(1 +3^{-u(\Sigma)}\right).
\]
Let $\Sigma_0$ be a group signature for an $H$-extension $F/k$. Then
$$
\sum_{\substack{\Sigma\\ \bar{\Sigma}=\Sigma_0}} M_\Sigma =2^{r_1(F)}
\quad \textrm{and}\quad
\sum_{\substack{\Sigma\\ \bar{\Sigma}=\Sigma_0}} M_\Sigma 3^{-u(\Sigma)} =\sum_{i=0}^{r_1(F)} \binom{r_1(F)}{i} 3^{-i-r_2(F)}=
\frac{2^{2r_1(F)}}{3^{r_1(F)+r_2(F)}}.
$$
where the sums are over group signatures $\Sigma$ for $G$-extensions, since the choices of $\Sigma$ with $i$ real places of $F$ split have $u(\Sigma)=i+r_2(F).$
We can then sum over all group signatures to obtain
\begin{align*}
\lim_{X \to \infty} \frac{1}{X} \sum_{K \in E_k(G,X)} h_3(K)
&=\sum_{\bar{\Sigma}}
\sum_{\substack{\Sigma\\ \textrm{given }\bar{\Sigma}}}
\sum_{F \in E_k^{\bar{\Sigma}}(H,\infty)}
\frac{M_\Sigma 2^{2^{m-1}} h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{
r_1(F)+
r_2(F)}\zeta_{F}(2)\Disc(F)^2 } \cdot \left(1 +3^{-u(\Sigma)}\right)
\\
&= \sum_{F \in E_k(H,\infty)} \frac{2^{2^{m-1}}h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{r_2(F)}\zeta_{F}(2)\Disc(F)^2 } \cdot
\left(1 + \frac{2^{r_1(F)}}{3^{r_1(F)+r_2(F)}}\right).
\end{align*}
To count each field $K$ exactly once (instead of once for each $G$ extension structure), we can divide the above sum by $|\Aut_{\operatorname{perm}}(G)|$.
We then sum over (isomorphism classes of ) $G$ of the form $C_2\wr H$, and by Lemma~\ref{lem:2-grp-with-transposition-is-wreath-product}
this will give a sum over $T_m$, the set of all (isomorphism classes of) transitive permutation $2$-groups $H$ of degree $ 2^{m-1}$. Using Theorem~\ref{thm:3-torsion-thin-2-ext}, we obtain
\begin{align*}
&\lim_{X \to \infty} \frac{1}{X} {\sum_{K \in E_k(m,X)} h_3(K)}\\
&= \sum_{\substack{H\in T_m } }
\frac{1}{|\Aut_{\operatorname{perm}}(C_2\wr H)|}
\sum_{F \in E_k(H,\infty)} \frac{2^{2^{m-1}}h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{r_2(F)}\zeta_{F}(2)\Disc(F)^2 } \cdot
\left(1 + \frac{2^{r_1(F)}}{3^{r_1(F)+r_2(F)}}\right).
\end{align*}
Above, each field $F$ appears in the sum $|\Aut_{\operatorname{perm}}(H)|$ times.
Next, we claim that the factors have value $|\Aut_{\operatorname{perm}}(H)|/ |\Aut_{\operatorname{perm}}(C_2\wr H)|=2^{-2^{m-1}}.$
The blocks of $B_G$ that $C_2\wr H$ acts on are the pairs that appear in $G$ as transpositions, so these are preserved by any permutation isomorphism of $C_2\wr H$, and we have a map $\Aut_{\operatorname{perm}}(C_2\wr H)\ra \Aut_{\operatorname{perm}}(H)$. It is easy to see this is a surjection and the kernel is given by all the permutations of $\{1, 2\}\times B_H$ that fix the $B_H$ coordinate, showing the claim.
Thus we have
$$
\lim_{X \to \infty} \frac{1}{X} {\sum_{K \in E_k(m,X)} h_3(K)}
=
\sum_{F \in E_k(m-1,\infty)}
\frac{h_3(F) \mathrm{Res}_{s=1} \zeta_{F}(s) }{2^{r_2(F)}\zeta_{F}(2)\Disc(F)^2 } \cdot \left(1 + \frac{2^{r_1(F)}}{3^{r_1(F)+r_2(F)}}\right).
$$
Combined with Theorem \ref{thm:counting-2-ext}, this yields the theorem.
(Note we may obtain the constants in Theorem \ref{thm:counting-2-ext} by the same argument as above, using the count of quadratic extensions of $F$ in Theorem~\ref{thm:cubic-count} in place of the average $3$-torsion result.)
\end{proof}
\subsection{Relative Class Group Averages}
For the families in Theorem \ref{thm:main}, the approach of the proof of Theorem \ref{thm:main} also permits us to determine the average size of the $3$-torsion subgroup of the relative class group $\Cl_{K/F}$,
where $F$ is the index two subfield of $K$, proving Theorem~\ref{T:rel} (when combined with Proposition~\ref{P:asCM}).
These averages are particularly nice when the Galois closure group and
$u := r_1(K)+r_2(K)-r_1(F)-r_2(F)$
is fixed.
For $G$ a transitive permutation $2$-group with a transposition, let $E_k^u(G,X)$ denote the set of $G$-extensions $K/k$ with $\Disc K\leq X$ and
$K/F$ of relative unit rank $u$, where $F/k$ is the unique index $2$ subfield of $K$.
\begin{theorem}\label{thm:relative-average}
Let $k$ be a number field, $m$ a positive integer, and $G \subseteq S_{2^m}$ a transtive $2$-group with a transposition. If $u$ is such that
$E_k^u(G,\infty)$ is non-empty, then
\[
\lim_{X \to \infty} \frac{1}{|E_k^u(G,X)|} \sum_{K \in E_k^u(G,X)} h_3(K/F)
= 1 + 3^{-u}.
\]
\end{theorem}
\begin{proof}
The proof uses the main input we have built for the proof of Theorem \ref{thm:main}
and
is nearly identical to that proof, so we will be brief.
Let $F/k$ be the index $2$ subfield of $K$.
In particular, using the trivial inequality $h_3(K/F) \leq h_3(K/k)$, it follows from Theorem \ref{thm:tail} that for any $X,Y>0$,
\[
\sum_{K \in E_k^u(G,X)} h_3(K/F)
= \sum_{\substack{ K \in E_k^u(G,X) \\ \Disc(F) \leq Y}} h_3(K/F)
+ O_{k,m,\epsilon}(X/Y^{1-\epsilon}+X^{1-\delta_m}).
\]
Again using the trivial inequality $h_3(K/F) \leq h_3(K/k)$, we see from Lemma~\ref{lem:KFbij} and Theorem \ref{thm:3-torsion-thin-2-ext} that
\[
\sum_{\substack{ K \in E_k^u(G,X) \\ \Disc(F) \leq Y}} h_3(K/F)
= \sum_{F \in E_k(H,X)} 2^{2^{m-1}} \sum_{\substack{ [K:F] = 2 \\ \Disc(K/F) \leq X / \Disc(F)^2 \\ \mathrm{rk}(\mathcal{O}_K^\times / \mathcal{O}_F^\times) = u}} h_3(K/F) + O_{m,k}(X^{1-\delta}),
\]
for some $\delta>0$ depending on $m$ and $k$.
Let $r$ be the number of group signatures of quadratic extensions of $F$ such that $K/F$ of that signature have $\mathrm{rk}(\mathcal{O}_K^\times / \mathcal{O}_F^\times) = u$.
We take $X_0(Y)$ as in the proof of Theorem \ref{thm:main}, and have, for $X\geq X_0$
$$
\sum_{\substack{ K \in E_k^u(G,X) \\ \Disc(F) \leq Y}} h_3(K/F)
= \sum_{F \in E_k(H,X)}
\frac{r2^{
2^{m-1}-r_1(F)-r_2(F)} \mathrm{Res}_{s=1} \zeta_{F}(s) }{ \zeta_{F}(2)\Disc(F)^2
} \cdot \left(1 + 3^{-u} \right)
+ O_{m,k}(X^{1-\delta}+X/Y)
$$
Proceeding as in the proof of Theorem \ref{thm:main},
the result follows.
\end{proof}
\section{Comparison to the Cohen--Martinet heuristics}\label{sec:CM}
\subsection{Cohen--Martinet prediction for the average of $h_3(K/F)$}
Theorem \ref{thm:relative-average} verifies new cases of the Cohen--Martinet heuristics, as we now show.
\begin{proposition}\label{P:asCM}
The Cohen--Martinet heuristics \cite[Hypoth\'ese 6.6]{Cohen1990} predict that the average of Theorem~\ref{thm:relative-average} is as proved in the theorem.
\end{proposition}
\begin{proof}
Let $G$ be a transitive permutation $2$-group with a transposition and $S_K$ the stabilizer of an element $1\in B_G$. Our $K\in \mathcal{F}^u(G)$ is then the $S_K$ fixed field of a Galois $G$ extension $\tilde{K}/k$ with Galois group $G$. By Lemma~\ref{lem:2-grp-with-transposition-is-wreath-product}, we have $G=C_2 \wr H$, where $H$ is a permutation group of degree $d$.
In this notation, we have that $S_K=(1\times C_2^{d-1}) \rtimes \Stab_H(1)$, where $1\in B_H$ is the image of $1\in B_G$.
Then we define $S_F=(C_2^d) \rtimes \Stab_H(1)$ (the fixed field of $S_F$ will be the field we call $F$, the index $2$ subfield of $K$).
Let $U$ be the sign representation of $S_F/S_K$ over $\F_3$.
We claim
$h_3(K/F)=|\Hom_{S_F} (\Cl_{\tilde{K}/k},U )|$, where $\Hom_{S_F}$ denotes morphisms of $S_F$ modules.
To see this, let $e\in \F_3[G]$ be the element $e=|S_K|^{-1}{\sum_{g\in S_K} g}$. We have a map $m_e:\Cl_{\tilde{K}/k}/3 \ra \Cl_{K/k}/3$ given by multiplication by $e$, which is a surjection (as inclusion of ideals gives a section). Since $e$ acts as the identity on $U$, we have that any element of
$\Hom_{S_F} (\Cl_{\tilde{K}/k},U )$ factors through $m_e$, so is determined by a map $\Hom_{S_F} (\Cl_{K/k}/3,U ).$ Now $\Cl_{K/k}/3$ is a product of $\pm 1$ eigenspaces for $S_F/S_K$, and $\Cl_{K/F}/3$ is exactly the $-1$ eigenspace, which proves the claim.
Let $V:= \Ind_{{S_F}}^{G} U$. We claim that $V$ is an irreducible representation of $G$. If $h_i$ are coset representatives of $\Stab_H(1)$ in $H$, then
$1\rtimes h_i$ are coset representatives for $S_F$ in $G$.
Then $V$ has a basis $e_i:= (1\rtimes h_i) e_1$. The action of $G$ on $V$ is $(x\rtimes s) e_i = x_{s(i)} e_{s(i)}$. So on the one hand, it is permuting the $e_i$ from the permutation action of $s$, and on the other hand, it can change the sign for any particular basis independently.
Thus if $\sum c_i e_i$ is a non-zero element of $V$, we can act by some $g\in G$ to obtain $gx=e_1$, and thus we can generate all of $V$ from $G$ actions, which
shows that $V$ is irreducible.
So,
$$|\Hom_{S_F} (\Cl_{\tilde{K}/k},U )|=|\Hom_G(\Cl_{\tilde{K}/k} ,V)|=1+|\Sur_G(\Cl_{\tilde{K}/k} ,V)|.$$
From \cite[Theorem 6.2 and Theorem 4.1]{Wang2021}, we have that Cohen and Martinet predict the average of $|\Sur_G(\Cl_{\tilde{K}/k} ,V)|$ to be $\prod_{v}=|V^{\sigma_v}|^{-1}$,
where the product is over infinite places of the base field $k$, and $\sigma_v$ is a decomposition group for $v$. (Note that we can subdivide our family, which only has a fixed relative unit rank of $K/F$, based on their decomposition groups at each of the infinite places of $k$. We will see that the prediction for each of these subfamilies is the same, and only depends on the relative unit rank of $K/F$.) The action of $\sigma_v$ on the elements that $H$ acts on, or equivalently the cosets $h_iS_F$, has fixed points corresponding to the infinite places of
$F$ (over $v$) that are split over $k$ and $2$-cycles corresponding to infinite places of $F$ (over $v$) that are ramified over $k$. We can write $\sigma_v$ as generated by $x_v \rtimes r_v$.
If $h_i$ corresponds to a fixed point of $r_v$, then $\sigma_v \F_3 e_i=\F_3 e_i$, and the corresponding infinite place of $F$ is split in $K$ if and only if the $i$th coordinate of $x_v$ is trivial, which happens if and only if $\F_3 e_i$ is a trivial representation of $\sigma_v$.
If $h_i$ and $h_j$ correspond to a $2$-cycle of $r_v$, then $\F_3 e_i +\F_3 e_j$ is the regular representation of $\sigma_v$ and has one dimension of trivial representation.
Thus $\sum_v \dim_{\F_3} V^{\sigma_v}$ is the number of infinite places of $F$ that are not ramified in $K$, which is exactly the relative unit rank as defined above.
\end{proof}
\subsection{Cohen--Martinet prediction for the average of $h_3(K/k)$}\label{SS:wrongpred}
Using a similar approach, we can determine the Cohen--Martinet prediction for the average of $h_3(K/k)$ for $G$-extensions $K/k$ of a particular enriched or group signature.
In the case when $G=D_4$, we give below a numerical computation of the proven average for comparison.
Let $G$ be a transitive permutation $2$-group with a transposition and $S_K$ the stabilizer of an element. We consider the representation $\Ind_{S_K}^G \F_3$
of $G$ over $\F_3$ (which is the permutation representation of $G$).
Define $S_K$, $S_F$, $U,$ and $V$ as in the proof of Proposition~\ref{P:asCM}. Then $\Ind_{S_K}^G \F_3=\Ind_{S_F}^{G} \Ind_{S_K}^{S_F} \F_3=\Ind_{S_F}^{G} \F_3 \times V.$ Let $W=\Ind_{S_F}^{G} \F_3$, which is the permutation representation of $H$.
We see that $W$ contains no copies of the irreducible representation $V$, since the action $G$ on
$W$ factors through $H$ and the action of $G$ on $V$ does not.
Suppose that $W=\prod_i V_i^{a_i},$ where the $V_i$ are irreducible representations of $G$ (and note they all have $G$-action that factors through $H$).
We note that by Frobenius reciprocity $W$ contains exactly one copy of the trivial representation, and we let $W'$ be the quotient of of $W$ by this trivial representation.
We have
$h_3(K/k)=|\Hom (\Cl_{K/k},\F_3 )|$.
Let $e,m_e$ be
as in the proof of Proposition~\ref{P:asCM}.
Since any element of $\Hom_{S_K} (\Cl_{\tilde{K}},\F_3 )$ factors through $m_e \Cl_{\tilde{K}/k}=\Cl_{K/k}$, we have
a natural bijection between $\Hom (\Cl_{K/k},\F_3 )$ and $\Hom_{S_K} (\Cl_{\tilde{K}/k},\F_3 )|$.
Also $|\Hom_{S_K} (\Cl_{\tilde{K}/k},\F_3 )|=|\Hom_{G} (\Cl_{\tilde{K}/k},W \times V )|$.
If we let $e'=|G|^{-1}\sum_{g\in G} g$, we have that the map $m_e: \Cl_{\tilde{K}} \ra \Cl_k$ is a map whose kernel is exactly the relative class group
$\Cl_{\tilde{K}/k},$ and it is also the map that gives the maximal trivial representation quotient of any $G$ representation.
Thus $\Cl_{\tilde{K}/k}$ has no trivial representation part, and
$h_3(K/k)=|\Hom_{G} (\Cl_{\tilde{K}/k},W '\times V )|.$
From \cite[Theorem 6.2, Theorem 4.1]{Wang2021}, for a $G$-representation $Z$ with no trivial component, we have that Cohen and Martinet predict the average of $|\Sur_G(\Cl_{\tilde{K}} ,Z)|$ to be $\prod_{v}=|Z^{\sigma_v}|^{-1}$,
where the product is over infinite places $v$ of the base field $k$, in a family of $G$-extensions where $\sigma_v\in G$ is an element of the conjugacy class of complex conjugation over $v$.
Thus the predicted average of $h_3(K/k)$ is
\begin{align*}
&\left(1+\prod_{v} |V^{\sigma_v}|^{-1}\right) \prod_i \left(
1+\prod_{v} |V_i^{\sigma_v}|^{-1} +\cdots + \prod_{v} |V_i^{\sigma_v}|^{-a_i}
\right),
\end{align*}
where the left product is the predicted average of $h_3(K/F)$ and the right product is the predicted average of $h_3(F/k)$ (as can be worked out similarly to the above) for families with the corresponding behavior at infinite places.
Let $F/k$ be the index two subfield of $K$.
The action of $\sigma_v$ on the permutation basis elements of $W$ has fixed points corresponding to the split places of $F$ (over $v$) and $2$-cycles corresponding to the ramified places of $F$ (over $v$). Thus $\prod_v |W^{\sigma_v}|=3^{r_1(F)+r_2(F)}$ and
$\prod_v |(W')^{\sigma_v}|=3^{r_1(F)+r_2(F)-r_1(k)-r_2(k)}$.
If $W'$ is irreducible, this gives a nice formula for the predicted average of $h_3(K/k)$, which is $(1+3^{-u(K/F)} )(1+3^{-u(F/k)} )$,
where $u(K/F)$ is the difference in unit ranks of $K$ and $F$ in the given family (and similarly for $u(F/k)$) . However, when $W$ is not irreducible, the formula can be more complicated.
\begin{example}
Let $H=C_2$, so $G=D_4=\langle (1234), (24)\rangle$ and $S_K=\langle (24)\rangle$ and $S_F=\langle (13),(24)\rangle$.
Here $W'$ is irreducible, the sign representation of $H$, and thus the average conjecture by Cohen and Martinet is $(1+3^{-u(K/F)} )(1+3^{-u(F/k)} )$.
For example, if $k=\ensuremath{{\mathbb{Q}}}$, in the following table we give the averages of $h_3(K)$ predicted by the Cohen--Martinet heuristics, and a numerical computation of the proven averages from Theorem~\ref{thm:main}, in families with the given group signatures.
\begin{center}
\begin{tabular}{ c|c|c|c|c|c|c }
$\sigma_v $& $F$ & $r_1(K)$ & $|V^{\sigma_v}|$ & $|(W')^{\sigma_v}|$ & CM predicted & Thm~\ref{thm:main} proven\\
& & & && average of $h_3(K)$ & average of $h_3(K)$\\
\hline
() & real & 4 & 9 & 3&$(1+1/9)(1+1/3)=40/27\approx 1.48$&$\approx 1.12$\\
(24) & real & 2 & 3 & 3&$(1+1/3)(1+1/3)=16/9\approx 1.78$&$\approx 1.34$\\
(13)(24) & real & 0 & 1 &3&$(1+1)(1+1/3)=8/3\approx 2.67$&$\approx 2.01$\\
(12)(34) & imag. & 0 & 3 & 1&$(1+1/3)(1+1)=8/3\approx 2.67$&$\approx 1.41$\\
\end{tabular}
\end{center}
\end{example}
Essentially, in the proof of Theorem~\ref{thm:main}, we see that the averages of the $h_3(K/F)$ and $h_3(F/k)$ factors are independent, and the
$h_3(K/F)$ has average as predicted by Cohen and Martinet.
When we order fields in a family up to discriminant $X$, and take a uniform average and then let $X\ra\infty$, we will call this a \emph{discriminant-uniform} average.
A consequence of Cohen and Martinet's conjecture is that the discriminant-uniform average \emph{over $G$ extensions $K$} of $h_3(F_K/k)$ is the same
as the discriminant-uniform average \emph{over $H$ extensions $F$} of $h_3(F/k)$ when the $G$ and $H$ extensions have corresponding behavior at infinite places (see \cite[Theorem 9.2]{Wang2021}).
In contrast, what we see in the proof of Theorem~\ref{thm:main} is that
discriminant-uniform average over $G$ extensions $K$ of $h_3(F_K/k)$ is provably a \emph{weighted} average of $h_3(F/k)$ over $F$, against the measure on $H$-extensions
in which a field $F$ has measure proportional to $\Res_{s=1}\zeta_F(s)/(\zeta_F(2)\Disc(F)^2)$. There is no particular reason to think that this weighted average of $h_3(F/k)$ over $F$ will give the same result as the discriminant-uniform average over $F$. Indeed, the weighted average is heavily influenced by the $F$ of small discriminant.
When $H=C_2$, the discriminant uniform average
over $C_2$-extensions $F$
of $h_3(F/k)$ is known by Datskovsky--Wright, and is as predicted by Cohen--Martinet.
Since the quadratic fields of small discriminant have very little $3$-torsion in their class groups, one expects the weighted average of $h_3(F/k)$ to be smaller than the discriminant-uniform average over $F$, and indeed that is what we see in the chart above.
This perspective also explains the counterexample to the Cohen-Lenstra-Martinet heuristics of Bartel and Lenstra~\cite{Bartel2020}. Let $f$ be the indicator function of whether the $3$-torsion in the class group of a quadratic field is trivial. For a cyclic quartic field $K$, let $F_K$ be the quadratic subfield.
Bartel and Lenstra prove that the discriminant-uniform average over cyclic quartic $K$ of $f(F_K)$ is a weighted average of $f(F)$ over quadratic $F$. On the other hand, the Cohen-Lenstra-Martinet heuristics predict that the discriminant-uniform average over cyclic quartic $K$ of $f(F_K)$ is the discriminant-uniform average of $f(F)$ over quadratic $F$.
\section{Averages for other groups}
\label{sec:other-groups}
Our methods apply to more groups than just $2$-groups. We focus here on results that may be obtained purely from the results of Section \ref{sec:uniform-cubic}, and whose proofs in particular do not rely on arguments from Sections \ref{sec:arakelov} and \ref{sec:zero-density}. This permits a proof of the following general result.
\begin{theorem}\label{thm:other-groups}
Let $H \subseteq S_n$ be transitive and set $G = C_2 \wr H$. Then $G$-extensions $K$ of a number field $k$ have a unique index two subfield $F_K/k$. Moreover
$F_K$ is an $H$-extension of $k$. For $u \in \mathbb{Z}$, let $E_k^u(G,X) \subseteq E_k(G,X)$ be the subset of those $K$ for which $\rk \mathcal{O}_K^* - \rk \mathcal{O}_{F_K}^* = u$.
1) If $E_k^u(G,\infty)$ is non-empty and $|E_k(H,X)| \ll_{k,H,\epsilon} X^{2/3+\epsilon}$ for every $X\geq 1$, then
\[
\lim_{X \to \infty} \frac{1}{E_k^u(G,X)} \sum_{K \in E_k^u(G,X)} h_3(K/F_K)
= 1 + 3^{-u}.
\]
2) If $E_k(H,\infty)$ is non-empty and
\[
\sum_{F \in E_k(H,X)} h_3(F/k)
\ll_{k,H,\epsilon} X^{2/3+\epsilon}
\]
for every $X \geq 1$, then there is an explicit constant $c_{k,G,3}$ such that
\[
\lim_{X \to \infty} \frac{1}{E_k(G,X)} \sum_{K \in E_k(G,X)} h_3(K)
= c_{k,G,3}.
\]
\end{theorem}
Before we discuss the proof of Theorem \ref{thm:other-groups}, we note that the hypotheses of the first case are satisfied, for example, by any group $H \neq C_2$ in its regular representation that occurs as a Galois group over $k$ \cite[Proposition 1.3]{EV06}, and by any nilpotent group $H$ without a transposition \cite[Corollary 1.8]{Alb}. The hypotheses of the second case are satisfied for any $p$-group with $p \neq 5$ odd, as follows from \cite[Corollary 7.3]{KlunersMalle} and the trivial bound on $h_3(F/k)$ for $p \geq 7$, and from \cite{KluWan} for $p=3$.
\begin{proof}
The claims about the subfield $F_K$ follows as in Lemma \ref{lem:2-grp-with-transposition-is-wreath-product}.
Suppose now that $E_k^u(G,\infty)$ is non-empty. We mimic the proof of Theorem \ref{thm:main}, indicating the necessary modifications. Let $X \geq 1$. From Corollary \ref{cor:general-bound}, it follows for any $Y\geq 1$ that
\[
\sum_{\substack{ K \in E_k^u(G,X) \\ \Disc(F_K) \geq Y}} h_3(K/F_K)
\ll_{[k:\mathbb{Q}],G,\epsilon} \sum_{F \in E_k(H,X^{1/2}) \setminus E_k(H,Y)} \frac{h_2(F)^{2/3} X}{\Disc(F)^{1-\epsilon}}
\ll_{[k:\mathbb{Q}],G,\epsilon} \frac{X}{Y^{\frac{1}{3n[k:\mathbb{Q}]}-\epsilon}},
\]
where the second inequality is by partial summation and \cite[Theorem 1.1]{BSTTTZ}. This provides an analogue of Theorem \ref{thm:tail}.
We now establish a soft analogue of Theorem \ref{thm:3-torsion-thin-2-ext}, with the remainder of the proof then proceeding as in those of Theorems \ref{thm:main} and \ref{thm:relative-average}.
Note that if $G^\prime \subseteq G$ surjects onto $H$ and contains an element conjugate to $\sigma = (1,0,\dots,0) \rtimes 1_H$ in $G$, then in fact $G^\prime = G$.
For any fixed $H$-extension $F$ with discriminant at most $Y$, let $\mathcal{P}$ be a finite set of primes of $k$ that split completely in $F$.
If some prime of $F$ above a prime $\mathfrak{p} \in\mathcal{P}$ is inert in a quadratic extension $K/F$, while all the other primes of
$F$ above $\mathfrak{p}$ are split in $K$, then by the note above we have that $K/k$ is a $C_2\wr H$ extension.
If this happens, then we say that $\mathfrak{p}$ is bad for $K/F$.
It follows from the methods of Datskovsky and Wright \cite{DW86,DW88} (see also \cite[Theorem 2]{BSW15}) that
\[
\sum_{ \substack{ [K:F] = 2 \\ \text{group signature $\Sigma$} \\ \Disc(K/F) \leq X \\ \forall \mathfrak{p} \in \mathcal{P}, \mathfrak{p} \textrm{ not bad} }} h_3(K/F)
\sim X \prod_{\mathfrak{p} \in \mathcal{P}} \left(1 - \frac{n \Nm_{k/\mathbb{Q}} \mathfrak{p}^n}{2^n(\Nm_{k/\mathbb{Q}} \mathfrak{p} + 1)^n} \right) \cdot \frac{\mathrm{Res}_{s=1}\zeta_F(s)}{2^{r_1(F)+r_2(F)}\zeta_F(2)}(1+3^{-u(\Sigma)}).
\]
The expressions in the product above are uniformly (in $\mathfrak{p}$) bounded away from $1$, and in particular the product is $O(\beta_n^{\#\mathcal{P}})$ for some constant $\beta_n <1$ depending at most on $n$. Since there are infinitely many primes that split completely in $F$, this product may be made arbitrarily small, and thus
\[
\sum_{ \substack{ [K:F] = 2 \\ \Disc(K/F) \leq X \\ \mathrm{Gal}(K/k) \not\simeq G } } h_3(K/F)
= o_F(X).
\]
Proceeding now as in the proof of Theorem \ref{thm:main}, the first claim follows. The second follows analogously.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction} \label{Intro}
The standard approach to model the Universe starts with two
basic assumptions. First, Weyl's principle~\cite{Weyl:1923yve,weyl2009republication} is postulated, which entails\footnote{We adopt here a formulation of Weyl's principle in which it is assumed
that world lines of \textit{fundamental particles} (galaxies, galaxy clusters)
form, on average, a congruence of non-intersecting diverging geodesics
emerging from the distant past and orthogonal to
space-like hypersurfaces $M_3$.
In this form the principle permits a comoving frame relative to which the
constituents of the universe are at rest
on average~\cite{Robertson:1933zz,Narlikar:2002wxx,Rugh:2010mv}.}
the existence of a cosmic
time $t$. Second, it is assumed that our
$3-$dimensional space is homogeneous and isotropic (cosmological principle).
The most general spacetime geometry that embodies these assumptions is
the Friedmann--Robertson--Walker (FRW) metric
\begin{equation} \label{RWmetric}
ds^2 = c^2 dt^2 - a^2 (t) \left[\, \frac{dr^2}{1-kr^2} +
r^2(d\theta^2 + \sin^2 \theta d\phi^2) \,\right]\,,
\end{equation}
where $c$ is the speed of light, $a(t)$ is the scale factor, and
the spatial curvature is specified by the constant $k$, which takes the
values $k=0,\pm 1$ for Euclidean, spherical and hyperbolic geometries,
respectively.
The metric~(\ref{RWmetric}) expresses locally the two above basic assumptions.
It does not specify the topology of the spacetime manifold $\mathcal{M}_4$
or of the corresponding spatial ($t= \mbox{const}$) section $M_3$.
However, the FRW metric~\eqref{RWmetric} is consistent with the global decomposition
$\mathcal{M}_4 = \mathbb{R} \times M_3$, which we assume in this work.
Regarding the spatial geometry, recent high precision cosmic microwave background
radiation (CMB) data from the Planck satellite~\cite{2016,2020}
have provided strong evidence that the universe
is very nearly flat with curvature parameter $|\Omega_k| < 0.003$,
which is compatible with the standard inflationary predictions
that the spatial curvature should be very small today.
These indications
support the assumption we make in this work that the spatial section $M_3$
is flat ($k=0$).
As to the topology of the spatial sections, we first note that the FRW
geometry~\eqref{RWmetric} constrains but does not specify the topology of $M_3$.
In this way, no classical geometric theory as, for example, general relativity can be
used to derive the $M_3$ topology.
However, for $k=0$, it is a mathematical fact that, in addition to the simply-connected
Euclidean space $\mathbb{E}^{3}$, there are $17$ topologically inequivalent quotient flat
manifolds with nontrivial topology~\cite{wolf1967spaces,Thurston+2014}.
Given this set of topological possibilities for $M_3$
and despite our present-day inability to infer the topology from a
fundamental theory, as for example quantum gravity~\cite{1987IAUS..124..461F},
to disclose the spatial topology of FRW spacetime we must rely ultimately on
cosmological observations (see the review articles~%
\cite{Ellis:1970ey,LACHIEZEREY1995135,Starkman:1998qx,LEVIN2002251,%
Reboucas:2004dv,Luminet:2016bqv}) or on local physical
experiments.%
\footnote{
For recent constraints on cosmic topology from CMBR data we refer the readers to
Refs.~\cite{2014,2016,Cornish:2003db,ShapiroKey:2006hm, Bielewicz:2010bh,Vaudrevange:%
2012da,Aurich:2013fwa}
For some limits on the circles-in-the-sky method designed for the searches of spatial
topology through CMBR see Ref.~\cite{Gomero:2016lzd}}
Topological properties precede the geometrical features of a manifold.
Thus, it is important to find out whether, how and to what extent physical
results depend on a nontrivial topology.
Nonstandard choices of the background spatial topology
affect the mean squared velocity of charged test particles
under quantum vacuum fluctuations of the electromagnetic field.
In fact, on the assumption that the net role played by the spatial topology
is more clearly ascertained in the static FRW flat spacetime, the question of
how a nontrivial topology of the spatial section of Minkowski spacetime
modifies the stochastic motions of a test charged particle and a point
electric dipole under quantum vacuum
fluctuations of the electromagnetic field was studied in recent
papers~\cite{Bessa:2019aar,Lemos:2020ogj}.
By the way, we mention that the case of a point particle coupled to
a massless field living in a topologically nontrivial space
was considered in Ref.~\cite{Matas:2015rba}.
Orientability is an important topological property of spacetime manifolds.
It is generally assumed that, being a global property, the orientability of
$3-$space cannot be tested locally. So, a test for spatial orientability
would require a global trip along some specific closed paths around the
whole $3-$space to check whether one returns with left- and right-hand
sides exchanged.
Since such a global expedition does not seem to be feasible at the
cosmological scale, theoretical arguments that combine universality of
physical experiments with local arrow of time, CP violation and CPT invariance
are usually invoked~\cite{1967JETPL...6..236Z,Hawking:1973uf,%
penrose1984spinors,1979grec.conf..212G} to support the choosing of time-
and space-orientable manifolds, although there are dissenting stances%
~\cite{Hadley:2002fr,HadleyM2018}.
The impossibility of having \textsl{globally} defined spinor fields
on non-orientable spacetime manifolds \cite{Geroch:1968zm,Geroch:1970uv}.
is another theoretical argument to support the choice of space-and-time
orientable manifolds.%
\footnote{One can certainly take advantage of theoretical arguments of this sort to
support such underlying assumptions, but not as a replacement to experimental and
observational evidence in physics.}
Since $8$ out of the $17$ possible flat $3-$manifolds, $M_3$, with
nontrivial topology are non-orientable~\cite{wolf1967spaces}, the question as to
whether velocity fluctuations could be also employed to locally reveal specific
topological properties such as orientability was examined in
Ref.~\cite{Lemos:2020ogj}.
It was shown that it is possible to \textsl{locally} access the spatial
orientability of Minkowski spacetime through the study of the stochastic motions
of a charged particle and a point electric dipole subject to these fluctuations
in Minkowski spacetime with orientable and non-orientable spatial topologies.
It was found that a characteristic inversion pattern exhibited by certain
statistical orientability indicator curves, constructed from the mean
square velocity of an electric dipole, can be used as a local physical
signature of non-orientability of the spatial section $M_3$ of Minkowski
spacetime.
Thus, a question that naturally arises is how these results are
modified in an expanding FRW universe whose curvature parameter is
within the bounds determined by Planck data~\cite{2014,2016},
which indicate that a flat geometry is a good approximation to model the
spatial section of the Universe in the framework of general relativity.
To tackle this question,
in this paper we study the stochastic motions of a charged particle and a
point electric dipole under quantum vacuum electromagnetic fluctuations
in a spatially-flat FRW geometry with spatial sections endowed with
an orientable and its counterpart non-orientable spatial topologies.
In so doing we extend the results of~\cite{Bessa:2019aar,Lemos:2020ogj}
from the static Minkowski spacetime to a dynamical FRW spacetime. Our result for the orientability
indicator is very general, in the sense that it does not depend on the underlying gravitational theory,
because it is obtained from first topological-geometrical principles alone.
The structure of the paper is as follows.
In Sections~\ref{TopSet} and~\ref{Syst-montion} we describe the topological
and dynamical settings, respectively. In Sections~\ref{disp-particleE17}
and~\ref{Dipole-motion} we derive
statistical
orientability indicators both for the charged particle and the point
electric dipole that are independent of any specific metrical
theory of gravity.
To concretely study the time evolution of the orientability indicators,
in Section~\ref{casestudy} we choose general relativity and a barotropic
perfect fluid as the matter content.
In the case of a charged particle, we show that it is possible to distinguish
the orientable from the non-orientable topology by comparing the time
evolution of an orientability indicator, defined from the stochastic motion
of the particle, in the orientable and non-orientable topologies.
We then turn to the more substantial problem of finding a way to decide about
the orientability of a given spatial manifold in itself, without having to
compare the results for a non-orientable space with those for its orientable
counterpart.
Motivated by a dipole's directional properties, we inquire whether the
stochastic motions of a point electric dipole would be more effective
to unveil the presumed non-orientability of a $3-$space in itself.
From the orientability indicators computed for the dipole we identify
a characteristic inversion pattern displayed by the orientability
indicator curves for the non-orientable topology,
implying that the putative non-orientability can be detected
per se.
In Section~\ref{Finals} we present our final remarks and summarize
our findings, which indicate that it may be possible to locally disclose
a conceivable spatial non-orientability of FRW spacetime
through the stochastic motions of point-like ``charged'' objects
under quantum vacuum fluctuations of the electromagnetic field.
\section{Topological fundamentals } \label{TopSet}
To make this work to a certain extent self-contained, in this section
we define the notation, give some basic definitions and present a few
results concerning the topology of flat three-dimensional manifolds
which are used in this paper.
We begin by recalling that in the standard cosmological model
the spacetime is a manifold $\mathcal{M}_4$ locally endowed
with a FRW metric~\eqref{RWmetric} and globally decomposable
as $\mathcal{M}_4 = \mathbb{R}\times M_3$.
Although the spatial section $M_3$, whose geometry we assume to be Euclidean,
is usually taken to be the simply-connected Euclidean space $\mathbb{E}^{3}$,
it can likewise be one of the possible $17$ topologically inequivalent
multiply-connected quotient manifolds $\mathbb{E}^3/\Gamma$ where $\Gamma$
is a discrete group of
isometries or holonomies acting freely on the covering manifold
$\mathbb{E}^{3}$~\cite{wolf1967spaces,Thurston+2014}.
The quotient manifolds are compact in at least one direction.
The action of $\Gamma$ tiles the noncompact covering space $\mathbb{E}^{3}$
into infinitely many identical copies of the fundamental domain (FD) or cell (FC).
Thus, the multiple connectedness of the quotient manifold gives rise to periodic
boundary conditions (repeated domains or cells) on the covering manifold
$\mathbb{E}^{3}$ that are determined by the action of the group $\Gamma$
on $\mathbb{E}^{3}$.
An example of flat quotient manifold is the so-called slab space,
denoted in the literature by $E_{16}$, which is open (noncompact) in two
independent directions and decomposed into
$E_{16} = \mathbb{R}^2 \times \,\mathbb{S}^1 =\mathbb{E}^3/\Gamma$,
where $\mathbb{R}^2$ and $\mathbb{S}^1$ stand for the real plane and
the circle, respectively.
A fundamental domain is
a slab with a pair of opposite faces (two infinite parallel planes)
identified through translations.
The simply-connected covering space $\mathbb{E}^3$ is tiled with
these equidistant parallel planes, which together with two
noncompact independent spatial directions form the FD of $E_{16}$.
The periodicity in the compact direction is
given by the circle $\mathbb{S}^1$, whereas the noncompact
independent directions form $\mathbb{R}^2$.
In forming the quotient manifolds $M_3$ an essential point is that
they are obtained from the covering manifold $\mathbb{E}^3$ through identification
of points that are equivalent under the action of the group $\Gamma$. In this way,
each point in the quotient manifold $M_3$ represents all the equivalent points in
the covering space. Thus, for example, for $E_{16}$ quotient space,
taking the $x$-direction as compact, one has that, for
$n_x \in \mathbb{Z}$ and compact length $L>0$, points $(x,y,z)$ and
$(x+n_xL, y, z)$ are identified. In terms of the covering isometry
$\gamma \in \Gamma$ one has
\begin{equation}
\label{ident-E16}
P = (x,y,z) \,\mapsto \, P^{\prime} = \gamma P = (x + n_x L,\, y, z )\,.
\end{equation}
Another example that we shall be concerned with in this paper is
the slab space with flip $E_{17}$, which involves an additional
inversion in a direction orthogonal to the compact direction, that is,
one direction in the tiling planes is flipped as one moves from
one plane to the next. Taking the $x$-direction as compact
and letting the flip be in the $y$-direction, in the covering space
$\mathbb{E}^3$ one has
\begin{equation}\label{ident-E17}
P = (x,y,z) \; \mapsto \; P^{\prime} = \gamma P = (x+n_xL,\,(-1)^{n_x}y,\,z),
\end{equation}
and the identification $P \equiv P^{\prime}$ defines the $E_{17}$ topology.
It should be noted that for each of the $17$ quotient manifolds,
$\mathbb{E}^3/\Gamma$, the associated periodic conditions on the covering
space $\mathbb{E}^3$ are determined by the group $\Gamma$, and clearly different
discrete isometry groups $\Gamma$ define different topologies for $M_3$, which in
turn give rise to different periodicities and associated tiling of the
covering space $\mathbb{E}^3$.
Unlike the local geometric concept of homogeneity, which is formulated in
terms of the action of the local group of isometries, in topological spaces
we have the concept of \textsl{global} or \textsl{topological homogeneity}.
A way to describe global topological homogeneity of the
quotient manifolds is through distance functions. In fact, for any
$\mathbf{x} \in M_3$ the distance function $\ell_\gamma (\mathbf{x})$
for a given isometry $\gamma \in \Gamma$ is def\/ined by
\begin{equation}
\label{dist-function}
\ell_\gamma(\mathbf{x}) = d(\mathbf{x}, \gamma \mathbf{x}) \; ,
\end{equation}
where $d$ is the Euclidean metric. The distance function
gives the length of the closed geodesic that passes through $\mathbf{x}$
and is associated with a holonomy $\gamma \in \Gamma$.
For a \textsl{globally homogeneous} manifold, endowed with a topology defined
by a group $\Gamma$, the distance function for any covering
isometry $\gamma \in \Gamma$ is constant. In globally inhomogeneous manifolds, i.e.
manifolds with inhomogeneous topologies, in contrast,
the length of the closed geodesic associated with at least one $\gamma$
is non-translational (screw motion or flip, for example) and the corresponding
distance depends on the point $\mathbf{x} \in M_3$, and then is not constant.
In this way, the slab space $E_{16}$ is globally homogeneous since all
$\gamma$s are translations, whereas the slab space with flip, $E_{17}$,
is globally inhomogeneous since the covering group $\Gamma$ contains a
flip, which clearly is a non-translational holonomy.
Another very important global (topological) property of manifolds that
we shall deal with in this paper is \textsl{orientability}, which measures whether
one can consistently choose a definite orientation for loops in a manifold.
An orientation-reversing path in a manifold $M_3$ is a path that brings a
traveler back to the starting point mirror-reversed. Manifolds that contain
an orientation-reversing path are \textsl{non-orientable}, whereas those
that do not have any such reversing path are called
\textsl{orientable}~\cite{weeks2020shape}.
In two dimensions, one has planes, cylinders and two-tori as examples
of orientable surfaces, whereas the M\"obius strip and Klein bottle are
non-orientable surfaces.
For three-dimensional quotient manifolds, when the covering group $\Gamma$
contains at least one holonomy $\gamma$ that is a reflection (flip) the
associated quotient manifold is non-orientable.
In this way, the slab space, $E_{16}$, is orientable while the slab space with
flip, $E_{17}$, is non-orientable. Clearly non-orientable manifolds are necessarily
topologically inhomogeneous as the covering group $\Gamma$ contains
a reflexion, which is a non-translational covering holonomy.
In Table~\ref{Tb-2-Orient_and_Non_orient} we collect the names and symbols
used to refer to the manifolds together with the number of compact
independent dimensions and information concerning their orientability
and global homogeneity.
In the next section we shall study the motions of a charged test particle
and a point electric dipole under quantum vacuum fluctuations of the
electromagnetic field in the expanding FRW spacetime whose spatial
sections are the manifolds given in Table~\ref{Tb-2-Orient_and_Non_orient}.
\begin{table}[h!]
\begin{tabular}{lcccc}
\hline\hline
Name & Symbol & $\quad$ Compact Dim.\ & $\quad$ Orientable &$\quad$ Homogeneous \\
\hline
Slab space & $E_{16}$ & $1$ & yes & yes \\
Slab space with flip & $E_{17}$ & $1$ & no & no \\
\hline
\end{tabular}
\caption{Names and symbols of two Euclidean orientable and
non-orientable quotient manifolds $M_3 =\mathbb{E}^3/\Gamma$ together with the number of compact
dimensions (Comp.), orientability and global (topological) homogeneity.}
\label{Tb-2-Orient_and_Non_orient}
\end{table}
Finally, we briefly mention a few results that are
used throughout this paper (for a detailed discussion we refer the reader
to Ref.~\cite{1979grec.conf..212G}).
All simply-connected spacetime manifolds are both time- and space-orientable.
The product of two manifolds is simply-connected if and only if
the factors are.
If the spacetime is of the form $\mathcal{M}_4 = \mathbb{R}\times M_3$ then
space-orientability of the spacetime reduces to orientability of the $3-$space
$M_3$. This applies to the spacetime endowed both with the $E_{16}$ (orientable)
and the $E_{17}$ (non-orientable) topology that we deal with in this paper.
\section{Non-orientability from electromagnetic fluctuations}
\label{Syst-montion}
As shown in \cite{Bessa:2019aar,Lemos:2020ogj}, nontrivial spatial
topologies influence the stochastic motions both of a charged particle and a point
electric dipole in the presence of quantum vacuum fluctuations of the electromagnetic
field in Minkowski spacetime.
Here we investigate how these results are modified if instead of static Minkowski
spacetime an expanding FRW flat universe is the background geometry for the motions of
the charged particle and the dipole. To this end we consider a spatially flat FRW
spacetime endowed with two inequivalent spatial topologies given in
Table \ref{Tb-2-Orient_and_Non_orient}, namely the orientable slab space ($E_{16}$)
and the non-orientable slab space with flip ($E_{17})$.
\subsection{The point charge case}
\label{Subsec-charge}
We first consider a nonrelativistic test particle with
charge $q$ and mass $m$ locally subjected to
vacuum fluctuations of the electric field ${\bf E}({\bf x}, t)$
in the topologically nontrivial spacetime
manifold
equipped with the spatially flat Friedmann-Robertson Walker (FRW) metric
\begin{equation}
\label{metric}
ds^2 = dt^2 -a^2(t)(dx^2 + dy^2 + dz^2)\, ,
\end{equation}
which is the particular case of equation (\ref{RWmetric}) with $k=0$ and Cartesian
instead of spherical coordinates. The covariant equation of motion
is~\cite{Weinberg:1972kfs}
\begin{equation}\label{eqmotioncov}
\frac{Du^{\mu}}{d\tau} \overset{\text{\scriptsize def}}{=} \frac{du^{\mu}}{d\tau}
+ \Gamma^{\mu}_{\alpha \beta}u^{\alpha} u^{\beta} = \frac{f^{\mu}}{m},
\end{equation}
where $u^{\mu}=dx^{\mu}/d\tau$ is the particle's four-velocity, $m$ is its mass,
$\tau$ is its proper time and $f^{\mu}$ is the nongravitational four-force acting on it.
Since we are interested in the motion of a charged particle in an electromagnetic field,
the four-force is $f^{\mu} = q F^{\mu \nu}u_{\nu}$ where $F^{\mu\nu}$ is the
electromagnetic field tensor.
In the nonrelativistic case, in which the particle's proper time is indistinguishable
from the cosmic time $t$, the equation of motion for the point charge
becomes~\cite{Bessa:2008pr}
\begin{equation}\label{eqmotion}
\frac{d{\bf u}}{dt} + 2\frac{\dot a}{a}{\bf u} = \frac{q}{m} \,{\bf E}({\bf x}, t)\,,
\end{equation}
which can be written as
\begin{equation}\label{eqmotion-simpl}
\frac{1}{a^2}\frac{d}{dt}( a^2{\bf u}) =\frac{q}{m} \,{\bf E}({\bf x}, t)\, ,
\end{equation}
where ${\bf E} =(E^1,E^2,E^3)$ with $E^i=F^{i0}$.
After integration this yields
\begin{equation}\label{eqmotion-integr}
a^2(t){\bf u}({\bf x},t) = \frac{q}{m}
\int_{t_i}^t a^2(t^{\prime}){\bf E}({\bf x}, t^{\prime})dt^{\prime}\,.
\end{equation}
where we have assumed that the particle is initially at rest: ${\bf u}({\bf x},t_i)=0$.
Since proper (physical) distances $d$ at time $t$ are related to coordinate distances
$r$ by $d=a(t) r$, the proper (physical) velocity $\bf v$ is related to the coordinate
velocity $\bf u$ at time $t$ by
\begin{equation}\label{physveloc}
{\bf v}({\bf x},t) = a(t){\bf u}({\bf x},t)\,.
\end{equation}
Therefore, in terms of the physical velocity, Eq.~(\ref{eqmotion-integr}) becomes
\begin{equation}\label{eqmotion-integr-phys-veloc}
{\bf v}({\bf x},t) = \frac{q}{m}\frac{1}{a(t)}
\int_{t_i}^t a^2(t^{\prime}){\bf E}({\bf x}, t^{\prime})dt^{\prime}\,,
\end{equation}
from which the one can write the dispersion of each velocity component as%
\footnote{By definition, $\,\langle \Delta v^i(\mathbf{x}, t)^2 \rangle =
\langle v^i(\mathbf{x}, t)^2 \rangle -\langle v^i(\mathbf{x}, t) \rangle^2$.}
\begin{equation}\label{phys-veloc-dispersion}
\Bigl\langle \Delta v^i({\bf x},t)^2 \Bigr\rangle =\frac{q^2}{m^2 a^2(t)}
\int_{t_i}^t\int_{t_i}^t a^2(t^{\prime})\,a^2(t^{\prime\prime}) \, \Bigl\langle
E^i({\bf x}, t^{\prime})E^i({\bf x},
t^{\prime\prime}) \Bigr\rangle_{FRW} \, dt^{\prime}dt^{\prime\prime}\,,
\end{equation}
where $i=1,2,3$ for the corresponding directions $x,y,z$
\footnote{For any three-vector $\bf b$ we write either ${\bf b}=(b^1,b^2,b^3)$
or ${\bf b} = (b_x,b_y,b_z)$ according to convenience.}
As in Ref.~\cite{Yu:2004tu,Lemos:2020ogj},
here we assume that $\bf x$ is constant, meaning that in the time scales of interest
the particle's position essentially does not change.
\subsection{Velocity dispersion in terms of conformal time}
\label{Conformaltime}
The Friedmann-Robertson-Walker correlation function that appears in Eq.~(\ref{phys-veloc-dispersion})
can be more easily computed in terms of the Minkowski spacetime correlation function by making use
of how the electromanetic field changes under a conformal transformation.
Indeed, in terms of the conformal
time $\eta$ defined by $dt = a(t) d\eta$ the FRW metric becomes
\begin{equation}\label{metric-conformal}
ds^2 = a^2(d\eta^2 - dx^2 - dy^2 - dz^2)\,.
\end{equation}
In these coordinates the electromagnetic field tensor in FRW spacetime is related
to the one in Minkowski spacetime $ M$ by~\cite{Wald:1984rg,Bessa:2008pr}
\begin{equation}\label{electromg-conformal-transf}
F^{\mu \nu}({\bf x}, \eta )_{F\!RW} = a^{-4} F^{\mu \nu}({\bf x}, \eta )_{M}\,.
\end{equation}
Taking this equation into account, noting further that the coordinate change $t \to \eta$
implies
\begin{equation}\label{F01-transf}
F^{i0}({\bf x}, \eta )_{F\!RW} = a^{-1} F^{i0}({\bf x}, t)_{F\!RW}\, ,
\end{equation}
and changing the integration variable to $\eta$ using $dt = a\, d\eta$,
Eq.~(\ref{phys-veloc-dispersion}) reduces to
\begin{equation}\label{veloc-dispersion-eta}
\Bigl\langle \Delta v^i({\bf x},t)^2 \Bigr\rangle =\frac{q^2}{m^2 a^2(t)}
\int_{\eta_i}^{\eta}
\int_{\eta_i}^{\eta} \Bigl\langle E^i({\bf x}, \eta^{\prime})E^i({\bf x}, \eta^{\prime\prime})
\Bigr\rangle_M \, d\eta^{\prime}d\eta^{\prime\prime}\,.
\end{equation}
Therefore, in order to compute the velocity dispersion all one needs is the correlation
function in Minkowski spacetime with the time coordinate replaced by $\eta$.
The result can be expressed in terms of the cosmic time $t$ as long as the scale factor
is known as a function of $t$ or $\eta$.
Since the correlation function in equation~\eqref{veloc-dispersion-eta} is in Min-kowski
spacetime $M$, it depends on the topology of the spatial section as discussed in
\cite{Bessa:2019aar,Lemos:2020ogj}. Then the above result for
the velocity dispersion~~\eqref{veloc-dispersion-eta} in FRW spacetime depends on
the topology of the spatial sections $M_3$ of the FRW spacetime, whose set of
possible nonequivalent topologies is identical to the corresponding set for
the spatial section $M_3$ of Minkowski spacetime $M$.
In the next section we shall use this result to derive the velocity dispersion,
and thus explicit expressions for a velocity dispersion orientability indicator
for manifolds endowed with non-orientable topology $E_{17}$, and
for its orientable counterpart $E_{16}$.
\section{Orientability indicators for $\mathbf{E_{17}}$ and $\mathbf{E_{16}}$
spatial topologies} \label{disp-particleE17}
In this section, we compute the orientability indicators
for a FRW spacetime with flat spatial sections equipped with the non-orientable
$E_{17}$ topology.
The corresponding results for spatial sections endowed
with $E_{16}$ topology follow from those for $E_{17}$ with no
need of additional calculations.
Following Yu and Ford~\cite{Yu:2004tu}, we assume that the electric field
$\mathbf{E}$ is a sum of classical $\mathbf{E}_c$ and quantum $\mathbf{E}_q$
parts. Since there are no quantum fluctuations of ${\bf E}_c$ and
$\langle {\bf E}_q\rangle =0$, the two-point function
$\bigl\langle E^i({\bf x}, \eta^{\prime})E^i({\bf x}, \eta^{\prime\prime})
\bigr\rangle_M$ in equation~\eqref{veloc-dispersion-eta} involves only
the quantum part of the electric field~\cite{Yu:2004tu}.
It can be shown \cite{Birrell:1982ix} that locally
\begin{equation}\label{eqdif-00}
\Bigl \langle E^i({\bf x}, \eta)E^i({\bf x}', \eta')\Bigl \rangle =
\frac{\partial }{\partial x_i}
\frac{\partial}
{\partial {x'}_i}D({\bf x}, \eta ; {\bf x}', \eta') - \frac{\partial }{\partial \eta }
\frac{\partial}
{\partial \eta'}D ({\bf x }, \eta; {\bf x'}, \eta')
\end{equation}
The topology of the spatial section $M _3$ is (globally) taken into account as follows.
When $M_3$ is simply-connected
the Hadamard function $D({\bf x}, \eta;{\bf x}', \eta')$ is given by
\begin{equation}\label{eqren}
D_0({\bf x}, \eta; {\bf x}', \eta') = \frac{1}{4\pi^2(\Delta \eta^2 -
|\Delta \mathbf{x}|^2)} \,.
\end{equation}
The subscript $0$ indicates standard Minkowski spacetime $M$,
$\Delta \eta = \eta - \eta'$ and $|\Delta \mathbf{x}| \equiv r $
is the spatial separation for topologically trivial Minkowski
spacetime:
\begin{equation}\label{separation-trivial}
r^2 = (x-x')^2 + (y-y')^2 + (z - z')^2 \,.
\end{equation}
However, when Minkowski spacetime in endowed with a topologically nontrivial
spatial section, the {spatial separation} $r^2$ takes a different form that
captures the periodic boundary conditions imposed on the covering space
$\mathbb{E}^{3}$ by the covering group $\Gamma$, which characterize the
spatial topology. In Table~\ref{Tb-Spatial-separation}
we collect the {spatial separations} for the topologically non-homeomorphic
Euclidean spaces we shall address in this paper.%
\footnote{The reader is referred to Refs.~\cite{LACHIEZEREY1995135,%
Riazuelo:2003ud,Fujii:2011ga} for pictures of the fundamental cells and
further properties of all possible three-dimensional Euclidean topologies.}
\begin{table}[h!]
\begin{tabular}{*2{l|c}}
\hline\hline
Spatial topology & $\quad$ Spatial separation $r^2$ for Hadamard function \\
\hline
$E_{16}$ - Slab space & $(x - x'- n_x\, L)^2 + (y - y')^2 + (z - z')^2 $ \\
\hline
$E_{17}$ - Slab space with flip &
\hspace{-.25cm}$\left(x-x'-n_x \,L \right)^{2}+\left(y -(-1)^{n_x} y'\right)^{2}
+ \left(z - z' \right)^{2}$
\\ \hline
\end{tabular}
\caption{Spatial separation in Hadamard function for the multiply-connected flat
orientable ($E_{16}$) and its non-orientable counterpart ($E_{17}$) quotient Euclidean
manifolds. The topological compact length is denoted by $L$.
The numbers $n_x$ are integers and run from $-\infty$ to $\infty$.
For each multiply-connected topology, when $n_x=0$ we recover the
spatial separation for the simply-connected Euclidean $3-$space.} %
\label{Tb-Spatial-separation}
\end{table}
To obtain the correlation function for the electric field
that is required to compute the velocity dispersion~\eqref{veloc-dispersion-eta} for
slab space with flip $E_{17}$, we replace in Eq.~\eqref{eqdif-00} the Hadamard
function $D({\bf x}, \eta; {\bf x}', \eta')$ by its renormalized version
given by~\cite{Bessa:2019aar}
\begin{equation}\label{Hadamard-ren}
D_{ren}({\bf x}, \eta; {\bf x}', \eta') =
D({\bf x}, \eta; {\bf x}', \eta') - D_0({\bf x}, \eta; {\bf x}', \eta')
=\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}\frac{1}{4\pi^2(\Delta \eta^2 - r^2)}\,,
\end{equation}
where here and in what follows $\sum_{}^{\;'}$ indicates that
the Minkowski contribution term $n_x = 0$ is excluded from the summation,
$\Delta \eta = \eta -\eta^{\prime}$, and the spatial separation $r$ for $E_{17}$
is given in Table~\ref{Tb-Spatial-separation}.
The term with $n_x=0$ that would be present in the sum~\eqref{Hadamard-ren} is the Hadamard function
$D_{0}({\bf x}, \eta; {\bf x}', \eta')$ for Minkowski spacetime with simply-connected
spatial section $\mathbb{E}^3$.
This term has been subtracted out from the sum
because it gives rise to an infinite contribution
to the velocity dispersion~\cite{Bessa:2019aar,Lemos:2020ogj}.
Thus, from equation (\ref{eqdif-00}) the renormalized correlation
functions
\begin{equation}\label{correlation-i-E17}
\Bigl \langle E^i({\bf x}, \eta)E^i({\bf x}', \eta')\Bigr \rangle_{ren}
=\frac{\partial }{\partial x_i} \frac{\partial}
{\partial {x'}_i}D_{ren}({\bf x}, \eta; {\bf x}', \eta') - \frac{\partial }{\partial \eta}
\frac{\partial}
{\partial \eta'}D_{ren} ({\bf x }, \eta; {\bf x'}, \eta')\,,
\end{equation}
where $D_{ren}({\bf x}, \eta; {\bf x}', \eta')$ depends on the
spatial topology through $r$ according to (\ref{Hadamard-ren}) and
Table~\ref{Tb-Spatial-separation}.
{}From equations (\ref{correlation-i-E17}), (\ref{Hadamard-ren}) with $r$ given
in Table~\ref{Tb-Spatial-separation} the electric field correlation functions
for $E_{17}$ topology are found to be given by
\begin{eqnarray}\label{correlation-x-E17}
&&\Bigl\langle E_x({\bf x}, \eta )E_x({\bf x}', \eta')\Bigl\rangle_{ren}^{E_{17}} =
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}
\frac{\Delta \eta^2 + r^2 -2r_x^2}{\pi^2 [\Delta \eta^2 - r^2]^3},
\\ \label{correlation-y-E17}
&&\Bigl\langle E_y({\bf x}, \eta )E_y({\bf x}', \eta')\Bigl\rangle_{ren}^{E_{17}} =
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}\bigg[
\frac{\left( 3-(-1)^{n_x}\right) \Delta \eta^2}{2\pi^2 [\Delta \eta^2 - r^2]^3}
+\frac{\left( 1+ (-1)^{n_x}\right) r^2
- 4(-1)^{n_x} r_y^2} {2\pi^2 [\Delta \eta^2 - r^2]^3}\bigg],
\\ \label{correlation-z-E17}
&&\Bigl\langle E_z({\bf x}, \eta )E_z({\bf x}', \eta' )\Bigl\rangle_{ren}^{E_{17}} =
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}
\frac{\Delta \eta^2 + r^2 -2r_z^2}{\pi^2 [\Delta \eta^2 - r^2]^3},
\end{eqnarray}
where $\Delta \eta = \eta -\eta^{\prime}$ and
\begin{equation}\label{r-components-E17}
r_x = x-x^{\prime}-n_xL,\;\;\;\; r_y = y-(-1)^{n_x}\, y^{\prime},\;\;\;\;
r_z = z-z^{\prime}, \;\;\;\;r= \sqrt{r_x^2 + r_y^2+ r_z^2}\,.
\end{equation}
The orientability indicator $\mbox{\large $I$}_{v^2_i}^{E_{17}}$ that we will
consider here is defined by replacing
the electric field correlation functions in Eq.~\eqref{veloc-dispersion-eta} by
their renormalized counterparts~(\ref{correlation-i-E17}) in which $r$ is given
in Table~\ref{Tb-Spatial-separation}, namely
\begin{equation} \label{indicator-E17}
\mbox{\large $I$}_{v^2_i}^{E_{17}} ({\bf x},t) = \frac{q^2}{m^2a^2(t)}
\int_0^{\eta}\int_0^{\eta}\Bigl\langle E^i({\bf x}, \eta') E^i({\bf x}, \eta'')\Bigr \rangle_{ren}^{E_{17}}\, d\eta' d\eta''\,.
\end{equation}
Equation~(\ref{Hadamard-ren}) makes it is clear that the orientability indicator
$\mbox{\large $I$}_{v^2_i}^{E_{17}}$ is the difference between the velocity
dispersion in $E_{17}$ and the one in Minkowski with trivial (simply connected) topology
\subsection{Orientability indicator: general definition}
\label{Indicatorgeneral}
For later use and for the sake of generality, some words of clarification are
in order at this point before proceeding to the
calculation of the components of the statistical indicator.
From equations~~\eqref{veloc-dispersion-eta} and~\eqref{Hadamard-ren}
a definition of the orientability indicator for a general multiply-connected
flat topology can be written in
the form \cite{Lemos:2020ogj}
\begin{equation} \label{new-ind}
\mbox{\large $I$}_{v^2_i}^{MC}
= \Bigl\langle\Delta v_i^2 \Bigr \rangle^{MC}
- \;\,\Bigl \langle\Delta v_i^2 \Bigr \rangle^{SC} ,
\end{equation}
where $\bigl\langle\Delta v_i^2 \bigr \rangle$ is the
mean square velocity dispersion, and the superscripts $MC$ and $SC$
stand for multiply- and simply-connected topologies, respectively.
The right-hand side of~(\ref{new-ind}) is defined in the following
way: one first takes the difference
of the two terms with ${\bf x}^{\prime} \neq {\bf x}$ and then sets
${\bf x}^{\prime} = {\bf x}$. Since $\mbox{\large $I$}_{v^2_i}^{MC}$ is not
simply the velocity dispersion $\bigl\langle\Delta v_i^2 \bigr\rangle^{MC}$ but
the difference~\eqref{new-ind}, the possibility
that it takes negative values
should not be a matter of worry,
a point that does not seem to have been appreciated in some previous works in which
this indicator was implicity used~\cite{Yu:2004tu,Bessa:2019aar,Yu:2004gs,%
Ford:2005rs,Yu:2006tn,Seriu_2008,Parkinson:2011yx,DeLorenci:2016jhd,jun2004vacuum}
together with the particular assumption that
the second term vanishes.\footnote{It is experimentally and theoretically unsettled whether
the simply-connected term on the right-hand side of (\ref{new-ind}) vanishes
or not.
We adopt the general view that it is nonzero \cite{Lemos:2020ogj}.
It is only under the very special assumption that it vanishes that one
stumbles upon the counterintuitive negative values for mean square velocities often found in the literature
\cite{Yu:2004tu,Bessa:2019aar,jun2004vacuum,Ford:2005rs,Seriu_2008,Parkinson:2011yx,DeLorenci:2016jhd,Yu:2004gs}.}
At first sight, the indicator \eqref{new-ind} does not seem measurable because it involves the difference of quantities associated with
two different spacetimes, but spacetime is unique. However, $\mbox{\large $I$}_{v^2_i}^{MC}$ is accessible by measurements performed only in our spacetime, which is to be tested for multiple connectedness, by the following procedure. First one would measure the velocity correlation function
$\bigl\langle \Delta v_i({\bf x},t) \Delta v_i({\bf x}^{\prime},t)\bigr\rangle^{exp}$ for
${\bf x} \neq {\bf x}^{\prime}$, then one would subtract out the correlation
function $\bigl\langle \Delta v_i({\bf x},t) \Delta v_i({\bf x}^{\prime},t)\bigr\rangle^{SC}$ that can be {\it theoretically computed} for
${\bf x} \neq {\bf x}^{\prime}$
for the corresponding topologically trivial (simply-connected) FRW spacetime, just as was done in the Appendix of Ref. \cite{Lemos:2020ogj} in the case of the Minkowski spacetime.\footnote{We do not present here the results for the topologically trivial FRW spacetime for the sake of brevity.}
Finally, the corresponding curve for the difference \eqref{new-ind} as a function of time would be plotted in the coincidence limit
${\bf x} = {\bf x}^{\prime}$ (see, for example, the figures in Section \ref{casestudy}). This approach is similar to one used in cosmic crystallography, which is intended to detect cosmic
topology from the distribution of discrete cosmic sources~\cite{Lehoucq:1996qe}.
A topological signature of any multiply connected
$3-$manifold of constant curvature is given by a constant times the
difference $\Phi_{exp}^{MC}(s_i) - \Phi^{SC}_{exp}(s_i)$ of the
expected pair separation histogram (EPSH) corresponding to the multiply
connected manifold and the EPSH for the underlying simply connected
covering manifold~\cite{Gomero:2001gq,Gomero:1998dz}, whose expression
{\it can be derived in analytical form}~\cite{Gomero:2001gq,Reboucas:2000du}.
It should be noticed that, although with the present techniques the second term on
the right-hand side of \eqref{new-ind} cannot be directly computed at
${\bf x}={\bf x}^{\prime}$, we can nonetheless presume that a rigorous theoretical
treatment may allow a consistent determination of this term. From our current knowledge
of the theory, one can also reasonably expect that a correct renormalization procedure
will not significantly change the qualitative features of the two terms, at least not
to the extent that it would affect comparison with observational data.
\subsection{Orientability indicators for charged point particle}
\label{Indicatorpointcharge}
Let us return to the calculation of the
components of the orientability indicator for ${E_{17}}$
topology.
Since the correlation functions~(\ref{correlation-x-E17}) to~(\ref{correlation-z-E17})
depend on $\eta$ and $\eta'$ only through their difference, the changes of integration
variables $\eta_1 = \eta^{\prime} - \eta_i$ and $\eta_2 = \eta^{\prime\prime} - \eta_i$
in Eq.~(\ref{veloc-dispersion-eta}) allow the components of the velocity dispersion
to be computed with the help of the integrals~\cite{Bessa:2019aar}
\begin{eqnarray}\label{integral1}
{\cal I} & = &\int_0^{\Delta \eta} \! \int_0^{\Delta \eta} d\eta_1d\eta_2
\frac{1}{[ (\eta_2 - \eta_1)^2 -r^2]^3} \nonumber \\
& = &\frac{\Delta \eta}{16r^5 (\Delta \eta^2-r^2)}\bigg\{4r\Delta \eta - 3(r^2-\Delta \eta^2)
\ln \frac{(r-\Delta \eta)^2}{(r+\Delta \eta)^2} \bigg\}
\end{eqnarray}
and
\begin{eqnarray}\label{integral2}
{\cal J} & = & \int_0^{\Delta \eta} \int_0^{\Delta \eta} d\eta_1 d\eta_2 \frac{(\eta_2
- \eta_1)^2}{[(\eta_2 - \eta_1)^2 -r^2]^3}
\nonumber \\
& = &\frac{\Delta \eta}{16r^3(\Delta \eta^2-r^2)}\bigg\{ 4r\Delta \eta + (r^2-\Delta \eta^2)
\ln \frac{(r-\Delta \eta )^2}{(r+\Delta \eta )^2} \bigg\}\,,
\end{eqnarray}
in which $\Delta \eta = \eta - \eta_i$.
Inserting equations~(\ref{correlation-x-E17}) to (\ref{integral2}) into
Eq.~(\ref{veloc-dispersion-eta})
and taking the coincidence limit ${\bf x}^{\prime} \to {\bf x}$ we find
\begin{eqnarray}\label{dispersion-x-E17}
{\mbox{\large $I$}_{v^2_x}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}} & = &
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}
\frac{q^2 \Delta\eta}{16\pi^2 m^2 r^5 (\Delta\eta^2 - r^2) a^2(t)}
\bigg\{ 4r\Delta\eta ({\bar r}_x^2 + r^2) \nonumber \\
& & \hspace*{2.5cm} +(\Delta\eta^2-r^2)(3{\bar r}_x^2 -r^2)\ln \frac{(r-\Delta\eta )^2}{(r+\Delta\eta )^2} \bigg\},
\\
\label{dispersion-y-E17}
{\mbox{\large $I$}_{v^2_y}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}} & = &
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}} \frac{q^2 \Delta\eta}{32\pi^2 m^2 r^5
(\Delta\eta^2 - r^2) a^2(t)}
\bigg\{ 4r\Delta\eta ({\bar r}_y^2 + (3-(-1)^{n_x})r^2) \nonumber \\
& & \hspace*{2.5cm} + (\Delta\eta^2-r^2)[3{\bar r}_y^2 -(3-(-1)^{n_x})r^2]
\ln \frac{(r-\Delta\eta )^2}{(r+\Delta\eta )^2} \bigg\}\,, \\
\label{dispersion-z-E17}
{\mbox{\large $I$}_{v^2_z}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}} & = &
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}\frac{q^2 \Delta\eta}{16\pi^2 m^2 r^5
(\Delta\eta^2 - r^2)a^2(t)}
\bigg\{ 4r\Delta\eta ({\bar r}_z^2 + r^2) \nonumber \\
& & \hspace*{2.5cm} + (\Delta\eta^2-r^2)(3{\bar r}_z^2 -r^2)
\ln \frac{(r-\Delta\eta )^2}{(r+\Delta\eta )^2} \bigg\}\,,
\end{eqnarray}
where
\begin{eqnarray}\label{r-ri-coincidence-r}
r & = & \sqrt{n_x^2L^2+2(1-(-1)^{n_x})y^2}, \\
\label{r-ri-coincidence-r-bar-x}
{\bar r}_x^2 & = & r^2- 2r_x^2 = -n_x^2 L^2 + 2(1-(-1)^{n_x})y^2, \\
\label{r-ri-coincidence-r-bar-y}
{\bar r}_y^2 & = & (1+ (-1)^{n_x}) r^2 - 8 (-1)^{n_x} (1-(-1)^{n_x})y^2 \nonumber \\
& = & (1+ (-1)^{n_x})n_x^2 L^2 + 8 (1-(-1)^{n_x})y^2,\\
\label{r-ri-coincidence-r-bar-z}
{\bar r}_z^2 & = & r^2- 2r_z^2 = r^2\,,
\end{eqnarray}
with the use of equations~(\ref{r-components-E17}) in the coincidence limit.
Note that the velocity dispersion depends not only on the time interval
$\Delta t =t -t_i$ (or $\Delta \eta = \eta - \eta_i$) but also on $t$ (or $\eta$)
itself. This was to be expected because in a dynamical universe invariance under
time translations is lost.
The orientability indicator for $E_{16}$ follows easily from the results for $E_{17}$.
The factors of $(-1)^{n_x}$ that appear in equations \eqref{dispersion-x-E17}
to~\eqref{dispersion-z-E17}
arise from derivatives with respect to $y^{\prime}$ in the separation $r$ given in
Table~\ref{Tb-Spatial-separation}.
Hence, the results for $E_{16}$ are immediately
obtained from those for $E_{17}$ by simply replacing $(-1)^{n_x}$ by $1$ everywhere
in Eqs.~(\ref{dispersion-x-E17}) to (\ref{dispersion-z-E17}). This leads to
\begin{eqnarray}\label{dispersion-orientable-x-E16}
{\mbox{\large $I$}_{v^2_x}^{E_{16}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}} & = &
-\frac{q^2 \Delta \eta}{4\pi^2 m^2 a^2(t)} \sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}
\frac{1}{n^3L^3} \ln \frac{(n_x L-\Delta \eta )^2}{(n_x L+\Delta \eta )^2}\,, \\
\label{dispersion-orientable-yz-E16}
{\mbox{\large $I$}_{v^2_y}^{E_{16}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}} & = &
{\mbox{\large $I$}_{v^2_z}^{E_{16}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}} \nonumber \\
& = & \frac{q^2 \Delta \eta}{8\pi^2 m^2 a^2(t)}
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}
\bigg\{ \frac{4\Delta \eta}{n_x^2L^2(\Delta \eta^2- n_x^2 L^2)} + \frac{1}{n_x^3L^3}
\ln \frac{(n_xL-\Delta \eta )^2}{(n_xL + \Delta \eta )^2} \bigg\}\,.
\end{eqnarray}
\section{NON-ORIENTABILITY WITH POINT ELECTRIC DIPOLE}
\label{Dipole-motion}
A noteworthy outcome of the previous section is that %
the time evolution of the orientability indicator for a charged particle
can be used to locally differentiate an orientable ($E_{16}$) from a
non-orientable ($E_{17}$) spatial section of Minkowski spacetime.
However, it cannot be used to decide whether a given $3-$space manifold per se
is or not orientable. As shown in \cite{Lemos:2020ogj}, the spatial orientability
of Minkowski spacetime in itself can be ascertained in principle by the motions of a
point electric dipole.
Therefore, it is reasonable to expect
that the orientability indicator for a dipole can potentially bring about unequivocal information
regarding non-orientability of the spatial sections of the spatially flat FRW spacetime.
To examine this issue we now turn our attention
to topologically induced motions of an electric dipole under quantum vacuum electromagnetic
fluctuations.
The spatial components of the four-force on a point electric dipole are
$f^i = p^j\partial_j E^i$ where ${\bf p}=(p^1,p^2,p^3)$ is the electric dipole
moment vector. Since the dipole is taken to be a bound system, it is not affected
by the expansion of the universe, which means that the dipole moment $\bf p$
is a constant vector.
Under the same assumptions as made for the point particle, replacing $qE^i$ by
$p^j\partial_j E^i$ and following the same steps that led from (\ref{eqmotion})
to (\ref{phys-veloc-dispersion}), the mean squared velocity in each of the three
independent directions $i = x, y, z$ is given by
\begin{equation}\label{eqdispersion-dipole}
\Bigl \langle\Delta v^i({\bf x},t)^2 \Bigr \rangle = \frac{p^jp^k}{m^2a^2(t)}
\int_{t_i}^t\int_{t_i}^t a^2(t^{\prime})a^2(t^{\prime\prime})
\Bigl\langle \bigl(\partial_j E^i({\bf x}, t')\bigr)
\bigl(\partial_k E^i({\bf x}, t'')\bigr)\Bigr\rangle_{FRW}\, dt' dt''\,,
\end{equation}
which can be conveniently rewritten as
\begin{equation}\label{eqdispersion-dipole-rewritten}
\Bigl \langle\Delta v^i({\bf x},t)^2 \Bigl\rangle = \lim_{{\bf x}^{\prime}
\to {\bf x}}\frac{p^jp^k}{m^2a^2(t)}\int_{t_i}^t\int_{t_i}^t a^2(t^{\prime})
a^2(t^{\prime\prime})
\partial_j\partial_k^{\,\prime} \Bigl\langle E^i({\bf x}, t')
E^i({\bf x}', t'')\Bigr \rangle_{FRW} dt' dt'' \,,
\end{equation}
where $\partial_l^{\,\prime}= \partial/\partial x_l^{\prime}$ and the
summation convention applies only to repeated upper and lower indices.
Now we proceed to the computation of the orientability indicator for a point dipole
in spaces $E_{17}$ and $E_{16}$. The space $E_{17}$ has two topologically
special directions:
the compact $x$-direction and the flip $y$-direction associated
with the non-orientability of $E_{17}$.
To probe the non-orientability of $E{_{17}}$ by means
of stochastic motions it seems most promising to choose a dipole oriented
in the $y$-direction, since the orientation of the dipole would also be flipped
upon every displacement by the topological length $L$ along the compact
direction~\cite{Lemos:2020ogj}.
For a dipole oriented along the $y$-axis we have ${\bf p}=(0,p,0)$
and
\begin{equation}\label{eqdispersion-dipole-y}
\Bigl \langle\Delta v_x({\bf x},t)^2 \Bigr\rangle^{(y)} = \lim_{{\bf x}^{\prime}
\to {\bf x}}\frac{p^2}{m^2a^2(t)}\int_{t_i}^t\int_{t_i}^t a^2(t^{\prime})
a^2(t^{\prime\prime})
\partial_y\partial_{y^{\prime}} \Bigl\langle E_x({\bf x}, t')
E_x({\bf x}', t'')\Bigr\rangle_{FRW}\, dt' dt'',
\end{equation}
where the superscript within parentheses indicates the dipole's orientation.
We now proceed as in the case of the charged particle and rewrite the above
integral in terms of the conformal time and the correlation function in
Minkowski spacetime. It follows that the above equation takes the form
\begin{equation}\label{eqdispersion-dipole-y-conformal}
\Bigl \langle\Delta v_x({\bf x},t)^2\Bigr\rangle^{(y)} =\lim_{{\bf x}^{\prime}
\to {\bf x}}\frac{p^2}{m^2a^2(t)}
\times \int_{\eta_i}^{\eta}\int_{\eta_i}^{\eta} \partial_y\partial_{y^{\prime}}
\Bigl\langle E_x({\bf x}, \eta')
E_x({\bf x}', \eta'')\Bigr\rangle_M\, d\eta' d\eta''\,.
\end{equation}
Upon replacing the eletric-field correlation function by its renormalized version
given by Eq.~(\ref{correlation-x-E17}), the above equation yields the orientability
indicator in the $x$-direction for $E_{17}$:
\begin{equation}\label{eqdispersion-dipole-yx}
{\mbox{\large $I$}_{v^2_x}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}}
= \lim_{{\bf x}^{\prime}
\to {\bf x}}\frac{p^2}{\pi^2m^2a^2(t)} \sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}
\int_{\eta_i}^{\eta}\int_{\eta_i}^{\eta}
\partial_y\partial_{y^{\prime}} \frac{\Delta \eta^2 + r^2 -2r_x^2}{({\Delta \eta}^2 - r^2)^3}\,.
\end{equation}
where $\Delta \eta = \eta'-\eta''$, while $r_x$ and $r$ are defined by Eq.~(\ref{r-components-E17}).
Making use of
\begin{equation}\label{partial-y-yprime-yx}
\partial_y\partial_{y^{\prime}} \frac{\Delta \eta^2 + r^2 -2r_x^2}{({\Delta \eta}^2 - r^2)^3}
= -4(-1)^{n_x} \bigg[ \frac{2}{({\Delta \eta}^2 -r^2)^3}
+ 3 \frac{r^2 - r_x^2+ 6r_y^2}{({\Delta \eta}^2 -r^2)^4}
+ 24 \frac{(r^2 - r_x^2) r_y^2}{({\Delta \eta}^2 -r^2)^5} \bigg].
\end{equation}
we find
\begin{equation}\label{eqdispersion-dipole-yx-final-E17}
{\mbox{\large $I$}_{v^2_x}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}} = -\frac{4p^2}{\pi^2m^2a^2(t)} \sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}(-1)^{n_x} \bigg\{ 2I_1
+ 3 (r^2 - r_x^2 + 6r_y^2)I_2+ 24 (r^2 - r_x^2) r_y^2I_3 \bigg\},
\end{equation}
where~\cite{Lemos:2020ogj}
\begin{eqnarray}\label{integral-1}
&&I_1 = {\cal I} = \int_0^{\Delta\eta} \int_0^{\Delta\eta}
\frac{d\eta_1d\eta_2}{[(\eta_2-\eta_1)^2 -r^2]^3}
= \frac{\Delta\eta}{16} \bigg[ \frac{4\Delta\eta}{r^4 (\Delta\eta^2-r^2)}
+ \frac{3}{r^5} \ln \frac{(r-\Delta\eta )^2}{(r+\Delta\eta )^2} \bigg],\\
&&I_2 = \int_0^{\Delta\eta} \int_0^{\Delta\eta} \frac{d\eta_1d\eta_2}{[(\eta_2 -\eta_1)^2 -r^2]^4}
= \frac{\Delta\eta}{96}\bigg[ \frac{4\Delta\eta (9r^2-7\Delta\eta^2)}{r^6 (\Delta\eta^2-r^2)^2}
- \frac{15}{r^7} \ln \frac{(r-\Delta\eta )^2}{(r+\Delta\eta )^2} \bigg],\label{integral-2}\\
&&I_3 = \int_0^{\Delta\eta} \int_0^{\Delta\eta} \frac{d\eta_1d\eta_2}{[(\eta_2 -\eta_1)^2 -r^2]^5}
\nonumber\\
&& \hspace*{3cm} = \frac{\Delta\eta}{768} \bigg[\frac{105}{r^9} \ln \frac{(r-\Delta\eta )^2}{(r+\Delta\eta )^2}
+ \frac{4\Delta\eta (57\Delta\eta^4- 136r^2\Delta\eta^2 + 87r^4)}
{r^8 (\Delta\eta^2-r^2)^3} \bigg],\label{integral-3}
\end{eqnarray}
in which $\Delta\eta = \eta - \eta_i$.
Similar calculations lead to
\begin{eqnarray}\label{eqdispersion-dipole-yy-final-E17}
{\mbox{\large $I$}_{v^2_y}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}}
& = & -\frac{2p^2}{\pi^2m^2a^2(t)}\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}} (-1)^{n_x}
\bigg\{ (5-3(-1)^{n_x}) I_1 \nonumber \\
& & \hspace*{0.2cm} + 6 [r^2 + (7-6(-1)^{n_x})r_y^2] I_2 + 48 [r^2 -(-1)^{n_x}r_y^2]r_y^2I_3 \bigg\}
\end{eqnarray}
and
\begin{equation}
{\mbox{\large $I$}_{v^2_z}^{E_{17}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}} =-\frac{4p^2}{\pi^2m^2a^2(t)} \sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}}(-1)^{n_x}\bigg\{ 2I_1
+ +3(r^2 + 6 r_y^2 )I_2 + 24 r^2 r_y^2I_3 \bigg\}.
\label{eqdispersion-dipole-yz-final-E17}
\end{equation}
Since the coincidence limit ${\bf x}^{\prime} \to {\bf x}$ has been taken, it follows
from Eq.~(\ref{r-components-E17}) that in Eqs. (\ref{eqdispersion-dipole-yx-final-E17}) to
(\ref{eqdispersion-dipole-yz-final-E17}) one must put
\begin{equation}
\label{r-rx-ry-coincidence-E17}
r= \sqrt{n_x^2L^2 + 2(1-(-1)^{n_x})y^2},\; \;
r_x^2 = n_x^2L^2,\;\; r_y^2 = 2(1-(-1)^{n_x})y^2.
\end{equation}
For the slab space $E_{16}$ the components of the dipole velocity
dispersion are obtained from those for $E_{17}$ by setting $r_x^2 = r^2, r_y=0 $,
and replacing $(-1)^{n_x}$ by $1$ everywhere. Therefore, we have
\begin{eqnarray}\label{eqdispersion-dipole-yx-final-E16}
{\mbox{\large $I$}_{v^2_x}^{E_{16}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}} & = & -\frac{8p^2}{\pi^2m^2a^2(t)}
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}} I_1 ,\\
\label{eqdispersion-dipole-yy-final-E16}
{\mbox{\large $I$}_{v^2_y}^{E_{16}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}} & = & -\frac{4p^2}{\pi^2m^2a^2(t)}
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}} ( I_1 + 3r^2I_2 ),\\
\label{eqdispersion-dipole-yz-final-E16}
{\mbox{\large $I$}_{v^2_z}^{E_{16}}}({\bf x},t)_{_{\mbox{\scriptsize $FRW$}}}^{^{\mbox{\scriptsize $(y)$}}} & = & -\frac{4p^2}{\pi^2m^2a^2(t)}
\sum\limits_{{n_x=-\infty}}^{{\infty\;\;\prime}} (2I_1 + 3r^2I_2 ),
\end{eqnarray}
in which $r=\vert n_x \vert L$.
It is worth pointing out that the general results obtained for the
orientability indicators are independent of a specific gravitation theory,
they depend only on the assumption that spacetime is endowed with the
spatially flat metric~\eqref{metric}. We have also found that the amplitude of the indicator components, both for the charged particle and the electric dipole, is inversely proportional to $a^2(t)$.
This was not expected from the outset, and had to be derived.
However, the indicators depend on the scale factor,
which is determined by the gravitational theory.
Therefore, in order to get definite expressions for the indicators we shall
consider a specific model within the framework of general relativity.
\section{A case study}\label{casestudy}
The purpose of this section is to study the time evolution of the orientability
indicators for spatially flat expanding universes with spatial sections endowed
with either of the nontrivial topologies $E_{16}$ or $E_{17}$.
A difficulty arises, however, because the orientability indicators found in the
two previous sections, both for the charged particle and the point dipole, are
expressed in terms of a mixture of the cosmic time $t$ and the conformal
time $\eta$, which makes it difficult to bring out their behavior.
In order to express the orientability indicators exclusively in terms
of $t$ or $\eta$ one has to know the scale factor as a function of
$t$ or $\eta$.
As we have mentioned in the Introduction, the metric \eqref{metric}
expresses the principle of spatial homogeneity and isotropy along with
the existence of a cosmic time $t$, with the additional observational
input from the Planck collaboration~\cite{2016,2020}
that provided strong support for a flat $3-$space ($|\Omega_k| < 0.003$).
To study the dynamics of the Universe another assumption is necessary,
namely that the large scale structure of the Universe is essentially
determined by gravitational interactions, and therefore can be described
by a metrical theory of gravity, which we assume to be general relativity.
These very general assumptions constrain the matter content of the Universe
to be described by a perfect fluid with energy-momentum tensor
\begin{equation}
\label{EM-Tensor}
T_{\mu\nu} = (\rho+p)\,u_\mu u_\nu - p \,g_{\mu \nu}\;,
\end{equation}
where $u_\mu$ is the fluid four-velocity, $\rho$ is the total energy density and $p$ is the
pressure. In the case of arbitrary space curvature, the Einstein field equations imply the
Friedmann equation
\begin{equation}
\label{Friedmann}
\frac{\dot{a}^2}{a^2} = \frac{8\pi G}{3} \rho - \frac{k}{a^2},
\end{equation}
where $G$ is the gravitational constant. The conservation law $\nabla_{\mu}T^{\mu\nu}=0$
leads to the fluid equation
\begin{equation}
\label{fluid}
{\dot \rho} + 3\frac{\dot a}{a}(\rho + p) = 0.
\end{equation}
Our main interest is to show how valuable the orientability indicators are.
Thus, for illustrative purposes we shall consider a model universe which is
spatially flat and whose matter content consists of a single-component
barotropic perfect fluid with equation of state $p = w \rho$, where
the pure number $w$ satisfies $\vert w \vert < 1$ .
In this case the expanding solution to the Friedmann equation~(\ref{Friedmann})
with $k=0$ and the fluid equation (\ref{fluid}) is \cite{Ryden:2002vsj}
\begin{equation}
a=\left(\frac{t}{t_0}\right)^{2/(3+3w)}.
\label{sol_1_componente}
\end{equation}
The age of this universe is
\begin{equation}
t_0=\frac{2}{3(1+w)} H_0^{-1}
\label{idade}
\end{equation}
where $H_0$ is the Hubble constant, the present value of $H= {\dot a}/a$.
According to the latest observational data from the Planck team \cite{2020},
$H_0\simeq 67.37$ km/(s Mpc) or $H_0^{-1}\simeq 14.5$ Gyr and the age of the Universe is
$t_0 \simeq 13.8$ Gyr. In order for equation \eqref{idade} to match these results
we choose the equation of state parameter as $w = -0.299$. Thus, from
Eq.~\eqref{sol_1_componente} we find
\begin{equation}
\label{tempo_conf}
\Delta \eta = \int_{t_i}^t \frac{dt^{\prime}}{a(t^{\prime})}
= \frac{3(1+w)}{1+3w}t_0\left[\left(\frac{t}{t_0}\right)^{\frac{1+3w}{3(1+w)}}
-\left(\frac{t_i}{t_0}\right)^{\frac{1+3w}{3(1+w)}}\right].
\end{equation}
By means of this result the orientability indicators found in
Sections~\ref{disp-particleE17} and \ref{Dipole-motion} are expressed
in terms of the proper time $t$ alone.
It is extremely hard analytically to figure out how the orientability indicators
behave as functions of cosmic time because their expressions are given by quite
involved infinite sums.
The visualization of their behavior can be much more easily accomplished by
numerical plots, which is what we shall concentrate on in the following.
For all plots we set $t_i=t_0=1$, which means that the fluctuations begin to
be measured by the orientability indicators at the same reference instant
at which the scale factor is unity, that is, today.
The compactification length discussed in \eqref{ident-E16}, \eqref{ident-E17} and
Table \ref{Tb-Spatial-separation} is also set equal to unity: $L=1$.
Following \cite{Lemos:2020ogj}, the choice $y=0$, which freezes out the global
inhomogeneity degree of freedom, is made in all plots but
Fig.~\ref{Fig1}(a), in which we take $y=1/2$ to illustrate the global
inhomogeneity effect.
The orientability indicators are computed from equation~%
~\eqref{dispersion-x-E17}~--\eqref{dispersion-orientable-yz-E16} for the charged
particle, and equations~\eqref{eqdispersion-dipole-yx-final-E17}~--%
~\eqref{eqdispersion-dipole-yz-final-E16} for the point dipole, as well
as~\eqref{tempo_conf} with $ w =-0.299$.
The infinite sums are rapidly convergent, and the summations are numerically
performed taking $n_x \neq 0$ ranging from $-50$ to $50$.
\subsection{Non-orientability: point charge case}
\begin{figure*}[htpb]
\begin{center}
\begin{tabular}{c c}
\resizebox{0.45\columnwidth}{!}{\includegraphics{Figura1.eps}}
& \resizebox{0.45\columnwidth}{!}{\includegraphics{Figura2.eps}} \\
(a) & (b)
\end{tabular}
\end{center}
\caption{Time evolution of orientability indicators for the point charge.
In (a) the curve for indicator $\mbox{\large $I$}_{v^2_x}^{E_{17}}$ for $y=0$, shown as
a dotted line, coincides with the curve for $\mbox{\large $I$}_{v^2_x}^{E_{16}}$,
depicted as a dashed line. For $y=1/2$, the solid curve for $\mbox{\large $I$}_{v^2_x}^{E_{17}}$
is now different from the one for $\mbox{\large $I$}_{v^2_x}^{E_{16}}$.
In (b) the orientability indicators $\mbox{\large $I$}_{v^2_y}^{E_{17}}$ and
$ \mbox{\large $I$}_{v^2_y}^{E_{16}}$ are different even for $y=0$.
\label{Fig1}}
\end{figure*}
Figure~\ref{Fig1} shows orientability indicators as functions of cosmic time for the point charge.
In panel (a) the time evolution of $ {\mbox{\large $I$}_{v^2_x}^{E_{16}}}$ given
by~\eqref{dispersion-orientable-x-E16} is shown as a dashed line and that of
${\mbox{\large $I$}_{v^2_x}^{E_{17}}}$, given by~\eqref{dispersion-x-E17}, is displayed as the
dotted line for $y=0$. These indicators coincide. In panel (a) we also show as the solid
line the indicator $ {\mbox{\large $I$}_{v^2_x}^{E_{17}}}$ for $y=1/2$, exhibiting the
global inhomogeneity effect for the $E_{17}$ manifold.
In panel (b) of Fig.~\ref{Fig1} the component of orientability indicator
$ {\mbox{\large $I$}_{v^2_y}^{E_{16}}}$ given by~\eqref{dispersion-orientable-yz-E16}
is represented by a dashed line, whereas the component ${\mbox{\large $I$}_{v^2_y}^{E_{17}}}$,
defined by Eq.~\eqref{dispersion-y-E17}, is depicted as a dotted line.
Now these indicator curves are distinct.
Thus, it is possible to tell the two topologies apart by means of this component of the indicator.
Nevertheless, as will be seen below, the dipole is much more sensitive to non-orientability.
The behaviors of the components ${\mbox{\large $I$}_{v^2_z}^{E_{17}}}$
and~$ {\mbox{\large $I$}_{v^2_z}^{E_{16}}}$ are not shown because the corresponding curves
coincide when $y=0$, as can be directly checked from equations~\eqref{dispersion-z-E17} and~\eqref{dispersion-orientable-yz-E16}, although they
are distinct if $y \neq 0$.
\begin{figure*}[h]
\begin{center}
\begin{tabular}{c c}
\resizebox{0.45\columnwidth}{!}{\includegraphics{Figura3a.eps}}
& \resizebox{0.45\columnwidth}{!}{\includegraphics{Figura3b.eps}} \\
(a) & (b)
\end{tabular}
\end{center}
\caption{Under the same conditions as in Fig.~\ref{Fig1}(b), but now for a point dipole,
panel (a) shows the indicator $\mbox{\large $I$}_{v^2_x}^{E_{17}}$, whereas panel (b) displays
$\mbox{\large $I$}_{v^2_x}^{E_{16}}$. We have intentionally made separate plots for the
$E_{16}$ and $E_{17}$ topologies to emphasize the repetitious inversion pattern
roughly resembling $\cup$ followed by $\cap$ in the case of the non-orientable
spatial topology.
The indicator $\mbox{\large $I$}_{v^2_x}^{E_{16}}$ exhibited in panel (b) displays no inversion pattern.
\label{Figura2}}
\end{figure*}
\begin{figure*}[htpb]
\begin{center}
\begin{tabular}{c c}
\resizebox{0.45\columnwidth}{!}{\includegraphics{Figura4.eps}}
& \resizebox{0.45\columnwidth}{!}{\includegraphics{Figura5.eps}} \\
(a) & (b)
\end{tabular}
\end{center}
\caption{Under the same conditions as in Fig.~\ref{Figura2}, and also for a point dipole,
panel (a) shows $\mbox{\large $I$}_{v^2_y}^{E_{17}}$ as a dashed line and
$\mbox{\large $I$}_{v^2_y}^{E_{16}}$ as a solid line. Panel (b) displays
$\mbox{\large $I$}_{v^2_z}^{E_{17}}$ as a dashed line and
$\mbox{\large $I$}_{v^2_z}^{E_{16}}$ as a solid line. For FRW spacetime with the
non-orientable spatial topology, $E_{17}$, there is an inversion pattern resembling
successive upward and downward ``horns". Similar alternating horn-like inversion
patterns also arise in static Minkowski spacetime with $E_{17}$
topology~\cite{Lemos:2020ogj}, but here their shape is modified by the dynamical scale factor.
\label{Fig3}}
\end{figure*}
\subsection{Non-orientability: point dipole case}
Figure~\ref{Figura2} shows the time evolution of orientability indicators for the point dipole.
Fig.\ref{Figura2}(a) displays the dipole indicator $ \mbox{\large $I$}_{v^2_x}^{E_{17}}$,
given by Eq.~\eqref{eqdispersion-dipole-yx-final-E17}, and Fig.\ref{Figura2}(b) shows
$ \mbox{\large $I$}_{v^2_x}^{E_{16}}$, obtained from equation~\eqref{eqdispersion-dipole-yx-final-E16}.
We have intentionally exhibited in separate plots the indicators for the manifolds $E_{17}$
and $E_{16}$ in order to highlight the repetitious pattern roughly resembling $\cup$ followed
by $\cap$ in the non-orientable case.
Similar repetitious inversion patterns are also present in Minkowski spacetime with
$E_{17}$ topology~\cite{Lemos:2020ogj}. Here the shape of the curves is
modified by the dynamical scale factor $a(t)$, though.
Just as in the Minkowski case, the orientability indicator is already sensitive to
non-orientability even for $y=0$. The inversion pattern exhibited by $E_{17}$ is
qualitatively different from the pattern for $E_{16}$, which is not characterized by
successive inversions, making it possible to identify the non-orientable case in itself.
Figure~\ref{Fig3} exhibits the two remaining components of the dipole indicators.
Fig.~\ref{Fig3}(a) displays as a dashed line the dipole indicator
$\mbox{\large $I$}_{v^2_y}^{E_{17}}$ given by~\eqref{eqdispersion-dipole-yy-final-E17},
together with $ \mbox{\large $I$}_{v^2_y}^{E_{16}}$, defined
by~\eqref{eqdispersion-dipole-yy-final-E16}, which is depicted as a solid line.
Panel (b) of Fig.~\ref{Fig3} shows as a dashed line the orientability indicator
$\mbox{\large $I$}_{v^2_z}^{E_{17}}$ given by~\eqref{eqdispersion-dipole-yz-final-E17},
as well as $\mbox{\large $I$}_{v^2_z}^{E_{16}}$, defined
by~\eqref{eqdispersion-dipole-yz-final-E16}, as a dotted line.
Both panels reveal the same inversion pattern roughly resembling successive
upward and downward ``horns".
Similar alternating horn-like inversion patterns emerge in static Minkowski
spacetime with $E_{17}$ topology~\cite{Lemos:2020ogj}. Now, however, the shape
of the curves is modified by the dynamical scale factor $a(t)$.
This distinctive pattern allows one to recognize the non-orientability of
$3-$space per se.
\subsection{Non-orientabily: summary of findings}
Stochastic motions of a point electric charge under quantum electromagnetic fluctuations
give rise to the orientability indicators defined by equations \eqref{dispersion-x-E17} to
\eqref{dispersion-orientable-yz-E16}. Fig. \ref{Fig1}(b) shows that the $y$-component of
the orientability indicator for $E_{17}$ is different from the one for $E_{16}$.
Therefore, one can distinguish one topology from the other. But the curve patterns
for both topologies are qualitatively the same. This means that the identification of
a putative non-orientable topology requires a quantitative comparison of its
evolution curves with those for the counterpart orientable topology.
In short, we are unable to spot a non-orientable topology \textsl{per se}
by means of the stochastic motions of a point charged particle.
Things become more appealing when one considers the stochastic motions
of a point electric dipole, whose corresponding non-orientability
indicators are given by equations~\eqref{eqdispersion-dipole-yx-final-E17}
to~\eqref{eqdispersion-dipole-yz-final-E16}. A comparison of the
$x$-component of the orientability indicators, exhibited in Fig.~\ref{Figura2},
already shows a repetitious inversion pattern of roughly resembling
$\cup$ followed by $\cap$
for $E_{17}$ that distinguishes it from $E_{16}$. This distinction is much
more pronounced when the remaining components of the indicators for
$E_{17}$ and $E_{16}$ are compared, as in Fig.~\ref{Fig3}.
The conspicuous pattern roughly
resembling alternating upward and downward ``horns"
for $E_{17}$, which is absent from the curves for $E_{16}$, enables
us to identify the non-orientability of $E_{17}$ by itself, without
the need to compare its indicator curves with those for its orientable
counterpart. This is the greatest advantage of the dipole over the
point charge in the search for detecting non-orientability through
stochastic motions of point-like objects under quantum vacuum
fluctuations of the electromagnetic field.
Contrarily to the dipole case, the point-charge case does
not exhibit such prominent qualitative features and hence appears to be
of hardly any practical use.
\section{Closing remarks and conclusions} \label{Finals}
In the framework of general relativity
the Universe is modelled as a four-dimensional differentiable manifold
$\mathcal{M}_4$ endowed with the FRW metric~\eqref{RWmetric} that expresses
geometrically two basic assumptions of the cosmological modeling, namely the
existence of a cosmic time $t$, which emerges from Weyl's principle, and the
cosmological principle, which in turn ensures that the $3-$dimensional space $M_3$
is geometrically homogeneous and isotropic.
The FRW metric does not specify the topology of the underlying spacetime manifold
$\mathcal{M}_4$ or of the corresponding spatial ($t= \mbox{const}$) sections $M_3$,
which can in principle be found through observations.
So far, however, direct searches for a nontrivial topology of $M_3$ using CMB data from
WMAP and Planck have found no convincing evidence of multiple connectedness
below the radius of the last scattering surface~\cite{2014,2016,%
Cornish:2003db,ShapiroKey:2006hm, Bielewicz:2010bh,Vaudrevange:2012da,Aurich:2013fwa}.%
\footnote{This does not exclude the possibility of a FRW universe with
a detectable nontrivial cosmic topology~\cite{Gomero:2016lzd,%
Bernui:2018wef,Aurich:2021ofm}. }
In this work, rather than focusing on determining the topology of the
spatial sections $M_3$ of FRW spacetime, we have investigated its global
property of orientability.
In the physics at daily and even astrophysical length and time scales we do
not find any sign or hint of non-orientability of $3-$space.
At the cosmological scale, in order to disclose spatial
non-orientability, global trips around the whole $3-$space would be needed
to check for orientation-reversing closed paths.
Since such global journeys across the Universe are not feasible one might
think that spatial orientability cannot be probed.
We note, however, that the determination of the spatial
topology through, for example, the so-called circles-in-the-sky,
would bring out as a bonus an answer as to $3-$space orientability
at the cosmological scale.
On the other hand, at a theoretical level of reasoning, it is often
assumed that the spacetime manifold is separately time and space orientable.
As we have mentioned, the arguments supporting orientability combine the
space-and-time universality of the local physical laws\footnote{We note that space universality can be looked upon as a
topological assumption of global homogeneity of $M_3$. So, all elements
of the covering group $\Gamma$ are translations, and therefore spatial
universality by itself rules out non-orientable $3-$spaces.}
with a well-defined local arrow of time along with some local results
from discrete symmetries in particle physics~\cite{Hawking:1973uf,1979grec.conf..212G}.
Of course one is free to resort to such reasonings, but it is reasonable
to expect that the ultimate answer to questions
regarding the orientability of spacetime should rely on cosmological
observations or local experiments, or might come from a fundamental
theory of physics.
In this paper we have investigated whether electromagnetic quantum vacuum
fluctuations can be used to access the spatial orientability of a FRW expanding
spacetime, extending therefore the results of the recent paper~\cite{Lemos:2020ogj},
where the question of spatial orientability of Minkowski (static) spacetime was
examined.
To this end, we have studied the stochastic motions of point-like objects
under quantum electromagnetic fluctuations in FRW flat spacetime with the
orientable $E_{16}$ (slab) and non-orientable $E_{17}$ (slab with flip)
space topologies (cf.\ Tables~\ref{Tb-2-Orient_and_Non_orient}
and~\ref{Tb-Spatial-separation}).
The statistical indicator $\mbox{\large $I$}_{v^2_i}^{MC}$ [Eq.~\eqref{new-ind}], which
measures the departure of the mean square velocity dispersion for point-like
objects in the multiply-connected topology from its value for the
simply-connected covering space, has been shown to be suitable to reveal
spatial orientability of flat FRW spacetime.
In the case of a charged particle, we have derived expressions
\eqref{dispersion-x-E17}~--~\eqref{dispersion-orientable-yz-E16}
for the orientability indicator $\mbox{\large $I$}_{v^2_i}^{MC}$
for the
$E_{17}$ and $E_{16}$ space topologies.
Similarly, we have derived expressions~\eqref{eqdispersion-dipole-yx-final-E17
~--~\eqref{r-rx-ry-coincidence-E17} for the indicator~\eqref{new-ind} corresponding to
a point electric dipole
oriented in the flip direction of $E_{17}$
topology, and also expressions~\eqref{eqdispersion-dipole-yx-final-E16}~--%
~\eqref{eqdispersion-dipole-yz-final-E16} for the dipole in $3-$space
with the orientable $E_{16}$ topology.
The expressions for the orientability indicators for the particle
and the dipole in $E_{17}$ and $E_{16}$ spatial topologies
hold for an arbitrary scale factor $a(t)$, which is determined by
the gravitational theory.
In this way, to concretely study the time evolution of the orientability
indicator $\mbox{\large $I$}_{v^2_i}^{MC}$
we have assumed in Section~\ref{casestudy} the general relativity theory
and also that the matter content consists of a
single-component barotropic perfect fluid with equation of state $p = w \rho$,
with the equation of state parameter $w$ such that $| w | < 1$.
Under these assumptions we have made figures~\ref{Fig1} to~\ref{Fig3}.
Figure~\ref{Fig1}(a) illustrates the topological inhomogeneity effect
of $E_{17}$ topology and makes it clear that $y=0$ is the appropriate
particle's position in $E_{17}$ for comparison of the evolution of
the orientability indicator in $E_{16}$ and $E_{17}$ topologies.
Figure~\ref{Fig1}(b) shows that it is possible to distinguish
the orientable from the non-orientable topology by comparing the time
evolution of the respective $y-$components of the orientability indicator:
they give rise to different evolution curves for distinct topologies.
A more ambitious goal is that of finding a way to decide about
the orientability of a given spatial manifold in itself, without
having to make a comparison of the results for a non-orientable
space with those for its orientable counterpart.
We have addressed this matter and have shown that the
stochastic motions of a point electric dipole can be used
to disclose the putative non-orientability of a generic
$3-$space per se.
To this end, under the premises put forward in Section~\ref{casestudy}
(flat spacetime in general relativity with perfect fluid source)
we have used expressions
\eqref{eqdispersion-dipole-yx}~--~\eqref{r-rx-ry-coincidence-E17} and
\eqref{eqdispersion-dipole-yx-final-E16}~--~\eqref{eqdispersion-dipole-yz-final-E16}
for the stochastic motions of the dipole in FRW spacetime with, respectively,
the non-orientable $E_{17}$ and orientable $E_{16}$ spatial topologies to
plot Figures~\ref{Figura2} and \ref{Fig3}. These figures show
that an inversion pattern for the orientability
indicator curves comes about in the case of the non-orientable $E_{17}$ topology,
implying that the non-orientability of $E_{17}$ can be detected per se.%
\footnote{The curves for $E_{16}$ display a repetition pattern, which, however
is not an inversion pattern.}
Our results for the non-orientability indicator hold for any scale factor
$a(t)$. For the case studied --- flat FRW geometry in the context of general
relativity; matter content described by a perfect fluid with equation of state $p = w\rho$ with
$w = -0.299$ fixed from the Planck satellite observational data --- a non-orientability signature in the form of an inversion pattern
was found.
All we can be sure of is that there is an inversion pattern
in this particular case. In the framework of some other metric
gravitational theory or for other scale factors one has to re-examine the
question in order to find out whether the scalar factor would preserve,
modify or even destroy the inversion pattern.
An expected result from the beginning of this work was that the role played
by the topology on the stochastic motions of particles would depend crucially
on the topological compact length $L$, which in turn gives rise
to a lower bound for the time scale required to test orientability.
However, the time scale involved in Figs. \eqref{Fig1}~--~\eqref{Fig3} makes it clear that
to access the inversion patterns of the orientability indicators one needs
a relatively long period of time, typically the time needed
to travel across quite a few $L$s. A small topological length scale
is expected, for example, in the primordial universe.
An open question is whether those velocity fluctuations would leave
traces that could be extracted from today's observational data, as for example
from CMB maps, making it potentially possible to unveil information on
$3-$space orientability. This is a nontrivial and important issue
beyond the scope of the present paper, though.
\begin{acknowledgements}
M.J. Rebou\c{c}as acknowledges the support of FAPERJ under a CNE E-26/202.864/2017 grant,
and thanks CNPq for the grant under which this work was carried out.
We thank C.H.G. Bessa for fruitful discussions.
M.J.R. is also grateful to A.F.F. Teixeira for interesting comments, and also for reading
the manuscript and indicating typos.
\end{acknowledgements}
|
\section{HNCA in a Contextual Bandit Setting}
\label{bandit}
We first formulate HNCA in a contextual bandit setting. In this setting, an agent interacts with an environment in a series of time-steps.\footnote{We suppress the time-step, for example writing the context as $X$ instead of $X_t$.} At each time-step, the environment presents the agent with an i.i.d. random context $X\in\mathcal{X}$ (for example the pixels of an image). The agent then selects an action from a discrete set of choices $A\in\mathcal{A}$ (for example a guess of what class the image belongs to). The environment responds to the agent's choice with a reward $R=R(X,A)$, where $R: \mathcal{X},\mathcal{A}\mapsto\mathbb{R}$ is an unknown reward function (for example a reward of 1 for guessing the correct class and 0 otherwise). The agent's goal is to select actions which result in as much reward as possible.
In our case, the agent consists of a network of stochastic computational units. Let $\Phi$ be a random variable corresponding to the output of a particular unit. For each unit, $\Phi$ is drawn from a parameterized \textit{policy} ${\pi_\Phi(\phi|b)\defeq\P(\Phi=\phi|\pa(\Phi)=b)}$ conditioned on $\pa(\Phi)=b$, its parents in the network.\footnote{Expectations and probabilities are taken with respect to all random variables in the network, and the context.} Each unit's policy is differentiably parameterized by a unique set of parameters $\theta_\Phi\in\mathbb{R}^d$. A unit's parents $\pa(\Phi)$ may include the output of other units, as well as the context $X$. We focus on the case where $\Phi$ takes values from a discrete set. We will use $\ch(\Phi)$ to refer to the children of $\Phi$, that is, the set of outputs of all units for which $\Phi$ is an input.\footnote{We may also apply $\ch(\cdot)$ or $\pa(\cdot)$ to sets, in which case it has the obvious meaning of the union of the elementwise applications.} We assume the network has a single output unit, which selects the action $A$ sent to the environment.
The goal is to tune the network parameters to increase $\E[R]$. Towards this, we will construct an unbiased estimator of the gradient $\frac{\partial\E[R]}{\partial\theta_\Phi}$ for the parameters of each unit, and update the parameters according to the estimator.
Directly computing the gradient of the output probability with respect to the parameters for a given input, as we might do with backpropagation for a deterministic network, is generally intractable for discrete stochastic networks. Instead, we can define a local REINFORCE estimator, $\hat{G}^{\text{RE}}_{\Phi}\defeq\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}R$. It is well known that $\hat{G}^{\text{RE}}_{\Phi}$ is an unbiased estimator of $\frac{\partial\E[R]}{\partial\theta_\Phi}$ (see Appendix~\ref{local_REIN_unbiased} for a proof). However, $\hat{G}^{\text{RE}}_{\Phi}$ tends to have high variance.
\subsection{HNCA Gradient Estimator}
HNCA exploits the causal structure of the network to assign credit to each unit's output based on how it impacts the output of its immediate children. Assume $\Phi$ is a nonoutput unit and define $\mb(\Phi)\defeq\{\ch(\Phi),\pa(\Phi),\pa(\ch(\Phi))\setminus\Phi\}$ as a notational shorthand. Note that $\mb(\Phi)$ is a Markov blanket~\citep{pearl1988probabilistic} for $\Phi$, meaning that conditioned on $\mb(\Phi)$, $\Phi$ is independent of all other variables in the network as well as the reward $R$. Beginning from the expression for $\hat{G}^{\text{RE}}_{\Phi}$, we can rewrite $\frac{\partial\E[R]}{\partial\theta_\Phi}$ as follows:
\begin{align*}
\frac{\partial\E[R]}{\partial\theta_\Phi}
&\stackrel{(a)}{=}\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}R\right]\\
&\stackrel{(b)}{=}\E\left[\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}R\middle|\mb(\Phi),R\right]\right]\\
&\stackrel{(c)}{=}\E\left[\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}\middle|\mb(\Phi)\right]R\right]\\
&\stackrel{(d)}{=}\E\left[\sum_\phi\frac{\P(\Phi=\phi|\mb(\Phi))}{\pi_{\Phi}(\phi|\pa(\Phi))}\frac{\partial\pi_\Phi(\phi|\pa(\Phi))}{\partial\theta_\Phi}R\right],\numberthis\label{HNCA_derivation_1}
\end{align*}
where $(a)$ follows from the unbiasedness of $\hat{G}^{\text{RE}}$, $(b)$ applies the law of total expectation, $(c)$ pulls $R$ out of the expectation and then uses the fact that $\mb(\Phi)$ forms a Markov blanket for $\Phi$, thus we can drop the conditioning on $R$ without loosing anything, and $(d)$ expands the inner expectation over $\Phi$ and rewrites the log gradient. This idea of taking a local expectation conditioned on a Markov blanket is similar to the LEG estimator proposed by~\citet{titsias2015local}. However, it is not immediately obvious how to compute this estimator efficiently. \citet{titsias2015local} provide a more explicit expression and empirical results for a fully factorized variational distribution. Here, we will go beyond this case to provide a computationally efficient way to compute the inner expression for more general networks of stochastic units. To begin, we apply Theorem 1 from Chapter 4 of the probabilistic reasoning textbook of~\citet{pearl1988probabilistic}, which implies that
\begin{equation}\label{markov_blanket_conditional_expression}
\P(\Phi=\phi|\mb(\Phi))=\rho_\Phi(\phi)\pi_\Phi(\phi|\pa(\Phi)).
\end{equation}
where $\rho_\Phi(\phi)=\frac{\prod\limits_{C\in\ch(\Phi)}\pi_C(C|\pa(C)\setminus\Phi,\Phi=\phi)}{\sum\limits_{\phi'}\pi_\Phi(\phi'|\pa(\Phi))\prod\limits_{C\in\ch(\Phi)}\pi_C(C|\pa(C)\setminus\Phi,\Phi=\phi')}$. Intuitively, $\rho_\Phi(\phi)$ is the relative counterfactual probability of the children of $\Phi$ taking the value they did had $\Phi$ been fixed to $\phi$. See Appendix~\ref{markov_blanket_conditional} for a full proof. Substituting this result into the expression within the expectation in Equation~\ref{HNCA_derivation_1}, we get that the following is an unbiased estimator of $\frac{\partial\E[R]}{\partial\theta_\Phi}$:
\begin{equation}\label{HNCA_estimator}
\hat{G}^{\text{HNCA}}_{\Phi}\defeq\sum_\phi\rho_\Phi(\phi)\frac{\partial\pi_\Phi(\phi|\pa(\Phi))}{\partial\theta_\Phi}R,
\end{equation}
which we call the HNCA gradient estimator. Equation~\ref{HNCA_estimator} applies only to units for which $\ch(\Phi)\neq\emptyset$ and thus excludes the output unit $A$. In our contextual bandit experiments, we use the REINFORCE estimator $\hat{G}^{\text{RE}}_\Phi(\phi)$ for the output unit, in Section~\ref{extended} we show how to improve upon this if we have access to the reward function.
HNCA assigns credit to a particular output choice $\phi$ based on the relative counterfactual probability of its children's outputs had $\phi$ been chosen, independent of the actual value of $\Phi$. Intuitively, this reduces variance, because each potential output choice of a given unit will get credit proportional to the difference it makes further downstream. On the other hand, REINFORCE credits whatever output happens to be selected, whether it makes a difference or not. This intuition is formalized in the following theorem:
\begin{theorem}\label{reduced_variance}
$\Var(\hat{G}^{\text{HNCA}}_{\Phi})\leq \Var(\hat{G}^{\text{RE}}_{\Phi})$, where $\Var(\vec{X})$ stand for the elementwise variance of random vector $\vec{X}$, and the inequality holds elementwise.
\end{theorem}
Theorem~\ref{reduced_variance} follows from the law of total variance by the proof available in Appendix~\ref{HNCA_action_value_low_var}.
\subsection{Efficient Implementation of HNCA}
We implement HNCA as a message-passing procedure. A forward pass propagates information from parents to children to compute the network output. A backward pass passes information from children to parents to compute the HNCA gradient estimator. The computational complexity of this procedure depends on how difficult it is to compute the numerators of $\rho_\Phi(\phi)$. We could naively recompute $\pi_C(C|\pa(C)\setminus\Phi,\Phi=\phi)$ from scratch for each possible $\phi$. When $C$ corresponds to a Bernoulli unit, which computes its output probability as a linear function of its inputs followed by sigmoid activation, this would require time $\mathcal{O}(|\pa(C)|N_{\Phi})$, where $N_{\Phi}$ is the number of possible values $\Phi$ can take (2 if $\Phi$ is also Bernoulli). To do this for every parent of every unit in a Bernoulli network would thus require $\mathcal{O}(2\sum_\Phi|\pa(\Phi)|^2)$. This is much greater than the cost of a forward pass, which takes on the order of the total number of edges in the network, or $\mathcal{O}(\sum_\Phi|\pa(\Phi)|)$. This contrasts with backpropagation where the cost of the backward pass is on the same order as the forward pass, an appealing property, which implies that learning is not a bottleneck.
\begin{figure}[htb]
\begin{algorithm}[H]
\begin{algorithmic}[1]
\STATE Receive $\vec{x}$ from parents\\
\STATE $l=\vec{\theta}\cdot\vec{x}+b$\\
\STATE $p=\sigma(l)$\\
\STATE $\phi\sim \textit{Bernoulli}(p)$\\
\STATE Pass $\phi$ to children\\
\STATE Receive $\vec{q}_1,\vec{q}_0,R$ from children\\
\STATE $q_1=\prod_{i}\vec{q}_1[i]$\\
\STATE $q_0=\prod_{i}\vec{q}_0[i]$\\
\STATE $\bar{q}=pq_1+(1-p)q_o$\\
\STATE $\vec{l}_1=l+\vec{\theta}\odot(1-\vec{x})$\\
\STATE $\vec{l}_0=l-\vec{\theta}\odot\vec{x}$\\
\STATE $\vec{p}_1=(1-\phi)(1-\sigma(\vec{l}_1))+\phi\sigma(\vec{l}_1)$\\
\STATE $\vec{p}_0=(1-\phi)(1-\sigma(\vec{l}_0))+\phi\sigma(\vec{l}_0)$\\
\STATE Pass $\vec{p}_1,\vec{p}_0,R$ to parents\\
\STATE $\vec{\theta}=\vec{\theta}+\alpha\sigma^{\prime}(l)\vec{x}\left(\frac{q_1-q_0}{\bar{q}}\right)R$\\
\STATE $b=b+\alpha\sigma^{\prime}(l)\left(\frac{q_1-q_0}{\bar{q}}\right)R$
\end{algorithmic}
\caption{HNCA (Bernoulli unit)}\label{bernoulli_alg}
\end{algorithm}
\addtocounter{algorithm}{-1}
\captionof{algorithm}{The forward pass in lines 1-5 takes input from the parents, uses it to compute the fire probability $p$ and samples $\phi\in\{0,1\}$. The backward pass receives two vectors of probabilities $\vec{q}_1$ and $\vec{q}_0$, each with one element for each child. Each element represents $\vec{q}_{0/1}[i]=\P\left(C_i\middle|\pa(C_i)\setminus\Phi,\Phi=0/1\right)$ for a given child $C_i\in \ch(\Phi)$. Lines 7 and 8 take the product of child probabilities to compute $\prod_i\pi_{C_i}(C_i|\pa(C_i)\setminus\Phi,\Phi=0/1)$. Line 9 computes the associated normalizing factor. Line 10-13 use the logit $l$ to efficiently compute a vector of probabilities $\vec{p}_1$ and $\vec{p}_0$. Each element corresponds to a counterfactual probability of $\phi$ if a given parent's value was fixed to 1 or 0. Here $\odot$ represents the elementwise product. Line 14 passes information to the unit's children. Lines 15 and 16 finally update the parameter using $\hat{G}^{\text{HNCA}}_{\Phi}$ with learning-rate hyperparameter $\alpha$.}
\end{figure}
Luckily, we can improve on this for cases where $\pi_C(C|\pa(C)\setminus\Phi,\Phi=\phi)$ can be computed from $\pi_C(C|\pa(C))$ in less time than computing $\pi_C(C|\pa(C))$ from scratch. This is indeed the case for linear Bernoulli units, for which the policy can be written $\pi_\Phi(\phi|\vec{x})=\sigma(\vec{\theta}\cdot\vec{x}+b)$ where $\vec{x}$ is the binary vector consisting of all parent outputs, $b$ is a scalar bias, $\vec{\theta}$ is the parameter vector for the unit, and $\sigma$ is the sigmoid function. Say we wish to compute the counterfactual probability of $\Phi=1$ given $\vec{x}[i]=1$, if we already have $\pi_\Phi(1|\vec{x})$. Regardless of the actual value of $\vec{x}_i$ we can use the following identity:
\begin{equation*}
\pi_\Phi(1|\vec{x}\setminus \vec{x}[i],\vec{x}[i]=1)=\sigma(\sigma^{-1}(\pi_\Phi(1|\vec{x}))+\vec{\theta}[i](1-\vec{x}[i])).
\end{equation*}
This requires only constant time, whereas computing $\pi_\Phi(\phi|\vec{x})$ requires time proportional to the length of $\vec{x}$. This simple idea is crucial for implementing HNCA efficiently. In this case, we can compute the numerator terms for every unit in a Bernoulli network in $\mathcal{O}(\sum_\Phi|\pa(\Phi)|)$ time. This is now on the same order as computing a forward pass through the network. Computing $\hat{G}_{\Phi}^{\text{HNCA}}$ for a given $\Phi$ from these numerator terms requires multiplying a scalar by a gradient vector with the same size as $\theta_\Phi$. For a Bernoulli unit, $\theta_\Phi$ has $\mathcal{O}(|\pa(\Phi)|)$ elements, so this operation adds another $\mathcal{O}(\sum_\Phi|\pa(\Phi)|)$, maintaining the same order of complexity.
Algorithm~\ref{bernoulli_alg} shows an efficient implementation of HNCA for Bernoulli units. Note that, for ease of illustration, the pseudocode is implemented for a single unit and a single training example at a time. In practice, we use a vectorized version which works with vectors of units that constitute a layer, and with minibatches of training data.
In Section~\ref{bandit_experiments}, we will apply HNCA to a model consisting of a number of hidden layers of Bernoulli units followed by a softmax output layer. Appendix~\ref{softmax_alg} provides an implementation and discussion of HNCA for a softmax output unit. Note that the output unit itself uses the REINFORCE estimator in its update, as it has no children, which precludes the use of HNCA. Nonetheless, the output unit still needs to provide information to its parents, which do use HNCA. Using a softmax unit at the output, we can still maintain the property that the time required for the backward pass is on the same order as the time required for the forward pass. If, on the other hand, the entire network consisted of softmax nodes with $N$ choices each, the HNCA backward pass would require a factor of $N$ more computation than the forward pass, we discuss this in Appendix~\ref{softmax_alg} as well.
\subsection{Contextual Bandit Experiments}
\label{bandit_experiments}
\begin{figure}[htb]
\includegraphics[width=\columnwidth]{img/contextual_bandit_plots.pdf}
\caption{Training stochastic networks on a contextual bandit version of MNIST. Each line represents the average of 5 random seeds with error bars showing $95\%$ confidence interval. Final values (train accuracy for the left plots) at the end of training are written beside each line. The left column shows the online training accuracy (or equivalently the average reward) as a dotted line, and the test accuracy as a solid line (though they essentially overlap). The right column shows the natural logarithm of the mean gradient variance. Mean gradient variance is computed as the mean of the per-parameter empirical variance over examples in a training batch of $50$. We find that, for each network depth, HNCA drastically reduces gradient variance, resulting in significantly improved performance on this task.}
\label{fig:contextual_bandit_plots}
\end{figure}
We evaluate HNCA against REINFORCE in terms of gradient variance and performance on a contextual bandit version of MNIST~\citep{lecun2010mnist}, with the standard train test split. Following~\citet{dong2020disarm}, input pixels are dynamically binarized, meaning that at each epoch they are randomly fixed to $0$ or $1$ with probability proportional to their intensity. For each training example, the model outputs a prediction and receives a reward of $1$ if correct and $0$ otherwise. We use a fully connected, feedforward network with 1, 2 or 3 hidden layers, each with 200 Bernoulli units, followed by a softmax output layer. We train using ADAM optimizer~\citep{kingma2014adam} with a learning rate fixed to $10^{-4}$ and batch-size of $50$ for $100$ epochs. Learning rate and layer size hyperparameters follow~\citet{dong2020disarm} for simplicity. We map the output of the Bernoulli units to one or negative one, instead of one or zero, as we found this greatly improved performance in preliminary experiments. We report results for HNCA and REINFORCE, both with and without an exponential moving average baseline subtracted from the reward. We use a discount rate of $0.99$ for the moving average.
Figure~\ref{fig:contextual_bandit_plots} shows the results, in terms of performance and gradient variance, for gradient estimates generated by HNCA and REINFORCE. We find that HNCA provides drastic improvement in terms of both gradient variance and performance over REINFORCE. Note that performance degrades with number of layers for both estimators, reflecting the increasing challenge of credit assignment. Subtracting a moving average baseline generally improves performance of both algorithms, except for HNCA in the single hidden layer case. The comparison between the two algorithms is qualitatively similar whether or not a baseline is used.
In Appendix~\ref{nonlinear_bandit}, we demonstrate that HNCA can also be used to efficiently train a stochastic layer as the final hidden layer of an otherwise deterministic network, this could be useful, for example, for learning a binary representation.
\section{Optimizing a Known Function}\label{extended}
In Section~\ref{bandit}, we introduced HNCA in a setting where the reward function was unknown, and dependent only on the input context and the output of the network as a whole. Here, we extend HNCA to optimize the expectation of a known function $f$, which may have direct dependence on every unit. We refer to this extension simply as $f$-HNCA. This setting is more in line with the setting explored by~\citet{titsias2015local}, and $f$-HNCA is distinguished from LEG mainly by its computationally efficient message passing implementation, which in turn facilitates its application to multi-layer stochastic networks.
We assume the function $f=\sum_if^i$ is factored into a number of \textit{function components} $f^i$, which we index by $i$ for convenience. This factored structure has two benefits, the first is computational. In particular, it will allow us to compute counterfactual values for each component with respect to changes to its input separately. The second is for variance reduction by realizing that we only need to assign credit to function components that lie downstream of the unit being credited. A similar variance reduction approach is also used by the NVIL algorithm of~\citet{mnih2014neural}.
Each function component $f^i$ is a deterministic function of a subset of the outputs of units in the network, as well as possibly depending directly on some parameters. Thus, $f^i=f^i(\widetilde{\pa}(f^i);\theta^i)$, where $\theta^i$ is a set of real valued parameters which may overlap with the parameters $\theta_\Phi$ for some subset of units in the network, and $\widetilde{\pa}(f^i)$ is the set of nodes in the network which act as input to $f^i$. Formally, $f^i$ without arguments will refer to the random variable corresponding to the output of the associated function. We use the notation $\widetilde{\pa}$, distinct from $\pa$, to make it clear that function components are not considered nodes in the network.
The goal in this setting is to estimate the gradient of $\E[f]$, so that we can maximize it by gradient ascent. By linearity of expectation, we can define unbiased estimators for each $\frac{\partial\E[f^i]}{\partial\theta_\Phi}$ separately and sum over $i$ to get an unbiased estimator of the full gradient.
\subsection{HNCA with a Known Function}\label{GHNCA}
We now discuss how to extend the HNCA estimator to construct an estimator of $\frac{\partial\E[f]}{\partial\theta_\Phi}$ for a particular unit $\Phi$ and function component in this setting. We begin by considering the gradient for a single function component $\frac{\partial\E[f^i]}{\partial\theta_\Phi}$. First, note that we can break the gradient into indirect and direct dependence on $\theta_\Phi$:
\begin{equation}\label{gradient_decomp}
\frac{\partial\E[f^i]}{\partial\theta_\Phi}=\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}f^i\right]+\E\left[\frac{\partial f^i}{\partial\theta_\Phi}\right].
\end{equation}
The direct gradient $\frac{\partial f^i}{\partial\theta_\Phi}$ is zero unless $\theta_\Phi\in\theta^i$, in which case it can be computed directly given we assume access to $f^i$. From this point on, we will focus on the left expectation.
The main added complexity in estimating $\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}f^i\right]$, compared to the contextual bandit case, arises if $f^i$ has a direct functional dependence on $\Phi$. In this case we can no longer assume that $f^i$ is separated from $\Phi$ by $\mb(\Phi)$. Luckily, this is straightforward to patch. Let $f^i_\Phi(\phi)$ be the random variable defined by taking the function $f^i(\widetilde{\pa}(f^i);\theta^i)$ and substituting the specific value $\phi$ instead of the random variable $\Phi$ into the arguments while keeping all other $\widetilde{\pa}(f^i)$ equal to the associated random variables. By design, $f^i_\Phi(\phi)$ is independent of $\Phi$ given $\mb(\Phi)$, which allows us to define the following unbiased estimator for $\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}f^i\right]$ (see Appendix~\ref{Generalized_HNCA_derivation} for the full derivation):
\begin{equation}\label{Generalized_HNCA_i}
\hat{G}^{\text{$f$-HNCA},i}_\Phi(\phi)\defeq\sum_\phi\rho_\Phi(\phi)\frac{\partial\pi_\Phi(\Phi|\pa(\Phi))}{\partial\theta_\Phi}f^i_\Phi(\phi),
\end{equation}
where $\rho_\Phi(\phi)$ is as in Equation~\ref{markov_blanket_conditional_expression}. As $\rho_\Phi(\phi)$ is defined with respect to $\ch(\Phi)$, this estimator is only applicable if $\Phi$ has children (i.e. $\ch(\Phi)\neq\emptyset$). In fact, even if $\Phi$ has children, we can ignore them if they have no downstream connection\footnote{More generally, if only a subset of $\ch(\Phi)$ lies in $\widetilde{\an}(f^i)$ we can replace $\ch(\Phi)$ in $\rho_\Phi(\phi)$ with $\ch^i(\Phi)=(\ch(\Phi)\cap\widetilde{\an}(f^i))$, but, we will not use this in our experiments in this work.} to $f^i$, as such children cannot influence $f^i$. Thus if $\ch(\Phi)\cap\widetilde{\an}(f^i)=\emptyset$ we instead define $\hat{G}^{\text{$f$-HNCA},i}_\Phi(\phi)\defeq\sum_\phi\frac{\partial\pi_\Phi(\Phi|\pa(\Phi))}{\partial\theta_\Phi}f^i_\Phi(\phi)$. In Appendix~\ref{Generalized_HNCA_low_variance}, we extend Theorem~\ref{reduced_variance} to apply to $f$-HNCA, showing that using $\hat{G}_{\Phi}^{\text{$f$-HNCA},i}$ results in a variance reduced estimator for $\E\left[\frac{\partial\log(\pi_\Phi(\Phi|\pa(\Phi)))}{\partial\theta_\Phi}f^i\right]$ compared to REINFORCE. The full $f$-HNCA gradient estimator is defined by summing up these components and accounting for any direct functional dependence of $f$ on network parameters:
\begin{equation}\label{generalized_HNCA}
\hat{G}_{\Phi}^{\text{$f$-HNCA}}\defeq\begin{multlined}[t]\sum_\phi\frac{\partial\pi_\Phi(\phi|\pa(\Phi))}{\partial\theta_\Phi}\Biggl(\rho_\Phi(\phi)\smashoperator{\sum_{i:\ch(\Phi)\cap\widetilde{\an}(f^i)\neq\emptyset}}f^i_\Phi(\phi)\hspace{15pt}+\\
\smashoperator{\sum_{i:\ch(\Phi)\cap\widetilde{\an}(f^i)=\emptyset}}f^i_\Phi(\phi)\hspace{10pt}\Biggr)+\sum_i\frac{\partial f^i}{\partial\theta_\Phi}.
\end{multlined}
\end{equation}
If $\Phi\not\in\widetilde{\an}(f^i)$ then $\frac{\partial\E[f^i]}{\partial\theta_\Phi}=\E\left[\frac{\partial f^i}{\partial\theta_\Phi}\right]$ as $\Phi$ cannot influence something with no downstream connection. Hence, in the two leftmost sums over $i$ in Equation~\ref{generalized_HNCA}, we implicitly only sum over $i$ such that $\Phi\in\widetilde{\an}(f^i)$.
In addition to the efficiency of computing counterfactual probabilities, for $f$-HNCA, we have to consider the efficiency of computing counterfactual function components $f^i_\Phi(\phi)$ given $f^i$. For function components with no direct connection to a unit $\Phi$, this is trivial as $f^i_\Phi(\phi)=f^i$. If $f^i$ is directly connected, then implementing $f$-HNCA with efficiency similar to HNCA will require that we are able to compute $f^i_\Phi(\phi)$ from $f^i$ in constant time. This is the case if $f^i$ is a linear function followed by some activation. For example, functions of the form $f^i=\log(\sigma(\vec{\theta}\cdot\vec{x}+b))$ which will appear in the ELBO function used in our variational auto-encoder (VAE; \citet{kingma2013auto,rezende2014stochastic}) experiments. More algorithmic details can be found in Appendix~\ref{VAE_implementation}.
\subsection{Variational Auto-encoder Experiment}\label{VAE_experiments}
\definecolor{purple}{HTML}{7570b3}
\definecolor{orange}{HTML}{d95f02}
\definecolor{pink}{HTML}{e7298a}
\definecolor{green}{HTML}{66a61e}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{img/generative_model.pdf}
\caption{An illustration of the ELBO for a 3 layer discrete hierarchical VAE broken down into function components for $f$-HNCA. $\vec{X}$ is the input to be encoded, each additional circle is the latent state from a layer of the encoder network. Each rectangle is a set of function components which contribute to the ELBO. The parameters of the encoder are trained to maximize the ELBO by $f$-HNCA. Consider the $f$-HNCA gradient estimator for $\vec{\Phi}_1$. The function components $H(q_1)$, marked in \textcolor{purple}{purple} are upstream of $\vec{\Phi}_1$, however, $H(q_1)$ depends directly on $\theta_{q_1}$ and thus $\frac{\partial H(q_1)}{\partial \theta_{q_1}}$ is nonzero, so the entire contribution of $H(q_1)$ to the gradient estimate $\hat{G}^{\text{$f$-HNCA}}_\Phi$ will come from this gradient. The function components marked in \textcolor{green}{green} have only direct connection with $\vec{\Phi}_1$, so they will receive credit via $\hat{G}^{\text{$f$-HNCA},i}_\Phi(\phi)\defeq\sum_\phi\frac{\partial\pi_\Phi(\Phi|\pa(\Phi))}{\partial\theta_\Phi}f^i_\Phi(\phi)$. The function components marked in \textcolor{orange}{orange} have both direct connections and downstream connections mediated by $\vec{\Phi}_2$, so they will receive credit via Equation~\ref{Generalized_HNCA_i}. Finally, the variables marked in \textcolor{pink}{pink} have only mediated connections to $\vec{\Phi}_1$ through $\vec{\Phi}_2$, so $f^i_\Phi(\phi)=f^i$, the estimator for these variables essentially reduces to the original HNCA estimator defined in Equation~\ref{HNCA_estimator}.}
\label{fig:generative_model}
\end{figure}
Here, we demonstrate how the $f$-HNCA approach described in Section~\ref{GHNCA} can be applied to the challenging task of training a discrete hierarchical VAE. Consider a VAE consisting of a generative model (decoder) $p$ and an approximate posterior (encoder) $q$, each of which consist of $L$ discrete stochastic layers. Samples $\vec{X}$ are generated by $p$ as
\begin{equation*}
\vec{X}\sim p_0(\vec{X}|\vec{\Phi}_1), \vec{\Phi}_1\sim p_1(\vec{\Phi}_1|\vec{\Phi}_2), ..., \vec{\Phi}_L\sim p_L(\vec{\Phi}_L),
\end{equation*}
while $q$ approximates the posterior $\P(\vec{\Phi}_L|X)$ as a distribution which can be sampled as
\begin{equation*}
\Phi_L\sim \begin{multlined}[t]q_L(\vec{\Phi}_L|\vec{\Phi}_{L-1}), \vec{\Phi}_{L-1}\sim q_{L-1}(\vec{\Phi}_{L-1}|\vec{\Phi}_{L-2}),...\\
,\vec{\Phi}_1\sim q_1(\vec{\Phi}_1|\vec{X}),
\end{multlined}
\end{equation*}
where, each $p_i$ and $q_i$ represents a vector of Bernoulli distributions, each parameterized as a linear function of their input (except the prior $p_L(\vec{\Phi}_L)$ which takes no input, and is simply a vector of Bernoulli variables with learned means). Call the associated parameters $\theta_{p_i}$ and $\theta_{q_i}$ We can train such a VAE by maximizing a lower bound on the log-likelihood of the training data, usually referred to as the evidence lower bound (ELBO) which we can write as $\E[f_E]$ where
\begin{equation}\label{ELBO_reward}
\begin{multlined}[b]
f_E\defeq\log(p_0(\vec{X}|\vec{\Phi}_1))+\sum_{l=1}^{L-1}\log(p_l(\vec{\Phi}_l|\vec{\Phi}_{l-1}))+\\\log(p_L(\vec{\Phi}_L))+H(q_1(\cdot|\vec{X}))+\sum_{l=1}^{L-1}H(q_{l+1}(\cdot|\vec{\Phi}_l))
\end{multlined},
\end{equation}
where $H$ is the entropy of the distribution, and the expectation is taken with respect to the encoder $q$ and random samples $\vec{X}$. Each $\vec{\Phi}_i$ is sampled from the associated encoder $q_i$. Note that each term in Equation~\ref{ELBO_reward} is a sum over elements in the associated output vector, we can view each element as a particular function component $f^i$. The resulting compute graph is illustrated in Figure~\ref{fig:generative_model}.
We compare $f$-HNCA with REINFORCE and several stronger methods for optimizing an ELBO of a VAE trained to generate MNIST digits. We focus on strong, unbiased, variance reduction techniques from the literature that do not require modifying the architecture or introduce significant additional hyperparameters. Since HNCA falls into this category, this allows for straightforward comparison without the additional nuance of architectural and hyperparameter choices. Specifically, we compare HNCA with REINFORCE leave one out (REINFORCE LOO; \citet{kool2019buy}) and DisARM~\citep{dong2020disarm}. Note that in the multi-layer case, both DisARM and REINFORCE LOO require sampling an additional partial forward pass beginning from each layer, which gives them a quadratic scaling in compute cost with the number of layers. By contrast, HNCA requires only a single forward pass and a backward pass of similar complexity.
Initially, we found that $f$-HNCA outperformed the other tested methods in the single layer discrete VAE case, but fell short in the multi-layer case. However, we found that a simple modification that subtracts a layer specific scalar baseline, similar to that used by~\citet{mnih2014neural}, significantly improved the performance of $f$-HNCA in the multi-layer case. Specifically, for each layer, we maintain a scalar running average of the sum of those components of f with mediated connections (those highlighted in pink and orange in Figure 2) and subtract it from the leftmost sum over i in Equation 6 to produce a centered learning signal.\footnote{Using such a baseline for components without mediated connections would analytically cancel.} We use a discount rate of $0.99$ for the moving average.\footnote{We used the first value we tried, we did not tune it.} We refer to this variant as $f$-HNCA with Baseline. We also tested subtracting a moving average of all downstream function components in REINFORCE to understand how much this change helps on its own. It's not obvious how to implement such a running average baseline for the other tested methods given they already utilize alternative methods to center the learning signal, thus a naive moving average baseline would have expectation zero.
As in Section~\ref{bandit_experiments}, we use dynamic binarization and train using ADAM optimizer with learning rate $10^{-4}$ and batch-size $50$. Following~\citet{dong2020disarm}, our decoder and encoder each consist of a fully connected, stochastic feedforward neural network with 1, 2 or 3 layers, each hidden layer has 200 Bernoulli units. We train for $840$ epochs, approximately equivalent to the $10^6$ updates used by~\citet{dong2020disarm}. For consistency with prior work, we use Bernoulli units with a zero-one output. For all methods, we train each unit based on downstream function components, as opposed to using the full function $f$. See Appendix~\ref{VAE_exp_details} for more implementation details.
Figure~\ref{fig:generative_model_plots}, shows the results in terms of ELBO and gradient variance, for gradient estimates generated by $f$-HNCA and the other methods tested. As in the contextual bandit case, we find that $f$-HNCA provides drastic improvement over REINFORCE. $f$-HNCA also provides a significant improvement over all other methods for the single-layer discrete VAE, but underperforms the other strong methods in the multi-layer case. On the other hand, $f$-HNCA with Baseline significantly improves on the other tested methods in all cases. REINFORCE with baseline outperforms ordinary $f$-HNCA in the multi-layer cases. Hence, this baseline subtraction is a fairly powerful variance reduction technique for REINFORCE, with strong complementary benefits with $f$-HNCA. In Appendix~\ref{test_set_ELBOs}, we additionally report multi-sample test-set ELBOs for the final trained networks, which reflect the same performance ordering as the training set ELBOs. In Appendix~\ref{ablation}, we perform an ablation experiments on $f$-HNCA with Baseline and find that the choice of whether to exclude children when $\ch(\Phi)\cap\widetilde{\an}(f^i)=\emptyset$ has a significant performance impact, while the additional impact of excluding upstream function components is fairly minimal.
\begin{figure}[htb]
\includegraphics[width=\columnwidth]{img/generative_model_plots.pdf}
\caption{Training discrete VAEs to generate MNIST digits. Each line represents the average of 5 random seeds with error bars showing $95\%$ confidence interval. Final values at the end of training are written near each line in matching color. The left column shows the online training ELBO. The right column shows the natural logarithm of the mean encoder gradient variance. Mean gradient variance is computed as the mean over parameters and batches of the per-parameter empirical variance over examples in a training batch of $50$. $f$-HNCA outperforms all other tested methods in the single-layer case, but underperforms in the multi-layer cases. $f$-HNCA with Baseline outperforms the other methods in the multi-layer case. $f$-HNCA with baseline is excluded from the single layer results as there are no mediated connections.}
\label{fig:generative_model_plots}
\end{figure}
\section{Discussion and Conclusion}
We introduced HNCA, an algorithm for gradient estimation in networks of discrete stochastic units. HNCA is inspired by Hindsight Credit Assignment~\citep{harutyunyan2019hindsight}, and can be seen as an instance of Local Expectation Gradients, extending the work of~\citet{titsias2015local} by providing a computationally efficient message passing algorithm and extension to multi-layer networks of stochastic units. Our computational efficient approach directly addresses concerns in the literature that LEG is inherently computationally expensive~\citep{tucker2017rebar,mnih16vimco}. We prove that HNCA is unbiased, and that it reduces variance compared to REINFORCE. Empirically, we show that HNCA outperforms strong methods for training a single-layer Bernoulli VAE, and when subtracting a simple moving average baseline also outperforms the same methods for the case of a multi-layer Hierarchical VAE.
It's worth highlighting that efficient implementation of HNCA is predicated on the ability to efficiently compute counterfactual probabilities or function components when a single input is changed. This is not always possible, for example, if $f$ is the result of a multi-layer deterministic network. An example of this situation is the nonlinear discrete VAE architecture explored by \citet{dong2020disarm} and \citet{yin2019arm} where the encoder and decoder are nonlinear networks with a single stochastic Bernoulli layer at the outputs. However, as we show in Appendix~\ref{nonlinear_bandit}, HNCA can be used to train a final Bernoulli hidden layer at the end of a nonlinear network.
In addition to optimizing a known function of the output of a stochastic network, we show in Section~\ref{bandit} that HNCA can be applied to train the hidden layers of a multi-layer discrete network in an online learning setting with unknown reward function. REINFORCE LOO and DisARM, which rely on the ability to evaluate the reward function multiple times for a single training example, cannot.
Future work could explore combining HNCA with other methods for complimentary benefits. One could also explore extending HNCA to propagate credit multiple steps which would presumably allow further variance reduction, but presents challenges as the relationships between more distant nodes in the network becomes increasingly complex.
HNCA provides insight into the challenges of credit assignment in discrete stochastic compute graphs, which has the potential to have an impact on future approaches.
\section*{Acknowledgments}
The author thanks Rich Sutton, Matt Taylor and Tian Tian for useful conversations, and anonymous reviewers for useful feedback. I also thank the Natural Sciences and Engineering Research Council of Canada and Alberta Innovates for providing funding for this work.
|
\section{Introduction}
Large-scale neuronal oscillations emerge due to the synchronous interplay
of ensembles of neurons. These oscillations are successfully replicated by
mathematical models of spiking neurons,
which also allow for a mechanistic understanding of neuronal rhythmogenesis~\cite{Wan10,Boe17}.
According to these theories, inhibitory synapses play a central role in setting up
neuronal synchronization either in isolation~\cite{WB96} or due to their
interplay with excitatory neurons~\cite{WC72}.
Additionally, inhibitory cells are very often coupled electrically, and this
coupling is usually mediated by so-called gap junctions~\cite{NPR18}.
Such electrical synapses are well-known to largely favor synchrony.
Recently, important efforts have been put forward to model the oscillatory dynamics of
so-called whole-brain networks~\cite{DTB+15,LB19,GCC+21}.
To facilitate both the analysis and the computational work many studies do not
use spiking neuron models, but apply the mathematical framework of the
Kuramoto model (KM);
see, e.g., Refs.~\onlinecite{GCC+21,LB19,BHD10,CHS+11,VMM14,PDH+15,SKS+15,SLR+15,PSP+16,ARM+16,PPJ18,RGA+19,
CM19,ZZV+20,NPG+20,JEP21,PPM+21,OKD21,PGR21,WDB+21}.
Yet, it remains unclear how to relate the parameters of the KM to bio-physically
meaningful parameters, such as synaptic strengths.
In this paper, we aim to theoretically substantiate the use of the KM for
neuronal modeling, by providing a mathematical link between
the quadratic integrate-and-fire (QIF) model and the KM.
We derive a KM for QIF neurons and subsequently justify its validity in two different ways:
First, we compare the predictions of the KM with
those of an exact mean-field model for QIF neurons---often
referred to as firing rate or neural mass model (NMM)~\cite{PDR+19,MP20,MPR15}.
Second, we use two populations of \emph{identical} Kuramoto
oscillators to find so-called chimera states~\cite{MKB04,AMS+08,Lai09,MBP16}.
In a chimera state, one of the two homogeneous populations displays in-phase synchrony, and
NMMs are useless in this case.
However, the KM for QIF neurons is perfectly suited to describe full synchrony,
and we exploit this to uncover the existence of chimera states in two-population networks
of QIF neurons.
Our derivation of the KM for QIF neurons
mainly builds on a previous work by Izhikevich~\cite{Izh07}
and also on Refs.~\onlinecite{PM14,MP18}. In Chap. 10 of Ref.~\cite{Izh07},
Izhikevich applied perturbation methods to derive
a simplified model that approximated the dynamics of two
identical QIF neurons, with \emph{either} chemical or electrical coupling
\footnote{More recently a similar derivation has been obtained
for populations of heterogeneous Winfree oscillators with
sinusoidal infinitesimal Phase Resetting Curves (iPRC)~\cite{PM14}.
When the oscillators have the iPRC of the QIF neuron, the
population of Winfree oscillators can be well approximated to
to a population of QIF neurons with (weak) chemical synapses~\cite{MP18}.}.
Here, we extend the work of Izhikevich and derive a model that
approximates the dynamics of an \emph{ensemble of heterogeneous} QIF neurons with
\emph{both} chemical and electrical coupling.
The approximated model turns out to be a well-known version of the
KM~\cite{SSK88,SK86} and is valid when both heterogeneities
and coupling strengths are weak.
This paper is organized as follows: In Section II, we introduce the QIF population model,
and in Section III, we describe the method to reduce the QIF model to the KM. In Section IV,
we analyze the dynamics of the KM and demonstrate that it correctly describes the
collective dynamics of populations of nearly identical QIF neurons, with weak electrical and
chemical synapses. In Section V, we exploit the KM to uncover the presence of
chimera states in coupled populations of identical QIF neurons. Finally, in Section IV,
we briefly discuss and summarize our results.
\section{Population of QIF neurons with electrical and chemical
synapses}
We investigate a population of $N$
quadratic integrate-and-fire (QIF) neurons $i=1,\dots,N$ interacting all-to-all via
both electrical and chemical synapses~\cite{KE04,Lai15,PDR+19}
\begin{equation}
\tau \dot V_i = V_i^2+\eta_i +\epsilon ~ I_{i,syn}(t), ~~\text{if } V_i>V_p, \text{ then } V_i \leftarrow V_r,
\label{qif}
\end{equation}
where $V_i$ is the membrane potential of neuron $i$,
$\tau$ is the membrane time constant of the neurons, and $\eta_i$ represents an
external current, which varies from cell to cell.
Due to the quadratic nonlinearity of the
QIF model, the membrane potential blows up in finite time, and a resetting
rule is needed: When the neurons reach the peak value $V_p$, they emit a spike and the
voltage is reset to $V_r$. We assume symmetric
spike resetting, $V_p=-V_r$ and $V_p \to \infty$, so that the QIF model
is equivalent to the so-called theta-neuron~\cite{EK86,Erm96}.
In addition, we consider $\eta_i>0$ and hence, in the absence of
synaptic inputs ($I_{j,syn}=0$), QIF neurons are self-sustained oscillators.
Finally, synaptic inputs (whose total strength is controlled by the \emph{small}
parameter $\epsilon\geq 0$) are composed of electrical and chemical synapses,
\begin{equation}
I_{i,syn}(t) = g (v(t)- V_i) + J \tau r(t).
\label{Isyn}
\end{equation}
Specifically, electrical synapses (of strength $\epsilon g\geq 0$) diffusively couple
each neuron with the mean membrane potential
\begin{equation}
v (t)= \frac{1}{N} \sum_{j=1}^{N} V_j(t).
\label{v}
\end{equation}
Electrical synapses mostly connect inhibitory neurons,
and hence the chemical synaptic strength, $J$,
is thought of as a negative parameter thereafter.
Finally, chemical synapses (of strength $\epsilon J$) are mediated by the mean firing rate
\begin{equation}
r (t)= \frac{1}{N} \sum_{j=1}^N\sum_k \delta\left(t- t_j^{(k)}\right),
\label{r}
\end{equation}
where $t_j^{(k)}$ is the time of the $k$th spike of the $j$th neuron and $\delta(t)$ is the Dirac delta function.
\section{Derivation of the Kuramoto model for populations of QIF neurons}
In the following we derive the Kuramoto model corresponding to Eq.~\eqref{qif}.
The derivation exploits well-known mathematical methods that are reviewed for example
in Refs.~\onlinecite{HI97,Izh07,PD19}.
We perform the derivation of the KM as follows: First, we obtain the
phase resetting curve (PRC) of the QIF model. Second, invoking weak coupling, we
derive the so-called Winfree model corresponding to the QIF model.
Finally, we assume weak heterogeneity and apply
the method of averaging to obtain the Kuramoto model corresponding to
Eq.~\eqref{qif}.
\subsection{Phase resetting curve (PRC) of a QIF neuron}
We first consider an isolated, regularly spiking QIF neuron; i.e.,~$I_{i,syn}=0$ and $\eta_i > 0$.
The solution of the QIF model immediately after a spike, $V_i(0)=-\infty$, is
\begin{equation}
V_i(t)=\sqrt{\eta_i} \tan \left( t\sqrt{\eta_i}/\tau -\pi/2\right).
\label{sol}
\end{equation}
The frequency of the oscillations is
\begin{equation}
\Omega_i=2 \sqrt{\eta_i}/\tau,
\label{Omega}
\end{equation}
and a phase variable $\theta_i$ can be defined in the interval $[0,2\pi)$ as
\begin{equation}
\theta_i=\Omega_i t=2\arctan( V_i/\sqrt{\eta_i})+\pi.
\label{phase}
\end{equation}
Next, we assume that the neuron is perturbed so that its membrane
potential instantaneously changes from $V_i$ to $V_i+\delta V$.
Then, the new phase after the perturbation is
$\theta_{i,new}=2 \arctan((V_i+\delta V)/\sqrt{\eta_i})+\pi$.
The PRC measures the
phase shift produced by the perturbation, i.e. PRC$=\theta_{i,new}-\theta_i$.
Hence, for the QIF model, the PRC is~\cite{Izh07}
\begin{equation}
\text{PRC}(\theta_i,\delta V)=2\arctan(\delta V/\sqrt{\eta_i}-\cot(\theta_i/2))+\pi-\theta_i.
\label{PRC}
\end{equation}
This function depends on both the strength of the perturbation
and the phase of the neuron at the instant of the perturbation.
The PRC Eq.~\eqref{PRC} is always positive, indicating that positive/negative
perturbations only produce positive/negative phase shifts. This
characterizes the so-called Class 1 neuronal oscillators~\cite{Erm96}.
\subsection{Weak coupling approximation and the Winfree model}
The PRC Eq.~\eqref{PRC} exactly characterizes the phase response of the
QIF neuron to a perturbation. Next, we invoke weak coupling, which allows for deriving
a new phase model---called the Winfree model---that
approximates the network Eq.~\eqref{qif} for $\epsilon \ll 1$.
Weak perturbations produce small changes in the membrane potential, $|\delta V| \ll 1$.
Then the PRC scales linearly with the strength of the perturbation~\cite{Win67}
$$\text{PRC}(\theta_i,\delta V) \approx Z(\theta_i) \delta V,$$
where $Z(\theta_i)$ is called phase sensitivity function or infinitesimal phase resetting curve (iPRC). For the QIF model, the iPRC is
\begin{equation}
Z(\theta_i)=
\left.\frac{\partial \text{PRC}(\theta_i,\delta V)}{\partial (\delta V)}\right|_{\delta V=0}
=\frac{1-\cos \theta_i}{\sqrt{\eta_i}}.
\label{iPRC}
\end{equation}
When weak perturbations are described by a continuous function $P(t)$
with $|P(t)|\ll 1$, the infinitesimal change in the phase due to the
perturbations is $d \theta = Z(\theta) P(t) dt$.
Accordingly,
assuming weak coupling $\epsilon \ll 1$,
the population of QIF neurons Eq.~\eqref{qif} is well approximated
by the Winfree model,
\begin{align}
\dot \theta_i &= \Omega_i + \frac{\epsilon}{\tau} (1-\cos \theta_i) \sum_{j=1}^N P(\theta_i,\theta_j),
\label{Winfree}
\end{align}
where perturbations to neuron $i$ are due to synaptic inputs from neuron $j$
and can be written in terms of the phase variables as
\begin{equation}
P(\theta_i,\theta_j)=\frac{g}{N} \left( \cot(\theta_i/2)-\sqrt{\frac{\eta_j}{\eta_i}} \cot(\theta_j/2)\right) +\frac{2J}{N} \sqrt{\frac{\eta_j}{\eta_i}} \delta\left(\theta_j\right).
\label{P}
\end{equation}
Recall that $\theta_j\in [0,2\pi)$ so that
the Dirac delta function in Eq.~\eqref{P} has argument zero whenever neuron $j$
fires a spike.
\subsection{Weak heterogeneity and the averaging approximation}
The Winfree model can be further simplified using the method of averaging.
We consider the external currents in Eq.~\eqref{qif} as a common
current $\bar \eta$ plus a weakly distributed parameter as
\begin{equation}
\eta_i=\bar \eta + \epsilon \chi_i.
\label{eps}
\end{equation}
In the derivation of the Winfree model, we already assumed weak coupling, $\epsilon \ll 1$. Therefore, the smallness of parameter $\epsilon$
implies now the smallness of both coupling terms and the level of heterogeneity.
This assumption allows for a separation of time scales so that the
phases $\theta_i$ can be written as
\begin{equation}
\theta_i=\Phi+\phi_i,
\label{sts}
\end{equation}
where $\Phi$ describes the fast, free-running oscillation of period
$$T=\tau \pi/\sqrt{\bar \eta},$$
whereas the phases $\phi_i$ describe slow phase drifts produced by
weak heterogeneities and synaptic inputs.
Substituting Eq.~\eqref{sts} into
the Winfree model Eqs.~(\ref{Winfree},\ref{P}) and collecting
terms of order $\epsilon$,
we find the evolution equation for the slow phases
\begin{align}
\dot \phi_i = \frac{ \epsilon\chi_i}{\tau \sqrt{\bar\eta}} + [1-\cos(\Phi+\phi_i)]
\frac{\epsilon}{\tau N}\sum_{j=1}^N p(\Phi+\phi_i,\Phi+\phi_i+\Delta_{ji} ).
\label{av0}
\end{align}
Here, we defined pairwise phase differences as $\Delta_{ji} = \phi_j-\phi_i$
and a function describing synaptic perturbations as
\begin{equation}
p( x,y ) = g\left[\cot(x/2)-\cot(y/2)\right]+2J \delta(y).
\nonumber
\end{equation}
To apply the method of averaging to Eq.~\eqref{av0},
we consider that in one period of the fast oscillation, $T$,
the slow phases $\phi_i$ can be assumed constant. Then, Eq.~\eqref{av0} reduces to
\begin{equation}
\dot \phi_i=\frac{\epsilon}{\tau N} \sum_{j=1}^N \Gamma( \Delta_{ji} ),
\label{av}
\end{equation}
where the coupling function $\Gamma$ is obtained by averaging
the r.h.s. of Eq.~\eqref{av0} over one period $T$.
This involves the evaluation of four integrals that can be
explicitly computed and yields the phase interaction function
\begin{equation}
\Gamma(\Delta_{ji})= \frac{ \chi_i}{\sqrt{\bar\eta}} +
g \sin \Delta_{ji}+ \frac{J}{\pi}
( 1-\cos \Delta_{ji}).
\label{gamma}
\end{equation}
\subsection{Kuramoto model for populations of QIF neurons}
Substituting Eq.~\eqref{gamma} into Eq.~\eqref{av} and expressing the result in terms of the
original phases Eq.~\eqref{sts}, we find the Kuramoto model
\begin{equation}
\dot \theta_i= \omega_i+
\frac{\epsilon}{\tau N} \sum_{j=1}^N \left[g \sin ( \theta_j-\theta_i)- \frac{J}{\pi}
\cos( \theta_j-\theta_i) \right] +\frac{\epsilon }{\tau} \frac{J}{\pi}
\label{KM}
\end{equation}
with natural frequencies
\begin{equation}
\omega_i = \frac{2 \sqrt{\bar\eta}}{\tau} +\epsilon \frac{\chi_i}{\tau\sqrt{\bar\eta}}.
\label{om}
\end{equation}
In the absence of electrical synapses $g=0$, Eqs.~\eqref{KM} essentially
\footnote{For $g=0$ the natural frequencies
in Refs.~\cite{PM14,MP18} differ from Eq.~\eqref{om}.
The reason for this discrepancy is that Refs.~\cite{PM14,MP18}
consider the Winfree model with distributed natural frequencies $\Omega_i$,
while here we study the QIF model with distributed currents $\eta_i$.}
reduces to the Kuramoto model with chemical synapses derived in Refs.~\onlinecite{PM14} and~\onlinecite{MP18}.
The KM for QIF neurons Eq.~\eqref{KM} generalizes the results
in Refs.~\onlinecite{PM14} and~\onlinecite{MP18} to networks with both electrical and
chemical coupling, and it is our main result.
Eq.~\eqref{om} is the linear approximation of Eq.~\eqref{Omega} for
weak heterogeneity---see also Eq.~\eqref{eps}.
The last term of Eq.~\eqref{KM}
describes the deviation of the natural frequencies due to synaptic coupling,
which exclusively depends on chemical coupling.
Excitatory coupling ($J>0$) speeds up the frequencies of the oscillators,
and inhibition ($J<0$) slows them down.
These frequency shifts do not qualitatively affect the collective dynamics of
Eq.~\eqref{KM}, but they may become relevant
if the oscillators are not all-to-all coupled~\cite{SSK88,BT05}
or in the case of
interacting excitatory and inhibitory populations~\cite{MP18}.
Alternatively, Eq.~\eqref{KM} can be cast in the more transparent
form~\cite{SSK88,SK86}
\begin{equation}
\dot \theta_i =\omega_i+\frac{K}{N}\sum_{j=1}^N
\left[\sin(\theta_j-\theta_i -\alpha)+\sin\alpha \right],
\label{KMp}
\end{equation}
with the coupling constant
\begin{equation}
K=\frac{\epsilon}{\tau}\sqrt{(J/\pi)^2+g^2},
\label{K}
\end{equation}
and the phase lag parameter
\begin{equation}
\alpha=\arctan\left(\frac{J/\pi}{g}\right).
\label{alpha}
\end{equation}
The coupling parameters $K$ and $\alpha$ satisfy a simple geometric relation with the
coupling parameters of the QIF model Eq.~\eqref{qif}, illustrated
in Fig.~\ref{Fig1}.
Given a particular choice of the QIF coupling parameters,
Fig.~\ref{Fig1} shows that
electrical coupling and chemical coupling (divided by a factor $\pi$)
contribute equally to the overall coupling strength $K$ of the KM,
and that $K$ is insensitive to the sign of the chemical coupling---in Fig.~\ref{Fig1} we consider an inhibitory network; i.e., $J<0$.
To lighten the notation, we consider $\epsilon=1$ thereafter. Hence, for the
KM Eq.~\eqref{KMp} to be a good approximation of Eq.~\eqref{qif},
in the following, the synaptic weights $J$ and $g$ need to be regarded as small quantities.
\begin{figure}[t]
\includegraphics[width=0.35\textwidth]{Fig1.pdf}
\caption{Geometric relation---determined by Eqs.~(\ref{K},\ref{alpha})---between the coupling parameters of the QIF model Eqs.~\eqref{qif}
and the Kuramoto model Eqs.~\eqref{KMp}. }
\label{Fig1}
\end{figure}
\section{Analysis of the Kuramoto model for Quadratic Integrate-and-fire neurons}
Using Fig.~\ref{Fig1}---or, equivalently, Eq.~\eqref{alpha}---we may infer how chemical and electrical synapses contribute to synchronization,
using well-known results for the KM.
For example, the phase constant $\alpha$
critically determines the synchronization behavior of Eq.~\eqref{KMp}~\cite{SK86}.
In the absence of electrical coupling, $g= 0$, we find $\alpha = +\pi/2$
for excitatory coupling and $\alpha=-\pi/2$ for inhibitory coupling. This
indicates that collective synchronization is unreachable---consistent
with the well-known fact that instantaneous chemical coupling
is unable to synchronize type 1 neuronal oscillators~\cite{Erm96,HMM95,DRM17}.
In contrast, in the absence of chemical coupling, one finds $\alpha=0$,
and Eq.~\eqref{KMp} reduces to the standard KM, in which
collective synchronization is
achieved at a critical degree of heterogeneity $\Delta=\Delta_c(K)$
that depends on the coupling strength~\cite{Kur84}.
Between these two extreme cases, that is in networks
with both electrical and chemical synapses,
we find the phase lag parameter $|\alpha| \in (0,\pi/2)$,
and synchronization generally depends on both $\alpha$ and the
overall shape of the distribution of natural frequencies~\cite{SK86}.
To validate the KM for QIF neurons, in Section IV A,
we obtain the mean-field model corresponding to Eqs.~\eqref{KMp} and compare its predictions with those of the mean-field model derived
in Ref.\cite{PDR+19}, which describes the dynamics of the QIF network
Eq.~\eqref{qif} exactly.
\subsection{Mean-field model}
\begin{figure}[b]
\includegraphics[width=85mm]{Fig2.pdf}
\caption{Synchronization boundary of the QIF network Eq.~\eqref{qif}
with the Lorentzian distribution of currents for various
values of the (scaled) inhibitory coupling strength, $J/\sqrt{\bar \eta}$.
The solid line corresponds to
the approximated critical width Eq.~\eqref{DeltaC}, which is independent of $J$.
Dashed lines correspond to the exact synchronization boundaries, obtained
using Eq.~(7) in Ref.~\cite{PDR+19}.}
\label{Fig2}
\end{figure}
In the thermodynamic limit ($N \to \infty$), the dynamics of
Eqs.~\eqref{KMp} are greatly simplified assuming
$\chi_i$ in Eq.~\eqref{om} to be Lorentzian-distributed
$$G(\chi)= \frac{\Delta/\pi}{\chi^2+\Delta^2},$$
where $\Delta$ is the half-width of the distribution.
Then, using the so-called Ott-Antonsen (OA) ansatz~\cite{OA08},
the KM Eqs.~\eqref{KMp} can be exactly reduced to a
mean-field model consisting of two differential equations
for the complex Kuramoto order parameter,
\begin{equation}
Z=R e^{i \psi}= \frac{1}{N}\sum_{j=1}^N e^{i \theta_j},
\label{kop0}
\end{equation}
in the limit $N\to\infty$. The mathematical approach to obtain the mean-field equations
corresponding to Eqs.~\eqref{KMp} is a standard procedure.
Here, we skip the mathematical details and refer the reader to, for example
Ref.~\cite{MP11}, where the
mean-field model corresponding to Eq.~\eqref{KMp} was derived in detail.
Accordingly, using Eqs.~(\ref{om},\ref{K},\ref{alpha}),
we obtain the mean-field equations
\begin{subequations}
\label{OA0}
\begin{eqnarray}
\dot R &=& \frac{R}{2\tau} \left(- \frac{2\Delta}{\sqrt{\bar\eta}}+g(1-R^2) \right),
\label{R}\\
\dot \psi &=& \frac{ 2 \sqrt{\bar\eta}}{\tau}+ \frac{J}{2\pi\tau } (1-R^2) ,
\label{psi}
\end{eqnarray}
\end{subequations}
which approximate the dynamics of the QIF model Eqs.~\eqref{qif}
for small $g$ and $J$.
The radial equation Eq.~\eqref{R} shows that the incoherent state ($R=0$)
is a stable fixed point above the critical width,
\begin{equation}
\Delta_c = g \sqrt{\bar \eta}/2 ,
\label{DeltaC}
\end{equation}
which is independent of chemical coupling, $J$.
At $\Delta=\Delta_c$ a stable nontrivial solution---corresponding to a partially
synchronized state---bifurcates from incoherence with
\begin{equation}
R= \sqrt{(\Delta_c-\Delta)/\Delta_c},
\nonumber
\end{equation}
and frequency
\begin{equation}
\Omega= \frac{2\sqrt{\bar \eta}}{\tau}+ \frac{\Delta}{\tau \pi\sqrt{\bar \eta}}\frac{J }{g }.
\label{OmC}
\end{equation}
The solid line in Fig.~\ref{Fig2} corresponds to the critical boundary Eq.~\eqref{DeltaC},
while dashed lines correspond to the exact synchronization
boundaries of the QIF network for various degrees of inhibitory coupling---see Eq.~(7) in Ref.~\onlinecite{PDR+19}.
Note that for weak electrical coupling and/or weak heterogeneity,
all boundaries approach Eq.~\eqref{DeltaC}. Furthermore, for weak heterogeneity,
the frequency of the synchronized cluster Eq.~\eqref{OmC} agrees
with Eq.~(8) in Ref.~\cite{PDR+19}, which describes the frequency of
the oscillations near their onset.
In sum, these results confirm
the validity of the Kuramoto model Eq.~\eqref{KMp} as an approximation of a
population of heterogeneous QIF neurons with electrical and chemical
coupling, Eq.~\eqref{qif}.
\section{Chimera states in coupled homogeneous populations of QIF neurons}
To further illustrate the appropriateness of the KM
to investigate the dynamics of QIF networks,
we investigate the presence of chimera states
in populations of QIF neurons. Our motivation is threefold:
\begin{enumerate}
\item Chimera states were originally uncovered
in a nonlocally coupled network of identical Kuramoto oscillators~\cite{KB02}.
Given that the phase dynamics Eqs.~\eqref{KMp} are an approximation of Eqs.~\eqref{qif}
valid for weak heterogeneity and weak coupling,
we expect QIF networks to display
similar chimera states, at least for weak coupling.
\item Several papers have been devoted to investigate
chimera states in networks of spiking neurons;
see, e.g., Refs.~\onlinecite{Sak06,OPT10,OOH+13,VHO+14,HKV+14,
PA15,MBG+19,Lai19,HKZ+16,MBG+19,GBS+20,LJH21}.
Some of them provide
numerical evidence that the presence of both chemical and electrical synapses favors the emergence of chimera states~\cite{HKZ+16,MBG+19}.
Yet, the relation between chimera states in spiking neuron networks
with the original chimera states uncovered in the KM~\cite{KB02,MKB04,AMS+08,Lai09}
is lacking~\cite{BGL+20,Ome18,Hau21}.
\item Recently, exact mean-field models for large populations of QIF
neurons (often called neural mass models, NMMs)
with electrical and chemical synapses have been put forward
\cite{Lai15,PDR+19,MP20,BRN+20}. However, such NMMs
have an important limitation when neurons are---as in a chimera state---identical and fully synchronized since both the
mean membrane potential and the
mean firing rate diverge at the instant of collective firing
\footnote{
NMMs for QIF neurons are derived after adopting the thermodynamic limit, $N \to \infty$, and
under the assumption $V_r \to -\infty$, and $V_p \to \infty$, see \cite{MP20,MPR15}.
If neurons are fully synchronized, ($V_i(t)=V_j(t), ~\forall i,j$)
the population of QIF neurons behaves as a single neuron.
Therefore, in QIF-NMM, the mean voltage Eq.~\eqref{v}
diverges when all neurons fire a spike---in Refs.~\cite{Lai15,Erm06} the mean-field variable $v$ is approximated to avoid this divergence, see also \cite{PDR+19}. }.
This divergence is avoided using the averaging approximation, and hence
the KM for QIF neurons, Eq.~\eqref{KM}, becomes singularly suited to study
collective behavior where neurons are fully synchronized.
\end{enumerate}
Chimera states were originally uncovered
in a ring of identical Kuramoto oscillators with nonlocal coupling
when $\alpha \lesssim \pi/2$~\cite{KB02}.
Shortly after their discovery, chimera states were also found in a simpler setup,
consisting of two populations of identical
Kuramoto oscillators~\cite{MKB04,AMS+08}.
Here, to investigate chimera states in networks of QIF neurons,
we adopt the two-population setup of Refs.~\onlinecite{MKB04,AMS+08}.
Specifically, we analyze the dynamics of two identical populations
(labeled $\sigma \in \{1,2\}$) of $n=N/2$ identical QIF neurons,
interacting all-to-all via both chemical and electrical synapses
\begin{equation}
\tau \dot V_i^{\sigma} = \left( V_i^{\sigma} \right)^2+\bar \eta +I_{i,syn,s}^{\sigma}+
I_{i,syn,c}^{\sigma},
\label{qif2}
\end{equation}
with the resetting rule of Eq.~\eqref{qif}. Synaptic inputs have a contribution
$I_{i,syn,s}^{\sigma}$ due to
self-interactions within each population $\sigma$
and another contribution $I_{i,syn,c}^{\sigma}$ due to cross-interactions
of population $\sigma=\{1,2\}$ with population $\sigma'=\{2,1\}$,
\begin{eqnarray*}
I_{i,syn,s}^{\sigma} &=& g_s (v^{\sigma}- V_i^{\sigma}) + J_s \tau r^{\sigma},\\
I_{i,syn,c}^{\sigma} &=& g_c (v^{\sigma'}- V_i^{\sigma}) + J_c \tau r^{\sigma'}.
\end{eqnarray*}
Here, $v^{\sigma}$ and $r^{\sigma}$ are the mean
membrane voltage and mean firing rate of population $\sigma$, respectively.
Using Eqs.~(\ref{KMp},\ref{K},\ref{alpha}), it is straightforward to write the
KM corresponding to Eqs.~\eqref{qif2} as
\begin{eqnarray}
\dot \theta_i^{\sigma}= \omega
&+& \frac{K_s}{n}\sum_{j=1}^n
\left[\sin\left(\theta_j^{\sigma}-\theta_i^{\sigma} -\alpha_s \right)
+\sin \alpha_s \right]\nonumber\\
&+& \frac{K_c}{n}\sum_{j=1}^n
\left[ \sin\left(\theta_j^{\sigma'}-\theta_i^{\sigma} -\alpha_c \right)
+\sin \alpha_c \right],
\label{KM2}
\end{eqnarray}
with $\omega = 2 \sqrt{\bar\eta}/\tau$ and
\begin{eqnarray}
K_{s,c}&=&\frac{1}{\tau}\sqrt{(J_{s,c}/\pi)^2+g_{s,c}^2}, \label{Ksc}\\
\alpha_{s,c}&=&\arctan\left(\frac{J_{s,c}/\pi}{g_{s,c}}\right). \label{alphasc}
\end{eqnarray}
The KM Eqs.~\eqref{KM2} is slightly more general than the model originally investigated
in Refs.~\cite{MKB04,AMS+08,Lai09}---which considered $\alpha_c=\alpha_s$.
In the QIF network, this equality of the phase lag parameters implies that the ratios of chemical to electrical coupling
\begin{equation}
\rho_{s} =\frac{J_{s}/\pi}{g_{s}},\quad \rho_{c} =\frac{J_{c}/\pi}{g_{c}},
\label{rho}
\end{equation}
are identical, $\rho_s=\rho_c.$ Recent work has also
considered the dynamics of chimera states in populations of Kuramoto oscillators
with distributed phase lags~\cite{MBP16,CRK16}.
Specifically, Martens et al.~\cite{MBP16} investigated chimera
states in the two-population model Eqs.~\eqref{KM2}.
\subsection{Mean-field model}
As we discussed previously, in the thermodynamic limit ($n=N/2 \to \infty$)
the KM can be exactly reduced to a low-dimensional mean-field model
using the OA ansatz. In the case of the
homogeneous,
two-population Kuramoto model Eqs.~\eqref{KM2},
the dynamics reduces to six ordinary differential equations using the
Watanabe-Strogatz ansatz~\cite{PR08,WS94}.
Assuming a particular set of initial conditions
for the phases, the system further reduces to four differential equations
and it is described by the OA ansatz~\cite{PR08,AMS+08}.
Such mean-field equations describe the evolution of the complex
Kuramoto order parameters of the two populations,
\begin{equation}
Z_\sigma=R_{\sigma}e^{i \psi_\sigma}= \frac{1}{n}\sum_{j=1}^n e^{i \theta^\sigma_j}.
\label{kops}
\end{equation}
Using the mean-field analysis in Refs.~\onlinecite{KNA+10,MBP16} and
Eqs.~(\ref{Ksc},\ref{alphasc},\ref{rho}), the mean-field equations for the complex Kuramoto
order parameters can be further reduced (by virtue of the rotational symmetry of the KM)
to the three dimensional system
\begin{subequations}
\label{OA}
\begin{eqnarray}
\frac{d R_1}{d\tilde t}&=& \frac{1-R_1^2}{2}\left[ R_1 + \frac{g_c}{g_s} R_2 \cos \Psi
- \rho_s\frac{J_c}{J_s} R_2 \sin \Psi \right], \label{R1}\\
\frac{d R_2}{d\tilde t} &=& \frac{1-R_2^2}{2}\left[ R_2 + \frac{g_c}{g_s} R_1 \cos \Psi
+ \rho_s\frac{J_c}{J_s} R_1 \sin \Psi \right],\label{R2} \\
\frac{d\Psi}{d\tilde t}&=& \rho_s \frac{ R_1^2-R_2^2 }{2R_1R_2}
\left(\frac{J_c}{J_s} \cos \Psi -R_1R_2 \right)-\nonumber \\
&&\frac{g_c}{g_s}\frac{R_1^2+R_2^2+2 R_1^2 R_2^2}{2R_1R_2}\sin \Psi,
\label{phi}
\end{eqnarray}
\end{subequations}
where the phase difference between the complex
order parameters Eq.~\eqref{kops} is defined as $\Psi=\psi_1-\psi_2$.
In addition, we have rescaled time as $\tilde t = g_s t/\tau$ so that the
dynamics of Eqs.~\eqref{KM2} depends only on three combinations of parameters:
the ratios of cross to self couplings $g_c/g_s$ and $J_c/J_s$
and the ratio of chemical to electrical coupling $\rho_s$ ---see Eq.~\eqref{rho}.
In contrast, the original QIF model Eq.~\eqref{qif2} can, after appropriate rescaling,
only be reduced to involve at least four parameters.
\subsection{Phase diagram of the mean-field model}
Chimera states in two-population Kuramoto networks
correspond to symmetry-broken
states where one of the populations is fully synchronized (i.e. $R_\sigma=1$),
while the other remains only partially synchronized ($R_{\sigma'}<1$).
In addition, chimera states in two-population networks of identical Kuramoto
oscillators coexist with the fully synchronized state,
$R_1=R_2=1$~\cite{MKB04,AMS+08,Lai09}.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Fig3.pdf}
\caption{Phase diagrams of the mean field model~Eqs.~\eqref{OA}, for two
values of the ratio $\rho_s$; see Eqs.~\eqref{rho}. Shaded and hatched regions
correspond to regions of steady and unsteady stable chimera states, respectively.
Red lines: Saddle-node (SN) bifurcations. Blue lines: Hopf bifurcations.
Green lines: Homoclinic bifurcations. Filled circles: Takens-Bogdanov
points.
In panel (a), the $+$ symbol corresponds to the coordinates used
for the numerical simulations of the QIF network depicted in Fig.~\ref{Fig4}(f,g,h):
$J_c/J_s=0.15$, $g_c/g_s=2.0$.
In panel (b), symbols correspond to the coordinates used
for the numerical simulations of the QIF network depicted in
Fig.~\ref{Fig4}(b,c,d) and Fig.~\ref{Fig5}:
$J_c/J_s=0.75$, $g_c/g_s=0.5$ ($+$ symbol)
and in Fig.~\ref{Fig6}: $J_c/J_s=0.5$, $g_c/g_s=0.5$ ($\times$ symbol). }
\label{Fig3}
\end{figure}
To obtain the phase diagrams depicted in Fig.~\ref{Fig3},
we set $R_2=1$ in Eqs.~\eqref{OA} and numerically continued~\cite{AUTO} chimera states
using initial conditions in their basin of attraction---see Fig.~\ref{Fig4}(a)
and Ref.~\onlinecite{AMS+08}.
The diagrams show the regions where steady (shaded) and
unsteady (hatched) chimera states are stable
for two different values of the ratio $\rho_s$; see Eqs.~\eqref{rho}.
These regions lie between a saddle-node (red) and a Homoclinic (green) bifurcation lines,
which---together with a Hopf (blue) bifurcation line
separating steady and unsteady chimera states---meet at two
Takens-Bogdanov (TB) points.
For decreasing $|\rho_s|$, the region of chimera states
shrinks and eventually disappears when the two TB points collide.
The phase diagrams in Fig.~\ref{Fig3} are qualitatively
identical to that of Fig.~4a in Ref.~\onlinecite{MBP16}, but here the
regions of chimeras are represented
in the parameter space of the QIF model
\footnote{Figures~4(b-f) in Ref.~\cite{MBP16} show a bifurcation scenario with
a transcritical bifurcation that is not observed in Fig.~\ref{Fig3}.
This bifurcation occurs for $\alpha_s>\pi/2$, and
these values of the phase lag parameter are unreachable in the QIF model,
where $\alpha_s\in [-\pi/2,0)$ for inhibitory coupling and
$\alpha_s \in (0,\pi/2]$ for excitatory coupling), see Eq.~\eqref{alphasc}.}.
This allows us to determine three
necessary conditions for the existence of chimera states in two-population
networks of QIF neurons:
\begin{enumerate}
\item Chimera states only exist in the
presence of \emph{both} chemical and electrical coupling.
\item Self-chemical coupling needs to be much larger
than self-electrical coupling, $|J_s| \gg g_s$ ---or, equivalently, $|\rho_s| \gg 1$.
Using Eq.~\eqref{alphasc}, this implies that $\alpha_s$ is close to
$\pm \pi/2$ in correspondence with
previous work~\cite{MBP16,MKB04,AMS+08}.
\item The modulus of self-chemical coupling needs to be larger than that of
cross-chemical coupling, $|J_s|>|J_c|$.
\end{enumerate}
These three conditions are not sufficient conditions to have chimera states in
Eqs.~\eqref{qif2} though. Indeed, Eq.~\eqref{KM2} and the corresponding
mean-field Eqs.~\eqref{OA} are an approximation of the QIF
Eqs.~\eqref{qif2} for weak coupling, but
it remains to be seen whether chimera states persist in QIF networks
when coupling strengths become stronger. We numerically explore this
issue in the next sections.
\begin{figure*}
\includegraphics[width=1.\textwidth]{Fig4.pdf}
\caption{Basins of attraction of chimera states in the two-population KM (dashed lines)
and in the two-population QIF model (green-dotted regions).
Panels (a-d): Phase portraits of the (a) KM and (b-d) QIF model, with
$J_c/J_s=0.75$, $g_c/g_s=0.5, \rho_s=-40/\pi$, see $\times$ symbol in Fig.~\ref{Fig3}(b).
Panels (e-h): Phase portraits of the (e) KM and (f-h) QIF model, with
$J_c/J_s=0.15$, $g_c/g_s=2, \rho_s=-120/\pi$, see $+$ symbol in Fig.~\ref{Fig3}(a).
Symbols and lines in panels (a,e): Solid/open diamond:
Stable/unstable chimera states, respectively;
solid dot: In-phase synchronized state.
Solid blue line: Unstable manifold of the saddle point.
Dashed black lines: Stable manifold of the saddle point,
basin of attraction of stable chimeras.
The basins of attraction of chimeras
have been transformed from the $(R_1 \cos \Psi,R_1\sin \Psi)$ coordinates (panels a,e)
to the $(r_1,v_1)$ coordinates (panels b-d and f-h) using Eq.~\eqref{conformal}.
Green dots corresponds to initial values leading to
chimera states after $t=2500$ time units in numerical simulations of
two populations of $n=200$ QIF neurons.
Parameters: $\tau=1$ and $\eta=1$.
Simulations using the Euler scheme with time step:
$dt=10^{-4}$ and symmetric resetting: $V_p=-V_r=1000$.}
\label{Fig4}
\end{figure*}
\subsection{Chimera states in populations of QIF neurons}
In the following, we numerically investigate the presence of chimera states in the
spiking neuron network model of QIF neurons
Eqs.~\eqref{qif2}. First, we confirm that, for weak coupling,
chimeras are present in QIF networks and they
exist in the parameter range predicted by the phase diagrams of the KM,
Fig.~\ref{Fig3}. However, then we show that
the basin of attraction of chimera states shrinks
as synaptic coupling strengths become stronger.
\subsubsection{Dynamics of chimera states}
Using the OA ansatz, the dynamics of the two-population model
Eqs.~\eqref{KM2} with $N\to \infty$ can be exactly reduced to the
three-dimensional system Eqs.~\eqref{OA}.
In a chimera state we may set $R_2=1$ so that
Eqs.~\eqref{OA} further reduce to a planar system with variables $R_1$ and $\Psi$.
Figures~\ref{Fig4}(a,e) show the basins of attraction (dashed lines) of (a) a steady
chimera state (solid diamond symbol) and of
(e) an unsteady chimera state (red limit cycle),
with parameters corresponding to $+$ symbols in Fig.~\ref{Fig3}.
As mentioned previously, chimera states coexist with the stable fully synchronized
solution (solid dot symbols), $Z_1=R_1=1,~\Psi=0$. The basin of attraction of the chimera
state is defined by the stable manifold of a saddle point (open diamond symbols)
~\cite{AMS+08}.
To set initial conditions leading to the chimera state of Fig.~\ref{Fig4}(a)
in the network of QIF neurons,
we considered the initial condition $Z_2=1$ and $Z_1=R_1 e^{i\Psi}$
---with $R_1$ and $\Psi$ such that the system is in
the basin of attraction of the steady chimera state.
Then we used the conformal map~\cite{MPR15}
\begin{equation}
\label{conformal}
\pi r_\sigma - i v_\sigma =\frac{1-Z_\sigma}{1+Z_\sigma},
\end{equation}
to transform the mean-field coordinates $Z_2$ and $Z_1$ into the mean firing rate $r$
and the mean membrane potential $v$ of the populations of QIF neurons---for population 2 we find $\pi r_2 + i v_2=0$. Then,
we initialized the membrane voltages of the populations according to the formula
$V_i(0)=v+(\pi \tau r) \tan[\pi/2 (2i-n-1)/(n+1)],$ for $i=1,\dots,n$~
\footnote{
For $n\to \infty$, this corresponds to a Lorentzian distribution of voltages
$\rho(V)=\pi \tau r /[ (V-v)^2+ (\pi \tau r)^2]$; see Ref.~\onlinecite{MPR15},
and to uniformly distributed
constants of motion in the Watanabe-Strogatz theory~\cite{PR08,Lai18,BGL+20}
}.
In Fig.~\ref{Fig5}, we show the results of a numerical simulation of the QIF network
($g_s=0.1$).
The raster plot in Fig.~\ref{Fig5}(a) clearly shows the signature of a chimera state:
Neurons in population 1 (blue) are only partially synchronized, while
neurons in population 2 remain fully synchronized.
The time evolution of the firing rate $r_1$ and the mean-membrane
potential $v_1$ for the incoherent group are displayed in
Figs.~\ref{Fig5}(b) and (c), respectively.
These collective variables indicate a periodic evolution of the incoherent population,
with fluctuations caused by the
finite resetting of the QIF neurons and by finite-size effects.
Finally, Fig.~\ref{Fig5}(d) shows the Kuramoto order parameter $R_1$ (blue)
obtained using the time series $r_1(t)$ and $v_1(t)$ and
the conformal map Eq.~\eqref{conformal}.
In contrast with the steady chimera state in the mean-field Eqs.\eqref{OA}
(black dotted line),
the chimera state in the network of QIF neurons is not stationary but oscillates
periodically in time. The same unsteady chimeras arise in
two-population networks of Winfree oscillators~\cite{PM14}
and are the consequence of the lack of rotational symmetry in the Winfree and QIF models.
In Fig.~\ref{Fig6}, we also explored how the unsteady chimera states
in the Kuramoto model Eqs.~\eqref{KM2}
translate to networks of QIF oscillators.
To this aim we set the parameters of the QIF model
in the hatched region of the bifurcation diagram in Fig.~\ref{Fig3}b ($\times$ symbol),
and used the same initial conditions as in the previous simulation.
Here, we find a more complex chimera state that seems to display
macroscopic quasiperiodic dynamics
\footnote{We cannot discount that the collective dynamics
of the QIF Eqs.\eqref{qif2} is chaotic, as it also occurs in coupled
populations of Winfree oscillators~\cite{PM14}.}:
both the firing rate
and the mean-membrane potential of the incoherent population oscillate with
two characteristic frequencies as can be appreciated in Figs.~\ref{Fig6}(b,c).
Again, the quasiperiodic chimera state in the QIF network corresponds to a
periodic chimera state in the Kuramoto model.
In Fig.~\ref{Fig5}(d) we used Eq.~\eqref{conformal}
to represent the Kuramoto order parameter for the QIF network (blue), which
roughly approximates the periodic dynamics of the corresponding
Eqs.~\eqref{OA}.
\begin{figure}[t]
\centerline{\includegraphics[width=85mm,clip=true]{Fig5.pdf}}
\caption{Periodic chimera state in a
two-population network of $N=5000$ identical
inhibitory QIF neurons ($n=2500$ neurons in each population).
(a) Raster plot of 500 randomly chosen neurons. Neurons in population 1 (blue) are
partially synchronized, and neurons on population 2 are in-phase synchronized (red).
(b) Time series of the mean firing rate $r_1$ of population 1,
computed averaging the firing rate at each time step
in time windows of $\delta t=0.05$.
(c) Time series of the mean membrane potential $v_1$
of population 1.
(d) Time series of the Kuramoto order parameter of population 1, $R_1$,
obtained from the mean-field quantities $r_1,v_1$ using the conformal map
Eq.~\eqref{conformal} (blue lines)
and from direct integration of Eqs.~(\ref{OA}) (black dots).
Parameters as in Fig.~\ref{Fig4}d (see also $+$ symbol in Fig.\ref{Fig3}b):
$J_c/J_s=0.75$, $g_c/g_s=0.5$, $\rho_s=-40/\pi$, and $g_s=0.1$.
Numerical simulations performed using the Euler scheme with time step:
$dt=10^{-4}$ and symmetric resetting: $V_p=-V_r=1000$.
}
\label{Fig5}
\end{figure}
\begin{figure}[t]
\centerline{\includegraphics[width=85mm,clip=true]{Fig6.pdf}}
\caption{Quasiperiodic chimera state in a two-population network of $N=5000$ identical
inhibitory QIF neurons. The description of the panels and parameters is as
in Fig.~\ref{Fig5} except $J_c/J_s=0.5$
(see also the $\times$ symbol in Fig.\ref{Fig3}b).
}
\label{Fig6}
\end{figure}
\subsubsection{Chimeras for strong coupling}
The derivation of Eqs.~\eqref{KM2} from
the QIF network Eqs.~\eqref{qif2} has been made under the assumption of
weak coupling. Yet, are chimera states in QIF networks robust against
stronger levels of coupling?
To investigate this issue, we used the conformal map Eq.~\eqref{conformal} to express
the boundary of the basin of attraction of Fig.~\ref{Fig4}(a) in terms of
$(r_1,v_1)$. This transformed boundary is represented as a dashed line in Figs.~\ref{Fig4}(b-d).
Then, we performed three sets of numerical simulations of the QIF network for
increasing values of $g_s$ while keeping the ratios $g_c/g_s$, $J_c/J_s$, and $\rho_s$ constant---note that this implies increasing all the other coupling parameters.
For very small values of $g_s$,
we expect the averaging approximation to hold, and so the stability boundary of
chimera states in the QIF model.
The green dotted region in Fig.~\ref{Fig4}(b) corresponds to the stability boundary
of the steady chimera state in the QIF network
for $g_s=0.025$. The boundary approximately agrees with that of the KM (dashed line),
although the region is slightly smaller in the QIF model.
Notably, further increases in coupling strength---see Figs.~\ref{Fig4}(c,d)---lead to a gradual reduction of the chimera's basin of attraction.
To further investigate the reduction of the stability boundary of chimeras in the QIF model,
in Figs.~\ref{Fig4}(f-h) we computed the basins of attraction of chimera states
in a different parameter regime---corresponding to the $+$ symbol in Fig.~\ref{Fig3}(a).
Here $|\rho_s|$ is three times larger than in the previous case, and hence,
chemical couplings are three times stronger than in Figs.~\ref{Fig4}(b-d).
Correspondingly, the reduction of the basin of attraction of chimera states in
Figs.~\ref{Fig4}(f-h) is more pronounced than in Figs.~\ref{Fig4}(b-d).
In fact, in Fig~\ref{Fig4}(h), we find the complete disappearance of the region of stable
chimera states in the QIF network. This suggests that chimera states in QIF networks are only
observable for weak coupling.
\section{Conclusions}
In this paper we have applied a perturbative approach to simplify
a weakly heterogeneous population of QIF neurons, with
weak all-to-all chemical and electrical coupling Eqs.~\eqref{qif}.
This approach leads to a classical variant of the Kuramoto model
Eq.~\eqref{KMp}~\cite{SSK88,SK86}, whose coupling parameters
satisfy a simple geometric relation with those of the QIF model
\footnote{A similar approach has been applied to the Leaky Integrate-and-Fire model.
In this case the approximated phase model is not the KM, but it
contains higher harmonics in the coupling function~\cite{PR15ii}}
Fig.~\ref{Fig1}.
The approximation of the QIF network by Eq.~\eqref{KMp}
allows one to use the framework of the KM to investigate the
role of chemical and electrical synapses in setting up synchronization.
For example, we find that in the absence of electrical coupling
the phase lag parameter of the KM is $\alpha=\pm \pi/2$,
which prohibits synchronization; see also Ref.~\onlinecite{MP18}.
Moreover, for Lorentzian distributions of currents
the synchronization threshold depends only on electrical coupling,
Eq.~\eqref{DeltaC}, whereas
the oscillation frequency Eq.~\eqref{OmC} is determined by the ratio
of chemical to electrical coupling.
These results are in consonance with the exact description
provided by so-called neural mass (or firing rate) models
for networks of QIF neurons~\cite{PDR+19,MP20}.
The framework of the KM
allows for uncovering and investigating dynamical states
that are not reachable using neural mass models for QIF neurons~\cite{MPR15,Lai15,PDR+19,MP20}.
Here, we analyzed the case of chimera states in two-population networks
of identical QIF neurons~\cite{MKB04,AMS+08}.
Despite the large number of studies devoted to investigate
chimera states in spiking neuron networks---see e.g.~\cite{OPT10,OOH+13,VHO+14,HKV+14,
PA15,MBG+19,Lai19,HKZ+16,MBG+19,GBS+20,LJH21}---,
the relation between such states and the original
chimera states uncovered in the KM~\cite{KB02,MKB04,AMS+08}
is lacking~\cite{BGL+20,Hau21}. We showed
that chimera states in QIF networks
emerge in the presence of both chemical and electrical couplings
but only if chemical coupling is much stronger than electrical coupling.
However, our numerical results suggest
that chimeras in QIF networks are not robust against stronger levels of coupling.
Finally, we introduced a framework for the analysis of QIF networks that can be
readily applied to a variety of extensions of Eqs.~\eqref{qif}.
In particular, the derivation of Eq.~\eqref{KMp} does not
impose constraints on the structure of the network or the
shape of the distribution of heterogeneities.
Given that network structure~\cite{SSK88,BT05}
and heterogeneities~\cite{Paz05,LCT10,OW12,MBS+09,PM09,PDD18}
greatly affect the dynamics of the KM,
it may be interesting to investigate how this translates to QIF networks.
\acknowledgments
The authors thank Diego Paz\'o for helpful discussions.
PC acknowledges financial support from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 101017716 (Neurotwin).
EM acknowledges support by the Agencia Estatal de Investigaci\'on
under the Project No.~PID2019-109918GB-I00.
\section*{Data availability}
Data sharing is not applicable to this article as no new data were created or analyzed in this study.
\section*{Author Declarations}
The authors have no conflicts to disclose.
|
\section{Introduction}
\label{sec:intro}
Given a large repository of searchable content, information retrieval (IR) systems control the exposure of the content to searchers. In many cases, creators value not only the exposure of their content but also an understanding of the search contexts in which their content is surfaced. To enable transparency about search contexts facilitated by a given system, search engines should identify which \emph{search queries expose each of the corpus documents in the ranking results}~\cite{biega2017learning}. Yet, search systems currently do not provide such transparency to content producers, as few non-brute-force methods exist that can be used for this task. Identification of such exposing queries is our focus in this paper.
\vspace{.3em}
\noindent\textbf{Exposing Query Use Cases.} Evidence that exposure transparency is important permeates multiple lines of work, including fairness and bias, privacy, search engine optimization, and even security. \citet{azzopardi2008retrievability} defined the \emph{retrievability bias} as the disparity in the weighted sum of ranks different documents are retrieved at over all possible queries. Finding which queries expose which documents is therefore necessary to quantify and audit this bias. Fair rankings typically focus on particular queries~\cite{diaz2020evaluating}, while there is virtually no way for an individual user to get an overview of which rankings overall expose their content. This is especially crucial in settings where exposure in rankings is highly monetizable, like hiring~\cite{geyik2019fairness} or online marketplaces~\cite{mehrotra2018towards}. \citet{biega2017learning} argued that exposing queries are an important \emph{privacy} tool in search systems, demonstrating a range of scenarios where \emph{exposure to sensitive queries} is problematic~\cite{peddinti2014cloak}. The \emph{right to be forgotten}~\cite{bertram2019five} is one of the key data protection rights: Knowledge of exposing queries might enable users to execute this right more precisely than removing themselves from search results completely.
Content creators could furthermore use the exposure transparency to aid in improving the performance of their documents in search results~\cite{raifer2017information,vasilisky2020studying}.
Beyond content creators, service providers may also care about understanding how different content is exposed~\cite{otterbacher2017competent, wilkie2017algorithmic} to better understand their systems.
Finally, controlling which queries expose certain types of content might prevent unintended leaking of information through search engines
~\cite{fbLeak}.
\vspace{.3em}
\noindent\textbf{Approaches.} An intuitive strategy for identifying exposing queries is to analyze past search logs to find which documents were actually returned in response to specific queries.
Yet, such logs quickly lose their relevance as a source of exposing queries as the search ecosystem is constantly changing and both the collection to be searched and the search models are updated.
Furthermore, in certain sensitive search contexts, the potential for exposure should ideally be captured before exposure occurs and is logged. Thus, exposing queries need to be computed based on a given search ranker rather than from logs. Exact brute-force computation assuming we have a universe of feasible queries---for instance, based on existing query logs and/or generated from the corpus documents---is practically inefficient as it involves issuing all these queries to the retrieval system and re-aggregating the ranking results. The question then becomes of whether we can do better.
\vspace{.3em}
\noindent\textbf{Contributions.}
In this paper, we explore the feasibility of approximate exposing query identification (EQI) as a \emph{retrieval} task in which the function of queries and documents is \emph{reversed}.
We study this idea in the context of two classes of search systems: dense embedding-based nearest neighbor search---\emph{e.g.},~\cite{huang2013learning, nalisnick2016improving, xiong2020approximate} and traditional BM25-based~\cite{robertson2009probabilistic} retrieval using inverted-index~\citep{zobel2006inverted}.
In case of embedding-based search systems, we show how retrieval-based EQI can be improved if we treat the problem as a form of metric learning in a joint query-document embedding space, transforming an embedding space of a document retrieval model such that a nearest neighbor search in the transformed embedding space approximately corresponds to a reverse nearest neighbor search over the original embeddings.
In search systems, whether a query exposes a document is a function of both the document relevance as well as how many other documents in the corpus are relevant.
Our intuition, therefore, is that we might be able to reorganize queries and documents in the embedding space by adequately selecting training data that captures the corpus characteristics of dense document regions.
We furthermore derive an evaluation metric to measure the quality of a ranking of exposing queries similar to two-sided metrics previously used for assessing the effectiveness of query suggestion algorithms~\cite{santos2013learning}.
The metric reveals that the exposure significance of a query is dependent on the behavior of both the document retrieval users and the EQI system user.
Overall, our work contributes a novel conception of transparency in search systems and explores a potential computational approach to achieving it.
\section{Related work}
\label{sec:related}
\vspace{.5em}
\noindent\textbf{Exposing query identification. }
\label{sec:related-exposure}
The exposing query identification problem was defined by in the context of content creator privacy~\citet{biega2017learning}.
The authors proposed an EQI approach that identifies unigram and bigram exposing queries in the specific case of a ranker that uses unigram language models with Dirichlet smoothing. The solution generates synthetic candidate queries from the corpus and then prunes this queryset using a ranker-specific heuristic. The output exposing query set is exact, although computed over a limited set (unigram and bigram) of not always realistic synthetic queries.
Instead, formulating EQI as a retrieval task over a corpus of queries allows us to:
\begin{enumerate*}[label=(\roman*)]
\item model content exposure more adequately by incorporating user browsing models; and
\item expand the space of solutions to include query retrieval methods that can handle arbitrary-length exposing queries.
\end{enumerate*}
EQI is closely related to the concept of \emph{retrievability} developed by~\citet{azzopardi2008retrievability}. Retrievability is a dot product score assigned to a document quantifying whether (1) a document is exposed within a rank a user is patient enough to inspect and (2) how likely a query is to be issued to the search engine. Our goal is to discover specific queries which expose a document to a searcher and could thus be used to estimate component (1) of retrievability.
Further related concepts include \emph{keyqueries}~\cite{gollub2013keywords}, which are the minimal queries retrieving a document in a given document retrieval system. By definition, keyqueries are thus a subset of all exposing queries.
\vspace{.5em}
\noindent\textbf{Search transparency and explainability. }
EQI is a form of search \emph{transparency}: the goal is to help users understand comprehensively what outcomes the search system yields for them. Transparency is closely related to interpretability and explainability, where the goal is to understand how and why a ranking algorithm returns certain results.
Explainability approaches include generation of post-hoc explanations for black-box models, both local (explaining a single ranking result)~\cite{verma2019lirme,singh2019exs} and global (explaining a ranking model)~\cite{singh2018posthoc}.
Model-oriented techniques include modifying models to be more interpretable and transparent~\cite{zhuang2020interpretable} or explaining through visualizations~\cite{chios2021helping,choi2020interpreting}, while user-oriented techniques include inferring the intent behind a search query that a model has assumed~\cite{singh2018interpreting} or tailoring explanations to user mental models of search~\cite{thomas2019investigating}, evaluating explanations through the lens of trust~\cite{polley2021towards}.
\vspace{.5em}
\noindent\textbf{Reversing the role of queries and documents. }
\label{sec:related-related}
Our approach to EQI is inspired by several IR problems in which the role of documents and queries is reversed. \citet{yang2009query} employ a query-by-document technique issuing the most informative phrases from a document as queries to identify related documents. In the context of EQI, the setup has a few differences: We retrieve exposing queries rather than other related documents, and are interested in an extensive coverage of exposing queries thus using the whole document as a query rather than just the most informative phrases. \citet{pickens2010reverted} index query result sets, where each retrievable item is a query and each query is described by a list of pseudo-terms that correspond to document IDs that the query retrieves using a chosen retrieval system.
This approach, referred to as \emph{reverted indexing}, allows for query-by-multiple-documents. In our context, where we are interested in retrieving queries that expose a single document, reverted indexing is akin to the exact brute-force method.
\citet{santos2013learning} develop a framework for suggesting queries which might retrieve more relevant results for a given query. \citet{nogueira2019document} develop a generative model which, for a given document, produces queries that could expand the document for more effective retrieval. Both approaches produce a set of queries to which a given document may be \emph{relevant}, but not those that would expose the document when issued specifically to a given document retrieval system.
We expect that the set of queries generated by these approaches to intersect with the ideal set of exposing queries, but as the behavior of the particular document retrieval system diverges from the labeled data, the intersection will be smaller; and the effectiveness may also vary between labeled-relevant, labeled-nonrelevant, and unlabeled documents.
It is important in these cases for EQI to be effective for documents that are actually exposed, irrespective of their relevance.
\section{Exposing query identification}
\label{sec:task}
We are given a \emph{document retrieval} system producing a ranked list of documents retrieved from a collection $\mathcal{D}$ in response to a query.
We define EQI as a complementary retrieval task where, given a document $d$, the system is responsible for retrieving a ranked list of queries from a query log $\mathcal{Q}$, ranked according to their probability of exposing the document in the document retrieval system.
Thus, EQI technically means reversing document retrieval.
We use subscripts $\textbf{\fbox{}}_{\;q \shortto d}$ and $\textbf{\fbox{}}_{\;d \shortto q}$ to denote variables corresponding to the document retrieval and the EQI task, respectively.
Table~\ref{tbl:notation} presents the notations.
\begin{table}
\small
\centering
\caption{List of notations.}
\label{tbl:notation}
\begin{tabular}{ll}
\toprule
\textbf{Notation} & \textbf{Description} \\
\midrule
$\mathcal{D}$ & A collection of documents \\
$\mathcal{Q}$ & A corpus of queries\\
$d$ & An individual document \\
$q$ & An individual query \\
\midrule
\multicolumn{2}{l}{\textbf{Document retrieval}} \\
$\sigma_q$ & A ranked list of documents retrieved in response to $q$ \\
$\sigma_q^*$ & The ideal set of documents for $q$\\
$n_{q \shortto d}$ & Number of documents retrieved per query\\
$\mu_{q \shortto d}$ & A user browsing model for inspecting $\sigma_q$ \\
$\gamma_{q \shortto d}$ & Persistence parameter for inspecting $\sigma_q$\\
$\rho(d,\sigma_q)$ & 0-based rank of $d$ in $\sigma_q$\\
\midrule
\multicolumn{2}{l}{\textbf{Exposing query retrieval}} \\
$\psi_d$ & A ranked list of queries retrieved in response to $d$ \\
$\psi_d^*$ & The ideal ranked list of queries for $d$ \\
$n_{d \shortto q}$ & Number of queries retrieved per document\\
$\mu_{d \shortto q}$ & A user browsing model for inspecting $\psi_d$ \\
$\gamma_{d \shortto q}$ & Persistence parameter for inspecting $\psi_d$\\
$\rho(q,\psi_d)$ & 0-based rank of $q$ in $\psi_d$\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Deriving an Evaluation Metric}
Retrieval metrics often take the form of a cumulative quality score summed over all items in the top-ranking positions weighted by the probability that a searcher inspects the document at a given rank:
$
\mathcal{M}(\sigma) = \sum_{d \in \sigma}{\mu(d, \sigma)\cdot g_d}.
$
Here, $\sigma$ is the result list, the quality score $g_d$ of a document $d$ is typically a function of its relevance to the query and $\mu(d, \sigma)$ is the probability of a user inspecting $d$ under the user browsing model $\mu$.
Metrics---such as, NDCG~\citep{jarvelin2002cumulated}---further normalize this value by the ideal value of the metric:
$
\text{Norm-}\mathcal{M}(\sigma) = \frac{\sum_{d_i \in \sigma}{\mu(d_i, \sigma)\cdot g_{d_i}}}{\sum_{d_j \in \sigma^*}{\mu(d_j, \sigma^*)\cdot g_{d_j}}}
\label{eqn:norm-ir-metric}
$
Where $\sigma^*$ is the ideal ranked list of results.
Similarly, in the exposing query retrieval scenario, we want to measure the quality of the ranked list of retrieved queries, which we refer to as the exposure list, in response to a document.
Towards that goal, we propose the Ranked Exposure List Quality (RELQ), which assumes a similar form as $\text{Norm-}\mathcal{M}$:
\begin{align}
\text{RELQ}_{\mu_{d \shortto q}}(\psi_d) &= \frac{\sum_{q_i \in \psi_d}{\mu_{d \shortto q}(q_i, \psi_d)\cdot g_{q_i}}}{\sum_{q_j \in \psi^*_d}{\mu_{d \shortto q}(q_j, \psi_d^*)\cdot g_{q_j}}}
\label{eqn:resq-prenorm}
\end{align}
Where $\psi_d$ and $\psi_d^*$ are the retrieved and ideal exposure list for $d$. \newline
The \emph{key observation} now is that in case of EQI, the quality value $g_q$ corresponds to the probability that $q$ exposes $d$ in the original retrieval system---which is given by $\mu_{q \shortto d}$:
\begin{align}
\text{RELQ}_{\mu_{d \shortto q}, \mu_{q \shortto d}}(\psi_d) &= \frac{\sum_{q_i \in \psi_d}{\mu_{d \shortto q}(q_i, \psi_d)\cdot \mu_{q \shortto d}(d, \sigma_{q_i})}}{\sum_{q_j \in \psi^*_d}{\mu_{d \shortto q}(q_j, \psi_d^*)\cdot \mu_{q \shortto d}(d, \sigma_{q_j})}}
\label{eqn:resq}
\end{align}
This derivation leads us to a metric that jointly accounts for the behavior of two distinct users:
\begin{enumerate*}
\item the searcher using the document retrieval system and to whom the content is exposed, and
\item the user of the EQI system.
Intuitively, the proposed metric has a higher value when the EQI user is exposed to queries that in turn prominently expose the target document in their corresponding rank lists.
\end{enumerate*}\footnote{
Similar two-sided metrics have previously been used in the task of query suggestion: the success of a query prediction system depends on the quality of the query ranking but also on the quality of the document ranking each query retrieves~\cite{santos2013learning}. EQI requires a similar multiplicative metric; unlike the query prediction metric, however, an exposure set quality metric needs to:
\begin{enumerate*}[label=(\roman*)]
\item quantify whether a query \emph{exposes} a document rather than whether the document is relevant; and
\item incorporate two different models of user behavior (EQI and document retrieval system users).
\end{enumerate*}}
\vspace{5px}
\noindent\textbf{Metric instantiations. }To compute the metric, we can plug in standard user behavior models for $\mu_{d \shortto q}$ and $\mu_{q \shortto d}$ that make different assumptions about how users interact with retrieved results under the document retrieval and the exposing query retrieval settings.
For example, if we assume that both $\mu_{d \shortto q}$ and $\mu_{q \shortto d}$ are based on the same user model as the Rank-Biased Precision (RBP) metric~\citep{Moffat:2008:RBP} with persistence parameters $\gamma_{d \shortto q}\in(0,1]$ and $\gamma_{d \shortto q}\in(0,1]$, we get the following variant:
\begin{align}
\text{RELQ}_\text{RBP, RBP}(\psi_d) &= \frac{\sum_{q_i \in \psi_d}{\gamma_{d \shortto q}^{\rho(q_i, \psi_d)}\cdot \gamma_{q \shortto d}^{\rho(d, \sigma_{q_i})}}}{\sum_{q_j \in \psi^*_d}{\gamma_{d \shortto q}^{\rho(q_j, \psi_d^*)}\cdot \gamma_{q \shortto d}^{\rho(d, \sigma_{q_j})}}}
\label{eqn:resq-rbprbp}
\end{align}
Where, $\rho(x, \Omega)$ is the $0$-based rank of $x$ in the ranked list $\Omega$.
Instead of RBP, we can plug in different combinations of user models, including NDCG or exhaustive (where a user inspects all ranking results). For example, if we
assume an exhaustive model for EQI, while adopting NDCG for the document retrieval:
\begin{align}
\text{RELQ}_\text{EXH, NDCG}(\psi_d) &= \frac{\sum_{q_i \in \psi_d} \vmathbb{1}\{d \in \sigma_{q_i}\} \cdot \frac{\rho(d, \sigma_{q_i})}{\log_2(i+1)}}
%
{\sum_{q_j \in \psi^*_d}{\vmathbb{1}\{d \in \sigma_{q_j}\}} \cdot \frac{\rho(d, \sigma_{q_j})}{\log_2(j+1)}}
\label{eqn:resq-exhrbp}
\end{align}
Where, $\vmathbb{1}\{.\}$ is the indicator function.
\subsection{Practical considerations}
\subsubsection{Prototypical users}
\label{sec:typical_users}
$RELQ_\text{RBP, RBP}$ is parametrized by two user patience parameters: $\gamma_{q \shortto d}$ and $ \gamma_{d \shortto q}$. Their values should be chosen to best reflect the behavior of users in a given underlying task. Absent behavioral data, we can instead represent certain prototypical users. Patient document retrieval users might be curious or malicious (in case of positive or negative query contexts, respectively). Patient EQI system users might be worried about their privacy or interested in monetization of their content.
\subsubsection{Query collections}
EQI assumes the existence of an exhaustive query collection. There are a number of approaches to creating such a collection. First, we can use an existing log from the document retrieval system that contains all the queries ever issued by searchers. The advantage of this approach is that the queries will be realistic. However, the model will not be able to capture instances of problematic exposure~\cite{biega2016r,biega2017learning,fbLeak} by unseen queries. Synthetically generating a collection of queries is an alternative. To limit the collection size when this approach is adopted, queries can be truncated at a certain length and further pruned according to various frequency heuristics ~\cite{azzopardi2008retrievability,biega2017learning,callan2001query}. The approach might allow the system to detect problematic exposure before it occurs, but it will prevent the system from surfacing exposure to many viable queries. A practitioner should choose an approach depending on whether it is important to report the common or worst-case exposure contexts in a given application.
\section{EQI for traditional search systems}
\label{sec:bm25}
We also study EQI in the context of BM25~\citep{robertson2009probabilistic}, an important traditional retrieval model.
BM25 estimates the query-document relevance score as a function of how often the term occurs in the text (term frequency or TF) and some measure of how discriminative the term is---\emph{e.g.}, inverse-document frequency (IDF)~\citep{robertson2004understanding}.
\begin{align}
s_{q,d} = \sum_{t \in q}{IDF(t) \cdot \frac{TF(t, d) \cdot (k_1+1)}{TF(t, d) + k_1 \cdot (1 - b + b \cdot \frac{|d|}{\text{AvgDL}})}}
\end{align}
Where, $k_1$ and $b$ are parameters of the model that can be tuned and AvgDL is the average document length in the collection.
A simple approach to retrieving exposing queries in the context of BM25 may therefore involve simply exchanging the role of query and documents under this setting.
Instead of precomputing term-document scores and indexing the document collection, we can instead precompute the term-query scores and index the query log.
Given the weighted term vector corresponding to a document, we can retrieve queries from the prepared index.
Real life IR systems often make practical assumptions about the queries being typically short and documents being longer in their design to achieve more efficient retrieval.
While we may need to work around some of these system constraints, the general framework for reversing a BM25-based system seems straightforward.
The term weighting functions can be further tuned, if necessary, for the exposing query retrieval setting independently of the scoring function employed for document retrieval.
\subsection{Indexing and retrieval}
\label{sec:bm25-implementation}
In our work, we adopt the Anserini toolkit~\cite{yang2017anserini} to index the query log and issue documents as queries.
We tune the model parameters, observing that the impact is limited under the reversed setting as the term-saturation is applied on the short indexed queries. We foresee future work that applies the term-saturation on the input to the retrieval system instead, but that consequently requires the retrieval system to support explicit term-weighting.
\section{EQI for embedding-based search systems}
\label{sec:ance}
The first family of retrieval systems that we consider in this work are embedding-based search systems, also referred to as dual-encoder systems, that learn independent vector representations of query and document such that their relevance can be estimated by applying a similarity function over the corresponding vectors.
Let $f_Q : \mathcal{Q} \to \mathbb{R}^n$ and $f_D : \mathcal{D} \to \mathbb{R}^n$ be the query and document encoders, respectively, and $\otimes: \mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R}$ be the vector similarity operator.
The score $s_{q,d}$ for a query-document pair is then given by,
\begin{align}
s_{q,d} &= f_Q(q) \otimes f_D(d)
\end{align}
Ideally, $s_{q,d_+} > s_{q,d_-}$, if $d_+$ is more relevant to the query $q$ than $d_-$.
While $\otimes$ is typically a simple function, such as dot-product or cosine similarity, the encoding functions $f_Q$ and $f_D$ are often learned by minimizing an objective of the following form:
\begin{align}
\mathcal{L}_{q \shortto d} &= \mathbb{E}_{q, d_+, d_- \sim \theta} [\ell_{q \shortto d}(s_{q,d_+} - s_{q,d_-})]
\label{eqn:loss-fwd}
\end{align}
Commonly used functions for the instance loss $\ell_{q \shortto d}$ include hinge \citep{herbrich2000large} or RanknNet \citep{burges2005learning} loss.
For example, in case of RankNet:
\begin{align}
\ell_{q \shortto d}(\Delta) &= \text{log}(1 + e^{-\Delta})
\label{eqn:inst-loss-fwd}
\end{align}
\textbf{Metric learning for EQI. }
For EQI, we want to learn a similarity metric such that a query $q^+$ should be more similar to document $d$ than $q^-$, if $d$ has a higher probability of exposure to the user on the search result page $\sigma_{q^+}$ compared to $\sigma_{q^-}$.
Similar to the embedding-based model for document retrieval, we now define two new encoders $h_Q : \mathcal{Q} \to \mathbb{R}^n$ and $h_D : \mathcal{D} \to \mathbb{R}^n$ for query and document, respectively, and write our new optimization objective as follows:
\begin{align}
\mathcal{L}_{d \shortto q} &= \mathbb{E}_{d, q_+, q_- \sim \theta} [\ell_{d \shortto q}(u_{d, q_+} - u_{d, q_-})]
\label{eqn:loss-rev}
\end{align}
Where, $ u_{d, q} = h_Q(q) \otimes h_D(d)$.
If we assume that exposure of a document depends only on the rank position at which it is displayed on a result page, then $q_+$ and $q_-$ should be sampled such that $\rho(d,\sigma_{q_+}) < \rho(d,\sigma_{q_-})$---\emph{i.e.}, $d$ is displayed at a higher rank position for $q_+$ than $q_-$.
Let, $X_{q \shortto d} = \{f_Q(q) | q \in \mathcal{Q}\} \bigcup \{f_D(d) | d \in \mathcal{D}\}$ and $X_{d \shortto q} = \{h_Q(q) | q \in \mathcal{Q}\} \bigcup \{h_D(d) | d \in \mathcal{D}\}$.
Then, $(X_{q \shortto d}, \otimes)$ and $(X_{d \shortto q}, \otimes)$ define two distinct metric spaces corresponding to the embedding-based document retrieval system and exposing query retrieval system, respectively.
To retrieve exposing queries for a document $d$, we need to perform a reverse nearest neighbor (rNN) search in the $(X_{q \shortto d}, \otimes)$ metric space.
However, rNN solutions are inefficient for high-dimensional and dynamic models.
We instead propose to learn a new metric space $(X_{d \shortto q}, \otimes)$ where a NN search would approximate the rNN search in the $(X_{q \shortto d}, \otimes)$ space.
\subsection{Training data generation}
\label{sec:ance-training-data}
Generating training data is a key challenge in this learning task.
To train the model, we need a dataset of documents and their exposing queries.
We thus need to devise a training data generation strategy which ensures that: (1) we only issue a limited number of queries to the document retrieval model to keep the approach computationally efficient, (2) we have at least one exposing query for each training document, and (3) the training data represents the distribution of queries around each training document. The pseudocode for the proposed procedure is outlined in Algorithm \ref{alg:train-data}.
Training data is ensured to be representative and high-quality under strict efficiency constraints as we only conduct document retrieval for the training queries (line~\ref{algo:line:5}). By selecting training documents from the candidate set $\mathcal{D}_{\text{candidate}}$ (line~\ref{algo:line:9}), we guarantee that each training document has at least one exposing query that retrieves it in top ranks. We then do a KNN search from $Q_{\text{train}}$ of the queries closest to each training document in the document retrieval embedding space (line~\ref{algo:line:12}). This gives us a sample distribution of queries around each document and the positive and negative document-query pairs will match the current errors in the document retrieval embedding space (where similarity between document and query does not necessarily reflect the exposure).
Note that from $\Psi_{\text{train}}$, as constructed in Algorithm \ref{alg:train-data}, we can sample two types of training instances of form $<d, q^+, q^->$, corresponding to:
\begin{enumerate*}[label=Case \arabic*:,leftmargin=4\parindent]
\item $\rho(d,\sigma_{q^+}) < \rho(d,\sigma_{q^-}) < n_{q \shortto d}$, and
\item $\rho(d,\sigma_{q^+}) < n_{q \shortto d} \le \rho(d,\sigma_{q^-})$.
\end{enumerate*}
During training, we randomly sample training instances for Case 1 and Case 2 with probability $\alpha$ and $1 - \alpha$, respectively. Intuitively, Case 1 helps the model better rank the exposing queries already retrieved by the document retrieval embedding space, while Case 2 helps the model learn to retrieve exposing queries that are far away from the document in the starting document embedding space.
\begin{algorithm}[t]
\caption{Training data generation}
\label{alg:train-data}
\begin{algorithmic}[1]
\STATE Randomly sample $\mathcal{Q}_{\text{train}}$ from $\mathcal{Q}$
\STATE Initialize $\mathcal{D}_{\text{candidate}} = \emptyset$
\STATE Initialize cache results $\mathcal{C}=\{\}$
\FORALL{$q_\text{train} \in \mathcal{Q}_{\text{train}}$}
\STATE Retrieve top $n_{q \shortto d}$ ranked list of documents $\sigma_{q_{\text{train}}}$ from $\mathcal{D}$ \label{algo:line:5}
\STATE Cache retrieval results $\mathcal{C}[q_{\text{train}}]=\sigma_{q_{\text{train}}}$
\STATE $\mathcal{D}_{\text{candidate}} \leftarrow \mathcal{D}_{\text{candidate}} \bigcup \; n_{q \shortto d} \; \text{documents in} \; \sigma_{q_{\text{train}}}$
\ENDFOR
\STATE Randomly sample $\mathcal{D}_{\text{train}}$ from $\mathcal{D}_{\text{candidate}}$ \label{algo:line:9}
\STATE Initialize training dataset $\Psi_{\text{train}} = \{\}$
\FORALL{$d_\text{train} \in \mathcal{D}_{\text{train}}$}
\STATE Retrieve top $n_{d \shortto q}$ ranked list of queries $\psi_{d_{\text{train}}}$ from ${\mathcal{Q}_{\text{train}}}$ \label{algo:line:12}
\STATE Initialize $\beta_{d_{\text{train}}}=\{\}$
\FORALL{$q \in \psi_{d_{\text{train}}}$ }
\IF{$q \in \mathcal{C}[q]$}
\STATE Get document rank $\rho(d_{\text{train}}, \sigma_q)$ from $\mathcal{C}[q]=\sigma_q$
\ELSE{}
\STATE Label $\rho(d_{\text{train}}, \sigma_q)$ as $\infty$
\ENDIF
\STATE $\;\;\;\; \beta_{d_{\text{train}}}[q]=\rho(d_{\text{train}}, \sigma_q)$
\ENDFOR
\STATE Add to the training dataset $\Psi_{\text{train}} [d_{\text{train}}]=\beta_{d_{\text{train}}}$
\ENDFOR
\RETURN $\Psi_{\text{train}}$
\end{algorithmic}
\end{algorithm}
\subsection{Model architecture}
\label{sec:ance-network}
We adopt the publicly available\footnote{\href{https://github.com/microsoft/ANCE}{https://github.com/microsoft/ANCE}}
BERT-based dual-encoder architecture by \citet{xiong2020approximate} as our dual-encoder search system.
The model is trained in an active metric learning setting~\citep{kumaran2018active, yang2012bayesian, ebert2012active}, optimizing towards the so-called approximate nearest neighbor negative contrastive estimation (ANCE) objective.
As far as we are aware, the model trained on the ANCE loss has demonstrated a state-of-the-art performance by a dual-encoder system. We therefore adopt this model as our base architecture for both the document and the exposing query retrieval stages.
Unless specifically mentioned, we use the same hyperparameters as in the original paper.
We anticipate that both models---for the document retrieval and the exposing query retrieval tasks---need to learn a fundamental notion of topical relevance between queries and passages that are largely similar in both cases.
Therefore, we pre-initialize the exposing query retrieval model with the learned parameters of the document retrieval model before training.
Furthermore, we add a few additional layers on top of the query and document encoders. However, efficiency constraints limit the amount of data we can use for training the extra layers. To avoid overfitting, we design the network architecture to only slightly transform the original embedding space. In particular, we explore following two settings:
\textit{Append:}
Let $\vec{x}$ be the vector output of the base encoder, which for our base model has a size of $768$.
Then, under the Append settings, the final embedding output is $\vec{y} = \big(\vec{x}, \text{FF}(\vec{x})\big)$, where $(,)$ denotes a concatenation of two vectors, and FF is a feedforward network.
Specifically, in our experiments we employ a four-layer feedforward neural network, with hidden layer sizes of $64$ and the last output dimension as $32$.
Consequently, after concatenation the final embedding vector dimensions is $800$.
For each hidden layer of the sub-network, we apply a ReLU nonlinearity, a dropout layer, and a layer normalization.
\textit{Residual:}
Using a similar notation, the output of the residual layer~\cite{he2016deep} can be written as $\vec{y} = \vec{x} + \text{FF}(\vec{x})$.
The feedforward network FF in this case comprises a single hidden layer of size $384$ and an output vector size of $768$.
We add ReLU nonlinearity, dropout, and layer normalization similar to the Append setting.
\subsection{Training and retrieval}
\label{sec:ance-train}
We train all models, including baselines, for 1K iterations, where each iteration consists of $100$ batches with $1000$ samples per batch. We use dot product as the similarity measurement as in the ANCE model. We use a pairwise optimization objective in the form of the cross entropy loss.
We set $\alpha$, as introduced in Section \ref{sec:ance}, to 0.5 and the dropout rate to $0.1$.
We use Adam optimizer with a fixed learning rate of 1e-4---and set the coefficients beta1 to 0.9, beta2 to 0.999, and eps to 1e-08. Unless otherwise specified, all results reported in Section \ref{sec:result} are obtained from models trained on 50K queries and 200K passages.
For experimentation agility under limited GPU budget, we freeze the base encoders and only optimize the parameters of the FF layers on top---under both Append and Residual settings---when training for the exposing query retrieval task.
We expect that additional improvements may be possible from fine-tuning the encoders end-to-end but we do not explore that direction in our current work.
We use the Faiss IndexFlatIP Index \cite{johnson2019billion}, which guarantees efficient computation of exact nearest neighbors, for nearest neighbor search in the context of both our baseline and proposed models, as well as for the ground truth results generation.
\section{Experiments}
\label{sec:experiment}
\subsection{Data}
\label{sec:method-data}
For EQI experimentation, we need a large-scale dataset accompanied by a large query collection. We thus use the MS MARCO passage retrieval benchmark~\citep{bajaj2016ms}, extensively used to compare neural methods for retrieval, including in the TREC Deep Learning track~\citep{craswell2020overview}.
The MS MARCO dataset contains $8,841,823$ passages and $532,761$ relevant query-passage pairs for training, as labeled by human assessors.
\subsection{Methods}
\label{sec:experiment-baseline}
\paragraph{Brute-Force (Exact EQI) Baseline}
An obvious baseline for our work corresponds to the brute-force approach of running every query in our log through the document retrieval system and recording all the retrieved results.
Then for a given passage, we can iterate over all the document retrieval results to produce an exact solution to the exposing query retrieval problem.
While the brute force method achieves perfect recall of all exposing queries, it also incurs both a heavy computation cost for the document retrieval step.
\paragraph{Original Retrieval Space Methods.}
For traditional search systems, an intuitive baseline model can index the query log as a document collection and use documents as queries. We refer to this method as BM25-reverse. We also tune the term weighting parameters $k_1$ and $b$, which we refer to as BM25-tuned. In the context of dual-encoder search systems,
we expect that, even in the document retrieval embedding space, exposing query and document pairs would be relatively close to each other, and a nearest neighbor search around a passage in this space would identify some of its exposing queries. We employ this approach as a baseline, which we refer to as the ANCE-reverse in the remainder of this paper.
\paragraph{Metric learning. } We also implement the metric learning approaches described in Sec.~\ref{sec:ance}, referring to the two architectures as ANCE-append and ANCE-residual.
\subsection{Ground truth end experimental protocol}
\label{sec:method-data}
In our experiments, we set $n_{q\shortto d}=100$ and $n_{d\shortto q}=100$.
To obtain the ground truth exposing queries we obtained a ranked list $\sigma^*_q$ of the top 100 passages for each query $q$ in the collection.
To evaluate the EQI models, for each model and each passage $d$ in the test set of size $20,000$, we retrieve the top 100 queries as the exposing ranked list $\psi_d$.
\section{Results}
\label{sec:result}
\subsection{Overall performance}
\label{sec:model_performance}
\begin{table}
\small
\begin{center}
\caption{Model performance on the EQI task benchmarked using the MS MARCO dataset.
We report the RELQ metric corresponding to four different user model combinations.
}
\label{tab:main_results}
\begin{subtable}{\columnwidth}
\begin{tabular}{l c c c c}
\toprule
& \multicolumn{3}{c}{$\text{RELQ}_\text{RBP,RBP}$ with ($\gamma_{q \shortto d}$, $\gamma_{d \shortto q}$)}& $\text{RELQ}_\text{EXH,NDCG}$\\
Model & $(0.5, 0.5)$ & $(0.5,0.9)$ & $(1,1)$ \\
\midrule
BM25-reverse & $0.441$ & $0.624$ & $0.840$ & $0.645$ \\
BM25-tuned & $0.442$ & $0.626$ & $0.845$ & $0.648$ \\
Brute force & $1.000$ & $1.000$ & $1.000$ & $1.000$\\
\bottomrule
\end{tabular}
\caption{EQI for BM25}
\end{subtable}
\begin{subtable}{\columnwidth}
\vspace{4px}
\begin{tabular}{l c c c c}
\toprule
& \multicolumn{3}{c}{$\text{RELQ}_\text{RBP,RBP}$ with ($\gamma_{q \shortto d}$, $\gamma_{d \shortto q}$)}& $\text{RELQ}_\text{EXH,NDCG}$\\
Model & $(0.5, 0.5)$ & $(0.5,0.9)$ & $(1,1)$ \\
\midrule
ANCE-reverse & $0.493$ & $0.685$ & $0.887$ & $0.698$ \\
ANCE-append & $0.624$ & $0.825$ & $0.982$ & $0.837$ \\
ANCE-residual & $0.633$ & $0.834$ & $0.984$ & $0.845$ \\
Brute force & $1.000$ & $1.000$ & $1.000$ & $1.000$ \\
\bottomrule
\end{tabular}
\caption{EQI for ANCE}
\end{subtable}
\end{center}
\end{table}
Table~\ref{tab:main_results} summarizes the EQI performance results on the MS MARCO benchmark, using different parametrizations of the RELQ metric.
The brute force baseline, by definition finding all exposing queries, achieves a perfect value of the RELQ metric.
In the context of traditional search models, BM25-tuned consistently shows a slightly better performance than BM25-reverse according to all metrics. In the context of dual-encoder models, both the proposed metric learning approaches (ANCE-append and ANCE-residual) significantly outperform the ANCE-reverse under different parametrizations of RELQ.
The improvements of both models against ANCE-reverse are statistically significant according to a one-tailed paired t-test ($p < 0.01$).
Of the two models, Residual performs better than Append, and the difference in their performance are again statistically significant based on one-tailed paired t-test ($p < 0.01$).
Therefore, we only use the Residual model in the remaining analyses.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{event_update.pdf}
\caption{Visualization of retrieved exposing query lists. The figure presents retrieved ranked lists of exposing queries for four test passages. Dotted lines represent the exposing queries retrieved by both models. The number reported is the $RELQ_{RBP,RBP}$ with $\gamma_{q\shortto d}=0.5, \gamma_{d\shortto q}=0.9$. The x-axis represents the ranking position in the exposing query retrieval model $\rho(q,\psi_d)+1$ and the y-axis represents the rank at which the query retrieves the test passage in the document retrieval system $n_{q\shortto d}-\rho(d,\sigma_q)$ (the higher, the more exposing).}
\label{fig:example}
\end{figure}
\paragraph{Illustration. }Beyond the mean metric values, we illustrate the behavior of ANCE-reverse and ANCE-residual on four selected test passages in Figure \ref{fig:example}.
These passages were selected to reflect different performance levels.
For each passage, we compare the ranked exposing query lists retrieved by the ANCE-Reverse and the Residual models. Recall that $n_{d\shortto q}=100$.
Vertical bars denote exposing queries, with the dotted vertical bars showing queries retrieved by both models.
The x-axis represents the ranking position of the query in the EQI model, while the y-axis represents the ranking position at which the document retrieval system retrieves the passage for a given query.
More precisely, we plot $n_{q\shortto d}-\rho(d,\sigma_q)$, such that higher bar indicates that the document was retrieved at a higher rank for that query.
An ideal ranked list should arrange the vertical bars from the highest (queries exposing the passage at top ranking positions) to the lowest with no gap in between.
Except for Passage 3, the Residual model retrieves more exposing queries and ranks them higher than the baseline Reverse model. This shift in exposing queries in the ranked lists indicates that the proposed learning algorithm can learn a new metric space that can recover more exposing queries.
For Passage 1 and Passage 2, the Residual model furthermore retrieves many exposing queries which are originally not within the 200 nearest neighbors under the reverse baseline model. This example illustrates clearly that the nearest neighbor problem and the reverse nearest neighbor problem are not symmetric.
Visualization of Passages 3 and 4 illustrates that, when a passage has fewer exposing queries overall, the RELQ metric tends to be more sensitive to the ranks $\rho(q,\psi_d)$ of highly exposing queries. This observations may, however, change under different user behavior assumptions for the RELQ metric. We analyze this aspect in the Appendix~\ref{subsec:user_model}.
\subsection{Impact of the training data size}\label{subsec:sample_size}
Approximate EQI trades exactness for performance; in the context of dual-encoder models, performance depends on the ability to train with little data. We examine the impact of training data size on model performance.
Table~\ref{tab:train_size} shows the results for the Residual model trained for $1000$ epochs.
We report $RELQ_{RBP,RBP}$ with $\gamma_{q \shortto d}=0.5, \gamma_{d \shortto q}=0.9$.
We find that increasing the number of both queries and training passages leads to diminishing returns. One-tailed paired t-tests are conducted along each column and each row. Except for the case, where the number of training queries is fixed at 500K and the number of training passages is increased from 400K to 600K, all increases in performance are statistically significant ($p < 0.01$) after Bonferroni correction. The Residual model trained on even a relatively small training data---$50$K queries and $200$K passages---shows good performance.
\begin{table}
\small
\begin{center}
\caption{Impact of the training sample size. Residual model performs better when either the sample size of training queries or training passages increases.}
\label{tab:train_size}
\begin{tabular}{l|l|c|c|c}
\hline
& & \multicolumn{3}{c}{Passages} \\
\hline
& & $200,000$ & $400,000$ & $600,000$ \\
\hline
\multirow{3}[0]{*}{\rotatebox{90}{Queries}} & $50,000$ & 0.834 & 0.854 & 0.859
\\
& $250,000$ & 0.872 & 0.877 & 0.880
\\
& $500,000$ & 0.879 & 0.882 & 0.883
\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.32\textwidth]{evolve_qsize.pdf}
\caption{Each line is generated from a Residual model trained on a particular size of training queries. The starting point of each line is to train and test on the same query log. Along the line, query log is gradually expanding.}
\label{fig:qset_evol}
\end{figure}
\vspace{.3em}
\begin{figure*}
\centering
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{prob_length_update.pdf}
\caption{By query length}
\label{fig:prob-length}
\end{subfigure}
\hfill
\begin{subfigure}{0.30\textwidth}
\centering
\includegraphics[width=\textwidth]{prob_df_update.pdf}
\caption{By document frequency of rarest query term}
\label{fig:prob-df}
\end{subfigure}
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=\textwidth]{legend_vertical.png}
\end{subfigure}
\caption{Probability of being an exposing query for one of the test passages vs probability of retrieval.}
\label{fig:prob}
\end{figure*}
\begin{figure}
\centering
\tiny
\includegraphics[width=0.42\textwidth]{log_log_two_models.pdf}
\caption{Rank-size distribution of the number of ground truth and retrieved exposing queries per passage. A few passages have a large number of exposing queries but there is also a long tail of passages with few or no exposing queries.}
\label{fig:log_log}
\end{figure}
\subsection{Impact of expanding query collections}\label{subsec:expand_query}
The search ecosystem is dynamic. Searchers may issue previously unknown queries, while content producers may create new documents.
Note that the evaluation in Section~\ref{sec:model_performance} already covers the scenario in which exposure needs to be computed for new documents, as none of the test passages were seen or present in the corpus during training. We additionally examine the effects of a growing query collection.
Our goal is to see how long the EQI model can sustain good retrieval performance without retraining as the query collection grows. This process is simulated by first sampling an initial set of queries available during training, and then test using a gradually expanding collection by embedding new queries without model retraining. The number of training passages is fixed as 200,000. Four Residual models are then trained for 1000 epochs on different sizes of a starting query collections: 50,000, 150,000, 250,000 and 350,000.
Note that this experimental protocol implies that new queries are drawn from the same distribution as existing queries.
Simulation results are illustrated in Figure \ref{fig:qset_evol}.
We report $RELQ_{RBP,RBP}$ with $\gamma_{q\shortto d}=0.5, \gamma_{d\shortto q}=0.9$.
\vspace{.5em}
\noindent\textbf{Decreasing concave trend along each line:} The decrease in model performance generally signals the need for model retraining.
In different real-world applications, service providers may have different tolerance levels to performance loss vs the retraining costs.
A visualization like the one in Figure \ref{fig:qset_evol} can be used as a monitoring tool for deciding when to retrain as the query collection expands.
The concave shape indicates that the model performance stabilizes. Recall that ANCE-Reverse achieves a RELQ of only 0.685 on the full query collection.
We reckon that the Residual model might eventually converge to a performance level still outperforming the Baseline model should the query collection increase further.
\vspace{.5em}
\noindent \textbf{Decreasing rate drop with larger training sample size:} The performance curves are flatter for models initially trained with more queries.
These models capture the characteristics of a fully expanded query collection better.
\vspace{.5em}
\noindent \textbf{Increasing task difficulty:} Both the decreasing trend along each curve and the decreasing starting value for each curve might suggest that the difficulty of the EQI task increases with the size of the query collection.
Interestingly, an opposite observation has been made about document retrieval systems by~\citet{hawking2003collection} who showed that the document retrieval precision increases with increasing collection size due to the sparsity of relevant documents in sub samples.
Reconciling this difference requires further analysis.
We hypothesize that the difference may be due to the fact that \citet{hawking2003collection} performed their analysis on term-matching-based retrieval models, whereas recent evidence shows that in the context of neural models an increase in the collection size may lead to a decrease in retrieval effectiveness~\citep{zamani2018neural}.
\subsection{Impact of query characteristics}
\label{subsec:query_analysis}
\vspace{.5em}
\noindent \textbf{Probability of retrieval vs. query length:}
We investigate whether the EQI models are more likely to retrieve longer queries.
As shown in Figure \ref{fig:prob-length}, under both BM25 and ANCE, the probability of being an exposing query for one of the test passages is slightly higher in case of shorter queries, \emph{i.e.}, queries with fewer than 10 words.
As for the top-10 retrieval probability, we find that both ANCE-reverse and ANCE-residual tend to slightly under-retrieve shorter queries and over-retrieve longer queries, while BM25-reverse retrieves equally regardless of query length. However, for the top-$100$ retrieval probability, we observe that all models are likely to significantly over-retrieve queries, in particular shorter ones.
\vspace{.5em}
\noindent \textbf{Probability of retrieval vs. query term frequency:}
We investigate whether the models are able to retrieve both rare and common queries.
We measure the rarity of a query by computing the minimum document frequency of the query terms. This analysis sheds a light on the usefulness of the approach in providing privacy-awareness, as rare queries can contain personally identifying information.
Figure~\ref{fig:prob-df} demonstrates that ANCE-reverse, ANCE-residual and BM25-reverse tend to retrieve such queries at a rate roughly proportional to their probability of being exposing for one of the test passages.
\vspace{.5em}
\noindent\textbf{Distributions of exposing queries per passage:}
Figure \ref{fig:log_log} shows a log-log plot of the distribution of the number of exposing queries (y-axis) per passage.
For all the methods, we find that there is a long tail of passages with very few or no exposing queries, while some passages may have tens, or even hundreds, of exposing queries. This result suggests the strong retrievability bias~\cite{azzopardi2008retrievability}, and is in line with prior work on search exposure~\cite{biega2017learning}.
\subsection{Discussion}
The results reported earlier in this section indicate that reversing the role of queries and documents can be a reasonable lightweight strategy for EQI, in particular when the model parameters are further optimized for the query retrieval task in the form of metric learning.
However, deploying an EQI system in the wild requires several other practical considerations.
While Section~\ref{subsec:sample_size} tells us that we can achieve reasonable effectiveness in exposing query retrieval using moderately sized training datasets (~50K queries), the required amount of training will depend on the size of the query corpus from which we are retrieving, as seen in Section~\ref{subsec:expand_query}.
As we can produce almost infinitely large training data for the query retrieval model automatically using the document retrieval model, the size of required training dataset itself may not be a big concern but may still have implications for training time of the query retrieval model and thus the attractiveness of EQI as a lightweight alternative to brute force aggregation.
Another important concern may be around potential systemic biases in retrieved queries.
In Section~\ref{subsec:query_analysis}, we notice that the methods we study retrieve queries roughly in proportion to their probability of being exposing queries independent of query term rarity which is promising in terms of our ability to retrieve tail queries, and thus downstream privacy applications.
However, these models disproportionately over-retrieve shorter queries which may have implications for downstream applications such as quantifying the retrievability bias, or enforcing the right to be forgotten. Studying these effects is an interesting direction for future work.
\section{Conclusions}
\label{sec:conclusion}
This paper focused on the search transparency problem of exposing query identification, that is, finding queries that expose documents in the top-ranking results of a search system. We studied the feasibility of approximate EQI as a retrieval problem for two classes of search systems (dual-encoder and traditional BM25 systems). We further demonstrated the potential of improving baseline EQI approaches through metric learning.
We also derived an evaluation metric called Ranked Exposure List Quality (RELQ) that jointly models the behavior of two users who influence the EQI outcomes: a user of the EQI system and the document retrieval user.
Many promising future directions remain to be explored. Examples include developing better network architectures, algorithms for EQI in black-box search systems or for document retrieval models that use external click signals for ranking, developing generative models for query collections better grounded in downstream exposure tasks, as well as domain-specific quantitative and qualitative application and user studies. Our work provides a conceptual and empirical foundation for further study of EQI, which can contribute to advances in various areas of search and ranking: bias and fairness, privacy, data protection, security, and SEO.
\section{APPENDIX}
\begin{figure*}[th!]
\centering
\begin{subfigure}{0.99\columnwidth}
\includegraphics[width=\textwidth]{baseline_heatmap.pdf}
\caption{ANCE-reverse}
\label{fig:baseline_heatmap}
\end{subfigure}
\begin{subfigure}{0.99\columnwidth}
\includegraphics[width=\textwidth]{model_heatmap.pdf}
\caption{ANCE-residual}
\label{fig:model_heatmap}
\end{subfigure}
\caption{Model performance (RELQ values) for different user behavior models (the EQI system user on the x-axis, and the document retrieval user on the y-axis). The darker the color, the better the performance.}
\label{fig:heatmap}
\end{figure*}
\subsection{Impact of metric parametrization}
\label{subsec:user_model}
Under different user behavior assumptions, the same exposing query list will achieve different absolute value of RELQ.
We explore the relationship between user behavior and model performance for different combinations of user patience parameters $\gamma_{q\shortto d}$ (document retrieval user) and $\gamma_{d\shortto q}$ (EQI user).
Figure~\ref{fig:heatmap} presents the resulting performance heatmaps for ANCE-reverse and ANCE-residual.
Both heatmaps exhibit similar patterns.
Higher $\gamma_{q \shortto d}$ generally correlates with a high value of RELQ, except for very high values of $\gamma_{d \shortto q}$, such as $0.9$ or $1.0$, where a higher value of RELQ is observed for lower values of $\gamma_{q \shortto d}$.
These patterns are expected. Having a more patient user of the document retrieval system (high $\gamma_{q \shortto d}$) will mean more exposing queries per document: as the user goes deeper down the rank list more documents will get exposed to the query. Intuitively then, when the user of the EQI system is impatient, the presence of many ground truth exposing queries makes the task easier.
However, when the EQI user is very patient and inspects the exposure set exhaustively, fewer ground truth exposing queries (impatient document retrieval user) do not lead to a higher task difficulty, and in this case, we indeed observe the highest absolute RELQ values.
ANCE-residual outperforms the ANCE-reverse model under all metric parametrizations.
Notably, however, ANCE-reverse performs very well for patient EQI users ($\gamma_{d \shortto q}=1.0$).
This result suggests that even though both models are able to retrieve many exposing queries, the ANCE-residual model tends to rank the exposing queries higher.
|
\section{Introduction}
Unsteady locomotion of flying animals such as birds and insects have inspired the development of micro air vehicles, which are less than \SI{15}{\centi\meter} in size and operate in a Reynolds number range from \numrange{10}{10000} \cite{shyy_flapping_1999,shyy_aerodynamics_2011,shyy_recent_2010}.
Their field of application includes information gathering in confined spaces, areal mapping, or the transport of small goods \cite{floreano_science_2015,taylor_experimental_2010}.
Flapping wing configurations are a suitable manner of propulsion for micro air vehicles, since they offer better efficiency at low Reynolds number ($\Rey < \num{100}$) and an improved manoeuvrability in comparison to fixed and rotating wing configurations \cite{bayiz_hovering_2018,hawkes_fruit_2016}.
The development of complex flapping wing kinematics for various flight scenarios poses a challenge in this field.
A similar challenge is the selection of optimal motion kinematics of other bio-inspired propulsion systems such as underwater vehicles mimicking fish propulsion.
A standardised optimisation procedure facilitates the identification of motions with desired characteristics.
The motion kinematics have to be parameterised by a finite number of variable parameters prior to optimisation to reduce the dimensionality of the solution space and allow for optimal solutions to be found in an affordable way within a reasonable timeframe.
The optimisation solution landscape can dependent strongly on the parameterisation and possible optima might be missed \cite{mandre_optimizing_2019}.
The ideal parameterisation has as little variable parameters as necessary to reduce the computational complexity of the optimisation problem without restricting the solution space or creating a bias in a certain direction.
Yet, any parameterisation will exclude possible solutions due to discretisation in comparison to a continuous solution in an infinite-dimensional solution space and it is vital to carefully select the most suitable approach for each optimisation problem at hand.
Past efforts to optimise bio-inspired motion kinematics initially focussed on simple linear and harmonic motions.
\citet{tuncer_optimization_2005} numerically optimised a flapping airfoil for combined maximum thrust and efficiency.
They parameterised the sinusoidal plunge and pitching motions with a variable amplitude and phase shift for a fixed frequency.
\citet{berman_energy-minimizing_2007} implemented a quasi-steady model of insects with a hybrid optimisation algorithm.
Their kinematic function enables a continuous transition of the pitching angle evolution $\beta$ with the parameter $C$ between sinusoidal and trapezoidal motions according to
\begin{equation}
\beta(t) = \frac{\kindex{\beta}{m}}{\tanh(C)} \tanh\left(C \sin(2\pi f t + \Phi)\right) + \kindex{\beta}{0}
\end{equation}
with $f$ being the flapping frequency, $\Phi$ the phase shift between the stroke and pitching motions, $\kindex{\beta}{0}$ the offset angle and a scaling factor $\kindex{\beta}{m}$ for further modulations.
The temporal evolutions of the stroke and elevation angles were defined similarly, yielding a total of twelve variable parameters.
Solutions for these twelve parameters were determined that maximised the aerodynamic efficiency while providing enough lift to support the body weight of different insects.
The aerodynamic performance can be further increased by more complex, non-harmonic, and asymmetric kinematics, which have to be defined using more parameters or different base functions.
The pitching kinematics of insects in nature are often asymmetric which could be an evolutionary adaption to improve the performance \cite{liu_size_2009}.
\citet{martin_experimental_2018} were among the first to implement a kinematic motion function with the possibility to represent asymmetric motions.
\citet{liu_size_2009} performed an optimisation of the motion of a pectoral fish fin which was defined as a complex combination of trigonometric expressions for all three spatial angles.
The motion kinematics were successfully optimised for a combined objective of minimal side-thrust and maximal efficiency.
They noticed an increasing difficulty in finding the global optimum with increasing number of parameters describing the kinematics.
\citet{mandre_optimizing_2019} investigated the influence of the parameterisation with a finite Fourier series on a heaving and pitching hydrofoil.
Small modifications of the parameterisation can lead to a significantly altered solution landscape with a bias towards small but pronounced local optima.
\citet{gehrke_phenomenology_2021} studied pitching motions with increased complexity, which are able to mimic more closely the motions observed in nature.
The stroke motion was fixed and the elevation motion was omitted.
The pitching angle was defined by a set of control points in the angle-time domain, which were connected with a third-degree spline.
The resulting optimal pitching kinematics forming the Pareto front in the stroke average lift versus efficiency diagram vary significantly and highlight the potential of asymmetric and more complex kinematics to push the performance envelope of bio-inspired micro aerial vehicles.
Here, we will present three different approaches to define and parameterise kinematics for optimisation studies.
The different approaches will be presented, compared, and evaluated for the example of the experimental optimisation of the pitching kinematics of a flapping wing, building upon the work presented in \citet{gehrke_genetic_2018} and \citet{gehrke_phenomenology_2021}.
An optimisation is performed for each kinematic function with $\num{6},\num{12}$ and $\num{18}$ parameters, leading to more than \num{30000} individual experiments in total.
The different approaches to parameterise motion kinematics for optimisation selected here are:
\begin{itemize}
\item control points connected by a spline interpolation,
\item a finite Fourier series, and
\item a linear combination of modes determined by a modal decomposition of kinematics created by a random walk
\end{itemize}
\section{Materials and Methods}
\subsection{Experimental set-up}
The kinematics for the different optimisation studies are evaluated with a robotic flapping wing mechanism immersed in an octagonal water tank (see also~\cite{gehrke_genetic_2018, gehrke_phenomenology_2021}).
The wing is a rectangular flat plate with a chord of $c = \SI{34}{\milli\metre}$ and a span of $R = \SI{107}{\milli\metre}$ (\cref{fig:experimental_setup}).
The flapping frequency of the system is $f = \SI{0.25}{\hertz}$ with a peak-to-peak stroke amplitude of $\phi = \ang{180}$.
The flapping frequency and stroke amplitude remain constant throughout this study and lead to a Reynolds number of $Re = 2 \phi f c \kindex{R}{2} / \nu = 4895$ at the radius of the second moment of area $\kindex{R}{2}$, for a kinematic viscosity of $\kindex{\nu}{\SI{20}{\celsius}} = \SI{1.00e-6}{\meter\squared\per\second}$ for all experiments~\cite{gehrke_phenomenology_2021, sum_wu_scaling_2019}.
These characteristic parameters are selected to match typical flapping wing vehicles and larger insects such as the hawk moth~\cite{liu_size_2009, shyy_recent_2010}.
The wing is actuated by two servo motors (Maxon motors, type RE35, \SI{90}{\watt}, \SI{100}{\newton\milli\meter} torque, Switzerland) at the top of a 2-axis shaft system which controls the pitch and stroke angles of the wing (\cref{fig:experimental_setup}).
The wing moves for a total tip-to-tip amplitude of \SI{0.47}{\metre} which results in a \SI{6.91}{c} tip clearance to the tank for an outer diameter of \SI{0.75}{\metre}.
Previous studies have confirmed that no wall effects are present for a tip clearance larger than \SI{5}{c}~\cite{manar_tip_2014, krishna_flowfield_2018}.
The flapping wing apparatus was specifically designed to run unsupervised for large amounts of time and to execute complex motion profiles with high repeatability.
Preliminary tests have determined the maximum error between prescribed motion and motor response on the mechanism to be $< \ang{0.1}$ for high acceleration kinematics~\cite{gehrke_phenomenology_2021}.
The motors are controlled by a motion control board (DMC-4040, Galil Motion Control, USA).
A six-axis force and torque transducer (Nano17, ATI Industrial Automation, USA) is installed at the wing root in \cref{fig:experimental_setup}b to capture the time resolved aerodynamic loads.
The sensor is calibrated with \textit{SI-12-0.12} to record at a resolution of \SI{3.13}{\milli\newton} for force and \SI{0.0156}{\newton\milli\meter} for torque measurements.
A data acquisition card (National Instruments, USA) is used to record the force and torque transducer signal at a sample frequency of \SI{1000}{\hertz}.
The lift and power data was filtered with a zero phase delay low-pass \num{5}th order digital Butterworth filter for the phase-averaged, time-resolved plots.
The filter's cut-off frequency was selected to be \num{12} times the hovering frequency of the system.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_angles_wing}
\caption{a. Flapping wing mechanism with stroke angle $\phi$ and pitch angle $\beta$ for the hovering motion, b. wing geometry and axis of rotation, c. pitch angle $\beta$ kinematics of various natural fliers, adopted from~\cite{liu_size_2009} and trapezoidal pitch angle kinematics from a robotic flapper~\cite{Bhat.2020}.}
\label{fig:experimental_setup}
\end{figure}
\subsection{Kinematic functions}
Three different approaches to parameterise motion kinematics are selected here:
\begin{inparaenum}[a)]
\item control points connected by a spline interpolation,
\item a finite Fourier series, and
\item a linear combination of modes determined by a modal decomposition of kinematics created by a random walk.
\end{inparaenum}
The different parameterisation were selected by their ability to mimic trapezoidal and sinusoidal pitch angle profiles commonly seen on robotic flapping wing devices as well as more complex pitch angle kinematics observed on nature's fliers (\cref{fig:experimental_setup}c).
We created kinematic functions using the three approaches with three different parameter counts ($p = 6,12, 18$).
The first $(p-1)$ parameters are used to modify the pitching angle evolution and the phase shift is controlled with the last parameter $\kindex{\Delta t}{0}$ for all kinematic functions.
The motions are described by the pitching angle $\beta$ (\cref{fig:experimental_setup}).
The sinusoidal stroke angle $\phi$ remains unchanged throughout the optimisations and the elevation angle is kept at zero due to its negligible influence on the lift production \cite{liu_size_2009}.
\subsubsection{Control points connected by splines}
\begin{figure*}[tb!]
\centering
\includegraphics[width=\linewidth]{fig_spline_kin}
\caption{Example of the parameterisation of the pitch angle kinematics using control points connected by a spline interpolation with $p=6$ parameters.
Five parameters describe the pitch angles ($\kindex{\beta}{1}, ..., \kindex{\beta}{5}$) of the five control points that are equidistantly distributed within a half stroke.
A fifth degree spline connects the five control points and the half half stroke profile is point mirrored to create the full symmetric pitch angle profile in grey.
The sixth parameter $\kindex{\Delta t}{0}$ introduced a phase shift leading to the final pitching profile in colour.}
\label{fig:pchip_definition}
\end{figure*}
As a first intuitive approach, we select a limited number of control points that will be connected using spline interpolation.
To start, we create a symmetric pitch profile by distributing $(p-1)$ control points equidistantly throughout a half-stroke which represents the most efficient use of the available parameters.
Freely spaced control points could lead to large gradients for higher parameter counts which can not be executed by our experimental set-up.
In a way, the temporal spacing of the control points acts as a build-in low-pass filter for the pitch angle gradients.
If the phase of the control points is not fixed, the phase bounds of the individual points depend on each other which makes the selection of the kinematics slightly more complicated.
Depending on the specific problem at hand, it is possible to place more control points in certain parts of the motion where larger variations are required to optimise and reduce the parameter count.
The beginning and end of the half-stroke are fixed at $\beta=\ang{0}$ and the control points are solely specified by their pitch angle \kindex{\beta}{n}, with $1\leq n\leq(p-1)$ as indicated in \cref{fig:pchip_definition}.
Next, the control points are connected by a fifth degree spline interpolation to minimise the local acceleration $\ddot{\beta}$.
By point-mirroring the first half-stroke around $(t/T,\beta)=(0.5,0)$, we obtain a symmetric pitching profile for the back and forth stroke, represented by the grey curve in \cref{fig:pchip_definition}.
Finally, we added a phase-shifted \kindex{\Delta t}{0} to allow for advanced and delayed wing rotation with respect to the stroke reversal, leading to the coloured curve in \cref{fig:pchip_definition}.
This procedure allows us to created complex non-linear and non-harmonic motions with high curvatures.
The main advantage of the control point parameterisation, is its intuitive and direct local control of the pitching profiles.
Local adjustments can be made by moving individual control points and the values of the control points are directly linked to the pitch angle values, which facilitates the formulation of pitch angle constraints due to mechanical and motor limitations.
The distribution of the control points along the time axis can be easily varied to adapt the approach to specific restrictions or requirements for a broad range of applications.
On the downside, we noticed that the creation of a random set of control point splines is computationally demanding and required $\SI{22.8}{\second}$ on average for motions with $12$ parameters on a typical desktop computer.
The computational time increases exponentially with the parameter count and takes a couple of minutes for the tests with $18$ parameters.
\subsubsection{Fourier-series kinematics}
\begin{figure*}[tb]
\centering
\includegraphics[width=\linewidth]{fig_fourier_kin}
\caption{Example of the parameterisation of the pitch angle kinematics using a finite Fouries series $p=6$, $12$, and $18$ parameters.
The last parameter $\kindex{\Delta t}{0}$ introduced the phase shift.}
\label{fig:fourier_kin}
\end{figure*}
In the second approach, we use a finite Fourier sine-cosine series to define our pitching kinematics:
\begin{equation}
\beta\left( t \right) = \sum\limits_{ k=1,3,5,...}^{N}\Big( \kindex{a}{k} \cos\left( 2\pi k\kindex{f}{0} t \right) + \kindex{b}{k} \sin\left(2\pi k\kindex{f}{0} t \right) \Big)
\label{eq.fourier}
\end{equation}
with $\kindex{a}{k}$ the cosine coefficients, $\kindex{b}{k}$ the sine coefficients, and $\kindex{f}{0}$ the flapping frequency.
For the current application, we use only the odd coefficients.
The first odd sine-coefficients $\kindex{b}{k}$ are responsible for the base oscillation whereas the odd cosine-coefficients $\kindex{a}{k}$ are used for smaller modulations.
The ability to represent higher curvatures, as they appear in trapezoidal motions, mainly depends on the number of Fourier terms included.
A finite Fourier series is not ideal to describe sudden jumps or discontinuities and tends to display an overshoot followed by decaying oscillations near sharp gradients.
This behaviour is known as the Gibbs phenomenon and can be reduced by adding a Lanczos-$\sigma$-factor \cite{hamming_numerical_1986}.
The Lanczos-$\sigma$-factor is defined as:
\begin{equation}
\sigma\left( \frac{ k }{ m } \right) = \frac{ \sin\left( \frac{ k }{ m } \pi\right) }{ \frac{ k}{ m }\pi }
\label{eq:lanczos-sigma-factor}
\end{equation}
with $m$ as the last plus one summing index of a finite Fourier series.
The Fourier series based pitching kinematics used in this paper are described by a finite Fourier series with Lanczos-$\sigma$ correction factor:
\begin{equation}
\begin{split}
\beta\left( t \right) = \sum\limits_{k=1,3,5,...}^{ (p-1)/2 } & \sigma \left( \frac{ 2k }{p+1} \right) \\ & \Big( \kindex{a}{k} \cos\left( 2\pi k\kindex{f}{0} t \right) + \kindex{b}{k} \sin\left(2\pi k\kindex{f}{0} t \right)\Big)
\end{split}
\label{eq:fourier_with_sigma}
\end{equation}
with $(p-1)$ parameters, $(p-1)/2$ cosines and $(p-1)/2$ sine coefficients.
The obtained motion kinematics are periodic, continuous in all derivatives and can easily be expressed in a closed form.
The half strokes can be asymmetric around the quarter stroke.
The phase-shift parameter \kindex{\Delta t}{0} is again applied at the end to create advanced or delayed wing rotations with respect to the stroke reversal.
Exemplary motions created using $6$, $12$, and $18$ parameters are presented in \cref{fig:fourier_kin}.
\subsubsection{Modal reconstruction}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_modal_kin}
\caption{Example of an individual random walk-based motion (top left) and the first $17$ eigenmodes of the proper orthogonal decomposition of a family of $2 \times 10^5$ random walk-based motions.
The first $5$ modes are presented with a green background, modes $6$ to $11$ with a red background, and modes $12$ to $17$ with a grey background.
Examples of new motions created as a linear combination of the first $5$, $11$, and $17$ modes are presented in the bottom left.
The half stroke motions are point-mirrored around $(T/2,0)$ to obtain a full stroke and the phase shift is applied to create the final motions in the bottom right.
}
\label{fig:modes}
\end{figure*}
In the third approach, we use a low-order eigenmode reconstruction.
The eigenmodes are obtained by a modal decomposition of a family of available arbitrary kinematics.
This parameterisation approach is particularly elegant if a family of kinematics is known.
For example, if a database of kinematics is available from direct observations of insects or fish motions, we can first reduce the dimensionality of the available kinematics using modal decomposition.
Measured and new kinematics can then be created by low-order reconstruction of the most dominant kinematic modes.
Here, we do not have a database of measured kinematics at our disposal.
To demonstrate the concept of eigenmode reconstruction as a parameterisation approach, we have artificially generated a family of randomised kinematics based on a biased random walk algorithm.
The random walk motion generator is not an essential part of the approach it merely serves as bypass to obtain a generalised motion database.
The random walk algorithm takes randomly sized steps in a discretised half-stroke.
The random expectation value distribution for the individual step sizes is defined such that their sum equals zero over one half-stroke.
Each step is defined as the expectation value plus a random fluctuation.
The motions starts at $\beta=\ang{0}$ and is set again to zero at the end of the half-stroke.
This creates a wider spread of values and higher gradients at the end of the half-stroke in comparison to the start of the half-stroke.
This bias is resolved by line-mirroring motions at mid-half-stroke ($T/4$).
An example of a random walk-based motion is presented in the top left panel of \cref{fig:modes}.
A large amount of motions ($\mathcal{O}(\num{e5})$) is created within a few minutes with this method.
The family of random walk-based motions are then decomposed using proper orthogonal modal decomposition (POD).
All modes start and end with a zero passage and are nearly symmetric (\cref{fig:modes}).
The symmetry is expected to improve with a larger data basis of random kinematics whereas the computational complexity of the POD is the limiting factor for the number of input motions.
The modes exhibit high similarity with Legendre polynomials starting from the second polynomial.
The POD ensures the optimality of the projection space for a given set of training motions and the optimal use of parameters.
A large variety of complex, asymmetric motions and oscillations can be defined by a linear combination of the first $(p-1)$ eigenmodes:
\begin{equation}
\beta(t)= \sum\limits_{n=1}^{p-1} \kindex{a}{n} \sqrt{2\kindex{\lambda}{n}} \,\kindex{\psi}{n}(t)
\end{equation}
with \kindex{\psi}{n} the POD eigenmodes, \kindex{\lambda}{n} the corresponding eigenvalues, and \kindex{a}{n} the parameterisation coefficients.
The first $17$ eigenmodes, \kindex{\psi}{1}-\kindex{\psi}{17}, are presented in \cref{fig:modes} for the proper orthogonal decomposition of the family of random walk based motions used in this paper.
The coefficients \kindex{a}{n} are considered normalised and have values confined between \num{-1} and \num{-1}.
The motions are point-mirrored around $(T/2,0)$ to obtain a full stroke and the phase shift is applied to create the final motion as depicted in the bottom row of \cref{fig:modes}.
Discontinuities in the gradient between the end and beginning of each half-stroke cannot be executed by the motors and are mitigated by a robust local regression smoothing, which is tuned to mainly affect the sections around stroke reversal.
Discontinuities in the velocity $\dot{\beta}$ appear for randomly created kinematics in the early populations and diminish when the convergence progresses.
\subsection{Genetic algorithm optimisation}
The pitching kinematics of our robotic flapping wing device have been optimised using the multi-objective genetic algorithm optimisation algorithm (\texttt{gamultiobj}) from the global optimisation toolbox of Matlab\textsuperscript{\tiny\textregistered} \cite{matlab_global_2020}.
A genetic algorithm is a meta-heuristic optimisation method, suited to find the global optimum in a non-linear solution space with a large number of degrees of freedom.
It mimics the process of natural selection known from evolution by recombination and mutation.
An initial population of $200$ individuals, with a uniform distribution of the parameters within their bounds is created to start the optimisation.
Each individual corresponds to one periodic flapping motion.
The fitness values for two optimisation objectives are directly measured during experiments for each motion.
The optimisation objectives are the maximum stroke average lift $\kindex{\overline{C}}{L}$ and the maximum stroke average efficiency $\overline{\eta}$.
The stroke average efficiency $\overline{\eta}$ is defined here as:
\begin{equation}
\overline{\eta} = \frac{\kindex{\overline{C}}{L}}{\kindex{\overline{C}}{P}}.
\end{equation}
The overline indicates stroke average quantities.
The lift and power coefficients,\kindex{C}{L} and \kindex{C}{P}, are defined as
\begin{equation}
\kindex{C}{L} = \frac{ L }{ \frac{ 1 }{ 2 } \rho R c \overline{U}^{2}}\quad, \quad
\kindex{C}{P} = \frac{ P }{ \frac{ 1 }{ 2 } \rho R c \overline{U}^{3}}\quad.
\end{equation}
The dimensional lift $L$ and power $P$ are normalised with the density $\rho$, the wing span $R$, the chord length $c$ and the stroke average velocity $\overline{U}$.
Each motion is executed over eight cycles.
The stroke average quantities are determined based on the last four cycles to exclude any influence of start-up transients.
At the selected Reynolds number, the load responses are extremely repeatable and the difference between the individual cycle averaged lift coefficients and the average coefficient over the last four cycles is below \SI{5}{\percent} after the first cycle and below \SI{2}{\percent} after three cycles.
Variations up to \SI{20}{\percent} in the lift coefficient are observed in the first cycle for the most efficient kinematics.
The evaluations of the individual experiments take $\SI{39}{\second}$ each, including a settling time for the water in the tank.
The fittest individuals produce offspring by recombination and mutation.
Recombined individuals account for \SI{60}{\percent} of the following population with the parameters calculated by a randomised linear combination $\textit{child} = \textit{parent 1} + R ( \textit{parent 2} - \textit{parent 1})$, with $R$ a randomly picked number from a uniform distribution covering the interval $[0,1]$ \cite{matlab_global_2020}.
The remaining \SI{40}{\percent} of the new population are created by random mutation of the individuals' parameters.
No clones from the elite are transferred into the next generation.
The fitness values of the new individuals are determined and the procedure is repeated until convergence is detected, after which the optimisation is stopped.
The final output of the genetic algorithm optimisation is a global Pareto front, which consists of all non-inferior solutions in the $\kindex{\overline{C}}{L}$ versus $\overline{\eta}$ space.
The solutions on the Pareto front are called non-inferior solutions if the value for each objective can only be improved by decreasing the value of another one.
The results of $9$ optimisations will be presented and compared here.
We have considered each of the parameterisation approaches with $p=6$, $12$, and $18$ parameters.
The experimental set-up is subjected to mechanical and safety constraints to protect the equipment.
The maximum pitching angle is limited by the range of free movement in the experimental set up such that $|\kindex{\beta}{max}| \leq \SI{90}{\degree}$.
The minimum angle of $\kindex{\beta}{min} = \ang{0}$ is imposed for the parameterisation of the initial half stroke before applying the phase-shift $\kindex{\Delta t}{0}$ to avoid ambiguous solutions where advance or delayed rotation is not governed entirely by the phase-shift.
The phase shift is bound by $[-\pi/2, \pi/2]$.
The minimum and maximum pitch angle bounds are easy to define for the control point approach by directly limiting the values of $\beta$ of the control points.
This does not apply for the Fourier series and modal reconstruction based parameterisations.
Here, the extreme angles are the result of a combination of different parameters.
We opted here to keep the allowed parameter ranges as wide as possible and to reject proposed kinematics that are not safe or not feasible for the experimental mechanism before execution.
The most important constraint for our robotic flapper is the pitch angle acceleration.
High accelerations put stress on the load cell due to the inertial forces on the wing and have to be limited to avoid damage.
The maximally allowed pitch angle accelerations are $|\kindex{\ddot{\beta}}{max}| \leq \SI{1800}{\degree\per\second\squared}$.
The majority of randomly created individuals for the control point and modal reconstruction approach exceed the acceleration constraint.
This leads to a validity rate of $ \leq \SI{10}{\percent}$ in combination with the constraint $|\kindex{\beta}{max}| \leq \SI{90}{\degree}$.
The parameter bounds are tightened for the Fourier series approach to keep the share of invalid motions within a range of \SIrange{10}{15}{\percent}.
A simple restriction of the parameter space does not have the same effect for the splines connecting control points and the modal reconstruction.
Here, a pre-selection of kinematics is necessary to obtain a full initial population of $200$ individuals.
The average and peak acceleration values of the randomly created motions increase with the parameter count and so does the share of randomly generated motions that cannot be executed.
The pre-selection process takes a couple minutes for the modal reconstruction with $6$ parameters and up to $\SI{4}{\hour}$ for the control points connected by splines with $12$ parameters on a standard desktop computer.
The control point approach with $18$ parameters has very tightly spaced control points, which lead to high local accelerations above the limit value but they only appear for a very short time such that their risk of damaging the set-up is deemed low.
The constraint of a minimum angle $\kindex{\beta}{min}\geq \SI{0}{\degree}$ in combination with the acceleration and velocity constraints is too strict for the modal reconstruction approach to create an initial population in a reasonable amount of time.
By loosening the lower angular constraint $\kindex{\beta}{min}$, and allowing $\kindex{\beta}{min}<0$, we obtain more kinematics that have a maximum acceleration below the mechanically allowed limit.
All parameterisation approaches required a customised implementation of the constraints and adapted selection of the parameter bounds.
The implementation of the experimental constraints was most straight forward for the control point based approach due to the more direct local access to the kinematics.
\subsection{Convergence criterion}
To monitor the progress of the optimisations and to decide when to consider the result to be converged, we have implemented a generational distance measure.
The generational distance allows for a quantitative comparison between the optimisations of the different kinematic functions possible.
It is a common measure to determine optimisation progress in genetic algorithms if no optimal Pareto front as reference is known \cite{audet_performance_2020}.
The generational distance (GD) is defined by \citet{veldhuizen_multiobjective_1999} as:
\begin{equation}\label{eq:gd}
\textrm{GD}\left( S,P \right)= \frac{ 1 }{ \left\lvert S \right\rvert } \Bigg( \sum\limits_{ s\in S }{ \,\min\limits_{r\in P}\parallel F \left( s \right)-F \left( r \right)\parallel^{ p }} \Bigg)^{ \frac{ 1 }{ p } },
\end{equation}
with $S$ the set of Pareto front members of the current generation, $P$ the Pareto set members of the previous generation, $F$ the objective function, and $p$ the selected distance norm.
The generational distance describes the average distance the Pareto front members are shifted between successive generations and represents the optimisation progress.
The individual distances are measured between the point of the current front and the closest point from the previous front.
All distance measures for the efficiency $\overline{\eta}$ and stroke over lift coefficient $\kindex{\overline{C}}{L}$ are normalised by the average range of values for the first ten generations (or all generations if less that ten generations have been tested).
The convergence speed is defined as the gradient of the generational distance.
An optimisation is considered converged if the convergence speed $\leq \num{5e-4}$ for two consecutive generations and if the generational distance $\leq \num{5e-3}$.
The changes of the Pareto front are assumed to be minor after the convergence criterion is fulfilled.
\section{Results and discussion}
Here, we will analyse and compare the performance of the three selected parameterisation approaches for optimisation applications by the example of a multi-objective optimisation of the pitching kinematics of a flapping wing.
The pitching kinematics are parametrised by
\begin{inparaenum}[a)]
\item control points connected by a spline interpolation,
\item a finite Fourier series, and
\item a modal reconstruction of POD eigenmodes.
\end{inparaenum}
Each of the approaches has been tested with $6$, $12$, and $18$ parameters.
This lead to more than $\num{30 000}$ experimental iterations that have been executed over a period of several weeks.
We will first compare how well the different approaches cover the motion space, discuss the optimisation process and convergence, and finally compare the Pareto front solutions and the optimised kinematics.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_initial_pop}
\caption{Probability density map of a randomly selected initial population consisting of $200$ individuals for the different parameterisations with $12$ parameters without applying a phase-shift.
The pitching angle is bound by $\kindex{\beta}{min} = \SI{0}{\degree}$ and $\kindex{\beta}{max} = \SI{90}{\degree}$.
First row shows the coverage of the solutions space for the different parameterisation methods: a) control points connected by a spline interpolation, b) finite Fourier series, c)
modal reconstruction, without taking into account an acceleration constraint.
The second row shows the coverages when the acceleration constraint ($\kindex{\ddot{\beta}}{max} \leq \SI{1800}{\degree \per \second \squared}$) is applied.
The lower angular bound was lifted for the modal reconstruction when applying the acceleration constraint to allow for a set of $200$ executable individuals to be found in a reasonable amount of time.}
\label{fig:initial_pop}
\end{figure*}
\begin{figure*}
\includegraphics[width=\linewidth]{fig_pop_prop}
\caption{Quantitative comparison of the characteristic properties of the initial population consisting of $200$ individuals for the different parametrised of the kinematics for $12$ parameters.
The brighter bars indicate values obtained without taking into account the acceleration constraint.
The darker bars include the influence of the acceleration constraint.
The diversity of the population is characterised by the maximum pitch rate $\kindex{\dot{\beta}}{max}$, maximum pitching acceleration $\kindex{\ddot{\beta}}{max}$,
the half-stroke-averaged standard deviation of the pitching angle $\overline{\sigma(\beta)}$, the half-stroke-averaged normalised standard deviation of the pitch rate $\overline{\sigma( {\dot{\beta}} )} / \overline{\sigma( {\beta} )} $ and the pitching acceleration $\overline{\sigma( {\ddot{\beta}} )} / \overline{\sigma( {\beta} )}$, and the correlation-coefficients for the pitch rate $ r ( \dot{ \beta } ) $
and pitching acceleration $r(\ddot{\beta})$.
}
\label{fig:pop_prop}
\end{figure*}
\subsection{Coverage of the kinematic solution space}
Suitable parameterisation approaches for optimisation studies should be able to represent a high variety of qualitative shapes and achieve good initial coverage of the solution space to not a priori exclude potential optimal solutions.
To quantify and compare the ability of the different approaches to cover the solution space, we randomly created an initial population consisting of $200$ individuals for the different parameterisations with $12$ parameters without any acceleration constraint and without applying the phase-shift.
The probability density maps for these initial populations of pitching kinematics are presented in the top row of \cref{fig:initial_pop} to give a visual impression of the coverage of the solution space.
The density maps for $p=18$ do not show significant differences and lead to the same conclusions as for $p=12$.
The splines connecting control points and the modal reconstruction functions cover almost all achievable angles in the first half-stroke (\cref{fig:initial_pop}a1,c1).
The Fourier-series have a slightly narrower band of admissible motions around stroke reversal with lower diversity (\cref{fig:initial_pop}b1).
The fixed phase-locations of the control points lead to the occurrence of regions in the density map with higher local probabilities indicating an inhomogeneous coverage and slight bias towards low angles at the control points.
A quantitative comparison is presented in \cref{fig:pop_prop}, where the brighter bars indicate the values without the application of the acceleration constraint.
The parameters we calculated to characterise the coverage of the solutions space and diversity of the parameterised motions kinematics include the maximum pitch rate $\kindex{\dot{\beta}}{max}$ and pitch acceleration $\kindex{\ddot{\beta}}{max}$, the half-stroke-averaged standard deviation of the pitching angle $\overline{\sigma(\beta)}$, the half-stroke-averaged normalised standard deviation of the pitch rate $\overline{\sigma( {\dot{\beta}} )} / \overline{\sigma( {\beta} )} $ and the pitch ac\-celera\-tion $\overline{\sigma( {\ddot{\beta}} )} / \overline{\sigma( {\beta} )} $.
The Fourier series-based kinematics have the lowest maximum pitch rate and pitch accelerations and a lower diversity based as indicated by lower values of the standard deviations of the pitch angle, rate, and acceleration compared to the control point approach and the modal reconstruction.
The highest gradients are measured among the modal reconstruction solutions if the acceleration constraint is not applied.
The control point kinematics exhibit the highest diversity closely followed by the modal reconstruction method if we do not apply the acceleration constraint.
If we apply the acceleration constraint to omit solutions that cannot or should not be executed by our experimental device, we obtain the probability density plots presented in the bottom row of \cref{fig:initial_pop}.
The quantitative measures after application of the acceleration constraint are depicted in \cref{fig:pop_prop} by the darker bars.
Here, the acceleration is limited to $\ddot{\beta} \leq \SI{1800}{\degree \per \second \squared}$.
The coverage achieved by the initial populations is now reduced for all approaches especially near the stroke reversal.
Motions with larger angles around stroke reversal are omitted due to the acceleration limit.
This affects the control point approach the most.
The splines connecting the control points tend to develop characteristic spikes when the distance between control points decreases which leads to accelerations above the allowed limit.
The initial population for Fourier series kinematics exhibits a small band of angles with very low diversity around stroke reversal and a poor coverage for smaller $\beta$.
This is caused by the different restriction of the parameter space for the Fourier series.
The best coverage is achieved for all kinematic functions at moderate angles between $t/T = 0.125-0.375$.
The modal reconstruction has the highest density at lower angles compared to the others (\cref{fig:initial_pop}c1).
Unfortunately, this increases the chance that a randomly selected individual does not meet the acceleration constraint and needs to be omitted and replaced when building he initial population.
To obtain a set of \num{200} executable individuals for the initial population within a reasonable amount of time, the lower angular constraint $\kindex{\beta}{min}$ is loosened.
This explains the non-zero probability for $\kindex{\beta}{min}<0$ in \cref{fig:initial_pop}c2.
All kinematic functions show a significant loss in diversity following the application of the acceleration constraint.
The modal reconstruction now outperforms the control point approach and the Fourier series in all measures.
Most of the diversity of the initial control point approach without constraint is created over the tightly spaced control points with a large spectrum of admissible angles.
This allows for a large variety of angles, velocities, and accelerations, but it also leads to higher number of omitted kinematics.
This could potentially be reduced by distributing the control points differently or by also considering their time coordinates as variable parameters.
Alternative measures for the diversity of the randomly selected individuals in an initial population are obtained by determining the correlation-coefficients for the pitch rate $r(\dot{\beta})$ and the pitch acceleration $r(\ddot{\beta})$.
They are calculated as the ensemble average correlation-coefficient $r$ for each individual with the rest of the population.
A higher diversity within a population now leads to a lower correlation coefficient $r$.
Overall, the diversity decreases upon application of the acceleration constraint as the correlation values increase.
The control point approach and the modal reconstruction show again a higher diversity than the population of Fourier series.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_similarity}
\caption{Cumulative distribution of the relative node degree $\kindex{n}{d}/N$ in the similarly networks describing the initial population of $N=200$ individuals for the three parameterisation approaches.
Distributions corresponding to the networks based on $\kindex{R}{i,i}$ are in the top row, those based on $\kindex{D}{i,i}$ are in the bottom row.
The light and dark shading correspond respectively to randomly selected populations without and with the application of the acceleration constraint.
}
\label{fig:similarity}
\end{figure*}
To more quantitatively compare the similarity between different randomly generated kinematic, we borrow ideas here from a network topology approach proposed by \citet{zhang_complex_2006} and more recently used e.g. by \citet{tandon_condensation_2021} for comparing the similarity between trajectories.
The similarity between kinematics can be quantified using the maximum cross-correlation coefficient $\kindex{R}{i,j}$ for each pair of kinematics $\kindex{\beta}{i}$ and $\kindex{\beta}{j}$.
The maximum cross-correlation coefficient is defined as:
\begin{equation}
\kindex{R}{i,j} = \max\limits_{\tau=0...T} \frac{cov\left(\kindex{\beta}{i}(0:T),\kindex{\beta}{j}(\tau:(T+\tau))\right)}{\sqrt{var\left(\kindex{\beta}{i}(0:T)\right) var\left(\kindex{\beta}{j}(\tau:(T+\tau)\right)}}
\end{equation}
with $T$ the period, $cov(.)$ the covariance, and $var(.)$ the variance operators.
By taking the maximum over all possible phase shifts $\tau$, the resulting value only evaluates the shape similarity and is not influence by a phase shift, e.g. introduced by $\kindex{\Delta t}{0}$ here.
Another metric proposed by \citet{zhang_complex_2006} is the mean absolute error $\kindex{D}{i,j}$ between two curves, which is defined here as:
\begin{equation}
\kindex{D}{i,j} = \min\limits_{\tau=0...T} \displaystyle\frac{1}{T}\sum\limits_{t=0}^{T} \left\lvert\kindex{\beta}{i}(t)-\kindex{\beta}{j}(t+\tau)\right\rvert \quad.
\end{equation}
We normalised \kindex{D}{i,j} by the period $T$ and expressed the result in degree such that $\kindex{D}{i,j}$ can be interpreted as an average pitch angle amplitude difference between two kinematics.
Two identical curves have a maximum cross-correlation coefficient $\kindex{R}{i,j}=1$ and a minimal mean absolute error $\kindex{D}{i,j}=0$.
Both quantities evaluate different aspects, $\kindex{R}{i,j}$ provides a measure for the shape similarity between curves and $\kindex{D}{i,j}$ provides a measure for the amplitude variations between curves.
We use both here to evaluate the degree of similarity within a randomly generated population.
After calculating $\kindex{R}{i,j}$ and $\kindex{D}{i,j}$ between all curves with the initial population, we construct two similarity networks.
Each kinematic motion is represented by a node in the network.
Different nodes are connected to each other if the shape similarity measure $\kindex{R}{i,j}$ is higher than a predefined threshold \kindex{R}{th} or if the amplitude dissimilarity measure $\kindex{D}{i,j}$ is lower than a predefined threshold \kindex{D}{th}.
We then determine the node degree for each node, which corresponds to the number of nodes it is connected to.
Kinematics with a low node degree show similarity only to a low number of other kinematics in the population.
The cumulative distributions of the node degrees (\kindex{n}{d}) among the initial population for the three different parameterisation approaches with $p=12$ are presented in \cref{fig:similarity}.
The light and dark shading correspond respectively to randomly selected populations without and with the application of the acceleration constraint.
The more diverse a population of kinematics the more kinematics we expect with a low degree of connectivity and the larger the coloured area in \cref{fig:similarity}.
The distributions in the top row have been obtained for the networks based on the shape similarity measure $\kindex{R}{i,j}$ and indicate how diverse the population is in terms of the shape of the kinematics.
The distributions in the bottom row have been obtained for the networks based on the amplitude difference measure $\kindex{D}{i,j}$ and indicate how diverse the population is in terms of the amplitudes of the kinematics.
The unconstraint populations for the control point and the modal reconstruction have a high degree of diversity as the majority of the kinematics have a low node degree and show similarities to only a few other kinematics in the population.
The Fourier series performs slightly lower than the two others.
By construction, the Fourier series have an inherent similarity in shape and it is not entirely unexpected to find more nodes with a higher node degree for this approach.
Almost all cumulative distributions in \cref{fig:similarity} shift to higher node degrees when the acceleration constraint is applied, indicating that the constraint decreases the diversity both in terms of shapes and in terms of amplitudes.
The only exception is the shape diversity for the modal reconstruction which remains nearly unaffected by the constraint.
The spline population is the most diverse for the unconstrained case and the modal population exhibits the highest diversity for the constrained case.
We have also used these metrics to analyse the ability of the different approaches to create the insect-like kinematics and trapezoidal motions presented in \cref{fig:experimental_setup}.
All approaches except the modal reconstruction are capable to mimic the presented kinematics with a minimum correlation coefficient $R\geq 0.993$ and a maximum mean absolute error $D\leq\ang{1}$.
The modal reconstruction does equally well for most of the motions except for the trapezoid with steep edges which is approximated with a larger mean absolute error of $D=\ang{2.3}$.
Based on all visual and quantitative comparison presented here, the control point method yields the best coverage of the solution space and the largest diversity within an initial population if no acceleration constraint is applied.
The modal reconstruction method performs better than the control point method when the constraint is applied.
The Fourier series populations are least diverse and have the lowest coverage, but still perform sufficiently well for our optimisation application.
\subsection{Optimisation progress and convergence}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_pop_size}
\caption{Number of executable motions or individuals per generation for all optimisations with $6$, $12$, and $18$ parameters and the three parameterisation approaches.}
\label{fig:individuals_per_generation}
\end{figure}
Due the specific implementation of the genetic algorithm for our flapping wing optimisation, we can not replace individual kinematics that are omitted based on the acceleration constraint.
This is not a general limitation of the approach and can be overcome for future applications.
However, in the current study, the total number of the executable individuals in first generations is significantly lower than the nominal $200$ individuals due to the implementation (\cref{fig:individuals_per_generation}).
This is in particular the case for the control point and modal reconstruction methods with $12$ or $18$ parameters.
The Fourier series method is less affected and also the optimisation with a lower number of control points has populations with consistently close to \SI{85}{\percent} of executable individuals.
The number of executable individuals per generation increases for most of the optimisations but never reaches full population size with exception of the control point approach with $6$ parameters.
The control point approach with $12$ parameters and the modal reconstruction approach have particularly low numbers of executable individuals in the early generations, but the situation improves in the course of the optimisation.
The lower number of executable individuals in the populations decrease the diversity which hampers the search for the optimum solutions and delays convergence.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_gen_distance}
\caption{Moving average over $10$ iterations for the generational distance (GD) versus the number of iterations.
Each iteration corresponds to the evaluation of one flapping motion.}
\label{fig:convergence}
\end{figure}
The convergence of the different approaches is analysed based on the generational distance introduced in \cref{eq:gd}.
The variation of the generational distance with the number of executed motions is presented in \cref{fig:convergence}.
Surprisingly, the parameter count is not the most important factor that influences the convergence behaviour.
The largest influence on the convergence is the parameterisation approach.
The optimisations using the same parameterisation form groups with similar convergence speed according to the generational distance metric.
The modal reconstructions exhibit the highest relative generational distance and convergence speed at the very beginning.
The modal reconstruction with $18$ parameters has the highest initial generational distance of \num{0.346} but also the modal approach with $12$ (GD=\num{0.129}) and $6$ parameters (GD=\num{0.125}) have initial values that are one magnitude larger than the values obtained for the other conditions.
The higher initial generational distance can be explained by a larger diversity of the initial population.
The convergence speed behaves proportionally to the generational distance such that all optimisations converge in comparable time.
The smallest relative generational distance and convergence speed is observed for the Fourier series.
The motions of the initial Fourier population are less diverse and closer to the final solution due to the different constriction of the parameter bounds.
The optimisation progress is slower because the differences between the motions are less pronounced and consequently harder to determine.
On the other side, higher diversity for the initial populations leads to higher convergence speed.
The total convergence time differs by \SI{40}{\percent} between the slowest (control point approach with $p=12$) and fastest (Fourier series approach with $p=6$) optimisation and the majority takes between $\num{2486}$ (Fourier series approach with $p=12$) and $\num{2742}$ (control point approach with $p=12$) iterations (\cref{tab:convergence}).
The optimisations are not deterministic and uncertainties are introduced due to random initialisation, recombination and mutation.
The differences in convergence time are small when considering that the computational complexity of the optimisations increases exponentially with the parameter count.
The Fourier series optimisations tend to reach convergence slightly faster than the modal and control point approaches.
\begin{table}
\centering
\caption{Number of iterations and generations required until convergence for the different parameterisation approaches.
Convergence is reached if the convergence speed $\leq \num{5e-4}$ and generational distance $ \leq \num{5e-3}$ for two consecutive generations.}
\begin{tabular}{@{}llll@{}}
\toprule
parameterisation & $p$ & iterations & generations\\
\toprule
control points & 6 & $2718$ & $14$\\
& 12 & $2742^\ast$ & $25^\ast$ \\
& 18 & $2968$ & $25$ \\
Fourier series & 6 & $2118$ & $11$\\
& 12 & $2486$ & $13$\\
& 18 & $2491$ & $13$\\
modal reconstruction & 6 & $2683$ & $18$ \\
& 12 & $2721$ & $23$ \\
& 18 & $2702$ & $29$\\
\toprule
\multicolumn{4}{l}{\scriptsize
\parbox{0.85\linewidth}{
$^{*}$ The convergence for this case is determined manually by observing the evolution of the Pareto kinematics.}}
\end{tabular}
\label{tab:convergence}
\end{table}{}
\subsection{Pareto fronts and optimal kinematics}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_pareto}
\caption{Global Pareto front for all kinematic function optimisations for the last generation.
Kinematics corresponding to the final global Pareto front solutions for the different kinematic functions.}
\label{fig:pareto_front}
\end{figure*}
The resulting Pareto fronts for all $9$ optimisations are presented in \cref{fig:pareto_front}.
The colours indicate the number of optimisation parameters and the symbols indicate the parameterisation approach.
The different Pareto fronts have the same shape but there is a shift between them and they reach slightly different extreme values.
The modal reconstruction approach with $p=6$ finds the most efficient solutions (up to $\eta\approx1.4$) and the Fourier series approach with $p=12$ finds the solutions that generate the highest stroke average lift (up to $\kindex{\overline{C}}{L}\approx 2.2$).
Overall, the optimisations with the highest number of parameters are most limited in finding highly efficient solutions.
The higher number of parameters might lead to more complex kinematics that require more power to achieve the same performance in lift.
For our current example, $p=12$ seems to provide the best balance across the different parameterisation approaches.
The additional complexity provided by $12$ optimisation parameters compared to the $p=6$ optimisations is beneficial in improving the aerodynamic performance without increasing the power requirements and penalising the efficiency.
Increasing the parameter count leads to slower convergence and kinematics with higher harmonics variations.
To aid the genetic algorithm to fine tune the parameters, we have conducted an additional optimisation where we have taken the results for $p=6$ as a start point for a $p=12$ optimisation using the Fourier series parameterisation approach.
The results did not lead to significant differences with respect to the regular $p=12$ optimisation and are not presented here.
Overall, the differences between Pareto fronts are of the same order of magnitude as the inherent experimental variations.
The influence of the number of parameters seems to have a stronger influence than the parameterisation approach.
In general, the performance of the three parameterisations is comparable and we can not identify a clear preference of any one of them.
The optimal kinematics that make up the Pareto front are included in \cref{fig:pareto_front}.
The columns with the same background colour correspond to optimisations with the same number of parameters.
The rows correspond to different parameterisation approaches.
The colour of the $\beta(t/T)$ curves for one half stroke varies from grey to green, red, or black with increasing efficiency.
All Pareto kinematics show a continuous transition from high lift motions to more efficient ones, which indicates that the solutions are converged and we obtained a well developed Pareto front.
For the same number of parameters, the Pareto kinematics show similar characteristic features for the three parameterisation approaches but also display some subtle differences that cause the shifts in the Pareto front.
A clear common feature is the preference for advanced rotation for most Pareto kinematics.
Advanced rotation means that the wing has rotated past its horizontal orientation ($\beta=0$) before the end of the stroke motion.
By doing so, it starts the next stroke with a lower and more favourable angle of attack.
A positive influence of the advancement of the rotation for lift production has been observed before on insects and their robotics counterparts \cite{Sane.2001, Sun.2002, Krishna:2019cb}.
The rotational advancement generates a higher stroke average lift but comes at a cost and there is a continuous decrease of the phase shift towards less advanced rotation with increasing efficiency in all our optimisations.
The pitch rate at the end of the stroke is strikingly similar for all cases.
The variation in phase shift between the highest lift generating and most efficient kinematics tends to decrease for optimisations with fewer parameters.
This suggests that lift increase can be generated at lower cost by a more complex pitching kinematics during the entire stroke instead of relying of the advanced rotation (\cref{fig:pareto_front}).
Another common feature is the increase in the maximum pitch angle with increasing efficiency.
A larger pitch angle corresponds to a lower angle of attack with leads to lower drag and more efficient kinematics.
The values of the maximum pitch angles are approximately the same for all parameterisations and parameter counts except for the Fourier series, which reach slightly lower maximum values of $\beta$.
The main differences between the Pareto optimal kinematics of the different cases are the number of local maxima and their timing.
The control point and modal reconstruction approaches with $p=6$ show two local maxima around $t/T = 0.125$ and $t/T = 0.375$ (\cref{fig:pareto_front}).
These two maxima are less pronounced or even absent in the Fourier series kinematics which look more like a smooth trapezoidal profile.
The two local maxima become more pronounced for the control point approach with $p=12$ and only the second one around $t/T = 0.375$ remains present when $p$ is increased to $18$ for this approach.
The modal reconstructions for $p=18$ are also characterised by a single peak that is located around $t/T = 0.32$.
The Fourier series with $p=12$ and $p=18$ have a distinct bump at the beginning of the stroke, around $t/T=0.1$.
Overall, the complexity of the motions increases with the parameter count and different shapes and kinematics can lead to similar Pareto fronts depending on the choice of the parameterisation and the parameter count.
This is due to the complex relationship between the flow development and the growth of the leading edge vortex for flapping wings motions \cite{Bhat.2020,gehrke_phenomenology_2021}.
In previous work, we identified the shear layer velocity $\kindex{u}{s}$ at the leading edge as a characteristic parameter to scale the force response of generalised flapping wing pitching kinematics \citep{gehrke_phenomenology_2021}.
The shear layer velocity act as a proxy for the feeding rate of the leading edge vortex.
Higher angles of attack or lower pitching angles lead to a higher shear layer velocity $\kindex{u}{s}$ which in turn leads to a faster growth of the leading edge vortex.
This is the common feature of high lift generating kinematics.
The stronger leading edge vortex leads to higher lift, but reaches its maximum size and circulation earlier in the cycle.
The maximum in lift and leading edge circulation is typically reached around mid-stroke for high lift kinematics.
After the vortex reaches its maximum circulation, it lifts off of the wing which leads to a drop in lift in second half of the stroke.
Efficient motions exhibit smaller angles of attack with pitching angle maxima around $\kindex{\beta}{max}=\SI{70}{\degree}$ (\cref{fig:pareto_front}).
They create a smaller leading edge vortex which grows continuously during the majority of the stroke and promote drag-minimised lift creation.
The power and lift coefficients are less dominated by the stroke motion as the angle of attack is smaller causing the shear layer velocity \kindex{u}{s} to be relatively constant during mid-stroke.
\section{Conclusion}
In this study, we presented three different approaches to parameterise motion kinematics, demonstrated their application, and evaluated their performance by the example of an experimental optimisation of the pitching kinematics of a robotic flapping wing in hover.
The pitch angle kinematics in our application are described by
\begin{inparaenum}[a)]
\item control points that are connected by a fifth-order spline interpolation,
\item a finite Fourier series, and
\item a linear combination of modes determined by a modal decomposition of kinematics created by a random walk.
\end{inparaenum}
Each of the parameterisation approaches is implemented with three different parameter counts: $6$, $12$ and $18$, which leads to a total of nine optimisations and more than $\num{30 000}$ experimental iterations conducted over a period of several weeks.
The performance of the different approaches has been analysed by comparing the diversity of the solutions and coverage of the motion space, the optimisation process and convergence, and the Pareto front solutions and the optimised kinematics.
The coverage of the solution space was qualitatively evaluated based on a probability density map of the random initial population of \num{200} individuals.
For a more quantitative comparison of the coverage and the diversity of the solutions, we introduced a number of diversity measures including the maximum pitch angle, rate, and acceleration, their standard deviations, and correlation coefficients for the pitch and acceleration.
To further quantify the diversity between the kinematics within the initial populations, we calculated similarity networks based on a mutual shape similarity measure and an amplitude error.
The cumulative distribution of the node degree in a population is an intuitive quantitative measure to compare the diversity of the kinematics based on their shape and amplitude.
Based on all qualitative and quantitative measures, the control point kinematics exhibit the highest diversity closely followed by the modal reconstruction method if no experimental constraints are in place.
The Fourier series-based kinematics have the lowest maximum pitch rate, pitch acceleration, and a lower diversity compared to the control point and the modal reconstruction approach.
In most experimental applications, the theoretical parameter space cannot be fully explored due to mechanical, electrical, or other constraints for example related to the measurement equipment.
In our case, the main limitation is imposed by the sensitivity and measurement range of the load cell and the performance envelope of the motor that controls the pitching motion.
These limitations lead to a pitch acceleration constraint that significantly reduces the coverage and the solution diversity for all three parameterisation approaches.
The practical implementation of an acceleration constraint is easier for the Fourier series approach than for the two other approaches.
The modal reconstruction parameterisation performs best in the various diversity measures, followed by the control point and the Fourier series approaches when the acceleration constraint is in place.
The generational distance criterion is implemented to monitor the progress of the optimisation and as a convergence measure to end the optimisation.
All optimisations converged in comparable time which did not depend on the parameter count.
In the presented optimisation study, it is desirable to have fewer parameters as more parameters increase complexity without a clear aerodynamic benefit.
The subtle changes in the kinematics lead to variations in the loads that are of the same order as the inherent experimental fluctuation of the system.
This can slow down the convergence of the genetic algorithm especially in later stages when only minor aerodynamic improvements are achieved.
The resulting Pareto fronts for the different optimisations are similar in shape and in the range of values of the two objective functions they cover.
The parameter count has a stronger influence on the final results than the parameterisation approach.
This is encouraging for the robustness of the three parameterisation approaches presented here.
The optimal kinematics corresponding to the Pareto fronts have slightly different shapes depending on the parameter count and the parameterisation approach selected but they yield comparable fitness values.
The most prominent common features of the kinematics are their shared preference for advanced rotation and the increase in the maximum pitch angle with increasing efficiency.
The main differences between the optimal kinematics of the different cases are the number of local maxima and their timing.
\begin{table*}[tb!]
\centering
\caption{Overview of the key characteristics for the different parameterisations.}
\begin{tabular}{l@{}cccc@{}}
\toprule
& Complexity & Handling & Interpretability & Applications \\
& & of constraints & of parameters & \\
\toprule
control points & - & + & + & intermittent kinematics, e.g. \\
& & & & burst and coast swimming \\
Fourier series & + & $\circ$ & - & vertical axis wind turbines, \\
& & & & flapping foil energy harvester \\
modal reconstruction & $\circ$ & - & $\circ$ & complex animal-like \\
& & & & locomotion \\
\toprule
\end{tabular}
\label{tab:conclusion}
\end{table*}
The goal of this paper was to present three different methods to parameterise complex motion kinematics for optimisation studies and to compare their performance on an experimental flapping wing system.
Each optimisation application has different objective functions, constraints, and specific challenges.
There is no one solution that fits all and the different approaches to parameterise the solutions have their specific advantages and disadvantages outlined in~\cref{tab:conclusion}.
The Fourier series approach is the easiest method to implement but the introduction of parameter constraints and the parameter bounds cannot be as directly contained as for the control point approach.
Nevertheless, the Fourier series has an analytical definition and continuous high-order derivatives which make it a promising parameterisation for structurally sensitive applications like vertical-axis wind turbines or flapping-foil energy harvesters.
By definition, the control points are more intuitive to constrain and their parameter values can be interpreted directly in terms of their influence on the temporal evolution and amplitude of the kinematics.
With the ability to create sharp and zero gradient curves, the control point method is most suitable for intermittent kinematics like burst and coast swimming of fish- and squid-like devices, and flapping wing flight.
The modal reconstruction is extremely versatile and adaptable.
The initial family of solutions can be generated by a random walk method as presented here, but can also be a family of kinematics measured in nature or in the lab.
Building a parameterisation upon an existing library of kinematics can be an extremely powerful tool.
Complex animal-like locomotions can be described by only a few modes which would require many parameters being constructed by analytical kinematic definitions.
Depending on the individual complexity of the mode shapes, it is almost impossible to restrict parameters bounds to meet experimental constraints.
Instead, motions have to be omitted after generation if they do not meet constrains such as start and end position and acceleration.
This can severely hamper the creation of initial populations for a genetic algorithm optimisation, for example.
In the presented application of an experimental optimisation of a robotic flapping wing mechanism, all three approaches yielded similar results in a comparable amount of time.
Other applications might have different challenges, but we expect that the main characteristics of the presented approaches are valid for different systems and that the choice of the most suitable parameterisation approach can be guided by the properties listed in \cref{tab:conclusion}.
\section*{Acknowledgments}
\addcontentsline{toc}{section}{Acknowledgments} %
The authors thank Guillaume de Guyon for the help in implementing the fifth-degree spline kinematic function.
Funding has been provided by the Swiss National Science Foundation under grant number 200021\_175792.
|
\section{Introduction}
Deep Neural Networks (DNNs) have revolutionized the fields of machine learning, computer vision, and natural language processing. As their name suggests, a key characteristic of DNNs is that they are deep. That is, they have a large depth, which can be defined as the length of the longest path from an input neuron to an output neuron. Often a neural network can be described as a linear sequence of layers, i.e.\@ groups of neurons with no intra-group connections. In such cases, the depth of a network is its number of layers.
It has been generally accepted that large depth is an essential component for high-performing networks because depth increases the representational ability of a network and helps learn increasingly abstract features~\citep{he2016deep}. In fact, one of the primary reasons given for the success of ResNets is that they allow training very deep networks with as many as 1000 layers~\citep{he2016deep}. As such, state-of-the-art performance is increasingly achieved by training models with large depth, and what qualifies as ``deep'' has shifted from ``2 or more layers'' in the early days of deep learning to the ``tens or hundreds of layers'' routinely used in today's models. For example, as shown in Figure~\ref{fig:acc_depth}, competitive benchmarks such as ImageNet are dominated by very deep models~\citep{he2016deep,huang2017densely,tan2019efficientnet} with at least 30 layers, whereas models with fewer than 30 layers perform substantially worse. The best-performing model with fewer than 20 layers has a top-1 accuracy of only $75.2$, substantially lower than accuracies achievable with 30 or more layers when evaluated with a single image crop~\citep{he2015delving,tan2019efficientnet}.
But is large depth always necessary? This question is worth asking because large depth is not without drawbacks. A deeper network leads to more sequential processing and higher latency; it is harder to parallelize and less suitable for applications that require fast responses.
In this paper, we study whether it is possible to achieve high performance with ``non-deep'' neural networks, especially networks with $\sim\!\! 10$ layers. We find that, contrary to conventional wisdom, this is indeed possible. We present a network design that is non-deep and performs competitively against its deep counterparts. We refer to our architecture as \method{} (\underline{Par}allel \underline{Net}works). We show, for the first time, that \emph{a classification network with a depth of just 12 can achieve accuracy greater than 80\% on ImageNet, 96\% on CIFAR10, and 81\% on CIFAR100}. We also show that \emph{a detection network with a low-depth (12) backbone can achieve an AP of $48\%$ on MS-COCO}. Note that the number of parameters in \method{} is comparable to state-of-the-art models, as illustrated in Figure~\ref{fig:acc_depth}.
A key design choice in \method{} is the use of parallel subnetworks. Instead of arranging layers sequentially, we arrange layers in parallel subnetworks. This design is ``embarrassingly parallel'',
\begin{wrapfigure}[32]{r}{0.6\textwidth}
\begin{center}
\includegraphics[width=0.58\textwidth]{figures/accuracies_depth.pdf}
\end{center}
\caption{Top-1 accuracy on ImageNet vs.\ depth (in log scale) of various models. \method{} performs competitively to deep state-of-the-art neural networks while having much lower depth. Performance of prior models is as reported in the literature. Size of the circle is proportional to the number of parameters. Models are evaluated using a single 224$\times$224 crop, except for ViTB-16 and ViTB-32~\citep{dosovitskiy2020image}, which fine-tunes at 384$\times$384 and PReLU-net~\citep{he2015delving}, which evaluates at 256$\times$256. Models are trained for 90 to 120 epochs, except for parameter-efficient models such as MNASNet~\citep{tan2019mnasnet}, MobileNet~\citep{sandler2018mobilenetv2}, and EfficientNet~\citep{tan2019efficientnet}, which are trained for more than 300 epochs. For fairness, we exclude results with longer training, higher resolution, or multi-crop testing.}
\label{fig:acc_depth}
\end{wrapfigure}
in the sense that there are no connections between the subnetworks except at the beginning and the end. This allows us to reduce the depth of the network while maintaining high accuracy. It is worth noting that our parallel structures are distinct from ``widening'' a network by increasing the number of neurons in a layer.
\method{} not only helps us answer a scientific question about the necessity of large depth, but also offers practical advantages. Due to the parallel substructures, \method{} can be efficiently parallelized across multiple processors. We find that \method{} can be effectively parallelized and \emph{outperforms ResNets in terms of both speed and accuracy}. Note that this is achieved despite the extra latency introduced by the communication between processing units. This shows that in the future, with possibly specialized hardware to further mitigate communication latency, \method{}-like architectures could be used for creating extremely fast recognition systems.
We also study the scaling rules for \method{}. Specifically, we show that \method{} can be effectively scaled by increasing width, resolution, and number of branches, all while keeping depth constant. We observe that the performance of \method{} does not saturate and increases as we increase computational throughput. This suggests that by increasing compute further, one can achieve even higher performance while maintaining small depth ($\sim\!\! 10$) and low latency.
To summarize, our contributions are three-fold:
\begin{itemize}\setlength\itemsep{0em}
\item We show, for the first time, that a neural network with a depth of only 12 can achieve high performance on very competitive benchmarks (80.7\% on ImageNet, 96\% on CIFAR10, 81\% on CIFAR100).
\item We show how parallel structures in \method{} can be utilized for fast, low-latency inference.
\item We study the scaling rules for \method{} and demonstrate effective scaling with constant low depth.
\end{itemize}
\section{Related Work}
\smallsec{Analyzing importance of depth}
There exists a rich literature analyzing the importance of depth in neural networks. The classic work of Cybenko et al.\ showed that a single-layer neural network with sigmoid activations can approximate any function with arbitrarily small error~\citep{cybenko1989approximation}. However, one needs to use a network with sufficiently large width, which can drastically increase the parameter count. Subsequent works have shown that, to approximate a function, a deep network with non-linearity needs exponentially fewer parameters than its shallow counterpart~\citep{liang2016deep}. This is often cited as one of the major advantages of large depth.
Several works have also empirically analyzed the importance of depth and came to the conclusion that under a fixed parameter budget, deeper networks perform better than their shallow counterparts~\citep{eigen2013understanding,urban2016deep}. However, in such analysis, prior works have only studied shallow networks with a linear, sequential structure, and it is unclear whether the conclusion still holds with alternative designs. In this work, we show that, contrary to conventional wisdom, a shallow network can perform surprisingly well, and the key is to have parallel substructures.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.65\textwidth}
\centering
\resizebox{\columnwidth}{!}{\includegraphics[trim=0.0cm 1.5cm 0.0cm 0.5cm, scale=0.4]{figures/SimpleNet_height.pdf}}
\vspace{1mm}
\caption{\method{}}
\label{fig:network}
\end{subfigure}%
~
\begin{subfigure}[t]{0.35\columnwidth}
\centering
\resizebox{\columnwidth}{!}{\includegraphics[trim=0.0cm 1.2cm 0.0cm 1.2cm, scale=0.5]{figures/RepVGG_blocks.pdf}}
\vspace{1mm}
\caption{The \method{} block}
\label{fig:block}
\end{subfigure}
\caption{Schematic representation of \method{} and the \method{} block. \method{} has depth 12 and is composed of parallel substructures. The width of each block in (a) is proportional to the number of output channels in \method{-M} and the height reflects output resolution. The \method{} block consists of three parallel branches: 1$\times$1 convolution, 3$\times$3 convolution and Skip-Squeeze-and-Excitation (SSE). Once the training is done, the 1$\times$1 and 3$\times$3 convolutions can be fused together for faster inference. The SSE branch increases receptive field while not affecting depth.}
\end{figure*}
\smallsec{Scaling DNNs}
There have been many exciting works that have studied the problem of scaling neural networks.
\citet{tan2019efficientnet} showed that increasing depth, width, and resolution leads to effective scaling of convolutional networks. We also study scaling rules but focus on the low-depth regime. We find that one can increase the number of branches, width, and resolution to effectively scale \method{} while keeping depth constant and low. \citet{zagoruyko2016wide} showed that shallower networks with a large width can achieve similar performance to deeper ResNets. We also scale our networks by increasing their width. However, we consider networks that are much shallower -- a depth of just 12 compared to 50 considered for ImageNet by~\citet{zagoruyko2016wide}~-- and introduce parallel substructures.
\smallsec{Shallow networks}
Shallow networks have attracted attention in theoretical machine learning. With infinite width, a single-layer neural network behaves like a Gaussian Process, and one can understand the training procedure in terms of kernel methods~\citep{jacot2018neural}. However, such models do not perform competitively when compared to state-of-the-art networks~\citep{li2019enhanced}. We provide empirical proof that non-deep networks can be competitive with their deep counterparts.
\smallsec{Multi-stream networks}
Multi-stream neural networks have been used in a variety of computer vision tasks such as segmentation~\citep{chen2016attention,chen2017deeplab}, detection~\citep{lin2017feature}, and video classification~\citep{wu2016multi}. The HRNet architecture maintains multi-resolution streams throughout the forward pass~\citep{WangSCJDZLMTWLX19}; these streams are fused together at regular intervals to exchange information. We also use streams with different resolutions, but our network is much shallower (12 vs.\ 38 for the smallest HRNet for classification) and the streams are fused only once, at the very end, making parallelization easier.
\section{Method}
In this section, we develop and analyze \method{}, a network architecture that is much less deep but still achieves high performance on competitive benchmarks. \method{} consists of parallel substructures that process features at different resolutions. We refer to these parallel substructures as \textit{streams}.
Features from different streams are fused at a later stage in the network, and these fused features are used for the downstream task. Figure~\ref{fig:network} provides a schematic representation of \method{}.
\subsection{\method{} Block}
\label{sec:block}
In \method{}, we utilize VGG-style blocks~\citep{simonyan2014very}. To see whether non-deep networks can achieve high performance, we empirically find that VGG-style blocks are more suitable than ResNet-style blocks (Table~\ref{tab:ablation2}). In general, training VGG-style networks is more difficult than their ResNet counterparts~\citep{he2016deep}. But recent work shows that it is easier to train networks with such blocks if one uses a ``structural reparameterization" technique~\citep{ding2021repvgg}. During training, one uses multiple branches over the 3$\times$3 convolution blocks. Once trained, the multiple branches can be fused into one 3$\times$3 convolution. Hence, one ends up with a plain network consisting of only 3$\times$3 block and non-linearity. This reparameterization or fusion of blocks helps reduce latency during inference.
We borrow our initial block design from Rep-VGG~\citep{ding2021repvgg} and modify it to make it more suitable for our non-deep architecture. One challenge with a non-deep network with only 3$\times$3 convolutions is that the receptive field is rather limited. To address this, we build a Skip-Squeeze-Excitation (SSE) layer which is based on the Squeeze-and-Excitation (SE) design~\citep{hu2018squeeze}. Vanilla Squeeze-and-Excitation is not suitable for our purpose as it increases the depth of the network. Hence we use a Skip-Squeeze-Excitation design which is applied alongside the skip connection and uses a single fully-connected layer. We find that this design helps increase performance (Table~\ref{tab:ablation1}). Figure~\ref{fig:block} provides a schematic representation of our modified Rep-VGG block with the Skip-Squeeze-Excitation module. We refer to this block as the \textit{RepVGG-SSE}.
One concern, especially with large-scale datasets such as ImageNet, is that a non-deep network may not have sufficient non-linearity, limiting its representational power. Thus we replace the ReLU activation with SiLU~\citep{ramachandran2017searching}.
\subsection{Downsampling and Fusion Block}
Apart from the RepVGG-SSE block, whose input and output have the same size, \method{} also contains \textit{Downsampling} and \textit{Fusion} blocks. The Downsampling block reduces resolution and increases width to enable multi-scale processing, while the Fusion block combines information from multiple resolutions.
In the Downsampling block, there is no skip connection; instead, we add a single-layered SE module parallel to the convolution layer. Additionally, we add 2D average pooling in the 1$\times$1 convolution branch. The Fusion block is similar to the Downsampling block but contains an extra concatenation layer. Because of concatenation, the input to the Fusion block has twice as many channels as a Downsampling block. Hence, to reduce the parameter count, we use convolution with group 2. Please refer to Figure~\ref{fig:block2} in the appendix for a schematic representation of the Downsampling and Fusion blocks.
\subsection{Network Architecture}
Figure~\ref{fig:network} shows a schematic representation of the \method{} model that is used for the ImageNet dataset. The initial layers consist of a sequence of Downsampling blocks. The outputs of Downsampling blocks 2, 3, and 4 are fed respectively to streams 1, 2, and 3. We empirically find 3 to be the optimal number of streams for a given parameter budget (Table~\ref{tab:streams}). Each stream consists of a series of RepVGG-SSE blocks that process the features at different resolutions. The features from different streams are then fused by Fusion blocks using concatenation. Finally, the output is passed to a Downsampling block at depth 11. Similar to RepVGG~\citep{ding2021repvgg}, we use a larger width for the last downsampling layer. Table~\ref{tab:arch} in the appendix provides a complete specification of the scaled \method{} models that are used in ImageNet experiments.
In CIFAR the images are of lower resolution, and the network architecture is slightly different from the one for ImageNet. First, we replace the Downsampling blocks at depths 1 and 2 with RepVGG-SSE blocks. To reduce the number of parameters in the last layer, we replace the last Downsampling block, which has a large width, with a narrower 1$\times$1 convolution layer. Also, we reduce the number of parameters by removing one block from each stream and adding a block at depth 3.
\subsection{Scaling \method{}}
With neural networks, it is observed that one can achieve higher accuracy by scaling up network size. Prior works~\citep{tan2019efficientnet} have scaled width, resolution, and depth. Since our objective is to evaluate whether high performance can be achieved with small depth, we keep the depth constant and instead scale up \method{} by increasing width, resolution, and the number of streams.
For CIFAR10 and CIFAR100, we increase the width of the network while keeping the resolution at 32 and the number of streams at 3. For ImageNet, we conduct experiments by varying all three dimensions (Figure~\ref{fig:scaling}).
\subsection{Practical Advantages of Parallel Architectures}
\label{sec:parallel}
The current process of 5 nm lithography is approaching the 0.5 nm size of the silicon crystal lattice, and there is limited room to further increase processor frequency. This means that faster inference of neural networks must come from parallelization of computation.
The growth in the performance of single monolithic GPUs is also slowing down~\citep{8192482}. The maximum die size achievable with conventional photolithography is expected to plateau at ${\sim\!\!800 \text{mm}^2}$~\citep{8192482}. On the whole, a plateau is expected not only in processor frequency but also in the die size and the number of transistors per processor.
To solve this problem, there are suggestions for partitioning a GPU into separate basic modules that lie in one package. These basic modules are easier to manufacture than a single monolithic GPU on a large die. Large dies have a large number of manufacturing faults, resulting in low yields~\citep{7856626}. Recent work has proposed a Multi-Chip-Module GPU (MCM-GPU) on an interposer, which is faster than the largest implementable monolithic GPU. Replacing large dies with medium-size dies is expected to result in lower silicon costs, significantly higher silicon yields, and cost advantages~\citep{8192482}.
Even if several chips are combined into a single package and are located on one interposer, the data transfer rate between them will be less than the data transfer rate inside one chip, because the lithography size of the chip is smaller than the lithography size of the interposer. Such chip designs thus favor partitioned algorithms with parallel branches that exchange limited data and can be executed independently for as long as possible. All these factors make non-deep parallel structures advantageous in achieving fast inference, especially with tomorrow's hardware.
\section{Results}
\begin{table}[t]
\begin{minipage}[]{0.49\columnwidth}
\centering
\caption{Depth vs.\ performance on ImageNet. We test on images with 224$\times$224 resolution. We rerun ResNets~\citep{he2016deep} in our training regime for fairness, thus boosting their accuracy. Our \method{} models perform competitively with ResNets while having a low constant depth.}
\label{tab:imagenet}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lccc}
\toprule
Model & Depth & Top-1 & Top-5\\
& (in M) & Acc. & Acc.\\
\midrule
ResNet & 18 & 69.57 & 89.24 \\
ResNet & 34 & 73.27 & 91.26 \\
ResNet-Bottleneck & 50 & 75.99 & 92.98 \\
ResNet-Bottleneck & 101 & 77.56 & 93.79 \\
\midrule
ResNet (ours)& 18 & 70.15 & 89.55 \\
ResNet (ours)& 34 & 74.12 & 91.89 \\
ResNet-Bottleneck (ours)& 26 & 77.53 & 93.87 \\
ResNet-Bottleneck (ours)& 101 & 79.63 & 94.68 \\
\midrule
\method{-S}& 12 & 75.19 & 92.29 \\
\method{-M}& 12 & 76.60 & 93.02 \\
\method{-L}& 12 & 77.66 & 93.6 \\
\method{-XL}& 12 & 78.55 & 94.13 \\
\bottomrule
\end{tabular}
}
\end{minipage} \hfill
\begin{minipage}[]{0.49\columnwidth}
\centering
\caption{Speed and performance of \method{} vs.\ ResNet. Because of parallel substructures, \method{} can be distributed across multiple GPUs. In spite of communication overhead, \method{} is faster than similar-performing ResNets.}
\label{tab:speed1}
\vspace{-1mm}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lccccc}
\toprule
Model & Depth & Top-1 & Speed & \# Param & Flops \\
& & Acc. & (samples/s) & (in M) & (in B)\\
\midrule
\method{-L} & 12 & \textbf{77.66} & \textbf{249} & 54.9 & 26.7 \\
ResNet34 & 34 & 74.12 & 248 & 21.8 & 7.3 \\
ResNet50 & 50 & 77.53 & 207 & 25.6 & 8.2 \\
\midrule
\method{-XL} & 12 & \textbf{78.55} & \textbf{230} & 85.0 & 41.5 \\
ResNet50 & 50 & 77.53 & 207 & 25.6 & 8.2 \\
\bottomrule
\end{tabular}
}
\begin{minipage}[]{\columnwidth}
\centering
\vspace{3mm}
\caption{Fusing features and parallelizing the substructures across GPUs improves the speed of \method{}. Speed was measured on a GeForce RTX 3090 with Pytorch 1.8.1 and CUDA 11.1.}
\label{tab:speed2}
\vspace{-1mm}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lccc}
\toprule
Model & Top-1 & Speed & Latency \\
& Acc. & (samples/sec) & (ms) \\
\midrule
\method{-L} (Unfused)& 77.66 & 112 & 8.95 \\
\method{-L} (Fused, Single GPU)& 77.66 & 154 & 6.50 \\
\method{-L} & 77.66 & \textbf{249} & \textbf{4.01} \\
\bottomrule
\end{tabular}
}
\end{minipage}
\end{minipage}
\end{table}
\smallsec{Experiments on ImageNet} ImageNet~\citep{deng2009imagenet} is a large-scale image classification dataset with high-resolution images. We evaluate on the ILSVRC2012~\citep{russakovsky2015imagenet} dataset, which consists of 1.28M training images and 50K validation images with 1000 classes. We train our models for 120 epochs using the SGD optimizer, a step scheduler with a warmup for first 5 epochs, a learning rate decay of 0.1 at every $30^{\text{th}}$ epoch, an initial learning rate of 0.8, and a batch size of 2048 (256 per GPU). If the network does not fit in memory, we decrease the batch size and the learning rate proportionally, for example, a decrease to a learning rate of 0.4 and a batch size of 1024. Unless otherwise specified, the network is trained at 224$\times$224 resolution. We train our networks with the cross-entropy loss with smoothed labels~\citep{szegedy2016rethinking}. We use cropping, flipping, color-jitter, and rand-augment~\citep{cubuk2020randaugment} data augmentations.
In Table~\ref{tab:imagenet}, we show the performance of \method{} on ImageNet. We find that one can achieve surprisingly high performance with a depth of just 12. For a fair comparison with ResNets, we retrain them with our training protocol and data augmentation, which improves the performance of ResNets over the official number. Notably, we find that \method{-S} outperform ResNet34 by over 1 percentage point with a lower parameter count (19M vs.\ 22M). \method{} also achieves comparable performance to ResNet with the bottleneck design, while having 4 to 8 times less depth. For example, \method{-L} performs as well as ResNet50 and gets a top-1 accuracy of 77.66 as compared to 77.53 achieved by ResNet50. Similarly, \method{-XL} performs comparably to ResNet101 and gets a top-5 accuracy of 94.13, in comparison to 94.68 achieved by ResNet101, while being 8 times shallower.
In Table~\ref{tab:speed1}, we find that \method{} performs favourably to ResNet when comparing accuracy and speed, however with more parameters and flops. For example, \method{-L} achieves faster speed and better accuracy than ResNet34 and ResNet50. Similarly, \method{-XL} achieves faster speed and better accuracy than ResNet50, however with more parameters and flops. This suggests that depending on the use case, one can use \method{}s instead of ResNets to trade off speed with more parameters and flops. Note that the high speed can be achieved by leveraging the parallel sub-structures that could be distributed across GPUs (more details below).
In Table~\ref{tab:speed2}, we test speed for three variants of \method{}: unfused, fused, and multi-GPU. The unfused variant consists of 3$\times$3 and 1$\times$1 branches in the RepVGG-SSE block. In the fused variant, we use the structural-reparametrization trick to merge the 3$\times$3 and 1$\times$1 branches into a single 3$\times$3 branch (Section~\ref{sec:block}). For both fused and unfused versions, we use a single GPU for inference, while for the multi-GPU version, we use 3 GPUs. For the multi-GPU version, each stream is launched on a separate GPU. When all layers in a stream are processed, the results from two adjacent streams are concatenated on one of the GPUs and processed further. For transferring data across GPUs we use the NCCL backend in Pytorch.
We find that \method{} can be effectively parallelized across GPUs for fast inference. This is achieved in spite of the communication overhead. With specialized hardware for reducing communication latency, even faster speeds could be achieved.
\begin{table}[tb]
\begin{minipage}[]{0.49\columnwidth}
\label{tab:coco}
\centering
\captionof{table}{Non-deep networks can be used as backbones for fast and accurate object detection systems. Speed is measured on a single RTX 3090 using Pytorch 1.8.1 and CUDA 11.1.}
\label{tab:coco}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lccc}
\toprule
Model & Backbone & MS-COCO & Latency \\
& Depth & AP & (in ms) \\
\midrule
YOLOv4-CSP (official) & 64 & 46.2 & 21.0 \\
YOLOv4-CSP (Ours) & 64 & 47.6 & 20.0 \\
ParNet-XL (Ours) & 12 & 47.5 & 18.7 \\
ParNet-XL-CSP (Ours) & 12 & \textbf{48.0} & \textbf{16.4} \\
\bottomrule
\end{tabular}
}
\end{minipage}
\hfill
\begin{minipage}[]{0.49\columnwidth}
\label{fig:accuracies}
\centering
\captionof{table}{A network with depth 12 can get $80.72\%$ top-1 accuracy on ImageNet. We show how various strategies can be used to boost the performance of \method{}.}
\label{tab:boost}
\resizebox{0.95\columnwidth}{!}{
\begin{tabular}{lcc}
\toprule
Model & Top-1 Acc. & Top-5 Acc. \\
\midrule
\method{-XL} & 78.55 & 94.13 \\
\quad + Longer Training & 78.97 & 94.51 \\
\quad + Train \& Test Res.\ 320 & 80.32 & 94.95 \\
\quad + 10-crop testing & 80.72 & 95.38 \\
\bottomrule
\end{tabular}
}
\end{minipage}
\end{table}
\begin{table*}[ht]
\caption{Performance of various architectures on CIFAR10 and CIFAR100. Similar-sized models are grouped together. \method{} performs competitively with deep state-of-the-art architectures while having a much smaller depth. Best performance is \textbf{bolded}. The second and third best performing model in each model size block are \underline{underlined}.}
\label{tab:cifar}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{l@{\hskip 4mm}c@{\hskip 4mm}c@{\hskip 4mm}c@{\hskip 4mm}c@{\hskip 4mm}}
\hline
Architecture & Depth & Params & CIFAR10 & CIFAR100 \\
& & (in Millions) & Error & Error \\
\hline
ResNet110 (official) & 110 & 1.7 & 6.61 & -- \\
ResNet110 (reported in~\citet{huang2016deep}) & 110 & 1.7 & 6.41 & 27.22 \\
ResNet (Stochastic Depth)~\citep{huang2016deep} & 110 & 1.7 & \underline{5.23} & 24.58 \\
ResNet (pre-act)~\citep{he2016identity} & 164 & 1.7 & 5.46 & \underline{24.33} \\
DenseNet~\citep{huang2017densely} & 40 & 1.0 & 5.24 & \underline{24.42} \\
DenseNet (Bottleneck+Compression)~\citep{huang2017densely} & 100 & 0.8 & \textbf{4.51} & \textbf{22.27} \\
\method{} (Ours) & 12 & 1.3 & \underline{5.0} & 24.62 \\
\hline
ResNet (Stochastic Depth)~\citep{huang2016deep} & 1202 & 10.2 & 4.91 & -- \\
ResNet (pre-act)~\citep{he2016identity} & 1001 & 10.2 & 4.62 & 22.71 \\
WideResNet~\citep{zagoruyko2016wide} & 40 & 8.9 & 4.53 & 21.18 \\
WideResNet~\citep{zagoruyko2016wide} & 16 & 11.0 & 4.27 & 20.43 \\
WideResNet (SE)~\citep{hu2018squeeze} & 32 & 12.0 & \underline{3.88} & \underline{19.14} \\
FractalNet (Compressed)~\citep{larsson2016fractalnet} & 41 & 22.9 & 5.21 & 21.49 \\
DenseNet~\citep{huang2017densely} & 100 & 7.0 & 4.10 & 20.20 \\
DenseNet (Bottleneck+Compression)~\citep{huang2017densely}& 250 & 15.3 & \textbf{3.62} & \textbf{17.60} \\
\method{} (Ours) & 12 & 15.5 & \underline{3.90} & \underline{20.02} \\
\hline
WideResNet~\citep{zagoruyko2016wide} & 28 & 36.5 & 4.00 & 19.25 \\
WideResNet (Dropout in Res-Block)~\citep{zagoruyko2016wide} & 28 & 36.5 & 3.89 & \underline{18.85} \\
FractalNet~\citep{larsson2016fractalnet} & 21 & 36.8 & 5.11 & 22.85 \\
FractalNet (Dropout+Drop-path)~\citep{larsson2016fractalnet} & 21 & 36.8 & 4.59 & 23.36 \\
DenseNet~\citep{huang2017densely} & 100 & 27.2 & \underline{3.74} & 19.25 \\
DenseNet (Bottleneck+Compression)~\citep{huang2017densely} & 190 & 25.6 & \textbf{3.46} & \textbf{17.18} \\
\method{} (Ours) & 12 & 35 & \underline{3.88} & \underline{18.65} \\
\hline
\end{tabular}}
\end{table*}
\smallsec{Boosting Performance} Table~\ref{tab:boost} demonstrates additional ways of increasing the performance of \method{}, such as using higher-resolution images, a longer training regime (200 epochs, cosine annealing), and 10-crop testing. This study is useful in assessing the accuracy that can be achieved by non-deep models on large-scale datasets like ImageNet. By employing various strategies we can boost the performance of \method{-XL} from 78.55 to 80.72. Notably, we reach a top-5 accuracy of 95.38, which is higher than the oft-cited human performance level on ImageNet~\citep{russakovsky2015imagenet}. Although this does not mean that machine vision has surpassed human vision, it provides a sense of how well \method{} performs. To the best of our knowledge, this is the first instance of such ``human-level'' performance achieved by a network with a depth of just 12.
\smallsec{Experiments on MS-COCO} MS-COCO~\citep{lin2014microsoft} is an object detection dataset which contains images of everyday scenes with common objects. We evaluate on the COCO-2017 dataset, which consists of 118K training images and 5K validation images with 80 classes.
To test whether a non-deep network such as \method{} can work for object detection, we replace the backbone of state-of-the-art single stage detectors with \method{}. Specifically, we replace the CSPDarknet53s backbone from YOLOv4-CSP~\citep{Wang_2021_CVPR} with ParNet-XL, which is much shallower (64 vs.\ 12). We use the head and reduced neck from the YOLOR-D6 model, and train and test these models using the official YOLOR code~\citep{wang2021learn}. We retrain YOLOv4-CSP~\citep{Wang_2021_CVPR} with our protocol (same neck, same head, same training setup) for fair comparison and it improves performance over the official model. Additionally, for fair comparison, we test the ParNet-XL-CSP model by applying the CSP~\citep{Wang_2021_CVPR} approach to ParNet-XL. We find that ParNet-XL and ParNet-XL-CSP are faster than the baseline even at higher image resolution. We thus use higher image resolution for ParNet-XL and ParNet-XL-CSP.
In Table~\ref{tab:coco} we find that even on a single GPU, \method{} achieves higher speed than strong baselines. This demonstrates how non-deep networks could be used to make fast object detection systems.
\smallsec{Experiments on CIFAR} The CIFAR datasets consist of colored natural images with 32$\times$32 pixels. CIFAR-10 consists of images drawn from 10 and CIFAR-100 from 100 classes. The training and test sets contain 50,000 and 10,000 images respectively. We adopt a standard data augmentation scheme (mirroring/shifting) that is widely used for these two datasets~\citep{he2016deep,zagoruyko2016wide,huang2017densely}. We train for 400 epochs with a batch size of 128. The initial learning rate is 0.1 and is decreased by a factor of 5 at 30\%, 60\%, and 80\% of the epochs as in~\citep{zagoruyko2016wide}. Similar to prior works~\citep{zagoruyko2016wide,huang2016deep}, we use a weight decay of 0.0003 and set dropout in the convolution layer at 0.2 and dropout in the final fully-connected layer at 0.2 for all our networks on both datasets. We train each network on 4 GPUs (a batch size of 32 per GPU) and report the final test set accuracy.
Table~\ref{tab:cifar} summarizes the performance of various networks on CIFAR10 and CIFAR100. We find that \method{} performs competitively with state-of-the-art deep networks like ResNets and DenseNets while using a much lower depth and a comparable number of parameters. \method{} outperforms ResNets that are 10 times deeper on CIFAR10 (5.0 vs.\ 5.23) while using a lower number of parameters (1.3M vs.\ 1.7M). Similarly, \method{} outperforms ResNets that are 100 times deeper on CIFAR10 (3.90 vs.\ 4.62) and CIFAR100 (20.02 vs.\ 22.71) while using 50\% more parameters (15.5M vs.\ 10.2M).
\method{} performs as well as vanilla DenseNet models~\citep{huang2017densely} with comparable parameter counts while using 3-8 times less depth. For example, on CIFAR100, \method{} (depth 12) achieves an error of 18.65 with 35M parameters and DenseNet (depth 100) achieves an error of 19.25 with 27.2M parameters. \method{} also performs on par or better than Wide ResNets~\citep{zagoruyko2016wide} while having 2.5 times less depth. The best performance on the CIFAR dataset under a given parameter count is achieved by DenseNets with the bottleneck and reduced width (compression) design, although with an order of magnitude larger depth than \method{}.
Overall, it is surprising that a mere depth-12 network could achieve an accuracy of 96.12\% on CIFAR10 and 81.35\% on CIFAR100. This further indicates that non-deep networks can work as well as deeper counterparts.
\begin{table}[tb]
\begin{minipage}[]{0.44\columnwidth}
\label{fig:simpleview}
\centering
\captionof{table}{Ablation of various choices for \method{}. Data augmentation, SiLU activation, and Skip-Squeeze-Excitation (SSE) improve performance.}
\label{tab:ablation1}
\resizebox{0.98\columnwidth}{!}{
\begin{tabular}{lcccc}
\toprule
Model & Params & Top-1 & Top-5 \\
& (in M) & Acc. & Acc. \\
\midrule
Baseline & 32.4 & 75.02 & 92.15 \\
\quad + Data Augmentation & 32.4 & 75.12 & 92.00 \\
\quad + SSE & 35.6 & 76.55 & 93.01 \\
\quad + SiLU & 35.6 & 76.60 & 93.07 \\
\bottomrule
\end{tabular}
}
\end{minipage}
\hfill
\begin{minipage}[]{0.54\columnwidth}
\centering
\captionof{table}{ParNet outperforms non-deep ResNet variants. At depth 12, VGG-style blocks outperform ResNet blocks, and three branches outperform a single branch.}
\label{tab:ablation2}
\resizebox{0.99\columnwidth}{!}{
\begin{tabular}{llcccc}
\toprule
Name & Block & Branch & Depth & Params & Top-1 \\
& & & & (in M) & Acc. \\
\midrule
ResNet12-Wide & ResNet & 1 & 12 & 39.0 & 73.52 \\
ResNet14-Wide-BN & ResNet-BN & 1 & 14 & 39.0 & 72.06 \\
ResNet12-Wide-BN & ResNet-SSE & 1 & 12 & 39.0 & 73.91 \\
ParNet-M-OneStream & RepVGG-SSE & 1 & 12 & 36.0 & 75.83 \\
ParNet-M & RepVGG-SSE & 3 & 12 & 35.9 & 76.6 \\
\bottomrule
\end{tabular}
}
\vspace{-1mm}
\end{minipage}
\end{table}
\begin{table}[t]
\caption{\method{} outperforms ensembles across different parameter budgets.}
\label{tab:ensemble}
\resizebox{\columnwidth}{!}{
\begin{tabular}{ccccccc}
\toprule
& ParNet-M & Single Stream & ParNet-L & Ensemble & ParNet-XL & Ensemble \\
& & & & (2 Single Streams) & & (3 Single Streams) \\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\cmidrule(lr){6-7}
\# Param & 35.9 & 36 & 55 & 72 & 85.3 & 108 \\
ImageNet Top-1 & \textbf{76.60} & 75.83 & \textbf{77.66} & 77.20 & \textbf{78.55} & 77.68 \\
\bottomrule
\end{tabular}
}
\end{table}
\begin{table}[t]
\centering
\caption{Performance vs.\ number of streams. For a fixed parameter budget, 3 streams is optimal.}
\label{tab:streams}
\begin{tabular}{ccccccccc}
\toprule
& \multicolumn{4}{c}{\# of Branches (Res. 224)} & \multicolumn{4}{c}{\# of Branches (Res. 320)} \\
\cmidrule(lr){2-5}
\cmidrule(lr){6-9}
& 1 & 2 & 3 & 4 & 1 & 2 & 3 & 4 \\ \midrule
ImageNet Top-1 & 75.83 & 76.1 & \textbf{76.75} & 76.34 & 76.91 & 77.12 & \textbf{77.56} & 77.46 \\
\bottomrule
\end{tabular}
\end{table}
\smallsec{Ablation Studies} To test if we can trivially reduce the depth of ResNets and make them wide, we test three ResNet variants: ResNet12-Wide, ResNet14-Wide-BN, and ResNet12-Wide-SSE. ResNet12-Wide uses the basic ResNet block and has depth 12, while ResNet14-Wide-BN uses the bottleneck ResNet block and has depth 14. Note that with the bottleneck ResNet block, one cannot achieve a depth lower than 14 while keeping the original ResNet structure as there has to be 1 initial convolution layer, 4 downsampling blocks (3 $\times$ 4 = 12 depth), and 1 fully-connected layer. We find that ResNet12-Wide outperforms ResNet14-Wide-BN with the same parameter count. We additionally add SSE block and SiLU activation to ResNet12-Wide to create ResNet12-Wide-SSE to further control for confounding factors. We find that \method{-M} outperforms all the ResNet variants which have depth 12 by 2.7 percentage points, suggesting that trivially reducing depth and increasing width is not as effective as our approach. We show that the model with three branches performs better than a model with a single branch. We also show that with everything else being equal, using VGG-style blocks leads to better performance than the corresponding ResNet architecture. Architectural choices in \method{} like parallel substructures and VGG-style blocks are crucial for high performance at lower depths.
Table~\ref{tab:ablation1} reports ablation studies over various design choices for our network architecture and training protocol. We show that each of the three decisions (rand-augment data augmentation, SiLU activation, SSE block) leads to higher performance. Using all three leads to the best performance.
In Table~\ref{tab:streams}, we evaluate networks with the same total number of parameters but with different numbers of branches: 1, 2, 3, and 4. We show that for a fixed number of parameters, a network with 3 branches has the highest accuracy and is optimal in both cases, with a network resolution of 224x224 and 320x320.
\smallsec{\method{} vs.\ Ensembles} Another approach to network parallelization is the creation of ensembles consisting of multiple networks. Therefore, we compare \method{} to ensembles. In Table~\ref{tab:ensemble}, we find that \method{} outperforms ensembles while using fewer parameters.
\smallsec{Scaling \method{}} Neural networks can be scaled by increasing resolution, width, and depth~\citep{tan2019efficientnet}. Since we are interested in exploring the performance limits of constant-depth networks, we scale \method{} by varying resolution, width, and the number of streams.
Figure~\ref{fig:scaling} shows that each of these factors increases the accuracy of the network. Also, we find that increasing the number of streams is more cost-effective than increasing the number of channels in terms of accuracy versus parameter count. Further, we find that the most effective way to scale \method{} is to increase all three factors simultaneously. Because of computation constraints, we could not increase the number of streams beyond 3,
but this is not a hard limitation. Based on these charts, we see no saturation in performance while scaling \method{s}. This indicates that by increasing compute, one could achieve even higher performance with \method{} while maintaining low depth.
\begin{figure}[t]
\begin{subfigure}[t]{0.5\textwidth}
\centering
\resizebox{\columnwidth}{!}{\includegraphics[trim=1.9cm 18.0cm 5.0cm 2.0cm, clip, scale=0.9]{figures/acc_param.pdf}}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.5\textwidth}
\centering
\resizebox{\columnwidth}{!}{\includegraphics[trim=1.9cm 18.0cm 5.0cm 2.0cm, clip, scale=0.9]{figures/acc_param_scale.pdf}}
\end{subfigure}
\vspace{-2mm}
\caption{Performance of \method{} increases as we increase the number of streams, input resolution, and width of convolution, while keeping depth fixed. The left plot shows that under a fixed parameter count, the most effective way to scale \method{} is to use 3 branches and high resolution. The right plot shows the impact on performance by changing only one of the aforementioned factors. We do not observe saturation in performance, indicating that \method{s} could be scaled further to increase performance while maintaining low depth.}
\label{fig:scaling}
\end{figure}
\section{Discussion}
We have provided the first empirical proof that non-deep networks can perform competitively with their deep counterparts in large-scale visual recognition benchmarks. We showed that parallel substructures can be used to create non-deep networks that perform surprisingly well. We also demonstrated ways to scale up and improve the performance of such networks without increasing depth.
Our work shows that there exist alternative designs where highly accurate neural networks need not be deep. Such designs can better meet the requirements of future multi-chip processors. We hope that our work can facilitate the development of neural networks that are both highly accurate and extremely fast.
|
\section{Introduction} \label{sec:intro}
When studying treatment effects in the social sciences, study units are often clustered together and form dependencies. For example, in educational assessment studies, such as the Trends in International Mathematics and Science Study (TIMSS), the Programme for International Student Assessment (PISA), or the Early Childhood Longitudinal Study (ECLS), students (i.e., the study units) are clustered at the classroom or school level. In the National Study of Learning Mindsets that was part of the workshop at the 2018 Atlantic Causal Inference Conference \citep{Carvalho2019}, students (i.e., the study units) are clustered at the school level. In cross-classified educational studies, students belong to two, non-nested clusters, say school and neighborhood. In all these examples, which are broadly known as multilevel studies, students' data from the same school or the same neighborhood may be correlated due to shared school or neighborhood characteristics and properly accounting for the correlation structure, ideally with minimal assumptions, is paramount to obtain valid and robust inference of treatment effects \citep{MLbook1}. The main goal of our paper is to study nonparametric, doubly robust estimation of treatment effects in multilevel studies where there are only nonparametric constraints on the observed multilevel data. A useful byproduct of our paper is one formal framework to leverage existing i.i.d.-based machine learning methods to more efficiently estimate treatment effects in clustered/multilevel settings.
There is a long and rich history of studying treatment effects in multilevel studies; see \citet{Review2}, \citet{Review1}, and references therein for the most recent reviews.
Broadly speaking, to account for correlation in multilevel studies, many of the popular methods use parametric mixed effect models or semiparametric estimating equations
\citep{GEE,Gelman2006,hongraudenbush06,ArpinoMealli2011,thoemmes2011ps,LZL2013, ArpinoCannas2016}. Some well-known methods include generalized linear mixed effects models (GLMMs) and generalized estimating equations (GEEs). However, a key limitation with these methods is that their respective parametric components are assumed to be correctly specified.
A bit more formally,
if $\mathbf{O}_{ij} = (Y_{ij}, A_{ij}, \mathbf{X}_{ij})$ denote the outcome, binary treatment, and pre-treatment covariates, respectively, of study unit $j$ in cluster $i$, GLMMs assume that the random effects, which model the correlation between study units in a cluster, are Normally distributed and the outcomes of a pair of units $ij$ and $ij'$, $j\neq j'$, follow a conditional bivariate Normal distribution with a constant off-diagonal covariance matrix.
GEEs and other semiparametric approaches do not make distributional assumptions, but the outcome model $\text{E} ( Y_{ij} \, \vert \, A_{ij}, \mathbf{X}_{ij})$ and/or the propensity score model $\text{E} ( A_{ij} \, \vert \, \mathbf{X}_{ij})$ \citep{Rosenbaum1983} are often assumed to be (semi-)parametric \citep{RobinsSNMM1994, Gray1998, Robins1999, Gray2000, RobinsMSM, Goetgeluk2008, Stephens2012, LZL2013, Zetterqvist2016, Liu2018, Review2, Review1, Lee2021}.
Unfortunately, all the aforementioned methods may lead to inconsistent estimates of treatment effects when the parametric component of their estimators are mis-specified.
Recently, there has been great interest in using supervised machine learning (ML) methods to estimate treatment effects where both of the outcome and the treatment are modeled nonparametrically; see \citet{SL2007}, \citet{Hill2011}, \citet{vvLaan2011}, \citet{AtheyImbens2016}, \citet{Victor2018}, \citet{grf}, \citet{Hahn2020}, \citet{NieWager2020}, and references therein. But, many of these estimators are designed for
i.i.d. data where $\mathbf{O}_{ij} \stackrel{\mathrm{iid}}{\sim} P$ for some distribution $P$ and it is not clear how to properly apply these flexible ML methods in multilevel settings where the data within each cluster are correlated, say when $P(\mathbf{O}_{ij},\mathbf{O}_{ij'}) \neq P(\mathbf{O}_{ij})P(\mathbf{O}_{ij'})$ for $j\neq j'$.
We present a new type of nonparametric, doubly robust and asymptotically Normal estimators of treatment effects in multilevel studies. At a high level, our estimator allows both (i) the outcome regression $\text{E} (Y_{ij} \, \vert \, A_{ij}, \mathbf{X}_{ij})$, (ii) the propensity score $\text{E} (A_{ij} \, \vert \, \mathbf{X}_{ij})$, and (iii) the correlation structure between study units to be nonparametric.
In other words, we allow both the ``first-order'' conditional moments (i.e., (i) and (ii)) and the ``second-order'' conditional moments (i.e., (iii)) to be nonparametric. To achieve this, the estimator uses two non-standard nuisance functions in causal inference that nonparametrically capture the correlation structure in the treatment and the outcome: (i) a conditional propensity score, which is the propensity score of a study unit given his/her peers' treatment status and pre-treatment covariates, and (ii) an outcome covariance model, which measures the correlation in the residualized outcomes of study units in the same cluster. Also, the new estimator allows investigators to use existing i.i.d.-based ML methods in causal inference, but in a manner that allows for the estimator to be doubly robust and potentially more efficient in multilevel settings. In real data, we find that our estimator consistently has dramatically smaller variance the variances of competing estimators, especially when studying cluster-specific treatment effects; see Section \ref{sec:application} for details.
\section{Setup} \label{sec:Setup}
\subsection{Review: Notation}
Let $i=1,\ldots,N$ index $N$ clusters and let $j=1,\ldots,n_{i}$ be the number of individuals (i.e., study units) within cluster $i$. In two-level studies, $i$ would denote clusters and $j$ would denote individuals within a cluster. In three-level observational studies, $i$ would denote clusters jointly defined by second and third levels and $j$ would denote individuals within this cluster; one can also expand the cluster index $i$ using double subscripts, say $st$ where $s$ indexes clusters at the second-level and $t$ indexes clusters at the third-level.
In cross-classified studies with non-nested clusters, $i$ would denote each cluster defined by interacting non-nested clusters, say school and neighborhood, and $j$ would denote individuals within the interacted cluster. More generally, we can expand the subscript $i$ to better denote two, three, or $k$-level hierarchical structures in a multilevel study. But, for simplicity, we suppress these double or triple subscripts.
For each study unit $j$ in cluster $i$, let $\mathbf{X}_{ij}=(\mathbf{W}_{ij}^{\intercal}, \mathbf{C}_i^{\intercal} )^{\intercal}$ denote the study unit's pre-treatment covariates where $\mathbf{W}_{ij}$ denotes unit-level pre-treatment covariates and $\mathbf{C}_i$ denotes cluster-level pre-treatment covariates. In educational assessment studies like the ECLS, unit-level covariates may include student's gender, age, prior test scores, and socioeconomic status. Cluster-level covariates may include type of school, say private or public schools, location of the school, say rural, suburban, or city, and cluster size $n_i$. Often, cluster-level covariates drive the dependence between study units in a cluster. For example, if the outcome is student test scores, school location may be one of the reasons why students from the same school may have similar test scores. In Section \ref{sec:result}, we utilize this observational about multilevel studies to help train the aforementioned outcome covariance model that captures the outcome correlation between study units.
Let $A_{ij} \in \{0,1\}$ denote treatment assignment where $A_{ij} = 1$ indicates that individual $j$ in cluster $i$ was treated and $A_{ij} = 0$ indicates that individual $j$ in cluster $i$ was untreated. Let $Y_{ij}$ denote the observed outcome of individual $j$ in cluster $i$. We use the following notation for vectors of observed variable: $\mathbf{Y}_i = (Y_{i1}, \ldots, Y_{in_i})^{\intercal}$, $\mathbf{A}_i = (A_{i1}, \ldots, A_{in_i})^{\intercal}$, and $\mathbf{X}_i = (\mathbf{X}_{i1}^{\intercal}, \ldots, \mathbf{X}_{in_i}^{\intercal})^{\intercal}$. Also, let $\mathbf{A}_ {i(-j)} $ be the vector of treatment variables for all study units in cluster $i$ except individual $j$. Overall, for each individual $j$ in cluster $i$, we observe $\mathbf{O}_{ij} = (Y_{ij}, A_{ij}, \mathbf{X}_{ij})$ and for each cluster $i$, we observe $\mathbf{O}_i = (\mathbf{O}_{i1},\ldots,\mathbf{O}_{i n_i})$.
We adopt the following notations for convergence, norms, and vectors. For a measurable function $f$, let $\|f\|_{P,2}$ denote the $L_2(P)$-norm of $f$, i.e., $\|f\|_{P,2}= \big\{ \int \| f(\mathbf{o}) \|_2^2 \, dP(\mathbf{o}) \big\}^{1/2}$.
Let $O(\cdot)$ and $o(\cdot)$ be the usual big-o and small-o notation, respectively. We denote $m$-dimensional vectors of ones and zeros as $\bm{1}_m$ and $\bm{0}_m$, respectively, and denote the $m \times m$ identity matrix as $I_m$.
\subsection{Review: Causal Estimands and Causal Identification}
We use the potential outcomes notation \citep{Neyman1923, Rubin1974}
to define causal effects. Let $\pot{Y_{ij}}{1}$ denote the potential outcome if individual $i$ in cluster $j$ were treated and let $\pot{Y_{ij}}{0}$ denote the potential outcome if the individual were untreated.
The target estimand of interest is the (weighted) average treatment effect and is denoted as $\tau^*= \text{E} [ w(\mathbf{C}_i) \{ \pot{Y_{ij}}{1} - \pot{Y_{ij}}{0} \} ]$.
The weight function $w(\mathbf{C}_i) \in \mathbb{R}$ is bounded and specified by the investigator based on the scientific question at hand.
For example, if $w(\mathbf{C}_i) =1$, $\tau^*$ becomes the usual (i.e., unweighted or equally-weighted) average treatment effect. But, if an investigator wants to over-represent certain clusters, say schools in rural areas are weighed more than schools in non-rural areas when taking the average across schools, $w(\mathbf{C}_i)$ can vary as a function of $\mathbf{C}_i$. Or, if investigators only want to study certain clusters, we can set $w(\mathbf{C}_i)=1$ for the clusters of interest, $w(\mathbf{C}_i) =0$ otherwise, and re-weigh the estimand to study the (conditional) average treatment effects among schools with $w(\mathbf{C}_i)=1$; see Section \ref{sec:application} for examples. For simplicity, we will refer to both the unweighted/equally-weighted and weighted/unequally-weighted treatment effects as the average treatment effect in the paper.
To identify the average treatment effect $\tau^*$ from observed data, we assume the following.
\begin{itemize}
\item[\hypertarget{(A1)}{(A1)}] (\textit{Stable Unit Treatment Value Assumption, \citet{Rubin1976, Rubin1978}}) $Y_{ij} = \pot{Y_{ij}}{A_{ij}}$.
\item[\hypertarget{(A2)}{(A2)}] (\textit{Conditional Ignorability}) $\pot{Y_{ij}}{1} , \pot{Y_{ij}}{0} \perp A_{ij} \, \vert \, \mathbf{X}_{ij}$.
\item[\hypertarget{(A3)}{(A3)}] (\textit{Overlap}) There exists a finite constant $\delta > 0$ such that $\delta \leq P(\mathbf{A}_{i} = \mathbf{a} \, \vert \, \mathbf{X}_{i}= \mathbf{x}) \leq 1- \delta$ for all $(\mathbf{a},\mathbf{x})$.
\end{itemize}
\noindent Briefly, Assumptions \hyperlink{(A1)}{(A1)} and \hyperlink{(A2)}{(A2)} are well-known identifying assumptions in causal inference. Notably,
similar to past works in multilevel studies (e.g., \citet{ArpinoMealli2011}, \citet{LZL2013}, \citet{ArpinoCannas2016}, \citet{Lee2021}), we assume that interference is absent. Assumption \hyperlink{(A3)}{(A3)} is a slight generalization of the usual overlap assumption to allow for a vector of treatment indicators in cluster $i$. Assumption \hyperlink{(A3)}{(A3)} implies the usual overlap assumption for the ``individual-level'' propensity score $e^*(a \, \vert \, \mathbf{x}) = P(A_{ij} = a \, \vert \, \mathbf{X}_{ij} = \mathbf{x})$, i.e., $c_e \leq P(A_{ij} = a \, \vert \, \mathbf{X}_{ij} = \mathbf{x}) \leq 1-c_e$ for a positive constant $c_e$. In the paper, we refer to $e^*(a \, \vert \, \mathbf{x})$ as the individual-level propensity score and $P(\mathbf{A}_{i} = \mathbf{a} \, \vert \, \mathbf{X}_{i}= \mathbf{x})$ as the cluster-level propensity score.
\citet{hongraudenbush06} and \citet{hong2013heterogeneous} discuss the practical merits of Assumptions \hyperlink{(A1)}{(A1)}-\hyperlink{(A3)}{(A3)} when estimating treatment effects in multilevel studies.
Under Assumptions \hyperlink{(A1)}{(A1)}-\hyperlink{(A3)}{(A3)}, the average treatment effect $\tau^*$ can be written as a function of observed data via
\begin{align} \label{eq-ident}
& \tau^* = \text{E} \big[ w(\mathbf{C}_i) \big\{ g^*(1,\mathbf{X}_{ij}) - g^*(0, \mathbf{X}_{ij}) \big\} \big]
\end{align}
where $g^*(1,\mathbf{x}) = \text{E} (Y_{ij} \, \vert \, A_{ij} = 1, \mathbf{X}_{ij}=\mathbf{x})$ and $g^*(0,\mathbf{x}) = \text{E} (Y_{ij} \, \vert \, A_{ij} = 0, \mathbf{X}_{ij}=\mathbf{x})$ are the true outcome regression models under treatment and control, respectively.
The rest of the paper focuses on estimating $\tau^*$, specifically the functional on the right-hand side of equation \eqref{eq-ident}, when the observed data $\mathbf{O}_{ij}$ comes from a multilevel study, i.e., when $\mathbf{O}_{ij} \stackrel{\mathrm{iid}}{\sim} P$ no longer holds.
\subsection{Review: Existing Estimators in Multilevel Studies} \label{sec:2-3}
To better frame our contribution, we briefly review existing estimators of $\tau^*$ in multilevel studies. Our review focuses on existing doubly robust estimators of $\tau^*$ as our proposed estimator is also doubly robust; for a more complete review of estimators in multilevel studies, including estimators based on inverse probability weighting, see \cite{MLbook1}, \citet{thoemmes2011ps}, \citet{Review2} and \citet{Review1}.
Broadly speaking, all existing doubly robust estimators of $\tau^*$ in multilevel settings are variations of the augmented inverse probability weighted (AIPW) estimator \citep{Robins1994, Scharfstein1999}, denoted as $\overline{\tau}$ below.
\begin{align} \label{eq-bartau}
& \overline{\tau} = \overline{\tau} (e, g)
= \frac{1}{N} \sum_{i=1}^{N} {\varphi}(\mathbf{O}_i, e,g) \ , \
\\
&
\varphi(\mathbf{O}_i, e , g)
= \frac{w(\mathbf{C}_i)}{n_i} \sum_{j =1}^{n_i} \bigg[ \frac{A_{ij} \big\{ Y_{ij} - g(1, \mathbf{X}_{ij}) \big\} }{e(1 \, \vert \, \mathbf{X}_{ij})} + g(1, \mathbf{X}_{ij})
- \frac{(1 - A_{ij}) \big\{ Y_{ij} - g(0, \mathbf{X}_{ij}) \big\} }{e(0 \, \vert \, \mathbf{X}_{ij})} - g(0, \mathbf{X}_{ij}) \bigg]. \nonumber
\end{align}
So long as either the outcome model $g$ or the individual-level propensity score $e$ is correctly specified, but not necessarily both, $\overline{\tau}$ is consistent for $\tau^*$ (i.e., doubly robust). Also, under addition assumptions, $\overline{\tau}$ is asymptotically Normal around $\tau^*$. For completeness, Section \ref{Supp-sec:ExistingMethod}
of the Supplementary Materials provides one proof of these statistical properties that allow for nonparametrically estimated $g$ and $e$. But, the citations we provide in the next paragraph provide alternative proofs of these statistical properties under different, often more restrictive, assumptions about $g$ and $e$.
Existing doubly robust estimators differ on how $g$ and $e$ in $\overline{\tau}$ are modeled. For example, one popular choice, especially in educational studies like our data example below, is to use generalized linear mixed effect models (GLMMs) \citep{Goldstein2002, hongraudenbush06, ArpinoMealli2011,thoemmes2011ps,ArpinoCannas2016,Shardell2018}, say
\begin{align*}
& Y_{ij} = \alpha_0 + A_{ij} \alpha_1 +
\mathbf{X}_{ij} ^{\intercal} \bm{\beta} + U_i + \epsilon_{ij} , \quad{}
&& U_i \stackrel{\mathrm{iid}}{\sim} N(0,\sigma_U^2), \quad{} \epsilon_{ij} \stackrel{\mathrm{iid}}{\sim} N(0, \sigma_\epsilon^2),
\\
&
P \big( A_{ij} = 1 \, \vert \, \mathbf{X}_i, V_i \big)
=
\text{expit}
\big( \mathbf{X}_{ij} ^{\intercal} \bm{\gamma} + V_i \big), \quad{}
&&
V_i \stackrel{\mathrm{iid}}{\sim} N(0,\sigma_V^2),
\nonumber \\
& g(A_{ij}, \mathbf{X}_{ij}) = \int P(Y_{ij} \, \vert \, A_{ij}, \mathbf{X}_{ij}, U_i = u) dP(u), \quad{}
&& e(1 \, \vert \, \mathbf{X}_{ij}) = \int P \big( A_{ij} = 1 \, \vert \, \mathbf{X}_{ij}, V_i = v \big) dP(v) \ .
\end{align*}
where $(\mathbf{X}_{i}, U_i, V_i) \perp \epsilon_{ij}$ and $\mathbf{X}_{i} \perp (U_i, V_i)$. The terms $U_i$ and $V_i$ are unobserved, Normally distributed random effects that govern the correlation between study units in a cluster.
Another popular choice is to use generalized estimating equations (GEEs) \citep{GEE, Gray1998, Robins1999, Gray2000, Goetgeluk2008, Stephens2012, Zetterqvist2016}, say $ g(A_{ij},\mathbf{X}_{ij}) = \alpha_0 + A_{ij} \alpha_1 + \mathbf{X}_{ij} ^{\intercal} \bm{\beta}$ and $e(1 \, \vert \, \mathbf{X}_{ij}) = \text{expit}
\big( \mathbf{X}_{ij} ^{\intercal} \bm{\gamma} \big)$. Unlike GLMMs which parametrize the within-cluster correlations, GEEs leave the correlation structure unspecified.
Unfortunately, both GLMMs and GEEs require parametric models for the outcome regression or the propensity score. If these models are mis-specified, $\overline{\tau}$ may be inconsistent for $\tau^*$.
Finally, a modern, but somewhat perplexing approach would be to use supervised ML models for the outcome regression $g$ and the propensity score $e$.
We say perplexing because commonly used supervised ML models in causal inference often assume that the data, specifically $\mathbf{O}_{ij}$, is i.i.d. even though in multilevel settings, it is not. But, as noted by \citet{Carvalho2019} from the aforementioned 2018 Atlantic Causal Inference Conference workshop in Section \ref{sec:intro}, using i.i.d.-based ML methods was common to study treatment effects in multilevel settings, likely due to its simplicity.
Yet, to the best of our knowledge, there is no theoretical justification as to whether this would be the most appropriate or efficient way of using i.i.d.-based supervised ML models to estimate $\tau^*$ in multilevel studies. Our proposed estimator below can be seen an answer to this question by proposing a different way to use i.i.d.-based modern ML methods to obtain a doubly robust, asymptotically Normal, and a more efficient estimator of $\tau^*$.
\section{Our Approach: A More Efficient, Nonparametric, Doubly Robust Estimator in Multilevel Studies} \label{sec:result}
\subsection{Key Ideas: Conditional Propensity Score and Outcome Covariance Model} \label{sec:overview_est}
At a high level, our proposed estimator, denoted as $\widehat{\tau}$, retains some of the favorable properties of the existing estimator $\overline{\tau}$ discussed above, notably double robustness. But, in settings where $\overline{\tau}$ is consistent and asymptotically Normal for $\tau^*$ in multilevel studies, $\widehat{\tau}$ can be more efficient than $\overline{\tau}$. For simplicity, this section introduces $\widehat{\tau}$ when the nuisance functions are known and subsequent sections discusses how to estimate the nuisance functions using i.i.d.-based ML methods.
The new estimator $\widehat{\tau}$ contains two new functions to model the correlation structure between study units: (i) the conditional propensity score, denoted as $\pi^* ( a \, \vert \, \mathbf{a}' , \mathbf{x}, \mathbf{x}' ) = P \big( A_{ij} = a \, \vert \, \mathbf{A}_ {i(-j)} = \mathbf{a}', \mathbf{X}_{ij} = \mathbf{x}, \mathbf{X}_{ {i(-j)} } = \mathbf{x}' \big)$, and (ii) the outcome covariance model $\beta(\mathbf{C}_i)$.
These two functions are used inside the proposed estimator $\widehat{\tau}$ as follows:
\begin{align*
& \widehat{\tau}(\pi, g, \beta) = \frac{1}{N} \sum_{i=1}^N \phi(\mathbf{O}_i,\pi,g,\beta), \\[-0.1cm]
\nonumber
&
\phi(\mathbf{O}_i, \pi, g, \beta)
= \frac{w(\mathbf{C}_i)}{n_i} \sum_{j=1}^{n_i} \bigg[ \frac{A_{ij} \big[ \{ Y_{ij} - g(1, \mathbf{X}_{ij}) \} - \beta(\mathbf{C}_i) \sum_{k\neq j} \{ Y_{ik} - g(A_{ik}, \mathbf{X}_{ik}) \} \big] }{\pi(1 \, \vert \, \mathbf{A}_{i(-j)}, \mathbf{X}_{ij}, \mathbf{X}_{i(-j)})} + g(1, \mathbf{X}_{ij})
\nonumber
\\[-0.1cm]
&
\hspace*{2cm}
-
\frac{(1-A_{ij}) \big[ \{ Y_{ij} - g(0, \mathbf{X}_{ij})\} - \beta(\mathbf{C}_i) \sum_{k\neq j} \{Y_{ik} - g(A_{ik}, \mathbf{X}_{ik})\}\big]}{\pi(0 \, \vert \, \mathbf{A}_{i(-j)}, \mathbf{X}_{ij}, \mathbf{X}_{i(-j)})} - g(0, \mathbf{X}_{ij})
\bigg]\ .
\nonumber
\hspace*{-1cm}
\end{align*}
Compared to the existing estimator $\overline{\tau}$, the new estimator $\widehat{\tau}$ captures the correlation structure in the treatment and the outcome variables. Specifically, for the treatment variable, $\widehat{\tau}$ uses the conditional propensity score $\pi$ to capture the (conditional) correlation between $ij$'s treatment $A_{ij}$ and his/her peers $\mathbf{A}_ {i(-j)} $ conditional on $ij$'s and his/her peers' covariates, $\mathbf{X}_{ij}$ and $\mathbf{X}_ {i(-j)} $, respectively. As we show below, the conditional propensity score reduces to the usual, individual-level propensity score in the absence of any correlation between treatment variables. More importantly, as we show in the theorems below, the conditional propensity score is, in some sense, the ``correct'' inverse probability weight to use in order to simultaneously maintain double robustness and achieve efficiency gains in multilevel settings.
For the outcome variable, $\widehat{\tau}$ uses the outcome covariance model $\beta(\mathbf{C}_i)$ to capture the relationship between $ij$'s (residual) outcome relative to his/her peers' (residual) outcomes; here, the residual refers to the leftover variation in an individual's outcome after conditioning on his/her own treatment and pre-treatment covariates. Specifically,
$\beta(\mathbf{C}_i)$ acts as a (functional) regression coefficient from regressing $ij$'s residual (i.e., $Y_{ij} - g(A_{ij}, \mathbf{X}_{ij})$) onto the sum of his/her peers' residuals (i.e., $\sum_{k\neq j} \{Y_{ik} - g(A_{ik}, \mathbf{X}_{ik}) \} $).
The new estimator $\widehat{\tau}$ collapses to the existing estimator $\overline{\tau}$ under additional assumptions about the conditional propensity score and the correlation structure. For example, if the $ij$'s treatment is independent of his/her peers' treatments and pre-treatment covariates conditional on his/her own pre-treatment covariates, $\pi$ reduces the usual, individual-level propensity score $e$ and the inverse probability weights are identical between $\widehat{\tau}$ and $\overline{\tau}$.
Or, if the residuals of everyone are independent of each other, or equivalently, everyone's outcomes are independent of each other conditional on their own treatment and pre-treatment covariates, a hypothetical regression between $ij$'s residuals onto his/her peers residuals would lead to an estimated coefficient of $\beta(\mathbf{C}_i) = 0$, matching the equivalent expression in equation \eqref{eq-bartau} for $\overline{\tau}$. Finally, if both independence conditions hold, $\widehat{\tau}$ equals $\overline{\tau}$. In short,
$\overline{\tau}$ can be seen as a special case of $\widehat{\tau}$ when there are no correlations between the outcome and the treatment of study units in the same cluster.
Given the conditional propensity score $\pi$ and the outcome regression $g$, let $\beta_{\pi,g}^*$ be the minimizer of the variance of the estimator $\widehat{\tau}(\pi,g,\beta)$ over some model space for the outcome covariance model $\mathcal{B}$ that contains 0, i.e., $\beta_{\pi,g}^* = \argmin_{\beta \in \mathcal{B} } \text{Var} \big\{ \widehat{\tau}(\pi,g,\beta) \}$ where $0 \in \mathcal{B}$; the next section discusses different model spaces $\mathcal{B}$. Lemma \ref{thm:2} highlights some simple, but useful characteristics of $\widehat{\tau}$ when we assume, for a moment, that the nuisance functions are fixed, non-random functions.
\begin{lemma} \label{thm:2} Suppose Assumptions \hyperlink{(A1)}{(A1)}-\hyperlink{(A3)}{(A3)} and the following assumptions hold.
\begin{itemize}
\item[\hypertarget{(M1)}{(M1)}] (\textit{Nonparametric Model}) For some constants ${\overline{M}}$, the observed data satisfies the following; see Section \ref{sec:M1} for examples of data generating processes in multilevel studies which satisfy \hyperlink{(M1)}{(M1)}.
\begin{align*}
&
P(\mathbf{O}_1,\ldots,\mathbf{O}_N) =
\prod_{i=1}^{N} P (\mathbf{O}_i) \ , \\
&
P (\mathbf{O}_i) \in \left\{ P \left|
\begin{array}{l}
n_i \leq {\overline{M}}, \
\text{E} \big( Y_{ij} \, \vert \, A_{ij} =a, \mathbf{X}_{ij} = \mathbf{x} \big) =
g(a,\mathbf{x}) , \\
\text{E} \big( A_{ij} =a \, \vert \, \mathbf{A}_{i(-j)} = \mathbf{a}', \mathbf{X}_{ij} = \mathbf{x}, \mathbf{X}_{ {i(-j)} } = \mathbf{x}' \big) = \pi( a \, \vert \, \mathbf{a}', \mathbf{x}, \mathbf{x}')
\end{array} \right.
\right\} \ .
\end{align*}
\item[\hypertarget{(M2)}{(M2)}] (\textit{Bounded Moments}) Let $\mathbf{g}^*(\mathbf{A}_i, \mathbf{X}_i) = (g^*(A_{i1}, \mathbf{X}_{i1}), \ldots, g^*(A_{in_i}, \mathbf{X}_{in_i}))^{\intercal}$.
For all $(\mathbf{A}_i,\mathbf{X}_i)$, $\text{E} \big\{ \big\| \mathbf{Y}_i - \mathbf{g}^*(\mathbf{A}_i, \mathbf{X}_i) \big\|_2^4 \, \vert \, \mathbf{A}_i, \mathbf{X}_i \big\}$ is bounded. Also, for any $\mathbf{X}_i$, there exists $\mathbf{A}_i$ such that the smallest eigenvalue of
$\text{E} \big[ \big\{ \mathbf{Y}_i - \mathbf{g}^*(\mathbf{A}_i, \mathbf{X}_i)\big\}^{\otimes 2} \, \vert \, \mathbf{A}_i, \mathbf{X}_i \big]$ is positive.
\end{itemize}
Then, the following properties hold for the proposed estimator $\widehat{\tau}$.
\begin{itemize}
\item[(1)] (\textit{Mean Double Robustness}) For any fixed outcome covariance model $\beta$, we have $\text{E} \big\{ \widehat{\tau}(\pi, g,\beta) \big\} = \tau^*$ as long as the conditional propensity score or the outcome model (but not necessarily both) is correct, i.e., $\pi=\pi^*$ or $g=g^*$.
\item[(2)] (\textit{Efficiency Gain Under Conditionally Independent Treatment Assignment}) Suppose the cluster-level propensity score is decomposable into the product of individual-level propensity scores, i.e., $P(\mathbf{A}_i \, \vert \, \mathbf{X}_i ) = \prod_{j = 1}^{n_i} P(A_{ij} \, \vert \, \mathbf{X}_{ij} )$. Then, we have $\text{Var} \big\{ \widehat{\tau}(\pi^*,g,\beta_{\pi^*,g}^* ) \big\} \leq \text{Var} \big\{ \overline{\tau}(e^*,g) \big\}$ for any outcome model $g$.
\end{itemize}
\end{lemma}
In words, part 1 of Lemma \ref{thm:2} states that the new estimator $\widehat{\tau}$ is doubly robust (in expectation) for any outcome covariance model $\beta$; that is, whatever the investigator specifies for the outcome covariance model $\beta$, $\widehat{\tau}$ will remain unbiased if either the conditional propensity score or the outcome regression, but not necessarily both, is correct. This property is very similar to a well-known property of the GEE estimator where, irrespectively of the specification of the weighting/covariance matrix, the GEE estimator remains consistent. Indeed, as we show in Section \ref{sec:properties}, with a few mild assumptions, notably on how the nuisance functions are nonparametrically estimated, we can also achieve the usual double robustness (in consistency) and critically, asymptotic Normality of $\widehat{\tau}$ for a working, but not necessarily ``correct'' outcome covariance model $\beta$.
Part 2 of Lemma \ref{thm:2} states that if the treatment assignment mechanism is independent conditional on $\mathbf{X}_i$ and does not depend on others' covariates, the new estimator $\widehat{\tau}$ is at least as efficient than the existing estimator $\overline{\tau}$. The conditional independence assumption is satisfied in most randomized experiments and some observational studies in multilevel settings where each individual is randomly assigned to treatment with equal probability or with probability that is a function of his/her own pre-treatment covariates. However, when the treatment assignment mechanism depends on peers' covariates or treatment, there is no general relationship between $\widehat{\tau}$ and $\overline{\tau}$ in terms of variance unless we make further assumptions about the correlation structure. This is, in part, because the variance term in the asymptotic expansion of $\widehat{\tau}$ may behave like a first-order bias term in clustered settings; see Section \ref{Supp-proof-thm:2}
of the Supplementary Materials for the exact details.
Regardless, even if the conditional independence assumption in part 2 of Lemma \ref{thm:2} is not satisfied and the relative efficiency between the two estimators cannot be theoretically characterized, subsequent sections numerically show that $\widehat{\tau}$ almost always have smaller variances than $\overline{\tau}$ or other existing estimators.
\subsection{Interpreting Assumption (M1)} \label{sec:M1}
We take a moment to discuss an important condition for our theoretical results, Assumption \hyperlink{(M1)}{(M1)}, and to explain that, despite its initial appearance, the assumption is not strong for estimating the our target estimand $\tau^*$.
At a high level, Assumption \hyperlink{(M1)}{(M1)} assumes that the cluster-level data $\mathbf{O}_i$ are i.i.d. from some distribution $P$, which belongs to a family of models $\mathcal{P}$ where the conditional expectation of the individual-level outcome and the individual-level treatment are the same for every unit $ij$. At first, this may seem like a strong assumption. Indeed, if the target quantity of interest is at the \emph{cluster-level}, say the average treatment effect for a cluster-level outcome and the treatment is assigned at the cluster-level, our work based on Assumption \hyperlink{(M1)}{(M1)} is a simple extension of the usual semiparametric theory in causal inference except its applied at the cluster-level; in fact, in this setting, the efficient influence function of the cluster-level estimand can be derived from existing work of \citet{Hahn1998}. However, our quantity of interest $\tau^*$ is at the \emph{individual-level}, specifically the average treatment effect of assigning treatment at the individual-level on an individual-level outcome, and Assumption \hyperlink{(M1)}{(M1)} is likely insufficient to derive the efficient influence function of $\tau^*$ because of the unknown correlation structure of $\mathbf{O}_{ij}$ within each cluster; see additional details below. To put it differently, the mismatch between our target estimand $\tau^*$, which is defined as a contrast at the \emph{individual-level}, and Assumption \hyperlink{(M1)}{(M1)} on the observed data, which is assumed to be i.i.d. at the \emph{cluster-level}, makes the problem more difficult than had we assumed that both the estimand and the data's i.i.d. assumption are at the same level (i.e., both individual-level or cluster-level).
Also, Assumption \hyperlink{(M1)}{(M1)} does not necessarily assume that all functionals of the individual-level observations $\mathbf{O}_{ij}$ are identical across study units; it only assumes that the individual outcome regression and the conditional propensity score model are the same. For example, it may be possible that the conditional variance of the outcome $\text{Var} ( Y_{ij} \, \vert \, \mathbf{A}_i, \mathbf{X}_i )$ may be different across study units where study unit $j=1$ in cluster $i$ may have $\text{Var} ( Y_{i1} \, \vert \, \mathbf{A}_{i}, \mathbf{X}_{i} ) = 1$ and study unit $j=2$ in cluster $i$ may have $\text{Var} ( Y_{i2} \, \vert \, \mathbf{A}_{i}, \mathbf{X}_{i} ) = A_{i2} + \| \mathbf{X}_i \|^2 + 1$. Similarly, the functional form of the usual, individual-level propensity score $e(a \, \vert \, \mathbf{X}_{ij})$ may actually be different across study units where $e(1 \, \vert \, \mathbf{X}_{ij}) = 0.5$ for study unit $j=1$ and $e(1 \, \vert \, \mathbf{X}_{ij}) = \text{expit} \big\{ 1 + 2 \big\| \mathbf{X}_{ij} \|^2 \big\}$ for study unit $j=2$.
Additionally, Assumption \hyperlink{(M1)}{(M1)} encompasses a large array of existing parametric, semiparametric, or nonparametric models that have been used to understand multilevel observational studies. The list below provides some common examples; all examples assume that the clusters are independent from each other and $n_i$ is bounded. Also, Section \ref{Supp-sec:modelM1}
of the Supplementary Materials provide formal proofs that these models satisfy \hyperlink{(M1)}{(M1)}.
\begin{enumerate}
\item \emph{GLMM and GEE models in Section \ref{sec:2-3}}: Under mild assumptions on the rank of $\mathbf{X}_{ij}$,
these existing parametric or semi-parametric models for multilevel studies are instances of the models allowed under Assumption \hyperlink{(M1)}{(M1)}.
\item \emph{Locally independent, nonparametric latent variable models of \citet{Lord1980} and \citet{Holland1986}}: Suppose there exists mutually independent latent variables $(\mathbf{U}_i, \mathbf{V}_i)$ that are also independent from pre-treatment covariates. These variables factorize the observed outcome and treatment variables as follows:
\begin{align} \label{eq-201}
&
P ( \mathbf{Y}_i \, \vert \, \mathbf{A}_i, \mathbf{X}_i, \mathbf{U}_i , \mathbf{V}_i )
=
\prod_{j=1}^{n_i} P(Y_{ij} \, \vert \, A_{ij} , \mathbf{X}_{ij} , \mathbf{U}_i ),
\nonumber
\\
&
P ( \mathbf{A}_i \, \vert \, \mathbf{X}_i, \mathbf{U}_i, \mathbf{V}_i )
=
\prod_{j=1}^{n_i} P(A_{ij} \, \vert \, \mathbf{X}_{ij} , \mathbf{V}_i ), \ \mathbf{X}_i \perp (\mathbf{U}_i, \mathbf{V}_i) \ .
\end{align}
\begin{figure}[!htb]
\centering
\begin{tikzpicture}[scale=0.7]
\tikzset{vertex/.style = {draw=none,fill=none,minimum size=1em}}
\tikzset{edge/.style = {->,> = latex'}}
\tikzset{edge2/.style = {-}}
\node[vertex] (v) at (0,0) {$\mathbf{V}_i$};
\node[vertex] (u) at (0,4) {$\mathbf{U}_i$};
\node[vertex] (x1) at (0,1) {$\mathbf{X}_{i1}$};
\node[vertex] (x2) at (0,3) {$\mathbf{X}_{i2}$};
\node[vertex] (a2) at (5,1) {$A_{i2}$};
\node[vertex] (a1) at (4,0) {$A_{i1}$};
\node[vertex] (y2) at (5,3) {$Y_{i2}$};
\node[vertex] (y1) at (4,4) {$Y_{i1}$};
\draw[edge2] (x1) to (x2);
\draw[edge] (x1) to (a1);
\draw[edge] (x2) to (a2);
\draw[edge] (x1) to (y1);
\draw[edge] (x2) to (y2);
\draw[edge] (a1) to (y1);
\draw[edge] (a2) to (y2);
\draw[edge] (v) to (a1);
\draw[edge] (v) to (a2);
\draw[edge] (u) to (y1);
\draw[edge] (u) to (y2);
\end{tikzpicture}
\caption{A graphical illustration of Assumption \protect\hyperlink{(M1)}{(M1)} under a cluster of size $n_i=2$ when there are latent variables $\mathbf{U}_i$ and $\mathbf{V}_i$.}
\label{Fig:DAG}
\end{figure}
Figure \ref{Fig:DAG} provide a visual illustration of equation \eqref{eq-201} when $n_i = 2$.
The latent variables $\mathbf{U}_i$ and $\mathbf{V}_i$ are not unmeasured confounders in that they do not simultaneously affect both the outcome and the treatment. Instead, the latent variables act similarly to random effects in GLMMs in that they are used to model dependence in the observed outcome or the treatment variables among study units in the same cluster. These models are also allowed under Assumption \hyperlink{(M1)}{(M1)}.
\item \emph{Nonparametric, conditionally i.i.d. models}: Conditional on the observed cluster-level variables $\mathbf{C}_i$, the individual-level observations $\mathbf{O}_{ij}$ become independent, i.e.,$P (\mathbf{O}_1, \ldots , \mathbf{O}_N ) = \prod_{i=1}^{N} P(\mathbf{C}_i) \prod_{j=1}^{n_i} P(\mathbf{O}_{ij} \, \vert \, \mathbf{C}_i)$. These conditionally independent nonparametric models are a subset of the models allowed under Assumption \hyperlink{(M1)}{(M1)} .
\item \emph{Nonparaemtric mixture models}: Suppose each cluster's data is generated from a mixture model of the form
\begin{align*}
P(\mathbf{O}_i) = \int P(\mathbf{O}_i \, \vert \, \bm{Z}_i = \bm{z} ) \, dP( \bm{z} ),
\end{align*}
where $\bm{Z}_i$ represents observed or unobserved mixture labels and the functional forms of $\text{E}(Y_{ij} \, \vert \, A_{ij} = a, \mathbf{X}_{ij} = \mathbf{x} , \bm{Z}_i = \bm{z} )$ and $\text{E} \big( A_{ij} =a \, \vert \, \mathbf{A}_{i(-j)} = \mathbf{a}', \mathbf{X}_{ij} = \mathbf{x}, \mathbf{X}_{ {i(-j)} } = \mathbf{x}', \bm{Z}_i = \bm{z} \big)$ are the same across all $(i,j)$.
We remark that \citet{Vaart1996} considered a variant of the above model where the density $P(\mathbf{O}_i \, \vert \, \bm{Z}_i = \bm{z} )$ was parametrized by a finite-dimensional parameter. Under additional assumptions on $\bm{Z}_i$, say the conditional distribution of $\bm{Z}_i \, \vert \, (h_1(\mathbf{O}_{ij}), h_2(\mathbf{O}_{ {i(-j)} }) )$ is identical across all $(i,j)$ for any fixed functions $h_1$ and $h_2$,
the mixture model satisfies \hyperlink{(M1)}{(M1)}; note that these are not the only assumptions on $\bm{Z}_i$ that would satisfy \hyperlink{(M1)}{(M1)} and see Section \ref{Supp-sec:modelM1}
of the Supplementary Materials for weaker conditions on $\bm{Z}_i$.
\end{enumerate}
Assumption \hyperlink{(M1)}{(M1)} assumes that the true data generating model can be any one of these models above, plus others that are not listed. Without knowing a priori which model is the correct data generating model or making more assumptions on top of Assumption \hyperlink{(M1)}{(M1)}, deriving the efficient influence function of $\tau^*$ may not be possible and the next best alternative is to develop an estimator that is at least as efficient as the existing ones.
Finally, Assumption \hyperlink{(M1)}{(M1)} bounds the cluster size $n_i$ and to the best of our knowledge, there is no established semiparametric efficiency theory that allows the dimension of $\mathbf{O}_i$ to grow to infinity while the elements of $\mathbf{O}_i$ remain dependent arbitrarily and asymptotically, i.e., the dependence does not vanish to zero as sample size increases. Instead, by bounding the cluster size, we allow for arbitrary dependence between units in a cluster. Practically, this implies that our estimator's theoretical guarantees in multilevel studies are realized when the cluster size is smaller than the number of clusters, i.e., $N \gg n_i$. But, in our numerical study in Section \ref{sec:varyingN}, we find that our theoretical guarantees remain plausible even if this is not the case, say if there are $N = 25$ clusters with each cluster having $n_i = 100$ study units.
\subsection{Nonparametric Estimation of Nuisance Functions $\pi, g$ and $\beta$} \label{sec:nfunc}
In this section, we discuss nonparametric estimation of the nuisance functions $\pi, g$ and $\beta$ inside the proposed estimator $\widehat{\tau}$ using existing, i.i.d.-based ML methods popular in causal inference. For estimating the conditional propensity score $\pi$ and the outcome regression $g$, we propose to use a slight modification of cross-fitting \citep{Victor2018} at the cluster-level where we randomly split the observed data into non-overlapping folds, say two folds $\mathcal{I}_1$ and $\mathcal{I}_2$, at the cluster level instead of at the individual-level; note that the original version of cross-fitting randomly splits the data at the individual-level and assumes that each study unit's data is i.i.d. One fold, referred to as the auxiliary data and denoted as $\mathcal{I}_k^c$, is used to estimate the nuisance functions and the other fold, referred to as the main data and denoted as $\mathcal{I}_k$, evaluates the estimated nuisance functions obtained from the first fold. We then switch the roll of the two folds to fully use the data for treatment effect estimation. We let $\widehat{\pi}^{(-k)}$ and $\widehat{g}^{(-k)}$ be the estimated conditional propensity score and the outcome regression based on the auxiliary data $\mathcal{I}_k^c$. We also let $\widehat{\pi}^{(-k)}( a \, \vert \, \mathbf{A}_{ {i(-j)} }, \mathbf{X}_{ij}, \mathbf{X}_{ {i(-j)} } )$ $(a=0,1)$ and $\widehat{g}^{(-k)} (A_{ij}, \mathbf{X}_{ij})$ be the evaluation of these functions using cluster $i$ in the main data $\mathcal{I}_k$.
There are some practical considerations to consider when estimating the conditional propensity score $\pi$ with existing i.i.d-based ML methods. Naively using the raw variables $(\mathbf{A}_ {i(-j)} , \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} )$ as independent variables in existing i.i.d.-based ML methods may not work well because each cluster may have different cluster size, leading to different number of independent variables. For example, without covariates, a cluster of size $2$ will have 1 independent variable while a cluster of size 10 will have 9 independent variables when we directly use $\mathbf{A}_ {i(-j)} $ as independent variables in a nonparametric regression estimator. Thankfully, there are many ways to resolve this issue in the literature. \citet{vvLaan2014}, \citet{Sofrygin2016}, and \citet{Ogburn2020arxiv} suggested making structural assumptions on $\pi$ where the investigator uses fixed, dimension-reducing summaries of peers' data $(\mathbf{A}_ {i(-j)} , \mathbf{X}_ {i(-j)} )$ as independent variables. Some popular examples of these dimension-reducing summaries include the proportion of treated peers, $\overline{A}_ {i(-j)} = \sum_{\ell \neq j} A_{i\ell}/(n_i-1)$, and the average of peers' covariates, $\overline{\mathbf{X}}_ {i(-j)} = \sum_{\ell \neq j} \mathbf{X}_{i\ell}/(n_i-1)$.
With the additional structural assumptions, estimating $\pi$ becomes a classification problem where $A_{ij}$ is the dependent variable and $(\overline{A}_ {i(-j)} , \mathbf{X}_{ij}, \overline{\mathbf{X}}_ {i(-j)} ) $ are the independent variables, respectively. Or, investigators can use additional summaries of peers' data as independent variables or more flexible models of dependence typical in network data \citep{Salathe2010, Keeling2011,Lorna2013, Imai2015}.
For estimating the outcome covariance model $\beta$, we only use the main sample from the cluster-level cross-fitting procedure above and solve the following optimization problem.
\begin{equation} \label{eq-genbeta}
\widehat{\beta}^{(k)} \in \argmin_{\beta \in \mathcal{B} } \frac{2}{N} \sum_{i \in \mathcal{I}_k}
\left[
\begin{array}{l}
w^2(\mathbf{C}_i) {\mathbf{I}}(\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} )^{\intercal} B \big( \beta(\bC_i) \big)
\\
\hspace*{1.5cm}
\times \widehat{\mathbf{S}}_{i}^{(-k)} B \big( \beta(\bC_i) \big) {\mathbf{I}}(\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} )^{\intercal}
\end{array}
\right]
\end{equation}
where $B \big( \beta(\bC_i) \big) = I_{n_i} - \beta(\mathbf{C}_i) \bm{1}_{n_i} \bm{1}_{n_i} ^{\intercal} , \quad{}
\widehat{\mathbf{S}}_{i}^{(-k)} = \big\{ \mathbf{Y}_i - \widehat{\mathbf{g}}^{(-k)}(\mathbf{A}_i, \mathbf{X}_i) \big\}^{\otimes 2}$ and
\begin{align*}
{\mathbf{I}} (\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)}) & =
\frac{1}{n_i}
\begin{bmatrix}
\frac{\mathbbm{1}(A_{i1} = 1)}{ \widehat{\pi}^{(-k)}(1 \, \vert \, \mathbf{A}_{i(-1)}, \mathbf{X}_{i1}, \mathbf{X}_{i(-1)}) }
-
\frac{\mathbbm{1}(A_{i1} = 0)}{ \widehat{\pi}^{(-k)} (0 \, \vert \, \mathbf{A}_{i(-1)}, \mathbf{X}_{i1}, \mathbf{X}_{i(-1)}) }
\\
\vdots
\\
\frac{\mathbbm{1}(A_{in_i} = 1)}{ \widehat{\pi}^{(-k)} (1 \, \vert \, \mathbf{A}_{i(-n_i)}, \mathbf{X}_{in_i}, \mathbf{X}_{i(-n_i)}) }
-
\frac{\mathbbm{1}(A_{in_i} = 0)}{ \widehat{\pi}^{(-k)} (0 \, \vert \, \mathbf{A}_{i(-n_i)}, \mathbf{X}_{in_i}, \mathbf{X}_{i(-n_i)}) }
\end{bmatrix} \ .
\end{align*}
The motivation for equation \eqref{eq-genbeta} comes from the sandwich variance/Huber-White variance estimator \citep{Huber1967, White1980} where $B \big( \beta(\bC_i) \big)$ and $ \widehat{\mathbf{S}}_{i}^{(-k)} $ correspond to the outer and inner part, respectively, of the sandwich estimator.
Also, from a theoretical perspective, the model space for $\beta$, denoted as $\mathcal{B}$ in \eqref{eq-genbeta}, only needs to be complete and $P$-Donsker with a finite envelope; see Theorem \ref{thm:3} for details. Some examples of complete and $P$-Donsker classes include a collection of bounded functions indexed by a finite-dimensional parameter in a compact set, a collection of bounded monotone functions, a collection of variationally bounded functions, $\alpha$-H\"older class with $\alpha> \text{dim}(\mathbf{C}_i)/2$, and Sobolev class; see Chapter 2 of \citet{VW1996}
for a textbook discussion about Donsker classes.
Theoretically speaking, it may be tempting to consider more complex models of $\beta$ beyond $P$-Donsker, say by using flexible ML methods. Indeed, if this is desired, we can use additional steps in the cluster-level cross-fitting procedure described above to estimate $\beta$ and remove the Donsker class assumption. However, in the process of developing the new estimator, we found that this strategy is only useful if the sample size is sufficiently large to estimate the underlying outcome covariance relationships. Also, since $\beta$ do not need to be ``correctly'' estimated to achieve consistency or asymptotic Normality of the proposed estimator (in contrast to the conditional propensity score or the outcome model; see below),
we find that using the $P$-Donsker class assumption is sensible for $\beta$.
In particular, in our numerical experiments, we find that the following, simple $P$-Donsker model space, denoted as $\mathcal{B}_\gamma$, leads to good performance in our empirical examples:
\begin{align} \label{eq-ParaB}
\mathcal{B}_\gamma
=
\bigg\{
\beta(\mathbf{C}_i) \, \bigg| \,
&
\beta(\mathbf{C}_i)
=
\sum_{\ell=1}^J \mathbbm{1}\{ L(\mathbf{C}_i) = \ell \} \gamma_\ell , \
\begin{array}{l}
\gamma_\ell \in [-B_0,B_0] \text{ for some } B_0>0 \ ,
\\
L: \mathbf{C}_i \rightarrow \{1,\ldots,J\}
\end{array}
\bigg\} \ .
\end{align}
Functions in $\mathcal{B}_\gamma$ are parametrized by a finite dimensional parameter $\bm{\gamma} = \big( \gamma_1,\ldots,\gamma_J \big)^{\intercal} \in [-B_0,B_0]^{\otimes J}$ where $\beta(\mathbf{C}_i) = \gamma_\ell$ if cluster $i$ belongs to stratum $\ell$ based on a user-specified label function $L$.
The label function $L$ can be a function about cluster size, say $L(\mathbf{C}_i) = \mathbbm{1}( n_i \leq M_1) + 2 \cdot \mathbbm{1}(M_1 < n_i \leq M_2) + 3 \cdot \mathbbm{1} (M_2 < n_i)$ with thresholds $M_1$ and $M_2$, or $L$ can be defined over a discrete covariate $C_i^{(d)}$, say $L( \mathbf{C}_i ) = \sum_{j=1}^J j \cdot \mathbbm{1}( C_i^{(d)} = d_j )$. Under $\mathcal{B}_\gamma$, estimation of $\beta(\mathbf{C}_i)$ is equivalent to solving the following estimating equation for $\bm{\gamma}$ for $\ell=1,\ldots,J$.
\begin{align*}
&
\hspace*{-0.05cm}
\sum_{i \in \mathcal{I}_k} \hspace*{-0.05cm} w(\mathbf{C}_i)^2 \mathbbm{1} \big\{ L(\mathbf{C}_i) = \ell \big\}
\Big[
2 {\mathbf{I}}(\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} ) ^{\intercal}
( I - \bm{1}\bm{1}^{\intercal} ) \widehat{\mathbf{S}}_i^{(-k)}
( I - \bm{1}\bm{1}^{\intercal} ) {\mathbf{I}}(\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} ) \widehat{\gamma}_\ell
\\
\nonumber
&
\hspace*{0.5cm}
-
{\mathbf{I}}(\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} ) ^{\intercal}
\Big\{ 2 \widehat{\mathbf{S}}_i^{(-k)} - \bm{1}\bm{1}^{\intercal} \widehat{\mathbf{S}}_i^{(-k)} - \widehat{\mathbf{S}}_i^{(-k)} \bm{1}\bm{1}^{\intercal} \Big\}
{\mathbf{I}}(\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} )
\Big]
= 0
\ .
\end{align*}
\noindent Once we have estimates of $\pi$, $g$, and $\beta$ from above, we can plug them inside $\widehat{\tau}$ to arrive at the final estimator of $\tau^*$, i.e.
\begin{align} \label{eq-tauhat}
&
\widehat{\tau}
=
\widehat{\tau} (\widehat{\pi}, \widehat{g}, \widehat{\beta}) =
\frac{1}{2} \sum_{k=1}^2 \frac{2}{N} \sum_{i \in \mathcal{I}_k}
\phi(\mathbf{O}_i, \widehat{\pi}^{(-k)}, \widehat{g}^{(-k)}, \widehat{\beta}^{(k)})
\ .
\end{align}
Section \ref{Supp-sec:practical}
of the Supplementary Materials provides additional implementation details and tips, especially for investigators who are more familiar with using parametric or semiparametric methods like GLMM and GEE. Much of these tips, including median adjustment to stabilize cross-fitted estimators, training ML models, and removing outlying clusters, already exist in the literature and the references in the Supplementary Materials contain additional details.
\subsection{Statistical Properties} \label{sec:properties}
To characterize the statistical properties of $\widehat{\tau}$, we make the following assumptions about the behavior of the estimated nuisance functions.
\begin{itemize}
\item[\hypertarget{(EN1)}{(EN1)}] (\textit{Estimated Conditional Propensity Score and Outcome Regression})
For all $k=1,2$, and $(\mathbf{A}_i,\mathbf{X}_i)$, we have\\
$\widehat{\pi}^{(-k)} \big( 1 \, \vert \, \mathbf{A}_ {i(-j)} , \mathbf{X}_{ij}, \mathbf{X}_{ {i(-j)} }) \in [c_\pi, 1-c_\pi]$, and $\big| \widehat{g}^{(-k)}(A_{ij}, \mathbf{X}_{ij}) - g^*(A_{ij}, \mathbf{X}_{ij}) \big| \leq c_g$ for some positive constants $c_\pi$ and $c_g$.
\item[\hypertarget{(EN2)}{(EN2)}] (\textit{Convergence Rate of Estimated Conditional Propensity Score and Outcome Regression})
We consider one of the following three conditions.
\begin{itemize}
\item[\hypertarget{(EN2.PS)}{(EN2.PS)}] We have $\big\| \widehat{\pi}^{(-k)} \big( 1 \, \vert \, \mathbf{A}_ {i(-j)} , \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} \big) - \pi^* \big( 1 \, \vert \, \mathbf{A}_ {i(-j)} , \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} \big) \big\|_{P,2} = O_P(r_{\pi,N})$ where $r_{\pi,N}$ is $o(1)$ and $
\big\| \widehat{g}^{(-k)}(A_{ij}, \mathbf{X}_{ij}) - g(A_{ij},\mathbf{X}_{ij}) \big\|_{P,2} = O_P(r_{g,N})$ where $g$ is some function satisfying $\big| g^*(A_{ij},\mathbf{X}_{ij}) - g(A_{ij},\mathbf{X}_{ij}) \big| \leq c_g $ for all $(A_{ij},\mathbf{X}_{ij})$ and $r_{g,N}$ is $o(1)$.
\item[\hypertarget{(EN2.OR)}{(EN2.OR)}] We have $\big\| \widehat{g}^{(-k)}(A_{ij}, \mathbf{X}_{ij}) - g^*(A_{ij},\mathbf{X}_{ij}) \big\|_{P,2} = O_P(r_{g,N})$ where $r_{g,N}$ is $o(1)$ and
$\big\| \widehat{\pi}^{(-k)} \big( 1 \, \vert \, \mathbf{A}_{ {i(-j)} }, \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} \big) - \pi \big( 1 \, \vert \, \mathbf{A}_{ {i(-j)} }, \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} \big) \big\|_{P,2} = O_P(r_{\pi,N})$ where $\pi$ is some function satisfying $\pi(1 \, \vert \, \mathbf{A}_{ {i(-j)} }, \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} ) \in [c_\pi,1-c_\pi]$ for all $(\mathbf{A}_{ {i(-j)} },\mathbf{X}_{i})$ and $r_{\pi,N}$ is $o(1)$.
\item[\hypertarget{(EN2.Both)}{(EN2.Both)}] We have $\big\| \widehat{\pi}^{(-k)} \big( 1 \, \vert \, \mathbf{A}_ {i(-j)} , \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} \big) - \pi^* \big( 1 \, \vert \, \mathbf{A}_ {i(-j)} , \mathbf{X}_{ij}, \mathbf{X}_ {i(-j)} \big) \big\|_{P,2} = O_P(r_{\pi,N})$ and $
\big\| \widehat{g}^{(-k)}(A_{ij}, \mathbf{X}_{ij}) - g^*(A_{ij},\mathbf{X}_{ij}) \big\|_{P,2} = O_P(r_{g,N})$
where $r_{\pi,N}$, $r_{g,N}$, and $N^{1/2}r_{\pi,N} r_{g,N}$ are $o(1)$.
\end{itemize}
\item[\hypertarget{(EN3)}{(EN3)}] (\textit{Estimated Outcome Covariance Model}) $\mathcal{B}$ is complete and $P$-Donsker with $\big\| \widehat{\beta}^{(k)}(\mathbf{C}_i) - \beta^{\dagger,(-k)} (\mathbf{C}_i) \big\|_{P,2} = O_P(r_{\beta,N})$ where $r_{\beta,N} = o(1)$ and
\begin{align} \label{eq-betadagger}
\beta^{\dagger,(-k)}
&
\in
\argmin_{\beta \in \mathcal{B} }
\text{E}
\Bigg[
\begin{array}{l}
w(\mathbf{C}_i)^2 {\mathbf{I}} (\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} ) ^{\intercal} B\big( \beta(\mathbf{C}_i) \big)
\\
\times
\big\{ \mathbf{Y}_i - \widehat{\mathbf{g}}^{(-k)} (\mathbf{A}_i, \mathbf{X}_i) \big\}^{\otimes 2}
B \big( \beta(\mathbf{C}_i) \big) {\mathbf{I}} (\mathbf{A}_i, \mathbf{X}_i , \widehat{\pi}^{(-k)} )
\end{array}
\, \Bigg| \,
\mathcal{I}_k^c
\Bigg] \ .
\end{align}
\item[\hypertarget{(EN4)}{(EN4)}] (\textit{Convergence Rate of Outcome Covariance Model}) We have $N^{1/2} r_{g,N} r_{\beta,N}$ is $o(1)$.
\end{itemize}
At a high level, Assumptions \hyperlink{(EN1)}{(EN1)} and \hyperlink{(EN2)}{(EN2)} are similar to assumptions about nuisance functions in modern, cross-fitted estimators of treatment effects in i.i.d. settings. In contrast, Assumption \hyperlink{(EN3)}{(EN3)} and \hyperlink{(EN4)}{(EN4)} impose conditions on the outcome covariance model
where the estimated outcome covariance model $\widehat{\beta}^{(k)}$ based on empirical risk minimization (i.e., equation \eqref{eq-genbeta}) converges to the minimizer of the population objective function (i.e., equation \eqref{eq-betadagger}), denoted as $\beta^{\dagger,(-k)}$. If we only desire consistency of $\widehat{\tau}$ without asymptotic Normality, we only need Assumption \hyperlink{(EN3)}{(EN3)}. On the contrary, if we desire asymptotic Normality of $\widehat{\tau}$, Assumption \hyperlink{(EN4)}{(EN4)} places additional restrictions on $r_{\beta,N}$ where the product of the convergence rate of $\widehat{\beta}^{(k)}$ and the outcome regression $\widehat{g}^{(-k)}$ is of order $N^{-1/2}$. Assumptions \hyperlink{(EN3)}{(EN3)} and \hyperlink{(EN4)}{(EN4)} are satisfied if investigators choose a sufficiently ``nice'' $\mathcal{B}$.
For example, we can choose the outcome covariance model as a singleton set, say $\mathcal{B} = \{0\}$, and Assumptions \hyperlink{(EN3)}{(EN3)} and \hyperlink{(EN4)}{(EN4)} will automatically hold with $r_{\beta,N}=0$. Also, our simple function class $\mathcal{B}_\gamma$ in equation \eqref{eq-ParaB} will satisfy these rate condition with $r_{\beta,N}=N^{-1/2}$; see Lemma \ref{Supp-lemma:consistency2}
in the Supplementary Material for a formal proof.
We make two remarks about the minimizer of the population objective function $\beta^{\dagger,(-k)}$. First, the population objective function fixes the nuisance functions $\widehat{\pi}^{(-k)}, \widehat{g}^{(-k)}$, because the outcome covariance model conditions on the estimated $(\widehat{\pi}^{(-k)}, \widehat{g}^{(-k)})$ from the auxiliary sample and uses the data from the main sample only to estimate $\widehat{\beta}^{(k)}$.
Second, $\beta^{\dagger,(-k)}$ does not have to be the ``true'' outcome covariance model; rather, $\beta^{\dagger,(-k)}$ is simply the best possible outcome covariance model when we solve the population objective function over some function class $\mathcal{B}$.
Theorem \ref{thm:3} formally characterizes the property of $\widehat{\tau}$.
\begin{theorem} \label{thm:3}
Suppose Assumptions \hyperlink{(A1)}{(A1)}-\hyperlink{(A3)}{(A3)}, \hyperlink{(M1)}{(M1)}-\hyperlink{(M2)}{(M2)}, \hyperlink{(EN1)}{(EN1)} and \hyperlink{(EN3)}{(EN3)} hold. Then the estimator $\widehat{\tau}$ in equation \eqref{eq-tauhat} satisfies the following properties.
\begin{itemize}
\item[(1)] (\textit{Double Robustness}) If either \hyperlink{(EN2.PS)}{(EN2.PS)} or \hyperlink{(EN2.OR)}{(EN2.OR)} holds, $\widehat{\tau}$ is consistent for the treatment effect $\tau^*$, i.e., $\widehat{\tau} = \tau^* + o_P(1)$.
\item[(2)] (\textit{Asymptotic Normality}) If \hyperlink{(EN2.Both)}{(EN2.Both)} and \hyperlink{(EN4)}{(EN4)} holds, $\widehat{\tau}$ has an asymptotically Normal distribution centered at the true treatment effect, i.e., $\sqrt{N} \big( \widehat{\tau} - \tau^* \big)$ weakly converges to \\
$ N \big( 0 , \text{Var} \big\{ \phi(\mathbf{O}_i, \pi^*, g^*, \beta^* ) \big\} \big)$ where
\begin{align*}
\beta^*
\in
\argmin_{\beta \in \mathcal{B} }
\text{E}
\left[
\begin{array}{l}
w(\mathbf{C}_i)^2 {\mathbf{I}} (\mathbf{A}_i, \mathbf{X}_i , \pi^*) ^{\intercal} B\big( \beta(\mathbf{C}_i) \big)
\\
\hspace*{1.5cm}
\times
\big\{ \mathbf{Y}_i - \mathbf{g}^*(\mathbf{A}_i, \mathbf{X}_i) \big\}^{\otimes 2}
B \big( \beta(\mathbf{C}_i) \big) {\mathbf{I}} (\mathbf{A}_i, \mathbf{X}_i , \pi^*)
\end{array}
\right]
\ .
\end{align*}
Also, a consistent estimator of the variance of $\widehat{\tau}$ is
\begin{align*}
&
\widehat{\sigma}^2
=
\frac{1}{N} \sum_{k=1}^2 \sum_{i \in \mathcal{I}_k}
\Big\{
\phi(\mathbf{O}_i, \widehat{\pi}^{(-k)}, \widehat{g}^{(-k)}, \widehat{\beta}^{(k)})- \widehat{\tau}_{k}
\Big\}^2
\ , \
&& \widehat{\tau}_{k} =
\frac{2}{N} \sum_{i \in \mathcal{I}_k}
\phi(\mathbf{O}_i, \widehat{\pi}^{(-k)}, \widehat{g}^{(-k)}, \widehat{\beta}^{(k)}) \ .
\end{align*}
\item[(3)] (\textit{Efficiency Gain Under Known Treatment Assignment Mechanism}) Suppose \hyperlink{(EN2.Both)}{(EN2.Both)} and \hyperlink{(EN4)}{(EN4)} hold. Additionally, if the same condition in part 2 of Lemma \ref{thm:2} concerning the treatment assignment mechanism holds, the asymptotic relative efficiency (ARE) between $\widehat{\tau}$ and $\overline{\tau}$ is ${\rm ARE}(\widehat{\tau},\overline{\tau}) = \text{Var}\big\{\varphi (\mathbf{O}_i, e^*, g^*) \big\} / \text{Var} \big\{ \phi(\mathbf{O}_i, \pi^*, g^*, \beta^*) \big\} \geq 1$.
\end{itemize}
\end{theorem}
Theorem \ref{thm:3} shows that the proposed estimator $\widehat{\tau}$ is consistent, doubly robust, and asymptotically Normal. Critically, we do not need our estimated outcome covariance model $\widehat{\beta}$ to converge to some underlying ``true'' model to guarantee double robustness, consistency, and asymptotic Normality of $\widehat{\tau}$; as mentioned earlier, this is similar to GEE estimators where any, arbitrary weighting matrix leads to consistent and asymptotically Normal estimates so long as the weights converges to some function. Also, similar to Lemma \ref{thm:2}, under some assumptions, the proposed estimator $\widehat{\tau}$ is theoretically guaranteed to be at least as efficient than the existing estimator $\overline{\tau}$.
\section{Simulation} \label{sec:sim}
\subsection{Finite-Sample Performance of $\widehat{\tau}$ and $\overline{\tau}$} \label{sec:varyingN}
We study the finite-sample performances of our proposed estimator $\widehat{\tau}$ and the existing estimator $\overline{\tau}$ through a simulation study. For pre-treatment covariates $\mathbf{X}_{ij}$, we use six variables $(W_{1ij}, W_{2ij}, W_{3ij}, C_{1i}, C_{2i}, n_i)$ where $W_{1ij}$, $W_{2ij}$, and $C_{1i}$ are from the standard Normal distribution and $W_{3ij}$ and $C_{2i}$ are from the Bernoulli distribution with mean parameter 0.3 and 0.7, respectively. The number of clusters varies from $N \in \{25,50,100,250,500\}$ and the cluster size takes on values between $2000/N$ and $3000/N$ so that the expected total number of individuals in the simulated dataset is 2500.
The treatment follows a mixed effects model with non-linear terms, i.e.,
\begin{align*}
&
\hspace*{-0.25cm}
P \big( A_{ij} = 1 \, \vert \, \mathbf{X}_{ij} , V_i)
=
\text{expit} \Bigg\{
\begin{array}{l}
-0.5 + 0.5 W_{1ij} - \mathbbm{1} \big( W_{2ij} > 1 \big) \quad
\\
\quad
+ 0.5 W_{3ij} - 0.25 C_{1i} + C_{2i} + V_i
\end{array}
\Bigg\},
\end{align*}
where $V_i \sim N(0,\sigma_V^2)$. When $\sigma_V=0$, the random effect $V_i$ in the propensity score model vanishes and each element of $\mathbf{A}_{i}$ are conditionally independent from each other given $\mathbf{X}_i$. But, when $\sigma_V \neq 0$, treatments in cluster $i$ are correlated with each other.
For the outcome, we use a mixed effects model with the following non-linear terms
\begin{align} \label{eq-Sim-OR}
&
Y_{ij} \, \vert \, (A_{ij}, \mathbf{X}_{ij}, U_i)
\sim
N \big(
3 + (2.1 + W_{2ij}^2 + 3W_{3ij} ) A_{ij}
+
2 W_{1ij} - C_{1i}^2 + W_{2ij} C_{2i} + U_i
\ , \ 1
\big),
\end{align}
where $U_i$ follows $N(0,\sigma_U^2)$. Similar to the treatment model, individuals' outcomes in cluster $i$ are correlated via $\sigma_U$.
The target estimand is the average treatment effect with $w(\mathbf{C}_i)=1$, which equals $4$ under the models above, i.e., $\tau^* = 4$.
We consider four different values of $(\sigma_V,\sigma_U)$: (i) no treatment correlation and weak outcome correlation where $(\sigma_V,\sigma_U)=(0,0.5)$;
(ii) no treatment correlation and strong outcome correlation where $(\sigma_V,\sigma_U)=(0,1.5)$;
(iii) strong treatment correlation and weak outcome correlation where $(\sigma_V,\sigma_U)=(1.5,0.5)$;
and (iv) strong treatment correlation and strong outcome correlation where $(\sigma_V,\sigma_U)=(1.5,1.5)$. To estimate the nuisance functions in both $\widehat{\tau}$ and $\overline{\tau}$, we use ensembles of multiple ML methods via the super learner algorithm \citep{SL2007, Polley2010}; see Section \ref{Supp-sec:Sim}
of the Supplementary Materials for the exact ML methods used in the super learner and other implementation details. For estimating the outcome covariance model $\beta$, we choose the function space $\mathcal{B}_\gamma$ in equation \eqref{eq-ParaB} where we use 1, 2, 3, 5, and 3 strata parameters (i.e., $\text{dim}(\bm{\gamma}) \in \{ 1,2,3,5,3 \}$) for cluster sizes $N = 25,50,100,250$, and $500$, respectively.
We repeat the simulation 200 times and report the empirical bias, empirical standard error, and the coverage rate of 95\% confidence intervals.
Also, to make the terms $\sigma_V$ and $\sigma_U$ governing the correlation structures more interpretable, we compute the intra-cluster correlation coefficient (ICC), a commonly used measure in multilevel studies to assess correlation between study units in the same cluster. For the outcome ICC, denoted as ${\rm ICC}_{Y}(\sigma_U)$, it varies from 0 to 0.8.
For the treatment ICC, denoted as ${\rm ICC}_A(\sigma_V)$, it varies varies from 0.05 to 0.36.
For the exact details on computing ICCs, especially for binary variables, see Section \ref{Supp-sec:Sim}
of the Supplementary Materials.
Table \ref{tab:0} summarizes the result. Broadly speaking, the proposed estimator $\widehat{\tau}$ achieve the smallest standard errors compared to the existing estimator $\overline{\tau}$ across all simulation scenarios, with $\overline{\tau}$ becoming less efficient under strong outcome correlation. Also, the coverage rates of confidence intervals based on $\widehat{\tau}$ are closer to the nominal coverage than those based on $\overline{\tau}$. Finally, even with small number of clusters where the bounded cluster condition in Assumption \hyperlink{(M1)}{(M1)} may be suspect, say $N = 25$ clusters with roughly 100 individuals per cluster, we see that $\widehat{\tau}$ has small bias and nominal coverage. Based on the result, we believe $\widehat{\tau}$ can be used in settings where each cluster has more study units compared to the total number of clusters, say if households are the study units and U.S. states are clusters.
\begin{table}[!htp]
\hspace*{-0.5cm}
\renewcommand{\arraystretch}{1.1} \centering
\footnotesize
\setlength{\tabcolsep}{2.5pt}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{Est.} & \multirow{2}{*}{$(\sigma_V,\sigma_U)$} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}($\text{ICC}_A$,\\ \ $\text{ICC}_Y$)\end{tabular}} & \multicolumn{3}{c|}{\begin{tabular}[c]{@{}c@{}}$N=25$\\ $\text{E}(n_i)=100$\end{tabular}} & \multicolumn{3}{c|}{\begin{tabular}[c]{@{}c@{}}$N=50$\\ $\text{E}(n_i)=50$\end{tabular}} & \multicolumn{3}{c|}{\begin{tabular}[c]{@{}c@{}}$N=100$\\ $\text{E}(n_i)=25$\end{tabular}} & \multicolumn{3}{c|}{\begin{tabular}[c]{@{}c@{}}$N=250$\\ $\text{E}(n_i)=10$\end{tabular}} & \multicolumn{3}{c|}{\begin{tabular}[c]{@{}c@{}}$N=500$\\ $\text{E}(n_i)=5$\end{tabular}} \\ \cline{4-18}
& & & Bias & SE & {\tiny Cover.} & Bias & SE & {\tiny Cover.} & Bias & SE & {\tiny Cover.} & Bias & SE & {\tiny Cover.} & Bias & SE & {\tiny Cover.} \\ \hline
\multirow{4}{*}{$\overline{\tau}$} & (0.0,0.5) & (0.05,0.20) & 0.40 & 8.50 & 0.995 & 0.39 & 7.48 & 0.965 & 0.18 & 7.52 & 0.940 & 0.98 & 6.53 & 0.960 & 0.19 & 6.54 & 0.955 \\ \cline{2-18}
& (0.0,1.5) & (0.05,0.69) & 1.72 & 27.10 & 0.990 & -0.48 & 12.07 & 0.985 & 1.01 & 12.71 & 0.950 & -0.17 & 10.69 & 0.965 & 1.00 & 11.00 & 0.950 \\ \cline{2-18}
& (1.5,0.5) & (0.27,0.20) & 7.37 & 88.19 & 0.990 & 3.12 & 31.34 & 0.980 & 1.35 & 16.92 & 0.975 & 0.69 & 8.59 & 0.975 & 0.48 & 7.11 & 0.960 \\ \cline{2-18}
& (1.5,1.5) & (0.27,0.69) & 7.64 & 222.3 & 0.990 & 7.31 & 85.20 & 0.965 & 0.14 & 45.70 & 0.945 & 1.00 & 19.79 & 0.970 & 0.75 & 13.42 & 0.950 \\ \hline
\multirow{4}{*}{$\widehat{\tau}$} & (0.0,0.5) & (0.05,0.20) & -0.43 & 6.22 & 0.920 & -0.12 & 6.13 & 0.965 & 0.18 & 6.65 & 0.920 & 0.55 & 6.25 & 0.935 & 0.20 & 6.20 & 0.950 \\ \cline{2-18}
& (0.0,1.5) & (0.05,0.69) & -0.08 & 6.58 & 0.945 & 0.01 & 6.21 & 0.955 & 0.07 & 6.19 & 0.960 & 0.27 & 6.59 & 0.970 & 1.05 & 7.58 & 0.950 \\ \cline{2-18}
& (1.5,0.5) & (0.27,0.20) & -0.38 & 6.63 & 0.930 & -0.15 & 7.38 & 0.905 & -0.29 & 7.15 & 0.925 & -0.17 & 7.11 & 0.920 & -0.06 & 6.14 & 0.980 \\ \cline{2-18}
& (1.5,1.5) & (0.27,0.69) & -0.51 & 7.09 & 0.940 & -0.26 & 7.10 & 0.935 & -0.08 & 7.09 & 0.955 & -0.09 & 7.23 & 0.950 & 0.84 & 7.79 & 0.965 \\ \hline
\end{tabular}
\caption{Finite-sample performance of $\widehat{\tau}$ and $\overline{\tau}$ under different number of clusters. Each row represents the values of $\sigma_V$ and $\sigma_U$. Each column shows the empirical biases, empirical standard errors, and coverages of 95\% confidence intervals. $N$ is number of clusters and the average number of individual per cluster (i.e., cluster size) is 2500/$N$. Bias and SE columns are scaled by 100.}
\label{tab:0}
\end{table}
\subsection{Comparison to Existing Nonparametric Methods} \label{sec:comparison}
Next, we compare the performance of $\widehat{\tau}$ to existing nonparametric estimators of treatment effects in the literature. The rationale behind this comparison is inspired by the discussion in \citet{Carvalho2019} where a practitioner may naively apply existing i.i.d.-based nonparametric methods in multilevel settings due to its simplicity. Comparing these estimators to the proposed estimator $\widehat{\tau}$, which is specifically tailored for multilevel settings, may shed some light on the cost of naively applying existing methods in practice.
For comparison, we use causal forests \citep{WA2018,grf} implemented in the \texttt{grf} R-package \citep{grfpackage}, and R/U-learners \citep{NieWager2020} with gradient boosting and lasso via the \texttt{rlearner} R package. Again, we remark that these competing ML methods were initially developed for i.i.d. data. But, \texttt{grf} provides some ways to accommodate clustering by using the options \texttt{weight.vector} and \texttt{clusters} options in the function \texttt{average\_treatment\_effect} \citep{grfvig}. We use these recommended tuning procedures for both causal forests and R/U-learners. The simulation model is the same as Section \ref{sec:varyingN} except we fix $N = 500$.
Table \ref{tab:1} summarizes the result.
Among competing ML methods initially developed for i.i.d. data, GRF with the recommended modifications for clustering seems to get close to the targeted nominal coverage, ranging from 91\% to 95.5\%. This suggests that the suggestions in \citet{grfvig} to tweak the original GRF to accommodate clustering has some merit and is worth theoretically investigating in the future. In contrast, R- and U-Learners' coverage rates are often much lower, ranging from 70\% to 92\%. Finally, our estimator $\widehat{\tau}$ maintain nominal coverage and achieves the smallest standard error among all estimators.
\begin{table}[!htp]
\renewcommand{\arraystretch}{1.1} \centering
\footnotesize
\setlength{\tabcolsep}{2pt}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$(\sigma_V,\sigma_U)$ & \multicolumn{3}{c|}{$(0.0,0.5)$} & \multicolumn{3}{c|}{$(0.0,1.5)$} & \multicolumn{3}{c|}{$(1.5,0.5)$} & \multicolumn{3}{c|}{$(1.5,1.5)$} \\ \hline
$({\rm ICC}_A,{\rm ICC}_Y)$ & \multicolumn{3}{c|}{$(0.05,0.20)$} & \multicolumn{3}{c|}{$(0.05,0.69)$} & \multicolumn{3}{c|}{$(0.27,0.20)$} & \multicolumn{3}{c|}{$(0.27,0.69)$} \\ \hline
Statistic & Bias & SE & Cover. & Bias & SE & Cover. & Bias & SE & Cover. & Bias & SE & Cover. \\ \hline
R-Learner-B & -1.61 & 7.25 & 0.875 & -1.86 & 9.90 & 0.920 & -5.30 & 7.12 & 0.805 & -9.88 & 12.43 & 0.725 \\ \hline
R-Learner-L & -11.45 & 9.62 & 0.750 & -11.43 & 12.78 & 0.780 & -8.88 & 10.67 & 0.885 & -9.30 & 15.09 & 0.885 \\ \hline
U-Learner-B & -3.28 & 7.37 & 0.860 & -3.66 & 9.96 & 0.900 & -6.38 & 7.31 & 0.755 & -11.13 & 12.59 & 0.700 \\ \hline
U-Learner-L & -11.40 & 9.44 & 0.810 & -11.41 & 12.79 & 0.800 & -8.85 & 10.74 & 0.900 & -9.30 & 15.11 & 0.885 \\ \hline
GRF & -0.10 & 7.58 & 0.930 & 0.81 & 11.18 & 0.915 & 0.07 & 7.61 & 0.955 & 0.18 & 12.97 & 0.910 \\ \hline
$\overline{\tau}$ & 0.19 & 6.54 & 0.955 & 1.00 & 11.00 & 0.950 & 0.48 & 7.11 & 0.960 & 0.75 & 13.42 & 0.950 \\ \hline
$\widehat{\tau}$ & 0.20 & \textbf{6.20} & 0.950 & 1.05 & \textbf{7.58} & 0.950 & -0.06 & \textbf{6.14} & 0.980 & 0.84 & \textbf{7.79} & 0.965 \\ \hline
\end{tabular}
\caption{Comparison of different estimators under varying correlation strength. Each row represents different estimators for the average treatment effect. Each column shows the empirical biases, empirical standard errors, and coverages of 95\% confidence intervals under different values of $\sigma_V$ and $\sigma_U$. Bias and SE columns are scaled by 100. The values in boldface are the smallest standard errors under each data generating process.}
\label{tab:1}
\end{table}
We extend the above comparisons by considering non-normal outcome regression random effect $U_i$. Specifically, we use the same outcome regression as equation \eqref{eq-Sim-OR} but the distribution of $U_i$ is generated from the mixture of the following four distributions: $N(0,0.5^2)$, $0.5 \cdot t(5)$, ${\rm Laplace}(0,0.5)$, and ${\rm Unif}(-0.25,0.25)$. The ANOVA-type estimate of the ICC of the outcome is 0.23. We keep the same propensity score model under $\sigma_V \in \{0,1.5\}$ and
Table \ref{tab:1-1} summarizes the result. Similar to Table \ref{tab:1}, under all settings, $\widehat{\tau}$ has the smallest standard error.
Moreover, $\widehat{\tau}$ achieves closer to nominal coverage compared to existing nonparametric estimators. Overall, the additional simulation supports the robustness and the efficiency gains of $\widehat{\tau}$ under different correlation structures among individuals.
\begin{table}[!htp]
\renewcommand{\arraystretch}{1.1} \centering
\footnotesize
\setlength{\tabcolsep}{3pt}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$\sigma_V$ & \multicolumn{3}{c|}{$0.0$} & \multicolumn{3}{c|}{$1.5$} \\ \hline
$({\rm ICC}_A,{\rm ICC}_Y)$ & \multicolumn{3}{c|}{$(0.05,0.23)$} & \multicolumn{3}{c|}{$(0.27,0.23)$} \\ \hline
Statistic & Bias & SE & Cover. & Bias & SE & Cover. \\ \hline
R-Learner-B & -2.01 & 6.86 & 0.910 & -5.56 & 8.40 & 0.745 \\ \hline
R-Learner-L & -11.50 & 10.00 & 0.755 & -8.97 & 11.75 & 0.850 \\ \hline
U-Learner-B & -3.82 & 7.02 & 0.855 & -6.54 & 8.47 & 0.730 \\ \hline
U-Learner-L & -11.41 & 10.01 & 0.795 & -8.91 & 11.84 & 0.845 \\ \hline
GRF & -0.41 & 7.80 & 0.910 & -0.09 & 8.69 & 0.905 \\ \hline
$\overline{\tau}$ & 0.04 & 6.37 & 0.965 & 0.47 & 7.61 & 0.965 \\ \hline
$\widehat{\tau}$ & -0.14 & \textbf{5.90} & 0.965 & -0.17 & \textbf{7.48} & 0.935 \\ \hline
\end{tabular}
\caption{Comparison of different estimators under non-normal $U_i$. Each row represents different estimators for the average treatment effect. Each column shows the empirical biases, empirical standard errors, and coverages of 95\% confidence intervals under different values of $\sigma_V$ and $\sigma_U$. Bias and SE columns are scaled by 100. The values in boldface are the smallest standard errors under each data generating process.}
\label{tab:1-1}
\end{table}
\section{Applications: Early Childhood Longitudinal Study} \label{sec:application}
We apply our method to estimate average treatment effects of center-based pre-school programs on children's reading score in kindergarten from the Early Childhood Longitudinal Study's Kindergarten Class of 1998-1999 (ECLS-K) dataset \citep{ECLSK2009}; note that a similar question with respect to children's math score was asked in \citet{Lee2021}. Briefly, the dataset consists of children followed from kindergarten through eighth grade in the United States. Children define the individual study units, kindergarten classrooms define clusters, and the subscript ${ij}$ indicates the $j$th child in the $i$th kindergarten. We let $A_{ij}=1$ if the child received center-based care before kindergarten and $A_{ij}=0$ otherwise (i.e., parental care); note that the treatment was assigned at the individual-level. The outcome $Y_{ij} \in [22,94]$ is the standardized reading score of each child, measured during the Fall semester and after treatment assignment. For cluster-level pre-treatment covariates, we include cluster size, region (northeast/midwest/south/west), kindergarten location (central city/urban/rural), and kindergarten type (public/private). For individual-level pre-treatment covariates, we include the child's gender, age, race, family type, parental education, and economic status.
We restrict our analysis
to children with complete data on the outcome, treatment, and pre-treatment covariates.
This results in 15,980 children in 942 kindergartens, which corresponds to 16.96 children per kindergarten. Cluster size varies between 1 and 25 children; the first, second, and third quartiles of cluster size are 14, 19, and 21, respectively;
see Section \ref{Supp-sec:Supp-ECLSK}
of the Supplementary Materials for additional details.
To assess the outcome and the treatment correlations, we compute the ICCs as follows. For the outcome ICC, we fit a linear mixed effects regression model where $Y_{ij}$ is the independent variable, $(A_{ij}, \mathbf{X}_{ij})$ are the dependent variables, and kindergarten classrooms serve as random effects. For the treatment ICC, we fit a logistic mixed effect regression model where $A_{ij}$ is the independent variable, $ \mathbf{X}_{ij}$ are the dependent variables, and kindergarten classrooms serve as random effects. Using these models, we find that the ICCs of the outcome and the treatment variables are 0.103 and 0.063, respectively, suggesting a moderate amount of correlation/dependencies between children in the same kindergarten classroom.
We focus on $\tau_{\text{ATE}}^*$, the overall average treatment effect with $w(\mathbf{C}_i) = 1$ (i.e., the average treatment effect across all clusters), and $\{ \tau_1^*,\ldots,\tau_{12}^*\}$, 12 weighted average treatment effects defined by region and kindergarten location with $w_1(\mathbf{C}_i) = \mathbbm{1} \{$kindergarten $i$ is in a central city in northeast region$\}$, $\ldots$, $w_{12}(\mathbf{C}_i) = \mathbbm{1} \{$kindergarten $i$ is a rural area in west region$\}$. The choice of the 12 weighted average treatment effects are based on the previous works \citep{ECLSK1, ECLSK2} indicating that educational attainment varies by region and location. We compute
the proposed estimator $\widehat{\tau}$ using the same procedure in Section \ref{sec:sim} except we repeat cross-fitting 100 times
and use 24 strata parameters (i.e., $\text{dim}(\bm{\gamma})=24$) to estimate the outcome covariance model $\beta$ in $\mathcal{B}_\gamma$. We remark that by Theorem \ref{thm:3}, as long as either the outcome model or the conditional propensity score is correctly specified, $\widehat{\tau}$ is consistent, irrespectively of the estimated outcome covariance model. Also, if the convergence rates are sufficiently fast, $\widehat{\tau}$ is asymptotically Normal.
Figure \ref{fig:1} summarizes the result where the 12 weighted average treatment effects are divided by the empirical proportion of each subgroup, i.e., $\{ \sum_{i=1}^N w_t(\mathbf{C}_i) / N \}^{-1} \widehat{\tau}_t $ for $t=1,\ldots,12$. As a consequence, the latter 12 estimates in the figure present the estimated subgroup average treatment effects, i.e. $\text{E} \{ \pot{ Y_{ij} }{1} - \pot{ Y_{ij} }{0} \, \vert \, w_t(\mathbf{C}_i) = 1 \}$; note that since the empirical proportions $\{ \sum_{i=1}^N w_t(\mathbf{C}_i) / N \}$ converge to the true proportions of $P(w_t(\mathbf{C}_i) = 1)$, by Slutsky's theorem, the estimator for the subgroup average treatment effect enjoys the same properties as the estimator for $\tau^*$. For comparison, we use the causal forest-based estimator and the existing estimator $\overline{\tau}$ in Section \ref{sec:comparison}. First, all three estimators show the similar overall effects around 1.6 and indicate that the effect is significant at level 0.05. Second, the subgroup average treatment effects seem to be heterogeneous across census blocks and locations, and the significance of the effect estimates is different depending on the estimators. In particular, for the subgroup average treatment effect among city schools in the northeast census region, the causal forest-based estimator and $\overline{\tau}$ show negative, but insignificant at level 0.05, effect estimates whereas the proposed estimator $\widehat{\tau}$ show a positive, significant at level 0.05 effect estimate. Similarly, the effect estimates in rural northeast, south, and midwest regions based on the causal forest-based estimator and $\overline{\tau}$ are insignificant at level 0.05 in whereas those based on the proposed estimator $\widehat{\tau}$ are significant at level 0.05. Lastly, the proposed estimator $\widehat{\tau}$ always yield smaller variances compared to the competing estimators. Specifically, our
proposed estimator $\widehat{\tau}$ yields 6.6 to 54.9 percent shorter confidence intervals and has 12.8 and 79.7 percent smaller variance compared to the other estimators.
\begin{figure}[!htb]
\centering
\includegraphics[width=1\textwidth]{Subgroup_Effects.png}
\caption{Summary of the data analysis from the ECLS-K dataset.
Each column shows the overall average treatment effect (labeled as overall) and the subgroup average treatment effect (labeled as Central city\_NE, $\ldots$, Urban\_MW) where NE and MW are shorthand for northeast and midwest, respectively. The black ($\bullet$), red ({\color{red}$\bullet$}), and blue ({\color{myblue}$\bullet$}) dots show the point estimates of the causal forest-based estimator (denoted as GRF), the existing doubly robust estimator (denoted as $\overline{\tau}$), and our proposed estimator (denoted as $\widehat{\tau}$), respectively. The vertical lines show 95\% confidence intervals.
}
\label{fig:1}
\end{figure}
\section{Conclusion} \label{sec:Conclude}
This paper presents a new estimator $\widehat{\tau}$ to infer treatment effects in multilevel studies. The new estimator is doubly robust, asymptotically Normal, and does not require parametric specifications of the outcome model, the treatment model, and the correlation structure. The new estimator $\widehat{\tau}$ leverages recent ML-based methods in causal inference in i.i.d. settings to estimate, among other things, two non-standard functionals in causal inference. These two functionals are designed to capture the correlation between the treatment and the outcome among study units in the same cluster. Also, the new estimator $\widehat{\tau}$ can be more efficient than existing estimators in multilevel studies; we present both theoretical conditions that guarantee efficiency improvements and numerical studies showing $\widehat{\tau}$ performing better than existing, nonparametric estimators.
We end by re-iterating some important limitations of the new estimator $\widehat{\tau}$ and, in light of these limitations, offer some guidelines on how to use $\widehat{\tau}$ in practice. First, while $\widehat{\tau}$ performs well in data and all the simulation studies we considered, we do not show whether $\widehat{\tau}$ is optimal in terms of semiparametric efficiency. As hinted in Sections \ref{sec:M1} and \ref{sec:nfunc},
in nonparametric, multilevel settings, the semiparametric efficiency bound of $\tau^*$ depends on both the smoothness of the conditional mean of the outcome as well as the covariance of the outcomes in a cluster as each part can have non-diminishing first-order terms in a locally linear, asymptotic expansion of the estimator around $\tau^*$. In light of this, we conjecture that $\widehat{\tau}$ can achieve the semiparametric efficiency bound under more stronger assumptions about the exchangeability of the elements of the outcome covariance matrix. Second, while our theory currently relies on the bounded cluster size assumption, as demonstrated in Section \ref{sec:varyingN}, our methods seem to perform well even if the cluster size is larger than the number of clusters. Of course, we do not expect this phenomena to hold in settings where the cluster is very large and there are a couple number of clusters.
Despite these limitations, we believe the proposed estimator provides a new approach to robustly and efficiently estimate treatment effects in multilevel settings where the study units exhibit dependence within clusters and can be a valuable tool for practitioners working with multilevel observational studies.
\newpage
|
\section{Introduction}
In 5G and beyond, wireless networks will be densified with multiple access points of different types \cite{1},\cite{2}. The access points need to be connected to the operators’ core network via a transport network. On a global scale, fiber and microwave technology are dominant backhauling techniques. Fiber is a reliable link immune to interference and environmental effects with high peak data rate. However, the installation/maintenance cost of fiber may be high and it may not be attainable to deploy it everywhere. Due to the local geometry and features in some locations, fiber installations may not be feasible. Also, there are some governing policies in certain locations that may not allow installing new infrastructures for the use of optical fiber \cite{2}.
Wireless backhaul, on the other hand, is a scalable and economical backhaul option that can meet the increasing requirements of 5G systems \cite{2}, although it is sensitive to, e.g., blockage, tree foliage, rain, and supports lower peak data rates compared to fiber. For this reason, microwave is a backhaul technology used by most mobile operators worldwide, and the trend is likely to continue in the future. Typical wireless backhaul links are designed for point-to-point communications at 10-80 GHz, with strong line-of-sight (LoS) signal components. Also, even though there are few microwave communication standards, the existing wireless backhaul technologies are mainly based on non-standardized solutions.
With 5G and beyond, the access links will operate in millimeter wave (mmWave) spectrum, the range which was previously used for backhauling. Thus, there may be a conflict of interest between the access and backhaul links, which requires standardization. On the other hand, with low-height access points installed on, e.g., lamp posts, there is a probability for blockage, and we also need to support non-line-of-sight (NLoS) communication in the backhaul links. These are the main motivations for the current integrated access and backhaul (IAB) networks \cite{2}. With IAB, the objective is to provide flexible wireless backhauling using 3rd generation partnership project (3GPP) new radio (NR) technology, and provide not only the existing cellular services but also backhaul in the same node and via the same hardware \cite{2},\cite{3}.
The performance of IAB networks have been studied in different works. In particular,\cite{2} provides the basics for IAB network architecture and studies the service coverage in downlink communication. Then, \cite{3} evaluates the coverage extension improvement in 28 GHz band with IAB deployment in 3GPP urban micro scenarios. Also, \cite{4} investigates the number of IAB nodes that are required for 5G IAB deployment.
Then, \cite{5} investigates the power allocation problem for a proposed in-band self-backhaul scheme using an iterative algorithm. Moreover, \cite{6} investigates power control for moving networks in mmWave based wireless backhaul, and \cite{7} proposes dynamic power control to improve NLOS transmission performance. Additionally, \cite{8} proposes an uplink power control scheme based on machine learning in 5G networks for near-optimum performance in terms of transmit power, data rate, and network energy. In \cite{9}, the IAB networks are studied in an end-to-end manner, in line with 3GPP Release 16 (Rel-16).
Interestingly, \cite{10} formulates a multi-hop scheduling problem to offer an efficient IAB network deployment. In \cite{11}, a genetic algorithm (GA) formulation is developed for both IAB node and non-IAB backhaul link distribution.
Finally, \cite{12} investigates the potentials and challenges of mobile IAB, and \cite{13} develops joint scheduling and rate allocation for maximizing the throughput.
In this paper, we study the effect of power allocation in the uplink performance of IAB networks. Considering mmWave channel characteristics and the power ranges agreed in 3GPP, we develop a GA-based power allocation scheme maximizing the network coverage probability. Here, combined with resource allocation, the UEs and the IAB nodes transmit powers are jointly optimized using GA. Also, we investigate the effect of the interference on the network performance. Moreover, our simulations verify the effect of separate access and backhaul transmission, compared to the cases with simultaneous access and backhaul transmission. Finally, we verify the effect of different parameters such as the minimum data rate requirement, the cell size and the transmit power on the service coverage probability.
Our simulations show that, with a power allocation scheme, the network coverage probability is improved, compared to the cases with non-optimized power allocation. For instance, consider a two-hop IAB network operating at 28 GHz and 400 MHz channel bandwidth. Then, with a service coverage probability of 70\% and typical parameter settings, the implementation of power control leads to a minimum of 5 dB SNR gain, compared to the cases with non-optimized power allocation. Moreover, we have considered and evaluated a case with dedicated slots for backhaul transmission and compared it with a case of having simultaneous access and backhaul transmission. We observed that having dedicated slots for backhaul transmission offers a higher service coverage probability. Finally, for a broad range of parameter settings and mmWave transmission, the effect of interference on the service coverage probability may be negligible, if the network deployment is properly planned.
\section{System Model}
IAB network consists of two types of nodes \cite{2},\cite{4}:
\begin{itemize}
\item IAB donor, consisting of the central unit (CU) and distributed unit (DU) which serves the UEs as well as the other IAB nodes. IAB donor is connected to the core network via a non-IAB, e.g., fiber, backhaul link.
\item IAB nodes, consisting of the DU and mobile termination (MT) units which serves the UEs and, possibly, other IAB nodes in the chain of multi-hop communications. The IAB nodes rely on IAB for backhauling.
\end{itemize}
The motivation for the CU/DU split in IAB donor, as initially suggested in 3GPP Rel. 15 for next generation NodeBs (gNBs), is that time-critical functionalities, such as scheduling, and fast retransmission, can be realized in the DU close to the radio and the antenna, while the less time-critical radio functionalities are centralized in the CU. In Rel. 16 IAB, both out-of-band and in-band backhauling are supported in which the access and backhaul operate in different and the same frequency bands, respectively. In-band backhauling gives the flexibility in resource allocation between the access and backhaul, at the cost of complexity/coordination \cite{14}. In this paper, we concentrate on in-band backhauling.
In an IAB multi-hop chain, the parent node connects to the downstream UEs and IAB nodes via the IAB-DU. The IAB-MT is the module connecting an IAB node to its parent IAB-DU. From many aspects, IAB-MT part of a node behaves like a UE in the sense that it connects to the parent IAB-DU like a UE.
On the other hand, from the UE perspective, the IAB-DU of a node appears as a normal DU.
\begin{figure}[h]
\vspace{-4mm}
\centering
\includegraphics[width=0.53\textwidth, height= 9cm]{figure/Slide1p2.png}
\vspace{-7mm}
\caption{Schematic diagram of the system model of an IAB network.
}
\label{fig.iab_networks}
\vspace{-1mm}
\end{figure}
Figure \ref{fig.iab_networks} shows our considered system model, assuming an outdoor two-hop IAB network. This is motivated by the fact that, as reported by e.g., \cite{2}, \cite{11}, although 3GPP does not limit the possible number of hops, in practice traffic aggregation in the backhaul links and latency become challenging as the number of hops increases.
The system model supports \textit{E} number of UEs that are randomly distributed within the specified coverage area of radius \textit{r}. The UEs are randomly distributed using the finite homogeneous Poisson point process (FHPPP) approach \cite{4}. Also, there are \textit{M} IAB nodes in fixed locations within the coverage area.
Two specific network arrangements are considered: (i) one IAB donor associated with a finite number of stationary IAB nodes, and (ii) two IAB donors associated with a finite number of stationary IAB nodes each, in adjacent macro cells, in order to investigate the effect of inter-cell interference. The wireless channel is modeled to include the effects of shadowing, interference, pathloss, fading and rainfall.
The following are the general assumptions made in this paper:
\begin{itemize}
\item A central time division duplex (TDD) scheduler governs the communication of the network nodes and UEs.
\item The transmit power of the UEs and IABs is taken as the effective isotropic radiated power (EIRP) as suggested by 3GPP \cite{15}.
\item Periodic access to channel state information (CSI) is available at the IAB donor, IAB node and UE.
\end{itemize}
\subsection{Channel Model}
The received power at either the IAB donor or IAB node is modelled as
\begin{equation}
P_r = P_t + G_t + G_r - L - \sigma - Y_R - \phi.
\label{eq.3.pr}
\end{equation}
Here, $P_t$ is the transmit power, $G_t$ is the gain of the transmitter, $G_r$ is the gain of the receiver, $L$ is pathloss, $\sigma$ is shadowing loss, $Y_R$ is rain loss, and $\phi$ is the channel fading effect, which is modeled as a Rayleigh flat fading. All values are in dB. In our work, the transmit power values used are EIRP defined in \cite{15}, where $P_\text{EIRP}$ is expressed as
\begin{equation}
P_{\text{EIRP}} = P_t + G_t.
\label{eq.3.eirp}
\end{equation}
A 3GPP urban macro (UMa) model is selected for the pathloss and shadowing \cite{16}. The model factors in the height of the IAB donor, IAB nodes and the UEs.
The UMa pathloss model is represented by
\begin{equation}
\begin{split}
L = 32.4 + 10 \alpha \log_{10}(d_\text{{3D}}) + 20{\log_{10}{f_c}}\\
- 10({(d^{'}_{\text{BP}})^2} + (h_{\text{BS}} - h_{\text{UE}})^2).
\label{eq.3.pathloss}
\end{split}
\end{equation}
Here, $\alpha$ is the pathloss exponent.
${d}_{3D}$ is the 3D distance calculated using trigonometric equation, which is the LoS distance from the top of every UE to the top of the base station (BS), i.e, IAB donor or IAB node. 2D distance is the horizontal distance from the BS to a UE. Also, $f_c$ is the carrier frequency, in GHz. The pathloss exponent is related to the signal blockage, either in LoS or NLoS use cases. Also, $h_{\text{BS}}$ is the height of the BS, and $h_{\text{UE}}$ is the maximum height of the UE, and $d^{'}_{\text{BP}}$ is the break point distance which is determined by the relationship
\begin{equation}
d^{'}_{\text{BP}} = \frac{4 * h^{'}_{\text{BS}}* h^{'}_{\text{UT}}*{f}_c}{ {c}},
\label{eq.3.breakpoint}
\end{equation}
where, $h^{'}_\text{BS}$ is the effective antenna heights of the BSs, $h^{'}_\text{UT}$ is the effective antenna heights of the UE, and $c$ is the speed of light.
In mmWave communication, the rain loss may not be negligible, depending on the frequency, distance and rain intensity. The International Telecommunications Union Radio communication (ITU-R) section has a model for corresponding signal attenuation for a given rain rate and operating frequency band \cite{17}. The system model is built on rain rate of between 15 mm/h and 20 mm/h. The ITU-R power-law relationship model is expressed as
\begin{equation}
Y_R = k * R^{\mathit{\Gamma}}.
\label{eq.3.rain}
\end{equation}
Here, $Y_R$ is the signal specific attenuation expressed in dB/km, while $R$ is the given rain rate in mm/h. Then, $k$ and $\mathit{\Gamma}$ are polarization coefficients that are determined based on the operating frequency band \cite{15}. Finally, using similar terminology as in \eqref{eq.3.pr}, the interference is expressed as
\begin{equation}
I_i = \sum_{\forall j\ne i} P_{\text{EIRP}_j} + G_{r_j} - L_j - \sigma_j - Y_{R_j} - \phi_j.
\label{eq.3.interferance}
\end{equation}
In this way, using \eqref{eq.3.pr}-\eqref{eq.3.interferance}, the received signal-to-interference-plus-noise ratio (SINR), $\gamma$, at an IAB node or IAB donor is given by
\begin{equation}
\text {$\gamma$} = \frac{P_r}{{I}+{ N_0}},
\label{eq.3.SINR}
\end{equation}
where, $N_0$ is the modeled channel noise.
Then, the achievable rate, in bits per second, is obtained by
\begin{equation}
R_b = \text{BW} * \log_{2}(1 + \text{$\gamma$}),
\label{eq.3.shanon}
\end{equation}
where BW is the UE channel bandwidth, in Hz, either to the IAB donor or IAB node. The same concept holds for every IAB node associated with an IAB donor, where one can derive the achievable data rate of the IAB-IAB donor link similar to \eqref{eq.3.shanon}. Moreover, the backhaul data rate of an IAB node is a summation of the data from the UEs that are associated with it, thereby guaranteeing a successful communication of the UEs to the IAB donor.
Then, given a target data rate, $R_b$, the minimum required SINR in a link is found as
\begin{equation}
{\gamma}_{\text{min}} = 2^\frac{R_b}{\text BW} - 1.
\label{eq.3.SINR.min}
\end{equation}
Our metric of interest is the service coverage probability. Here, with a two-hop setup, a UE can either connect directly to the IAB donor, or its message is forwarded to the donor IAB via an intermediate IAB node. Then, a transmission fails and a UE is out of coverage if its message can be transferred to the donor IAB in none of these paths, which can be found out via the comparison of the SINR of the links and the minimum required SINR.
\section{Uplink Power Optimization}
In this section, we formulate the optimization problem regarding the uplink power control in IAB networks. The problem formulation is based on determining appropriate transmit powers for the UEs and associated IAB nodes within a range to meet a pre-defined service requirement determined by a baseline data rate and a corresponding SINR requirement.
To reduce the optimization complexity, we consider a sequential procedure for resource association and power allocation. Firstly, we start with the UE and IAB node association rule. Then, we optimize the power allocation based on the considered node association.
Motivated by the 3GPP Rel. 17 discussions on interference management, we evaluate the system performance in two distinct cases where either the access and backhaul transmissions are separated in different time slots, or they can be performed in the same slot. Also, 3GPP considers different ranges of possible transmit powers for the UEs and BSs \cite{15}.
We apply GA for power control. Here, the GA takes the CSI and the baseline service requirements to implement power allocation for all UEs and IAB nodes, within the specified transmit power range, such that the service requirements of the nodes are satisfied.
In words, the proposed GA follows the following procedure. Let us denote the number of iterations by $N$, determined by the designer.
First, $K$ possible random solutions are selected. This set is created by containing randomly selected transmit powers, with each value coming from the pre-defined range of transmit power values supported by the UE and IAB node. For each set of the $K$ possible solutions, the SINR value of the links and the network coverage probability are calculated. The best solution, leading to maximum coverage probability, which we refer to as Queen in the following, is determined and kept for the next generation. Keeping the Queen for the next generation guarantees the continuous improvement in different iterations of the GA.
In the next step, $S$ $<<$ $K$ sets of transmit power values are generated around the Queen, by making few mutations on it, i.e., changing few elements inside the Queen, such that the UEs/IABs transmit powers remain within their acceptable range. Then, $V=K-S-1$ random set of solutions are generated. The algorithm iterates $N$ times and the final Queen is returned as the best solution maximising the coverage probability.
As an advantage, the proposed algorithm is generic in the sense that it can be applied in different channel models and UEs/IABs transmit power ranges. Moreover, as we show in the following, the proposed GA converges with a few iterations. This is important because, it is straightforward to show that the considered power allocation is an NP-hard problem with no closed-form or easy-to-search solution. As opposed, the proposed GA only requires $NK$ solution checkings to reach a (semi) optimal solution.
\vspace{-1mm}
\begin{algorithm}
\caption{GA-Transmit Power Control Algorithm}\label{alg:euclid}
For each instance, with a set of UEs and IAB nodes, do the followings:
\begin{algorithmic}[1]
\STATE Create $K$, e.g., $K=10$, sets of transmit powers, randomly selected from the range between the minimum and maximum transmit powers allowed for the UEs and the IAB nodes. Each element in the set corresponds to either a UE or an IAB node transmit power.
\STATE For each set, calculate the received power, SINR and coverage probability.
\STATE Find the set with the highest coverage probability, and call it as the Queen. The Queen is kept for the next generation.
\STATE Create $S{\ll}K$, e.g., $S=5$, sets of transmit powers around the Queen. These sets are created by small modifications in the Queen such that the UEs/IABs transmit powers remain within their possible ranges.
\STATE Create $V = K - S -1$ sets of random transmit powers, with each element in the set related to the allowed transmit power range for the UE and the IAB nodes.
\STATE Go to back to \textbf {Step} \textbf 2 and continue for $N$ iterations.
\STATE Finally, return the Queen as the optimal power solution which has highest coverage probability after $N$ iterations.
\end{algorithmic}
\end{algorithm}
\section{Simulation Results}
The model is deployed for an urban macro environment with 3GPP and ITU-R parameters detailed in Table \ref{Parameters} \footnote{The parameter settings are not necessarily aligned with the Ericsson parameters of interest.}.
The simulation results are presented in different parts as follows.
\begin{table}[h!]
\vspace{-2mm}
\centering
\small\addtolength{\tabcolsep}{-4pt}
\caption{Simulation Parameters.}
\label{Parameters}
\begin{tabular}{|l|r|l|r|}
\hline
\textbf{Parameter} & \textbf{Value} & \textbf{Parameter} & \textbf{Value} \\ \hline
$f_c$ & 28 GHz & Cell radius ($r$) & 200 m \\ \hline
$BW$ & 400 MHz & Shadowing ($\sigma$) & 4 dB \\ \hline
Sub-carrier spacing & 120 kHz & Pathloss Exp.($\alpha$) & 4 \\ \hline
Min. RBs & 24 & Eff. Ant. Height & 1 m \\ \hline
Max. RBs & 270 & Ref. distance ($d_0$) & 1 m \\ \hline
Thermal noise ($T_{0}$) & -174 + BW (dB) & Num. IAB nodes & 4 per cell \\ \hline
${\text{UE}}_{\text{Noise figure}}$ ($N_f$) & 5 dB & Min. Data Rate & 64 kbps \cite{18} \\ \hline
Noise ($N_0$) & $T_0$ + $N_f$ (dB) & UE EIRP & \{23 - 43\} dBm \\ \hline
${\text{Height}}_{\text{IAB Donor}}$ & 25 m & IAB node EIRP & \{35 - 53\} dBm \\ \hline
${\text{Height}}_{\text{IAB Node}}$ & \{21 - 24\} m & Receiver gain & 25 dB\cite{18} \\ \hline
UE height & 1.5 m & Rain rate ($R$) & \{15 - 20\} mm/h \\ \hline
\end{tabular}
\end{table}
\begin{figure}[h]
\vspace{-1mm}
\centering
\includegraphics[width=0.5\textwidth, height= 6cm]{figure/GA_coverage.png}
\vspace{-7mm}
\caption{Examples of the convergence of the GA.}
\label{fig.gaconverge}
\vspace{-2mm}
\end{figure}
Figure \ref{fig.gaconverge} shows different examples of the convergence of the GA.
Here, we set $N$ = 200, $K$ = 20 and $S$ = 10. As seen in Fig. \ref{fig.gaconverge}, the algorithm converges in a ladder form. The observed ladder form is because the GA may not necessarily find the best solution in every iteration, and it may be trapped for a while in a local minimum. However, due to Step 5 of the Algorithm, GA can always avoid a local minimal and reach the global optimum if sufficiently large number of iterations is considered \cite{19}, \cite{20}. In GA, the larger the number of iterations, the more accurate the final solutions but at a cost of the running time. Then, as seen in Fig. \ref{fig.gaconverge} and our various non-included simulations, in different channel realizations the proposed GA converges with a few iterations. This reduces the optimal complexity, compared to, e.g., exhaustive search based solutions, significantly.
In Fig. \ref{fig.SC2COMB}, we investigate the number of UEs that can be served within the cell, with and without uplink power optimization. We also investigate the effect of increasing the resource blocks (RBs) per UE when the access and backhaul links are operating at different times. Here, we consider two adjacent cells, with a fixed cell radius.
\begin{figure}[h]
\vspace{-3mm}
\centering
\includegraphics[width=0.5\textwidth, height=6cm]{figure/SC2COMBn.png}
\vspace{-7mm}
\caption{Service coverage probability versus the number of UEs with 2 and 4 RBs per UE.}
\label{fig.SC2COMB}
\vspace{-2mm}
\end{figure}
From Fig. \ref{fig.SC2COMB}, it is observed that an optimal power allocation makes it possible to serve considerably higher number of UEs. For instance, consider the parameter settings of Fig. \ref{fig.SC2COMB} and service coverage probability of $70\%$. Then, with a minimum data rate 1 Mbps and 2 RBs per UE, the number of supported UEs increases from 7 with non-optimized power allocation to 19 UEs in the cases with optimal power control. This is indeed at the cost of dynamic coordination at the IAB donor, which may not be possible in practice. The relative gain of power control increases with the number of RBs per UE. Also, when the number of concurrent UEs within the cell increases, the service coverage probability decreases.
This may be attributed to the increase in the interference within the cell since the number of transmitters, i.e., the UEs, increases.
\begin{figure}[h]
\vspace{-3mm}
\centering
\includegraphics[width=0.5\textwidth, height=6cm]{figure/SINR_S2andS4_C4.png}
\vspace{-7mm}
\caption{Service coverage probability versus the SINR with 4 RBs per UE.}
\label{fig.sinrdifftdd}
\vspace{-2mm}
\end{figure}
In Fig. \ref{fig.sinrdifftdd}, we study the system performance in the cases with simultaneous and separated operation of the access and backhaul links in the uplink slots. Here, the results are presented for both cases with non-optimized and optimized power allocation.
Figure \ref{fig.sinrdifftdd} shows that, when access and backhaul are separated, with service coverage probability $70\%$, there is a minimum of 5 dB gain in SINR after the implementation of power optimization, compared to non-optimized power allocation.
Furthermore, when the access and backhaul links are working simultaneously, there is a decrease in the SINR.
With the parameter setting of Fig. \ref{fig.sinrdifftdd} and service coverage probability 70\%, the simultaneous operation of the access and backhaul in the uplink slots reduces the SINR, compared to the cases with separated access and backhaul operation, by 1 dB. Here, it should be noted that although separating the access and backhaul in the uplink slots saves the uplink signals of the IAB-connected UEs from the high transmit power of the IAB-MTs, still the neighbour non-IAB networks will suffer from the IAB-MTs high transmit powers.
\begin{figure}[h]
\vspace{-3mm}
\centering
\includegraphics[width=0.5\textwidth,height=6cm]{figure/intercell.png}
\vspace{-7mm}
\caption{Inter-cell interference. This figure shows a comparison between service coverage with one cell in comparison to the cases with two-cell set-up.}
\label{fig.intercell}
\vspace{-2mm}
\end{figure}
Figure \ref{fig.intercell} shows the effect of inter-cell interference on the number of UEs that can be served in the cell, with and without uplink power optimization. The figure compares service coverage probability of a single-cell set-up with two adjacent cell set-up. There is not much difference in the results from the two considered scenarios. Notably with a sufficient cell radius, coming from, e.g., network planning, the effect of inter-cell interference is minimal. This is because at 28 GHz, the signal strength decreases rapidly with the distance, whereby the effect of inter-cell interference is minimal if the distance between the cells is sufficiently large.
\begin{figure}[h]
\vspace{-3mm}
\centering
\includegraphics[width=0.5\textwidth, height=5.4cm]{figure/cdf_combine.png}
\vspace{-7mm}
\caption{CDF for UEs and IABs transmit powers.}
\label{fig.cdfsbs}
\vspace{-2mm}
\end{figure}
Figure \ref{fig.cdfsbs} studies the cumulative distribution function (CDF) of the UEs and the IAB nodes transmit powers, for different data rates. The results are presented for the cases with the UEs and IAB nodes transmit powers optimized via the GA. Also, the transmit powers are optimized within the range specified by 3GPP \cite{15}.
According to the figure, the required transmit power increases slightly with the data rate. However, there is relatively the same transmit power distribution of all the UEs irrespective of their associated BS, either IAB donor or IAB nodes.
Also, the transmit power variation of the UEs increases as the data rate increases. .
On the other hand, the transmit power variation of the IABs is low for the considered set of UEs data rates, as the IABs use almost the same transmit power. That is,
compared to the IABs, the UEs transmit power is more sensitive to the data rate.
\section{Conclusion}
We studied the effect of uplink power control on the service coverage probability of two-hop IAB networks. We developed a GA-based scheme for power control, and studied the effect of different parameters on the network performance.
Our results show that compared to the cases with non-optimized power allocation, our proposed sequential node association and power control algorithm improves the energy efficiency and the service coverage probability of IAB networks. Also, for different network configurations, the proposed GA converges with a few iterations, which reduces the optimization complexity compared to, e.g., exhaustive search based schemes, significantly.
Moreover, separating the IAB and UEs transmissions in the uplink slots may give the chance to improve the service coverage probability, as the IAB donor-connected UEs signals are not affected by the IABs high transmission powers. Finally, with proper network planning and mmWave communications, the effect of interference on the network performance is not considerable.
\section{Acknowledgement}
This work was supported in part by VINNOVA (Swedish Government Agency for Innovation Systems) within the VINN Excellence Center ChaseOn and in part by the European Commission through the H2020 project Hexa-X (Grant Agreement no. 101015956).
|
\subsection*{Massive scalar}
Equation of motion
\begin{equation}
(\mathfrak{D}^2 -2) \phi = 0.
\end{equation}
\subsection*{Massive spin-1/2}
Equations of motion
\begin{align}
&i \mathfrak{D}_n (\bar{\sigma}^n \xi_1)^{\dot{\alpha} } + \sqrt{2} \bar{\psi}^{\dot{\alpha} }_1 = 0, \\
&i \mathfrak{D}_n (\sigma^n \bar{\xi}_2)_{\alpha} + \sqrt{2} \psi_{2 \alpha}= 0, \\
&i \mathfrak{D}_n (\sigma^n \bar{\psi}_1)_{\alpha} + \sqrt{2} \xi_{1 \alpha} = 0, \\
& i \mathfrak{D}_n (\bar{\sigma}^n \psi_2)^{\dot{\alpha} } + \sqrt{2} \bar{\xi}^{\dot{\alpha} }_2 =0.
\end{align}
\subsection*{Massive vectors}
Equation of motion
\begin{equation}
(\mathfrak{D}^2 -2)C_m + 2i \varepsilon_{mn} C^n = 0.
\end{equation}
Lorenz condition
\begin{equation}
\mathfrak{D}^m C_m = 0.
\end{equation}
The same expressions hold for $a_m$.
\subsection*{Massive spin-3/2}
Equations of motion
\begin{align}
&i\mathfrak{D}_n (\bar{\sigma}^n \chi_{1m})^{\dot{\alpha} } + \sqrt{2} \bar{\lambda}_{1m}^{\dot{\alpha} } =0, \\
&i \mathfrak{D}_n (\sigma^n \bar{\chi}_{2m})_{\alpha} + \sqrt{2} \lambda_{2 \alpha m} = 0, \\
&i\mathfrak{D}_n (\sigma^n \bar{\lambda}_{1m})_{\alpha} + \sqrt{2} \chi_{1\alpha m} - i \sqrt{2} \varepsilon_{mn} \chi^n_{1\alpha} = 0, \\
& i\mathfrak{D}_{n} ( \bar{\sigma}^n \lambda_{2m})^{\dot{\alpha} } + \sqrt{2} \bar{\chi}_{2m}^{\dot{\alpha} } - i \sqrt{2} \varepsilon_{mn} \bar{\chi}^{\dot{\alpha} n}_2 = 0.
\end{align}
Lorenz gauge conditions
\begin{align}
&\mathfrak{D}^m \chi_{1 \alpha m} - 2 \varepsilon_{rs} (\sigma^{rs} \xi_1)_{\alpha} = 0, \\
& \mathfrak{D}^m \bar{\chi}^{\dot{\alpha} }_{2m} + 2 \varepsilon_{rs} (\bar{\sigma}^{rs} \bar{\xi}_2)^{\dot{\alpha} } = 0, \\
& \mathfrak{D}^m \bar{\lambda}^{\dot{\alpha} }_{1m} -i \frac{\sqrt{2}}{2} \tilde{\varepsilon}^{mn} (\bar{\sigma}_n \chi_{1m})^{\dot{\alpha} } + 2 \varepsilon_{rs} (\bar{\sigma}^{rs} \bar{\psi}_1)^{\dot{\alpha} } - 2 \sqrt{2} i \varepsilon^{mn} \mathfrak{D}_m (\bar{\sigma}_n \xi_1)^{\dot{\alpha} } = 0, \\
& \mathfrak{D}^m \lambda_{2 \alpha m} + i \frac{\sqrt{2}}{2} \tilde{\varepsilon}^{mn} (\sigma_n \bar{\chi}_{2m})_{\alpha} - 2 \varepsilon_{rs} (\sigma^{rs} \psi_2)_{\alpha} + 2 \sqrt{2} i \varepsilon^{mn} \mathfrak{D}_m (\sigma_n \bar{\xi}_2)_{\alpha} = 0.
\end{align}
Sigma trace constraints
\begin{align}
&(\bar{\sigma}^m \chi_{1m})^{\dot{\alpha} } + 4 \sqrt{2} \bar{\psi}_1^{\dot{\alpha} } = 0, \\
& (\sigma^m \bar{\chi}_{2m})_{\alpha} - 4 \sqrt{2} \psi_{2 \alpha} = 0, \\
&(\sigma^m \bar{\lambda}_{1m})_{\alpha} - 4 \sqrt{2} \xi_{1 \alpha} - 2 \sqrt{2} i \varepsilon_{rs} (\sigma^{rs} \xi_1)_{\alpha} = 0, \\
&(\bar{\sigma}^m \lambda_{2m})^{\dot{\alpha} } + 4 \sqrt{2} \bar{\xi}_2^{\dot{\alpha} } + 2 \sqrt{2} i \varepsilon_{rs} (\bar{\sigma}^{rs} \bar{\xi}_2)^{\dot{\alpha} } = 0.
\end{align}
\subsection*{Massive spin-2}
Equation of motion
\begin{equation}
( \mathfrak{D}^2 -2) v_{(mp)} + 2i \Big( \varepsilon_m^{\;\;n} v_{(np)} - v_{(mn)}\varepsilon^n_{\;\;p}\Big) = 0.
\end{equation}
Lorenz condition
\begin{equation}
\mathfrak{D}^m v_{(mn)} = - 2i \tilde{\varepsilon}_{nm} C^m - 2i \tilde{\varepsilon}_{nm} a^m -2i \varepsilon_{nm} \mathfrak{D}^m \phi.
\end{equation}
Other constraints for $v_{mn}$
\begin{align}
&v^n_{\;\; n} + 8 \phi = 0, \\
&v_{[mn]} = 2 \epsilon_{mnpq} \Big( \mathfrak{D}^p C^q + \mathfrak{D}^p a^q \Big) + 2i \varepsilon_{mn} \phi.
\end{align}
\newsec{Conclusions}[conclu]
In this work, we first showed how to couple the hybrid formalism for the open superstring to a constant electromagnetic background and derived expressions for the worldsheet variables in terms of the oscillator modes. We then computed the open superstring field theory action and equations of motion for the compactification-independent massless sector of the charged open superstring in a constant background, and demonstrated that the result is consistent with previous calculations from bosonic string theory. Finally, we constructed the open superstring field theory action for the first massive compactification-independent states of the charged open superstring in a constant background and showed that the equations of motion in Lorenz gauge describe 12 complex bosonic and 12 complex fermionic propagating degrees of freedom.
The most obvious application of our results is to try to relate our action with alternative actions for a charged massive spin-$3/2$ field in four dimensions. Although we have shown that the massive spin-$3/2$ field in open superstring field theory propagates causally, it is coupled in four dimensions through the electromagnetic background with a massive spin-$1/2$ field. It is unclear if our action can be generalized in a manner to decouple these massive spin-$3/2$ and spin-$1/2$ fields, except in $d=10$ where there is no massive spin-$1/2$ field in the spectrum of the uncompactified superstring.
To compute the equations of motion of the massive component fields, we used Lorenz gauge to simplify the computations. In principle, one can explicitly expand the action of \actionmassiveback in terms of the component fields and obtain a gauge-invariant description of the equations of motion. Although it might be complicated, this gauge-invariant component action could be useful for looking for generalizations which describe the massive spin-$2$ and massive spin-$3/2$ fields without coupling to massive fields of lower spin. For example, in $d=10$, one expects to be able to describe both the spin-$3/2$ and spin-$2$ fields without coupling to fields of lower spin. It would be interesting to see how this is accomplished, and to compare with the bosonic string field theory action of Argyres-Nappi \cite{Argyres:1989cu} in which the massive spin-$2$ field is described in $d=26$ without massive fields of lower spin.
\section*{\centering Acknowledgements}
KB would like to thank ICTP-SAIFR for hospitality at the start of this project. His work is supported in part by the Agence Nationale de Recherche under grant ANR-15-CE31-0002 ``HiggsAutomator''. NB would like to acknowledge partial financial support from CNPq grant number 311434/2020-7 and FAPESP grant numbers 2016/01343-7, 2019/24277-8 and 2019/21281-4. CAD acknowledges FAPESP grant number 2020/10183-9 for financial support. ML acknowledges FAPESP grant numbers 2019/17805-8 and 2016/16824-0 for financial support. We would also like to thank Massimo Porrati for useful discussions.
\newpage
\begin{appendices}
\newsec{Results in Spinor and Oscillator Algebra} \label{spinorconventions}
Some relations and definitions about the supersymmetric modes that were used in the calculations of this work are given below.
For a more complete list of identities and conventions for the sigma matrices and spinorial indices, see the appendices of \cite{Wess:1992cp}.
\begin{equation}
(\chi^{\alpha} \psi_{\alpha}) = (\chi \psi), \quad (\bar{\chi}_{\dot{\alpha} } \bar{\psi}^{\dot{\alpha} }) = (\bar{\chi} \bar{\psi}), \quad (\chi \psi)^{\dagger} = (\bar{\chi} \bar{\psi}).
\end{equation}
\begin{equation}
\theta_0^{\alpha}\theta_{\beta 0}= \frac{1}{2} \delta^{\alpha}_{\beta}(\theta_0 \theta_0), \quad \bar{\theta}_{\dot{\alpha} 0} \bar{\theta}_0^{\dot{\beta}}= \frac{1}{2} \delta^{\dot{\beta}}_{\dot{\alpha} } (\bar{\theta}_0\bar{\theta}_0).
\end{equation}
\begin{align}
&[p_{\alpha 0}, (\theta_0 \theta_0)] = -2 \theta_{\alpha 0}, \quad [\bar{p}_{\dot{\alpha} 0}, (\bar{\theta}_0 \bar{\theta}_0)]= -2 \bar{\theta}_{\dot{\alpha} 0}, \\
& [(p_0 p_0),(\theta_0 \theta_0)]=-4 -4(\theta_0 p_0), \quad [(\bar{p}_0 \bar{p}_0), (\bar{\theta}_0 \bar{\theta}_0)] = -4 -4 (\bar{\theta}_0 \bar{p}_0).
\end{align}
\begin{align}
\alpha_{0n} \alpha_{0p} (\bar{\sigma}^n \sigma^p)^{\dot{\alpha} }_{\;\; \dot{\beta}} & = -\alpha_{0n} \alpha^n_0 \delta^{\dot{\alpha} }_{\dot{\beta}} - i \varepsilon_{np} (\bar{\sigma}^{np})^{\dot{\alpha} }_{\;\;\dot{\beta}}, \\
\alpha_{0n} \alpha_{0p} (\sigma^n \bar{\sigma}^p)_{\alpha}^{\;\; \beta} & = -\alpha_{0n} \alpha^n_0 \delta_{\alpha}^{\beta} - i \varepsilon_{np} (\sigma^{np})_{\alpha}^{\;\; \beta}.
\end{align}
\begin{equation}
[d^{\alpha}_n,\Pi_{\alpha \dot{\alpha} m}] = 8i \partial \bar{\theta}_{\dot{\alpha} m+n}, \quad [\bar{d}^{\dot{\alpha} }_n, \Pi_{\alpha \dot{\alpha} m}]= -8i \partial \theta_{\alpha m+n}.
\end{equation}
\begin{equation}
\begin{split}
[d_{\alpha m}, \bar{d}_{\dot{\alpha} n} \bar{d}^{\dot{\alpha} }_n] &= 4i \Pi_{\alpha \dot{\alpha} m+n} \bar{d}^{\dot{\alpha} }_n + 16 \partial \theta_{\alpha m+2n} \\
& =4i \bar{d}^{\dot{\alpha} }_n \Pi_{\alpha \dot{\alpha} m+n} - 16 \partial \theta_{\alpha m+2n}
\end{split}
\quad
\begin{split}
[\bar{d}_{\dot{\alpha} m}, d^{\alpha}_n d_{\alpha n}]&= -4i \Pi_{\alpha \dot{\alpha} m+ n} d^{\alpha}_n + 16 \partial \bar{\theta}_{\dot{\alpha} m+2n} \\
&= - 4i d^{\alpha}_n \Pi_{\alpha \dot{\alpha} m+n} - 16 \partial \bar{\theta}_{\dot{\alpha} m+2n}.
\end{split}
\end{equation}
\begin{equation}
d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 = -\bar{d}^{\dot{\alpha} }_0 \Pi_{\alpha \dot{\alpha} 0} d^{\alpha}_0 -4i\Pi^n_0 \Pi_{n 0} +8i \partial \theta^{\alpha}_0 d_{\alpha 0} + 8i \partial \bar{\theta}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0.
\end{equation}
\begin{equation}
\begin{split}
[d^{\alpha}_0 d_{\alpha 0}, \bar{d}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0] &= -8i\bar{d}^{\dot{\alpha} }_0 \Pi_{\alpha \dot{\alpha} 0} d^{\alpha}_0 + 16 \Pi^n_0 \Pi_{n 0} -32 \partial \bar{\theta}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 -32 \partial \theta^{\alpha}_0 d_{\alpha 0} \\
& = 8i d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 -16 \Pi^n_0 \Pi_{n0} + 32 \partial \bar{\theta}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 + 32 \partial \theta^{\alpha}_0 d_{\alpha 0} \\
& = 4i \Pi_{\alpha \dot{\alpha} 0} [d^{\alpha}_0, \bar{d}^{\dot{\alpha} }_0] + 32 d^{\alpha}_0 \partial \theta_{\alpha 0} -32 \partial \bar{\theta}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0.
\end{split}
\end{equation}
\begin{align}
&d_{\alpha 0} f(\theta_0, \bar{\theta}_0)= \big[- p_{\alpha 0} - (\sigma^m \bar{\theta}_0)_{\alpha} \alpha_{0m} - \frac{i}{2} (\bar{\theta}_0 \bar{\theta}_0) (\varepsilon \cdot \sigma)_{\alpha}^{\; \beta} \theta_{\beta 0}\big] f(\theta_0, \bar{\theta}_0), \\
&\bar{d}_{ \dot{\alpha} 0} f(\theta_0, \bar{\theta}_0)= \big[ -\bar{p}_{\dot{\alpha} 0} + (\theta_0 \sigma^m)_{\dot{\alpha} } \alpha_{0m} + \frac{i}{2} (\theta_0 \theta_0) (\varepsilon \cdot \bar{\sigma})^{\dot{\beta}}_{\; \dot{\alpha} } \bar{\theta}_{\dot{\beta} 0}\big]f(\theta_0, \bar{\theta}_0), \\
&\Pi^m_0f(\theta_0, \bar{\theta}_0) = \big[- i \alpha_0^m + \frac{i}{2} \epsilon^{mrsn} \varepsilon_{rs} (\theta_0 \sigma_n \bar{\theta}_0)\big]f(\theta_0, \bar{\theta}_0).
\end{align}
\begin{align}
&d^{\alpha}_0 d_{\alpha 0} f(\theta_0, \bar{\theta}_0)=\big[ p_0^2 -2 (\bar{\theta}_0 \bar{\sigma}^m)^{\alpha} \alpha_{0m} p_{\alpha 0} + (\bar{\theta}_0 \bar{\theta}_0) \alpha^n_0 \alpha_{0n} -i (\bar{\theta}_0 \bar{\theta}_0) (\varepsilon \cdot \sigma)_{\beta}^{\;\alpha} \theta_0^{\beta} p_{\alpha 0}\big]f(\theta_0, \bar{\theta}_0), \\
&\bar{d}_{\dot{\alpha} 0}\bar{d}^{\dot{\alpha} }_0 f(\theta_0, \bar{\theta}_0)= \big[\bar{p}_0^2 -2 (\bar{\sigma}^m \theta_0)^{\dot{\alpha} } \alpha_{0m} \bar{p}_{\dot{\alpha} 0} + (\theta_0 \theta_0) \alpha^n_0 \alpha_{0n} - i (\theta_0 \theta_0) (\varepsilon \cdot \bar{\sigma})^{\dot{\beta}}_{\;\dot{\alpha} } \bar{\theta}_{\dot{\beta} 0} \bar{p}_0^{\dot{\alpha} }\big]f(\theta_0, \bar{\theta}_0), \\
& \Pi^m_0 \Pi_{m0} f(\theta_0, \bar{\theta}_0)=\big[- \alpha^m_0 \alpha_{0m} + \alpha_{0m} \epsilon^{mrsn} \varepsilon_{rs} (\theta_0 \sigma_n \bar{\theta}_0) - \frac{1}{2} \varepsilon_{rs} \varepsilon^{rs} (\theta_0 \theta_0) (\bar{\theta}_0 \bar{\theta}_0)\big]f(\theta_0, \bar{\theta}_0).
\end{align}
\begin{equation}
\{ d_{\alpha 0}, \partial \theta^{\beta}_0 \} = \frac{i}{2} (\varepsilon \cdot \sigma)_{\alpha}^{\; \beta}, \quad \{ \bar{d}^{\dot{\alpha} }_0, \partial \bar{\theta}_{\dot{\beta} 0} \} = \frac{i}{2} (\varepsilon \cdot \bar{\sigma})^{\dot{\alpha} }_{\;\dot{\beta}}.
\end{equation}
\begin{equation}
[d_0^2, \partial \theta^{\beta}_0] = i (\varepsilon \cdot \sigma)_{\alpha}^{\;\beta} d_0^{\alpha}, \quad [\bar{d}_0^2, \partial \bar{\theta}_{\dot{\beta} 0}] = i (\varepsilon \cdot \bar{\sigma})^{\dot{\alpha} }_{\; \dot{\beta}} \bar{d}_{\dot{\alpha} 0}.
\end{equation}
\begin{equation}
\label{Deltas}
\Delta_{\alpha}^{\; \beta} = \delta^{\beta}_{\alpha} + \frac{i}{2} (\varepsilon \cdot \sigma)^{\;\beta}_{\alpha}, \quad \bar{\Delta}^{\dot{\alpha} }_{\;\dot{\beta}}= \delta^{\dot{\alpha} }_{\dot{\beta}} + \frac{i}{2} (\varepsilon \cdot \bar{\sigma})^{\dot{\alpha} }_{\;\dot{\beta}}.
\end{equation}
\begin{align}
&(-d_0^2 \bar{d}_1^2 - d_1^2 \bar{d}_0^2 + d_2^2 \bar{d}_{-1}^2) \varphi = (-\bar{d}_1^2 d_0^2 + \bar{d}_2^2 d_{-1}^2 - \bar{d}_0^2 d_1^2) \varphi, \\
& (3 d_0^2 \bar{d}_1^2 -3 d_1^2 \bar{d}_0^2 + d_2^2 \bar{d}_{-1}^2) \varphi = (-3 \bar{d}_0^2 d_1^2 + 3 \bar{d}_1^2 d_0^2 - \bar{d}_2^2 d_{-1}^2) \varphi, \\
& (d_{-1}^2 \bar{d}_1^2 -2 d_0^2 \bar{d}_0^2 + d_1^2 \bar{d}_{-1}^2) \varphi = (\bar{d}_{-1}^2 d_1^2 + \bar{d}_1^2 d_{-1}^2 -2 \bar{d}_0^2 d_0^2) \varphi, \label{relationsforeommodes} \\
& \bar{d}_1^2 d_{-1}^2 B = d_1^2 \bar{d}_{-1}^2 B -2 [d_0^2, \bar{d}_0^2]B, \\
& (64i d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 -19 d_0^2 \bar{d}_0^2 -3 \bar{d}_0^2 d_0^2)C =\big[-11\{d_0^2, \bar{d}_0^2 \} + 128 \Pi^n_0 \Pi_{n0} -256 (\partial \bar{\theta}_0 \bar{d}_0) \nonumber \\
& \hspace{6.11cm}-256 ( \partial \theta_0 d_0) \big]C.
\end{align}
\begin{align}
d_0^2 \bar{d}_1^2 \varphi &= 2(d_0 d_0) (\bar{d}_0 \bar{d}_1) \varphi, \\
d_1^2 \bar{d}_0^2 \varphi &= (8i d^\alpha_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}_1^{\dot{\alpha} } + 2 d^{\alpha}_0 (\bar{d}_0 \bar{d}_0) d_{\alpha 1} + 8i d^{\alpha}_0 \bar{d}^{\dot{\alpha} }_0 \Pi_{\alpha \dot{\alpha} 1} -32 d^{\alpha}_0 \partial \theta_{\alpha 1} ) \varphi, \\
d_2^2 \bar{d}_{-1}^2 \varphi &= (-16 \Pi_{\alpha \dot{\alpha} 0} \Pi^{\dot{\alpha} \alpha}_1 - 8i\Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 d^{\alpha}_1 + 8i d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_1 + 8i d^{\alpha}_0 \bar{d}^{\dot{\alpha} }_0 \Pi_{\alpha \dot{\alpha} 1} \nonumber\\
&- 32 \partial \bar{\theta}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_1 -32 \bar{d}_{\dot{\alpha} 0} \partial \bar{\theta}^{\dot{\alpha} }_1 -64 d^{\alpha}_0 \partial \theta_{\alpha 1} )\varphi, \\
d^2_{-1} \bar{d}_1^2 \varphi &= 4(d_{-1} d_0)(\bar{d}_0 \bar{d}_1) \varphi, \\
d_1^2 \bar{d}_{-1}^2 \varphi &= ( 8i d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 + 8i \Pi_{\alpha \dot{\alpha} -1} d^{\alpha}_0 \bar{d}^{\dot{\alpha} }_1 - 16 \Pi_{\alpha \dot{\alpha} -1} \Pi^{\dot{\alpha} \alpha}_1-8i \Pi_{\alpha \dot{\alpha} -1} \bar{d}^{\dot{\alpha} }_0 d^{\alpha}_1\nonumber \\
& -8i \bar{d}^{\dot{\alpha} }_{-1} d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 1} -4 \bar{d}_{\dot{\alpha} -1} d^{\alpha}_0 \bar{d}^{\dot{\alpha} }_0 d_{\alpha 1} + 8i d^{\alpha}_{-1} \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_1 + 8i d^{\alpha}_{-1} \bar{d}^{\dot{\alpha} }_0 \Pi_{\alpha \dot{\alpha} 1} \nonumber \\
& -64 d^{\alpha}_{-1} \partial \theta_{\alpha 1} -64 \partial \bar{\theta}_{\dot{\alpha} -1} \bar{d}^{\dot{\alpha} }_1) \varphi, \\
d_0^2 \bar{d}_0^2 \varphi &= ( d^{\alpha}_0 d_{\alpha 0} \bar{d}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 + 2 \bar{d}_{\dot{\alpha} -1} d^{\alpha}_0 d_{\alpha 0} \bar{d}^{\dot{\alpha} }_1 + 2 d^{\alpha}_{-1} \bar{d}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 d_{\alpha 1} + 8i d^{\alpha}_{-1} \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_1 \nonumber\\
& + 8i d^{\alpha}_{-1} \bar{d}^{\dot{\alpha} }_0 \Pi_{\alpha \dot{\alpha} 1} + 8i \Pi_{\alpha \dot{\alpha} -1} d^{\alpha}_0 \bar{d}^{\dot{\alpha} }_1 -32 \partial \bar{\theta}_{\dot{\alpha} -1} \bar{d}^{\dot{\alpha} }_1 -32 d^{\alpha}_{-1} \partial \theta_{\alpha 1}) \varphi.
\end{align}
\begin{align}
d_0^2 \bar{d}_0^2 B &= (d_0 d_0) (\bar{d}_0 \bar{d}_0) B,\\
d_1^2 \bar{d}_{-1}^2 B &= 8i d^{\alpha}_0 \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 B, \\
d_{-1}^2 \bar{d}_0^2 B & = 2 d^{\alpha}_{-1} d_{\alpha 0} \bar{d}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 B, \\
d_0^{2} \bar{d}^2_{-1} B & = ( 2 \bar{d}_{\dot{\alpha} -1} d^{\alpha}_0 d_{\alpha 0} \bar{d}^{\dot{\alpha} }_0 + 8i d^{\alpha}_{-1} \Pi_{\alpha \dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 + 8i \Pi_{\alpha \dot{\alpha} -1} d^{\alpha}_0 \bar{d}^{\dot{\alpha} }_0 -32 \partial \bar{\theta}_{\dot{\alpha} -1} \bar{d}^{\dot{\alpha} }_0) B, \\
\bar{d}^2_{-1} d^2_0 B &= 2 \bar{d}_{\dot{\alpha} -1} \bar{d}^{\dot{\alpha} }_0 d^{\alpha}_0 d_{\alpha 0} B, \\
\bar{d}^2_0 d^2_{-1} B & = ( 2 d^{\alpha}_{-1} \bar{d}_{\dot{\alpha} 0} \bar{d}^{\dot{\alpha} }_0 d_{\alpha 0} + 8i \bar{d}_{\dot{\alpha} -1} \Pi^{\dot{\alpha} \alpha}_0 d_{\alpha 0} + 8i \Pi_{\alpha \dot{\alpha} -1} \bar{d}^{\dot{\alpha} }_0 d^{\alpha}_0 -32 \partial \theta^{\alpha}_{-1} d_{\alpha 0} ) B.
\end{align}
\end{appendices}
\newpage
|
\section{Introduction}
\label{S-Introduction}
Prominence seismology is a powerful diagnostic tool that combines the observation of prominence oscillations with theoretical models to infer coronal plasma parameters and to analyse the dynamical response associated with restoring forces involved in the motion \citep[see e.g.,][]{2009tripathiSSRv149,2018arregieLRSP15}. In particular, large-amplitude oscillations of prominences can be classified into longitudinal oscillations, with the motion parallel to the prominence's axis, and transverse oscillations, moving perpendicular to the axis direction. Transverse oscillations can also be divided into horizontal and vertical oscillations, and have been related to the winking filament phenomenon observed at H$\alpha$ \citep{1966AJ.....71..197R}. Ranges of periods, velocity amplitudes and damping times of 11--29~min, 6--41~km s$^{-1}$ and 25--180~min \citep{2014shenApJ795}, respectively, have been reported for transverse oscillations. Whereas longitudinal oscillations are associated with larger periods, velocity amplitudes and damping times in the ranges of 44--160~min, 30--100~km s$^{-1}$ and 115--600~min, respectively. In this work we will use equivalently the words prominence and filament.
The determination of driving mechanisms for large-amplitude oscillations are still subject to research. Nevertheless, several studies have suggested that the triggering disturbances are fast MHD shocks, such as chromospheric Moreton waves \citep{2002PASJ...54..481E,2013franchileAA552} and coronal EUV waves produced by distant flares and/or coronal mass ejections \citep{2004ApJ...608.1124O,2006A&A...449L..17I,2012asaiApJ745,2012ApJ...754....7S,2014shenApJ786,2014shenApJ795}. Subflares or jets are also able to produce large-amplitude oscillations \citep{2003ApJ...584L.103J,2021lunaApJ912}. \citet{2013franchileAA552} pointed out that the activation time of two faraway winking filaments coincided with the passage through them of a large-scale wavefront, not detected at coronal heights. Also, \cite{2014shenApJ786} analysing four winking filaments, excited by a weak coronal EUV wave coming from an X-flare region, suggested that the oscillation parameters resulted from the normal mode excitation of each filament triggered by a single disturber. Previously, \cite{1966AJ.....71..197R} proposed that the oscillation parameters depend on the intrinsic properties of the prominence rather than on the external disturbance features. In fact, if the external driver is a pulse, i.e. not a forced oscillation mechanism, the system should respond to the perturbation with its natural frequency. Moreover, considering the fluctuation-dissipation theorem \citep{1966RPPh...29..255K}, small amplitude oscillations will be spontaneously excited at the natural frequencies of the system.
Several observational studies of the properties of filament oscillations contributed to understand the physical mechanisms involved in the motion. For instance, \cite{2014shenApJ795} suggested that a prominence oscillates like a linear vertical solid body with one end tied on the solar surface. The analysis performed by \cite{2013ApJ...773..166L} and \cite{2015RAA....15.1713P} on the measurements of transverse oscillations also supported the idea that filaments oscillate as a whole. However, \cite{2011A&A...531A..53H}, by analysing this type of arched prominence oscillations, suggested that the prominence presented a global kink mode, although there were some discrepancies indicating the prominence oscillates as a collection of separated but interacting threads rather than like a rigid body. Concerning the forces, \cite{2014shenApJ795} proposed that the restoring forces of the transverse oscillations are most likely due to the coupling of gravity and magnetic tension of the supporting magnetic field. Furthermore, by analysing transverse oscillations, \cite{2008gilbertApJ685} sustained that the main restoring force is the magnetic tension.
From another perspective, several numerical studies have been performed to contribute to the understanding of prominence seismology. Among several models available to emulate prominence dynamics, the main ones are two simplified cases associated with the magnetic structures of \citet{1957ZA.....43...36K} and \citet{1974A&A....31..189K}. In the first model the prominence is sustained by the magnetic field lines (magnetic dip), whereas in the second one it is totally contained in the interior of a closed helical magnetic structure that isolates the prominence mass. For example, performing 3D MHD simulations, \citet{2020adroverA&A633} analysed prominence oscillations generated by instantaneous velocity perturbations on a relaxed system based on the Kippenhahn-Schl\"uter model, identifying the magnetic force as the restoring process for transverse oscillations. Also, they showed that periods increase with the prominence density and width, but decreases with the magnetic field strength. On the other hand, in line with Kuperus and Raadu, \citet{2018zhouApJ856} carried out 3D ideal MHD simulations of a prominence in an initially relaxed state where the oscillations were triggered by perturbing the velocity field independently in each direction. These authors found that periods of horizontal and vertical transverse oscillations are different, and attributed this result to the shape of the prominence, which is wider in the vertical direction. In addition, they determined for transverse oscillations that the main restoring force is the magnetic tension. \citet{2020liakhA&A637} used 2.5D simulations to model oscillations excited independently by horizontal and vertical internal perturbations, and also by an external perturbation. In the former case, considering a single prominence the authors found a weak dependence of transverse oscillation periods on the density contrast and on the shear angle, suggesting that periods are almost constant with height. For the latter case, they found that the displacement and deformation of the magnetic field lines produced by a coronal shock wave generates the prominence oscillation. In turn, \citet{2021lunaApJ912} by 2.5D MHD modelling demonstrated that coronal jets are able to produce large-amplitude oscillations. Alternatively on an analytical study representing a filament by a current-currying wire, which is similar to the present model, \citet{2016kolotkovA&A590} developed a linear model of the transverse oscillations of a prominence determining periods and stability conditions. This article was followed by a weakly nonlinear and a fully nonlinear study considered in \citet{2018kolotkovJASTP172}. Beyond the differences between the models or the proposed disturbances, there is a consensus that the magnetic force plays an important role as a restoring force characterising the oscillations. However, the analysis of oscillation properties is a topic still open to discussions.
With regards to the damping of oscillations, several theoretical mechanisms were proposed. An important aspect is the interaction between the moving filament and the ambient coronal plasma. This motion gives rise to a process where a transfer of energy from longer scales to smaller ones takes place until it finally dissipates due to viscous effects on smaller scales. Also, viscous processes were described by \citet{1966ZA.....63...78H}, and the damping contribution by emission of magnetoacoustic waves, known as wave leakage, was described by \citet{1969SoPh....6...72K} and \citet{1992SoPh..142..113V}. For example, \cite{1999A&A...345.1038S}, considering filament oscillations in terms of a solid body, suggested that vertical oscillations lead to fast wave emission carrying momentum away from the filament and so damping the motion, whereas horizontal oscillations would lead to the emission of slow waves but in a less effective process. In this sense, \citet{2018arregieLRSP15} noticed that as emitted slow waves propagate along magnetic field lines are unable to take energy out of an environment with closed magnetic lines. Under these conditions, wave leakage is only possible by fast waves. Another important damping mechanism is the resonant wave process \citep[see e.g.,][]{2011SSRv..158..289G}, where the energy of the global oscillation is transferred to an inhomogeneous transition layer between the filament and the solar corona. For instance, \cite{2011A&A...531A..53H} observationally studying large-amplitude transverse oscillations of a filament, where periods and damping times for different heights were determined, found that the damping times showed a linear dependence with the periods, where the resonant absorption was suggested as the main damping mechanism. Particularly, throughout 3D numerical simulations \cite{2016ApJ...820..125T} showed evidence that the energy transfer from global oscillations to continuum Alfv\'en modes at the filament's edge is the main damping mechanism.
With the aim to contribute to the understanding of quiescent filament oscillations, in this paper we perform 2.5D ideal MHD simulations. Our motivation is to study large-amplitude transverse oscillations driven by an external coronal shock wave produced by a distant event, e.g., a jet or a flare. For this purpose, we consider different scenarios varying the height, size and mass of the filament. The interaction between the filament and the coronal shock wave is analysed by means of the main motion characteristics, i.e., period of oscillation, amplitudes, damping time and restoring forces.
The paper is organised as follows: in Section 2 we present the governing equations, the solar atmospheric scenario and the magnetic model used to represent the filament, detailing how the initial equilibrium conditions were found in a stratified medium with variable gravity. The main features of the numerical code used to run different configurations and the perturbing method are developed. In Section 3 we present the results and discussion, starting with the analysis of the filament dynamics in terms of the forces involved to continue with a parametric study evaluating the different configurations. Finally in Section 4 the main conclusions are exposed.
\section{The model}
\label{S-Model}
By numerical experiments we study transverse oscillatory properties of quiescent filaments triggered by a large-scale coronal wave coming from a remote site. The scenario is depicted in Figure~\ref{f:esquema}. The quiescent filament is assumed to be located in a coronal quiet-region, being long enough such that it can be adequately represented by a 2.5D modelling in a perpendicular plane of symmetry. Based on the model of \cite{1990forbesJGR95}, the scenario includes: a filament floating in the corona at a certain height, a variable gravity and a stratified coronal background in hydrostatic equilibrium, with the chromosphere and a thin transition region at the base. The large-scale coronal wave perturbation is emulated by a blast. This scenario allows us to carry on oscillatory analysis measuring periods of oscillations, amplitudes and damping times.
\begin{figure}
\centerline{\includegraphics[width=0.8\textwidth]{fig1b.pdf}}
\caption{Our schematic scenario of a quiescent filament floating in a stratified atmosphere hit transversely by a coronal shock wave. The plane $(x,y)$ is considered in numerical simulations.}
\label{f:esquema}
\end{figure}
An ideal MHD scenario is considered with a compressible plasma in presence of a gravitational field. In conservative form the ideal MHD equations read (CGS units):
\begin{eqnarray}
\frac{\partial \rho}{\partial t} + \mathbfit{\nabla}\cdot\left(\rho\mathbfit{v}\right) &=& 0, \nonumber \\
\frac{\partial \rho\mathbfit{v}}{\partial t} + \mathbfit{\nabla}\cdot \left(\rho\mathbfit{v}\mathbfit{v}-\frac{1}{4\pi}\mathbfit{B}\mathbfit{B}\right) &=& -\mathbfit{\nabla}p + \frac{1}{c}\mathbfit{j}\times\mathbfit{B} + \rho\mathbfit{g}, \nonumber \\
\frac{\partial\mathbfit{B}}{\partial t} + \mathbfit{\nabla} \cdot \left(\mathbfit{vB}-\mathbfit{Bv}\right) &=& 0, \nonumber \\
\frac{\partial E}{\partial t} + \mathbfit{\nabla}\cdot \left[(E+p+\frac{B^2}{8\pi})\mathbfit{v}-\frac{1}{4\pi}\mathbfit{B}(\mathbfit{v}\cdot\mathbfit{B})\right] &=& \rho\mathbfit{g}\cdot\mathbfit{v},
\label{e:mhd}
\end{eqnarray}
where $\rho$ is mass density, $\mathbfit{v}$ is plasma velocity, $\mathbfit{B}$ is magnetic field, $\mathbfit{j}$ is current density and $\mathbfit{g}$ is acceleration of gravity. Also, $E$ is total energy, $e$ is internal energy and $p$ is thermal pressure, whose expressions are
\begin{eqnarray}
&& E = \frac{1}{2}\rho v^2 + e + \frac{B^2}{8\pi}, \nonumber \\
&& e = \frac{p}{\left(\gamma -1\right)}, \nonumber \\
&& p = \frac{R_{\rm{g}}}{\bar{\mu}}\rho T,
\label{e:ener}
\end{eqnarray}
with $T$ the plasma temperature, $\gamma=5/3$ and $R_{\rm{g}}$ the gas constant. The ideal plasma is assumed to be fully ionised and with a solar abundance\footnote{Solar abundance: $70.7\%~$ H $+$ $27.4\%~$ He $+$ $1.9\%~$ heavier elements.} (implying a mean atomic mass $\bar{\mu}=0.613$). The Amp\`ere's law is $\mathbfit{j}=\frac{c}{4\pi} {\bf \nabla} \times\mathbfit{B}$, with $c$ the speed of light. Furthermore, let us note in the conservative MHD Equations~\ref{e:mhd} appear the tensor products $\mathbfit{v}\mathbfit{v}$, $\mathbfit{vB}$, $\mathbfit{Bv}$ and $\mathbfit{BB}$ (the magnetic part of the Maxwell stress tensor).
\subsubsection{Stratified atmosphere}
\label{sub:sa}
To describe our model set-up two reference systems were used with different origins. First, a Cartesian coordinates $(x,y,z)$ was used to describe the set-up as a whole with its origin fixed at the base of the chromosphere, the $y$-axis pointing radially away from the solar surface and the $x$-, $z$-axes parallel to the surface (neglecting the surface curvature). Second, to locally describe the filament structure it is convenient to use a cylindrical coordinate system $(r,\phi,z')$ with its origin fixed at the filament's centre, for the radial, azimuthal and axial directions, respectively, see Figure~\ref{f:esquema}. Note that as the $z'$- and $z$-axes are parallel, the $z'$ coordinate can be omitted.
The simulated solar atmosphere is in hydrostatic equilibrium with a gravitational stratification in the vertical direction including a temperature profile for the chromosphere, transition region and corona. The chromosphere has a height $h_{\rm {cr}}=2.5~$Mm with a constant temperature $T_{\rm {cr}}=1.2\times 10^4~$K, the transition region extends up to a height $h_{\rm {tr}}=0.5~$Mm with a linearly varying temperature, and the corona is based at $y=h_{\rm {cr}}+h_{\rm {tr}}$ extending upwards with a constant temperature $T_0=1~$MK. Therefore, the adopted atmospheric temperature is
\begin{equation}
T(y)=\left\{
\begin{array}{ll}
T_{\rm {cr}} & \ {\rm at~} 0\le y \le h_{\rm {cr}}, \vspace{0.2cm} \\
\frac{\left(T_0-T_{\rm {cr}}\right)}{h_{\rm {tr}}} \left(y-h_{\rm {cr}}\right) + T_{\rm {cr}} & \ {\rm at~} h_{\rm {cr}}< y < h_{\rm {cr}}+h_{\rm {tr}}, \vspace{0.2cm} \\
T_0 & \ {\rm at~} y \geq h_{\rm {cr}}+h_{\rm {tr}}.
\end{array} \right.
\label{e:Tatm}
\end{equation}
The initial pressure due to the hydrostatic equilibrium condition, $\frac{{\rm d}p}{{\rm d}y}=-\rho g$, is
\begin{equation}
p(y)=\left\{
\begin{array}{ll}
p_{\rm {tr}} \exp\left(\frac{\alpha}{T_{\rm {cr}}} \left[ \frac{1}{y+R_{\odot}} - \frac{1}{h_{\rm {cr}}+R_{\odot}} \right] \right) & \ {\rm at~} 0\le y \le h_{\rm {cr}}, \vspace{0.25cm} \\
p_0\exp\left(-\alpha \int_{h_{\rm {cr}}+h_{\rm {tr}}}^{y} \frac{{\rm d}y'}{T(y')\left(y'+R_{\odot}\right)^2} \right) & \ {\rm at~} h_{\rm {cr}}< y < h_{\rm {cr}}+h_{\rm {tr}}, \vspace{0.25cm} \\
p_0\exp\left(\frac{\alpha}{T_0} \left[\frac{1}{y+R_{\odot}} - \frac{1}{h_{\rm {cr}}+h_{\rm {tr}}+R_{\odot}}\right] \right) & \ {\rm at~} y \geq h_{\rm {cr}}+h_{\rm {tr}},
\end{array} \right.
\label{e:Patm}
\end{equation}
with the reference pressure $p_0=\frac{R_{\rm{g}}}{\bar{\mu}}\rho_0 T_0$ fixed at the coronal base and
\begin{equation}
p_{\rm {tr}} = p_0\exp\left(-\alpha \int_{h_{\rm {cr}}+h_{\rm {tr}}}^{h_{\rm {cr}}} \frac{{\rm d}y'}{T(y')\left(y'+R_{\odot}\right)^2} \right)
\label{e:Ptr}
\end{equation}
being the extrapolated pressure at the base of the transition region. The gravity acceleration is $\mathbfit{g}(y)=-\frac{G M_{\odot}}{(y+R_{\odot})^2}\hat{\mathbfit{j}}$. $\alpha=\frac{\bar{\mu}GM_{\odot}}{R_{\rm{g}}}$ is a constant, with $G$ the gravitational constant, $M_{\odot}$ the solar mass and $R_{\odot}$ the solar radius. The density $\rho$ and internal energy $e$ in the atmosphere are determined by the Equations of state \ref{e:ener}. Finally, the integral in Equations~\ref{e:Patm} and \ref{e:Ptr} has a closed analytic solution if $T(y)$ is a linear function, which is
\begin{eqnarray}
\int_{y_0}^{y} \frac{{\rm d}y'}{T(y')\left(y'+R_{\odot}\right)^2} &\!=\!& \frac{1}{R_{\odot}a-b}\left(\frac{1}{y+R_{\odot}}-\frac{1}{y_0+R_{\odot}}\right)
\nonumber \\
&&
+ \frac{a}{\left(R_{\odot}a-b\right)^2} \ln\left(\frac{T(y)}{T(y_0)}\frac{y_0+R_{\odot}}{y+R_{\odot}} \right).
\end{eqnarray}
Constants $a$ and $b$ are obtained by writing the linear temperature of the transition region in Equation \ref{e:Tatm} as $T(y)=ay+b$, having $a\equiv \frac{T_0-T_{\rm {cr}}}{h_{\rm {tr}}}$ and $b\equiv T_{\rm {cr}}-a h_{\rm {cr}}$.
\subsubsection{The filament}
\label{sub:fil}
In line with \citet{1990forbesJGR95}, but adding the gravity force, the flux rope is modelled as an electric current-carrying wire of radius $R$ floating in the corona. The magnetic configuration consists of three components: the current-carrying wire located at height $y=h$; a mirror current at depth $y=-h$; and a line dipole of relative strength $M_{\rm dip}$ at depth $y=-d$, representing the photospheric field.
The initial magnetic field, considering each contribution respectively, is:
\begin{eqnarray}
B_x(x,y) &=& - \frac{(y - h)}{r_-}B_\phi(r_-) + \frac{(y + h)}{r_+}B_\phi(r_+) \nonumber \\
&& + \frac{\left(x^2 - (y + d)^2\right)}{r_d^4} M_{\rm{dip}} d R_3 B_\phi (R_3) , \nonumber \\
B_y(x,y) &=& \frac{x}{r_-}B_\phi(r_-) - \frac{x}{r_+}B_\phi(r_+) + \frac{2x(y + d)}{r_d^4} M_{\rm{dip}} d R_3 B_\phi (R_3), \nonumber \\
B_z(x,y) &=& B_z(r),
\label{e:bxy}
\end{eqnarray}
with $r_{-}$, $r_{+}$ and $r_d$ being the distances from the filament, the mirror current and the dipole, respectively, to the observed point, where
\begin{eqnarray}
r_\pm &=& \sqrt{x^2 + (y \pm h)^2}, \\
r_d &=& \sqrt{x^2 + (y + d)^2}.
\label{e:rB}
\end{eqnarray}
$B_\phi$, $B_z$, $R_3$ and $M_{\rm dip}$ are defined next.
We assume a transition layer of thickness $\Delta$ to relate the internal structure of a filament of radius $R$ with the outer coronal region. Thus, three filament zones are distinguished: an inner zone (Z1) for radius $0\le r \le R_2$; a transition zone (Z2) for radius $R_2 < r < R_3$; and the external zone (Z3) for $r\ge R_3$. $R_2=R-\frac{\Delta}{2}$ and $R_3=R+\frac{\Delta}{2}$. High-resolution observations have revealed the existence of inhomogeneous fine structures of filaments \citep{2005SoPh..226..239L}, making reasonable to include in the model a non-uniform layer between the filament inner zone and the coronal external zone \citep[e.g., as done by][and references therein]{2009solerApJL}.
The assumed electric current density inside the filament is $\mathbfit{j}(r)=j_\phi${\boldmath$\hat{\phi}$}$+ j_z \hat{\mathbfit{k}}$, with components:
\begin{equation}
\! j_z(r) \!=\! \left\{
\begin{array}{ll}
j_0 & \vspace{0.2cm} \\
\! \frac{j_0}{2} \left(\cos\left[\frac{\pi}{\Delta}\left(r-R_2\right)\right] + 1 \right) & \vspace{0.2cm} \\
0 &
\end{array} \right.
{\rm {and}} \quad
j_\phi(r) \!=\! \left\{
\begin{array}{ll}
\! \frac{j_1 r}{2\sqrt{R_2^2-r^2}}, & \ : \ {\rm at~Z1}, \vspace{0.2cm} \\
\! 0 & \ : \ {\rm at~Z2}, \vspace{0.2cm} \\
0 & \ : \ {\rm at~Z3},
\end{array} \right.
\label{e:jz}
\end{equation}
where $j_0$ and $j_1$ are constants. The azimuthal component of the magnetic field $B_\phi(r)$ produced by the axial current density $j_z$ is
\begin{equation}
B_\phi(r)=\left\{
\begin{array}{ll}
\frac{2\pi j_0}{c} r & \ {\rm at~Z1}, \vspace{0.2cm} \\
\frac{2\pi j_0}{c}\frac{1}{r} \left\lbrace \frac{r^2}{2} + \frac{R_2^2}{2} - \left(\frac{\Delta}{\pi}\right)^2 + \left(\frac{\Delta}{\pi}\right)^2 \cos\left[\frac{\pi}{\Delta}\left(r-R_2\right)\right] \right. \\
\left. +\frac{\Delta}{\pi}r \sin\left[\frac{\pi}{\Delta}(r-R_2)\right] \right\rbrace & \ {\rm at~Z2}, \vspace{0.2cm} \\
\frac{2\pi j_0}{c}\frac{1}{r} \left(\frac{R_3^2}{2} + \frac{R_2^2}{2} - 2\left(\frac{\Delta}{\pi}\right)^2\right) & \ {\rm at~Z3},
\end{array} \right.
\label{e:bphi}
\end{equation}
and the axial component is
\begin{equation}
B_z(r)=\left\{
\begin{array}{ll}
\frac{\sqrt{8}\pi}{c}j_1\sqrt{R_2^2-r^2} & \ {\rm at~Z1}, \vspace{0.2cm} \\
0 & \ {\rm at~Z2}, \vspace{0.2cm} \\
0 & \ {\rm at~Z3}.
\end{array} \right.
\label{e:bz}
\end{equation}
The initial thermal pressure inside the filament, intending to be in equilibrium as close as possible with its neighbourhood and satisfying $\mathbfit{\nabla}p - \frac{1}{c}\mathbfit{j}\times\mathbfit{B}\approx 0$, results in
\begin{equation}
p(r) = p_{\rm c}(r) - \frac{1}{c}\int_r^\infty j_\phi(r')B_z(r') \, {\rm d} r' + \frac{1}{c}\int_r^\infty j_z(r')B_\phi(r') \, {\rm d}r',
\label{e:p1}
\end{equation}
where $p_{\rm c}$ is the background coronal pressure given by the last expression of Equation \ref{e:Patm}. Considering the different shells in the radial direction throughout the filament, the pressure is
\begin{equation}
p(r)=\left\{
\begin{array}{ll}
p_{\rm c}(r) + \frac{\pi}{c^2}\left(j_0^2-j_1^2\right)\left(R_2^2-r^2\right)
+ \frac{1}{c}\int_{R_2}^{R_3} j_z(r')B_\phi(r') \, {\rm d}r' & \ {\rm at~Z1}, \vspace{0.2cm} \\
p_{\rm c}(r) + \frac{1}{c}\int_{r}^{R_3} j_z(r')B_\phi(r') \, {\rm d}r' & \ {\rm at~Z2}, \vspace{0.2cm} \\
p_{\rm c}(r) & \ {\rm at~Z3}.
\end{array} \right.
\label{e:Pfil}
\end{equation}
This balance pressure was obtained taking into account the azimuthal magnetic field $B_{\phi}$ generated by the filament, written in Equation~\ref{e:bphi}, but neglecting the external contributions to the azimuthal component by the mirror wire and the dipole. The prescribed filament temperature is
\begin{equation}
T(r)=\left\{
\begin{array}{ll}
T_{\rm {fil}} & \ {\rm at~Z1}, \vspace{0.2cm} \\
\frac{\left(T_0-T_{\rm {fil}}\right)}{\Delta} \left(r-R_2\right) + T_{\rm {fil}} & \ {\rm at~Z2}, \vspace{0.2cm} \\
T_0 & \ {\rm at~Z3},
\end{array} \right.
\label{e:Tfil}
\end{equation}
with the constant inner temperature of the filament $T_{\rm {fil}}=0.3~$MK. Finally, the density $\rho(r)$ of the filament is obtained introducing the expressions for the pressure and temperature, Equations \ref{e:Pfil} and \ref{e:Tfil}, into Equation \ref{e:ener}.
Figure~\ref{f:init_cond} shows zoom-in initial conditions for a particular-filament case floating in the stratified atmosphere. The density pattern is shown in panel \textsf{(a)}, and the thermal pressure with the total magnetic field lines superimposed is shown in panel \textsf{(b)}.
\begin{figure}
\centerline{
\hspace*{0.015\textwidth}
\includegraphics[width=0.59\textwidth]{dens_init.png}
\hspace*{-0.05\textwidth}
\includegraphics[width=0.58\textwidth]{pres_init.png}}
\caption{Zoom-in snapshots of the initial conditions for the filament floating in a stratified atmosphere. \textsf{(a)} The density pattern is shown. \textsf{(b)} Thermal pressure with total magnetic field lines superimposed (black-solid lines). The thick-dashed line in panel (a) indicates the slice used to evaluate the distance-magnetic field plot. This filament configuration corresponds to Case 3 defined below in Section \ref{Sub:sim}.}
\label{f:init_cond}
\end{figure}
\subsubsection{Equilibrium condition}
\label{sub:equi}
In order to establish equilibrium conditions, the forces acting on the filament are analysed. The forces are: the upward force exerted by the mirror current $F_{\rm mir}$; and the downward forces of the dipole $F_{\rm dip}$ and the weight $F _{\rm g}$. Thus, the initial total force $F$, per unit length $L$, in the vertical direction is
\begin{eqnarray}
\frac{F}{L} &=& F_{\rm mir} - F _{\rm dip} - F_{\rm g} \nonumber \\
&\approx& \frac{1}{c^2}\frac{I^2}{h} - \frac{2}{c^2}\frac{Im}{\left(h+d\right)^2} - \left(m_{\rm fil} - m_{\rm buo}\right)g,
\label{e:force}
\end{eqnarray}
where approximations were made on the magnetic force expressions and are valid for a small radius $R$. $I$ is the electric current flowing inside the filament and its mirror, $m$ is the dipole strength, $m_{\rm fil}$ is the mass of the filament (per unit length) and $m_{\rm buo}$ is the background mass enclosed in a filament-like volume. The weight term includes the buoyancy force $m_{\rm buo}g$ that is smaller than the weight, $\sim$1$\%$. The axial electric current going through the transverse filament area $A_{\rm fil}$ is
\begin{equation}
I=\int_{A_{\rm fil}} j_z \hat{\mathbfit{k}}\cdot{\rm d}\mathbfit{A'}= \frac{\pi j_0}{2}\left(R_3^2+R_2^2-\left(\frac{2\Delta}{\pi}\right)^2 \right).
\label{e:curr}
\end{equation}
\begin{figure}
\centerline{\includegraphics[width=0.6\textwidth]{equil.pdf}}
\caption{The locations of equilibria for the ratio $\frac{h}{d}$ as a function of the relative strength of the dipole $M_{\rm{dip}}$. Black-solid portion of the line represents stable equilibria, while black-dashed portion unstable ones. Red lines display the same example but neglecting the gravity. The equilibria indicated by the black lines and the stable point chosen at $M_{\rm{dip}}=2.25$ correspond specifically to Case 3 defined in Section \ref{Sub:sim}.}
\label{f:equil}
\end{figure}
To satisfy the balance ($F=0$) the following relation must hold:
\begin{equation}
\frac{h}{d}=\frac{\left(M_{\rm{dip}}-(1-M_{\rm{g}})\right)\pm \sqrt{M_{\rm{dip}}^2-2M_{\rm{dip}}(1-M_{\rm{g}})}}{1-M_{\rm{g}}},
\label{e:balan}
\end{equation}
which gives the equilibria in turns of the height and depth as a function of two parameters, $M_{\rm{dip}}$ and $M_{\rm{g}}$. The dimensionless parameter $M_{\rm{dip}}=\frac{m}{Id}$ gives the relative strength between the dipole force $F_{\rm dip}$ and the mirror force $F_{\rm mir}$. Similarly, $M_{\rm{g}}=\frac{c^2 h}{I^2}F_{\rm g}$ is the relative strength between the weight and the mirror forces. Solutions with minus (plus) sign represent stable (unstable) equilibria \citep{1990forbesJGR95}. An example of these equilibria are shown in Figure~\ref{f:equil}, where the radio $\frac{h}{d}$ between the filament height and the dipole depth is plotted as a function of $M_{\rm{dip}}$. The black-solid portion of the line represents stable locations while the black-dashed portion unstable ones. To have an idea of the weight contribution, the red curve represents the equilibria neglecting the weight force. Note that the weight force allows to consider smaller values of the relative dipole strength $M_{\rm{dip}}$, i.e. considering gravity, the minimum stable location occurs for $M_{\rm{dip}}=1.24$, while neglecting gravity, the minimum stable location occurs for $M_{\rm{dip}}=2$.
When searching for equilibrium configurations the mass of the filament plays an important role, but its value $m_{\rm fil}=\int_{V_{\rm fil}} \rho(\mathbfit{r'}) \, {\rm d}V'$ has not a closed expression. Therefore, we obtained the following approximated expression
\begin{equation}
m_{fil} \approx \frac{1}{T_{\rm fil}} \left[\pi\rho_0 T_0 R^2 + \epsilon R^4(j_0^2 - \frac{j_1^2}{2}) \right],
\label{e:mfil}
\end{equation}
which is helpful to understand the mass dependence on the set-up parameters. With $\epsilon=\frac{\pi\bar{\mu}}{2R_{\rm g}c^2}$ being a constant. The filament mass increases for larger values of the radius $R$ and the axial current density $j_0$, while decreases with $j_1$. In addition, the mass is inversely proportional to its temperature $T_{\rm fil}$, resulting in heavier filaments for colder temperatures (an vice-versa). In a lesser degree the mass also depends on the coronal background density $\rho_0$ and temperature $T_0$.
\subsubsection{Perturbation}
\label{sub:pert}
After a relaxation period, the filament was perturbed using a blast mechanism \citep{2004balsaraApJS151}, placed far in the domain's border. This configuration emulates a standing quiescent filament perturbed by an external large-scale coronal wave coming, for example, from a remote energetic flaring site. More details of the simulations are given below in Section \ref{Sub:code}. The wavefront, with a small inclination angle from the horizontal axis towards the solar surface, hits the filament exciting simultaneously horizontal and vertical transverse oscillations. The downward inclination of the wave is in line with observations \citep{2013ApJ...773..166L}. An inclination angle of $\sim$7$^{\circ}$ was used in all cases. The instantaneously applied blast was imposed at $t=30.08~$min, placed at $(x,y)=(95,h+6)~$Mm with a diameter of $20~$Mm and releasing an energy of $7\times 10^{19}~$erg. The average propagation speed of the shock wave was of $486~$km s$^{-1}$. Which in terms of the Alfv\'en Mach number $M_{\rm A}$, for the different cases carried on (see Section \ref{Sub:sim} below) involving different magnetic field strengths (Figure~\ref{f:B_cut}), represented $M_{\rm A}=1.8$ for filaments with small radius and $M_{\rm A}=0.85$ for big radius filaments (in both cases measured at coordinates $(x,y)=(49,19)~$Mm where the ambient magnetic field magnitude behind the wave-front was $3.~$G on the former case and $5.8~$G on the latter).
\cite{2018zhouApJ856} discarded simulations of shock waves as external perturbations on the filament, by assuming that perturbing the velocity field would be representative of anything that results in a bulk motion of the filament. Following this idea, as an alternative to the blast mechanism, we also tried perturbing the velocity field inside and around the filament, which is a much less numerically demanding mechanism. This mechanism is supposed to mimic the velocity impulsive phase exerted by the blast on the filament. However, we found that the resulting oscillations were not equivalent to those obtained from the blast, being the main difference that the damping is stronger when the velocity mechanism was used. Hence, to model large-scale coronal wave perturbations, we discarded the velocity mechanism.
\subsection{Numerical simulations}
\subsubsection{Code}
\label{Sub:code}
To carry out 2.5D ideal MHD numerical simulations the set of Equations~\ref{e:mhd} was solved using the {\sc Flash} code \citep[][release 4.5]{2013DubeyIJHPCA28}. This code uses Godunov-type schemes in a co-located regular grid of finite volumes to solve the coupled compressible MHD equations with adaptive mesh refinement (AMR) and high performance computing capabilities.
For our simulations we chose the unsplit staggered mesh solver with a second-order MUSCL-type reconstruction. This solver, which is based on the scheme by \citet{2009leeJCoPh228}, uses a directionally unsplit technique for the evaluation of numerical fluxes and implements the constrained transport method and the corner transport upwind method for the treatment of the magnetic field, which leads to a better numerical behaviour of the divergence free condition ($\mathbfit{\nabla}\cdot\mathbfit{B}=0$). The Riemann's problems at interfaces of computational cells were calculated using the Roe's solver and the MC slope-limiter was used for reconstructions of cell-centred variables.
The 2D physical domain was delimited by $[-120, 120]~\rm{Mm}\times [0, 120]~\rm{Mm}$ and discretised by a Cartesian grid with eight levels of refinement with $20 \times 10$ cells per block, which gives a maximal spatial resolution of $\delta x = \delta y = 0.094~$Mm. Boundary conditions were set as follows: for the left and right borders were chosen zero-gradient (outflow) conditions for thermodynamic variables and velocity, allowing waves to leave the domain without reflection. For the magnetic field at lateral and upper ends we extrapolated the initial condition to ghost cells in order to avoid spurious magnetic forces generated by the violation of the force-free condition of the border background magnetic field. Obviously, this implementation is valid provided that the magnetic perturbations do not reach these borders, although relatively small numerically induced forces can be tolerated since the region of interest is at the prominence location. The bottom and upper borders require more detailed treatment due to the presence of gravity and the tying of the magnetic field lines to the chromosphere. For the vertical extrapolation of thermodynamic variables we implemented the hydrostatic boundary conditions proposed by \citet{2019krauseAA631} with constant temperature, while the zero-gradient condition is imposed for the velocity at the upper end. On the other hand, the line-tied boundary conditions proposed by \citet{1987SoPh..114..311R} were applied to the magnetic field and velocity at the bottom.
Taking into account that our study requires an accurate model of the prominence equilibrium at the unperturbed condition, we needed to improve the capability of the numerical code to preserve the hydrostatic equilibrium. As explained by \citet{2019krauseAA631}, the traditional MUSCL-type scheme are not able to satisfy that condition since the vertical exponential decay of the pressure cannot be exactly approximated by a polynomial reconstruction. In addition, the numerical pressure gradient is not cancelled by the gravitational source term when a simple cell-centred evaluation is used. Consequently, spurious momentum fluxes are numerically induced producing non-physical velocities. To avoid this undesirable behaviour, we implemented the local hydrostatic reconstruction scheme and we improved the equilibrium preservation during the simulation.
\subsubsection{Simulations}
\label{Sub:sim}
The overall simulations were composed by two stages. The first one was the relaxation period beginning from the initial condition at $t_0 = 0$ (see Figure~\ref{f:init_cond}) and extending for 30 min to allow an equilibrium adjustment between the filament and its background environment. This time lapse is necessary because initially the filament is not in exact equilibrium. After this time lapse the system reaches a pseudo-equilibrium position, stationary enough to be considered still. By the end of this first period the non-equilibrium velocity of the filament is smaller than $1~$km s$^{-1}$, that means an order of magnitude smaller than the velocity exerted by the blast. The second stage corresponds to the oscillating scenario, which starts when the blast is discharged, at $t_{\rm pert}=30.08~$min, and lasts until the simulation ends. During this stage the shock wave generated by the blast passed through the filament and the oscillating motion is triggered.
Aiming to make a seismological study of the system as a function of height, size and mass of the filament, seven simulations were performed, six of them for three different heights ($h=[14,19,25]~$Mm) and two radii ($R=[2.5,3.5]~$Mm), and the remaining one varying the mass by means of the current density $j_1$. These three parameters were assumed as the independent ones, while the remaining parameters, i.e. $j_0$, $\rho_0$, $\Delta$, $M_{\rm dip}$, $T_0$, $T_{\rm cr}$, $T_{\rm fil}$, $h_{\rm cr}$ and $h_{\rm tr}$, have fixed values for all cases. The details of each case and the fixed parameters are listed in Table~\ref{t:fits} (Set-up parameters). For all cases: $j_0 = 600~$statA cm$^{-2}$; $\rho_0 = 3.2\times 10^{-15}~$g cm$^{-3}$; $\Delta = 1.25~$Mm; and $M_{\rm{dip}} = 2.25$. Note that the azimuthal current density $j_1=600~$statA cm$^{-2}$ is the same for all cases, except in Case 5 that is equal to $j_1=797.5$ statA cm$^{-2}$. It should be taken into account that the equilibrium black lines displayed in Figure~\ref{f:equil} are different for all cases.
\begin{figure}
\centerline{\includegraphics[width=0.6\textwidth]{Bcut.pdf}}
\caption{Distance-magnetic field plot along the thick-dashed line of Figure~\ref{f:init_cond}\textsf{(a)}. The values corresponds to the magnetic field magnitude in the plane $(x,y)$ for the filaments of Case 3 (black-solid line) with small radius $R=2.5~$Mm and Case 4 (red-solid line) with big radius $R=3.5~$Mm.}
\label{f:B_cut}
\end{figure}
For all cases, the dipole depth $d$ and strength $m$ are dependent parameters determined by the balance Equation \ref{e:balan}. Thus, to set up equilibrium configurations, higher filaments correspond to deeper magnetic dipoles but interestingly with larger strengths $m$. On the other hand, to increase the filament radius $R$ implies larger values of the magnetic field in the filament surroundings (as displayed in Figure~\ref{f:B_cut}) and larger repelling mirror magnetic forces $F_{\rm mir}$. In addition, a larger radius produces a shallower dipole with larger strength $m$, and also increases the filament mass $m_{\rm fil}$, thus balancing the repelling mirror force. Finally, given the total pressure balance between the filament's interior and exterior by Equation \ref{e:Pfil}, a rise in the azimuthal current density $j_1$, as in Case 5, results in a decrease of the filament mass, generating a shallower dipole with smaller strength $m$.
\begin{figure}
\centerline{
\hspace*{0.015\textwidth}
\includegraphics[width=0.65\textwidth,angle=0]{pres0.png}
\hspace*{-0.034\textwidth}
\includegraphics[width=0.65\textwidth,angle=0]{pres1.png}}
\centerline{
\hspace*{0.015\textwidth}
\includegraphics[width=0.65\textwidth,angle=0]{pres2.png}
\hspace*{-0.034\textwidth}
\includegraphics[width=0.65\textwidth,angle=0]{pres3.png}}
\caption{Temporal evolution of the interaction between the quiescent filament and a large-scale coronal wave coming from a remote site. The images correspond to zoom-in snapshots of Case 3. \textsf{(a)} In colour map it is plotted the pressure pattern at time $t=32~$min before the filament-shock interaction begins. The black lines in (a) display the field lines of the total magnetic field. \textsf{(b)} It is shown just a moment after the interaction begins at $t=35~$min. The arrow points the inclination angle of the shock and the dash line represents the horizontal direction. \textsf{(c)} When the filament oscillation reaches its maximum left horizontal displacement at $t=41~$min. Here the vertical dashed line indicates the slice used to evaluate the pressure stack Figure~\ref{f:damping}\textsf{(b)}. \textsf{(d)} Same as (c) but for the maximum right horizontal displacement at $t=59~$min. (See Case 3's movie attached).}
\label{f:pres_patt}
\end{figure}
\section{Results and Discussions}
\subsection{Dynamics}
As a result of the interaction with the shock wave, the filament exhibits a damped oscillatory motion coupled in the horizontal and vertical directions. Figure~\ref{f:pres_patt} displays the evolution of the wavefront--filament interaction (see the attached Case 3's movie). Panel \textsf{(a)} shows the shock wave triggered by the blast, travelling at a coronal level and disturbing the chromosphere. In the figure, black lines represent the total magnetic field and the colour map the thermal pressure. On the right side, the presence of waves compressing the plasma and deforming the magnetic field is noticeable. At $t=33~$min, both the main coronal wavefront and its subsequent chromospheric reflection impact the filament and destabilise its equilibrium state (see panel \textsf{(b)}). The main wavefront hits the filament with a small downward inclination angle. Then, the shock wave continues propagating beyond the filament position, dragging it towards the bottom-left direction pointed by the arrow on panel \textsf{(b)}. Later on, the filament starts oscillating during a few cycles (panels \textsf{(c)}--\textsf{(d)}). By $t= 42.5~$min the external perturbation has left the domain.
\begin{figure}
\centerline{
\hspace*{0.015\textwidth}
\includegraphics[width=0.515\textwidth]{c3dispX.pdf}
\hspace*{-0.03\textwidth}
\includegraphics[width=0.515\textwidth]{c3dispY.pdf}}
\centerline{
\hspace*{0.015\textwidth}
\includegraphics[width=0.515\textwidth]{c3speedX.pdf}
\hspace*{-0.03\textwidth}
\includegraphics[width=0.515\textwidth]{c3speedY.pdf}}
\caption{The filament oscillations for Case 3. \textsf{(a)}--\textsf{(b)} Horizontal and vertical displacements are shown as a function of time, just before and after the interaction was triggered: black-squared lines represents the simulation data while red lines represent the fitted curves. Note the displacements are relative to the initial position. \textsf{(c)}--\textsf{(d)} Displayed corresponding horizontal and vertical oscillation speeds.}
\label{f:disp_osc}
\end{figure}
To analyse the filament oscillatory motion, the displacement and velocity of the centre of mass were measured as a function of time for each case. The horizontal and vertical components are represented by $\Delta x$, $\Delta y$, $v_x$ and $v_y$, respectively. Afterwards, the displacements were fitted using an exponential decayed harmonic function $f_i(t)=A_i\sin(\frac{2\pi (t-t_f)}{P_i} +\phi)\exp(-\frac{t-t_f}{\tau_i})+C_i(t-t_f)+f_{0i}$, where $f_i$ holds for the displacements $\Delta x$ and $\Delta y$, $A_i$ is the amplitude of the oscillation, $P_i$ the period, $\tau_i$ the decay constant, $C_i$ the slope, $f_{0i}$ the initial position and $t_f=35~$min the initial fitted time. The index $i$ indicates the fitted variable. The results obtained for all cases are exhibited in Table~\ref{t:fits} (Fitted Parameters).
As an example, Figure~\ref{f:disp_osc} shows in black-squared lines the displacements and speeds measured for Case 3 along with the fitted curves in red lines. The speed fits are obtained by the time derivative of the displacement fitted functions d$f_i(t)$/d$t$, however the corresponding speed amplitudes displayed in Table~\ref{t:fits} are calculated as the half difference between the minimum and maximum values. The fits show that there is a main frequency describing the oscillation with a certain degree of accuracy. Note that the displacements and speeds are almost null before the interaction begins and then show a damped oscillation with the vertical motion exhibiting a small upward slope. This upward slope is mainly due to a slight reduction of the filament density with time because of numerical diffusion. Comparing the left panels \textsf{(a)}--\textsf{(c)} with the right ones \textsf{(b)}--\textsf{(d)}, it is clearly seen that the horizontal period is larger than the vertical one. In the same way, the horizontal damping time is larger than the vertical one (see Table~\ref{t:fits}).
This behaviour is qualitatively similar for all cases.
Figure~\ref{f:traj} displays the trajectories for Cases 3 and 4 during the oscillating stage. As the displacements are referred to the initial position, both trajectories almost begin at the origin indicated by their corresponding times $t=30~$ min. The numeric points are time sequenced to indicate the temporal evolution. Comparing the trajectories with those determined analytically by \citet[][see their figures 3 and 4]{2018kolotkovJASTP172} utilising a similar model as ours, we can see that the trajectories in Figure~\ref{f:traj} resemble Lissajous-like curves of an hourglass shape suggesting that the coupled transverse oscillations are close to a resonant regimes. Note the patterns shown in Figure~\ref{f:traj} are distorted from the symmetric Lissajous-like curves because the filament is finite; it oscillates embedded in an inhomogeneous plasma; and due to the action of dissipative effects (numerical diffusion and damping, see Section \ref{Sub:diff}). Not all simulated cases exhibit a Lissajous-like pattern, specially Case 5 which is strongly damped.
\begin{figure}
\centerline{
\hspace*{-0.02\textwidth}
\includegraphics[width=0.53\textwidth]{c3traj.pdf}
\hspace*{-0.04\textwidth}
\includegraphics[width=0.53\textwidth]{c4traj.pdf}}
\caption{Filament trajectories during the oscillating stage corresponding to Case 3 (panel \textsf{(a)}) and Case 4 (panel \textsf{(b)}). The displacements are referred to the initial positions $(x,y)=(0,h)~$Mm. The sequential numeric annotations indicate the times (in min) corresponding to those positions helping to understand the temporal evolution.}
\label{f:traj}
\end{figure}
\subsubsection{Numerical Diffusion and Damping}
\label{Sub:diff}
It is important to verify that the filament attenuated oscillations (e.g., as can be appreciated from Figure~\ref{f:disp_osc} for Case 3) are not due to the artificial viscosity produced by truncation errors of the numerical scheme used, whose magnitude depends on the order of accuracy of the method and is generally reduced by an increase in the grid resolution. However, a direct evaluation of the numerical diffusion is not a simple task because the modified set of Equations \ref{e:mhd} (i.e., the governing equations affected by the truncation error terms of the numerical method used) is too hard to obtain for high-order MUSCL type schemes. A numerical estimation is also difficult since the diffusion effects are different along the domain due to gradient changes and variations of the spatial resolutions when using AMR meshes.
For the analysis we chose Case 3 as a reference run and performed simulations with different spatial resolutions to evaluate changes in the numerical results. If the damping time is a consequence of the numerical diffusion, it will be shorter for coarser grids because the numerical diffusion coefficient increases when the grid resolution is reduced. This means that the filament motion should be damped in fewer cycles for coarser grids than for finer ones. Figure~\ref{f:resol} displays the numerical results for the displacements using four different progressively refined grids ($\delta x=\delta y=[0.37, 0.19, 0.094, 0.059]~$Mm), corresponding to relative coarser grids of $300\%$, $100\%$ and $0\%$, with respect to the standard resolution of Case 3 (see Section \ref{Sub:code}), and a relative finer grid of $-40\%$. The AMR meshes used for the runs $300\%$, $100\%$ and Case 3 were of six, seven and eight levels of refinement, respectively, with $20\times 10$ cells per block; whereas for the run $-40\%$ was of eight levels with $32\times 16$ cells per block. Taking into account that the upward slopes exhibited by the vertical displacements $\Delta y$, which are smaller for higher grid resolutions, are due to the numerical imbalances between the filament and its background, not affecting the oscillatory part of the motion, we concluded that the numerical results of Case 3 are not significantly influenced by the artificial viscosity. In fact, we can see that there is no direct correlation between the damping level and the grid resolution since, for example, the motion in the $x$-axis for the reference run damps a bit faster than for the coarser ones. Therefore, for all studied cases using the reference grid resolution, we can assume that the numerical diffusion has a minimum impact on the damping process and can be reasonably ignored.
\begin{figure}
\centerline{
\hspace*{-0.02\textwidth}
\includegraphics[width=0.52\textwidth]{dx.pdf}
\hspace*{-0.04\textwidth}
\includegraphics[width=0.52\textwidth]{dy.pdf}}
\caption{Test to determine weather the damped motion exhibited by the filament is due to numerical diffusion or it has a physical origin. Four runs were performed varying the spatial resolution, corresponding to relative coarser grids of $300\%$, $100\%$ with respect to the standard resolution of Case 3, and a relative finer grid of $-40\%$. We concluded that the acting damping process has a significant physical meaning. \textsf{(a)--\textsf{(b)}} Show horizontal and vertical displacements for the different resolutions.}
\label{f:resol}
\end{figure}
With regards to the filament attenuated oscillations, since we are not considering non-adiabatic processes (as radiative losses or heat conduction) and according to the geometry of the problem, we discuss two main possible damping mechanisms: resonant absorption and wave leakage.
First, we consider the resonant absorption of infinitely long thread oscillations \citep[see e.g.,][and references therein]{2018arregieLRSP15}. This mechanism relies on the energy transfer from transverse kink modes to small scale Alfv\'en waves because of the plasma inhomogeneity located at the transition layer between the filament and the surrounding corona. Assuming a small transition layer ($\Delta/R <<1$), analytical expressions for the damping time scale were obtained for the kink modes \citep[see e.g.,][]{1992SoPh..138..233G,2002A&A...394L..39G,2002ApJ...577..475R}, claiming a linear dependence between periods and damping times. In order to explore a possible relation, we fit a linear function $\tau=aP+b$ using the values in Table \ref{t:fits}. Figure~\ref{f:damping}\textsf{(a)} shows the damping times against the periods for horizontal (circles) and vertical (diamond) oscillations. Case 5 is pointed out with a cross and an asterisk. The error bars are in grey colour, for the periods these errors are comparable to the symbol size. Small circles and diamonds correspond to cases with filament radius $R=2.5~$Mm and big symbols to $R=3.5~$Mm. The black-dashed line displays the linear fit with $a=(1.6\pm0.5)$ and $b=(20.8\pm 13.7)~$min. For the analysis we excluded the damping times of Cases 6 and 7 because they are out of range. This crescent linear tendency looks similar to the behaviour by \citet{2011A&A...531A..53H} suggesting that the resonant absorption might be acting as a damping mechanism. In addition as a complementary study, taking into account the work of \citet{2016ApJ...820..125T} in which they found an energy enhancement in a thin region near the filament's edge attributed to resonant absorption, we explored the evolution of the kinetic energy in a shell surrounding the filament's core ($R_2<r<5~ {\rm Mm}$). The results did not show a definite kinetic energy enhancement associated with the surrounding shell to support the linear tendency found in Figure~\ref{f:damping}\textsf{(a)}.
Thus, more studies are required to investigate to what extent the resonant absorption may be responsible for the oscillation damping.
\begin{figure}
\centerline{
\hspace*{-0.02\textwidth}
\includegraphics[width=0.49\textwidth]{tau_vs_p.pdf}
\hspace*{-0.02\textwidth}
\includegraphics[width=0.5265\textwidth]{pmap_y3565_vs_t.pdf}}
\caption{\textsf{(a)} Damping times vs periods for the cases exhibited in Table~\ref{t:fits} except Cases 6 and 7. The values for the horizontal (vertical) direction are displayed with circle (diamond) symbols. Case 5 is displayed with cross symbols. The widths of symbols are according to the radius of the filament. The grey lines correspond to the errors of the variables. The black-dashed line shows a linear fit. \textsf{(b)} Pressure stack plot of Case 3 along the vertical dash line in Figure~\ref{f:pres_patt}\textsf{(d)}. The black solid lines outline the speed of slow waves propagating away from the filament. The colour map represent $ln(p)$}
\label{f:damping}
\end{figure}
Second, to study weather the wave-fronts observed emanating from the filament (such as in Figures \ref{f:pres_patt}\textsf{(c)}--\textsf{(d)}) are evidence of wave leakage, we constructed for Case 3 a pressure stack plot for each time measured along a vertical slice located at $(x,y)=(0,35$--$65)~$Mm (highlighted in Figure~\ref{f:pres_patt}\textsf{(c)} by a dash line). The stack plot results are shown in Figure~\ref{f:damping}\textsf{(b)} where wave signatures propagating with speeds of $\sim$52~km s$^{-1}$ at an early time, and later with $\sim$25~km s$^{-1}$ and $\sim$8~km s$^{-1}$ can be appreciated. Also, the periods of these waves, which can be deduced from the separation in time between the fronts, is of $\sim$20~min, hence associated with the periods of the filament oscillation. On the slice domain the spatial average sound speed $\bar{c}_{\rm s}=150~$km s$^{-1}$, the average Alfv\'en speed $\bar{v}_{\rm A}=185~$km s$^{-1}$ and the average plasma $\bar{\beta}=0.94$. From Figures~\ref{f:pres_patt}\textsf{(c)}--\textsf{(d)} and \ref{f:damping}\textsf{(b)} we identify compressive, slow magneto-acoustic waves emanating from the filament with an oblique propagation angle with respect to the magnetic field, which seems to be the reason why the wave-front speeds are markedly smaller than the sound speed. Thus it seems that, while the filament oscillates, it acts as a forcing driver transferring energy into the surrounfing plasma. On the other hand, during the oscillating stage the filament lost into the environment $\sim$3$\%$ of its mass. We thus conclude that the main process responsible of the oscillation damping seems to be the wave leakage.
\begin{table}
\caption{Set-up and fitted parameters for all numerical cases. Varying parameters: the initial filament height $h$, filament radius $R$, azimuthal current density $j_1$ and maximum simulation time $t_{\rm {max}}$. $\bar{\rho}_{\rm fil}$ is the averaged filament density. Fix parameters for all cases: the axial current density $j_0$, density at the coronal base $\rho_0$, filament transition layer $\Delta$, dipole intensity $M_{\rm{dip}}$, coronal temperature $T_0$, chromospheric temperature $T_{\rm cr}$, filament temperature $T_{\rm{fil}}$, chromospheric width $h_{\rm{cr}}$ and transition region width $h_{\rm{tr}}$. Fitted parameters in the horizontal $x$-axis: the period $P_x$, displacement amplitude $A_x$, the damping time $\tau_x$ and the speed amplitude $A_{v_x}$. Same is displayed for the vertical $y$-axis.}
\label{t:fits}
\begin{tabular}{lccccccc}
\hline
& \tiny{Case 1} & \tiny{Case 2} & \tiny{Case 3} & \tiny{Case 4} & \tiny{Case 5} & \tiny{Case 6} & \tiny{Case 7}\\
\hline
\underline{Set-up parameters}& & & & & & & \\
$h~$[Mm] & 14. & 14. & 19. & 19. & 19. & 25. & 25. \\
$R~$[Mm] & 2.5 & 3.5 & 2.5 & 3.5 & 2.5 & 2.5 & 3.5 \\
$j_1~$[statA cm$^{-2}$] & 600. & 600. & 600. & 600. & 797.5 & 600. & 600. \\
$t_{\rm {max}}~$[min] & 120. & 120. & 120. & 120. & 120. & 150. & 150. \\
$\bar{\rho}_{\rm fil}~$[g cm$^{-3}\times 10^{-13}$] &4.0&7.2&4.0&7.2&2.5&4.0 &7.1 \\
$j_0~$[statA cm$^{-2}$] & \multicolumn7c{600.} \\
$\rho_0~$[g cm$^{-3}$] & \multicolumn7c{$3.2\times 10^{-15}$} \\
$\Delta~$[Mm] & \multicolumn7c{1.25} \\
$M_{\rm{dip}}$ & \multicolumn7c{$2.25$} \\
$T_0~$[MK] & \multicolumn7c{1.} \\
$T_{\rm cr}~$[K] & \multicolumn7c{$1.2\times 10^4$} \\
$T_{\rm{fil}}~$[MK] & \multicolumn7c{0.3} \\
$h_{\rm{cr}}~$[Mm] & \multicolumn7c{2.5} \\
$h_{\rm{tr}}~$[Mm] & \multicolumn7c{0.5} \\
\hline
\underline{Fitted parameters} & & & & & & & \\
& & & &\underline{$x$-axis} & & & \\
$P_x~$[min] & 21.0 & 22.6 & 44.9 & 35.3 & 26.9 & 66.9 & 51.7 \\
$A_x~$[Mm] & 6.0 & 1.5 & 11.2 & 3.1 & 15.2 & 22.2 & 6.7 \\
$\tau_x~$[min] & 22.1 & 56.2 & 63.6 & 92.8 & 16.1 & 197.2 & 354.4 \\
$A_{v_x}~$[km s$^{-1}$] & 19.8 & 7.2 & 27.1 & 9.9 & 36.0 & 29.9 & 13.5 \\
& & & &\underline{$y$-axis} & & & \\
$P_y~$[min] & 13.7 & 12.5 & 18.8 & 16.1 & 15.6 & 27.5 & 23.0 \\
$A_y~$[Mm] & 2.7 & 1.4 & 3.3 & 2.2 & 5.1 & 2.6 & 2.9 \\
$\tau_y~$[min] & 32.7 & 47.5 & 45.7 & 27.1 & 17.4 & 66.8 & 53.7 \\
$A_{v_y}~$[km s$^{-1}$] & 15.6 & 8.9 & 14.5 & 9.1 & 18.7 & 6.7 & 9.7 \\
\hline
\end{tabular}
\end{table}
\subsubsection{Restoring Forces}
\label{Sub:force}
\begin{figure}
\centerline{
\hspace*{-0.01\textwidth}
\includegraphics[width=0.53\textwidth]{Fx.pdf}
\hspace*{-0.04\textwidth}
\includegraphics[width=0.53\textwidth]{Fy.pdf}}
\caption{Forces acting on the oscillating filament. \textsf{(a)} Averaged forces per unit volume in the horizontal direction: a reference square-black-dashed line depicts the displacement with its scale modified; triangle-up-blue-solid line represents the thermal pressure; circle-red-solid line the magnetic pressure; pentagon-cyan-solid line the magnetic tension and asterisk-green-solid curve the net force. \textsf{(b)} Similar to panel (a) but corresponding to the vertical component: square-black-dashed line represents vertical displacement with a modified scale; triangle down-orange-solid line is the almost constant weight force.}
\label{f:fuerzas}
\end{figure}
Figures \ref{f:fuerzas}\textsf{(a)}--\textsf{(b)} display the calculation of the thermal pressure force $-\mathbfit{\nabla}p$ (blue lines) and the Lorentz force $\frac{1}{c}\mathbfit{j}\times\mathbfit{B}$, decomposed into the magnetic tension $\frac{\mathbfit{B}\cdot\mathbfit{\nabla}\mathbfit{B}}{4\pi}$ (cyan lines) and magnetic pressure force $-\mathbfit{\nabla}(\frac{B^2}{8\pi})$ (red lines), along the horizontal and vertical axes, respectively. In green lines we represented the net forces. Also, as a reference, in black lines the filament displacements were added with an appropriate scale, to facilitate a visual comparison. The Figure~\ref{f:fuerzas} corresponds to Case 3 and displays the averaged forces over the grid-cells belonging to the filament structure ($r<R_3$) evolving in time. With respect to the horizontal direction, panel \textsf{(a)} shows that, before the interaction begins (by $t<33~$min), all forces are almost null, as expected. Then the forces behave noisily during a very short time lapse, attributed to the strong interaction with the shock waves. Later on, the magnetic tension results the main restoring force (more smoothly drawn by the net force in green colour). Meanwhile, for early times, the thermal pressure force seems to follow the displacement and the magnetic pressure one does not play a significant role. Regarding the forces acting on the vertical direction, panel \textsf{(b)} shows that the magnetic tension is always positive, pointing upwards, opposing the almost constant and negative weight force (orange line). It is noticeable the correspondence between the magnetic tension and the vertical displacement. Note that the main restoring force in the upward direction is the magnetic tension, while in the downward direction is the weight. In turn, the thermal and magnetic pressure forces play secondary roles, the thermal pressure force becomes negative for later times, while the magnetic pressure forces pushes upwards until finally vanishes, by $t\approx 80~$min.
\subsection{Parametric study}
Figure~\ref{f:results} summarised the fitted parameters of all cases for the horizontal and vertical displacements (also listed in Table~\ref{t:fits}). Left (right) panel corresponds to the $x$-($y$-)axis. In each panel, circles indicate periods as a function of height, and the colour map describes the amplitudes of oscillation. Small circles represent the triad of Cases 1--3--6 with filament small radius $R=2.5~$Mm, connected by dashed lines; while the big circles indicate the triad of Cases 2--4--7 with large radius $R=3.5~$Mm, linked by dotted lines. We also show the damping times, depicted by squared symbols following equal logic. Case 5, with a different filament mass, is highlighted with different symbols, a star for the period and a diamond for the damping time.
In the following we itemise the main trends obtained for the seven cases. Beginning with the $x$-direction represented by Figure~\ref{f:results}\textsf{(a)}:
\noindent
\begin{enumerate}
\item Comparing cases of same height but different radii: large radii are generally associated with smaller periods, smaller amplitudes (darker blue) and larger damping times than small radius cases. This tendency is seen comparing the pairs of Cases 1--2, 3--4, and 6--7, except for the pair 1--2 of similar periods.
\item Comparing cases of same radius but different heights: larger heights are associated with larger periods, larger amplitudes and larger damping times. This occurs for the triads of Cases 1--3--6 and 2--4--7.
\item Comparing cases of same radius and height but different filament masses: Case 5 shows a smaller period, a larger amplitude and a considerable smaller damping time than the heavier Case 3.
\end{enumerate}
With respect to the $y$-direction (see Figure~\ref{f:results}\textsf{(b)}), the fitted parameters show similar trends as just described for the $x$-direction but having a less regular behaviour:
\begin{enumerate}
\item Comparing cases of same height but different radii: the pairs of Cases 1--2, 3--4 and 6--7 show that the increase of the radius implies smaller periods, smaller amplitudes (except the pair 6--7) and smaller damping times (except the pair 1--2).
\item Comparing cases of same radius but different heights: the triads of Cases 1--3--6 and 2--4--7 show that the increase in height commonly denotes larger periods, larger amplitudes (except the pair 3--6) and larger damping times (except the pair 2--4).
\item Comparing cases of same radius and height but different filament masses: Case 5 exhibits a smaller period, a larger amplitude and a quite smaller damping time than the heavier Case 3.
\end{enumerate}
Also, comparing the results between the horizontal and vertical oscillations, we note that the horizontal amplitudes, periods and damping times are larger than the vertical ones, in agreement with \citet{1999A&A...345.1038S}.
\begin{figure}
\centerline{
\hspace*{0.0\textwidth}
\includegraphics[width=0.51\textwidth]{Ax.pdf}
\hspace*{-0.02\textwidth}
\includegraphics[width=0.52\textwidth]{Ay.pdf}}
\caption{Fitted parameters as a function of height for all cases. Panels \textsf{(a)}--\textsf{(b)} show periods, amplitudes of oscillation and damping times obtained for the horizontal and vertical displacements, respectively. The circles indicate the periods and the colour map exhibits the amplitudes. Small circles connected with a dashed line represent the cases with lower filament radius, and big circles, joined by a dotted line, identify cases with larger radius. The squares display the damping times. The squared symbol sizes follow the same rules as the circles. The period for Case 5 is pointed out with a star and the damping time with a diamond symbol.}
\label{f:results}
\end{figure}
Summarising, there is a general tendency of periods, displacement amplitudes and damping times to increase with height. This tendency is clearly observed for horizontal oscillations, although it is not always valid for vertical ones. The same behaviour is followed by the velocity amplitudes. The general tendency to increase with height is in agreement with the fact that the magnitude of the net force acting on the filament, along with its restoring frequency, diminish with height (the net force frequency can be appreciate, e.g., in Figure~\ref{f:fuerzas}).
To increase the radius in general implies shorter periods and smaller amplitudes. But, while it clearly implies longer damping times considering the horizontal oscillations, we found shorter damping times when considering vertical oscillations. Note that the increase of the radius produces an increase on the net force frequency acting over the filament, which explains the shorter periods found.
Finally, exploring variations of the filament mass, while keeping fixed the remaining independent parameters of the set-up, the lower mass Case 5 exhibits shorter periods and larger amplitudes in both axes, although more noticeable is that Case 5 shows a stronger damped motion.
As exposed above, the horizontal and vertical oscillation properties of the filament are, in general, different. Thus, in observations using the Doppler technique to measure oscillation properties, projection effects along the line-of-sight may play a significant role in the results. For instance, in the framework drawn in Figure~\ref{f:esquema} a particular observer whose line-of-sight is that of a prominence seen at the limb ($\phi=0$) will measure the horizontal component of the oscillation; whereas, an observer with a top-down view of a filament at the centre of the solar disc ($\phi=90$) will measure the vertical component. The Doppler velocity measured by intermediate observers will correspond to the filament velocity vector projected along the line-of-sight of the observer. Therefore, as the winking phenomenon related to the Doppler shift is more likely to be detected in filaments with large velocity amplitudes, in the framework of our results this will correspond to Cases $[1,3,5,6]$ and with line-of-sights near the limb, that is when the horizontal velocity component, which is frequently larger than the vertical one, contributes more to the Doppler effect.
\section{Conclusions}
\label{S-Conclusion}
In order to contribute to the understanding of the oscillatory phenomenon of quiescent filaments, we performed a parametric study of large-amplitude transverse oscillations carrying out 2.5D ideal MHD simulations. The model of \citet{1990forbesJGR95} was adapted to numerically represent a filament floating in a gravitationally stratified atmosphere. After the set-up initialisation and a relaxation process, the quiescent filament was perturbed using a blast device that excites coupled horizontal and vertical oscillations for a few cycles.
By analysing the forces, even though our model differs from other authors, either in the magnetic configuration (Kippenhan and Schl\"uter or Kuperus and Raadu) or in the perturbation mechanism, we also found that the main restoring force is the magnetic tension. As suggested by \citet{2014shenApJ795}, along the vertical direction the main restoring force is the coupling between the gravity and magnetic tension. Whereas, along the horizontal direction the main force is the magnetic tension. In agreement with \citet{2018zhouApJ856} we found that the magnetic pressure plays a secondary role and the gas pressure contribution is negligible.
The oscillation properties present different behaviours along the horizontal and vertical axes. The periods, amplitudes and most of the damping times are larger in the horizontal direction. The filament exhibits a vertical oscillation frequency being approximately twice larger than the horizontal one (listed in Table~\ref{t:fits}), corresponding to coupled transverse oscillations, which in some cases seem close to a resonant regime (see Figure~\ref{f:traj}) indicated by the resemblance of the trajectory pattern with the Lissajous-like curves obtained by \citet{2018kolotkovJASTP172}. This behaviour may indicate that the vertical motion is a reaction to the horizontal oscillation, as pointed out by \citet{2020liakhA&A637}.
Also, our results (Cases 1 and 5) are in good agreement with the results in \citet{2009tripathiSSRv149} and \citet{2014shenApJ795} who found a range of periods of 11--29~min, velocity amplitudes of 6--41~km s$^{-1}$, and damping times of 25--180~min. Whereas Cases 6 and 7 keep some agreement with the events studied by \citet{2015RAA....15.1713P}, who reported values of 61--67~min, 12--22~km s$^{-1}$ and 92--117~min, respectively.
In order to explore the role of different parameters on the oscillatory motion, a parametric study was carried out performing seven simulations varying the height, size and mass of the filament. The fitted parameters revealed, as a first general rule, that periods, amplitudes and damping times increase with height (Figure~\ref{f:results}). This tendency is clearly observed in the horizontal direction, but has some exceptions in the vertical one. Given that the net force exerted onto the filament decreases with height and its restoring frequency as well, this results in larger periods and amplitudes measured on higher filaments.
This tendency of periods to increase with height partially agrees with the analytical model of \citet[][see their figure 3]{2016kolotkovA&A590} who considered uncoupled linear oscillations. For both the vertical periods increase with height, whereas for the horizontal periods these authors found a tendency to diminish with height.
Our first general tendency is in agreement with observational cases, for example: \cite{2013franchileAA552} analysed a filament with a low height of $\sim$7~Mm \citep{2008gilbertApJ685} oscillating with a short period of $\sim$4 min; \cite{2014shenApJ795} analysing a higher filament ($40~\textrm{Mm}$) measured a larger period of $\sim$13 min and an amplitude of $10~$Mm; moreover, \cite{2013ApJ...773..166L} reported oscillations of an even higher filament ($63~$Mm) with an even larger period $\sim$30 min and an amplitude of ($22~$Mm). Considering the damping time, this trend is also seen in observational studies: \citet{2012ApJ...761..103G} for a prominence at a height of $36~$Mm reported a transverse oscillation period of $28~$min and a damping time of $44~$min. Following the increasing tendency, \cite{2013ApJ...773..166L} measured global oscillations in a filament of height $63~$Mm determining a period of $30~$min and a damping time of $47~$min. This trend is also followed by the prominence studied at EUV by \cite{2011A&A...531A..53H}, who for a height of $80~$Mm reported a period of $100~$min and a damping time of $240~$min.
As a second general rule, in both directions the results also showed that larger filaments are associated to shorter periods and smaller amplitudes. As before, this tendency partially agrees with the linear model of \citet{2016kolotkovA&A590}; taking into account that in our model an increase of the filament radius implies an increase of the filament current $I$ (Equation \ref{e:curr}) these authors obtained that vertical periods diminish for larger filament currents $I$ but, conversely, horizontal periods increase with current $I$. On the other hand, this tendency is more difficult to compare with observations due to the lack of joint measurements of radius, period and amplitude. It seems that as larger filaments are subject to stronger net forces of higher restoring frequencies, this would imply that the filament oscillates with shorter periods and smaller amplitudes.
Concerning the damping mechanism obtained from the simulations, we examined the main processes discussed by several authors and found that the wave leakage seems to be the main damping mechanism.
\begin{acks}[Acknowledgements]
We want to thank an anonymous referee who helped us make a significant improvement to this work. EZ is grateful with FAPESP to have financed this research by the grant 2018/25177-4. MC, GK and AC are members of the Carrera del Investigador Cient\'ifico (CONICET). MC and GK acknowledge support from ANPCyT under the grant PICT No. 2016-2480. MC also acknowledge support from the SECYT-UNC grant No. 33620180101147CB. MVS thanks support from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No 724326). GGC thanks CNPq for support with a Productivity Research Fellowship. The research leading to these results has received funding from CAPES grant 88881.310386/2018-01, FAPESP grant 2013/24155-3. EZ and GGC are also grateful with Mackenzie Research Funding Mackpesquisa for the received support. GGC is Correspondent Researcher of the Consejo Nacional de Investigaciones Cient\'ificas y T\'ecnicas (CONICET) for the Instituto de Astronom\'ia y F\'isica del Espacio (IAFE), Argentina. Simulations were ran on the IATE's clusters and CRAAM's cluster {\sc wintermute}, we thank system managers Dar\'io Gra\~na and Tiago Giorgetti. {\sc Flash} code: ``The software used in this work was developed in part by the DOE NNSA ASC- and DOE Office of Science ASCR-supported Flash Center for Computational Science at the University of Chicago''. We also thank the {\sc VisIt} team for developing the graphical tool \citep{2012Harrison2012python}.
\end{acks}
\bibliographystyle{spr-mp-sola}
|
\section{Introduction}
With the increasing concern for data privacy and security of personal data, the demand for creating trusted methods of privacy-preserving communication is rapidly increasing. Recent advances in machine learning have also spurred interest in a variety of privacy-sensitive domains such as healthcare informatics \cite{ref:healthcare-fl} and coronavirus disease 2019 (COVID-19) diagnoses \cite{ref:covid-19-diagnosis-fl}. However, the privacy of personal data used to train these models has been called into question. Traditional methods of training require data to be centrally located. However, this is challenging or impossible with privacy regulations.
Federated learning (FL) \cite{ref:federated-learning} is a proposed solution allowing sensitive data to remain distributed among edge devices. This framework will enable devices to train models locally on-device and then upload their local model updates to a trusted central server to be aggregated, resulting in a global model that is then distributed. This process is repeated until the desired objective is reached, such as the convergence of model accuracy.
Additional improvements have been made to the aggregation strategies involved in traditional FL, notably in reducing the communication overhead to the central aggregation server. Solutions such as HierFavg \cite{ref:hierfavg} introducing a hierarchical scaling solution, where model updates are instead sent to local edge servers performing partial model aggregation. This combines the benefits of both centralized cloud and edge computing, resulting in both reduced latency and communication overhead.
While the traditional FL framework has demonstrated its viability in real-world applications, it has several notable issues. First, model updates can indirectly leak the local data used to train them. Specifically, users can be de-anonymized through data reconstruction via statistical analysis and data mining. An example of this was with the Netflix Prize Dataset \cite{ref:deanonymize-netflix} where data was correlated to users using an auxiliary IMDb dataset. The second issue is the centralized nature of aggregating the model updates, which creates a single point of failure and relies on trusted authorities.
Differential privacy \cite{ref:differential-privacy, ref:differential-privacy-impossibility}, a mathematical framework for evaluating dataset privacy of a given model, can be employed to solve the first issue, for example, by adding a certain noise to the model update. A solution to the second issue is to utilize blockchain technologies to decentralize the task of aggregating the local model updates. However, using a decentralized system incurs several more issues by removing the presence of a centrally trusted entity which gives rise to, for example, the problem of detecting malicious clients sending false model updates.
Additionally, because local datasets do not necessarily represent the global distribution, as they are non-IID (independent and identically distributed), two honest clients could submit opposing model updates. This problem can be addressed by validating each transaction. One possible solution is to allow each node to execute validation procedures in a smart contract, as described in BAFFLE \cite{ref:baffle}. However, delegating this task to every node of the network is computationally expensive considering the high overhead, local model training, and the consensus task \cite{ref:committee-consensus}. Another solution is to select a committee of nodes each round to perform this validation task \cite{ref:committee-consensus}.
We can define the current number of clients training each round as $\mathcal{C}$, and the number of committee nodes (endorsing peers) as $\mathcal{P}_E$. In each round, a total of $\mathcal{C} \times \mathcal{P}_E$ computations must be completed to validate all transactions.
Here, we propose a sharding mechanism extending the committee consensus method. As the network expands, clients will be assigned to different shards within the network, and a committee will be elected for each shard. After every phase, each shard's committee will publish a local shard-aggregated model, which will then be globally aggregated and redistributed to all clients. If we have $\mathcal{S}$ many shards, this would reduce the global computations from $\mathcal{C} \times \mathcal{P}_E$ to $\frac{\mathcal{C} \times \mathcal{P}_E}{\mathcal{S}^2}$ for each shard, or $\frac{\mathcal{C} \times \mathcal{P}_E}{\mathcal{S}}$ globally.
To preserve the security of the system across shards, we propose a flexible protocol allowing for a pluggable poisoning mitigation, and detection strategies \cite{ref:roni, ref:krum, ref:fools-gold, ref:attack-adaptive-aggregation}. This allows proposed models to adopt suitable strategies, while allowing the system evolves with recent advances.
Our {\em key contributions} of this paper can be summarized as follows:
\begin{itemize}
\item a \textbf{sharding framework} called ScaleSFL, which supports a sharding solution for blockchain-based federated learning. We formally discuss the design of our algorithm for the sharding mechanism based on two consensus levels: shard-level consensus and mainchain consensus.
\item an \textbf{implementation} of a proof-of-concept prototype system using Hyperledger Fabric \cite{ref:fabric} to run local experiments and evaluate the solution effectiveness.
\item \textbf{benchmarks} that we demonstrate the performance of our implementation by using Hyperledger Caliper benchmarking tools \cite{ref:hyperledger-caliper}, showing our framework scales linearly with the number of shards.
\end{itemize}
The remainder of this paper is organized as follows. The next section provides background and related works. Our ScaleSFL framework, a sharding solution for blockchain-based federated learning, is introduced in Section~\ref{sec3}. Evaluation and discussion are presented in Sections~\ref{sec4} and \ref{sec5}, respectively. Finally, we conclude our work, and describe future works in Section~\ref{sec6}.
\section{Background and Related Works}
Our current work integrates the ideas of federated learning and differential privacy, as well as blockchain-based federated learning. In this section, we introduce individual topics and corresponding hybrid solutions to provide background.
\subsection{Federated Learning and Differential Privacy}
Privacy-preserving machine learning combines the use of federated learning with differential privacy \cite{ref:privacy-preserving-deep-learning, ref:differential-privacy-sgd, ref:cit21}. Federated learning allows sensitive data from edge devices (e.g., smartphones, tablets, IoT devices) to remain on-device while training is done locally (distributed training on edge devices) \cite{ref:federated-learning}. In the original FL environment, a centralized server manages the training process. The server broadcasts the current model during each round to a subset of participating nodes to activate local training (requesting a model update). After all model update submission responses are received, the server aggregates the local models into a global model and prepares for the next round \cite{ref:federated-learning, ref:committee-consensus}. While FL offers many privacy benefits, differential privacy in conjunction provides stronger privacy guarantees \cite{ref:federated-learning} by mitigating data reconstruct-ability via random model update noise.
McMahan et al. \cite{ref:federated-learning} proposed the Federated Averaging algorithm (FedAvg). Their algorithm is a generalization of federated stochastic gradient descent (FedSGD) \cite{ref:differential-privacy-sgd} allowing each node to perform multiple batch updates while exchanging the models' parameters, with which we based our FL model training approach. Experiments demonstrate that FL can be used for training a variety of high-quality machine learning models at reasonable communication cost, exemplifying high practicality \cite{ref:federated-learning}. Therefore, we choose to apply differential privacy and use an adapted version of FedAvg for our current work.
Liu et al. \cite{ref:hierfavg} introduced HierFavg, a hierarchical scaling solution introducing multiple edge servers, each performing partial model aggregation, combining the benefits of both centralized cloud and edge computing. While centralized servers benefit from the vast reach of data collection, the communication overhead and latency of downloading and evaluating such updates limit this approach. Instead, model updates are sent to local edge servers to benefit from fast and efficient model updates, reducing the number of local iterations and total runtime. These local updates are then sent to a cloud server allowing for a sampling of the global data distribution to be considered and outperforming the local edge server performance. Combining these approaches is nontrivial, and the authors performed convergence analysis describing qualitative guidelines in applying their approach. While our approach is primarily motivated by increasing the throughput of model update metadata to blockchain ledgers, the composition of shards can benefit from the strategies found in hierarchical solutions.
\subsection{Blockchain-Based Federated Learning}
Blockchain technology can be utilized to decentralize the aforementioned centralized server in federated learning. The aggregation and facilitation logic previously done by the central server can be replaced by implementing smart contracts \cite{ref:committee-consensus}.
In addition to Ramanan and Nakayama's blockchain-based aggregator free federated learning (BAFFLE) \cite{ref:baffle} and Li et al.'s blockchain-based federated learning framework with committee consensus \cite{ref:committee-consensus} referenced earlier, numerous strategies for integrating blockchain technology with federated learning to achieve decentralization have been considered \cite{SatybaldyN20}. However, the approaches proposed by previous works do not directly focus on the performance of their chosen underlying consensus mechanism with which their proposed frameworks are built upon.
\subsubsection{Blockchain Integration}
Kim et al.'s blockchain-based federated learning (BlockFL) \cite{ref:bc-on-device-fl} was implemented using the proof-of-work consensus mechanism and focused on analyzing the end-to-end latency of their system. Ma et al.'s proposed blockchain-assisted decentralized FL (BLADE-FL) \cite{ref:fl-meets-bc} was able to address the single point of failure issue while analyzing the learning performance. Liu et al.'s proposed federated learning with asynchronous convergence (FedAC) method \cite{ref:bfl-async} seeks to improve communication performance through integrating asynchronous global aggregation. While these works focus on the performance of various system components, they do not consider the scalability requirements for large numbers of participants and the impact on the consensus mechanism's throughput.
Additionally, many blockchain-based federated learning applications with domain-specific frameworks have also been proposed. For example, healthcare \cite{ref:bfl-healthcare}, IoT \cite{ref:bfl-iot} and 6G networks \cite{ref:bfl-6g}. Similarly, these works focus on cutting-edge implementation plausibility rather than scalability or details regarding the performance of their consensus mechanisms.
Gai et al. \cite{ref:bfl-iiot} proposed a model that utilizes a blockchain-based solution to develop a privacy-preserving environment for task allocations in a cloud/edge system. Employing blockchain enhances trustworthiness in edge nodes and security in communications. Utilizing smart contracts also contributes toward reaching optimal task allocations. The model applies a differential privacy algorithm (Laplace distribution mechanism \cite{ref:differential-privacy}) to process energy cost data in edge nodes.
Li et al. \cite{ref:committee-consensus} proposed a blockchain-based FL framework with committee consensus in response to consensus mechanism efficiency and framework scalability. The committee consensus mechanism works as follows: A subset of nodes constitutes a committee responsible for validating model updates and new block generation. Meanwhile, the other nodes only execute their local model training and send their updates to the committee nodes. This committee is re-elected every round based on scores from the previous round, or alternatively, re-election is randomized for implementation simplicity. Li et al. demonstrated that committee consensus is more efficient and scalable while remaining secure.
\subsubsection{Sharding Solutions}
Sharding allows for scalability by splitting the network into smaller subgroups. This reduces storage and communication requirements while increasing throughput as the number of shards increases.
Yuan et al. \cite{ref:bfl-shard} proposed ChainFL, a two-layered blockchain-driven FL system that splits an IoT network into multiple shards. Also using Hyperledger Fabric \cite{ref:fabric}, ChainFL adopts a direct acyclic graph (DAG)-based mainchain to achieve parallel (and asynchronous) cross-shard validation. While their approach is DAG-based, we focus on a hierarchical FL solution, and are able to use existing off-chain FL frameworks to increase interoperability.
\subsection{Model Update Poisoning}\label{sec:model-defence}
A notable issue when aggregating model updates from untrusted and potentially adversarial clients are model poisoning attacks. In order to disrupt or bias the training process, and adversary may submit biased, or corrupted model updates. Because model updates must be verified prior to being accepted to the ledger, we consider each defence mechanism applied to each endorsing peer rather than a single centralized server. One possibility to make the training more robust is to apply norm constraints to each model update, such as clipping model updates and applying random noise to the aggregated model to prevent overfitting \cite{ref:fl-survey}. However, further mechanisms are required to counter adversarial examples and backdoors \cite{ref:fl-backdoor}. A subset of model poisoning is data poisoning attacks. These occur when the clients' dataset has been altered, such as a targeted attack by adversarial manipulation of edge devices, or untargeted such as a faulty data collection process, resulting in corrupted samples (e.g., incorrect labels). Here, we focus on readily applicable works related to the general detection of poisoned model updates.
Barreno et al. \cite{ref:roni} proposed a technique to determine which samples from a training set are malicious. To do so, they present the reject on negative influence (RONI) defence. This method measures the effect each training sample has on the accuracy of the model. This defence strategy is appropriately modified to suit the needs of a FL setting. Instead, we measure the influence of each model update on the accuracy of the global model. This can be done by utilizing a held-out testing set, potentially unique to each endorsing peer. This defence is not suitable for general FL applications due to the nature of client data being non-IID and incorrectly rejected due to misrepresentation by the test set. Additionally, malicious updates may go unnoticed due depending on the threshold of acceptance.
Blanchard et al. \cite{ref:krum} introduced a method designed to counter adversarial updates in a FL setting. Multi-Krum is a byzantine-resilient defence strategy that selects $n$ vectors furthest from the calculated mean, removing them from the aggregated gradient update. These distances are calculated based on the euclidean distance between gradient vectors. This strategy is effective for attacks with up to 33\% compromised/adversarial clients. However, the calculated mean can be influenced by the Sybil updates and may fail for an attacker with more significant influence over the network.
Fung et al. \cite{ref:fools-gold} proposed a novel defence strategy to discriminate between honest clients and Sybils based on the diversity of their respective gradient updates. The intuition behind this defence is to identify Sybils based on a shared objective, and thus their models' updates are likely to have low variance. This works by using a cosine similarity, looking for similarity in terms of indicative features (i.e., those that contribute to the correctness of the update). This approach can be further augmented with other defence methods such as Multi-Krum to determine the poisoned updates.
Wan and Chen \cite{ref:attack-adaptive-aggregation} introduced an attack-adaptive defence utilizing an attention-based neural network to learn plausible attacks. Instead of using similarity-based mechanisms, this introduces tailored defence strategies to prevent backdoor attacks.
Ma et al. \cite{ref:dp-data-poisoning} explored differential privacy as a data poisoning defence based on creating a robust learning algorithm. While this defence is ideal for systems already using differential privacy, it enforces the use of noise in the learning algorithm, hurting performance. Additionally, this assumes a small subset of malicious clients, failing to defend against larger coordinated attacks.
\begin{figure*}[!th]
\centering
\includegraphics[width=0.9\linewidth]{sharded-tasks.png}
\caption{Global overview of our FL sharding approach}
\label{fig:sharding-architecture}
\end{figure*}
Related attacks, such as model update plagiarism \cite{ref:fl-meets-bc}, may additionally occur, leading to a misleading distribution of data amongst clients. A malicious user may attempt to submit the same model update multiple times through several clients, known as lazy clients. They may then attempt to reap the rewards of contribution for each of these submissions. These can be detected by using a PN-sequence \cite{ref:fl-meets-bc, ref:blade-fl-analysis}, whereby the initial model update is submitted with random noise applied to the update and is then further verified by other clients using their own PN-sequence, checking for correlation.
\section{Our Sharding Solution}\label{sec3}
This section describes our approach and outlines the architecture for our proposed sharding solution. An outline of the workflow can be seen in Figure~\ref{fig:sharding-architecture}. We can break this into two primary components, the mainchain, which coordinates aggregated models from shards, and the individual shard chains.
Our framework follows a few core principles. Our solution is modular such that (a) the policies that govern model acceptance are abstracted, for example, model update poisoning, and mitigation strategies (Section~\ref{sec:model-defence}), are supported to determine model acceptance. Our framework additionally (b) supports external FL coordination, such that new paradigms such as data-centric FL are supported. This means the source of models updates is agnostic.
With these principles, our solution acts as an additional component to existing FL workflows such that it handles model provenance, the evolution of security and poisoning defences, scalability through sharding, and the possibility of reward distribution for future works. Therefore, we solve the initial problem of FL systems being dependent on a central service to aggregate updates, providing a trusted system to pin verified model updates. We will further discuss the components of this system below.
\subsection{FL Algorithm}
The underlying problem of federated learning is to produce a global model from client updates trained on local privacy-sensitive datasets. We can define each client $\mathcal{C}_k$ to have their own set of data $D_k \subseteq D, \forall k, 0 \leq i \leq K$. Using this, we can define the global objective to optimize as a modification of the FedAvg algorithm \cite{ref:federated-learning} as
\begin{equation} \label{eq:global-shard-objective-problem}
\underset{w \in R^d}{min} f(w) \quad \textrm{and} \quad \underset{w \in R^d}{min} F_k(w) \quad \textrm{for each client}
\end{equation}
where the batch loss for each client is defined as
\begin{equation}
f_{b_{k,i}}(w) = \frac{1}{|b_k|} \sum_{x_i \in b_{k,i}} l(x_i, w_k)
\end{equation}
and $b_{k,i}$ being a random subset from $D_k$. Each client updates their local objective with respect to their dataset $D_k$ by
\begin{equation}
w_k \leftarrow w_k - \eta_k \nabla f_{b_{k,i}}(w_k)
\end{equation}
where $\eta_k$ is the clients learning rate. The clients objective then becomes
\begin{equation} \label{eq:fl-client-objective}
F_k(w) = \frac{1}{|D_i|} \sum_{b_{k,i} \in D_i} f_{b_k}(w)
\end{equation}
and so the shards objective will become
\begin{equation} \label{eq:fl-shard-objective}
G_s(w) = \sum_k^K \frac{|D_k|}{D} F_k(w)
\end{equation}
This can be defined as weight updates by
\begin{equation}
w_s \leftarrow w_s + \sum_k^K \frac{|D_k| \Delta w_k}{D}
\end{equation}
This happens in parallel across all shards, where the global objective will become the sum of the results produced across all shards
\begin{equation} \label{eq:global-shard-objective}
f(w) = \sum_s^S \frac{|D_s|}{D} G_s(w)
\end{equation}
We should note that this described approach follows the original FedAvg algorithm and can be modified in terms of fairness \cite{ref:q-fed-avg} among other averaging strategies.
\subsection{Shard Level Consensus} \label{sec:shard-consensus}
Each task deployed on the network can run an independent consensus mechanism. This allows for various task-related workloads to be accommodated by adapting the consensus algorithm. For example, when training large models where the performance of a single node may be limited, mechanisms such as PBFT \cite{ref:pbft} can be used. However, in shards with a fewer number of clients, we can make use of Raft consensus \cite{ref:raft-consensus}. This may benefit settings where model complexity is lower, and so a peer may be able to handle the throughput bottleneck of Raft consensus, determined by node performance.
Within the chosen consensus mechanism, we apply a pluggable policy to determine the acceptance of model writes. This policy can be substituted to handle alternative schemes (Section~\ref{sec:model-defence}). These schemes handle the identification of malicious clients by finding poisoned model updates or detecting Sybil attacks. It should be noted that while these plugins are applied at a task level to maintain the network's security, they can be upgraded with the smart contract that they are governing.
Another quality to consider is the detection of lazy clients. Detection of these clients may use a PN-sequence, which is applied to the model updates. This technique could then be applied within the shard level consensus to verify the clients' PN-sequence.
Because each shard contains a subset of the total participants for each task, the algorithmic complexity of consensus is significantly reduced to $\frac{\mathcal{C} \times \mathcal{P}_E}{\mathcal{S}^2}$. As a result, the network overhead needed to communicate model updates is also significantly reduced, which is the primary bottleneck in peer-to-peer federated learning.
\subsection{Mainchain Consensus} \label{sec:mainchain-consensus}
\begin{figure}[!b]
\centering
\includegraphics[width=\linewidth]{global-aggregation.png}
\caption{Aggregation of sharded models}
\label{fig:global-aggregation}
\end{figure}
The mainchain is responsible for coordinating all verified aggregated model updates from each shard. This chain contains all participants across all shards; however, the activity on this chain is limited to shard level aggregation results and task proposals. The consensus for this follows the same pluggable consensus policy as the Shard level consensus (Section~\ref{sec:shard-consensus}); however, the submitting peers are limited only to the subset of endorsing peers on the shard chains. In this way, we limit the endorsing peers to those in possession of an aggregated model and are able to verify its authenticity.
All accepted model updates to the mainchain will have been endorsed by each shard endorsing peers, which can safely be distributed and aggregated. If multiple models from a single shard have been accepted by disagreements between a shards endorsing peers, the model with more endorsements will win. So, if the endorsing peers submit the same model (model hashes are identical), the endorsing peers can safely evaluate the model update once. Figure~\ref{fig:global-aggregation} shows this process. The final step here is to post the final global model of the mainchain. This model represents the result after a single round of FL, and each of the shards will start the next round with these parameters.
We note that a new round may begin earlier within a shard by assuming the model aggregated by a shard endorsing peers is valid before it is accepted to the mainchain. If this model is not accepted, then the round within that shard must be restarted. This should only happen if the majority of endorsing peers within a shard are compromised, and as such, the disruption will only affect the compromised shard.
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth]{transaction-flow.png}
\caption{Transaction flow for shard level model submission}
\label{fig:shard-level-model-submission}
\end{figure*}
\subsection{Workflow}
This section discusses the entire workflow of our proposed framework, discussing the steps required to produce a trained model.
Firstly, we define the participants' category in our system. We can break this into three categories: clients, peers, and endorsing peers, and described as follows.
\begin{enumerate}
\item \textbf{Clients:} In a traditional FL setting, all participants are considered to be clients, where they are comprised of typical IoT devices such as smartphones, laptops, appliances, and vehicles, among others. Clients hold the private datasets that will be used for training models and are they are responsible for producing model updates. We consider the same assumption here, where clients do not validate models but rely on more powerful devices to ensure the system's security. This is the lightest class of participants in terms of resource consumption.
\item \textbf{Peers:} Peers are responsible for holding copies of the ledger. From this pool of peers, a committee can be elected to reduce the evaluation overhead. This helps take the load off of endorsing peers and allows peers to offer API services for clients.
\item \textbf{Endorsing peers:} Members of the committee are responsible for evaluating model updates during consensus. They hold their own private dataset to endorse other model updates and can submit their own model updates, although they are not required to do so. Furthermore, the policy for evaluation of model weights can be modified, and these peers additionally must check for valid authentication of the write-set. Finally, committee members must participate in the mainchain consensus to determine which shard level updates will be accepted.
\end{enumerate}
These participants will then participate in one or more tasks, as managed by the mainchain smart contract. Starting from a fresh network, we will describe the following process.
\subsubsection{Task Proposal} The procedure begins with a task proposal on the application layer. It will be proposed on the mainchain in the form of a smart contract, and it outlines the requirements and task description. We leave the details of this specification open, as we may not know the complete set of features or optimization level objectives as such in a vertical FL (VFL) scenario \cite{ref:federated-learning-concepts}. Once this proposal has garnered enough interest in client registration, a new shard will be provisioned, and shard-related smart contracts will be deployed to the newly created shard(s). These processes can be thought of as an open request for participants to collaborate on a specific task.
\subsubsection{Client Training}
Within a specific shard, an off-chain coordination network will be formed as defined by the proposal. This could utilize existing FL frameworks and best practices, and new paradigms. As such, the fraction of clients fit within a specific round will be determined off-chain by this process and exclude, for example, currently unavailable clients due to battery or network coverage issues. These selected clients are then responsible for producing a model update to the global model weights $w$. Each required client will then compute their local updates as seen in Eq.~\eqref{eq:fl-client-objective}. We can see this as Step~1 in Figure~\ref{fig:shard-level-model-submission}.
\subsubsection{Off-Chain Model Storage}
A client will then optionally upload a model to an off-chain cache such as InterPlanetary File System (IPFS) \cite{ref:ipfs}. At this stage, the model can be downloaded publicly; however, it has not yet been verified nor recorded on the ledger yet. This step is used to record historical models or take the network load off of serving local model updates to peers. We can see this as Step~2 in Figure~\ref{fig:shard-level-model-submission}.
\subsubsection{Model Submission}
Once the model has been uploaded and is available for download, the client will submit the update metadata by invocation of the shard level smart contract. This update should include information such as the model's hash value and a link from which the model can be downloaded. This can be seen in Step~3 in Figure~\ref{fig:shard-level-model-submission}.
\subsubsection{Peer Endorsement}
The peers on the shard must endorse this model update according to the shard level consensus mechanism (Section~\ref{sec:shard-consensus}). This work could be offloaded to the peers' local workers, where additional computing resources may be present. This can be seen in Steps~4 and 5 in Figure~\ref{fig:shard-level-model-submission}.
\subsubsection{Model Evaluation}
The peers' local workers request the model weight update from the link specified in the model update. The downloaded model will then be verified against the submitted hash to ensure integrity. Once the model is verified, it can be evaluated by various techniques; for example, the peer may test the model against its local dataset and reject the update on sufficient degradation in performance, such as in the RONI defence. This would be ideal in cases where the data distribution is known to be IID. Alternative solutions such as the FoolsGold scheme may be used for non-IID data, which could be used to identify poisoning attacks from Sybils. The results of this will be returned from the worker back to the peer. Here, we have summarized Steps~6-8 in Figure~\ref{fig:shard-level-model-submission}.
\subsubsection{Shard Aggregation}
Endorsed model updates are written to the ledger according to the consensus quorum. The off-chain FL process can continue by aggregating the models from the current round if they have been accepted into the ledger. These models will then be downloaded and aggregated by the coordinating servers according to Eq.~\eqref{eq:global-shard-objective}. These aggregated models will then be published to the mainchain. Similarly, these aggregated shard updates are only be accepted if the committee of endorsing peers reaches a consensus on the update governed by the mainchains consensus mechanism (Section~\ref{sec:mainchain-consensus}).
\subsubsection{Global Aggregation}
Finally, all accepted shard level model updates accepted to the mainchain will be downloaded and aggregated by the shard level servers. The peers can easily verify these models by checking the hash of the new distributed global model against the posted hash from the mainchain for the finalized FL round.
\section{Experiments}\label{sec4}
To demonstrate the effectiveness of our proposed sharding solution, we evaluate the performance in a Proof of Concept (PoC) implementation\footnote{https://github.com/blockchain-systems/ScaleSFL}. The purpose of experimentation here is first to measure the throughput of a FL workload scale with respect to the number of shards. Secondly, we measure the sharding impact on model performance in both IID and non-IID scenarios. It is important to ensure the quality of models and keep the rate of convergence low when a potential subset of peers are malicious.
To run these experiments, we use Hyperledger Fabric \cite{ref:fabric}, a permissioned distributed ledger platform. This platform offers a modular architecture allowing for plug-n-play consensus and membership services. Hyperledger Fabric offers an execute-order-validate architecture for transactions, allowing each peer to evaluate models in parallel within each shard. This is opposed to most public blockchain platforms that first order transactions and execute them sequentially. Fabric additionally offers modular endorsement and validation policies which we will use to implement custom consensus logic within an endorsement plugin. We will read in a proposal response payload, sending the given transaction read-write set to the peers' worker for validation. This read-write set contains the necessary information to endorse or reject the proposed model update. This will be built directly into the Fabric peers.
Our solution workflow may be adapted to alternative permissioned platforms such as Quorum \cite{ref:quorum}, or public platforms such as Ethereum \cite{ref:ethereum}, which may allow for additional reward allocation schemes. Implementations on these platforms may consider the use of child chains utilizing existing layer-2 scaling solutions. We choose Hyperledger Fabric for our PoC primarily for the flexibility of consensus and membership services during development.
\begin{table*}
\caption{Experimental Configuration}\label{tab:experimental-configuration}
\begin{tabular}{cccccc}
\toprule
Component&Type&CPU&GPU&RAM&Disk (SSD)\\
\midrule
Caliper Benchmark&Caliper 0.4.2&Ryzen 7 3800X@3.9 GHz&NVIDIA RTX 3080&64GB&1TB\\
Fabric Peer (Shard)&Fabric 2.3.3&Ryzen 7 3800X@3.9 GHz&NVIDIA RTX 3080&64GB&1TB\\
Fabric Peer Worker&Python 3.9, PyTorch 1.10.1&Ryzen 7 3800X@3.9 GHz&NVIDIA RTX 3080&64GB&1TB\\
\bottomrule
\end{tabular}
\end{table*}
Tests here are run locally on a single machine simulating a Fabric test-network with a single orderer running Raft \cite{ref:raft-consensus}, and eight peers, each with their own organization, along with a certificate authority. Each peer worker runs on a single thread, allowing each peer to act independently. The experimental configuration can be seen in Table~\ref{tab:experimental-configuration}.
We implement two smart contracts to implement both the mainchain consensus and shard level consensus mechanisms. Smart contracts, otherwise known as chaincode in Fabric, will be deployed to a specific channel within Fabric. We use channels to simulate shards, where each channel operates independently with the ability for different membership and endorsement policies per channel. Additionally, we relax the previous definitions of endorsing peers, such that $P\colon \{Peers\} = P_E\colon \{Endorsing~peers\}$. We deploy a shard level smart contract we term the ``models" chaincode to each channel. Each participant operating within a shard will be assigned to a shard by our managing contract. We deploy the mainchain consensus on its own channel, in which every peer from every shard will join. We term this the catalyst contract, where it is responsible for aggregating a global model from each of the shards. As mentioned previously, model updates are stored off-chain. In this implementation, each peers worker hosts the models locally with a gRPC server.
We coordinate the FL processes off-chain while pinning updates back to each shard's ledger. This coordination is abstracted from the approach discussed previously; in this case, we use the Flower \cite{ref:flower} an open-source extensible FL framework. In this implementation, during aggregation, submitted models are verified to have been endorsed, which can be checked by querying the peers' local ledger to reduce a round of communication. This is conducted by creating a custom strategy within the Flower server and modifying the aggregated fit to filter out any updates which are not present on-chain, by querying the models' smart contract. We additionally implement differential privacy using Opacus \cite{ref:opacus}, applied during client training. We use an $(\epsilon, \delta)$ target of (5, $1\mathrm{e}{-5}$), a noise multiplier of 0.4, and a max gradient norm of 1.2.
\subsection{Evaluation}
For evaluation, we employ Hyperledger Caliper \cite{ref:hyperledger-caliper}, an open source benchmarking tool allowing for various workloads to be conducted for our system under test (SUT). These tests are performed by an independent process that generates transactions to be sent to the SUT by a specified configuration, monitoring for responses to determine metrics such as latency, throughput and success rate. We conduct tests with varying numbers of workers, transactions counts, and transactions per second (TPS) to evaluate the limits of the system.
We run several workloads to test the effects of sharding on transaction throughput and model performance. The benchmark tasks include:
\begin{itemize}
\item \textbf{Update creation throughput:} We first run an experiment to test the throughput of model creation. To do this, we generate a number of models updates, make the parameters available locally, and have the endorsing peers evaluate them during consensus. In this way, we can measure how sharding affects the performance of the consensus mechanism by parallelizing the workload among shards. We aim to verify our initial claim that the global computations required with become $\frac{\mathcal{P} \times \mathcal{C}}{\mathcal{S}}$.
\item \textbf{Model performance:} We evaluate the performance of the models across two scales: (1) the global number of epochs computed and (2) the number of gradients, which provide more accurate measures of performance for decentralized training \cite{ref:bfl-shard}. While we are primarily concerned with the performance of integrating existing FL solutions into the blockchain, sharding should allow a larger fraction of clients to be fit, allowing for model performance to be increased.
\end{itemize}
\subsection{Datasets}
For experimentation, we use three datasets, namely the MNIST dataset \cite{ref:mnist} consisting of 60,000 28x28 images of handwritten digits, and the CIFAR-10 dataset \cite{ref:cifar10} which consists of a dataset of 60,000 32$\times$32 images separated into 10 classes with 6,000 images per class. We use these to represent IID data, where these classes of data are split evenly between clients.
For non-IID scenarios, we use the LEAF dataset \cite{ref:leaf-dataset} which can be used to test model performance by partitioning the data between peers in a reproducible way. In this setting, the classes of data may not be evenly split. For example, in the case of the Federated Extended MNIST (FEMNIST) dataset, characters were split by the writer of the digit or character. This can be a strong benchmark to model real-world distributions of data.
These datasets are used as a common baseline for comparison and evaluation against previous solutions \cite{ref:federated-learning, ref:bfl-async, ref:bfl-shard}, showing the addition of a sharding solution maintains competitive results on model performance. While a non-IID assignment of data across shards may influence the generated global model during the global aggregation stage, the selection of mainchain defence mechanisms plays the largest role in model convergence. Thus, the results generated by the chosen datasets demonstrate the applicability of sharding solutions to existing Blockchain-based methods.
\subsection{Results}
We created multiple workloads for the \textit{Update Creation Throughput} benchmark to explore the effectiveness of our sharding solution. These workloads are conducted without malicious updates using the MNIST \cite{ref:mnist} dataset to evaluate the performance of model creation transactions. We set the timeout period for each transaction to 30 seconds, allowing us to consider the number of failed transactions as stale (i.e., not malicious). In this experiment, each of the clients evaluated the update against their entire local dataset, in which we have used the entire test split of the MNIST dataset for each client.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{shards-by-throughput.png}
\caption{\#shards vs. system throughput (TPS)}
\label{fig:benchmark-shards-by-throughput}
\end{figure}
We first measure the throughput with respect to the number of shards, as seen in Figure~\ref{fig:benchmark-shards-by-throughput}. This workload uses two Caliper workers, evaluated over 200 transactions, where we can see the throughput for each independently operating shard which allows the global throughput to scale linearly. This directly reflects the time for the evaluation of a model that is the primary bottleneck. We note here that the sent TPS for each number of shards is set just above its throughput in order to saturate the system for the evaluation. In this workload, we used the same number of sent transactions for consistency; however, higher numbers of shards have shown improvements in throughput for increased transactions sent, while a lower number of shards may begin to fail earlier.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{tps-by-throughput-avglatency.png}
\caption{Sent TPS vs. system throughput (TPS) \& average response latency}
\label{fig:benchmark-tps-by-throughput-avglatency}
\end{figure}
To test the maximum throughput achieved by our system, we measure the sent TPS against the system throughput and average latency. See Figure~\ref{fig:benchmark-tps-by-throughput-avglatency}. This workload is run with 2 caliper workers over 200 transactions. As we increase the sent TPS, the system will reach a point where it becomes saturated and is unable to handle a higher TPS. We can see this by looking at the point where the average latency begins to increase in the figure, where the throughput is maximized. We conduct this workload by measuring sent TPS in increments of 3, starting from 3 TPS. We can see each additional shard increases the overall system throughput similarly in Figure~\ref{fig:benchmark-shards-by-throughput} while demonstrating the maximum throughput limit.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{txcnt-by-fail-avglatency.png}
\caption{Transaction count sent vs. average response latency \& failure count}
\label{fig:benchmark-txcnt-by-fail-avglatency}
\end{figure}
To explore the limits of a usage surge, we measure the number of transactions sent by system throughput, average latency, and failure rate. In this benchmark, we are interested in the system's behaviour under a workload with a higher sent TPS than throughput. This workload is run with 2 caliper workers and a sent TPS just above the maximum throughput as seen in the previous workload. See Figure~\ref{fig:benchmark-txcnt-by-fail-avglatency}. When the system cannot handle the number of transactions currently queued, the average latency will spike, and the number of failed requests will increase because the system begins timing out requests. The system will then go through a ``flush" period where many long-running requests begin to fail, so the average latency of successful requests will drop. We note here that the average latency peaks at around 16 seconds due to the average between the maximum latency (the timeout limit) and the minimum latency, which corresponds to the time it takes to evaluate a model fully.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{txcnt-by-throughput.png}
\caption{Transaction count sent vs. System throughput (TPS)}
\label{fig:benchmark-txcnt-by-throughput}
\end{figure}
Figure~\ref{fig:benchmark-txcnt-by-throughput} shows the throughput over this workload, where the throughput of the system decreases when it is no longer able to handle the number of incoming requests, as the overhead to process these additional requests limits the number of successful requests processed. The decrease in throughput corresponds to the increase in average latency previously observed.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{workers-by-throughput-avglatency.png}
\caption{Caliper workers vs. system throughput (TPS) \& average response latency}
\label{fig:benchmark-workers-by-throughput-avglatency}
\end{figure}
We finally test how the system handles concurrent requests. To do so, we run multiple workloads, varying the number of caliper workers, and measuring the system throughput and average latency. This workload configuration sends 200 transactions, with the sent TPS equal to the previously-mentioned maximum throughput. This can be in Figure~\ref{fig:benchmark-workers-by-throughput-avglatency}. We observe that the throughput of the system with respect to the number of workers is quite noisy, however, there is a general downward trend in the system throughput. Increasing the number of caliper workers here allows us to scale the workload generation, however since the endorsement workers evaluating the model operate sequentially in these experiments (limited to a single thread), The throughput is limited. Similarly, we can see an upward trend in average latency caused by each transaction's time waiting for execution in the queue. We can see that the number of shards plays the largest factor here as workloads with more than 2 shards are tightly grouped with respect to average latency due to these workloads being able to operate in parallel across shards.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{model-performance.png}
\caption{Training loss and testing accuracy of CNN model on the MNIST dataset}
\label{fig:benchmark-model-performance}
\end{figure}
\begin{table}
\caption{Best accuracy with varying minibatch size (B) and number of local epochs (E)}\label{tab:benchmark-model-performance}
\begin{tabular}{cccccc}
\toprule
$B$&$E$&FedAvg (Accuracy)&ScaleSFL (Accuracy)\\
\midrule
10&1&00.8784&0.9835 \\
10&5&0.9511&0.9897 \\
10&15&0.9654&0.9896 \\ \hline
20&1&00.7249&0.9758 \\
20&5&0.9251&0.9881 \\
20&15&0.9544&0.9889 \\
\bottomrule
\end{tabular}
\end{table}
To measure \textit{model performance}, we measure the effect of various training parameters. We use minibatch sizes of $B \in \{10, 20\}$, and local epoch sizes of $E \in \{1, 5, 15\}$. We present results here from the non-IID MNIST dataset, however similar results hold for both CIFAR-10, and LEAF benchmarks. The effects of both minibatch size and local epoch sizes can be seen in Figure~\ref{fig:benchmark-model-performance}, and Table~\ref{tab:benchmark-model-performance}, where we compare our ScaleSFL solution and the traditional FedAvg algorithm. This comparison is performed under the assumption all clients are honest, and thus we are simply comparing the effect of sharding. Both training loss and accuracy are considered, and workflows are run for 15 global epochs, where the number of ScaleSFL shards is set to 8, with each shard sampling eight clients each round, making for a total of 64 clients. The FedAvg algorithm is comparably run with 64 clients, with both methods using a client learning rate of $\eta_k = 1e-2$.
In Figure~\ref{fig:benchmark-model-performance}, we observe the convergence rate of ScaleSFL is faster than the FedAvg algorithm, generally reaching an accuracy of $0.98$ within the 15 global epochs considered. This is due to the parallelized nature of ScaleSFL, where each shard can simultaneously filter and produce viable updates with reasonable client sampling rates, combining them to produce a stronger global model at each step.
\section{Discussion}\label{sec5}
\paragraph{\textbf{Hierarchical Sharding}}
Since each shard operates independently, several additional improvements over a traditional FL workflow can be achieved. For example, a typical device selection step is usually performed, sampling a subset of the available devices for each round. However, since the population of each shard is much smaller than the global population, each shard's ability to update and maintain active devices becomes more efficient, and during aggregation, a much larger sampling of devices can be used from the global population.
Additional optimizations could be achieved by placement of clients based on region \cite{ref:hierfavg}. Since each client may submit their model to an off-chain cache such as IPFS, shards with a population based within the same region could reduce overhead latency. In this way, the overhead only occurs during the global aggregation phase. The possibility of single-shard takeover attacks may be introduced by not using a random sampling scheme. However, as noted before, this disruption will only affect the compromised shard since all endorsing peers from all shards must agree on which shard level updates are aggregated. This may allow for region based sampling algorithms to be used for the assignment of participants. These strategies should be chosen depending on the underlying task, potentially chosen by the task submitter, or consensus among task participants prior to shard creation.
Alternative participant sampling can be considered in the case of cross-silo or consortium based settings such as medical or financial organizations. For example, it may be beneficial to allow clients of a particular organization to be grouped under a single shard \cite{ref:bfl-shard}. Fine-grained control oversampling and permissions are delegated independently within each shard.
\paragraph{\textbf{Model Provenance}}
By pinning model information back to the mainchain, we provide an open model hub for sharing and distributing trained models in a decentralized fashion. Model submission to IPFS additionally allows for transparent model check-pointing and disaster recovery. In the case that a bug has been introduced at some point in the FL process, either intentionally through malicious attack, or by data bias or incompatibility, previous model checkpoints may be restored, and a new task may be initiated using this saved model checkpoint.
\paragraph{\textbf{Rewards Allocation}}
While our implementation of ScaleSFL uses Hyperledger Fabric, alternative implementations may consider platforms such as Ethereum \cite{ref:ethereum} for implementation of reward mechanisms. Since training models in a permissionless fashion may result in decreased contributions, due to the availability of models through a model hub, and without expending computing resources, rewards become an integral part of the workflow. While access to the models should remain free, contributing computing resources should be rewarded by distributing rewards to the contributing clients. Additional incentives could be provided by task contributors or interested clients, to ``sweeten the pot", encouraging community contributions. Additionally, with the incorporation of a native cryptocurrency, submitting models transactions could incur a small gas fee. This would prevent attempted Denial of Service (DOS) attacks against the system since the rewards for model contributions are only realized for non-malicious updates. However, common DOS attacks may generate random model updates or send updates with dead cache links.
\paragraph{\textbf{Alternative Attacks}}
Up to this point, we have discussed malicious clients as clients which send poisoned model updates. However, malicious clients may attempt various alternate types of attacks, such as uploading a different model than specified in the task request or making the model update very large, attempting to perform a DOS attack on a shard. These attacks may be prevented by checking the size of the model uploaded to the cache prior to download. However, a more clever client may attempt to submit a legitimate model many times, attempting to either reap the rewards for training, or to perform a DOS attack against the system. These clients are known as lazy nodes \cite{ref:fl-meets-bc}. They additionally attempt to copy another client's model update during a gossip phase. This problem has severe negative effects on global model performance, so updates from these lazy nodes should be discarded. One way to recognize these lazy nodes is to publish model updates with a pseudo-noise (PN) sequence, and then subsequently publish the PN sequence after each client has published their model update. Lazy nodes will be spotted by checking the correlation between the model updates and the published PN sequence \cite{ref:fl-meets-bc}. Gas fees should deter spotted clients, and Sybils from attempting to submit further model updates. Incorporating the defence into our workflow would modify the off-chain model storage step, ensuring all models submitted during model submission have a valid PN signature.
\section{Conclusions}\label{sec6}
In this paper, we proposed a sharding solution for blockchain based federated learning. We described a modular workflow, allowing for pluggable defence mechanisms and poisoning mitigations strategies. This workflow supports off-chain FL frameworks to be used, while pinning verified model updates to the blockchain. We additionally implement a proof-of-concept of ScaleSFL using Hyperledger fabric. Our experimental results demonstrated that ScaleSFL improves validation performance linearly with the addition of shards. This helps address the prevalent scalability issue related to blockchain consensus by demonstrating the impact of sharding.
As {\em ongoing and future work}, we further explore the details of a sharded solution applied to the blockchain.
Specifically, we aim to simulate malicious attacks on the system via model poisoning updates.
This would show the effectiveness of a blockchain solution to provide security against bad actors.
Additionally, we will explore the application of a reward mechanism for participating in the model training process for our current system. This could be explored on other chains such as Ethereum \cite{ref:ethereum}, where the use of native tokens is well-supported. Finally, we would like to implement a more full-featured implementation of this PoC, including dynamic shard creation and allowing model proposition through our catalyst contract.
\begin{acks}
This work is partially supported by NSERC (Canada) and University of Manitoba.
\end{acks}
\bibliographystyle{unsrt}
|
\section{Introduction}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{BT_illu_dim1.pdf}
\caption{An illustration of the attention mechanisms on channel, spatial, and batch dimensions. Specifically, vision Transformer usually treats image patches as a sequence, and then uses a self-attention layer upon all image patches, \textit{i.e.}, the spatial attention. Different from previous work, we propose to also treat a mini-batch of samples as a sequence, where a typical self-attention layer can thus work on the batch dimension.}
\label{fig:illu}
\end{figure}
In the past five years, we have witnessed great success of deep neural networks with attention mechanisms, especially the Transformer architecture~\cite{vaswani2017attention}, for natural language processing (NLP). Recently, the Transformer architecture has been successfully extended to visual recognition tasks known as vision Transformer or ViT~\cite{dosovitskiy2020image}, suggesting its great potential as a general deep architecture for learning through different modalities. Therefore, following previous attention mechanisms used in CNNs, such as channel and spatial attentions~\cite{hu2018squeeze,wang2018non,woo2018cbam,carion2020end,gberta_2021_ICML}, recent vision Transformer variants mainly focus on exploring more efficient and effective local/global attention structures~\cite{liu2021swin,yuan2021tokens,chu2021twins,chen2022regionvit}. However, from the perspective of information propagation, current vision Transformer architectures can only enable the spatial information propagation among pixels/patches within the same image, ignoring the possible information propagation among different data samples within each training mini-batch. An intuitive example showing the differences between channel, spatial, and batch attentions is illustrated in Figure~\ref{fig:illu}.
To enable vanilla vision Transformer with the ability of sample relationship modeling, Hou~\textit{et al.}~present to incorporate a Transformer encoder module between the last hidden layer and the final classifier to explore image-level sample relationships using the attention mechanism, which is referred to as BatchFormer~\cite{hou2022batch} and achieves significant improvements over a wide range of data-scarce image classification tasks. However, BatchFormer mainly focuses on exploring sample relationships at the image-level for classification, leaving sample relationships for dense representations at the pixel/patch level poorly investigated. Inspired by this, we devise a new general batch Transformer module, BatchFormerV2, which further enables the sample relationship learning for dense representations, aiming to facilitate the research on a more powerful and general vision Transformer with not only the spatial attention but also the batch attention, from image-level to pixel/patch-level representation learning.
Exploring sample relationships within each mini-batch using deep neural networks is usually non-trivial, since we do not always have a mini-batch of data for testing, which is also known as the training-and-testing inconsistency. For example, batch normalization requires to track running statistics (\textit{i.e.}, the mean and variance of each mini-batch) during training in a momentum way~\cite{ioffe2015batch}, which is then used as batch statistics for testing. When applying a parametric attention module in the batch dimension, it is even more difficult to handle the above-mentioned inconsistency problem. To this end, Hou~\textit{et al.}~\cite{hou2022batch} introduce a shared classifier positioned both before and after the batchformer module, aiming to learn a batch-invariant classifier. In this paper, we devise a more general batch Transformer module to further enable the information propagation between samples at different levels, \textit{i.e.}, both pixel/patch and image levels, by learning batch-invariant representations.
Therefore, the proposed batchformer module can be applied to different hidden layers/blocks in typical vision Transformers.
To achieve this, we explore a two-stream pipeline for training, one stream with a batchformer module and the other stream without, where all other layers/blocks are shared by these two streams. By doing this, all these shared layers/blocks are trained to generalize well for the input with or without the batch Transformer module. During testing, we can directly remove the batch Transformer module without sacrificing performance.
In this paper, with the proposed simple yet effective batch Transformer module, we further enable the information propagation along the pixels/patches of different samples within each mini-batch to facilitate dense representation learning. To evaluate the proposed batch Transformer module, named as BatchFormerV2, we perform extensive experiments on several popular visual recognition tasks, including image classification, object detection, and panoptic segmentation, where the experimental results show that the proposed module can be a general representation learning solution via sample relationship learning. The main contributions of this paper are summarized as follows.
\begin{itemize}
\item We introduce a new plug-and-play module, BatchFormerV2, as a general solution for sample relationship learning at different levels, which further facilitates robust dense representation learning.
\item We develop a two-stream training strategy, making it possible to improve vision Transformers without any extra inference cost by directly removing the proposed BatchFormerV2 module during testing.
\item We perform extensive experiments on three popular visual recognition tasks, \textit{i.e.,} image classification, object detection, and panoptic segmentation, demonstrating remarkable generalizability of the proposed method.
\end{itemize}
\section{Related Work}
\subsection{Vision Transformer}
Transformers were first presented by Vaswani \textit{et al.} \cite{vaswani2017attention} for machine translation based on multi-head self-attention mechanism. As a core part of Transformers, attention mechanism \cite{bahdanau2014neural} aggregates information from the entire input sequence and then update it. In the past several years, Transformer-based architectures have dominated in natural language processing (NLP). For example, large-scale Transformer-based models, \textit{e.g.}, BERT~\cite{devlin2018bert}, show superior performance among massive down-stream NLP tasks. Besides, self-attention also demonstrates the powerful modeling ability of non-grid data~\cite{velickovic2017graph}, and improves graph representation learning.
Recently, Transformer models present new paradigms for computer vision tasks, including classification~\cite{dosovitskiy2020image,liu2021swin}, detection~\cite{carion2020end,zhu2020deformable,meng2021-CondDETR,gao2021fast}, segmentation~\cite{carion2020end,cheng2021per,zheng2020rethinking,strudel2021,wang2021pyramid,xie2021segformer}, and representation learning~\cite{chen2021empirical,bao2021beit,he2021masked}. Specifically, Dosovitskiy \textit{et al.}~\cite{dosovitskiy2020image} introduced a pure Transformer model, termed as ViT, to apply a sequence of image patches and achieve comparable performance on image classification tasks. Recently, Visual Transformer~\cite{dosovitskiy2020image,liu2021swin} has gradually become a new backbone for visual tasks, and massive large models based on Transformers have emerged in computer vision, including CLIP~\cite{radford2021learning}, MoCo~\cite{he2020momentum,chen2021empirical}, DINO~\cite{caron2021emerging}, DALL-E~\cite{ramesh2021zero}, BEiT~\cite{bao2021beit}, and MAE~\cite{he2021masked}. Except for the backbone, Transformer-based models, \textit{e.g.}, DETR~\cite{carion2020end}, have also reformed the pipeline of detection and segmentation. DETR~\cite{carion2020end}, constructed upon the encoder-decoder Transformer architecture, demonstrates a clear set-based method for detection and greatly simplifies the traditional pipeline which includes many hand-designed components. Recently, Zhu \textit{et al.} ~\cite{zhu2020deformable} present Deformable DETR, which largely accelerates the convergence and improves the performance.
Though the great success of Transformer in computer vision, current approaches merely investigate the spatial self-attention, while ignoring the pixel/patch relationships among samples. Recently, Hou~\textit{et al.}~\cite{hou2022batch} present to explore the sample relationships in the image level for data scarcity challenges, which however is impractical to dense prediction tasks. In this paper, we introduce a novel Batch Transformer, BatchFormerV2, to enable information propagation among samples in the pixel/patch level, and largely facilitate dense representation learning for dense prediction tasks, \textit{e.g.}, Object Detection and Panoptic Segmentation.
\subsection{Sample Relationship Learning}
Mini-Batch stochastic gradient descent optimization is one of core paradigm of training deep neural networks~\cite{goodfellow2016deep,krizhevsky2012imagenet,lecun2015deep}. In order to accelerate the network training, Ioffe \& Szegedy~\cite{ioffe2015batch} present Batch Normalization to reduce internal covariate shift via normalizing intermediate representations. Batch Normalization has inspired a large number of cross-batch techniques~\cite{wu2018group,chang2019domain}, and has become one of cornerstones for modern deep networks. Recently, Hou~\textit{et al.}~\cite{hou2022batch} introduce a BatchFormer module in the penultimate layer with a shared classifier for data scarcity tasks. Inspired by them, BatchFormerV2 proposed in this paper is pluggable into different layers in Visual Transformer and DETR, which first enables applying batch Transformer module for dense prediction tasks. Here, we compare the proposed BatchFormerV2 with other cross-batch techniques. First, Batch Normalization actually enables the sample information propagation in a linear way, and stables the network training. Differently, we think the proposed BatchFormerV2 propagates information among samples in the mini-batch in a non-linear way via multi-head self-attention. Meanwhile, transductive inference~\cite{liu2018learning,satorras2018few} is also popular in few-shot approaches and recent approaches~\cite{kossen2021self,mondal2021mini} introduce to reason the relationships between samples, while those techniques require batch inference and thus related applications are limited. In addition, Mixup~\cite{zhang2017mixup} is a popular augmentation technique, which draws virtual samples from the vicinity distribution of the training examples. Similar to Mixup~\cite{zhang2017mixup}, our BatchFormerV2 also changes the feature space according to the mini-batch training samples. Differently, BatchFormerV2 is able to insert different layers in Transformer modules, and learns the transform implicitly. Therefore, BatchFormerV2 can be easily applied to dense prediction tasks.
\section{Method}
In this section, we first revisit the attention mechanisms in vision Transformer. We then introduce the proposed BatchFormerV2. Lastly, we describe the two-stream training pipeline for the optimization of vision Transformers with the proposed method.
\subsection{A Revisit of Vision Transformer}
The great success of Transformer architecture in NLP has recently spread to almost every region of computer vision known as vision Transformers~\cite{dosovitskiy2020image,liu2021swin,carion2020end}. The Transformer architecture not only achieves superior performance in different vision tasks~\cite{liu2021swin} (\textit{e.g.}, image classification, object detection, and semantic segmentation), but also bring novel paradigms for some fundamental tasks (\textit{e.g.}, DETR~\cite{carion2020end} for object detection and MAE~\cite{he2021masked} for self-supervised learning). Among typical vision Transformers, the overall model usually consists of a stack of multiple Transformer encoder blocks~\cite{vaswani2017attention}, where each Transformer encoder block contains a multi-head self-attention layer (MHSA) followed by a feed-forward network (FFN). Specifically, the self-attention mechanism used in vision Transformer can be described as follows. Given $\mathbf{Q},~\mathbf{K},~\mathbf{V}\in R^{N\times C}$ as the query, key, and value, respectively, where $N$ is the number of image patches (or tokens) and $C$ is the embedding dimension. We then have the output $\mathbf{Z}$ for the self-attention module:
\begin{equation*}
\mathbf{Z} = \text{softmax}(\frac{\mathbf{Q}\mathbf{K}^\top}{\sqrt{C}})\mathbf{V},
\end{equation*}
where $\mathbf{Q}$, $\mathbf{K}$ and $\mathbf{V}$ are learned from the same input. Specifically, multi-head self-attention module applies attention by spliting the input into multiple representation subspaces and then concatenates the representations from different heads. From a perspective of information propagation, the Transformer architecture aggregates the feature of the tokens via spatial attention. Different from typical attention layers used in vision Transformer, BatchFormer performs a self-attention on the batch dimension, \textit{i.e.}, it aggregates the feature of the tokens from different samples within each mini-batch in an end-to-end learning way.
\subsection{BatchFormerV2}
\begin{figure}[!ht]
\centering
\includegraphics[width=.85\textwidth]{bt2_pbt2.pdf}
\caption{Illustration of BatchFormerV2. For the $i$-th spatial position in the feature map, we apply a Transformer block to the feature map where the length of each sequence is the batch size.}
\label{fig:pbt}
\end{figure}
To generalize the batch attention mechanism into pixel/patch level feature maps for dense representation learning, we devise BatchFormerV2 as follows. Given $\mathbf{Q},~\mathbf{K},~\mathbf{V} \in R^{B\times N\times C}$, we then have
\begin{equation}
\mathbf{Z}_i = \text{softmax}(\frac{\mathbf{Q}_{i}\mathbf{K}_i^{\top}}{\sqrt{C}})\mathbf{V}_i,~~~\mathbf{Z} = \text{concat} (\mathbf{Z}_1,\dots,\mathbf{Z}_N),
\end{equation}
where $\mathbf{Q}_i,~\mathbf{K}_i,~\mathbf{V}_i \in R^{B\times C}$ and $\mathbf{Z}\in R^{B\times N\times C}$.
As illustrated in Figure~\ref{fig:pbt}, given the input for a specific layer/block with the spatial dimensions $H,~W$, \textit{i.e.}, the number of image patches is $N=H\times W$. During training, at each spatial position $i=1,\dots, N$, we treat the batch of patch features in current position as a sequence, \textit{i.e.}, we have $N$ sequences each with the length of $B$. All above-mentioned sequences are then feed into a shared Transformer block.
Two reasons for using a shared Transformer block are as follows: 1) it will increase the computation and memory consumption considerably if we use different blocks at different spatial positions; 2) it will be difficult to dense prediction with different sizes of input images, which is in line with the motivation of fully convolutional networks (FCNs) for dense prediction~\cite{long2015fully} as well as the convolution operations~\cite{lecun1995convolutional} and channel-wise attentions~\cite{hu2018squeeze}.
Therefore, we share the Transformer block among the spatial dimensions in BatchFormerV2. In addition, by doing this, the proposed BatchFormerV2 can be implemented by simply transposing the spatial and batch dimensions before the standard multi-head self-attention layers. Noticeably, as illustrated in Figure~\ref{fig:alg}, BatchFormerV2 can be easily implemented with a few lines of code using popular deep learning packages such as PyTorch~\cite{paszke2019pytorch}.
\begin{figure}[!ht]
\centering
\begin{lstlisting}[language=Python]
def batch_former_v2(x, encoder, is_training, is_first_layer):
# x: input features with the shape (B, N, C).
# encoder: TransformerEncoderLayer(C, nhead, C, 0.5, batch_first=False)
if not is_training:
return x
orig_x = x
if not is_fist_layer:
orig_x, x = torch.split(x, len(x)//2)
x = encoder(x)
x = torch.cat([orig_x, x], dim=0)
return x
\end{lstlisting}
\caption{Python code of BatchFormerV2 based on PyTorch.}
\label{fig:alg}
\end{figure}
\subsection{Two-Stream Training}
\begin{figure}[!ht]
\centering
\includegraphics[width=.95\textwidth]{bt2_arch2.pdf}
\caption{The two-stream training pipeline for the proposed BatchFormerV2. For example, here the input for DETR~\cite{carion2020end} indicates the feature map from the backbone network, while the input for ViT~\cite{dosovitskiy2020image} is the feature map after the patch embedding layer. We copy the feature and input into BatchFromerV2 stream. The outputs of the two streams are inputted into shared prediction module. Particularly, the Transformer Blocks and prediction module, \textit{e.g.}, Transformer Decoder in DETR~\cite{carion2020end} and classification head in ViT~\cite{dosovitskiy2020image}, are shared between the two streams. Noticeably, during inference, the BatchFormer stream is removed and thus there are no additional computation.}
\label{fig:arch}
\end{figure}
One of the most significant challenge for applying batch attention is the inconsistency between training and testing~\cite{kossen2021self,hou2022batch}. Specifically, in BatchFormer~\cite{hou2022batch}, Hou~\textit{et al.}~ address the inconsistency problem by introducing a shared classifier, which enables to remove BatchFormer module during inference. Inspired by this, we generalize this solution to dense representation learning by utilizing the similar idea, \textit{i.e.}, batch-invariant learning. Therefore, we introduce a new two-stream training strategy for using BatchFormerV2 as follows.
When applying the proposed BatchFormerV2 module to a specific block of vision Transformers, we create a new siamese stream followed by a BatchFormerV2 module, leaving the original stream unchanged. That is, both two streams share the same Transformer block. By doing this, during training, all shared blocks are trained on a mixture of the distributions with or without a BatchFormerV2 module. Therefore, during testing, the original stream can work well for both with and without a mini-batch of testing data available. To avoid introducing any extra inference load, we thus remove the BatchFormerV2 module for testing. In addition, from the perspective of regularization, the proposed BatchFormerV2 also serves as a strong regularization during training, which has turned out to be very useful in vision Transformers.
Lastly, with the proposed two-stream training strategy, a BatchFormerV2 module can be easily integrated into existing Transformer architectures for different visual applications, such as ViT~\cite{dosovitskiy2020image} for image classification and DETR~\cite{carion2020end}/Deformable-DETR~\cite{zhu2020deformable} for object detection, in a plug-and-play manner.
\section{Experiments}
In this section, we evaluate the proposed BatchFormerV2 module for dense representation learning on popular dense prediction tasks, including object detection and panoptic segmentation. We also perform comprehensive ablation studies on common object detection. Lastly, as a more general solution than vanilla BatchFormer, we also perform experiments on image classification, demonstrating the generalizability of BatchFormerV2. Please refer to Appendix for the effectiveness of BatchFormerV2 on Long-Tailed Recognition, 3D Hand Mesh Reconstruction, and Self-Supervised Learning.
\subsection{Object Detection}
{\bf Implementation Details}.
We use the popular DETR/Deformable-DETR as our baseline methods, and apply the proposed BatchFormerV2 for object detection. We perform all experiments on the most popular common object detection benchmark dataset, COCO 2017~\cite{lin2014microsoft}, which contains 118k training images and 5k validation images. During training, the backbone network is initialized from the weights pretrained on ImageNet-1K~\cite{deng2009imagenet}.
We run experiments on eight Nvidia V100 GPUs (16GB) for DETR, and eight Nvidia A100 GPUs (40GB) for Deformable DETR. If not otherwise stated, the batch-size for DETR, Conditional DETR and SMCA is 16 and the default batch size for Deformable DETR is 24. If not stated, we insert the BatchFormer module in the first Transformer encoder layer in experiments. The number of heads in BatchFormerv2 is 4. For fair comparisons, all other hyperparameters follow the default configurations described in DETR~\cite{carion2020end},Conditional DETR~\cite{meng2021-CondDETR}, SMCA~\cite{gao2021fast} and Deformable DETR~\cite{zhu2020deformable}. An ablation study on the batch size can also be found in Section~\ref{sec:ab_obj}.
\setlength{\tabcolsep}{5.5pt}
\begin{table}
\begin{center}
\begin{tabular}{@{}lccccccc@{}}
\toprule
Method & Backbone & $AP$ & $AP_{50}$ & $AP_{75}$ & $AP_S$ & $AP_M$ & $AP_L$ \\
\hline
DETR~\cite{carion2020end}& ResNet-50 & 34.8 & 55.6 & 35.8 & 14.0 & 37.2 & 54.6 \\
+ BatchFormerV2 & ResNet-50 & {\bf 36.9} & {\bf 57.9} & {\bf 38.5} & {\bf 15.6} & {\bf 40.0} & {\bf 55.9 }\\
\hline
Conditional DETR~\cite{meng2021-CondDETR} & ResNet-50 & 40.9 & 61.8 & 43.3 & 20.8 & 44.6 & 59.2 \\
+BatchFormerV2 & ResNet-50 & {\bf 42.3} & {\bf 63.2} & {\bf 45.1} & {\bf 21.9} & {\bf 46.0} & {\bf 60.7} \\
\hline
SMCA (single scale)~\cite{gao2021fast} & ResNet-50 & 41.0 & - & - & 21.9 & 44.3 & 59.1 \\
+BatchFormerV2 & ResNet-50 & {\bf 42.3} & {\bf 63.5} & {\bf 45.4} & {\bf 22.5} & {\bf 45.7} & {\bf 60.1} \\
\hline
Deformable DETR~\cite{zhu2020deformable}& ResNet-50 & 43.8 & 62.6 & 47.7 & 26.4 & 47.1 & 58.0 \\
+ BatchFormerV2 & ResNet-50 & {\bf 45.5} & {\bf 64.3} & {\bf 49.8} & {\bf 28.3} & {\bf 48.6} & {\bf 59.4} \\
\hline
Deformable DETR$^*$~\cite{zhu2020deformable}& ResNet-50 & 45.4 & 64.7 & 49.0 & 26.8 & 48.3 & 61.7\\
+ BatchFormerV2 & ResNet-50 & {\bf 46.7} & {\bf 65.6} & {\bf 50.5} & {\bf 28.8} & {\bf 49.7} & {\bf 61.8} \\
\hline
Deformable DETR~\cite{zhu2020deformable}& ResNet-101 & 44.5 & 63.7 & 48.3 & 25.8 & 48.6 & 59.6 \\
+ BatchFormerV2 & ResNet-101 & {\bf 46.0} & {\bf 65.2} & {\bf 50.5} & {\bf 28.4} & {\bf 49.8} & {\bf 60.7} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Illustration of BatchFormerV2 on common object detection based on Deformable DETR~\cite{zhu2020deformable} and DETR~\cite{carion2020end}. Following Deformable DETR~\cite{zhu2020deformable}, we train all models with 50 epochs using the released code. Deformable DETR$^*$ indicates Deformable DETR with iterative bounding box refinement.}
\label{tab:obj_det}
\end{table}
\noindent {\bf Results}. As shown in Table~\ref{tab:obj_det}, BatchFormerV2 significantly improves the corresponding baseline methods. For example, without bells and whistles, BatchFormerV2 improves DETR by 2.1\% and Deformable DETR by 1.7\% when using a ResNet-50 backbone. We observe consistent improvement on Conditional DETR and SMCA. Moreover, we find that BatchFormerV2 mainly improves the object detection performance on small and medium objects. For example, BatchFormerV2 increases Deformable DETR in APS by 1.9\% and APM by 1.5\%, respectively. For DETR, BatchFormerV2 increases $AP_S$ and $AP_M$ by 1.6\% and 2.8\%, respectively. We think via building Transformer along the pixel of the feature map in the batch dimension, BatchFormerV2 utilizes features from other images to facilitates object detection in current image. For small objects which is usually challenging to detect, BatchFormerV2 is able to incorporate objects from other images to detect (refer to the Visualization in Section~\ref{sec:ab_obj}). Therefore, BatchFormerV2 significantly improves corresponding baselines.
\subsection{\bf Ablation Studies on Object Detection}
\label{sec:ab_obj}
To better understand the proposed method for dense representation learning, we perform ablation studies on some key factors that may have influences on the BatchFormerV2 performance. Please also refer to supplementary materials for more results.
\begin{table}
\begin{center}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{@{}cccccccc@{}}
\toprule
Batch Size & Epochs & $AP$ & $AP_{50}$ & $AP_{75}$ & $AP_S$ & $AP_M$ & $AP_L$ \\
\hline
16 & 50 & 44.7 & 63.5 & 48.9 & 27.3 & 48.1 & 59.1 \\
24 & 50 & {\bf 45.1} & {\bf 64.1} & {\bf 49.3} & {\bf 28.5} & {\bf 48.4} & 59.4 \\
32 & 50 & 44.9 & 63.8 & 48.8 & 27.7 & 48.3 & {\bf 60.0} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Ablation study on the training mini-batch size. We use Deformable DETR as our baseline and insert BatchFormerV2 module in the last Transformer layer.}
\label{tab:ab_obj_bs}
\end{table}
\noindent\textbf{Mini-Batch Size}. Considering that BatchFormerV2 aims to learn sample relationships among each mini-batch during training, we evaluate the influence of different mini-batch size on BatchFormerV2 as follows. As shown in Table~\ref{tab:ab_obj_bs}, 1) when increasing the batch-size from 16 to 24, the performance can be further improved with a small margin; 2) when further increasing the batch-size to 32, the performance is comparable, \textit{i.e.}, no additional improvements. Here, we maintain other hyper-parameters when increasing the batch size. We consider that it may require to tune other hyperparameters after increasing the batch size to achieve further improvements.
\begin{table}
\setlength\tabcolsep{4.pt}
\begin{center}
\begin{tabular}{@{}lccccccccccc @{}}
\toprule
& L1-2 & L1-3 & L3-6 & L4-6 & L5-6 & L1 & L2 & L3 & L4 & L5 & L6 \\
\hline
Accuracy(AP) & 45.2 & 44.9 & 45.4 & 44.9 & 45.0 & {\bf 45.5} & 45.3 & 45.2 & 45.2 & 45.2 & 45.1 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Ablation study on the insert position. Specifically, ``L1-3'' indicates that we insert BatchFormerV2 modules from the first layer to the third layer. ``L1" indicates that we only insert a BatchFormerV2 model in the first layer. Since the GPU memory limitation, we did not insert BatchFormerV2 in all layers. Instead, we selected some combinations that include all layers.}
\label{tab:ab_obj_layers}
\end{table}
\noindent{\bf Insert Position}. We show the influence of different insert positions for BatchFormerV2. Specifically, we use Deformable DETR as the baseline, which contains six Transformer layers. As shown in Table~\ref{tab:ab_obj_layers}, we find that: 1) the insert positions do have an important effect on the performance; and 2) the number of BatchFormer modules does not have significant influence on the final performance, \textit{i.e.}, more BatchFormerV2 modules cannot further improve the object detection performance; and 3) inserting BatchFormerV2 modules in early layers seems to be more effective for dense prediction tasks.
\setlength{\tabcolsep}{4pt}
\begin{table}
\begin{center}
\begin{tabular}{@{}lccccccc@{}}
\toprule
Method & Epochs & $AP$ & $AP_{50}$ & $AP_{75}$ & $AP_S$ & $AP_M$ & $AP_L$ \\
\hline
Deformable DETR & 50 & 43.8 & 62.6 & 47.7 & 26.4 & 47.1 & 58.0 \\
+ BatchFormerV2 (shared) & 50 & 44.9 & 63.6& 49.1 & 27.7 & 47.9 & 59.6 \\
+ BatchFormerV2 (non-shared) & 50 & {\bf 45.4} & {\bf 64.3} & {\bf 49.5} & {\bf 28.6} & {\bf 48.5} & {\bf 59.9} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Ablation study on shared BatchFormerV2 modules. Specifically, we show the influence when sharing the same BatchFormerV2 module among different layers. Here, we insert BatchFormerV2 modules from the third layer to the sixth layer.}
\label{tab:ab_obj_shared}
\end{table}
\noindent\textbf{Shared Modules}.
Furthermore, we test whether sharing or not sharing the BatchFormerV2 module among different layers would benefit the dense preidiction tasks. As shown in~Table~\ref{tab:ab_obj_shared}, we find that not-shared BatchFormerV2 could bring 0.5\% improvement compared to the shared scheme. It may suggest that the dense sample relationships are varying among different layers/levels, which also explains that, for different vision recognition tasks, the proposed BatchFormerV2 may add into different layers.
\begin{figure}
\centering
\includegraphics[width=.99\textwidth]{bt_attens.png}
\caption{Visualization of self-attention maps. Here, we insert a BatchFormerV2 module into the first Transformer layer in Deformable DETR.}
\label{fig:bt_atten}
\end{figure}
\noindent\textbf{Visualization}.
BatchFormerV2 are inserted into Transformer Encoder layers along each spatial position in the batch dimension, it enables the information propagation among samples in a mini-batch via the attention mechanism. Here, we visualize the attention of each slot on other slots along with the same position among the batch dimension. Particularly, we choose the largest scale feature map in Deformable DETR in the first Transformer Encoder layer to visualize.
Figure~\ref{fig:bt_atten} shows BatchFormerV2 mainly focuses on the objects (\textit{e.g.}, person, chairs), while paying less attention on the background (\textit{e.g.}, play ground, grass). This actually demonstrates BatchFormerV2 improves the object localization via the self-attention among different samples. We think this is verified to the result of Panoptic Segmentation (Section \ref{sec:seg}) in which the improvement of BatchFormerV2 on the segmentation of things is significantly better than that on the stuff segmentation. In addition, we observe self-attention highlights all the regions of the objects with clear boundaries, which explains the significant improvement of BatchFormerV2 on panoptic segmentation. Last but not least, in Figure~\ref{fig:bt_atten} the mini-batch samples in the test set are randomly constructed, which illustrates objects from different categories can be also mutually beneficial for the localization. Please refer to Appendix for more visualized comparisons.
\subsection{Panoptic Segmentation}
\label{sec:seg}
\textbf{Implementation Details}. We evaluate the proposed BatchFormerV2 for panoptic segmentation, which is a combination of instance and semantic segmentation tasks, on MS-COCO dataset. Specifically, we use the panoptic annotation provided by~\cite{kirillov2019panoptic}, which contains 53 \textit{stuff} categories in addition to 80 \textit{things} categories from the original MS-COCO dataset. We use DETR~\cite{carion2020end} as our baseline for panoptic segmentation, \textit{i.e.}, we utilize a mask head to generate panoptic segmentation results by treating both \textit{stuff} and \textit{things} classes in a unified way~\cite{kirillov2019panoptic}. Following~\cite{carion2020end}, we first train the model with BatchFormerV2 modules for object detection to predict bounding boxes around \textit{stuff} and \textit{things} classes 300 epochs. After that, we finetune the new mask head for extra 25 epochs.
\setlength{\tabcolsep}{4pt}
\begin{table}
\small
\setlength\tabcolsep{3pt}
\begin{center}
\begin{tabular}{@{}l|ccc|ccc|ccc|c@{}}
\toprule
Method & $PQ$ & $SQ$ & $RQ$ & $PQ^{th}$ & $SQ^{th}$ & $RQ^{th}$ & $PQ^{st}$ & $SQ^{st}$ & $RQ^{st}$ & $AP$ \\
\hline
DETR~\cite{carion2020end} & 43.4 & 79.3& 53.8 &48.2 &79.8 &59.5& 36.3 &78.5 & 45.3 & 31.1 \\
+BatchFormerV2 & {\bf 45.1} & {\bf 80.3} & {\bf 55.3} & {\bf 50.5} & {\bf 81.1} & {\bf 61.5} & {\bf 37.1} & {\bf 79.1} & {\bf 46.0} & {\bf 33.4} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Panoptic segmentation with DETR~\cite{carion2020end} on the COCO val dataset. $PQ^{th}$ and $PQ^{st}$ indicate the results on \textit{things} and \textit{stuff} classes, respectively.}
\label{tab:def_detr_seg}
\end{table}
\noindent\textbf{Results}. We report the panoptic quality ($PQ$) and the breakdown performances on things ($PQ^{th}$) and stuff ($PQ^{st}$) in Table~\ref{tab:def_detr_seg}. Specifically, we observe that BatchFormerV2 significantly improves $AP$ by 2.3\% and $PQ$ by 1.7\%. We also notice the improvement on $PQ^{th}$ is much larger than $PQ^{st}$. That is, BatchFormerV2 improves $PQ^{th}$ by 2.3\%, while the improvement on $PQ^{st}$ is only 0.8\%. This result is consistent with the results of object detection: by enabling the information propagation, BatchFormerV2 mainly facilitates object detection and instance segmentation.
Furthermore, following~\cite{carion2020end}, we actually freeze the bounding box branch and Transformer layers (include BatchFormerV2) when finetuning the mask head, we find that the performance of panoptic segmentation is also significantly improved. A possible explanation is that BatchFormerV2 improves the optimization of the backbone and the Transformer encoder for better object detail modeling for bounding box detection and subsequently facilitates the segmentation performance when finetuning the mask head.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{final_36918_18.png}
\caption{A visual comparison between DETR with and without using the proposed BatchFormerV2. Specifically, the first row is the original image, the second row is the panoptic segmentation result without BatchFormerV2, and the last row indicates the panoptic segmentation result with BatchFormerV2.}
\label{fig:panoptic_result}
\end{figure}
\noindent\textbf{Visualization}. We find BatchFormerV2 improves the details of segmentation and improves the segmentation on small objects. For example, we find with BatchFormer, the segmentation boundary of door is significantly improved, while the baseline model (\textit{i.e.}, DETR~\cite{carion2020end}) segments the door as wall by mistaking in the left area of Figure~\ref{fig:panoptic_result}. Meanwhile, the legs of the desk are also more clear. In the second figure, we can find the segmentation of airplane achieves better details (\textit{e.g.}, the tail and front wheels) with BatchFormerV2. In the last subfigure, the baseline model ignores the segmentation of grass (\textit{i.e.}, small stuffs), while model with BatchFormerV2 can correctly segment the grass.
\subsection{Image Classification}
\noindent {\bf Implementation Details}. We evaluate BatchFormerV2 for image classification using the most popular visual Transformer (ViT) using the same training strategy with DeiT~\cite{touvron2021deit}. We perform image classification on two popular benchmarks, \textit{i.e.}, CIFAR-100 and ImageNet. Specifically, we train all models for 300 epochs on ImageNet with an initial learning rate 1e-3 and the batch size 1024. For CIFAR-100, following~\cite{chen2021empirical}, we train all models with the initial learning rate 6e-4 and the batch size 1024. All experiments are conducted on a cluster with eight Nvidia A100 GPUs (40GB). When applying the proposed BatchFormerV2, we use the same number of heads with the corresponding baseline model. For ImageNet dataset, we insert BatchFormerV2 in the eighth layer. Emperically, we observe frequent crashes during training if we insert BatchFormerV2 module in very early layers. For CIFAR-100, we insert BatchFormerV2 module for all layers. More details and analysis are provided in supplementary materials.
\begin{table}
\begin{center}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{@{}lcccc@{}}
\toprule
Model & Params & Input & Top-1 & Top-5 \\
\hline
DeiT-Ti~\cite{touvron2021deit} & 5M & 224$^2$ & 72.2 & 91.1\\
+ BatchFormerV2& 5M & 224$^2$ & {\bf 72.7} & {\bf 91.5}\\
\hline
DeiT-S~\cite{touvron2021deit} & 22M & 224$^2$ & 79.8 & 95.0 \\
+ BatchFormerV2& 22M & 224$^2$ & {\bf 80.4} & {\bf 95.2} \\
\hline
DeiT-B~\cite{touvron2021deit} & 86M & 224$^2$ & 81.7 & 95.5 \\
+ BatchFormerV2 & 86M & 224$^2$ & {\bf 82.2} & {\bf 95.8}\\
\bottomrule
\end{tabular}
\end{center}
\caption{Image classification on ImageNet. We follow the same experimental setups described in DeiT~\cite{touvron2021deit}.}
\label{tab:imgnet}
\end{table}
\noindent\textbf{Results on ImageNet}. Table~\ref{tab:imgnet} demonstrates BatchFormerV2 consistently improves the performance among different ViT models. We observe BatchFormerV2 achieves similar improvement, \textit{i.e.}, around 0.5\%, compared to the baseline. Compared to the improvement on object detection and panoptic segmentation, the improvement on image classification is relatively small. It might be because dense prediction requires to localize the objects in the images, \textit{i.e.}, there are multiple targets in the image, while image classification treats the whole image as the target and requires to recognize the image. Furthermore, we think the strong data augmentation in classification might be also a limitation for BatchFormerV2 on ImageNet. We also analyze the effect of cutmix~\cite{yun2019cutmix} and mixup~\cite{zhang2018mixup} on BatchFormerV2 in supplementary materials. \\
\begin{table}
\begin{center}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{@{}lcccc@{}}
\toprule
Model & Params & Input & Epochs=100 & Epochs=300 \\
\hline
DeiT-Ti~\cite{touvron2021deit} & 5M & 224$^2$ & 49.2 & 69.2 \\
+ BatchFormerV2& 5M & 224$^2$ & {\bf 58.7} & {\bf 73.4} \\
\hline
DeiT-S~\cite{touvron2021deit} & 22M & 224$^2$ & 57.5 & 72.5 \\
+ BatchFormerV2& 22M & 224$^2$ & {\bf 68.5} & {\bf 75.2} \\
\hline
DeiT-B~\cite{touvron2021deit} & 86M & 224$^2$ & 52.2 & 71.8 \\
+ BatchFormerV2 & 86M & 224$^2$ & {\bf 66.6} & {\bf 74.8}\\
\bottomrule
\end{tabular}
\end{center}
\caption{Image classification on CIFAR-100. Following the experimental setups in~\cite{chen2021empirical}, we train all models from scratch and report the top-1 accuracy (\%). Specifically, DeiT-Ti, DeiT-S, and DeiT-B have the same architectures with ViT-Ti, ViT-S, and ViT-B, respectively.}
\label{tab:cifar100}
\end{table}
\noindent\textbf{Results on CIFAR-100}. Current vision Transformer architectures (\textit{e.g.}, ViT~\cite{dosovitskiy2020image}) usually require a large amount of training data or strong regularization to avoid severe overfitting problems. Therefore, it is still chanllenging for vision Transformer to train from scratch on a small dataset. In this paper, we also find that the proposed BatchFormerV2 module can significantly improves the performance of vision Transformer on small datasets. As illustrated in Table~\ref{tab:cifar100}, BatchFormerV2 significantly improves the performance of DeiT-B from 52.2\% to 66.6 \% by {\bf 14.4\%}, DeiT-S from 57.5\% to 68.5\% by {\bf 11\%}, DeiT-Ti from 49.2\% to 58.7\% by {\bf 9.5\%}. When we train all models with more epochs, \textit{i.e.}, 300 epochs, the improvement is also considerable. The possible reason is that BatchFormerV2 enables the information propagation among patches in different images, which benefits the optimization and generalization when learning on small datasets. Particularly, we find that DeiT-B does not achieve better performance compared to DeiT-S. This is possibly because DeiT-B is a too large model for a very small dataset, \textit{e.g.}, CIFAR-100.
\section{Limitations and Future Work}
Though the proposed BatchFormerV2 module is very effective and efficient for common vision tasks, including image classification, object detection and panoptic segmentation, we also empirically observe some frequent training crashes (i.e., model diverged with loss=nan) when we insert the BatchFormerV2 module into very early layers, especially in image classification task. In future, we will further explore how to smoothly use the BatchFormerV2 module in different layers of image classification models.
\section{Conclusion}
Attention mechanisms have attracted intensive interests from the communities of natural language processing and computer vision. Previous approaches mainly investigate 1) self-attentions on either channel or spatial dimensions; 2) batch attentions for image-level sample relationship learning under data scarce settings. In this paper, we present a more general batch Transformer module, termed as BatchFormerV2, to explore sample relationships for dense representation learning. Specifically, the proposed BatchFormerV2 module can be easily integrated into existing vision Transformer architectures for sample relationship modeling within each training mini-batch from either pixel/patch- or image-levels. Meanwhile, we further propose a two-stream training pipeline for BatchFormerV2, where two streams share all other layers/blocks except the BatchFormerV2 modules. By doing this, BatchFormerV2 can thus be a plug-and-play module and easily integrated into different vision Transformers without introducing any extra inference cost. Extensive experiments on a variety of visual recognition tasks, including image classification, objection detection, and panoptic segmentation, demonstrate the effectiveness of the proposed BatchFormerV2 module for robust representation learning.
\bibliographystyle{splncs04}
|
\section{Introduction}
The occurrence of geomagnetic storms is associated with input of the energy of the solar wind and subsequent coupling of the same with the Magnetosphere-Ionosphere (MI) system. In general, these storms occur when the north-south component ($B_z$) of the frozen-in Interplanetary Magnetic Field (IMF), emanating from the Sun and carried by the solar wind into the heliosphere, turns completely southward (or negative $B_z$ by convention) and remains in this state for several hours while reconnecting with the geomagnetic field [1-2]. These phenomena occur mainly following a Coronal Mass Ejection (CME) or successive CMEs and the type of magnetic storms driven by this solar transient are generally refereed to as CME-driven geomagnetic storms. Additionally, these type of storms occur from components like the strong magnetic field in the ejecta, the strong magnetic field of the sheath and the interplanetary shocks. It is to be noted that these storms mainly occur in the ascending phase and solar maximum phase of solar cycles (or solar activity cycle of $\approx$ 11 years duration). The Storm Sudden Commencements (SSC), before the onset of the main phase of these type of storms, are associated with the interplanetary shocks that are driven by the CMEs while the recovery phase lasts for only about one-two days [3-4]. A geomagnetic storm's beginning is noted with large variations of $B_z$ in addition to the transmission of magnetospheric convection electric field (with an associated equivalent current system having average periods of the order of few minutes to about 3 hours) from the higher to the lower-equatorial latitudes, also called as the Prompt Penetration Electric Fields (PPEF and eastward during dawn-to-dusk) which cause severe perturbations to the low-latitude electrodynamics [5-10].
The ionosphere of the low-latitude region is characterized by a phenomenon known as the Equatorial Ionization Anomaly (EIA), that starts around 09 Local Time (LT) and moves towards higher latitudes up to 15-20${^\circ}$ magnetic dip latitude around 16 LT and returns back to the magnetic equator around 24 LT. This phenomena is caused as a result of the two plasma motions: one perpendicular to the geomagnetic field and generating drifts upward the zonal electric field (eastward during daytime and westward during nighttime) produced by E region dynamo and the geomagnetic field and the other parallel to the geomagnetic field, that cause the plasma to diffuse down following the geomagnetic field line under influence of ambipolar diffusion drift related to gravity and pressure gradients [11-16].
The present work investigates the changes in the plasma drift, near the dip equator and the EIA regions, that were caused due to the influence of the CME-driven geomagnetic storm of October 13, 2016 by utilizing simulations from the Thermosphere-Ionosphere-Electrodynamics General Circulation Model (TIE-GCM) developed by NCAR. Firstly, the analysis is performed by simulating the vertical plasma drifts at Indore:IDR (a location near the EIA, magnetic dip of 32.23${^\circ}$N) and Tirunelveli: TIR (a dip equatorial location, magnetic dip of 0.2${^\circ}$N) under no variations of the IMF $B_z$ (i.e: $B_z$ = 0 nT) and then by simulating the same for these two locations, with observed values of $B_z$ during the entire day of October 13, 2016. The motivation of the work comes from the need of understanding these physics-based models and their performance such that these models can be reliably run across locations where actual data from ionosondes or GNSS receivers are not present. The paper is divided starting with a general discussion of the NCAR TIEGCM model followed by the results, the corresponding discussions and the summary sections.
\section{Data: The TIEGCM}
The National Center for Atmospheric Research (NCAR) TIE-GCM (https://www.hao.ucar.edu/modeling/tgcm/tie.php) is a 3D, non-linear and first principles representation of the coupled ionosphere and thermosphere system. It consists of self-consistent solution of the mid-and low-latitude dynamo field. It solves energy, continuity and momentum equations for ion and neutral species at every time-step (typically 120 s) using a semi-implicit, $\nth{4}$-order, finite difference method on each pressure surface in a vertical grid. The standard parameters in this model are: latitudinal and longitudinal extent from -87.5${^\circ}$ to +87.5${^\circ}$ and -180${^\circ}$ to +180${^\circ}$ respectively, in steps of 5${^\circ}$, altitude-wise pressure level from -7 to +7 in steps of $\frac{H}{2}$ with lower boundary $\approx$97 km and upper boundary $\approx$500-700 km (depending on the solar activity condition). The inputs to this model are: daily solar radio flux (F10.7) and its 81-day centered mean (F10.7$\_$81) while the output parameters (specified in 3D spatial and time dimensions) are: Height of pressure surfaces in cm, electron, ion and neutral temperatures in K, compositions like N2+, NO+, N+, Ne, NO, N(4S), N(2D), O, O2, O+, O2+, meridional, vertical and zonal drifts in m/s and potentials both geographic and geomagnetic coordinates. Several researchers [17-22] have shown the capability of the TIEGCM to understand geomagnetically disturbed as well as quiet-time conditions' ionospheric electrodynamic processes thus making it an ideal simulation platform that closely reproduces and resembles the same.
\section{Results and Discussions}
\subsection{Interplanetary and geomagnetic conditions}
A CME erupted into the space as a result of the eruption of a magnetic filament from the northern hemisphere of the Sun around 16 UT on October 08, 2016. The CME arrived at L1-point around 21:21 UT on October 12, 2016. Following the events, a G2 class ($K\_p$ = 6, NOAA scales: http://swpc/noaa.gov/noaa) geomagnetic storm commenced on October 13, 2016 at around 08:15 UT (https://www.spaceweather.com/archive). Figure \ref{sc01} from top panel to bottom panel shows the IMF $B_z$ (in nT), the average solar wind flow speed Vsw (in km/s), the Interplanetary Electric Field (IEF) $E_y$ (in mV/m) and the one-minute average of Dst index: SYM-H index (in nT) during October 12-14, 2016. The shaded region (24-48 UT) shows the day (October 13, 2016) of the main phase of the geomagnetic storm. The IMF $B_z$ had been below -10 nT during October 13, 2016 from 08:26 UT to 21:58 UT. It reached minimum value of -21 nT on 15:18 UT of the same day, thus signifying strong coupling between the solar wind and the MI system. The average Vsw values on October 12, 2016 had been around 367 km/s, rose to values around 411 km/s during the day of the storm main phase (October 13, 2016) and dropped back to about 369 km/s on October 14, 2016. The IEF $E_y$ had also showed a peak on October 13, 2016 around 16:18 UT with a value of 8.77 mV/m, as a result of higher values of Vsw and strong southward IMF $B_z$ conditions on October 13, 2016. The SYM-H here shows minimum value of -114 nT on 23:45 UT, thus designating the event to be a strong/intense (Dst/SYM-H $<$-100 nT) geomagnetic storm [23].
\begin{figure}[htbp]
\centering
\includegraphics[width=85mm]{omni-plot.png}
\caption{From top to bottom: IMF $B_z$ (in nT), Vsw (in km/s), IEF $E_y$ (mV/m) and the SYM-H (in nT) from October 12 through 14, 2016. The shaded rectangle (24-48 UT) marks October 13, 2016: the period of main phase of the geomagnetic storm.}
\label{sc01}
\end{figure}
\subsection{TIEGCM results}
Figure \ref{sc02} shows the TIEGCM simulation results on October 13, 2016. As discussed in the introduction section, the locations observed are IDR and TIR which are locations near the EIA and the geomagnetic equator passing the Indian subcontinent. The top panel shows the vertical drift variation of both IDR (in blue) and TIR (in red) when IMF $B_z$ was set to 0 nT (scenario 1). The bottom panel shows the same with the observed IMF $B_z$ values as obtained from the omniweb database (scenario 2). On investigating the simulation results, it can be seen that when the model is fed with actual or observed data, the difference in the eastward (positive values) peak of the vertical plasma drift reduces while there is an increase of the same in the westward (negative values) peak. Furthermore, feeding of the observed IMF $B_z$ reduces the overall level of the eastward drift but increases the westward drift after the Pre-Reversal Enhancement (PRE) around 13 UT (or 18:30 LT). It is also to be noted that at an off-equatorial location (IDR), the PRE gets suppressed in both the simulation scenarios, however there is no effect of the same on equatorial location's (TIR) PRE levels. The possible reason of these suppressions could be attributed to the disturbance dynamo effect which is getting reflected in an off-equatorial location. From this figure it is evident that when applying real data of the IMF $B_z$, the overall level of the drift is getting reduced (more westward). Thus indicating a possible mechanism that is affecting the vertical drifts of both off-equatorial and equatorial locations under geomagnetic storm effect.
\begin{figure}[htbp]
\centering
\includegraphics[width=85mm]{tiegcm-plot.png}
\caption{Variation of the vertical plasma drifts over an equatorial location (TIR in red) and an off-equatorial location (IDR in blue) on the day (October 13, 2016) of main phase of the geomagnetic storm event, when Top panel (Scenario 1): IMF $B_z$ is set to zero throughout the TIEGCM run and bottom panel (Scenario 2): when actual values of $B_z$ is fed as inputs to the model throughout the main phase period.}
\label{sc02}
\end{figure}
\section{Summary}
The low-latitude ionosphere is dynamic in nature, especially in and around the EIA, where almost two-thirds of the global ionization gets concentrated. In addition, the perturbations might increase or decrease when there is an external driver such as geomagnetic storms as a result of solar outburst and subsequent strong coupling of the solar wind and MI system. There arises a need to understand the effect of the IMF parameters (especially $B_z$) during such events on the plasma motion in the form of drift from equatorial to off-equatorial locations. This study compares the TIEGCM simulation results of the vertical plasma drift over both EIA and magnetic equator in the Indian sector under the scenario of no coupling of the solar wind with MI system ($B_z$ = 0 nT) and a scenario with actual observations of $B_z$ fed to this model. Observations brought forward the fact that there had been significant difference in the variation of the vertical plasma drift, in the form of suppression of the entire profile towards westward, when the model was fed with actual observations of IMF $B_z$. Further studies with variations of the IMF conditions, under geomagnetically disturbed as well as quiet-time conditions, fed to the TIEGCM model could bring out new physical mechanisms that are taking place at such highly geosensitive locations over low-to-equatorial latitudes, especially in and around the EIA and the magnetic equator. This initial work will be followed up by thorough investigation of the effects of the other IMF components on the global plasma drift motions over these dynamic locations.
\section{Acknowledgements}
SC acknowledges the World Data Center (WDC) for Geomagnetism, Kyoto for the Dst index value. The solar wind and interplanetary magnetic field and the SYM-H index data are obtained from NASA’s SPDF omniweb service https://omniweb.gsfc.nasa.gov/form/omni\_min.html. This work is supported by the Department of Space, Government of India.
|
\section{Didier's comments:}
Here are some thoughts that you could implement such that this paper is officially finalized within the H2020 time limits:
You could implement small changes in the abstract to emphasize the added value of your paper: 1) you distinguish among various types of technologies, 2) you have a large number of papers that cover quite a long time period of publications.
Same hold for the intro.
There, I would be more systematic when indicating how yours adds to other studies.
I am including pictures of the 1st 3 pages with specific suggestions.
For a next version, since I guess you’d like to submit to a Journal at some point:
I think it would be helpful to come up with some overall quantitative measure of the effects sizes reported in literature. This is not easy, but maybe finding out a way to standardize the effect sizes would help. I happy to think along w you how to do this.
In Section 3 or in the Appendix, you could report the count of articles by main WoS categories and Jnl titles.
I could look more carefully at the next version before you submit to a Journal.
\newpage
\section*{To dos and notes for revisions}
\begin{itemize}
\item K: Maybe if we want to have a critical note on the research question itself, we can refer to \url{https://www.aeaweb.org/articles?id=10.1257/aer.20151720}: The authors basically say that you have to work less when you live in a rich country. This interacts with the real income effect.
\item Final checks: Hyphenation. Yes, no?
\item List of hyphenation issues: -level, bias, -skill, -induced and -inducing, non- (e.g. non-production), production function based, saving, augmenting.
Note: When using Ctr+F and replace, go step-wise through the search results and verify every replacement manually. Not every occurrence of level, skill etc. needs hyphen (in case we decide to use hyphens).
\item Minor: order multiple references based on year or author names?
\item We could also capitalize Replacement, Reinstatement, Real income; but not sure. If we do this, we should also capitalize macro, meso, micro and then maybe even supports, depends, weak, no support, etc.
\end{itemize}
Post-Technequality submission:
\begin{itemize}
\item To help the reader, we might consider to create a table that summarizes our definitions of the different categories (technology, level of analysis, method).
\item Reclassify: \citet{padalino1997employment,hoskins2000,madariaga2018employment,reijnders2018tech,jiang2019subst} (or at least most of them) from other/indirect to TFP-style
\item Think about merge of ``weak'' and ``depends'' to ``ambiguous''
\item Not sure, but maybe it is also interesting to have another look at the subset of papers that talk about all three effects simultaneously.\footnote{\textcolor{orange}{These are 16 studies in total and they include key studies in the literature.}} Do the offer the ``gold standard'' to study the impact of techn. on employment? What are the limitations to this type of study? Maybe we can focus here on policy conclusions and we should discuss the net-employment effect. We may structure this part again along the three effects. Here we should also talk about the publication bias and cite the Ugur et al. paper. MS: approach of this section: compare ``support studies'' for replacement effect VS ``support studies'' for reinstatement + real income effect. And maybe end every paragraph with studies that look at net employment? I also try to check whether level of analysis matters.}
\item add in the existing surveys to the literature in the intro \url{https://cepr.org/active/publications/discussion_papers/dp.php?dpno=16868#}
\item Also there is this article which is relevant and in line with our findings, maybe we can add this in the end of intro or towards the conclusion \url{https://www.economist.com/finance-and-economics/2022/01/22/economists-are-revising-their-views-on-robots-and-jobs?utm_medium=social-media.content.np&utm_source=facebook&utm_campaign=editorial-social&utm_content=discovery.content&fbclid=IwAR033nOauO0_iagqp6xCdohn_Oo2oESA2ZkO6O9H5wSic1psRADSfcSZ5D0}
\end{itemize}
Comments Melline
\begin{itemize}
\item: Maybe we should consider to discuss the different types of technologies in more depth. At the moment, we do not discuss whether we expect to find different employment effects across the different technology types. Also, we should maybe provide short descriptions of what the technology categories entail. E.g. what is it that robots can do? Do the different categories overlap? E.g. is process innovation related to automation? And what is organizational innovation? Also, how are R&D and patents different from the innovation indicators? Maybe a short description of what total factor productivity is? In the end, we do not find very different results for the different types of technologies (except for product innovation maybe), so it would be good to reflect on this I think.
\item: I would like to improve the contribution part in the introduction section. There are some recent papers we should probably add, and I would like to have a better look into previous papers. I did not spend too much time on this before we submitted the paper for Techenquality.
\item Especially in case we need to shorten the paper, maybe we should consider to put the main emphasis on the "support" sections in the replacement, reinstatement and real income effect chapters. Also because there is a publication bias of course. And people will sell a paper as a "replacement paper" once they find support for this, I would expect. We could maybe shorten the sections on "no support", because it partly overlaps with the other two "support" sections.
\item I'm still wondering whether we should talk more about the "productivity" effect, in addition to the reinstatement effect. So with the productivity effect I refer to the complementarity between a machine and human labor. E.g. a computer saves a high-skill worker a lot of time when it comes to searching information (routine task). Therefore, the high-skill worker can spend more time on non-routine tasks in which it has a comparative advantage. The worker becomes more productive, and the demand for this type of worker might increase. It's somewhat different from the reinstatement effect, because here it's about working with the machine. And the reinstatement effect rather refers to additional labor around the machine, e.g. maintenance, the production of it. I know that most studies wont be able to distinguish the two effects (only some can). I also dont know whether the reinstatement or productivity/complementarity effect typically dominates, but it could still be worthwhile to discuss.
\item As suggested by Didier, we could consider to also analyze the effect sizes reported by studies.
\item A more general remark, although I'm not sure how to quickly address it: I miss some overall reflections in the paper and maybe some extra policy implications. The paper now mainly summarizes the results in the literature, but an additional critical reflection would be nice. But this is always the hardest part in review studies.
\item Adapt for journal submission? Check author guidelines. Do we need to shorten it for the JEL?
\item Minor: Address red comments in text
\end{itemize}
\begin{itemize}
\item Contributions:
\begin{itemize}
\item Systematic overview of studies of technological change: How to measure technologies and what data sources are available? This provides guidance for future empirical research aimed to understand the interplay between technology and the economy.
\item We introduce a comprehensive conceptual framework of technology-labor interactions that is systematically applied to the existing empirical literature. This framework helps researchers to embed their own research within the complex multi-order interplay of technology and labor.
\end{itemize}
\item Limitations:
\begin{itemize}
\item Real income effect most difficult to study. Here, we limited the review to those studies that explicitly study technology-labor interactions while simultaneously taking the productivity and wealth effects of technology into account. We are aware that this excludes numerous studies that focus on individual channels of the real income effect. [MAYBE THIS SENTENCE SHOULD BE MENTIONED ABOVE IN METHODS SECTION]
\end{itemize}
\end{itemize}
\begin{itemize}
\item Journal of Economic Literature (JEL): for non-commissioned articles we need to write a 5-10 page proposal and submit for consideration (see: \url{https://www.aeaweb.org/journals/jel/about-jel/editorial-policy}). However, we could possibly reach out to the editorial board (\url{https://www.aeaweb.org/journals/jel/submissions/} ) and suggest whether we can directly submit the paper instead of a proposal.
MS: I tried to look up author guidelines for the JEL, but could not find anything on maximum number of words. I guess for some journals we might have to shorten the paper.
\item Journal of Economic Perspectives (JEP): also a tough one to get in. For non-solicited articles, proposals for topics and authors should be directed to the journal office (see: \url{https://www.aeaweb.org/journals/jep/about-jep} )
\item Annual Review of Economics (ARE): maybe a bit more clubby in terms of style and potentially our paper is a bit off content-wise (see: \url{https://www.annualreviews.org/page/authors/editorial-policies})
\item Oxford Economic Papers: a non-review journal that could accommodate such an article (see: \url{https://academic.oup.com/oep/pages/About})
\item Journal of Economic Surveys: a definite match (see: \url{https://onlinelibrary.wiley.com/journal/14676419})
\end{itemize}
\section[{Re-coding of table entries}
\begin{itemize}
\item Discuss merge weak/indirect. Maybe even only two categories: Support - no (clear) support.
\item \cite{padalino1997employment}: Check this again; maybe rather TFP-style that "other".
\item Discuss TFP - Real income relationship
\end{itemize}
\end{comment}
\newpage
\section{Introduction}
The fear that technological advancement will make human labor obsolete is not new
\citep{mokyr2015history}. During the first Industrial Revolution, the adoption of power looms and mechanical knitting frames gave rise to the Luddite movement protesting against this technology by destroying textile machinery out of fear of job losses and skill obsolescence. The idea that technology can render workers redundant, at least in the short run, has also been supported by influential economists like Karl Marx and David Ricardo in the nineteenth century \citep{marx1988economic, ricardo}. Others, such as Thomas Mortimer, believed that machines could displace labor more permanently \citep{mortimer1772elements}. In the same spirit, more recent concerns about massive job losses \citep{smith2017automation} stem from improved computing power, decreasing costs, and advances in machine learning and robotics \citep{brynjolfsson2014second}.
The current debate on the labor market impact of technological change has also been fed by a number of influential studies predicting striking job losses in occupations that are susceptible to automation \citep{arntz2017, frey2017future, nedelkoska2018automation}.
However, this view is controversial and has been questioned in recent research highlighting the positive employment effects of technology \citep{aghion2022effects, economist2022robots}.
At the aggregate level, there is little evidence that technological change has led to widespread unemployment over the centuries. But technology-induced job losses can be significant. Many technologies are designed to save labor by replacing human workers by machinery. Think of tractors substituting manual labor, assembly lines replacing human handiwork and computer assisted technologies that increasingly replace workers in performing explicit and codifiable tasks \citep{nordhaus2007two}.
However, economic theory points out that several compensating mechanisms can counterbalance the initial labor-saving impact of new technologies \citep[e.g.][]{vivarelli2014innovation, acemoglu2019automation,baldwin2021jobs}. First, technological change can increase the demand for labor by creating new tasks and jobs that are directly associated with the new technology.
Furthermore, if automation technology increases productivity it releases production resources which can raise the demand for labor in non-automated tasks within the same firm or industry.
Second, technology can also increase the demand for labor through the demand side when new technologies boost productivity growth leading to lower production costs and consumer prices. Moreover, new technologies can raise the marginal product of labor and capital resulting in higher wages and returns to capital. These two effects contribute to a rise in real income. If demand is sufficiently elastic and positively responds to increases in income and decreases in prices, technologies may lead to a demand-induced expansion of output \citep{bessen2020autom}.
In this paper, we referred to these channels through which technology affects the demand for labor as (1) replacement, (2) reinstatement, and (3) real income effect, and analyzed the empirical basis for each effect through a systematic literature search. We systematically identified and reviewed 127 studies published between 1988-2021 that provide evidence on technological change in industrialized economies in the post-1980s. To uncover potential underlying heterogeneity, we pinned down five broad technology groups that are predominantly studied in the literature distinguishing between (1) information and communication technology (ICT); (2) robot-diffusion; (3) innovation surveys; (4) productivity; and (5) a residual category that contains various alternative indicators.
Overall, we found that the number of studies supporting the labor replacement effect is more than offset by the number of studies supporting the labor-creating reinstatement and real income effect. This observation is also supported by the studies analyzing the net employment effect of technological change suggesting the net impact of technology on labor to be rather positive than negative.
The findings for the five distinct technology categories show broadly similar patterns, but with some subtle differences. For ICT, we found no evidence that the replacement effect dominates the reinstatement and real income effect combined. However, our results show that the reinstated jobs qualitatively differ from the jobs replaced. The diffusion of ICT mostly had positive employment implications for high-skill workers, non-routine labor, and service jobs.
Also for robot-diffusion we found that the labor-saving impact is generally offset by robot-driven reinstatement of labor. In contrast to the ICT studies, robot studies remain silent on the complementarity between robots and human labor when performing tasks. Hence, the labor-creating effect of robots is most likely related to the production and maintenance of this type of technology.
For studies that rely on innovation as a measure of technology, the employment impact depends on the type of innovation. While product innovation is mostly labor-creating, the evidence on the employment impact of process innovation is somewhat mixed.
For the fourth category of studies relying on productivity as technology measure, we found a roughly equal balance between the number of studies suggesting support for the replacement and the two labor-creating mechanisms. The employment effects have been mostly favorable for non-production workers, high-skill labor, and service jobs. These studies support the idea that technological change leads to structural change with a reallocation of economic activity down the supply chain from more primary towards increasingly processed sectors and services \citep{kruger2008productivity, syrquin1988patterns}.
However, the net employment effects observed by these studies are rather negative than positive.
Lastly, the findings from studies that rely on other/indirect measures of technology indicate that the labor replacing effect is offset by the labor-creating effect. The employment effects have been mostly positive for non-production labor, yet some studies also found positive employment effects for low-skilled workers, particularly in service jobs.
Overall, although we find larger support for the labor-creating effects of technological change, we are careful in concluding that technology has a net positive effect on employment. However, we do safely conclude that the labor replacing effect of technology is typically more than offset by a range of compensating mechanisms suggesting that the widespread anxiety over technology-driven unemployment lacks its empirical base.
Our study is not the first to review the existing evidence on the effect of technology on employment, but our systematic approach contributes to the literature in several ways. First, most reviews are narrative which may be subject to the authors’ bias \citep{brown2002tech, bruckner2017frontier, calvino2018innovation, goos2018tech, mondolo2021composite, vivarelli2014innovation, aghion2022effects}. Second, earlier studies that provide systematic reviews of the literature focused on shorter time periods, i.e. 1980-2013 in \citet{ugur2018technological}, 2005-2017 in \citet{ballister2018}, 2000-2021 in \citet{perezacre2021participation}, while our review covers thirty-three years (1988-2021) which allows us to capture a more complete picture of the effects for each of the technologies examined. Third, in contrast to \citet{ugur2017litreview} who synthesized the evidence on less developed countries, we focused our review on developed economies aiming to capture the impact of technological change at the frontier. Fourth, while most earlier reviews restricted their analysis to specific types of technology, we analyzed how employment effects differ across five alternative measures of technological change. Fifth, to be as inclusive as possible, we did not limit our analysis to specific measurements of employment. Although evidence from different models (e.g. derived labor demand, skill/wage share, and decomposition analyses) do not necessarily yield comparable estimates, they are informative about the direction of the employment effect.
This analysis provides an empirical basis for the political and scientific debate on technology and labor, whereby the anxiety of a jobless future appears to be overstated given the poor empirical basis for this claim. Nevertheless, our results also suggest that low-skill, production, and manufacturing workers have been adversely affected by technological change. Hence, effective up- and reskilling strategies should remain at the forefront of policy making. Many occupations in the lower end but also in the middle part of the skill distribution will continue to evolve and demand a changing set of skills due to technological progress. Hence, the employment perspectives of relatively vulnerable groups can be significantly improved by investing in the right set of skills. Nonetheless, some workers who experience job loss might not be able to engage in upskilling or make the transition to new jobs. For these groups, targeted social support systems will be important.
This paper is structured as follows. Section \ref{sec:conceptual_framework} presents the conceptual framework and discusses the mechanisms through which technological change can increase or decrease the demand for labor. In Section \ref{sec:methods}, we discuss our methodology and Section \ref{sec:results} presents the results. Section \ref{sec:discussion} discusses the findings and Section \ref{sec:conclusion} concludes.
\section{Conceptual framework}
\label{sec:conceptual_framework}
We apply a three-stage framework to disentangle the interplay between technology and employment. Along these stages, we review the current state of empirical knowledge.
We rely on a generic understanding of technology as the capability to transform a given set of inputs into outputs. Technological change happens when the quantity and/or quality of inputs or output change \citep{saviotti2013co, ruttan1959usher}. For example, new technologies may make production processes more efficient, enabling firms to produce the same goods with less labor or material inputs. It can also be reflected in the output when technologies enable firms to bring new products to the market.
How does this interact with labor? Here, we focus on three key mechanisms which are incrementally more indirect. To illustrate these mechanisms, we introduce a stylized model with a generic production function:
\begin{equation}
Q = A^Q f(A^L L, A^X X)
\end{equation}
where $Q$ is output, $L$ is the amount of labor used along with other inputs $X$ to produce a quality-adjusted output $Q$. Other inputs can be capital goods, material and intermediate inputs, or different forms of labor (e.g. different occupations or differently skilled workers). Note that the level of output is quality-adjusted and it can increase either through a higher number of units or a higher quality of a given unit. The parameters $A^L$, $A^X$, and $A^Q$ represent the production technology. The production function is non-decreasing in its arguments $A^L$, $L$, $A^X$, and $X$, whereby a higher level of technology or production inputs leads to an increase in output $Q$. Technological change may enter in different forms by changing $A^L$, $A^X$, and/or $A^Q$.
\subsection{The replacement effect}
The most direct impact of technology on employment is the so-called ``replacement effect''. This effect occurs when the adoption of a new technology enables a firm to reduce labor inputs for a given quantity of output.
In the stylized model above, pure replacement happens if $A^L$ increases and $Q$ is constant, i.e. $dQ = 0$. This means that less labor is used but everything else remains equal. However, not every type of technological change leads to an increase in $A^L$, and even if this is the case, it only replaces labor if output $Q$ does not expand sufficiently.
Other forms of technological change can lead to an increase in $A^X$ which means that the same amount of output can be produced with lower input requirements $X$.
Technological change may also lead to an increase in $A^Q$ which increases the level of output $Q$ while keeping the input requirements constant. Examples are efficiency improvements reflected in a higher total factor productivity (TFP). Also product innovations such as the introduction of a new design can be captured by $A^{Q}$, if they enable firms to bring new and better products to the market while not changing their input requirements.
Empirically, it is challenging to measure whether or not technological change is labor replacing. In particular, this effect is likely heterogeneous across industries and occupations, and often difficult to draw causal inference at a sufficiently granular level, especially when new jobs are created at the same time.
For example, the introduction of a product innovation may coincide with changing input requirements reflected in the amount and type of labor. It may also be that labor-saving technological change does not necessarily lead to layoffs, but those employees that are no longer required to produce $Q$ find other useful tasks within the firm.
In this research, we try to find out whether technological change has been labor-replacing in the post 1980s, i.e. whether technological change reflected by $A^L$ or $A^Q$ had a negative impact on the demand for labor. We use various indicators that allow us to draw conclusions about the existence of the replacement effect. At different levels of aggregation, changes in employment constitute our key indicator. Empirical support for the replacement effect exists if we observe a technology-induced decrease in employment in those firms, industries, and countries where the technology is adopted. Measures of employment include the employment rate, number of workers, and hours worked. The labor share of income is also indicative but not sufficient to provide evidence for the labor-saving impact of technology. Moreover, a number of studies at different aggregation levels examine changes in the relative employment of different occupational groups which we also consider as indicative for the replacement effect, such as a technology-induced increase in the ratio of high- over low-skill labor use. Again, this is only an indicative but not sufficient piece of evidence.\footnote{For example, technology-induced changes in the relative demand for low-skill labor at the aggregate level do not necessarily mean that labor was replaced. It may be that technological change enabled the emergence of a new industry that uses skilled labor differently. This may induce a shift in the relative demand, but in absolute terms not a single worker was replaced by machinery.}
We also interpret micro (worker or firm) level studies that assess the relationship between the type of tasks performed by workers and the likelihood of being displaced as indicative of the replacement effect, as some tasks may be more susceptible to automation.
Another indicator of the replacement effect includes changes in the elasticity of substitution of labor and other inputs $X$. Technological change may alter this elasticity. A technology-induced increase in the elasticity indicates that the technological possibilities to replace labor by other inputs have improved. We also interpret this as supporting evidence.
\subsection{The reinstatement effect}
The reinstatement effect is the next indirect effect of technological change. It occurs if the adoption of a new technology induces the creation of new jobs that are directly linked to the new technology, regardless of whether or not technological change happens via $A^L$, $A^X$, or $A^Q$.
The reinstatement effect is often associated with an increase in $Q$, otherwise technological change would be only input-saving even though the effects may be heterogeneous across different groups of employees.
The creation of new jobs may be driven by different mechanisms that are empirically difficult to disentangle: Workers performing tasks that cannot be automated may experience a boost in productivity which increases the demand for these jobs. Furthermore, new jobs may be created if technology enables new fields of economic activity.
For example, an input-saving ($A^L$ or $A^X$) technology may induce the creation of new jobs within the same firm for the operation and maintenance of the technology. A firm may also start supplying goods to new customers if input-saving technological change made the outputs more affordable, or if technological change affected the quality of the output which expanded its range of applications. For instance, the introduction of computers at the workplace creates new complementary tasks related to programming, hard- and software maintenance, and data management.
Depending on the level of aggregation, the reinstatement effect also refers to jobs created upstream or downstream the supply chain, i.e. jobs associated with the production of $X$. For example, the suppliers of capital or intermediate inputs required to operate the new technology may increase their demand for labor if $X$ is used more intensively. Downstream industries may expand output output if upstream innovation reduces prices of intermediates. Hence, the reinstatement effect exists if $\partial L / \partial A > 0$ for any $A = A^L, A^X, A^Q$.
Here, we screen the empirical literature on whether or not it provides supporting evidence for the existence of this effect. Again, the measurement is complex as the technology-induced reinstatement of new jobs may happen at different levels of aggregation, i.e. in the same firm and/or in other industries. Hence, studies limited to a subset of firms or industries cannot capture the reinstatement of labor elsewhere.
An increasing demand for labor is the key indicator of supporting evidence for the reinstatement effect. This is reflected in lower unemployment, an increasing number of employees, and hours worked.
Note that the reinstatement effect does not need to be equally distributed across different types of labor and may co-exist with the replacement effect. To support the existence of this effect, it is sufficient if we observe an increase for at least one group. We also consider changes in the relative demand for labor as suggestive evidence for the existence of the reinstatement effect, as it may be driven by an increase in the demand for certain types of labor.
Whether the net impact of technology on employment is positive or negative depends on the balance between labor replacement and reinstatement.
\subsection{The real income effect}
The two effects introduced above mainly refer to the direct impact of technology on the production side when it changes the use of inputs in absolute and relative terms. Technological change also affects labor through an indirect channel that mostly operates through the demand side.
Assuming rational technology adoption decisions, technological change is always associated with productivity improvements; otherwise it would be irrational to adopt a new technology. Productivity improvements enable firms to produce a given value of output at lower costs which would be reflected in lower consumer prices $P$ if input costs savings are transmitted to consumers. Moreover, if technological change raises the marginal product of certain types of labor, we expect wages $w$ to rise. If technological change raises the marginal product of capital $K$, we also expect higher rents to capital which are another source of income. All these effects (lower prices $P$, higher wages $w$, higher returns to capital $r$) contribute to a rise in real income $I = \tfrac{wL + rK}{P}$. If demand is elastic and positively responds to increases in income ($\tfrac{\partial Q}{\partial (wL+rK)} \geq 0$) and decreases in prices ($\tfrac{\partial Q}{\partial p} \leq 0$), we can expect to observe an expansion of aggregate output $dQ \geq 0$.
However, it should be noted that the real income gains are not necessarily equally distributed. This may have an impact on the demand reaction as the propensity to consume is heterogeneous across income groups and products.
The expansion of output driven by a technology-induced real income effect may lead to a higher demand for labor.
As the real income effect on labor is very indirect, we interpret a study as empirically supporting the real income effect if it provides empirical evidence for at least one of the underlying mechanisms, namely: an increase in (1) productivity, (2) lower prices, (3) higher levels of income and wages, and (4) rising levels of output and a positive relationship between labor and output.
We interpret studies that report insights on at least one of these mechanisms as supportive for the real income effect while being aware that support for one of these mechanisms does not necessarily imply that the full chain of arguments holds. For example, productivity gains may not be forwarded to consumers in terms of lower prices if distorted competition prevents this, and rising levels of income do not necessarily imply a higher demand for consumption.
\section{Methods}
\label{sec:methods}
The aim of this review is to answer the question: \emph{What is the net employment effect of technological change since the 1980s in developed countries?} To answer this question, we reviewed the empirical evidence of studies published between 1988 - April 2021.
\subsection{Search strategy}
We closely followed the PRISMA 2020 guidelines to ensure the quality of the systematic search process \citep{page2021litreview}. A scoping review was used to identify relevant search terms in widely cited studies.
Subsequently, a computerized search was performed using the search terms that appeared either in the title, abstract, or list of keywords of studies, namely, i.e. `technolog*' combined with `labo\$r' or `employment'. The search was conducted in the Web of Science (WoS) Core Collection database.\footnote{A detailed description of the strategy and exact search strings used can be found in Appendix~\ref{appendix:searchstrategy}.}
The initial search resulted in 8,748 studies published between 1 January 1988 and 21 April 2021.
Given the large search outcome, six independent researchers were initially involved in screening the relevant records to ensure the timely and unbiased completion of the process. Based on the title and abstract, studies were considered relevant if the independent variable is related to technology and the dependent variable is related to (un)employment.
As a next step, the remaining 252 studies were assigned to three researchers, the authors of this paper, who independently assessed an equal amount of the remaining records based on the following inclusion criteria: (1) the study examines the effect of technological change on employment;
(2) the study makes a significant empirical contribution, i.e. excluding purely theoretical studies; and (3) the study covers at least one developed country and the period after the 1980s
Overall, the systematic search led to the inclusion of 127 studies. These studies were coded along a scheme that was iteratively refined throughout the process. We recorded the countries studied, the period covered, the outcome variable(s), and extracted descriptive information about the empirical operationalization of technological change, the measurement of the employment effects, the insights that can be gained about the three effects (replacement, reinstatement, and real income), and, if applicable, information about the net employment effect of technologies. We also extracted information about the level of analysis (e.g. country, region, firm, employee, occupation), data sources, sample size, methodology, potential heterogeneity, if applicable, and bibliometric information about the author, publication year and outlet.
\subsection{Coding scheme}
Throughout the process of screening the selected studies, we developed a simplified coding scheme for the effect considered, technology type, method, and level of analysis to cluster studies relying on similar measures and approaches. The coding scheme consists of categories (e.g. different technology types) and is based on common thematic patterns across the papers.
First, we classified a study as ``support'' if it provides results that offer data-based support for the existence of the effect of interest examined, i.e. replacement, reinstatement, and real income. Some studies find that the effects vary depending on the type of technology or the (sub)sample analyzed, for example distinguishing between industries, demographic groups, occupations, and types of labor. In this case we assigned the study to the category ``depends''. Studies that report negligible effects in terms of the magnitude or effects of low statistical significance were classified as ``weak''. Finally, studies were labeled as ``no support'' if they find that the effect of interest is insignificant or opposite to what was hypothesized.
Note that a study can investigate more than one type of effect, but the assigned labels, i.e. support, depends, weak or no support, are mutually exclusive within each effect. In the same spirit, we classify studies as ``positive'', ``negative'', ``depends'', and ``weak'' based on the extent to which they provide evidence on net employment effects.
The studies in our sample differ by the technologies examined and the empirical indicators used to measure them. We identified five broader classes of measures of technological change: ICT, robots, innovation, TFP-style, and a residual category other/indirect. The residual category accounts for technology types that are used by a small number of studies and are rather too heterogeneous to form separate technology groups.
The increased availability of ICT was one major technology trend of the post-1980s.
Empirical studies investigating its economic impact rely on different measures of ICT-diffusion such as
ICT investments or capital, as for example found in \citet{EUKLEMS2019data} and other comparable databases which are mostly publicly available. Other studies rely on survey data of computer use and ICT investment at the firm-level or occupation-level.
Another technology that raised much attention in economic research is the impact of robots.
In most studies, robot-diffusion is measured using data for industrial robots from the IFR \citep{IFR2020data} which, to our knowledge, is the only source consistently covering a large set of countries, time-periods and all relevant industries where industrial robots are adopted. A few other studies use country-specific data from the Japanese and Danish national associations for robot manufacturers on robot sales \citep{dekle2020robot, graetz2018robots}, or trade in robots \citep{blanas2019who}. Two studies rely on survey data about industry level robot use \citep{camina2020autom, edler1994tech}, and one study relies on survey data on self-reported job losses in response to the adoption of robots at the workplace \citep{dodel2020perceptions}. Robots received much attention because they can be interpreted as a pure automation technology, substituting human workers performing manual tasks.
Next to these direct technology-diffusion measures, we group studies relying on innovation as indicator of technological change. These studies often use data from the Community Innovation Survey (CIS) or comparable surveys for non-European countries. These surveys are regularly conducted by statistical offices to assess the innovativeness of firms and regions. Typically, the surveys allow to distinguish between process and product innovation and, in some cases, organizational innovation.
Process innovation is measured by survey questions asking firms to report whether they implemented a new improved production method, and product innovation is evaluated based on a question asking firms whether they recently introduced a new product.
Organizational innovation is measured through a question asking for the implementation of new organizational methods in business practices, workplace organization or external relations \citep{arundel2013history}.
As a fourth technology cluster, we code studies as TFP-style if they rely on measures of technological change that are inferred from production functions and input uses. These measures are for example estimates of productivity (TFP or labor productivity) or changes in substitution elasticities.
Some studies use readily available estimates provided by statistical offices or other relevant external data sources. Others explicitly estimate productivity or substitution elasticities on the basis of otherwise unexplained variation in the production function and sometimes also distinguish between different forms of input biases of technological change.
As a remainder, we use the residual category other/indirect to label studies. This category comprises measures that are only used by a small number of papers. Some of these studies rely on indirect indicators measuring the risk of automation and exposure to technological change in certain industries, demographic groups, and regions. These studies often rely on metrics developed by \citet{frey2017future} which provide estimates of the probability that tasks in certain occupations performed by human labor can be automated.
Another set of papers that we code as Other/indirect use trends in capital, high-tech equipment, and R\&D investments, and patents as proxies for technological change.
This category also includes a variety of other measures such as shifts in the age of the capital stock, assuming that more recent capital indicates more dynamic patterns of technological change, or changes in the occupational efficiency.
To cluster the papers by level of aggregation, we introduced four levels of analysis: Macro, meso, micro, and regional. Macro-level studies rely on country-level data and variation over time and/or countries. Meso-level studies usually include industry or sector-level data, while micro-level studies are at the more granular level ---mostly at the firm or individual employee-level. Finally, Regional level studies use variation across regional dimensions (e.g. commuting zones, NUTS regions, counties, etc.).
To classify papers by their methodology, we used three categories distinguishing between descriptive and conceptual analyses that link macro-level stylized facts to empirically reported technology trends (labeled ``descriptive''); regressions and similar forms of inferential statistics (labeled ``regression''); and studies that rely on simulation or calibration exercises (labeled ``simulation''). Simulation studies were only included if a substantial part of the study includes a significant amount of empirical data analysis, e.g. to motivate, calibrate, and/or estimate a simulation model.
Some papers are classified into multiple categories, e.g. when more than one type of technology is studied, or when the authors performed the analyses on different levels of aggregation.
\begin{table}[H]
\centering
\caption{Studies across various categories}
\label{tab:categories}
\begin{tabu}{p{0.3\textwidth}*4{c}} \toprule
\csname @@input\endcsname{inputs/stats_byeffect_short}
\end{tabu}
\begin{tabu}{l*{6}{c}} \toprule
\csname @@input\endcsname{inputs/stats_bytech_short}
\bottomrule
\end{tabu}
\begin{tabu}{p{0.35\textwidth}*{5}{>{\centering}p{0.095\textwidth}}}
\csname @@input\endcsname{inputs/stats_bylevel_short}
\bottomrule
\end{tabu}
\begin{tabu}{p{0.3\textwidth}*{3}{>{\centering}p{0.19\textwidth}}}
\csname @@input\endcsname{inputs/stats_effect_bymethod_short}
\bottomrule
\end{tabu}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Panel A presents the share and number (\#) of studies examining the replacement, real income, and reinstatement effect, respectively. The last column presents the share and \# of studies exploring at least two of the effects in the previous columns.
Panel B presents the share and number (\#) of studies examining ICT, robots, innovation, TFP-style and Other technology groups, respectively. `Other' refers to technologies measured indirectly through prices, automation risks, etc. The last column presents the share and \# of studies examining at least two of the technology groups in the previous columns. Panel C presents the share and number (\#) of studies where the analysis is conducted at the macro (e.g. country, over-time), meso (e.g. sectors, industries), micro (e.g. firms, individuals), and regional (e.g. regions, states, cities) level of data aggregation, respectively. The last column presents the share and \# of studies where the analysis is conducted in at least two of the previous levels of analysis. Panel D presents the share and number (\#) of studies by the primary type of empirical methodology used in each study to identify the effect(s) of interest. `Descriptive' refers to studies using descriptive statistics and conceptual analyses that link macro-level stylized facts to empirically reported technology-trends at the micro-level. `Regression' refers to any regression-based analysis or other quantitative inferential methods with empirical foundation. `Simulation' captures simulation methods, e.g. DSGE. Note that there is no overlap in methods reported, i.e. more than one primary method used in each study, and thus the shares across columns add up to one, up to rounding. The total \# of studies is 127.
\end{tablenotes}
\end{table}
\section{Results}
\label{sec:results}
We begin with an overview of the basic descriptive statistics of the full sample of 127 studies. Subsequently, we describe and contextualize three subsets of studies that report empirical results for each of the three effects: Replacement, reinstatement, and real income (see Section \ref{sec:conceptual_framework}).
\subsection{Overview}
Panel A in Table \ref{tab:categories} provides an overview of the studies by effect covered. The vast majority of studies (81\%) is related to the replacement effect. Another 62\% report results about the reinstatement and 26\% about the real income effect. More than half of the studies (59\%) simultaneously report results for at least two of the three effects and only 16 studies (13\%) for all three effects.
Panel B in Table \ref{tab:categories}, shows that roughly one third (35\%) of the studies analyzed the impact of ICT on employment. 13\% Studied the impact of robots and another 13\% examined the impact of innovation. About 14\% rely on TFP-style measures and 30\% fall into the residual category other/indirect. The distribution of Macro, Meso and micro-level studies is roughly balanced with 35\%, 32\%, and 30\%, respectively, while only 12\% have a regional focus. More than 80\% rely on regressions, 14\% on descriptives, and only 6\% used simulations. This result holds irrespective of the effect explored (see Appendix Table~\ref{tab:effect_bymethod}).
\begin{figure}[H]
\caption{Share of studies by type of result reported for each effect examined}
\label{fig:share_results_reported}
\includegraphics[width=1\textwidth]{inputs/Figure_Effects}
\scriptsize
Source: Author's calculations based on 127 studies collected from systematic literature review.\\
Notes: Panels (a)-(d) present the share of studies by each type of result reported for the replacement, reinstatement, real income and net employment effect, respectively. Specifically, in panels (a)-(c) a study is classified as `support' if it finds a significant empirical effect that supports the effect of interest examined, i.e. replacement, reinstatement, real income, respectively. `Depends' refers to the set of studies that find varying effects depending on the type of technology or the (sub)sample analyzed (e.g. type of sector or labor). Studies reporting effects that are negligible in terms of the magnitude were classified as `weak'. Studies were labeled as `no support' if they investigate the effect of interest, but documented insignificant or opposite effects. Similarly, in panel (d), `positive' and `negative' refer to studies which find for net employment a significant empirical effect that is positive and negative, respectively, while `depends' and `weak' are defined similarly to those above. A study can investigate more than one type of effect, but the assigned groups of results (i.e. support, depends, weak or no support and positive, depends, weak or negative) are mutually exclusive within each effect explored. In each panel, `n' is the number of studies examining the relevant effect.
\end{figure}
In Figure \ref{fig:share_results_reported}, we provide an overview of whether or not the empirical findings of the studies support an effect. Roughly two-thirds (67\%) of the studies that report results on the replacement effect find support for this effect, while only 16\% provide no support.
Panel (b) shows that among those studies that report results on the reinstatement effect, 81\% support the idea that technological change also creates new jobs either within the same firm, the same industry or elsewhere in the economy. Only a small fraction of the studies (2.5\%) find no support.
The share of supporting studies (70\%) is similarly high for the real income effect (Panel (c)). It should be highlighted that the number (n) of studies in this subsample is much smaller compared to the rest of the effects, i.e. 33 versus 103 and 79 studies.\footnote{Note that the real income effect interacts with some studies that rely on positive productivity shocks (TFP-style). These studies assume that productivity increases are an indicator of technological change. Our search for support of the real income effect asks whether this assumption holds true and we aim to understand the channels through which productivity interacts with demand as potential source of labor reinstatement.}
Panel (d) in Figure \ref{fig:share_results_reported} summarizes the results from papers that provide empirical evidence on the net effect. In total, 29\% of studies document a net positive effect, 18\% a negative effect, and 18\% report ambiguous or inconclusive results. The relative majority of studies (35\%) shows that the net employment effect depends.
\FloatBarrier
\subsection{Replacement effect}
\label{sec:results_replacement}
\subsubsection{Overview, methods and technical issues}
Does technology replace human labor? The majority of studies exploring the Replacement effect suggests that it does, but we also find a relatively small share of studies that do not support or suggest ambiguous effects (see Table \ref{tab:replacement}). Here, we systematize the empirical evidence by result (support, no support, depends, weak) and other characteristics of each study.
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on replacement}
\label{tab:replacement}
\begin{tabular}{l*{4}{c}} \toprule
& (1) & (2) & (3) & (4) \\
\csname @@input\endcsname{inputs/stats_replace_byfind}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Columns (1)-(4) present the share and number (\#) of studies with empirical results that support, depend (on various characteristics, e.g. technology type, analysis level, etc.), are weak, and do not support the presence of a replacement effect, respectively. The total \# of studies in the sample is 127 of which 103 examine the replacement effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
Many studies build on the neoclassical framework of skill- or task-biased technological change as a basis for the empirical analysis.
The majority of studies (83\%) use regressions, fifteen studies (15\%) rely on descriptive analyses, and the remaining three studies (3\%) use simulation methods (see Appendix Table \ref{tab:effect_bymethod}).
In column (1) of Table \ref{tab:tech_byfinding_replacement}, we present the fraction of studies by technology type (ICT, robots, innovation, TFP-style, other/indirect). The other columns display for each technology whether or not empirical support for the replacement effect is found.
Most studies (36\%) examine the impact of ICT, followed by TFP-style (17\%), robots (15\%), and innovation (12\%). 29\% use other measures of technology that fall in our residual category other/indirect.
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on replacement effect for each technology group considered}
\label{tab:tech_byfinding_replacement}
\begin{tabular}{lc|*{4}{c}} \toprule
& (1) & (2) & (3) & (4) & (5) \\
& \multicolumn{1}{c}{}& \multicolumn{4}{c}{by finding} \\
\cmidrule{3-6}
& \multicolumn{1}{c}{total}& \multicolumn{1}{c}{support}&\multicolumn{1}{c}{depends}&\multicolumn{1}{c}{weak}&\multicolumn{1}{c}{no support}\\ \midrule
\csname @@input\endcsname{inputs/stats_tech_byfind_replace}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes:
Column (1) presents the share and number (\#) of studies by the technology group considered in each row panel relative to the total number of studies examining the replacement effect. The row-sum of shares in column (1) does not add up to one since there are studies considering more than one technology group. Columns (2)-(4) present the share and \# of studies by the set of findings reported on the replacement effect for each type of technology considered in each row panel. For findings, a study is classified as `support' if it finds a significant empirical effect that supports the replacement effect examined. `Depends' refers to the set of studies that find varying effects depending on the type of technology or the (sub)sample analyzed (e.g. type of sector or labor). Studies reporting effects that are negligible in terms of the magnitude were classified as `weak'. Studies were labeled as `no support' if they investigate the effect of interest, but documented insignificant or opposite effects. The technology types reported include ICT, robots, innovation, TFP-style and Other types of technologies (e.g. indirectly measured through prices, automation risks), respectively. Note that there is no overlap in findings reported, i.e. more than one primary finding in each study, and thus the shares across columns add up to one, up to rounding. The total \# of studies in the sample is 127 of which 103 examine the replacement effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
Column (2) in Table \ref{tab:tech_byfinding_replacement} shows high support rates across all technologies except for innovation. The strongest support comes from studies that use robots (87\%), TFP-style (76\%), other/indirect (73\%), and ICT (62\%).
The findings from ICT studies are the most controversial showing simultaneously high numbers of supporting (62\%) and non-supporting studies (24\%), but only a few studies that report conditional or weak effects.
Innovation studies show the weakest support rate for the replacement effect, but seem consensual about the conditionality of the effect. Only one-quarter (25\%) supports this effect while 17\% does not find any empirical support. The majority (58\%) reports that the effect is conditional on the type of innovation (e.g. product or process innovation) and other relevant dimensions, such as the characteristics of the employees and firms.
\FloatBarrier
\subsubsection{Studies supporting the replacement effect}
The highest absolute number of papers (n=23) supporting the replacement effect studies the impact of ICT \citep{autor2002computers,autor2003skill,autor2013polar, baddeley2008structural,balsmeier2019time,dengler2018tech, diaz2002inno, eden2018, fonseca2018polar, fossen2021digital, morrisonpaul2001tech, wolff2009computer, addison2000tech, autor2015there, berman1994changes, luker1997employment, autor2015trade, blanas2019who, goaied2019, kristal2013capitalist, kaiser2001, jerbashian2019autom, morrison1997}.
These papers mostly build on theories of skill- or task-biased technological change and suggest that ICT replaces human labor in low-skill jobs and occupations or (regions with) industries that are intensive in routine tasks. Most studies are at the meso-level of analysis, followed by studies at the micro-level \citep{balsmeier2019time, dengler2018tech, fonseca2018polar, fossen2021digital, addison2000tech, kaiser2001} and macro-level \citep{baddeley2008structural,eden2018,wolff2009computer,autor2015there,goaied2019}. Only \citet{autor2013polar, autor2015trade} found ICT-induced replacement effects at the regional level.
The highest relative support (87\%) for the replacement effect comes from studies on robots (n=13) \citep{borjas2019, camina2020autom, devries2020rise, edler1994tech, faber2020robot, jung2020tech, compagnucci2019robotization, acemoglu2019automation, acemoglu2020robots, dodel2020perceptions, blanas2019who, graetz2018robots, labaj2020automation}.
Amongst these, \citet{compagnucci2019robotization} reported a positive effect on wages which is outweighed by employment losses, yielding a net negative effect on the wage bill. \citet{faber2020robot, acemoglu2020robots} studied the impact of robots on the regional (commuting zone) employment-population ratio.
\citet{acemoglu2020robots} used a long difference regression approach and report a negative relationship between robot exposure and employment where the effects were strongest in routine manual occupations and blue collar jobs. \citet{borjas2019} rely on similar data-sets to examine the impact of robots and immigrants on hourly wages and employment at the state-industry-level. They observed a negative association between robots and employment and wages.
\citet{faber2020robot} examined the impact of robot usage in an offshoring country. He found that robot adoption in offshoring industries in the US had a negative employment effect on regions in Mexico that heavily rely on exports to the US. This result supports the idea of cross-regional spillovers. When regions' comparative advantage is based on low labor costs in the production of tradable goods, they might lose this advantage as the tasks performed by cheap labor can now be performed by machinery.
One macro-level robot study examined the impact of robots, measured by the number of robots per thousand workers, on changes in the country-level wage bill \citep{labaj2020automation}. The authors found evidence for the existence of the replacement effect, but simultaneously reported that negative employment effects are overcompensated by the reinstatement of new labor.
With only three studies, the empirical support for the replacement effect based on innovation studies is weak \citep{vivarelli1996inno,cirillo2017tech,dachs2017inno}.
\citet{vivarelli1996inno} studied the impact of product and process innovation in Italian manufacturing firms during the 1980-1990s. They found support for labor displacement driven by the dominant role of process innovation.
\citet{cirillo2017tech} and \citet{dachs2017inno} made the same observation in different European countries, but also highlighted the presence of heterogeneity in their findings with stronger effects in high-tech industries and in Northern Europe.
Thirteen TFP-style studies support the Replacement effect empirically, mostly reporting shifts in labor demand, particularly across industries and across different types of labor e.g. from production to non-production labor \citep{angelini2020wage, ergul2020tech, gregory2001jobs, ho2008invest, baltagi2005skill, bloch2010technological, chen2014total, autor2018lshare, bessen2020autom, graham2000, whelan2009shocks, huh2019, kim2020lshare}.
\citet{angelini2020wage, ho2008invest, baltagi2005skill, gregory2001jobs} found that technological change is skill-biased by showing that lower skilled labor tends to be replaced by high-skilled labor.
\citet{bessen2020autom, chen2014total, bloch2010technological} evaluated the factor bias of technological change and documented patterns of labor-saving and capital using technologies over the past decades with considerable heterogeneity across industries and countries. \citet{autor2018lshare} documented that industry-level TFP-growth is associated with a decrease in wages, hours worked, the wage bill, and the labor share in major industrial countries since the 1970s.
\citet{whelan2009shocks, huh2019} studied cyclical macroeconomic fluctuations showing positive productivity shocks to be negatively related to hours worked. \citet{ergul2020tech, kim2020lshare} suggested that technology-induced shocks are associated with decreases in the labor share of income, albeit potentially being only transitory. \citet{graham2000} found that some of the industry-region employment losses can be attributed to technological change.
The studies using other/indirect technology measures (n=22) can be roughly grouped into three categories. A first bundle of studies \citep{peng2018it, gardberg2020digit, blien2021occup, grigoli2020automation, arntz2017} exploited variation in regional, industrial and/or occupational susceptibility to automation and provide evidence that automation risk indices help to explain employment losses and the longer unemployment spells for workers in certain jobs.
A second category of other/indirect technology measures includes capital and high-tech equipment investments as proxies for technological change \citep{morrison1997, wemy2019lshare, ho2008invest}. Also \citet{gera2001tech} infer the pace of technological change from shifts in the age of the capital stock, as more recent capital indicates more dynamic patterns of technology investment.
\citet{kim2020lshare, gera2001tech, morrisonpaul2001tech, vainiomaki1999tech} approximated technological change via R\&D spending and \citet{gera2001tech, feldmann2013} used patents.
Furthermore, \citet{gera2001tech} provide support for skill-biased technological change, i.e. more negative consequences in terms of employment and income for low-skill labor.
The skill bias is also supported by the descriptive analyses of \citet{oesch2010, buyst2018polar} made a detailed description of occupational shifts across industries attributing them to technology trends. Similarly, the detailed historical analysis from \citet{padalino1997employment} documents changes in the employment-GDP elasticity in G7 countries for the period 1964-94 and found - for the late 20th century - a negative correlation between GDP growth and employment in manufacturing, but not for the whole economy.
The third category of studies focuses on skill-biased technological change. They are all at the macro-level except for one regional level study \citep{reshef2013tech, padalino1997employment, manning2004tech, hoskins2000, hoskins2002, madariaga2018employment, reijnders2018tech}. The measures of technology employed are heterogeneous and vary from the use of the Leontief inverse matrix as a proxy for changes in production techniques to changes in occupational efficiency. All studies provide suggestive evidence that technological change decreases the (relative) demand for unskilled workers or routine jobs.
\subsubsection{Studies not supporting the replacement effect}
The number of studies that provide evidence against any significant impact of technology on employment is somewhat smaller (n=17). Nine of them are on ICT, two on robots, two on innovation, one on TFP-style study, and three studies use other/indirect technology measures.
\citet{fung2006} proxied ICT by expenses on IT and computer data processing and examined whether ICT had a labor-saving impact on firms in the US banking industry between 1992-2002. The author rejected the presence of labor-saving effects for the technologies considered, observing that the more technology-intensive firms increased their employment. Similarly, \citet{auberttarby2018} analyzed whether digitilization in the French newspaper and magazine press industry creates or destroys jobs.
Overall, they found digitalization to be associated with higher wages and a reduced probability to be laid off.
\citet{pantea2017ict, atasoy2016, gaggl2017computer} made similar observations for EU countries and Turkey looking at different firm-level ICT usage indicators. They estimated either insignificant or positive effects of these technologies on employment in SMEs. \citet{gaggl2017computer} showed that the introduction of these technologies is associated with changes in the composition of tasks performed by workers.
In line with that, \citet{biagi2017ict, ivus2015broadband} did not find any significant effect of ICT on employment in European industries and Canadian regions, respectively.
\citet{behaghel2016it} did not observe that ICT-diffusion increases the probability of dismissals in French cities. If anything, they provided only weak support of increased job instability for high-school dropouts.
\citet{borland2017robot} provided a descriptive historical analysis documenting a decrease in routine labor which does not deviate from historical patterns. They could not identify any noteworthy relationship between ICT-diffusion and employment changes.
\citet{dekle2020robot} and \citet{fu2021} evaluated the impact of robots on employment (n=2).
\citet{fu2021} studied a sample of seventy-four countries, including the EU-27 and twenty-nine developing countries. They found insignificant effects in developing countries and positive employment effects accompanied with productivity increases in developed countries. The authors also reported heterogeneity by gender, but they did not find evidence for robot-induced replacement. \citet{dekle2020robot} analyzed the impact of robot-diffusion during 1979-2012 in twenty Japanese industries, but could not identify any significant negative effects on employment.
Among innovation studies (n=2), \citet{calvino2018innovation} concluded in a literature review that the relationship between product innovation and employment is rather positive if statistically significant at all. The impact of process innovation remains controversial and the authors emphasized that the existing evidence is insufficient to support the existence of a replacement effect. \citet{fung2006} studied - in addition to ICT - the impact of in-house process innovation in the banking sector. But again, they could not find any support that these innovations reduce employment.
Three studies considered other/indirect technology measures. Using regressions, \citet{scholl2020autom} did not find any significant impact of automation risks on pay or employment.
\citet{sargent2000unempl} assumed that changes in economic regularities during the 20th century, such as the relationship between unemployment rates and vacancies, are an indication of technological change.
The author found descriptive evidence that such shifts occurred and linked these patterns to employment shifts across industries, occupations and educational groups, but he could not find any evidence of increased unemployment.
\citet{raval2019micro} explored plant-level information to examine whether capital-labor substitution possibilities have changed between 1987-2007. They found that capital-labor elasticities have been very persistent over the time period covered.
\subsubsection{Studies with ambiguous findings and indirect evidence}
Ambiguous results are reported in twenty studies coming in the following decreasing order from innovation (n=7), ICT (n=5), other/indirect (n=5), and TFP-style (n=3).
No robot study reports ambiguous results.
Among the studies with ambiguous results, some of these---classified as ``weak''---provide only indirect evidence that may be consistent with the replacement effect. The remaining studies---classified as ``depends''---provide contradicting evidence when using different specifications. In both cases and irrespective of the technology, all of them remain inconclusive on whether or not technology replaces labor.
Among the ICT studies, \citet{gallipoli2018it} provided indirect support for the replacement of non-IT by IT-intensive occupations with a higher wage premium.
\citet{breemersch2019polar} showed ICT adoption to be associated with increased polarization towards high pay jobs within manufacturing only, but with minimal contributions between manufacturing and non-manufacturing industries.
\citet{maurin2004skill} found that the share of skilled labor increased in response to ICT adoption which may be accompanied with the simultaneous replacement of unskilled labor.
\citet{autor1998computers} observed rapid skill upgrading processes since 1970 and provided suggestive evidence that these processes evolve faster in computer intense industries.
These studies point to a declining employment share of certain types of jobs, i.e. mostly routine task intensive, low-skill, and low-wage jobs. However, it is not necessarily true that changes in employment shares are accompanied by the replacement or ``phasing out'' of certain occupations.
\citet{omahony2021lshare} studied the impact of ICT and innovation-related capital investments, but did not find any conclusive results on whether the labor share is positively or negatively affected. Intangible investments related to innovation tend to have a positive impact, while investments related to firm organization tend to show the opposite pattern. Low- and middle-skill workers appear to be more negatively affected.
Three studies of those relying on TFP-style measures report ambiguous results \citep{boyle2002trade, dixon2020decline, fort2018}.
\citet{fort2018} made a detailed descriptive analysis showing a negative relationship between TFP-growth and employment in some but not all manufacturing industries. They reported that this does not realize through the shut-down of existing plants but through lower shares of labor input in plants that are new entrants.
\citet{boyle2002trade, dixon2020decline} did not directly evaluate the impact of TFP on labor, but showed that the decline in the labor share can be partly attributed to labor-saving technological change.
Seven studies analyzed the impact of innovation \citep{bogliacino2010inno, falk2015inno, cozzarin2016inno, evangelista2012inno, pellegrino2019rnd, vanreenen1997inno, kwon2015inno}.
All of them are firm-level analyses - except for the industry-level study by \citet{bogliacino2010inno} - and explored whether there are any effects of process/product innovation on labor demand. \citet{falk2015inno, evangelista2012inno, kwon2015inno} also studied the impact of organizational innovation. None of these studies could identify any significant impact of product and organizational innovation on labor. For process innovation, \citet{pellegrino2019rnd, vanreenen1997inno} either found only weakly significant negative effects or showed that the result depends on the size of firms.
\citet{falk2015inno} showed a negative association between process innovation and labor, but only for a subset of industries.
\citet{cozzarin2016inno} found ambiguous results depending on the type of innovation and technology considered: For example, he found no effect of advanced manufacturing technologies, a negative effect of product innovation, and a positive effect of process innovation.
\citet{evangelista2012inno} found weak evidence of a labor-saving impact of organizational product innovation in manufacturing but not in services. Next to the labor-saving effect, they also observed an innovation-induced increase in sales which may offset the negative effect on employment.
\citet{kwon2015inno} reported a negative impact of process but not product innovation.
Among those studies using other/indirect technology measures (n=5), three used investment in high-tech capital or R\&D as technology proxies \citep{breemersch2019polar, flug2000, idris2021tech} and \citet{omahony2021lshare} used knowledge-based assets. \citet{idris2021tech} could not identify any impact of high-tech on employment and \citet{flug2000, breemersch2019polar} documented technology-induced wage and skill polarization patterns in the labor market. However, \citet{breemersch2019polar} emphasized that technology plays a rather minor role in explaining labor market polarization.
\citet{omahony2021lshare} found heterogeneous results depending on the type of the technology considered: R\&D-based knowledge investments seem to mitigate the ICT-driven declining trend in the labor share, while innovation-related intangible investments are related to a rising labor share, opposite to the effect found for organization-related investments.
Finally, \citet{green2015} provide a descriptive analysis of technology trends and explored polarization in the Canadian labor market since the 1970s. They cannot confirm any hypothesis about skill-biased technological change. Rather, the drivers of the polarization differ from those in the US and other countries. The authors documented some occupations to be shrinking but argued this to be driven by the resource boom in Canada, rather than by technological change.
\FloatBarrier
\subsection{Reinstatement effect}
\label{sec:results_reinstatement}
\subsubsection{Overview, methods and technical issues}
Does the introduction of new technologies create new jobs?
In total, seventy-nine studies in our sample offer empirical insights on the reinstatement effect.
Among those studies that report empirical evidence on the reinstatement effect, 81\% support the existence of this effect, while 17\% report ambiguous findings and only two studies (3\%) document insignificant effects (see Table \ref{tab:reinstatement}).
Almost all of the reinstatement studies (82\%) simultaneously report findings on the replacement effect and seventeen studies (22\%) on the real income effect.
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on reinstatement}
\label{tab:reinstatement}
\begin{tabular}{l*{4}{c}} \toprule
& (1) & (2) & (3) & (4) \\
\csname @@input\endcsname{inputs/stats_reinst_byfind}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Columns (1)-(4) present the share and number (\#) of studies with empirical results that support, depend (on various characteristics, e.g. technology type, analysis level, etc.), are weak, and do not support the presence of a reinstatement effect, respectively. The total \# of studies in the sample is 127 of which 79 examine the reinstatement effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
In column (1) of Table \ref{tab:tech_byfinding_reinstatement}, we show that the largest fraction of studies looking at reinstatement focus on the impact of ICT (38\%) and other/indirect technology measures (28\%), followed by Innovation (18\%), TFP-style (14\%), and Robots (13\%).
Similarly to the replacement effect, the highest ambiguity (weak and depends) is observed for innovation (43\%) and ICT (20\%). However, even in these cases the support rates remain high.
The vast majority of studies (n=63) relies on regressions, thirteen studies rely on descriptive analyses, and only three are based on simulations (see Appendix Table \ref{tab:effect_bymethod}).
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on reinstatement effect for each technology group considered}
\label{tab:tech_byfinding_reinstatement}
\begin{tabular}{lc|*{4}{c}} \toprule
& (1) & (2) & (3) & (4) & (5) \\
& \multicolumn{1}{c}{}& \multicolumn{4}{c}{by finding} \\
\cmidrule{3-6}
& \multicolumn{1}{c}{total}& \multicolumn{1}{c}{support}&\multicolumn{1}{c}{depends}&\multicolumn{1}{c}{weak}&\multicolumn{1}{c}{no support}\\ \midrule
\csname @@input\endcsname{inputs/stats_tech_byfind_reinst}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item
Notes: Column (1) presents the share and number (\#) of studies by the technology group considered in each row panel relative to the total number of studies examining the reinstatement effect. The row-sum of shares in column (1) does not add up to one since there are studies considering more than one technology group. Columns (2)-(4) present the share and \# of studies by the set of findings reported on the reinstatement effect for each type of technology considered in each row panel. For findings, a study is classified as `support' if it finds a significant empirical effect that supports the reinstatement effect examined. `Depends' refers to the set of studies that find varying effects depending on the type of technology or the (sub)sample analyzed (e.g. type of sector or labor). Studies reporting effects that are negligible in terms of the magnitude were classified as `weak'. Studies were labeled as `no support' if they investigate the effect of interest, but documented insignificant or opposite effects. The technology types reported include ICT, robots, innovation, TFP-style and other types of technologies (e.g. indirectly measured through prices, automation risks), respectively. Note that there is no overlap in findings reported, i.e. more than one primary finding in each study, and thus the shares across columns add up to one, up to rounding. The total \# of studies in the sample is 127 of which seventy-nine examine the reinstatement effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
\subsubsection{Studies supporting the reinstatement effect}
Among those studies that support the reinstatement effect, twenty-three look at the impact of ICT \citep{atasoy2013broadband,auberttarby2018,autor2003skill,autor2013polar,baddeley2008structural,balsmeier2019time,fossen2021digital,fung2006,morrisonpaul2001tech,behaghel2016it,autor2015there,berman1994changes,luker1997employment,autor2015trade,blanas2019who,gaggl2017computer,gallipoli2018it,ivus2015broadband,kristal2013capitalist,kaiser2001,jerbashian2019autom,maurin2004skill,morrison1997}.
At the city-level, \citet{behaghel2016it} studied the impact of ICT on skill upgrading and job-to-job transitions showing that ICT adoption is associated with an increased demand for skilled labor, but not with higher dismissal rates.
\citet{gaggl2017computer} and \citet{ivus2015broadband} argued that the impact of ICT on net employment depends on various dimensions. \citet{gaggl2017computer} found a positive relationship between ICT adoption and the demand for non-routine cognitive labor, but did not observe effects on job replacement. However, they found that the positive effect diminishes over time which may indicate that it is only a transitory phenomenon.
The detailed descriptive analysis of industrial job creation and destruction dynamics by \citet{borland2017robot} provides evidence that the positive employment effects in Canada are heterogeneous across regions: ICT-diffusion only exhibits significant positive interactions with labor in rural regions with stronger effects in ICT-intensive industries.
\citet{gallipoli2018it} listed numerous empirical stylized facts for the US between 1980-2013 using micro-level census data and detailed occupational employment statistics. They documented the emergence of new and well-paid IT occupations, mostly in services. They showed that productivity growth can be mostly attributed to services while the employment share in manufacturing declined. However, their findings do not allow to draw conclusions about the existence of simultaneous replacement or net employment effects.
\citet{maurin2004skill} explored employee-level data on different types of computer technologies used at work. They found a positive correlation between the diffusion of computers and the share of high-skilled labor.
The authors also showed that the impact of ICT on labor demand is conditional on the type of ICT, and on the occupations and their task content.
Only the study by \citet{atasoy2013broadband} reports findings exclusively for the reinstatement effect, but not on any of the other two effects. The authors studied the impact of broadband deployment on county level labor markets in the US between 1999-2007. They found a positive effect of broadband access on employment and wages. The positive wage effects are larger in counties with more skilled labor supporting theories of skill-biased technological change.
In an economic history essay, \citet{autor2015there} documented in detail how the demand for labor for certain occupations in services, such as managers, personal care, food services and others, continuously increased since the 1980s until the financial crisis when the patterns of growth slowed down. Furthermore, he illustrated patterns of skill polarization reflected in the highest growth rates in the lowest and highest skill percentile.
Similarly, \citet{luker1997employment} documented shifts from manufacturing to services. They showed that a net increase in high-tech industries can be mostly attributed to services. Generally, high-tech service employment appears to grow faster than employment in the rest of the economy.
Thirteen regression studies on ICT simultaneously support the replacement and reinstatement of labor \citep{autor2003skill,autor2013polar,baddeley2008structural,balsmeier2019time,fossen2021digital,morrisonpaul2001tech,berman1994changes,autor2015trade,blanas2019who,kristal2013capitalist,kaiser2001,jerbashian2019autom,morrison1997}.
The two Regional level studies by \citet{autor2015trade, autor2013polar} found support for routine-biased technological change and observed increasing employment in abstract and manual tasks which neutralizes the negative employment effects of routine-task replacement.
\citet{autor2013polar} also documented a rise in polarization, i.e. both a rise in low-skill service jobs and a differential wage growth across occupations.
The study by \citet{baddeley2008structural} is a macro-level analysis confirming that computerization, next to financialization, was associated with employment shifts from manufacturing to services in the UK between 1979-2005.
At the meso-level, \citet{autor2003skill, morrisonpaul2001tech, berman1994changes, blanas2019who, kristal2013capitalist, jerbashian2019autom, morrison1997} observed that ICT adoption is associated with the creation of new jobs. \citet{autor2003skill} found an increase in non-routine jobs. \citet{morrison1997} showed that investments in computers and R\&D are associated with skill polarization patterns, i.e. an increased demand for high- and low-skill workers.
\citet{berman1994changes} documented a rise in non-production labor which coincides with a skill upgrading process.
A similar skill bias is confirmed by \citet{blanas2019who} who showed that robots and ICT are associated with higher employment in high- and middle-skill jobs and in services.
Relatedly, \citet{jerbashian2019autom} showed a positive correlation between falling IT prices and employment in high-wage occupations.
\citet{balsmeier2019time} and \citet{kaiser2001} provided evidence at the firm-level and \citet{fossen2021digital} at the level of individual employees.
The two firm-level studies found that increased investment in IT is associated with the creation of high-skill jobs. \citet{fossen2021digital} found evidence for the creation of another class of jobs showing that digitalization significantly increases the probability that high-skill workers engage in entrepreneurial activity.
Nine reinstatement supporting studies analyzed the impact of Robots \citep{devries2020rise, dekle2020robot, edler1994tech, acemoglu2019automation, blanas2019who, graetz2018robots, leigh2019robots, gentili2020machines, labaj2020automation}.
Two thirds of them reported ambiguous effects on net employment and the remaining three \citep{leigh2019robots, gentili2020machines, dekle2020robot} found positive effects. None of these studies documents a clear negative impact of robots on net employment. This suggests that whenever evidence for robot-driven reinstatement is found, it tends to (over)compensate the replacement of labor.
\citet{edler1994tech} built on an empirical input-output model and found a higher demand for skilled labor in response to robot-diffusion.
\citet{gentili2020machines} performed a descriptive clustering analysis and attributed changes in the robot intensity to changes in employment measured by hours worked finding that industries with the highest robot intensity experienced the highest productivity and employment gains. The effects are clustered in high-tech industries which account for a small share of total employment.
\citet{labaj2020automation} offer a macro-level regression analysis exploring how variations in the economy wide wage bill and labor share are explained by robot-diffusion rates in the US and Europe. They decomposed aggregate changes in the task content of production into a reinstatement and replacement effect and found evidence for both. However, the authors emphasized that whether reinstatement or replacement dominates varies across countries (especially between the US and EU) and that this variation cannot be explained by robot-diffusion.
For six meso-level regression-based analyses \citep{devries2020rise, dekle2020robot, acemoglu2019automation, blanas2019who, graetz2018robots,leigh2019robots} we observed that they mostly report increases in high-skill and in service jobs. \citet{dekle2020robot} found similar effects for total employment, i.e. not only for high-skill and service jobs. \citet{leigh2019robots} also showed that in US manufacturing the impact of robots tends to be positive.
Continuing, eight studies relying on innovation offer support for the reinstatement effect.
\citet{calvino2018innovation} made a literature review on the impact of innovation, captured by different indicators (R\&D intensity, CIS, patents) on employment at the firm and sector-level. They documented that the employment effect is mostly positive, but also depends on the sector and the type of innovation (i.e. process or product). They confirmed a mostly positive effect of product innovation and a negative effect of process innovation.
\citet{tether1998employment} descriptively explored employment creation by small innovative firms at the micro-level, where innovative firms are defined as inventor award winning firms.
The authors showed that these firms have faster than average employment growth patterns, but it remains unclear to which extent this observation can be generalized beyond this specific setting.
Three studies are micro-level regressions using firm-level data \citep{fung2006, vivarelli1996inno, dachs2017inno}. \citet{fung2006} studied labor-saving product and process innovation in the banking sector and found: (a) a positive association between labor-saving technologies and higher firm-level employment and (b) positive spillovers from patented process innovations on labor demand in non-innovating banks. The latter effect suggests the presence of technology-induced employment externalities to non-technology-adopting sectors and how these channels contribute to the creation of labor in the economy.
Both \citet{vivarelli1996inno} and \citet{dachs2017inno} observed that product innovation is positively associated with sales and labor demand, but \citet{vivarelli1996inno} found this pattern only for a subset of sectors which are characterized by higher design and engineering intensities and higher percentages of product innovations.
\citet{cirillo2017tech} and \citet{xiang2005} performed industry-level analyses. \citet{cirillo2017tech} found a positive relationship between the share of firms introducing product innovations, industry-level demand, and employment growth.
Similarly, \citet{xiang2005} showed that the introduction of new goods is positively associated with the relative demand for skilled labor in the US manufacturing industry.
\citet{capello2013inno} confirmed the positive association of product innovation with employment using regional-level data for the EU. Applying spatial regressions, they showed a positive relationship between the share of firms that introduce product innovations and regional employment and wage growth in sub-national regions (NUTS2) that are characterized by a high share of blue collar workers. The authors also highlighted that the effect in regions with low shares of blue collar workers may be negative, but with no insights on whether these relationships have changed over time.
Ten studies rely on TFP-style measures. In a descriptive analysis, \citet{angelini2020wage} inferred technological change from shifts in the skill content of production and provided evidence for the reinstatement of service jobs reflected in employment shifts from manufacturing to services. \citet{fort2018} conducted a detailed descriptive analysis of the impact of labor productivity changes on employment shifts at the industry and plant-level emphasizing that displacement and net effects are industry-specific.
The other eight papers are regressions, largely at the industry or macro-level. They rely on different proxies of productivity changes, mostly captured by TFP or labor productivity, and study labor demand in absolute and relative terms. \citet{bessen2020autom} investigated the role of the elasticity of demand with respect to productivity for three industries since the 19\textsuperscript{th} century. The author highlighted that final demand was historically a key driver of the reinstatement of labor in the steel, textiles and automobile industry.
The studies by \citet{ho2008invest, baltagi2005skill, kim2020lshare, sala2018effects} showed that the reinstatement of labor may be biased, as reflected in an increasing demand for skilled and non-production labor.
The other three studies provide similar findings in different settings. \citet{boyle2002trade} found that capital accumulation and technological change are key drivers of employment growth. \citet{autor2018lshare} investigated TFP shocks in upstream industries and showed that this is positively associated with hours worked and employment in downstream industries. \citet{graham2000} showed that regional employment increases can be attributed to technology.
Twenty studies fall into the residual category other/indirect.
Among these, \citet{morrisonpaul2001tech, feldmann2013, vainiomaki1999tech, vanroy2018tech, yildirim2020, kim2020lshare, fagerberg1997tech} used R\&D expenditures or patents as a measure of technological change. \citet{morrisonpaul2001tech} and \citet{kim2020lshare} observed a positive relationship between R\&D expenditures and employment of high-skill labor. \citet{yildirim2020, fagerberg1997tech, vanroy2018tech} all documented a positive relationship between R\&D and labor, whereas \citet{vanroy2018tech} found that this only holds in high- but not in low-tech industries. \citet{vainiomaki1999tech} found that high-technology sectors, measured by R\&D intensity, have the highest job creation rates.
\citet{feldmann2013} provided only indirect support for the reinstatement effect: In the short term, increased innovation reflected in patent applications had a negative employment effect, but this effect diminished over time which indicates that employment was reinstated after a labor reducing technology shock.
Three studies used different indicators of specific types of capital investment as proxies for technological change.
\citet{ho2008invest} used a price index-based approach to capture quality improvements in equipment and found that these enhancements are associated with an increase in demand for non-production labor in US manufacturing.
\citet{morrison1997} considered investments in specific high-tech equipment and \citet{raval2019micro} used the evolution of the capital stock as a technology measure. They all reported a positive relationship between labor demand and the capital indicators, suggesting a complementarity between these factors. The results by \citet{morrison1997}, similar to \citet{ho2008invest}, suggest that this relation is particularly strong for non-production labor.
Two studies with support for the reinstatement effect used an indirect automation risk-based approach \citep{gardberg2020digit, oesch2010}. Both confirmed at the country-level, and \citet{gardberg2020digit} additionally at the individual employee-level, that displacement is less likely in occupational groups with low automation risk. In contrast, employment has even increased in these jobs.
\citet{padalino1997employment,hoskins2000,madariaga2018employment,reijnders2018tech,jiang2019subst}
are macro-level studies using indirect approaches to capture technological change, such as decomposition analyses and substitution elasticities.
\citet{padalino1997employment,jiang2019subst} found a positive relationship between technological change and employment, while the other two found it only for certain sectors and jobs. However, also \citet{padalino1997employment} observed that the employment-GDP relationship is opposite in manufacturing, suggesting the simultaneous existence of replacement in manufacturing and reinstatement in non-manufacturing industries in recent decades.
Finally, four non-regression analyses in the sample of studies relying on the residual technology category found support for the reinstatement effect.
\citet{reshef2013tech} relied on an empirically estimated simulation model calibrated on US data. The results confirm the hypothesis of skill-biased technological change in the US during 1963-2005 documenting an increasing demand for skilled labor at the expense of unskilled.
For Canada, \citet{green2015} could not confirm the skill bias hypothesis in their detailed descriptive analysis. The analyses by \citet{buyst2018polar} and \citet{oesch2010} also suggest a bias of technological change being associated with increases in high paid occupations. Both studies also provide evidence for a rise in the demand for labor in the lowest skill group, but \citet{oesch2010} showed that this effect is heterogeneous across countries.
\subsubsection{Studies not supporting the reinstatement effect}
The two non-supporting studies are the studies by \citet{acemoglu2020robots, goaied2019}. While they do not support the reinstatement effect, they offer support for the replacement and real income effect. Both studies rely on regression analyses and report an overall negative impact of technology on net employment. \citet{acemoglu2019automation} studied the impact of robots on the regional employment to population ratio in 1990-2007 in the US. The authors found that robot-diffusion is associated with an increase in unemployment driven by employment losses in manufacturing and routine manual, blue collar occupations. The effects are larger for men than women.
Additionaly, they reported positive productivity effects and increases in capital income.
\citet{goaied2019} studied the long term impact of ICT-diffusion in 167 countries grouped into five regions using the number of mobile phone and internet users as a technology indicator. Distinguishing between long and short term effects, the authors reported negative short and long term employment effects. They reported a positive relationship between GDP and employment, but cannot attribute this to technology.
\subsubsection{Studies with ambiguous findings and indirect evidence}
Fifteen studies report ambiguous or only indirect evidence for the reinstatement effect. Six of them rely on ICT, six on innovation, one on TFP-style measures and two more looked, next to ICT, into the role of R\&D intensity (coded as other/indirect).
The TFP-style study by \citet{dupaigne2009shock} estimated the macroeconomic employment effect of positive labor productivity shocks using a vector auto-regressive model. The authors found that the impact, whether positive or negative, is heterogeneous across countries and also depends on how technology shocks are measured.
Among the ICT studies, four of them rely on regression analyses \citep{autor1998computers, fonseca2018polar, breemersch2019polar, omahony2021lshare}, one study conducted a simulation exercise \citep{charalampidis2019} and another one relies on descriptive statistics \citep{borland2017robot}.
\citet{charalampidis2019} studied automation technology shocks (while not providing much detail about the measurement of these shocks) and analyzed their interaction with aggregate fluctuations of the labor share. While this study found that technology shocks explain a large share of the fluctuations, it remains inconclusive with regards to the longer term impact of these shocks. However, the author argued that labor reinstatement may be insufficient to offset job losses.
The analysis by \citet{borland2017robot} relies on a detailed descriptive analysis of those industries in Australia that reported the largest changes in employment during the past few decades. They linked these observations to ICT-diffusion curves, but failed to identify any clear impact of technology on labor. The aggregate demand for labor was roughly constant over the considered period, but they also observed an increase in occupations that are intensive in non-routine tasks.
Two of the four regression analyses in this sample investigated job polarization. \citet{fonseca2018polar} found support for technology as a driver of job polarization, which may be indicative for the reinstatement of certain types of jobs. \citet{breemersch2019polar} observed country-level employment growth and cross-industrial shifts, but failed to attribute this to technological change. \citet{autor1998computers} documented skill upgrading processes and higher wage premia, but did not show any clear impact on the demand for labor. The study by \citet{omahony2021lshare} investigated the relationship between ICT-diffusion and the labor share. They found a negative relationship between ICT capital investments and the labor share which may indicate lack of reinstatement. Interestingly, the authors also studied the impact of R\&D investments and observed the opposite effect.
The six innovation papers \citep{evangelista2012inno, pellegrino2019rnd, piva2005innovation, piva2004change, vanreenen1997inno, kwon2015inno} report ambiguous or only weak support for the reinstatement effect. All of them are firm-level regression analyses and confirm a positive effect of product innovation on employment, but a negative effect for process innovation. Further, all of these studies, except for \citet{piva2004change, piva2005innovation}, simultaneously report ambiguous findings for the replacement effect.
\FloatBarrier
\subsection{Real income effect}
\label{sec:results_realincome}
\subsubsection{Overview, methods and technical issues}
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on real income}
\label{tab:realincome}
\begin{tabular}{l*{4}{c}} \toprule
& (1) & (2) & (3) & (4) \\
\csname @@input\endcsname{inputs/stats_realinc_byfind}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Columns (1)-(4) present the share and number (\#) of studies with empirical results that support, depend (on various characteristics, e.g. technology type, analysis level, etc.), are weak, and do not support the presence of a real income effect, respectively. The total \# of studies in the sample is 127 of which 33 examine the real income effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
With thirty-three studies in total, the number of papers providing empirical insights on the real income effect is much smaller compared to the other two effects. Table \ref{tab:realincome} shows that among these, twenty-three papers found support, six report ambiguous (weak or depends) results, and four found no support.
The majority of studies investigate the impact of ICT (36\%), followed by robots (27\%), TFP-style (18\%) and other/indirect measures (15\%). Only two papers (6\%) study innovation (see column (1) from Table \ref{tab:tech_byfinding_realincome}).
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on real income effect for each technology group considered}
\label{tab:tech_byfinding_realincome}
\begin{tabular}{lc|*{4}{c}} \toprule
& (1) & (2) & (3) & (4) & (5) \\
& \multicolumn{1}{c}{}& \multicolumn{4}{c}{by finding} \\
\cmidrule{3-6}
& \multicolumn{1}{c}{total}& \multicolumn{1}{c}{support}&\multicolumn{1}{c}{depends}&\multicolumn{1}{c}{weak}&\multicolumn{1}{c}{no support}\\ \midrule
\csname @@input\endcsname{inputs/stats_tech_byfind_realinc}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Column (1) presents the share and number (\#) of studies by the technology group considered in each row panel relative to the total number of studies examining the real income effect. The row-sum of shares in column (1) does not add up to one since there are studies considering more than one technology group. Columns (2)-(4) present the share and \# of studies by the set of findings reported on the real income effect for each type of technology considered in each row panel. For findings, a study is classified as `support' if it finds a significant empirical effect that supports the real income effect examined. `Depends' refers to the set of studies that find varying effects depending on the type of technology or the (sub)sample analyzed (e.g. type of sector or labor). Studies reporting effects that are negligible in terms of the magnitude were classified as `weak'. Studies were labeled as `no support' if they investigate the effect of interest, but documented insignificant or opposite effects. The technology types reported include ICT, robots, innovation, TFP-style and other types of technologies (e.g. indirectly measured through prices, automation risks), respectively. Note that there is no overlap in findings reported, i.e. more than one primary finding in each study, and thus the shares across columns add up to one, up to rounding. The total \# of studies in the sample is 127 of which thirty-three examine the real income effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
The existence of the real income effect is empirically supported if studies found empirical support for at least one of the channels of technology-labor interactions (see Section \ref{sec:conceptual_framework}). The majority of papers (68\%) provides empirical evidence about the productivity channel, followed by income (29\%), and output (18\%) (see Table \ref{tab:realincome_byeffect}). Despite the high number of papers reporting results on the productivity effects of technological change, only 9\% are informative about the impact on prices, which is a relevant indicator to evaluate whether or not consumers benefit from the productivity gains of technological change.
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by type of real income effect examined}
\label{tab:realincome_byeffect}
\begin{tabular}{l*{5}{c}} \toprule
& (1) & (2) & (3) & (4) & (5) \\
\csname @@input\endcsname{inputs/stats_grealinc_byfind}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Columns (1)-(5) present the share and number (\#) of studies focusing on productivity, prices, income, output, and other real income effects, respectively. The total \# of studies in the sample is 127 of which 33 examine the real income effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
\subsubsection{Studies supporting the effect}
Among the real income supporting studies, ten investigated the impact of ICT \citep{autor2002computers, baddeley2008structural, chun2015it, oulton2002ict, strohmaier2016ict, vu2013ict, autor2015there, berman1994changes, blanas2019who, goaied2019}. Seven of them reported a positive impact of ICT on productivity, three found positive income effects, and two provide evidence of a positive relationship between output growth and labor demand.
The three studies by \citet{autor2002computers, oulton2002ict, autor2015there} rely on descriptive analyses using industry-level case studies, macro-level growth accounting methods, and macroeconomic history.
The other seven are regressions are macro- or meso-level, as well as one firm-level study by \citet{chun2015it}.
Another six studies examined the impact of robots using Meso or regional-level regressions \citep{dekle2020robot, jung2020tech, acemoglu2019automation, acemoglu2020robots, blanas2019who, graetz2018robots}. They all provide evidence of the productivity increasing effect of robots. \citet{blanas2019who} also found a positive effect of robots on the wage bill of high-skill, old and middle aged men which may indicate an increase in the demand for certain types of jobs like engineers and managers.
Five papers rely on TFP-style measures \citep{boyle2002trade, chen2014total, autor2018lshare, bessen2020autom, graham2000}.
All of them are regressions, mostly at the macro-level except for the study by \citet{bessen2020autom} who performed a detailed analysis of the steel, auto and textile industry at a historical scale, and the region-industry-level study by \citet{graham2000}. Four of these studies reported positive productivity effects. \citet{boyle2002trade} and \citet{autor2018lshare} documented a positive impact on wage income, \citet{bessen2020autom} additionally found decreasing prices in response to labor productivity growth, and \citet{graham2000} confirmed that higher output is positively associated with the demand for labor.
\citet{cirillo2017tech} examined the impact of innovation and found at the industry-level that an innovation-induced expansion of output and sales is positively associated with labor demand, especially for clerk and manual workers.
The studies by \citet{fagerberg1997tech, padalino1997employment} used technology measures that fall into our residual category other/indirect. \citet{fagerberg1997tech} used R\&D as a technology indicator and \citet{padalino1997employment} used changing growth-employment elasticities as an indirect proxy of technological change. They both documented a positive relationship between aggregate output and employment.
\subsubsection{Studies not supporting this effect}
Only four studies did not find supporting evidence for the real income effect \citep{badescu2009impact,colombo2013ict,samaniego2006tech,dixon2020decline} not finding any positive impact of technology on productivity. \citet{dixon2020decline} provided additional explorations on income effects.
\citet{badescu2009impact} and \citet{colombo2013ict} studied the impact of ICT-diffusion at the micro-level. \citet{dixon2020decline} used TFP-style measures, and \citet{samaniego2006tech} relied on indirect measures of technological change to study the impact on macro-level TFP-growth.
\citet{badescu2009impact} and \citet{colombo2013ict} could not confirm any significant relationship between productivity and ICT-diffusion, while \citet{dixon2020decline} and \citet{samaniego2006tech} reported even negative effects. \citet{dixon2020decline} additionally found that the impact of technology on income is negative.
Based on an empirically estimated general equilibrium simulations, \citet{samaniego2006tech} studied the impact of productivity shocks. The author argued that the negative productivity effect of a new technology can be explained by the incompatibility of an existing technology with a new one. However, \citet{samaniego2006tech} also argued that this may be only temporary.
\subsubsection{Studies with ambiguous findings and indirect evidence}
Six studies in our sample show ambiguous findings or only indirect evidence for the real income effect \citep{camina2020autom,fu2021,blien2021occup,compagnucci2019robotization,cozzarin2016inno,oesch2010}.
The studies by \citet{camina2020autom, fu2021, compagnucci2019robotization} studied the impact of robot-diffusion. \citet{camina2020autom} showed that whether robots are productivity-enhancing depends on the exact type of robot-based technology; for example, some technologies like data-driven control can even exhibit a negative effect.
\citet{fu2021} reported cross-country differences: While robots show a positive effect on labor productivity in developed economies, no significant effects are found for developing countries.
\citet{compagnucci2019robotization} found ambiguous results: They observed a negative effect of robots on wages, but a negative one on prices.
\citet{blien2021occup, oesch2010} relied on other/indirect measures. In particular, \citet{blien2021occup} looked at the routine-intensity of occupations and evaluated its interaction with the income of certain occupations. They observed that the effects on employees' income are heterogeneous across occupations: Employees in jobs with high routine-intensity experience income losses after a job layoff. The authors also showed that routine-intensity was a less significant predictor of income losses during the 1980s compared to more recent periods. In a descriptive macro-level analysis, \citet{oesch2010} used the changes in the wage growth of various occupations that vary in their task content. They examined the evolution of wages for different occupations but found only weak effects.
Finally, \citet{cozzarin2016inno} studied the impact of innovation on wages and productivity in manufacturing and only found weak aggregate effects. In tendency, \citet{cozzarin2016inno} observed a positive association between process innovation, wages and productivity, but a negative one for product innovation and productivity.
\section{Discussion}
\label{sec:discussion}
What is the net employment effect of technological change? It depends on which of the labor-saving or the labor-creating effects of technological change dominate. We systematically collected evidence for technology-induced labor replacement and creation by disentangling three mechanisms of labor-technology interactions: 1) direct replacement; 2) direct reinstatement; and 3) indirect real income effects. The first mechanism is labor-saving, while the latter two labor-creating.
In this section, we discuss the extent to which our analysis enables us to draw conclusions about the \emph{net} employment effects of technology.
\subsection{Evidence for net employment effect}
In total, we found at least as many studies supporting the labor-creating effects (n=64 for the reinstatement and n=23 for the real income) as studies supporting the replacement effect (n=69).
One concern is the existence of a publication bias, whereby studies with significant support for an effect are more likely to be published \citep{ugur2019inno}.
We found that within the three subsets of studies reporting evidence for one of the effects, the shares of studies supporting the respective effect are roughly comparable, i.e. 67\% for the replacement, 81\% for the reinstatement, and 70\% for the real income effect.
These numbers suggest that if there was a publication bias, the bias would be roughly equal across all effects.
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Studies by findings on net employment}
\label{tab:totalemp}
\begin{tabular}{l*{4}{c}} \toprule
& (1) & (2) & (3) & (4) \\
\csname @@input\endcsname{inputs/stats_total_byfind}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Columns (1)-(4) present the share and number (\#) of studies with empirical results reporting positive, depends (on various characteristics, e.g. technology type, analysis level, etc.), weak, and negative effects on net employment, respectively. The total \# of studies is 127.
\end{tablenotes}
\end{threeparttable}
\end{table}
The heterogeneity in the design of studies does not allow us to systematically compare the sizes of the effects reported. However, some studies provide direct insights on the net employment effect. Specifically, Table \ref{tab:totalemp} shows that a larger share finds support for an overall positive (29\%) than a negative employment effect (18\%).
Overall, our findings strongly suggest that technological progress has not resulted in a negative net employment effect in the past decades.
In the next section, we summarize and contextualize the insights gained from our analysis about the net employment effect for each technology.
\begin{table}[H] \centering \footnotesize
\begin{threeparttable}
\caption{Each type of technology by findings on net employment effect}
\label{tab:tech_byfinding_total}
\begin{tabular}{lc|*{4}{c}} \toprule
& (1) & (2) & (3) & (4) & (5) \\
& \multicolumn{1}{c}{}& \multicolumn{4}{c}{by finding} \\
\cmidrule{3-6}
& \multicolumn{1}{c}{total}& \multicolumn{1}{c}{positive}&\multicolumn{1}{c}{depends}&\multicolumn{1}{c}{weak}&\multicolumn{1}{c}{negative}\\ \midrule
\csname @@input\endcsname{inputs/stats_tech_byfind_total}
\bottomrule
\end{tabular}
\begin{tablenotes}[flushleft]
\scriptsize \item Notes: Column (1) presents the share and number (\#) of studies by the technology group considered in each row panel relative to the total number of studies examining the net employment effect. The row-sum of shares in column (1) do not add up to one since there are studies considering more than one technology group.
Columns (2)-(4) present the share and \# of studies by the set of findings reported on the net employment effect for each type of technology considered in each row panel. For findings, a study is classified as `positive' if it finds a significant positive empirical effect that supports the net employment effect examined. `Depends' refers to the set of studies that find varying effects depending on the type of technology or the (sub)sample analyzed (e.g. type of sector or labor). Studies reporting effects that are negligible in terms of the magnitude were classified as `weak'. Studies were labeled as `negative' if they investigate the effect of interest, but documented negative effects. The technology types reported include ICT, robots, innovation, TFP-style and Other types of technologies (e.g. indirectly measured through prices, automation risks), respectively. Note that there is no overlap in findings reported, i.e. more than one primary finding in each study, and thus the shares across columns add up to one, up to rounding. The total \# of studies is 127 of which seventy-five examine the net employment effect.
\end{tablenotes}
\end{threeparttable}
\end{table}
\subsubsection{Net employment effects of ICT}
The most ambiguous results are found in ICT studies.
For the replacement effect, we observed simultaneously a high number supporting (n=23) and a high number not (n=9) or only weakly (n=5) supporting the effect.
For the reinstatement effect, we found twenty-three supporting it. These studies confirm that ICT reinstates labor in new ICT-related occupations. The findings also suggest that the dissemination of ICT is related to an employment increase for high-skill workers and non-routine cognitive labor suggesting that ICT complements workers in the performance of non-routine tasks.
We also found a sample of studies reporting an ICT-related employment growth in the service sector.
For the real income effect, we identified seven studies reporting a positive effect of ICT on productivity, and a few other studies reporting positive income effects and a positive relationship between output growth and labor.
Most studies support both the Replacement and Reinstatement effect at the same time.
In line with these observations, eleven studies (42\%, Table \ref{tab:tech_byfinding_total}) that report net employment effects find that the impact depends on other factors, with the most important being the type of labor.
We also found a larger share reporting a positive (27\%) rather than a negative (8\%) effect on net employment.
We conclude that ICT does not appear to induce an overall net negative effect, but our results show that the jobs reinstated qualitatively differ from the jobs replaced. These findings have important policy and management implications. Although our results suggest that the employment perspectives have developed more favorably for high-skill workers whose skills complement ICTs, investing in the right type of skills will also benefit (future) workers in low- and middle-skill occupations.
Despite the fact that many occupations in the middle - and also some in the lower end - of the skill distribution are intensive in tasks that can be performed by ICTs, many jobs in these segments will continue to require a changing set of skills due to increasingly digitalized work environments \citep{autor2015there, levy2010technology}.
Think for example of automotive technicians who use computer systems to diagnose cars. These workers are required to have basic digital skills to access computer systems, but also the ability to interpret information in digital environments. The need to invest in the development of digital skills is also stressed by the 2020 European Skills Agenda \citep{european2020european}. One of the European Skill Agenda's objective is to significantly increase the share of adults with a basic level of digital skills. Moreover, various EU initiatives have been introduced to promote the development of digital skills in vocational education and training institutions and systems \citep{european2022vocational}.
Our results also indicate a shift of work from production to services within and across sectors. Hence, the skills demanded in the digitalized economy are not only \textit{prima facie} digital skills, but also social and emotional skills are required to perform well in service tasks \citep{autor2013polar}. The growing importance of social skills in the labor market is also reflected in the increasing wage returns for these skills \citep{deming2017growing}.
\subsubsection{Net employment effects of robots}
The highest relative support for the replacement effect comes from studies analyzing the impact of robots.
Out of the studies investigating the replacement effect of robots, thirteen (87\%, Table \ref{tab:tech_byfinding_replacement}) found support for this mechanism.
Nine studies support the reinstatement effect of robots and six studies support the real income effect finding positive productivity and income effects.
A large share of the reinstatement studies also reports net employment effects, but none of them finds clear evidence for a negative effect of robots on net employment.
Also when looking at the full sample of robot studies (Table \ref{tab:tech_byfinding_total}), we found that a larger share provides support for a positive net employment effect (36\%) rather than a negative (21\%).
A substantial share of robot studies exploring the total employment effect report negligible effects.
Overall, our findings suggest that the labor-saving impact of robots tends to be offset by robot-driven reinstatement of labor.
These results confirm a recent turn in the academic debate about the impact of robots and automation on jobs, suggesting that the fear of a jobless future \citep{smith2017automation, mokyr2015history} may be exaggerated and lack an empirical base \citep{economist2022robots, aghion2022effects}. This finding has important political implications; for example, it undermines the rationale of a robot tax which has been put forward as a means to cope with automation-induced unemployment.
Although robots and ICTs have replaced workers in the performance of routine tasks, robots differ from ICTs as they are particularly suitable to perform manual tasks. Moreover, in comparison to ICTs, robots are mainly considered as a pure automation technology with relatively clear-cut effects on labor demand while the effects of ICT are more ambiguous.
\subsubsection{Net employment effects of innovation}
The least support for the replacement effect is found by innovation studies. Only three studies support the real income effect (25\%, Table \ref{tab:tech_byfinding_replacement}).
The innovation studies showing support for the replacement effect found that the labor-saving effect is mostly driven by process innovation, but overall, the available evidence is insufficient to conclude that process innovation is labor-saving.
Those studies that did not support the replacement effect or reported ambiguous evidence suggest a positive effect of product innovation on employment.
However, a high number of studies remain inconclusive and found no significant or only weak effects of product, organizational or process innovation on employment.
For the reinstatement effect, we identified eight (57\%, Table \ref{tab:tech_byfinding_reinstatement}) studies supporting the effect. Another four studies (29\%) indicate that the reinstatement effect depends on the type of innovation. In general, studies seem to find a positive employment effect in the case of product innovation and a negative or ambiguous effect of process innovation.
Thirteen innovation studies provided insights on the net employment effect, but none of them found a negative effect. Six studies suggest that the impact is positive, while seven report ambiguous results.
In line with other reviews \citep{vivarelli2014innovation, ugur2017litreview, calvino2018innovation}, our findings suggest that the distinction between process and product innovation is essential when speaking about the employment effects of innovation.
It is not surprising that process innovation can be labor-saving as it is by definition an innovation aimed to make production processes more efficient, which is mostly equivalent with input-saving. However, process innovation is not necessarily labor-saving for three reasons: (1) Process innovation can also save other inputs rather than labor; (2) efficiency improvements may lead to lower prices and an expansion of output which can be associated with the creation of new jobs; and (3) process innovation can coincide with the introduction of technologies that complement human labor, creating labor in industries where new machines are produced.
It is also not surprising that product innovation tends to be labor-creating. The successful launch of a product innovation is associated with the acquisition of new markets which may lead to an expansion of economic activity of a firm, and thus with the creation of new jobs.
\subsubsection{Net employment effects of TFP-style technology}
Among the TFP-style studies, we found strong support for the simultaneous destruction of production labor, mostly in manufacturing, and creation of new non-production labor, mostly in services.
We identified thirteen studies (76\%, Table \ref{tab:tech_byfinding_replacement}) that support the replacement effect. Many of them showed labor demand shifts, particularly from production to non-production labor within and across industries, and from low- to high-skill workers.
For the reinstatement effect, we found ten studies (91\%, Table \ref{tab:tech_byfinding_reinstatement}) supporting it by documenting employment shifts from manufacturing to services and employment increases for skilled non-production labor.
Moreover, one reinstatement study showed that TFP shocks in upstream industries are associated with employment increases in downstream industries.
These observations support the idea that technological change leads to structural change with a reallocation of economic activity down the value chain, i.e. from more primary and secondary industries towards increasingly processed sectors and services \citep{kruger2008productivity, syrquin1988patterns}.
We also identified five studies supporting the real income effect that report rising wage incomes and decreasing prices. One study also confirmed a positive relationship between income and employment. The number of TFP-style studies finding support for the replacement effect is roughly balanced compared to the number of studies finding support for one of the two labor-creating mechanisms.
Among the TFP-style papers, only thirteen provided insights on net employment effects, among which only one suggests a positive impact, five suggest a negative impact, and the remaining found that the net impact is weak or depends.
The theory of structural change suggests a reallocation of labor from
high to low productivity growth industries with increasing capital intensity and automation as major drivers of productivity growth \citep{syrquin1988patterns}. Theoretical models of structural change further suggest that the elasticity of demand is decisive for the sufficient reinstatement of labor \citep{ngai2007structural} which we aimed to capture through the assessment of the real income effect.
Our results for the net employment effect are rather inconclusive, but suggest that the net impact is more negative than positive. This finding indicates that the reinstatement of labor in more downstream and service-related industries may be insufficient to offset the labor-saving impact.
However, as a critical note, it should be flagged that historically productivity and welfare increases coincided with a reduction of weekly working hours, which may be seen as a very desirable effect \citep{bick2018hours}.
\subsubsection{Net employment effects of other/indirect technology}
The support for the labor-saving and labor-creating effects from studies relying on other/indirect measures is roughly balanced.
We found twenty-two studies supporting the replacement effect (73\%, Table \ref{tab:tech_byfinding_replacement}) and twenty studies (91\%, Table \ref{tab:tech_byfinding_reinstatement}) supporting the reinstatement effect. Another two studies (40\%, Table \ref{tab:tech_byfinding_realincome}) provide support for the Real income effect. {Furthermore, we find that three studies do not report any of the effects discussed above, but provide results on net employment.\footnote{These are macro-level studies by \cite{samaniego2008tech, balleer2012new, karabarbounis2013lshare}, with the former reporting positive net employment effects, while the remaining studies suggest the presence of negative net employment effects.}}
A first set of studies analyzed the relationship between automation risk indicators and employment. Studies supporting the replacement effect largely found negative employment effects in occupations with high automation risks, while studies that support the reinstatement effect reported employment increases in low automation risk jobs.
A second set of studies used capital and high-tech equipment investments as measures of technological change. Also these studies found support for both the replacement and reinstatement effect.
The same holds for studies looking at the impact of R\&D expenditures and patents.
One study on the real income effect also reported a positive relation between R\&D, aggregate output and employment.
\textcolor{red}{}
Studies on the replacement effect found that especially low-skill labor is negatively affected. However, we also found a few descriptive studies providing evidence that the demand for low-skill labor can also be positively affected, especially in service jobs.
Overall, the results suggest a bias of technological change that leads to a reallocation of labor towards non-production labor and high-tech industries, which is often, but not necessarily, high-skill. These findings provide both support for the skill-biased technological change hypothesis \citep{acemoglu2002technical} as well as for the routine-biased technological change hypothesis that predicts a polarization of the labor market \citep{autor2003skill}.
Furthermore, we found that a number of indirect measures of technology only provide evidence for one of the three employment mechanisms. One replacement effect study reports a negative association between GDP growth and employment in manufacturing, while a number of decomposition-like and substitution elasticity-based studies found support for the reinstatement effect.
Finally, one study on the real income effect uses growth-employment elasticities as an indirect measure for technology and reports a positive relationship between aggregate output and employment. Again, we conclude that indirect measures of technological change provide a roughly equal amount of evidence for the labor-saving effect and for the labor-creating effect.
In total, twenty-nine studies in the category other/indirect also investigated the net employment effect. The share of papers reporting an overall positive net employment effect (28\%) is larger than the share of papers documenting an overall negative effect (21\%). A substantial share of papers report that the overall employment effect depends on other factors (41\%). Hence, the set of studies using alternative measures do not support the idea that technological change has resulted in a net destruction of jobs.
These indirect measures relying on automation risks have a number of disadvantages. For instance, the primary error engendered by automation risk indicators is that they implicitly assume that tasks that can be performed by machines will automatically substitute for human labor.
Moreover, occupation-level automation risk indicators are likely to overestimate the share of automatable jobs as they disregard task heterogeneity within occupations as well as the adaptability of jobs in response to technological change \citep{arntz2017}.
\section{Conclusions}
\label{sec:conclusion}
This study systematically reviewed the available evidence on the impact of technological change on employment. Overall, we found that a substantially larger number of studies provide evidence for the labor-creating impact of technological change than for the labor-displacing impact. Several studies providing support for the labor-creating impact of technology report positive effects on productivity and prices, income, and final demand or output. Through these channels, technological change is expected to indirectly increase the demand for labor. However, only a few studies measure the actual employment implications of these mechanisms.
Although we are careful in concluding that, if anything, technological change has a positive net effect on employment, we do conclude that the replacement effect is likely to be more than offset by the labor-creating effect of technology. Hence, there does not appear to be an empirical foundation for the fear of technology-driven unemployment. We also investigated whether different types of technology have a differential effect on total employment. For almost all technology measures (i.e. robots, ICT, TFP-style, other/indirect measures) we found a comparable number of studies that support the labor-displacing as well as labor-creating effect of technology. Only for innovation measures, the empirical evidence suggests that product innovation is mostly labor-creating, while the evidence concerning process innovation remains inconclusive.
Despite the fact that we found no evidence for a negative net employment effect in quantitative terms, the qualitative impact of technological change on employment cannot be neglected. As our systematic review has pointed out, different types of technology have adverse effects on predominantly low-skill and production labor, manufacturing jobs, and workers performing routine tasks. Given the considerable labor-saving potential of technology, reskilling or retraining workers whose jobs are susceptible to automation is essential. However, not only individuals whose jobs have disappeared due to automation would benefit from (re-)training. Research shows that technological change also has induced within-occupational changes in skill requirements. In fact, according to estimates of the OECD \citep{arntz2017}, the share of jobs that will face changes in the task content due to automation is higher than the share of jobs that is of high risk of being automated. Hence, upskilling is also of utmost importance for workers whose jobs are not directly threatened by technological change. Furthermore, some displaced workers may not be able to make the transition to new jobs after experiencing job loss. These workers might have to rely on social support systems.
Our systematic review is subject to a number of limitations. First, empirical studies can only cover the impact of technologies that are already available today, but the scope of tasks that may be automated in the near future continuously expands \citep{brynjolfsson2014second, frey2017future}. Hence, empirical evidence on the impact of artificial intelligence, quantum computing, virtual reality, biotechnology, nanotechnology, renewable energy, and other emerging technologies that will soon impact our economy remains limited. In fact, none of the studies in our review assessed the impact of this new wave of technological innovation. To that end, it is unclear to which extent our findings can be extrapolated into the future. Therefore, real-time monitoring and ongoing research is needed to more fully understand the emerging impact of the introduction of new technologies on the future of work \citep{baldwin2019globotics}.
Second, our study also faces an inherent methodological challenge. Generally, it should be noted that it is not surprising to observe high support rates for each of the effects as studies with insignificant results are rarely published. Hence, these results are subject to a reporting bias, similar to the publication selection bias discussed in \citet{ugur2019inno}. This appears to be a more general problem of published empirical studies which also exists in other disciplines such as medicine, psychology, and experimental economics research. In these fields, pre-registrations were introduced with a detailed protocol of the empirical approach that has to be approved before the experimental study is conducted. While similar procedures are more difficult to establish in statistical analyses that often need to be adapted during the workflow, it should be kept in mind that the results reported are likely to suffer from a reporting bias in favor of the analyzed theory. Nevertheless, there is no reason to assume that the size of the bias differs across the three employment effects that we analyze in our study.
\newpage
\printbibliography
\newpage
\FloatBarrier
|
\section{Introduction and main results}
Since the ground breaking work \cite{harnack1887basics}, Harnack-type inequalities have become an essential tool in the analysis of partial differential equations (\textsc{PDE}). In \cite{MR1729395,MR3522349,MR3481178}, Harnack inequalities have been applied to antisymmetric functions---functions that are odd with respect to reflections across a given plane, for more detail see Section \ref{LenbLVkk}---in conjunction with the method of moving planes to obtain stability results for local overdetermined problems including \emph{Serrin's overdetermined problem} and the \emph{parallel surface problem}; see also \cite{MR3932952}.
In recent years, there has been consolidated interest in overdetermined problems driven by nonlocal operators, particularly the fractional Laplacian; however, standard Harnack inequalities for such operators are incompatible with antisymmetric functions since they require functions to be non-negative in all of \(\mathbb{R}^n\). In this paper, we will address this problem by proving a Harnack inequality for antisymmetric \(s\)-harmonic functions with zero-th order terms which only requires non-negativity in a halfspace. By allowing zero-th order terms this result is directly applicable to symmetry and overdetermined problems for semilinear equations driven by the fractional Laplacian.
\subsection{Background} Fundamentally the original Harnack inequality for elliptic \textsc{PDE} is a quantitative formulation of the strong maximum principle and directly gives, among other things, Liouville's theorem, the removable singularity theorem, compactness results, and Hölder estimates for weak and viscosity solutions, see~\cite{MR1814364,MR2597943}. This has led it to be extended to a wide variety of other settings. For local \textsc{PDE} these include:
linear parabolic \textsc{PDE}~\cite{MR2597943,MR1465184}; quasilinear and fully nonlinear
elliptic \textsc{PDE}~\cite{MR170096,MR226198,MR1351007}; quasilinear and fully nonlinear
parabolic \textsc{PDE}~\cite{MR244638,MR226168}; and in connection with curvature and geometric
flows such as Ricci flow on Riemannian manifolds~\cite{MR431040,MR834612,MR2251315}.
An extensive survey on the Harnack inequality for local \textsc{PDE} is~\cite{MR2291922}.
For equations arising from jump processes, colloquially known as nonlocal \textsc{PDE}, the first Harnack inequality is due to Bogdan~\cite{MR1438304} who proved the boundary Harnack inequality for the fractional Laplacian. The fractional Laplacian is the prototypical example of a nonlocal operator and is defined by \begin{align}
(-\Delta)^s u(x) &= c_{n,s} \mathrm{P.V.} \int_{\mathbb{R}^n} \frac{u(x)-u(y)}{\vert x - y \vert^{n+2s}} \dd y \label{ZdAlT}
\end{align}
where~$s\in(0,1)$, \(c_{n,s}\) is a positive normalisation constant (see~\cite{MR3916700} for more details, particularly Proposition~5.6 there) and \(\mathrm{P.V.}\) refers to the Cauchy principle value. The result in~\cite{MR1438304} was only valid for Lipschitz domains, but this was later extended to any bounded domain in~\cite{MR1719233}.
Over the proceeding decade there were several papers proving Harnack inequalities for more general jump processes including~\cite{MR1918242,MR3271268}, see also~\cite{MR2365478}. For fully nonlinear nonlocal \textsc{PDE}, a Harnack inequality was established in~\cite{MR2494809}, see also~\cite{MR2831115}. More recently, in~\cite{MR4023466} the boundary Harnack inequality was proved for nonlocal \textsc{PDE} in non-divergence form.
As far as we are aware, the only nonlocal Harnack inequality for antisymmetric functions in the literature is in~\cite{ciraolo2021symmetry} where a boundary Harnack inequality was established for antisymmetric \(s\)-harmonic functions in a small ball centred at the origin. Our results generalise this to arbitrary symmetric domains and to equations with zero-th order terms.
\subsection{Main results} \label{K4mMv} Let us now describe in detail our main results. First, we will introduce some useful notation. Given a point \(x=(x_1,\dots , x_n) \in \mathbb{R}^n\), we will write \(x=(x_1,x')\) with \(x'=(x_2,\dots , x_n)\) and we denote by \(x_\ast\) the reflection of \(x\) across the plane \(\{x_1=0\}\), that is, \(x_\ast = (-x_1,x')\). Then we call a function \(u:\mathbb{R}^n \to \mathbb{R}\) \emph{antisymmetric} if \begin{align*}
u(x_\ast) = -u(x) \qquad \text{for all } x\in \mathbb{R}^n.
\end{align*} We will also denote by \(\mathbb{R}^n_+\) the halfspace \(\{x\in \mathbb{R}^n \text{ s.t. } x_1>0\}\) and, given \(A\subset \mathbb{R}^n\), we let \(A^+ := A \cap \mathbb{R}^n_+\). Moreover, we will frequently make use of the
functional space \(\mathscr A_s(\mathbb{R}^n)\) which we define to be the set of all antisymmetric functions \(u\in L^1_{\mathrm {loc}}(\mathbb{R}^n)\) such that \begin{align}
\Anorm{u}:= \int_{\mathbb{R}^n_+} \frac{x_1 \vert u(x) \vert}{1+ \vert x \vert^{n+2s+2}}\dd x \label{QhKBn}
\end{align} is finite.
Our main result establishes the Harnack inequality in a general symmetric domain \(\Omega\).
\begin{thm} \thlabel{Hvmln}
Let \(\Omega \subset \mathbb{R}^n\) and \(\tilde \Omega \Subset \Omega\) be bounded domains that are symmetric with respect to \(\{x_1=0\}\), and let \(c\in L^\infty(\Omega^+)\). Suppose that \(u \in C^{2s+\alpha}(\Omega)\cap \mathscr{A}_s(\mathbb{R}^n) \) for some \(\alpha >0\) with \(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\), and satisfies \begin{align*}
(-\Delta)^s u +cu&=0 \qquad \text{in }\Omega^+.
\end{align*}
Then there exists \(C>0\) depending only on \(\Omega\), \(\tilde \Omega\), \(\| c\|_{L^\infty(\Omega^+)}\), \(n\), and \(s\) such that \begin{align*}
\sup_{\tilde \Omega^+} \frac{u(x)}{x_1} \leqslant C \inf_{ \tilde \Omega^+} \frac{u(x)}{x_1} .
\end{align*} Moreover, \(\inf_{\tilde \Omega^+} \frac{u(x)}{x_1}\) and \(\sup_{\tilde \Omega^+} \frac{u(x)}{x_1}\) are comparable to \(\Anorm{u}\).
\end{thm}
Here and throughout this document we use the convention that if \(\beta>0\) and \(\beta\) is not an integer then \(C^\beta (\Omega)\) denotes the Hölder space \(C^{k,\beta'}(\Omega)\) where \(k\) is the integer part of \(\beta\) and \(\beta'=\beta-k \in (0,1)\).
One can compare the Harnack inequality in Theorem~\ref{Hvmln} here
with previous results in the literature, such as Proposition~6.1 in~\cite{MR4308250},
which can be seen as a boundary Harnack inequality for antisymmetric functions,
in a rather general fractional elliptic setting. We point out that Theorem~\ref{Hvmln} here
holds true without any sign assumption on~$c$ (differently from Proposition~6.1 in~\cite{MR4308250}
in which a local sign assumption on~$c$ was taken), for all~$s\in(0,1)$ (while the analysis of Proposition~6.1 in~\cite{MR4308250}
was focused on the case~$s\in\left[\frac12,1\right)$), and in every dimension~$n$ (while Proposition~6.1 in~\cite{MR4308250} dealt with the case~$n=1$).
We would like to emphasise that, in contrast to the standard Harnack inequality for the fractional Laplacian, \thref{Hvmln} only assumes \(u\) to be non-negative in the halfspace \(\mathbb{R}^n_+\) which is a more natural assumption for antisymmetric functions. Also, note that the assumption that \(\tilde \Omega \Subset \Omega\) allows~\(\tilde \Omega^+\) to touch \(\{x_1=0\}\), but prevents \(\tilde \Omega^+\) from touching the portion of \(\partial\Omega\) that is not on \(\{x_1=0\}\); it is in this sense that we will sometimes refer to \thref{Hvmln} (and later \thref{C35ZH}) as a \emph{boundary} Harnack inequality for antisymmetric functions.
Interestingly, the quantity \(\Anorm{u}\) arises very naturally in the context of antisymmetric functions and plays the same role that \begin{align}
\| u\|_{\mathscr L_s (\mathbb{R}^n)} := \int_{\mathbb{R}^n} \frac{\vert u (x) \vert }{1+\vert x \vert^{n+2s}} \dd x \label{ACNAzPn8}
\end{align} plays in the non-antisymmetric nonlocal Harnack inequality, see for example \cite{MR4023466}. To our knowledge, \(\Anorm{\cdot}\) is new in the literature.
A technical aspect of \thref{Hvmln}, however, is that the fractional Laplacian is in general not defined if we simply have \(u\) is \(C^{2s+\alpha}\) and \(\Anorm{u} < +\infty\). This leads us to introduce the following new definition of the fractional Laplacian for functions in \(\mathscr A_s(\mathbb{R}^n)\) which we will use throughout the remainder of this paper.
\begin{defn} \thlabel{mkG4iRYH}
Let \(x\in \mathbb{R}^n_+\) and suppose that \(u \in \mathscr A_s(\mathbb{R}^n)\) is \( C^{2s+\alpha}\) in a neighbourhood of \(x\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer. The fractional Laplacian of \(u\) at \(x\), which we denote as usual by \((-\Delta)^su(x)\), is defined by \begin{align}
(-\Delta)^su(x) = c_{n,s} \lim_{\varepsilon\to 0^+} \int_{\mathbb{R}^n_+ \setminus B_\varepsilon(x)} \left(
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}}\right) \big(u(x)-u(y)\big)\dd y + \frac {c_{1,s}} s u(x)x_1^{-2s} \label{OopaQ0tu}
\end{align} where \(c_{n,s}\) is the constant from~\eqref{ZdAlT}.
\end{defn}
We will motivate \thref{mkG4iRYH} in Section \ref{LenbLVkk} as well as verify it is well-defined. We also prove in Section \ref{LenbLVkk} that if \(\| u\|_{\mathscr L_s(\mathbb{R}^n)}\) is finite, $u$ is \( C^{2s+\alpha}\) in a neighbourhood of \(x\), and antisymmetric then \thref{mkG4iRYH} agrees with~\eqref{ZdAlT}.
See also~\cite{MR3453602, MR4108219, MR4030266, MR4308250} where maximum principles in the setting
of antisymmetric solutions of integro-differential equations have been established by exploiting
the antisymmetry of the functions as in~\eqref{OopaQ0tu}.
It is also worth mentioning that the requirement that \(u\) is antisymmetric in \thref{Hvmln} cannot be entirely removed. In Appendix \ref{j4njb}, we construct a sequence of functions which explicitly demonstrates this. Moreover, \thref{Hvmln} is also false if one only assumes \(u\geqslant0\) in \(\Omega^+\) as proven in \cite[Corollary 1.3]{RoleAntisym2022}.
To obtain the full statement of \thref{Hvmln} we divide the proof into two parts: an interior Harnack inequality and a boundary Harnack inequality close to \(\{x_1=0\}\). The interior Harnack inequality is given as follows.
\begin{thm} \thlabel{DYcYH}
Let \(\rho\in(0,1)\) and \(c\in L^\infty(B_\rho(e_1))\). Suppose that \(u \in C^{2s+\alpha}(B_\rho(e_1))\cap \mathscr{A}_s(\mathbb{R}^n) \) for some \(\alpha >0\) with \(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\), and satisfies \begin{align*}
(-\Delta)^s u +cu =0 \qquad\text{in } B_{\rho}(e_1).
\end{align*}
Then there exists \(C_\rho>0\) depending only on \(n\), \(s\), \(\| c \|_{L^\infty(B_\rho(e_1))}\), and \(\rho\) such that \begin{align*}
\sup_{B_{\rho/2}(e_1)} u \leqslant C_\rho \inf_{B_{\rho/2}(e_1)} u.
\end{align*} Moreover, both the quantities \(\sup_{B_{\rho/2}(e_1)} u \) and \(\inf_{B_{\rho/2}(e_1)} u \) are comparable to \(\Anorm{u}\).
\end{thm}
For all \(r>0\), we write \(B_r^+ := B_r \cap \mathbb{R}^n_+\). Then the following result is the antisymmetric boundary Harnack inequality.
\begin{thm} \thlabel{C35ZH}
Let \(\rho>0\) and \(c\in L^\infty(B_\rho^+)\). Suppose that \(u \in C^{2s+\alpha}(B_\rho)\cap \mathscr{A}_s(\mathbb{R}^n) \) for some \(\alpha >0\) with \(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\), and satisfies \begin{align*}
(-\Delta)^s u +cu&=0 \qquad \text{in } B_\rho^+.
\end{align*}
Then there exists \(C_\rho>0\) depending only on \(n\), \(s\), \(\| c \|_{L^\infty(B_\rho^+)}\), and \(\rho\) such that\begin{align*}
\sup_{x\in B_{\rho/2}^+} \frac{u(x)}{x_1} \leqslant C_\rho \inf_{x\in B_{\rho/2}^+} \frac{u(x)}{x_1}.
\end{align*} Moreover, both the quantities \(\sup_{x\in B_{\rho/2}^+} \frac{u(x)}{x_1} \) and \(\inf_{x\in B_{\rho/2}^+} \frac{u(x)}{x_1} \) are comparable to \(\Anorm{u}\).
\end{thm}
To our knowledge, Theorems~\ref{DYcYH} and~\ref{C35ZH} are new in the literature. In the particular case \(c\equiv0\), \thref{C35ZH} was proven in \cite{ciraolo2021symmetry}, but the proof relied on the Poisson representation formula for the Dirichlet problem in a ball which is otherwise unavailable in more general settings.
The proofs of Theorems~\ref{DYcYH} and~\ref{C35ZH} are each split into two halves: for \thref{DYcYH} we prove an interior weak Harnack inequality for super-solutions (\thref{MT9uf}) and interior local boundedness of sub-solutions (\thref{guDQ7}). Analogously, for \thref{C35ZH} we prove a boundary weak Harnack inequality for super-solutions (\thref{SwDzJu9i}) and boundary local boundedness of sub-solutions (\thref{EP5Elxbz}). The proofs of Propositions~\ref{MT9uf}, \ref{guDQ7}, \ref{SwDzJu9i}, and~\ref{EP5Elxbz} make use of general barrier methods which take inspiration from \cite{MR4023466,MR2494809,MR2831115}; however, these methods require adjustments which take into account the antisymmetry assumption.
Once Theorems~\ref{DYcYH} and \ref{C35ZH} have been established, \thref{Hvmln} follows easily from a standard covering argument. For completeness, we have included this in Section \ref{lXIUl}.
Finally, in Appendix~\ref{4CEly} we provide an alternate elementary proof of \thref{C35ZH} in the particular case \(c\equiv0\). As in \cite{ciraolo2021symmetry}, our proof relies critically on the Poisson representation formula in a ball for the fractional Laplacian, but the overall strategy of our proof is entirely different. This was necessary to show that \(\sup_{x\in B_{\rho/2}^+} \frac{u(x)}{x_1} \) and \(\inf_{x\in B_{\rho/2}^+} \frac{u(x)}{x_1} \) are comparable to \(\Anorm{u}\) which was not proven in \cite{ciraolo2021symmetry}. Our proof makes use of a new mean-value formula for antisymmetric \(s\)-harmonic functions, which we believe to be interesting
in and of itself.
The usual mean value formula for \(s\)-harmonic functions says that if \(u\) is \(s\)-harmonic in \(B_1\) then we have that \begin{align}
u(0) = \gamma_{n,s} \int_{\mathbb{R}^n\setminus B_r} \frac{r^{2s} u(y)}{(\vert y \vert^2-r^2)^s\vert y \vert^n}\dd y \qquad \text{for all }r\in(0, 1]. \label{K8kw5rpi}
\end{align} {F}rom antisymmetry, however, both the left hand side and the right hand side of~\eqref{K8kw5rpi} are zero irrespective of the fact \(u\) is \(s\)-harmonic. It is precisely this observation that leads to the consideration of \(\partial_1u(0)\) instead of \(u(0)\) which is more appropriate when \(u\) is antisymmetric.
\begin{prop} \thlabel{cqGgE}
Let \(u\in C^{2s+\alpha}(B_1) \cap \mathscr L_s(\mathbb{R}^n)\) with \(\alpha>0\) and \(2s+\alpha\) not an integer. Suppose that~\(u\) is antisymmetric and \(r\in(0, 1]\). If \((-\Delta)^s u = 0 \) in \(B_1\) then \begin{align*}
\frac{\partial u}{\partial x_1} (0) &= 2n \gamma_{n,s} \int_{\mathbb{R}^n_+ \setminus B_r^+} \frac{r^{2s}y_1 u(y)}{(\vert y \vert^2 - r^2)^s\vert y \vert^{n+2}} \dd y .
\end{align*} Here
\begin{equation}\label{sry6yagamma098765} \gamma_{n,s}:= \frac{\sin(\pi s)\Gamma (n/2)}{\pi^{\frac n 2 +1 }}.
\end{equation}
\end{prop}
\subsection{Organisation of paper}
The paper is organised as follows. In Section \ref{LenbLVkk}, we motivate \thref{mkG4iRYH} and establish some technical lemmata regarding this definition. In Section~\ref{q7hKF}, we
prove \thref{DYcYH} and, in Section~\ref{TZei6Wd4}, we prove \thref{C35ZH}. In Section~\ref{lXIUl}, we prove the main result \thref{Hvmln}. In Appendix~\ref{j4njb}, we construct a sequence which demonstrates the antisymmetric assumption in \thref{Hvmln} cannot be removed. Finally, in Appendix~\ref{4CEly}, we prove
Proposition~\ref{cqGgE}, and with this we provide an elementary alternate proof of \thref{C35ZH} for the particular case~\(c\equiv0\).
\section{The antisymmetric fractional Laplacian} \label{LenbLVkk}
Let \(n\) be a positive integer and \(s\in (0,1)\). The purpose of this section is to motivate \thref{mkG4iRYH} as well as prove some technical aspects of the definition such as that it is well-defined and that it coincides with~\eqref{ZdAlT} when \(u\) is sufficiently regular. Recall that, given a point \(x=(x_1,\dots , x_n) \in \mathbb{R}^n\), we will write \(x=(x_1,x')\) with \(x'=(x_2,\dots , x_n)\) and we denote by \(x_\ast\) the reflection of \(x\) across the plane \(\{x_1=0\}\), that is, \(x_\ast = (-x_1,x')\). Then we call a function \(u:\mathbb{R}^n \to \mathbb{R}\) \emph{antisymmetric} if \begin{align*}
u(x_\ast) = -u(x) \qquad \text{for all } x\in \mathbb{R}^n.
\end{align*} It is common in the literature, particularly when dealing with the method of moving planes, to define antisymmetry with respect to a given hyperplane \(T\); however, for our purposes, it is sufficient to take~\(T=\{x_1=0\}\).
In order for the fractional Laplacian of \(u\), given by~\eqref{ZdAlT}, to be well-defined in a pointwise sense, we need two ingredients: \(u\) needs enough regularity in a neighbourhood of \(x\) to overcome the singularity at \(x\) in the kernel \(\vert x-y\vert^{-n-2s}\), and \(u\) also needs an integrability condition at infinity to account for the integral in~\eqref{ZdAlT} being over \(\mathbb{R}^n\). For example, if \(u\) is \(C^{2s+\alpha}\) in a neighbourhood of \(x\) for some~\(\alpha >0\) with \(2s+\alpha\) not an integer and \(u\in\mathscr L_s(\mathbb{R}^n)\) where \(\mathscr L_s (\mathbb{R}^n)\) is the set of functions \(v\in L^1_{\mathrm{loc}}(\mathbb{R}^n)\) such that \(\| v\|_{\mathscr L_s(\mathbb{R}^n)} <+\infty\) (recall \(\| \cdot \|_{\mathscr L_s(\mathbb{R}^n)}\) is given by~\eqref{ACNAzPn8}) then \((-\Delta)^su\) is well-defined at \(x\) and is in fact continuous there, \cite[Proposition 2.4]{MR2270163}.
In the following proposition, we show that if \(u\in \mathscr L_s(\mathbb{R}^n)\) is antisymmetric and \(u\) is \(C^{2s+\alpha}\) in a neighbourhood of \(x\) for some \(\alpha >0\) with \(2s+\alpha\) not an integer then \(u\) satisfies~\eqref{OopaQ0tu}. This simultaneously motivates \thref{mkG4iRYH} and demonstrates this definition does in fact generalise the definition of the fractional Laplacian when the given function is an antisymmetric function.
\begin{prop} \thlabel{eeBLRjcZ}
Let \(x\in \mathbb{R}^n_+\) and \(u \in \mathscr L_s(\mathbb{R}^n)\) be an antisymmetric function that is \(C^{2s+\alpha}\) in a neighbourhood of \(x\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer. Then~\eqref{ZdAlT} and \thref{mkG4iRYH} coincide.
\end{prop}
\begin{proof}
Let~$x=(x_1,x')\in \mathbb{R}^n_+$ and take~$\delta\in\left(0,\frac{x_1}2\right)$ such that~$u$
is~$C^{2s+\alpha}$ in~$B_\delta(x_1)\subset\mathbb{R}^n_+$. Furthermore, let \begin{align}
(-\Delta)^s_\delta u(x):= c_{n,s}\int_{\mathbb{R}^n \setminus B_\delta(x)} \frac{u(x)-u(y)}{\vert x - y\vert^{n+2s}} \dd y .\label{WUuXb0Hk}
\end{align}
By the regularity assumptions on \(u\), the integral in~\eqref{ZdAlT} is well-defined and \begin{align*}
\lim_{\delta\to0^+}(-\Delta)^s_\delta u(x) = c_{n,s} \mathrm{P.V.} \int_{\mathbb{R}^n}& \frac{u(x)-u(y)}{\vert x - y \vert^{n+2s}} \dd y =(-\Delta)^s u(x).
\end{align*}
Splitting the integral in~\eqref{WUuXb0Hk} into two integrals over \(\mathbb{R}^n_+ \setminus B_\delta(x)\) and \(\mathbb{R}^n_-\) respectively and
then using that \(u\) is antisymmetric in the integral over \(\mathbb{R}^n_-\), we obtain
\begin{align}
(-\Delta)^s_\delta u(x) &= c_{n,s}\int_{\mathbb{R}^n_+ \setminus B_\delta(x)} \frac{u(x)-u(y)}{\vert x - y\vert^{n+2s}} \dd y +c_{n,s} \int_{\mathbb{R}^n_+} \frac{u(x)+u(y)}{\vert x_\ast - y\vert^{n+2s}} \dd y \nonumber \\
&= c_{n,s}\int_{\mathbb{R}^n_+ \setminus B_\delta(x)} \left(
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}} \right)\big(u(x)-u(y)\big) \dd y \nonumber \\
&\qquad +2c_{n,s}u(x) \int_{\mathbb{R}^n_+ \setminus B_\delta(x)} \frac{\dd y }{\vert x_\ast - y \vert^{n+2s}} + c_{n,s} \int_{B_\delta(x)} \frac{u(x)+u(y) }{\vert x_\ast - y \vert^{n+2s}}\dd y . \label{cGXQjDIV}
\end{align}
We point out that if~$y\in\mathbb{R}^n_+$ then~$|x_\ast-y|\ge x_1$,
and therefore
\begin{eqnarray*}
&& \left \vert \int_{B_\delta(x)} \frac{u(x)+u(y) }{\vert x_\ast - y \vert^{n+2s}}\dd y \right \vert\le
\int_{B_\delta(x)} \frac{2|u(x)|+|u(y)-u(x)| }{\vert x_\ast - y \vert^{n+2s}}\dd y\\
&&\qquad\qquad \le \int_{B_\delta(x)} \frac{2|u(x)|+C|x-y|^{\min\{1,2s+\alpha\}} }{\vert x_\ast - y \vert^{n+2s}}\dd y\le C\big ( \vert u(x) \vert + \delta^{\min\{1,2s+\alpha\}} \big ) \delta^n,
\end{eqnarray*}
for some~$C>0$ depending on~$n$, $s$, $x$, and the norm of~$u$ in a neighbourood of~$x$.
As a consequence, sending \(\delta \to 0^+\) in~\eqref{cGXQjDIV}
and using the Dominated Convergence Theorem,
we obtain \begin{align*}
(-\Delta)^s u(x) &= c_{n,s}\lim_{\delta \to 0^+}\int_{\mathbb{R}^n_+ \setminus B_\delta(x)} \left( \frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}} \right)\big(u(x)-u(y)\big) \dd y \\
&\qquad +2c_{n,s}u(x) \int_{\mathbb{R}^n_+} \frac{\dd y }{\vert x_\ast - y \vert^{n+2s}}.
\end{align*}
Also, making the change of variables \(z=(y_1/x_1 , (y'-x')/x_1)\) if \(n>1\) and \(z_1=y_1/x_1\) if \(n=1\), we see that \begin{align*}
\int_{\mathbb{R}^n_+} \frac{\dd y }{\vert x_\ast - y \vert^{n+2s}} &= x_1^{-2s} \int_{\mathbb{R}^n_+} \frac{\dd z}{\vert e_1+ z \vert^{n+2s}}.
\end{align*}
Moreover, via a direct computation \begin{equation}\label{sd0w3dewt95b76 498543qdetr57uy54u}
c_{n,s} \int_{\mathbb{R}^n_+} \frac{\dd z}{\vert e_1+ z \vert^{n+2s}} = \frac {c_{1,s}}{2s}
\end{equation} see \thref{tuMIrhco} below for details.
Putting together these considerations, we obtain the desired result.
\end{proof}
\begin{remark} \thlabel{tuMIrhco}
Let \begin{align}
\tilde c_{n,s} := c_{n,s} \int_{\mathbb{R}^n_+} \frac{\dd z}{\vert e_1+ z \vert^{n+2s}}. \label{CA7GPGvx}
\end{align}The value of \(\tilde c_{n,s}\) is of no consequence to \thref{mkG4iRYH} and so, for the purposes of this paper, is irrelevant; however, it is interesting to note that \(\tilde c_{n,s}\) can be explicitly evaluated and is in fact equal to \((2s)^{-1}c_{1,s}\) (in particular \(\tilde c_{n,s}\) is actually independent of \(n\) which is not obvious from~\eqref{CA7GPGvx}). Indeed, if \(n>1\) then \begin{align*}
\int_{\mathbb{R}^n_+} \frac{\dd z}{\vert e_1+ z \vert^{n+2s}} &= \int_0^\infty \int_{\mathbb{R}^{n-1}} \frac{\dd z_1\dd z'}{\big ( (z_1+1)^2 + \vert z' \vert^2\big )^{\frac{n+2s}2}}
\end{align*} so, making the change of variables \(\tilde z'= (z_1+1)^{-1} z'\) in the inner integral, we have that \begin{align*}
\int_{\mathbb{R}^n_+} \frac{\dd z}{\vert e_1+ z \vert^{n+2s}} &= \int_0^\infty \frac 1 {(z_1+1)^{1+2s}} \left( \int_{\mathbb{R}^{n-1}} \frac{\dd \tilde z'}{\big ( 1 + \vert \tilde z' \vert^2\big )^{\frac{n+2s}2}} \right) \dd z_1.
\end{align*} By \cite[Proposition 4.1]{MR3916700}, \begin{align*}
\int_{\mathbb{R}^{n-1}} \frac{\dd \tilde z'}{\big ( 1 + \vert \tilde z' \vert^2\big )^{\frac{n+2s}2}} &= \frac{\displaystyle
\pi^{\frac{n-1}2} \Gamma \big ( \frac{1+2s}2 \big )}{\displaystyle \Gamma \big ( \frac{n+2s}2 \big )}.
\end{align*} Hence, \begin{align*}
\int_{\mathbb{R}^n_+} \frac{\dd z}{\vert e_1+ z \vert^{n+2s}} &= \frac{\pi^{\frac{n-1}2} \Gamma \big ( \frac{1+2s}2 \big )}{\Gamma \big ( \frac{n+2s}2 \big )} \int_0^\infty \frac {\dd z_1} {(z_1+1)^{1+2s}} = \frac{\pi^{\frac{n-1}2} \Gamma \big ( \frac{1+2s}2 \big )}{2s \Gamma \big ( \frac{n+2s}2 \big )}.
\end{align*} This formula remains valid if \(n=1\). Since \(c_{n,s} = s \pi^{-n/2}4^s \Gamma \big ( \frac{n+2s}2 \big )/ \Gamma (1-s)\), see \cite[Proposition~5.6]{MR3916700}, it follows that \begin{align*}
\tilde c_{n,s} &= \frac{2^{2s-1} \Gamma \big ( \frac{1+2s}2 \big )}{\pi^{1/2} \Gamma(1-s) } = \frac { c_{1,s}}{2s}.
\end{align*}
In particular, this proves formula~\eqref{sd0w3dewt95b76 498543qdetr57uy54u},
as desired.
\end{remark}
The key observation from \thref{eeBLRjcZ} is that the kernel \( {\vert x - y \vert^{-n-2s}} - {\vert x_\ast- y \vert^{-n-2s}}\) decays quicker at infinity than the kernel \(\vert x-y \vert^{-n-2s}\) which is what allows us to relax the assumption~\(u\in \mathscr L_s (\mathbb{R}^n)\). Indeed, if~$x$, $y\in \mathbb{R}^n_+ $ then \(\vert x_\ast - y \vert \geqslant \vert x-y \vert\) and \begin{align}
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}} &= \frac{n+2s} 2 \int_{\vert x - y \vert^2}^{\vert x_\ast -y \vert^2} t^{-\frac{n+2s+2}2} \dd t \nonumber \\
&\leqslant \frac{n+2s} 2 \big ( \vert x_\ast -y \vert^2-\vert x -y \vert^2 \big ) \frac 1 {\vert x - y \vert^{n+2s+2}} \nonumber \\
&= 2(n+2s) \frac{x_1y_1}{\vert x - y \vert^{n+2s+2}}. \label{LxZU6}
\end{align} Similarly, for all~$x$, $y\in \mathbb{R}^n_+$,\begin{align}
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}} &\geqslant 2(n+2s) \frac{x_1y_1}{\vert x_\ast - y \vert^{n+2s+2}}. \label{buKHzlE6}
\end{align} Hence, if \(\vert x\vert <C_0\) and if \(|x-y|>C_1\) for some \(C_0,C_1>0\) independent of \(x,y\), we find that \begin{align}
\frac{C^{-1} x_1y_1}{1+\vert y \vert^{n+2s+2}} \leqslant \frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}} \leqslant \frac{Cx_1y_1}{1+\vert y \vert^{n+2s+2}} \label{Zpwlcwex}
\end{align} which motivates the consideration of \(\mathscr A_s(\mathbb{R}^n)\).
Our final lemma in this section proves that \thref{mkG4iRYH} is well-defined, that is, the assumptions of \thref{mkG4iRYH} are enough to guarantee \eqref{OopaQ0tu} converges.
\begin{lem} \thlabel{zNNKjlJJ}
Let \(x \in \mathbb{R}^n_+\) and \(r>0\) be sufficiently small so that \(B:=B_r(x) \Subset\mathbb{R}^n_+\). If \(u \in C^{2s+\alpha}(B) \cap \mathscr A_s(\mathbb{R}^n)\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer then \((-\Delta)^su(x)\) given by \eqref{OopaQ0tu} is well-defined and there exists \(C>0\) depending on \(n\), \(s\), \(\alpha\), \(r\), and \(x_1\) such that \begin{align*}
\big \vert (-\Delta)^s u(x) \big \vert &\leqslant C \big ( \| u\|_{C^{\beta}(B)} + \Anorm{u} \big )
\end{align*} where \(\beta = \min \{ 2s+\alpha ,2\}\).
\end{lem}
We remark that the constant \(C\) in \thref{zNNKjlJJ} blows up as \(x_1\to 0^+\).
\begin{proof}[Proof of Lemma~\ref{zNNKjlJJ}]
Write \( (-\Delta)^s u(x) = I_1 +I_2 + s^{-1} c_{1,s} u(x) x_1^{-2s}\) where \begin{align*}
I_1 &:= c_{n,s}\lim_{\varepsilon \to 0^+} \int_{B\setminus B_\varepsilon(x)} \left(
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}}\right) \big(u(x)-u(y)\big)\dd y\\
\text{ and } \qquad I_2 &:= c_{n,s}\int_{\mathbb{R}^n_+\setminus B} \left(\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}}\right) \left(u(x)-u(y)\right)\dd y.
\end{align*}
\emph{For \(I_1\):} If \(y\in \mathbb{R}^n_+\) then \({\vert x - y \vert^{-n-2s}} - {\vert x_\ast- y \vert^{-n-2s}} \) is only singular at \(x\), so we may write \begin{align*}
I_1 &= c_{n,s}\lim_{\varepsilon \to 0^+} \int_{B\setminus B_\varepsilon(x)}\frac {u(x)-u(y)} {\vert x - y \vert^{n+2s}}\dd y - \int_B \frac {u(x)-u(y)} {\vert x_\ast- y \vert^{n+2s}}\dd y.
\end{align*}If \(\chi_B\) denotes the characteristic function of \(B\) then \(\bar u := u\chi_B\in C^{2s+\alpha}(B)\cap L^\infty(\mathbb{R}^n)\), so \((-\Delta)^s\bar u(x)\) is defined and \begin{align*}
(-\Delta)^s\bar u(x) &= c_{n,s}\mathrm{P.V.} \int_{\mathbb{R}^n} \frac{u(x) - u(y)\chi_B(y)}{\vert x-y\vert^{n+2s}} \dd y \\
&= c_{n,s} \mathrm{P.V.} \int_B \frac{u(x) - u(y)}{\vert x-y\vert^{n+2s}} \dd y + c_{n,s} u(x)\int_{\mathbb{R}^n\setminus B} \frac{\dd y }{\vert x-y\vert^{n+2s}}\\
&= c_{n,s}\mathrm{P.V.} \int_B \frac{u(x) - u(y)}{\vert x-y\vert^{n+2s}} \dd y + \frac 1{2s}c_{n,s} n \vert B_1 \vert r^{-2s} u(x) .
\end{align*}Hence, \begin{align*}
I_1&= (-\Delta)^s\bar u(x) -\frac 1{2s}c_{n,s} n \vert B_1 \vert r^{-2s} u(x) -c_{n,s} \int_B \frac {u(x)-u(y)} {\vert x_\ast- y \vert^{n+2s}}\dd y
\end{align*} which gives that \begin{align*}
\vert I_1 \vert &\leqslant \big \vert (-\Delta)^s\bar u(x) \big \vert +C \| u\|_{L^\infty(B)} +2\| u\|_{L^\infty(B)} \int_B \frac 1 {\vert x_\ast- y \vert^{n+2s}}\dd y \\
&\leqslant \big \vert (-\Delta)^s\bar u(x) \big \vert +C\| u\|_{L^\infty(B)}.
\end{align*}
Furthermore, if \(\alpha <2(1-s)\) then by a Taylor series expansion \begin{align*}
\big \vert (-\Delta)^s\bar u (x) \big \vert &\leqslant C \int_{\mathbb{R}^n} \frac{\vert 2\bar u(x)-\bar u(x+y) - \bar u(x-y) \vert }{\vert y \vert^{n+2s}} \dd y \\
&\leqslant C [\bar u]_{C^{2s+\alpha}(B)} \int_{B_1} \frac{\dd y }{\vert y \vert^{n-\alpha}} + C \| \bar u \|_{L^\infty(\mathbb{R}^n)} \int_{\mathbb{R}^n\setminus B_1} \frac{\dd y }{\vert y \vert^{n+2s}} \\
&\leqslant C \big ( \| \bar u\|_{C^{2s+\alpha}(B)} + \| \bar u \|_{L^\infty(\mathbb{R}^n)} \big ) \\
&=C \big ( \| u\|_{C^{2s+\alpha}(B)} + \| u \|_{L^\infty(B)} \big ) .
\end{align*} If \(\alpha \geqslant 2(1-s)\) then this computation holds with \([\bar u]_{C^{2s+\alpha}(B)} \) replaced with \(\|u\|_{C^2(B)}\).
\emph{For \(I_2\):} By~\eqref{LxZU6}, \begin{align*}
\vert I_2 \vert &\leqslant Cx_1\int_{\mathbb{R}^n_+\setminus B} \frac {y_1\big (\vert u(x) \vert + \vert u(y) \vert \big ) } {\vert x - y\vert^{n+2s+2}} \dd y \\
&\leqslant C \| u \|_{L^\infty(B)}\int_{\mathbb{R}^n_+\setminus B} \frac {y_1} {\vert x - y\vert^{n+2s+2}} \dd y + C \Anorm{u} \\
&\leqslant C \big (\| u \|_{L^\infty(B)}+ \Anorm{u} \big ) .
\end{align*}
Combining the estimates for \(I_1\) and \(I_2\) immediately gives the result.
\end{proof}
\section{Interior Harnack inequality and proof of Theorem~\ref{DYcYH}}\label{q7hKF}
The purpose of this section is to prove \thref{DYcYH}. Its proof is split into two parts: the interior weak Harnack inequality for super-solutions (\thref{MT9uf}) and interior local boundedness for sub-solutions (\thref{guDQ7}).
\subsection{The interior weak Harnack inequality}
The interior weak Harnack inequality for super-solutions is given in \thref{MT9uf} below.
\begin{prop} \thlabel{MT9uf}
Let \(M\in \mathbb{R} \), \(\rho \in(0,1)\), and \(c\in L^\infty(B_\rho(e_1))\). Suppose that \(u \in C^{2s+\alpha}(B_\rho(e_1))\cap \mathscr{A}_s(\mathbb{R}^n)\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\), and satisfies \begin{align*}
(-\Delta)^su +cu &\geqslant -M \qquad \text{in } B_\rho(e_1).
\end{align*}
Then there exists \(C_\rho>0\) depending only on \(n\), \(s\), \(\| c \|_{L^\infty(B_\rho(e_1))}\), and \(\rho\) such that \begin{align*}
\Anorm{u} &\leqslant C_\rho \left( \inf_{B_{\rho/2}(e_1)} u + M \right) .
\end{align*}
\end{prop}
The proof of \thref{MT9uf} is a simple barrier argument which takes inspiration from \cite{MR4023466}. We will begin by proving the following proposition, \thref{YLj1r}, which may be viewed as a rescaled version of \thref{MT9uf}. We will require \thref{YLj1r} in the second part of the section when we prove the interior local boundedness of sub-solutions (\thref{guDQ7}).
\begin{prop} \thlabel{YLj1r}
Let \(M\in \mathbb{R} \), \(\rho \in(0,1)\), and \(c\in L^\infty(B_\rho(e_1))\). Suppose that \(u \in C^{2s+\alpha}(B_\rho(e_1))\cap \mathscr{A}_s(\mathbb{R}^n)\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\), and satisfies \begin{align*}
(-\Delta)^su +cu &\geqslant -M \qquad \text{in } B_\rho(e_1).
\end{align*}
Then there exists \(C>0\) depending only on \(n\), \(s\), and~\(\rho^{2s}\| c \|_{L^\infty(B_\rho(e_1))}\),
such that \begin{align*}
\frac 1 {\rho^n} \int_{B_{\rho/2}(e_1)} u(x) \dd x &\leqslant C \left( \inf_{B_{\rho/2}(e_1)} u + M \rho^{2s}\right) .
\end{align*} Moreover, the constant \(C\) is of the form \begin{align}
C = C' \left(1+ \rho^{2s} \| c \|_{L^\infty(B_\rho(e_1))}\right) \label{CyJJQHrF}
\end{align} with \(C'>0\) depending only on \(n\) and \(s\).
\end{prop}
Before we give the proof of \thref{YLj1r}, we introduce some notation.
Given~$ \rho \in(0,1)$, we define
$$ T_\rho:= \left\{x_1=-\frac1{\rho}\right\}\qquad{\mbox{and}}\qquad
H_\rho:=\left\{x_1>-\frac1{\rho}\right\}.$$
We also let~$Q_\rho$ be the reflection across the hyperplane~$T_\rho$, namely
$$Q_\rho(x):=x-2(x_1+1/\rho)e_1\qquad{\mbox{for all }} x\in \mathbb{R}^n. $$
With this, we establish the following lemma.
\begin{lem} \thlabel{rvZJdN88}
Let \(\rho \in(0,1)\) and \(\zeta\) be a smooth cut-off function such that \begin{align*}
\zeta \equiv 1 \text{ in } B_{1/2}, \quad \zeta \equiv 0 \text{ in } \mathbb{R}^n \setminus B_{3/4}, \quad{\mbox{ and }}
\quad
0\leqslant \zeta \leqslant 1.
\end{align*} Define \(\varphi^{(1)} \in C^\infty_0(\mathbb{R}^n)\) by \begin{align*}
\varphi^{(1)}(x) := \zeta (x) - \zeta \big ( Q_\rho(x) \big ) \qquad \text{for all } x\in \mathbb{R}^n.
\end{align*}
Then \(\varphi^{(1)}\) is antisymmetric with respect to \(T_\rho:= \{x_1=-1/\rho\}\) and there exists \(C>0\) depending only on \(n\) and \(s\) (but not on \(\rho\)) such that \(\| (-\Delta)^s \varphi^{(1)} \|_{L^\infty(B_{3/4})} \leqslant C\).
\end{lem}
\begin{proof}
Since \(Q_\rho(x)\) is the reflection of \(x\in \mathbb{R}^n\) across \(T_\rho\), we immediately obtain that \(\varphi^{(1)} \) is antisymmetric with respect to the plane \(T_\rho\). As \(0\leqslant \zeta \circ Q_\rho \leqslant 1\) in \(\mathbb{R}^n\) and \(\zeta \circ Q_\rho=0\) in \(B_{3/4}\), from~\eqref{ZdAlT} we have that \begin{align*}
\vert (-\Delta)^s (\zeta \circ Q_\rho)(x) \vert &=C \int_{B_{3/4}(-2e_1/\rho)} \frac{(\zeta \circ Q_\rho)(y)}{\vert x - y\vert^{n+2s}} \dd y \leqslant C \qquad \text{for all }x \in B_{3/4}
\end{align*} using also
that \(\vert x- y\vert \geqslant 2(1/\rho -3/4) \geqslant 1/2\). Moreover, \begin{align*}
\| (-\Delta)^s \zeta \|_{L^\infty(B_{3/4})} &\leqslant C ( \| D^2\zeta \|_{L^\infty(B_{3/4})} + \| \zeta \|_{L^\infty(\mathbb{R}^n)} ) \leqslant C,
\end{align*} for example, see the computation on p.~9 of \cite{MR3469920}. Thus,
$$\| (-\Delta)^s \varphi^{(1)} \|_{L^\infty(B_{3/4})} \leqslant \| (-\Delta)^s \zeta\|_{L^\infty(B_{3/4})} + \| (-\Delta)^s (\zeta \circ Q_\rho)\|_{L^\infty(B_{3/4})} \leqslant C,$$
which completes the proof.
\end{proof}
Now we give the proof of \thref{YLj1r}.
\begin{proof}[Proof of \thref{YLj1r}]
Let \(\tilde u(x):=u (\rho x+e_1)\) and \(\tilde c(x) := \rho^{2s}c(\rho x+e_1)\). Observe that \((-\Delta)^s\tilde u+\tilde c \tilde u \geqslant - M\rho^{2s}\) in \(B_1\) and that \(\tilde u \) is antisymmetric with respect to \(T_\rho\).
Let \(\varphi^{(1)}\) be defined as in \thref{rvZJdN88} and suppose that \(\tau\geqslant 0\) is the largest possible value such that~\(\tilde u \geqslant \tau \varphi^{(1)} \) in the half space~\(H_\rho\). Since \(\varphi^{(1)} =1\) in \(B_{1/2}\), we immediately obtain that \begin{align}
\tau \leqslant \inf_{B_{1/2}} \tilde u = \inf_{B_{\rho/2}(e_1)} u. \label{mzMEg}
\end{align}
Moreover, by continuity, there exists \(a \in \overline{B_{3/4}}\) such that \(\tilde u(a)=\tau\varphi^{(1)} (a)\). On one hand, using \thref{rvZJdN88}, we have that \begin{align}
(-\Delta)^s(\tilde u-\tau \varphi^{(1)})(a)+\tilde c (a) (\tilde u-\tau \varphi^{(1)})(a) &\geqslant -M\rho^{2s} - \tau \big( C+ \|\tilde c\|_{L^\infty(B_1)}\big) \nonumber \\
&\geqslant -M\rho^{2s} - C\tau \big( 1+ \rho^{2s}\|c\|_{L^\infty(B_\rho(e_1))}\big) .\label{aDLwG3pX}
\end{align} On the other hand, since \(\tilde u-\tau\varphi^{(1)}\) is antisymmetric with respect to \(T_\rho\), \(\tilde u - \tau \varphi^{(1)} \geqslant 0\) in \(H_\rho\), and~\((\tilde u-\tau \varphi^{(1)})(a)=0\), it follows that \begin{align}
&(-\Delta)^s (\tilde u-\tau \varphi^{(1)})(a)+c_\rho(a)(\tilde u-\tau \varphi^{(1)})(a) \nonumber \\
&\leqslant - C \int_{B_{1/2}} \left(
\frac 1 {\vert a - y \vert^{n+2s}} - \frac 1 {\vert Q_\rho(a) - y \vert^{n+2s}} \right) \big(
\tilde u(y)-\tau \varphi^{(1)}(y)\big) \dd y. \label{wPcy8znV}
\end{align} For all \(y \in B_{1/2}\), we have that \(\vert a - y \vert \leqslant C\) and \(\vert Q_\rho(a) - y \vert \geqslant C\) (the assumption \(\rho <1\) allows to choose this \(C\) independent of \(\rho\)), so \begin{align}
(-\Delta)^s (\tilde u-\tau \varphi^{(1)})(a)+c_\rho(a)(\tilde u-\tau \varphi^{(1)})(a) &\leqslant - C \int_{B_{1/2}} \big(\tilde u(y)-\tau \varphi^{(1)}(y)\big) \dd y \nonumber \\
&\leqslant -C \left( \frac 1 {\rho^n} \int_{B_{\rho/2}(e_1)} u(y) \dd y - \tau \right). \label{ETn5BCO5}
\end{align} Rearranging \eqref{aDLwG3pX} and \eqref{ETn5BCO5} then using \eqref{mzMEg}, we obtain \begin{align*}
\frac 1 {\rho^n} \int_{B_{\rho/2}(e_1)} u(y) \dd y &\leqslant C \Big( \tau \big( 1+ \rho^{2s}\|c\|_{L^\infty(B_\rho(e_1))}\big) + M\rho^{2s} \Big) \\
&\leqslant C \big( 1+ \rho^{2s}\|c\|_{L^\infty(B_\rho(e_1))}\big) \left(
\inf_{B_{\rho/2}(e_1)} u + M \rho^{2s}\right)
\end{align*} as required.
\end{proof}
A simple adaptation of the proof of \thref{YLj1r} leads to the proof of \thref{MT9uf} which we now give.
\begin{proof}[Proof of \thref{MT9uf}]
Follow the proof of \thref{YLj1r} but instead of \eqref{wPcy8znV}, we write \begin{align*}
&(-\Delta)^s (\tilde u-\tau \varphi^{(1)})(a)+c_\rho(a)(\tilde u-\tau \varphi^{(1)})(a)\\
&\qquad= - C \int_{H_\rho} \left(
\frac 1 {\vert a - y \vert^{n+2s}} - \frac 1 {\vert Q_\rho(a) - y \vert^{n+2s}} \right) \big(
\tilde u(y)-\tau \varphi^{(1)}(y)\big) \dd y.
\end{align*} Then, for all \(x,y\in H_\rho\), \begin{align}
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert Q_\rho(x) - y \vert^{n+2s}} &= \frac{n+2s} 2 \int_{\vert x-y\vert^2}^{\vert Q_\rho(x)-y\vert^2} t^{- \frac{n+2s+2}2} \dd t \label{gvKkSL1N}\\
&\geqslant C \Big (\vert Q_\rho(x)-y\vert^2-\vert x-y\vert^2 \Big ) \vert Q_\rho(x) - y \vert^{- (n+2s+2)} \nonumber\\
&= C \frac{(x_1+1/\rho)(y_1+1/\rho)}{\vert Q_\rho(x) - y \vert^{n+2s+2}}, \nonumber
\end{align} so using that \(\tilde u - \tau \varphi^{(1)} \geqslant 0\) in \(H_\rho\), we see that \begin{align*}
(-\Delta)^s (\tilde u-\tau \varphi^{(1)})(a)+c_\rho(a)(\tilde u-\tau \varphi^{(1)})(a)
&\leqslant - C \int_{H_\rho} \frac{(y_1+1/\rho)(\tilde u(y)-\tau \varphi^{(1)}(y)) }{\vert Q_\rho(a) - y \vert^{n+2s+2}}\dd y \\
&\leqslant - C_\rho \left( \int_{H_\rho} \frac{(y_1+1/\rho)\tilde u(y) }{\vert Q_\rho(a) - y \vert^{n+2s+2}}\dd y + \tau \right) .
\end{align*} Making the change of variables \(z=\rho y +e_1\), we have that \begin{align*}
\int_{H_\rho} \frac{(y_1+1/\rho)\tilde u(y) }{\vert Q_\rho(a) - y \vert^{n+2s+2}}\dd y &= \rho^{-n-1} \int_{\mathbb{R}^n_+} \frac{z_1 u(z) }{\vert Q_\rho(a) - z/\rho +e_1/\rho \vert^{n+2s+2}}\dd z.
\end{align*} Thus, since \( \vert Q_\rho(a) - z/\rho +e_1/\rho \vert^{n+2s+2} \leqslant C_\rho (1+ \vert z \vert^{n+2s+2} ) \), we conclude that
\begin{align*}
\int_{H_\rho} \frac{(y_1+1/\rho)\tilde u(y) }{\vert Q_\rho(a) - y \vert^{n+2s+2}}\dd y &\geqslant C_\rho \int_{\mathbb{R}^n_+} \frac{z_1 u(z) }{1+ \vert z \vert^{n+2s+2}} \dd z=C_\rho \Anorm{u}.
\end{align*} As a consequence,
\begin{align}
(-\Delta)^s (\tilde u-\tau \varphi^{(1)})(a)+\tilde c(a)(\tilde u-\tau \varphi^{(1)})(a) &\leqslant- C_\rho \Anorm{u}+C_\rho \tau . \label{Pzf2a}
\end{align} Rearranging~\eqref{aDLwG3pX} and~\eqref{Pzf2a} then using~\eqref{mzMEg} gives \begin{align*}
\Anorm{u} &\leqslant C_\rho ( \tau + M ) \leqslant C_\rho \left( \inf_{B_{\rho/2}(e_1)} u + M \right),
\end{align*}
as desired.
\end{proof}
\subsection{Interior local boundedness}
The second part of the proof of \thref{DYcYH} is the interior local boundedness of sub-solutions given in \thref{guDQ7} below.
\begin{prop} \thlabel{guDQ7}
Let \(M \geqslant 0\), \(\rho\in(0,1/2)\), and \(c\in L^\infty(B_\rho(e_1))\). Suppose that \(u \in C^{2s+\alpha}(B_\rho(e_1))\cap \mathscr A_s(\mathbb{R}^n)\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer, and \(u\) satisfies \begin{equation}\label{sow85bv984dert57nb5}
(-\Delta)^su +cu \leqslant M \qquad \text{in } B_\rho(e_1).
\end{equation}
Then there exists \(C_\rho>0\) depending only on \(n\), \(s\), \(\| c \|_{L^\infty(B_\rho(e_1))}\), and \(\rho\) such that \begin{align*}
\sup_{B_{\rho/2}(e_1)} u &\leqslant C_\rho ( \Anorm{u} +M ) .
\end{align*}
\end{prop}
The proof of \thref{guDQ7} uses similar ideas to \cite[Theorem~11.1]{MR2494809} and \cite[Theorem~5.1]{MR2831115}. Before we prove \thref{guDQ7}, we need the following lemma.
\begin{lem} \thlabel{ltKO2}
Let \(R\in(0,1)\) and \(a\in \overline{ B_2^+}\). Then there exists \(C>0\) depending only on \(n\) and \(s\) such that, for all \(x\in B_{R/2}^+(a)\) and \(y \in \mathbb{R}^n_+ \setminus B_R^+(a)\), \begin{align*}
\frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast - y \vert^{n+2s}} &\leqslant C R^{-n-2s-2} \frac{x_1y_1}{1+\vert y \vert^{n+2s+2}}.
\end{align*}
\end{lem}
\begin{proof} Let \(\tilde x := (x-a)/R\in B_{1/2}\) and \(\tilde y = (y-a)/ R\in \mathbb{R}^n\setminus B_1\). Clearly,
we have that~\(\vert \tilde y - \tilde x \vert \geqslant 1/2\). Moreover, since \(\vert \tilde x \vert <1/2 < 1/(2 \vert \tilde y \vert)\), we have that \(\vert \tilde y - \tilde x \vert \geqslant \vert \tilde y \vert -\vert \tilde x \vert \geqslant (1/2) \vert \tilde y \vert\). Hence, \begin{align*}
\vert \tilde y - \tilde x \vert^{n+2s+2} \geqslant \frac 1 {2^{n+2s+2}} \max \big \{ 1 , \vert \tilde y \vert^{n+2s+2} \big \} \geqslant C \big ( 1+ \vert \tilde y \vert^{n+2s+2} \big )
\end{align*} for some \(C\) depending only on \(n\) and \(s\). It follows that
\begin{align}
\vert x -y \vert^{n+2s+2} = R^{n+2s+2} \vert \tilde x - \tilde y \vert^{n+2s+2}
\geqslant C R^{n+2s+2} \big ( 1 + R^{-n-2s-2} \vert y-a \vert^{n+2s+2} \big ). \label{TO9DRBuX}
\end{align}
Finally, we claim that there exists \(C\) independent of \(R\) such that \begin{align}
\vert y - a\vert \geqslant CR \vert y \vert \qquad \text{for all } y\in \mathbb{R}^n \setminus B_R(a) . \label{tdNyAxBN}
\end{align} Indeed, if \(y \in \mathbb{R}^n \setminus B_4\) then \begin{align*}
\vert y - a\vert \geqslant \vert y \vert -2 \geqslant \frac12 \vert y \vert>\frac R 2 \vert y \vert,
\end{align*} and if \(y \in (\mathbb{R}^n \setminus B_R(a) ) \cap B_4\) then \begin{align*}
\vert y -a\vert \geqslant R > \frac R 4 \vert y \vert
\end{align*} which proves \eqref{tdNyAxBN}.
Thus, \eqref{TO9DRBuX} and \eqref{tdNyAxBN} give that, for all \( x \in B_{R/2}(a)\) and \( y \in \mathbb{R}^n_+ \setminus B_R(a)\), \begin{align*}
\vert x -y \vert^{n+2s+2} &\geqslant C R^{n+2s+2} \big ( 1 + \vert y\vert^{n+2s+2} \big ) .
\end{align*} Then the result follows directly from~\eqref{LxZU6}.
\end{proof}
With this preliminary work, we now focus on the proof of \thref{guDQ7}.
\begin{proof}[Proof of \thref{guDQ7}]
We first observe that,
dividing through by~\( \Anorm{u} +M\), we may
assume that \((-\Delta)^su +cu\leqslant 1\) in \(B_\rho(e_1)\) and~\(\Anorm{u} \leqslant 1\).
We also point out that if~$u\le0$ in~$B_\rho(e_1)$, then the claim in \thref{guDQ7}
is obviously true. Therefore, we can suppose that
\begin{equation}\label{upos44567890}
\{u>0\}\cap B_\rho(e_1)\ne \varnothing.\end{equation}
Thus, we let~\(\tau\geqslant 0 \) be the smallest possible value such that \begin{align*}
u(x) &\leqslant \tau (\rho-\vert x-e_1 \vert )^{-n-2} \qquad \text{for all } x\in B_\rho(e_1).
\end{align*} Such a \(\tau\) exists in light of~\eqref{upos44567890}.
To complete the proof, we will show that \begin{align}
\tau &\leqslant C_\rho \label{svLyO}
\end{align}with \(C_\rho\) depending only on \(n\), \(s\), \(\| c\|_{L^\infty(B_\rho(e_1))}\),
and \(\rho\) (but independent of \(u\)). Since \(u\) is uniformly
continuous in~$B_\rho(e_1)$, there exists \(a\in B_\rho(e_1)\) such that \(u(a) = \tau (\rho-\vert a-e_1\vert)^{-n-2}\).
Notice that~$u(a)>0$ and~\(\tau >0\).
Let also~\(d=:\rho-\vert a-e_1\vert\) so that \begin{align}
u(a) = \tau d^{-n-2} ,\label{CEnkR}
\end{align} and let \begin{align*}
U:= \bigg \{ y \in B_\rho (e_1) \text{ s.t. } u(y) > \frac{u(a)} 2 \bigg \} .
\end{align*} Since \(\Anorm{u}\leqslant 1\), if \(r\in(0,d)\) then \begin{align*}
C_\rho &\geqslant \int_{B_\rho (e_1)} |u(x)| \dd x
\ge \int_{ U\cap B_r (a)} u(x) \dd x \geqslant \frac{u(a)}2 \cdot \vert U \cap B_r (a) \vert .
\end{align*} Thus, from~\eqref{CEnkR}, it follows that \begin{align}
\vert U \cap B_r (a) \vert \leqslant \frac{C_\rho d^{n+2}}\tau \qquad \text{ for all } r\in(0,d). \label{W5Ar0}
\end{align}
Next, we make the following claim.
\begin{claim}
There exists~$\theta_0\in(0,1)$ depending only on \(n\), \(s\), \(\| c\|_{L^\infty(B_\rho(e_1))}\), and \(\rho\)
such that if~\(\theta\in(0,\theta_0]\) there exists~\(C>0\) depending only on \(n\), \(s\), \(\| c\|_{L^\infty(B_\rho(e_1))}\), \(\rho\), and~$\theta$ such that \begin{align*}
\big \vert B_{\theta d /8} (a) \setminus U \big \vert \leqslant \frac 1 4 \vert B_{\theta d /8} \vert +C \frac{d^n }{\tau} .
\end{align*} In particular, neither \(\theta\) nor \(C\) depend on \(\tau\), \(u\), or \(a\).
\end{claim}
We will withhold the proof of the claim until the end. Assuming the claim is true, we complete the proof of the \thref{guDQ7} as follows. By~\eqref{W5Ar0} (used here with~$r:={\theta_0 d}/8$)
and the claim, we have that \begin{align*}
\frac{C_\rho d^{n+2}}\tau \geqslant \big \vert B_{\theta_0 d /8} \big \vert -\big \vert B_{\theta_0 d /8} (a) \setminus U \big \vert
\geqslant \frac 3 4 \vert B_{\theta_0 d/8} \vert -C \frac{d^n }{\tau} .
\end{align*} Rearranging gives that \( \tau \leqslant C( d^2 +1 ) \leqslant C\), which proves \eqref{svLyO}.
Accordingly, to complete the proof of \thref{guDQ7}, it remains to establish
the Claim. For this, let \(\theta\in(0,1)\) be a small constant to be chosen later. We will prove the claim by applying \thref{YLj1r} to an appropriate auxiliary function. For \(x\in B_{\theta d/2} (a)\), we have that \(\vert x -e_1 \vert \leqslant \vert a -e_1 \vert + \theta d/2=\rho-(1-\theta/2)d \), so, using~\eqref{CEnkR}, \begin{align}
u(x) \leqslant \tau \bigg (1-\frac \theta2\bigg )^{-n-2} d^{-n-2} = \bigg (1-\frac\theta 2\bigg )^{-n-2} u(a) \qquad \text{for all } x\in B_{\theta d/2}(a). \label{sS0kO}
\end{align}
Let \(\zeta \) be a smooth, antisymmetric function such that~$\zeta \equiv 1$ in~$\{x_1 > 1/2 \}$
and~$0\leqslant \zeta \leqslant 1$ in~$\mathbb{R}^n_+$,
and consider the antisymmetric function \begin{align*}
v(x) &:= \left(1-\frac\theta 2\right)^{-n-2} u(a) \zeta(x) -u (x) \qquad \text{for all } x\in \mathbb{R}^n.
\end{align*} Since \(\zeta \equiv 1\) in \(\{x_1 > 1/2 \} \supset B_{\theta d/2}(a)\) and \(0\leqslant \zeta \leqslant 1\) in \(\mathbb{R}^n_+\), it follows easily from~\eqref{OopaQ0tu} that \((-\Delta)^s \zeta \geqslant 0\) in \(B_{\theta d/2}(a)\). Hence, in \(B_{\theta d/2}(a)\), \begin{align*}
(-\Delta)^s v +cv &\geqslant -(-\Delta)^s u-cu+ c\bigg (1-\frac\theta 2\bigg )^{-n-2} u(a) \\
&\geqslant -1 -C\| c^-\|_{L^\infty(B_\rho(e_1))}\bigg (1-\frac\theta 2\bigg )^{-n-2} u(a) .
\end{align*} Taking \(\theta\) sufficiently small, we obtain \begin{align}
(-\Delta)^s v +cv &\geqslant -C (1 + u(a) ) \qquad \text{in } B_{\theta d/2}(a). \label{7fUyX}
\end{align}
The function \(v\) is almost the auxiliary function to which we would like to apply \thref{YLj1r}; however, \thref{YLj1r} requires \(v \geqslant 0\) in \(\mathbb{R}^n_+\) but we only have \(v \geqslant 0\) in \(B_{\theta d/2}(a)\) due to \eqref{sS0kO}. To resolve this issue let us instead consider the function \(w\) such that \(w (x) = v^+(x)\) for all \(x\in \mathbb{R}^n_+\) and~\(w(x) = -w(x_\ast)\) for all \(x\in \overline{\mathbb{R}^n_-}\).
We point out that~$w$ coincides with~$v$ in~$B_{\theta d/2}(a)$, thanks to~\eqref{sS0kO},
and therefore it is as regular as~$v$ in~$B_{\theta d/2}(a)$, which allows us to write the fractional Laplacian of~$w$ in~$B_{\theta d/2}(a)$ in a pointwise sense.
Also we observe that we have~\(w\geqslant0\) in \(\mathbb{R}^n_+\) but we no longer have a lower bound for \((-\Delta)^sw+cw\). To obtain this, observe that for all \(x\in \mathbb{R}^n_+\), \begin{align*}
(w-v)(x)
&= \begin{cases}
0, &\text{for all } x\in \{v>0\} \cap \mathbb{R}^n_+ \\
u(x)- (1-\theta /2 )^{-n-2} u(a) \zeta(x), &\text{for all } x\in \{v \leqslant 0\}\cap \mathbb{R}^n_+.
\end{cases}
\end{align*} In particular, \(w-v\leqslant |u|\) in \(\mathbb{R}^n_+\). It follows that for all \(x\in B_{\theta d/2}(a)\),
\begin{align*}
(-\Delta)^s (w-v)(x) &\geqslant -C \int_{\mathbb{R}^n_+ \setminus B_{\theta d/2}(a) } \left( \frac 1 {\vert x-y\vert^{n+2s}} - \frac 1 {\vert x_\ast - y \vert^{n+2s}} \right)| u(y)|\dd y.
\end{align*} Moreover, by \thref{ltKO2}, for all~$x\in B_{\theta d/4}(a)$,\begin{align}
(-\Delta)^s (w-v)(x) &\geqslant -C (\theta d)^{-n-2s-2} \Anorm{u}
\geqslant -C (\theta d)^{-n-2s-2} . \label{Fy8oddTC}
\end{align}Thus, by~\eqref{7fUyX} and~\eqref{Fy8oddTC}, for all \(x\in B_{\theta d/4}(a)\), we have that \begin{align}
(-\Delta)^sw(x)+c(x)w(x) &= (-\Delta )^s v (x)+c(x) v(x) + (-\Delta)^s (w-v)(x)
\nonumber \\
&\geqslant -C \big (1 + u(a) +(\theta d)^{-n-2s-2} \big ) \nonumber \\
&\geqslant -C \big ( (\theta d)^{-n-2s-2}+u(a) \big ) \label{c1kcbe0C}
\end{align} using that \(\theta d<1\).
Next let us consider the rescaled and translated functions \(\tilde w(x) := w(a_1 x+(0,a'))\) and \(\tilde c (x) := a_1^{2s} c(a_1 x+(0,a'))\) (recall that \(a' = (a_2,\dots, a_n) \in \mathbb{R}^{n-1}\)). By~\eqref{c1kcbe0C} we have that
\begin{align*}
(-\Delta)^s \tilde w+\tilde c\, \tilde w \geqslant -Ca_1^{2s} \big ((\theta d)^{-n-2s-2} + u(a) \big ) \qquad \text{in } B_{\theta d /(4a_1)}(e_1 ).
\end{align*} On one hand, by \thref{YLj1r}, we obtain \begin{align*}
\left(\frac{\theta d } 8 \right)^{-n} \int_{B_{\theta d /8}(a )} w( x) \dd x
&= \left(\frac{\theta d }{8a_1} \right)^{-n} \int_{B_{\theta d/(8a_1)}(e_1 )} \tilde w ( x) \dd x \\
&\leqslant C \Big( \tilde w(e_1) + (\theta d)^{-n-2} +u(a) (\theta d )^{2s} \Big)\\
&= C \left( \left(1-\frac\theta 2\right)^{-n-2} -1 \right)u(a) +C (\theta d)^{-n-2} +u(a) (\theta d )^{2s} .
\end{align*}We note explicitly that, by~\eqref{CyJJQHrF}, the constant in the above line is given by \begin{align*}
C= C'\big ( 1 +(\theta d)^{2s} \| c\|_{L^\infty(B_{\rho}(e_1))} \big ),
\end{align*} so using that \(\theta d<1\) it may be chosen to depend only on \(n\), \(s\), and \( \| c\|_{L^\infty(B_{\rho}(e_1))}\). On the other hand, \begin{align*}
B_{\theta d /8} (a) \setminus U &\subseteq\left\{ w \geqslant \left( \left( 1- \frac \theta 2\right)^{-n-2} -\frac12 \right) u(a) \right\} \cap B_{\theta d /8} (a) ,
\end{align*}so we have that \begin{align*}
(\theta d )^{-n} \int_{B_{\theta d/8}(a)} w(x) \dd x &\geqslant (\theta d)^{-n} \bigg ( \bigg ( 1- \frac \theta 2\bigg )^{-n-2} -\frac12 \bigg ) u(a) \cdot \big \vert B_{\theta d /8} (a) \setminus U\big \vert \\
&\geqslant C (\theta d)^{-n} u(a) \cdot \big \vert B_{\theta d /8} (a) \setminus U\big \vert
\end{align*} for \(\theta\) sufficiently small. Thus, \begin{align*}
\big \vert B_{\theta d /8} (a) \setminus U \big \vert
&\leqslant C (\theta d)^n \bigg ( \bigg (1-\frac\theta 2\bigg )^{-n-2} -1 \bigg ) +C(u(a))^{-1} (\theta d)^{-2} + (\theta d )^{n+2s} \\
&\leqslant C (\theta d)^n \bigg ( \bigg (1-\frac\theta 2\bigg )^{-n-2} -1 +\theta^{2s}\bigg ) +C \frac{\theta^{-2} d^n }{\tau}
\end{align*} using~\eqref{CEnkR} and that \(d^{n+2s}<d^n\) since \(d<1\). At this point we may choose \(\theta\) sufficiently small such that \begin{align*}
(\theta d)^n \bigg ( \bigg (1-\frac\theta 2\bigg )^{-n-2} -1 +\theta^{2s}\bigg ) \leqslant \frac 1 4 \vert B_{\theta d/8} \vert .
\end{align*} This proves the claim, and thus completes the proof of
Proposition~\ref{guDQ7}.
\end{proof}
\section{Boundary Harnack inequality and proof of \thref{C35ZH}} \label{TZei6Wd4}
In this section, we give the proof of \thref{C35ZH}. Analogous to the proof of \thref{DYcYH}, the proof
of \thref{C35ZH} is divided into the boundary Harnack inequality for super-solutions (\thref{SwDzJu9i}) and
the boundary local boundedness for sub-solutions (\thref{EP5Elxbz}). Together these two results immediately
give \thref{C35ZH}.
\subsection{The boundary weak Harnack inequality}
Our next result is the antisymmetric boundary weak Harnack inequality.
\begin{prop} \thlabel{SwDzJu9i}
Let \(M\in \mathbb{R}\), \(\rho>0\), and \(c\in L^\infty(B_\rho^+)\). Suppose that \(u\in C^{2s+\alpha}(B_\rho)\cap \mathscr{A}_s(\mathbb{R}^n)\) for some \(\alpha > 0\) with \(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\) and satisfies \begin{align*}
(-\Delta)^s u +cu \geqslant -Mx_1 \qquad \text{in } B_\rho^+.
\end{align*}
Then there exists \(C_\rho>0\) depending only on \(n\), \(s\), \(\| c \|_{L^\infty(B_\rho^+)}\), and \(\rho\) such that \begin{align*}
\Anorm{u} &\leqslant C_\rho \left( \inf_{B_{\rho/2}^+} \frac{u(x)}{x_1} +M \right).
\end{align*}
\end{prop}
As with the interior counter-part of \thref{SwDzJu9i}, that is
\thref{MT9uf}, we will actually prove the following rescaled version of \thref{SwDzJu9i},
namely \thref{g9foAd2c}. This version is essential to the proof of the boundary local boundedness for sub-solutions. Once \thref{g9foAd2c} has been proven, \thref{SwDzJu9i} follows easily with some minor adjustments.
\begin{prop} \thlabel{g9foAd2c}
Let \(M\in \mathbb{R}\), \(\rho >0\), and \(c\in L^\infty(B_\rho^+)\). Suppose that \(u\in C^{2s+\alpha}(B_\rho)\cap \mathscr{A}_s(\mathbb{R}^n)\) for some \(\alpha >0\) with~\(2s+\alpha\) not an integer, \(u\) is non-negative in \(\mathbb{R}^n_+\) and satisfies \begin{align*}
(-\Delta)^s u +cu \geqslant -Mx_1\qquad \text{in } B_\rho^+.
\end{align*}
Then there exists \(C>0\) depending only on \(n\), \(s\), and \(\rho^{2s}\| c \|_{L^\infty(B_\rho^+)}\) such that \begin{align*}
\frac 1 {\rho^{n+2}} \int_{B_{\rho/2}^+} y_1 u(y) \dd y &\leqslant C \left( \inf_{B_{\rho/2}^+} \frac{u(x)}{x_1} + M \rho^{2s}\right).
\end{align*} Moreover, the constant \(C\) is of the form \begin{align*}
C=C' \big(1+\rho^{2s} \| c \|_{L^\infty(B_\rho^+)} \big)
\end{align*} with \(C'\) depending only on \(n\) and \(s\).
\end{prop}
Before we prove \thref{g9foAd2c}, we require some lemmata.
\begin{lem} \thlabel{tZVUcYJl}
Let \(M\geqslant0\), \(k \geqslant 0\), and suppose that \(u\in C^{2s+\alpha}(B_1)\cap \mathscr{A}_s(\mathbb{R}^n)\) for some~\(\alpha>0\) with~\(2s+\alpha\) not an integer. \begin{enumerate}[(i)]
\item If \(u\) satisfies \begin{align*}
(-\Delta)^s u +k u \geqslant -Mx_1\qquad \text{in } B_1^+
\end{align*} then for all \(\varepsilon>0\) sufficiently small there exists \(u_\varepsilon \in C^\infty_0(\mathbb{R}^n)\)
antisymmetric and such that \begin{equation}
(-\Delta)^s u_\varepsilon +k u_\varepsilon \geqslant -(M+\varepsilon) x_1\qquad \text{in } B_{7/8}^+. \label{EuoJ3En6}
\end{equation}
\item If \(u\) satisfies \begin{align*}
(-\Delta)^s u +k u \leqslant Mx_1\qquad \text{in } B_1^+
\end{align*} then for all \(\varepsilon>0\) sufficiently small there exists \(u_\varepsilon \in C^\infty_0(\mathbb{R}^n)\)
antisymmetric and such that \begin{align*}
(-\Delta)^s u_\varepsilon +k u_\varepsilon \leqslant (M+\varepsilon) x_1\qquad \text{in } B_{7/8}^+.
\end{align*}
\end{enumerate} In both cases the sequence \(\{ u_\varepsilon\} \) converges to \(u\) uniformly in \(B_{7/8}\).
Additionally, if \(u \) is non-negative in \(\mathbb{R}^n_+\) then \(u_\varepsilon\) is also non-negative in \(\mathbb{R}^n_+\).
\end{lem}
For the usual fractional Laplacian, \thref{tZVUcYJl} follows immediately by taking a mollification of \(u\) and in principle this is also the idea here. However, there are a couple of technicalities that need to be addressed. The first is that here the fractional Laplacian is defined according to \thref{mkG4iRYH} and it remains to be verified that this fractional Laplacian commutes with the convolution operation as the usual one does. As a matter of fact, \thref{mkG4iRYH} does not lend itself well to the Fourier transform which makes it difficult to prove such a property. We overcome this issue by first multiplying \(u\) by an appropriate cut-off function which allows us to reduce to the case \((-\Delta)^s\) as given by the usual definition.
The second issue is that, directly using the properties of mollifiers, we can only expect to
control~\(u_\varepsilon\) in some~\(U \Subset B_1^+\) and not up to~\(\{x_1=0\}\). We are able to relax this thanks to the antisymmetry of \(u\).
\begin{proof}[Proof of Lemma~\ref{tZVUcYJl}]
Fix \(\varepsilon>0\). Let \(R>1\) and let~\(\zeta\) be a smooth radial cut-off function such that $$
\zeta \equiv 1 \text{ in } B_R, \quad \zeta \equiv 0 \text{ in } \mathbb{R}^n \setminus B_{2R},
\quad{\mbox{and}}\quad 0\leqslant \zeta \leqslant 1 .
$$
Let also \(\bar u := u \zeta\).
Now let us define a function \(f:B_1\to\mathbb{R}\) as follows: let \(f(x)=(-\Delta)^s u(x) + k u(x)\) for all \(x\in B_1^+\), \(f(x) = 0\) for all \(x\in B_1\cap \{x_1=0\}\), and \(f(x)=-f(x_\ast)\). We also define \(\bar f:B_1\to \mathbb{R}\) analogously with \(u\) replaced with \(\bar u\). By definition, both \(f\) and \(\bar f\) are antisymmetric\footnote{Note that the definition of antisymmetric requires the domain of \(f\) and \(\bar f\) to be \(\mathbb{R}^n\). For simplicity, we will still refer to \(f\) and \(\bar f\) as antisymmetric in this context since this technicality has no effect on the proof.}, but note carefully that, \emph{a priori}, there is no reason to expect any regularity of \(f\) and \(\bar f\) across \(\{x_1=0\}\) (we will in fact find that \(\bar f \in C^\alpha(B_1)\)).
We claim that for \(R\) large enough (depending on \(\varepsilon\)), \begin{align}
\vert \bar f(x) - f(x) \vert \leqslant \varepsilon x_1\qquad \text{for all } x\in B_1^+. \label{8LncZIti}
\end{align} Indeed, if \(x\in B_1^+\) then \begin{align*}
\bar f(x) - f(x) =(-\Delta)^s(\bar u - u ) (x) &= C \int_{\mathbb{R}^n_+\setminus B_R} \bigg ( \frac 1 {\vert x - y \vert^{n+2s}} - \frac 1 {\vert x_\ast- y \vert^{n+2s}} \bigg ) ( u -\bar u ) (y) \dd y .
\end{align*} {F}rom~\eqref{LxZU6}, it follows that \begin{align*}
\vert \bar f(x) - f(x) \vert \leqslant Cx_1 \int_{\mathbb{R}^n_+\setminus B_R} \frac{y_1\vert u(y) -\bar u(y)\vert}{\vert x - y \vert^{n+2s+2}} \dd y \leqslant Cx_1 \int_{\mathbb{R}^n_+\setminus B_R} \frac{y_1\vert u(y) \vert }{1+ \vert y \vert^{n+2s+2}} \dd y .
\end{align*} Since \(u\in \mathscr A_s(\mathbb{R}^n)\), taking \(R\) large we obtain~\eqref{8LncZIti}.
Next, consider the standard mollifier \(\eta(x) := C_0 \chi_{B_1}(x) e^{-\frac 1 {1-\vert x\vert ^2}}\) with \(C_0>0\) such that \(\int_{\mathbb{R}^n} \eta(x) \dd x =1\) and let~\(\eta_\varepsilon(x) := \varepsilon^{-n} \eta (x/\varepsilon)\). Also, let \(u_\varepsilon := \bar u \ast \eta_\varepsilon\) and \(f_\varepsilon := \bar f \ast \eta_\varepsilon\).
Notice that~$u_\epsilon\in C^\infty_0(\mathbb{R}^n)$ and it is antisymmetric. Additionally, we show that~\eqref{EuoJ3En6}
holds true in case~$(i)$ (case~$(ii)$ being analogous).
To this end, we observe that, since \(\bar u\) has compact support, we have that~\(\bar u \in \mathscr L_s(\mathbb{R}^n)\), so by \thref{eeBLRjcZ}, \((-\Delta)^s \bar u\) can be understood in the usual sense in \(B_1\), that is, by~\eqref{ZdAlT}. Moreover,
by~\cite[Propositions~2.4-2.6]{MR2270163}, we have that~\((-\Delta)^s\bar u\in C^\alpha(B_1)\) which gives that \(\bar f \in C^\alpha(B_1)\) and \begin{align*}
(-\Delta)^s\bar u +k \bar u &= \bar f \qquad \text{in }B_1.
\end{align*} In particular, we may use standard properties of mollifiers to immediately obtain \begin{align*}
(-\Delta)^s u_\varepsilon +k u_\varepsilon &= f_\varepsilon \qquad \text{in }B_{7/8}.
\end{align*}
Also, since \(\bar f\) is antisymmetric, it follows that \begin{align*}
f_\varepsilon(x) &= \int_{\mathbb{R}^n } \bar f (y) \eta_\varepsilon (x-y) \dd y = \int_{\mathbb{R}^n_+ } \bar f (y)\big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y.
\end{align*} Observe that, since \(\eta \) is monotone decreasing in the radial direction and \(\vert x- y \vert \leqslant \vert x_\ast - y \vert\) for all~\(x,y\in \mathbb{R}^n_+\), \begin{align}
\eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \geqslant 0 \qquad \text{for all } x,y\in \mathbb{R}^n_+. \label{DPmhap5t}
\end{align} Moreover, by~\eqref{8LncZIti}, we see that~\(\bar f (x) \geqslant -(M+\varepsilon)x_1\) for all \(x\in B_1^+\), so if \(x\in B_{7/8}^+\) and \(\varepsilon>0\) is sufficiently small (independent of \(x\)) then it follows that \begin{equation}\label{fjrehgeruig009887}
f_\varepsilon(x) =\int_{B_\varepsilon^+(x)} \bar f (y)\big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y \geqslant -(M+\varepsilon)\int_{B_\varepsilon^+(x)} y_1 \big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y.
\end{equation}
Next, we claim that \begin{align}
\int_{B_\varepsilon^+(x)} y_1 \big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y &\leqslant C \varepsilon x_1 \label{Hp0lBCzB}
\end{align} with \(C>0\) depending only on \(n\) and~\(s\).
To prove~\eqref{Hp0lBCzB}, we will consider the cases~\(x_1 \geqslant 2 \varepsilon\) and~\(x_1 < 2 \varepsilon\) separately.
Suppose that \(x_1 \geqslant 2 \varepsilon\). In this case, \begin{align*}
\int_{B_\varepsilon^+(x) } y_1\big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y &= \int_{B_\varepsilon(x) } y_1 \eta_\varepsilon (x-y)\dd y \leqslant C_0 \varepsilon^{-n} \int_{B_\varepsilon(x) } y_1\dd y.
\end{align*} Using that \(B_\varepsilon^+(x) \subset ( x_1-\varepsilon,x_1+\varepsilon ) \times Q_\varepsilon (x')\) where \( Q_\varepsilon (x')\subset \mathbb{R}^{n-1}\) is a square centred at \(x'\) with side lengths \(2\varepsilon\), we estimate \begin{align*}
\int_{B_\varepsilon(x) } y_1\dd y &\leqslant C\varepsilon^{n-1} \int_{x_1-\varepsilon }^{x_1+\varepsilon} y_1 \dd y_1 = C \varepsilon^n x_1 ,
\end{align*} which gives~\eqref{Hp0lBCzB}.
Now suppose that \(x_1<2\varepsilon\). Using that \(\vert \nabla \eta_\varepsilon (x) \vert \leqslant C \varepsilon^{-n-1}\), the Mean Value Theorem gives that \begin{align*}
\eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \leqslant C \varepsilon^{-n-1} \vert x-x_\ast \vert = 2C \varepsilon^{-n-1}x_1.
\end{align*} It follows that \begin{align*}
\int_{B_\varepsilon^+(x) } y_1\big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y & \leqslant C\varepsilon^{-n-1} x_1\int_{B_\varepsilon^+(x) } y_1 \dd y.
\end{align*} With \( Q_\varepsilon (x')\) as above, we have that \(B_\varepsilon^+(x) \subset ( 0,x_1+\varepsilon ) \times Q_\varepsilon (x')\), so \begin{align*}
\int_{B_\varepsilon^+(x) } y_1 \dd y \leqslant C\varepsilon^{n-1} \int_0^{x_1+\varepsilon} y_1 \dd y_1 =C\varepsilon^{n-1}(x_1+\varepsilon)^2 \leqslant C\varepsilon^{n+1}
\end{align*} from which we again obtain~\eqref{Hp0lBCzB}.
{F}rom~\eqref{fjrehgeruig009887} and~\eqref{Hp0lBCzB}, we obtain that
$$ f_\epsilon(x)\ge -C\epsilon(M+\epsilon)x_1\ge -(M+\epsilon)x_1$$
for all \(x\in B_{7/8}^+\), as soon as~$\epsilon$ is taken sufficiently small.
This is the desired
result in~\eqref{EuoJ3En6}.
Finally, it follows immediatey from the properties of mollifiers that \( u_\varepsilon \to u \) uniformly in \(B_{7/8}\) as~\(\varepsilon\to 0^+\). Moreover, if \(u \geqslant 0\) in \(\mathbb{R}^n_+\) then from antisymmetry, \begin{align*}
u_\varepsilon (x) = \int_{\mathbb{R}^n_+ } \bar u(y)\big ( \eta_\varepsilon (x-y) - \eta_\varepsilon (x_\ast-y) \big )\dd y \geqslant 0 \qquad \text{for all } x\in \mathbb{R}^n_+
\end{align*} using~\eqref{DPmhap5t}.
\end{proof}
Our second lemma is as follows.
\begin{lem} \thlabel{6fD34E6w}
Suppose that \( v\in C^\infty_0(\mathbb{R}^n)\) is an antisymmetric function satifying \(\partial_1 v (0)=0\). Then \begin{align*}
\lim_{h\to 0 }\frac{(-\Delta)^sv(he_1)} h = -2c_{n,s}(n+2s) \int_{\mathbb{R}^n_+} \frac{y_1 v(y)}{\vert y \vert^{n+2s+2}} \dd y .
\end{align*}
\end{lem}
Note that since \(v\in C^\infty_0(\mathbb{R}^n)\), the fractional Laplacian is given by the usual definition as per \thref{eeBLRjcZ}.
\begin{proof} [Proof of Lemma~\ref{6fD34E6w}]
We will begin by proving that \begin{align}
\lim_{h\to 0 }\frac{(-\Delta)^sv(he_1)} h = (-\Delta)^s \partial_1 v(0). \label{JMT019eU}
\end{align} For this, consider the difference quotient \begin{align*}
\partial^h_1v(x) := \frac{v(x+he_1)-v(x)}h
\end{align*} for all \(x\in \mathbb{R}^n\) and~\(\vert h \vert>0\) small. Since \(v\) is antisymmetric, \(v(0)=0\), so \begin{align*}
\frac{2v(h e_1)-v(he_1+y)-v(he_1-y)}h&= 2\partial^h_1v(0)-\partial^h_1v(y)-\partial^h_1v(-y) - \frac{v(y)+v(-y)}h
\end{align*} for all \(y\in \mathbb{R}^n\). Moreover, the function~\(y \mapsto v(y)+v(-y)\) is odd with
respect to~$y'$, and so \begin{align*}
\int_{\mathbb{R}^n} \frac{v(y)+v(-y)}{\vert y \vert^{n+2s}}\dd y&=0.
\end{align*} It follows that \begin{align*}
\frac{(-\Delta)^sv(he_1)} h &= \frac{c_{n,s}} 2 \int_{\mathbb{R}^n } \bigg ( 2\partial^h_1v(0)-\partial^h_1v(y)-\partial^h_1v(-y) - \frac{v(y)+v(-y)}h \bigg ) \frac{\dd y }{\vert y \vert^{n+2s}}\\
&= (-\Delta)^s \partial^h_1v(0) .
\end{align*} {F}rom these considerations and the computation at the top of p. 9 in \cite{MR3469920}, we have that \begin{align*}
\bigg \vert \frac{(-\Delta)^sv(he_1)} h -(-\Delta)^s \partial_1 v(0) \bigg \vert &= \big \vert (-\Delta)^s ( \partial^h_1v-\partial_1v)(0) \big \vert \\
&\leqslant C \Big( \|\partial^h_1v-\partial_1v\|_{L^\infty(\mathbb{R}^n)} + \|D^2\partial^h_1v-D^2\partial_1v\|_{L^\infty(\mathbb{R}^n)} \Big) .
\end{align*} Then we obtain~\eqref{JMT019eU} by sending \(h\to 0\), using that \(\partial^h_1v \to \partial_1v\) in \(C^\infty_{\mathrm{loc}}(\mathbb{R}^n)\) as \(h\to 0\).
To complete the proof, we use that \(\partial_1v(0)=0\) and integration by parts to obtain \begin{align*}
(-\Delta)^s \partial_1 v(0) &= -c_{n,s} \int_{\mathbb{R}^n} \frac{\partial_1v(y)}{\vert y \vert^{n+2s}} \dd y \\
&= c_{n,s} \int_{\mathbb{R}^n} v(y) \partial_1\vert y \vert^{-n-2s} \dd y \\
&= -c_{n,s} (n+2s) \int_{\mathbb{R}^n} \frac{y_1v(y)}{\vert y \vert^{n+2s+2}} \dd y \\
&= -2c_{n,s} (n+2s) \int_{\mathbb{R}^n_+} \frac{y_1v(y)}{\vert y \vert^{n+2s+2}} \dd y
\end{align*} where the last equality follows from antisymmetry of \(v\).
\end{proof}
We are now able to give the proof of \thref{g9foAd2c}.
\begin{proof}[Proof of \thref{g9foAd2c}] Since \(u\) is non-negative in \(\mathbb{R}^n_+\),
we have that~\( (-\Delta)^s u + \| c\|_{L^\infty(B_\rho^+)} u \geqslant -Mx_1\) in \(B_\rho^+\).
Define~\(\tilde u(x):= u(\rho x)\) and note that \begin{align*}
(-\Delta)^s \tilde u +\rho^{2s} \| c\|_{L^\infty(B_\rho^+)} \tilde u \geqslant -M\rho^{2s+1} x_1 \qquad \text{in } B_1^+.
\end{align*} By way of \thref{tZVUcYJl} (i), we may take a \(C^\infty_0(\mathbb{R}^n)\) sequence of functions approximating \(\tilde u\) which satisfy the assumptions of \thref{g9foAd2c} with \(M\) replaced with \(M+\varepsilon\), obtain the estimate, then pass to the limit. In this way we may assume \(\tilde u \in C^\infty_0(\mathbb{R}^n)\).
Let \(\zeta\) be a smooth radially symmetric cut-off function such that\begin{align*}
\zeta \equiv 1 \text{ in } B_{1/2}, \quad \zeta \equiv 0 \text{ in } \mathbb{R}^n \setminus B_{3/4},\quad {\mbox{and}}
\quad 0\leqslant \zeta \leqslant 1,
\end{align*} and define \(\varphi^{(2)} \in C^\infty_0(\mathbb{R}^n)\) by \(\varphi^{(2)}(x):= x_1 \zeta (x)\) for all \(x\in \mathbb{R}^n\). Suppose that \(\tau \geqslant 0\) is the largest possible value such that \(\tilde u \geqslant \tau \varphi^{(2)}\) in \(\mathbb{R}^n_+\). Since \(\varphi^{(2)}(x)=x_1\) in \(B_{1/2}\), we have that \(x_1 \tau \leqslant \tilde u(x)\) for all~\(x\in B_{1/2}\), so \begin{align}
\tau \leqslant \inf_{B_{1/2}^+} \frac{\tilde u(x)}{x_1} = \rho \inf_{B_{\rho/2}^+} \frac{u(x)}{x_1}. \label{pUJA2JZI}
\end{align} Since \(\tilde u\) is \(C^1\) in \(B_1\), there are two possibilities that can occur: either
there exists \(a\in B_{3/4}^+\) such that~\( \tilde u(a) = \tau \varphi^{(2)} (a)\); or there exists \(a \in B_{3/4} \cap \{ x_1=0\}\) such that \(\partial_1 \tilde u(a) = \tau \partial_1 \varphi^{(2)} (a)\).
First suppose that there exists \(a\in B_{3/4}^+\) such that \( \tilde u(a) = \tau \varphi^{(2)} (a)\). Since \(\varphi^{(2)}\in C^\infty_0(\mathbb{R}^n)\) and is antisymmetric, \((-\Delta)^s\varphi^{(2)}\) is antisymmetric and~$\partial_1 \varphi^{(2)}=0$ in~$\{x_1=0\}$, we can
exploit Lemma~\ref{6fD34E6w} to say that~\((-\Delta)^s\varphi^{(2)}(x)/x_1\) is bounded in \(\mathbb{R}^n\).
On on hand, using that \((\tilde u-\tau \varphi^{(2)})(a)=0\), we have that \begin{align}
(-\Delta)^s(\tilde u-\tau \varphi^{(2)})(a) &= (-\Delta)^s(\tilde u-\tau \varphi^{(2)})(a) +\rho^{2s} \| c\|_{L^\infty(B_\rho^+)}
(\tilde u-\tau \varphi^{(2)})(a) \nonumber \\
&\geqslant -M\rho^{2s+1}a_1 -\tau \big (C + \rho^{2s}\| c \|_{L^\infty(B_\rho^+)} \big ) a_1 \nonumber \\
&\geqslant -M\rho^{2s+1}a_1 -C \tau \big (1 + \rho^{2s} \| c \|_{L^\infty(B_\rho^+)} \big ) a_1 . \label{doW9AF3Y}
\end{align} On the other hand, since \(\tilde u-\tau \varphi^{(2)}\) is antisymmetric, non-negative in \(\mathbb{R}^n_+\), and \((\tilde u-\tau \varphi^{(2)})(a) = 0\), we have by \thref{mkG4iRYH} that \begin{align*}
(-\Delta)^s(\tilde u-\tau \varphi^{(2)})(a) = -C \int_{\mathbb{R}^n_+} \bigg (\frac 1 {\vert a - y \vert^{n+2s}} - \frac 1 {\vert a_\ast- y \vert^{n+2s}}\bigg )(\tilde u-\tau \varphi^{(2)})(y) \dd y .
\end{align*} It follows from~\eqref{buKHzlE6} that \begin{align}
(-\Delta)^s(\tilde u-\tau \varphi^{(2)})(a) &\leqslant -C a_1 \int_{B_{1/2}^+} \frac{y_1 (\tilde u-\tau \varphi^{(2)})(y)} {\vert a_\ast- y \vert^{n+2s+2}}\dd y \nonumber \\
&\leqslant -C a_1 \bigg ( \int_{B_{1/2}^+} y_1 \tilde u(y) \dd y -\tau \bigg ) \nonumber \\
&=-C a_1 \bigg ( \frac 1 {\rho^{n+1}} \int_{B_{\rho/2}^+} y_1 u(y) \dd y -\tau \bigg ). \label{retlxLlR}
\end{align} Rearranging~\eqref{doW9AF3Y} and~\eqref{retlxLlR}, and recalling~\eqref{pUJA2JZI} gives \begin{align}
\frac 1 {\rho^{n+1}} \int_{B_{\rho/2}^+} y_1 u(y) \dd y &\leqslant C \Big(
\tau (1 + \rho^{2s} \| c \|_{L^\infty(B_\rho^+)} ) + \rho^{2s+1}M \Big) \nonumber \\
&\leqslant C \rho (1 + \rho^{2s} \| c \|_{L^\infty(B_\rho^+)} ) \left( \inf_{B_{\rho/2}^+} \frac{u(x)}{x_1} + M \rho^{2s}\right), \label{7O4TL1vF}
\end{align}
which gives the desired result in this case.
Now suppose that there exists \(a \in B_{3/4} \cap \{ x_1=0\}\) such that \(\partial_1 \tilde u(a) =\tau \partial_1 \varphi^{(2)} (a)\). Let \(h>0\) be small and set \(a^{(h)}:= a+ he_1\). On one hand, as in~\eqref{doW9AF3Y}, \begin{align*}
(-\Delta)^s(\tilde u-\tau \varphi^{(2)})(a^{(h)}) &\geqslant -M\rho^{2s+1}h -C \tau \big ( 1+ \rho^{2s}\|c\|_{L^\infty(B_\rho^+)} \big )h .
\end{align*} Dividing both sides by \(h\) and sending \(h\to 0^+\), it follows from \thref{6fD34E6w} (after a translation) that \begin{align*}
-M\rho^{2s+1} -C \tau \big ( 1+ \rho^{2s}\|c\|_{L^\infty(B_\rho^+)} \big ) &\leqslant -C \int_{\mathbb{R}^n_+} \frac{y_1 (\tilde u-\tau \varphi^{(2)})(y)}{\vert y-a\vert^{n+2s+2}} \dd y \\
&\leqslant-C \bigg ( \frac 1{\rho^{n+1}} \int_{B_{\rho/2}^+} y_1 u(y) \dd y - \tau \bigg ).
\end{align*} Rearranging as before gives the desired result.
\end{proof}
Next, we give the proof of~\thref{SwDzJu9i}.
\begin{proof}[Proof of~\thref{SwDzJu9i}]
Let \(\varphi^{(2)}\), \(\tau\), and \(a\) be the same as in the proof of \thref{g9foAd2c}. The proof of \thref{SwDzJu9i} is identical to the proof of \thref{g9foAd2c} except for the following changes.
In the case \(a\in B_{3/4}^+\) and \(\tilde u(a) = \tau \varphi^{(2)}(a)\), we use~\eqref{buKHzlE6} to obtain \begin{align*}
(-\Delta)^s (\tilde u - \tau \varphi^{(2)})(a) \leqslant-C a_1 \int_{\mathbb{R}^n_+} \frac{y_1(\tilde u-\tau \varphi^{(2)})(y)}{\vert a_\ast - y \vert^{n+2s+2}} \dd y \leqslant - C_\rho a_1 \big ( \Anorm{u} -\tau \big )
\end{align*} where we have also used that \begin{align}
\vert a_\ast - y \vert^{n+2s+2} \leqslant C (1+\vert y \vert^{n+2s+2} \big ) \qquad \text{for all } y\in \mathbb{R}^n_+. \label{BWqeWX33}
\end{align} Moreover, in the case \(a\in B_{3/4}\cap \{x_1=0\}\) and \(\partial_1\tilde u(a) = \tau \partial_1\varphi^{(2)}(a)\),
we have that~\(a=a_\ast\) so~\eqref{BWqeWX33} also gives \begin{align*}
\int_{\mathbb{R}^n_+} \frac{y_1 (\tilde u-\tau \varphi^{(2)})(y)}{\vert y-a\vert^{n+2s+2}} \dd y &\geqslant C_\rho \big (\Anorm{u} -\tau \big ) .
\qedhere
\end{align*}
\end{proof}
\subsection{Boundary local boundedness}
We now prove the boundary local boundedness for sub-solutions.
\begin{prop} \thlabel{EP5Elxbz}
Let \(M \geqslant 0\), \(\rho\in(0,1)\), and \(c\in L^\infty(B_\rho^+)\). Suppose that \(u \in C^{2s+\alpha}(B_\rho)\cap \mathscr A_s (\mathbb{R}^n)\) for some \(\alpha>0\) with \(2s+\alpha\) not an integer, and \(u\) satisfies \begin{align*}
(-\Delta)^su +cu &\leqslant M x_1 \qquad \text{in } B_\rho^+.
\end{align*}
Then there exists \(C_\rho>0\) depending only on \(n\), \(s\), \(\| c \|_{L^\infty(B_\rho^+)}\), and \(\rho\) such that \begin{align*}
\sup_{x\in B_{\rho/2}^+} \frac{ u(x)}{x_1} &\leqslant C_\rho ( \Anorm{u} +M ) .
\end{align*}
\end{prop}
Before we prove \thref{EP5Elxbz}, we prove the following lemma.
\begin{lem} \thlabel{KaMmndyO}
Let \(\varphi \in C^\infty(\mathbb{R})\) be an odd function such that \(\varphi(t)=1\) if \(t>2\) and \(0\leqslant \varphi(t) \leqslant 1\) for all \(t\geqslant0\). Suppose that \(\varphi^{(3)} \in C^s(\mathbb{R}^n) \cap L^\infty(\mathbb{R}^n)\) is the solution to \begin{align}
\begin{PDE}
(-\Delta)^s \varphi^{(3)} &= 0 &\text{in } B_1, \\
\varphi^{(3)}(x) &= \varphi(x_1) &\text{in } \mathbb{R}^n \setminus B_1.
\end{PDE} \label{pa7rDaw7}
\end{align}
Then \(\varphi^{(3)}\) is antisymmetric and there exists \(C>1\) depending only on \(n\) and \(s\) such that \begin{align*}
C^{-1}x_1 \leqslant \varphi^{(3)}(x) \leqslant Cx_1
\end{align*} for all \(x\in B_{1/2}^+\).
\end{lem}
\begin{proof}
Via the Poisson kernel representation, see \cite[Section 15]{MR3916700}, and using that \(\varphi\) is an odd function, we may write \begin{align*}
\varphi^{(3)}(x) &= C \int_{\mathbb{R}^n \setminus B_1} \bigg ( \frac{1-\vert x \vert^2}{\vert y \vert^2-1}\bigg )^s \frac{\varphi(y_1)}{\vert x -y \vert^n} \dd y \\
&= C \int_{\mathbb{R}^n_+ \setminus B_1^+} \bigg ( \frac{1-\vert x \vert^2}{\vert y \vert^2-1}\bigg )^s \bigg ( \frac 1 {\vert x -y \vert^n} - \frac 1 {\vert x_\ast -y \vert^n}\bigg ) \varphi(y_1)\dd y.
\end{align*} {F}rom this formula, we immediately obtain that \(\varphi^{(3)}\) is antisymmetric (this can also be argued by uniqueness of solutions to~\eqref{pa7rDaw7}). Then, by an analogous computation to~\eqref{LxZU6} (just replacing~$n+2s$
with~$n$), \begin{align*}
\varphi^{(3)}(x) &\leqslant C x_1 \int_{\mathbb{R}^n_+ \setminus B_1^+} \frac{ y_1\varphi(y_1)}{(\vert y \vert^2-1)^s\vert x-y \vert^{n+2}} \dd y \\
&\leqslant C x_1 \left[\int_{B_2^+ \setminus B_1^+} \frac{ y_1}{(\vert y \vert^2-1)^s } \dd y
+\int_{\mathbb{R}^n_+ \setminus B_2^+} \frac{ y_1}{(\vert y \vert^2-1)^s(\vert y \vert - 1)^{n+2}} \dd y\right] \\
&\leqslant Cx_1
\end{align*} for all \(x\in B_{1/2}^+\). Similarly, using now~\eqref{buKHzlE6} (replacing~$n+2s$
with~$n$), we have that
\begin{align*}
\varphi^{(3)}(x) &\geqslant C x_1 \int_{\mathbb{R}^n_+ \setminus B_1^+} \frac{ y_1\varphi(y_1)}{(\vert y \vert^2-1)^s\vert x_\ast -y \vert^{n+2}} \dd y \\
&\geqslant C x_1 \int_{\{y_1>2\}} \frac{ y_1}{(\vert y \vert^2-1)^s(\vert y \vert +1)^{n+2}} \dd y \\
&\geqslant C x_1
\end{align*} for all \(x\in B_{1/2}^+\).
\end{proof}
Now we can give the proof of \thref{EP5Elxbz}.
\begin{proof}[Proof of \thref{EP5Elxbz}]
Dividing through by \(\Anorm{u}+M\), we can also assume that \((-\Delta)^s u + cu \leqslant x_1\) in \(B_\rho^+\) and \(\Anorm{u}\leqslant 1\). Moreover, as explained at the start of the proof of \thref{g9foAd2c}, via \thref{tZVUcYJl} (ii) (after rescaling), it is not restrictive to assume \(u \in C^\infty_0(\mathbb{R}^n)\) and it is antisymmetric.
Furthermore, we point out that the claim in \thref{EP5Elxbz} is obviously true if~$u\le0$
in~$B^+_\rho$, hence we suppose that~$\{u>0\}\cap B^+_\rho\ne\varnothing$.
Let \(\varphi^{(3)}\) be as in \thref{KaMmndyO} and let \(\zeta(x) := \varphi^{(3)}(x/(2\rho)) \). Suppose that \(\tau \geqslant 0\) is the smallest value such that \begin{align*}
u(x) &\leqslant \tau \zeta (x) (\rho - \vert x \vert )^{-n-2} \qquad \text{in } B_\rho^+.
\end{align*}
Notice that~$\tau>0$.
To complete the proof we will show that \(\tau \leqslant C_\rho\) with \(C_\rho\) independent of \(u\). Since \(u\) is continuously differentiable, there are two possibilities that can occur: \begin{itemize}
\item[Case 1:] There exists \(a \in B_\rho^+\) such that \begin{align*}
u(a) &= \tau \zeta (a) (\rho - \vert a \vert )^{-n-2}.
\end{align*}
\item[Case 2:] There exists \(a\in B_\rho \cap \{x_1=0\}\) such that \begin{align*}
\partial_1 u(a) = \tau \partial_1 \big \vert_{x=a} \big ( \zeta (x) (\rho - \vert x \vert )^{-n-2} \big )= \tau (\partial_1 \zeta (a)) (\rho - \vert a \vert )^{-n-2} .
\end{align*}
\end{itemize}
Let \(d:= \rho - \vert a\vert \) and define \(U\subset B_\rho^+\) as follows: if Case 1 occurs let \begin{align*}
U := \bigg \{ x \in B_\rho^+ \text{ s.t. } \frac{u(x)}{\zeta(x)} > \frac{u(a)}{2\zeta (a)} \bigg \};
\end{align*} otherwise if Case 2 occurs then let \begin{align*}
U:=\bigg \{ x \in B_\rho^+ \text{ s.t. } \frac{u(x)}{\zeta(x)} > \frac{\partial_1u(a)}{2 \partial_1\zeta (a)} \bigg \}.
\end{align*} Since \(u(a) = \tau \zeta(a) d^{-n-2}\) in Case 1 and \(\partial_1u(a) = \tau \partial_1\zeta(a) d^{-n-2}\) in Case 2, we may write \begin{equation}\label{ei395v7b865998cn754mx984zUUU}
U = \bigg \{ x \in B_\rho^+ \text{ s.t. } u(x)> \frac 1 2 \tau d^{-n-2} \zeta(x) \bigg \}
\end{equation}which is valid in both cases.
Then, we have that, for all \(r\in(0,d)\),
\begin{align*}
C_\rho &\geqslant \int_{B_\rho^+} y_1| u(y)| \dd y \geqslant \frac 1 2 \tau d^{-n-2} \int_{U \cap B_r(a)} y_1 \zeta(y) \dd y .
\end{align*}
As a consequence, by \thref{KaMmndyO}, we have that \begin{align}
\int_{U \cap B_r(a)} y_1^2 \dd y
\le C_\rho\int_{U \cap B_r(a)} y_1 \zeta(y) \dd y
\leqslant \frac{ C_\rho d^{n+2} } \tau \qquad \text{for all } r\in(0,d). \label{TTWpDkie}
\end{align} Next, we make the following claim.
\begin{claim}
There exists~$\theta_0\in(0,1)$ depending only on \(n\), \(s\), \(\| c\|_{L^\infty(B_\rho(e_1))}\), and \(\rho\) such that
if~$\theta\in(0,\theta_0]$ there exists~\(C>0\)
depending only on \(n\), \(s\), \(\| c\|_{L^\infty(B_\rho(e_1))}\), \(\rho\), and~$\theta$ such that
\begin{itemize}
\item In Case 1:
\begin{enumerate}[(i)]
\item If \(a_1 \geqslant \theta d/16 \) then \begin{align*}
\big \vert B_{(\theta d)/64}(a) \setminus U \big \vert &\leqslant \frac 1 4 \big \vert B_{(\theta d)/64} \big \vert +\frac{C d^n} \tau .
\end{align*}
\item If \(a_1 < \theta d/16 \) then \begin{align*}
\int_{B_{(\theta d)/64}^+(a) \setminus U } x_1^2 \dd x &\leqslant \frac 1 4 \int_{B_{(\theta d)/64}^+(a) } x_1^2 \dd x + \frac{C d^{n+2}}\tau .
\end{align*}
\end{enumerate}
\item In Case 2: \begin{align*}
\int_{B_{(\theta d)/64}^+(a)\setminus U } x_1^2\dd x &\leqslant\frac 1 4 \int_{B_{(\theta d)/64}^+ (a)} x_1^2 \dd x + \frac{ C d^{n+2}}\tau.
\end{align*}
\end{itemize} In particular, neither~\(\theta\) nor~\(C\) depend on \(\tau\), \(u\), or \(a\).
\end{claim}
We withhold the proof of the claim until the end. Assuming that
the claim is true, we complete the proof of \thref{EP5Elxbz} as follows.
If Case 1(i) occurs then for all \(y\in B_{(\theta_0 d)/64} (a)\)
we have that~\(y_1>a_1-( \theta_0 d)/64 \geqslant C d\), and so \begin{align*}
\int_{U \cap B_{(\theta_0 d)/64}(a)} y_1^2 \dd y \geqslant C d^2 \cdot \big \vert U \cap B_{(\theta_0 d)/64}(a) \big \vert .
\end{align*} Hence, from~\eqref{TTWpDkie} (used here with~$r:=(\theta_0d)/64$), we have that \begin{align*}
\vert U \cap B_{( \theta_0 d)/64}(a) \big \vert \leqslant \frac{C_\rho d^n}{\tau}.
\end{align*}
Then using the claim, we find that \begin{align*}
\frac{C_\rho d^n}{\tau} \geqslant \big \vert B_{(\theta_0 d)/64} \big \vert - \big \vert B_{(\theta_0 d)/64}(a) \setminus U \big \vert
\geqslant \frac 3 4 \big \vert B_{(\theta_0 d)/64} \big \vert -\frac{C d^n} \tau
\end{align*}which gives that \(\tau \leqslant C_\rho\) in this case.
If Case 1(ii) or Case 2 occurs then from~\eqref{TTWpDkie} (used here with~$r:=(\theta_0d)/64$)
and the claim, we have that \begin{equation}\label{dewiotbv5748976w4598ty}
\frac{C_\rho d^{n+2}}{\tau} \geqslant \int_{B_{(\theta_0 d)/64}^+(a) } x_1^2 \dd x - \int_{B_{( \theta_0 d)/64}^+(a) \setminus U } x_1^2 \dd x
\geqslant \frac 3 4 \int_{B_{( \theta_0 d)/64}^+(a) } x_1^2 \dd x - \frac{C d^{n+2}}\tau .
\end{equation}
We now observe that, given~$r\in(0,d)$, if~$x\in B_{r/4}\left(a+\frac34 re_1\right)\subset B^+_{r}(a)$
then~$x_1\ge a_1+\frac34 r-\frac{r}4\ge\frac{r}2$,
and thus
\begin{equation}\label{sdwet68980poiuytrkjhgfdmnbvcx23456789}
\int_{B_{r}^+(a) } x_1^2 \dd x\ge
\int_{B_{r/4}(a+(3r)/4 e_1) } x_1^2 \dd x\ge \frac{r^2}4\,|B_{r/4}(a+(3r)/4 e_1) |=C r^{n+2},\end{equation}
for some~$C>0$ depending on~$n$. Exploiting this formula with~$r:=(\theta_0 d)/64$ into~\eqref{dewiotbv5748976w4598ty},
we obtain that
$$ \frac{C_\rho d^{n+2}}{\tau} \geqslant C \left(\frac{ \theta_0 d}{64}\right)^{n+2 }- \frac{C d^{n+2}}\tau,$$
which gives that \(\tau \leqslant C_\rho\) as required.
Hence, we now focus on the proof of the claim.
Let \(\theta\in(0,1)\) be a small constant to be chosen later. By making a translation, we may also assume without loss of generality that \(a'=0\) (recall that~\(a'=(a_2,\dots,a_n)\in \mathbb{R}^{n-1}\)).
For each \(x\in B_{\theta d/2}^+(a)\), we have that~\(\vert x \vert \leqslant \vert a \vert +\theta d/2=\rho- (1-\theta/2)d\). Hence, in both Case 1 and Case 2,
\begin{align}
u(x) \leqslant \tau d^{-n-2} \bigg (1- \frac \theta 2 \bigg )^{-n-2}\zeta (x) \qquad \text{in }B_{\theta d/2}^+(a). \label{ln9LcJTh}
\end{align} Let \begin{align*}
v(x):= \tau d^{-n-2} \bigg (1- \frac \theta 2 \bigg )^{-n-2}\zeta (x)-u(x) \qquad \text{for all } x\in \mathbb{R}^n.
\end{align*} We have that \(v\) is antisymmetric and \(v\geqslant 0\) in \(B_{\theta d/2}^+(a) \) due to~\eqref{ln9LcJTh}. Moreover, since \(\zeta\) is \(s\)-harmonic in \(B_\rho^+ \supset B_{\theta d/2}^+(a)\), for all \(x\in B_{\theta d/2}^+(a)\), \begin{align*}
(-\Delta)^s v(x) +c(x)v(x) &= -(-\Delta)^s u(x) - c(x) u(x) +c(x) \tau d^{-n-2} \bigg (1- \frac \theta 2 \bigg )^{-n-2}\zeta (x)\\
&\geqslant -x_1 -C \tau d^{-n-2} \|c^-\|_{L^\infty(B_\rho^+)} \bigg (1- \frac \theta 2 \bigg )^{-n-2}\zeta (x).
\end{align*} Taking \(\theta\) sufficiently small and using that \(\zeta (x)\leqslant C x_1\) (in light of~\thref{KaMmndyO}),
we obtain \begin{align}
(-\Delta)^s v(x) +c(x)v(x)&\geqslant -C \big ( 1 + \tau d^{-n-2} \big )x_1 \qquad \text{in } B_{\theta d/2}^+(a). \label{ulzzcUwf}
\end{align}
Next, we define \(w(x):= v^+(x)\) for all \(x\in \mathbb{R}^n_+\) and \(w(x):= -w(x_\ast)\) for all \(x\in \overline{\mathbb{R}^n_-}\).
We point out that, in light of~\eqref{ln9LcJTh}, $w$ is as regular as~$v$ in~$B_{\theta d/2}^+(a)$, and thus
we can compute the fractional Laplacian of~$w$ in~$B_{\theta d/2}^+(a)$ in a pointwise sense.
We also observe that \begin{align*}
(w-v)(x) (x) &= \begin{cases}
0 &\text{if } x\in \mathbb{R}^n_+ \cap \{ v\geqslant 0\} ,\\
u(x)-\tau d^{-n-2} \big (1- \frac \theta 2 \big )^{-n-2}\zeta (x),&\text{if } x\in \mathbb{R}^n_+ \cap \{ v< 0\}.
\end{cases}
\end{align*} In particular, \(w-v\leqslant |u|\) in \(\mathbb{R}^n_+\).
Thus, for all \(x\in B_{\theta d/2}^+(a)\), \begin{align*}
(-\Delta)^s (w-v)(x)&\geqslant -C \int_{\mathbb{R}^n_+\setminus B_{\theta d/2}^+(a)} \left(\frac 1 {\vert x -y \vert^{n+2s}} - \frac 1 {\vert x_\ast -y \vert^{n+2s}} \right) |u(y)|\dd y .
\end{align*} Moreover, by \thref{ltKO2}, for all~$x \in B_{\theta d/4}^+(a)$,
\begin{equation}
(-\Delta)^s (w-v)(x) \geqslant -C(\theta d)^{-n-2s-2} \Anorm{u}
x_1 \geqslant -C(\theta d)^{-n-2s-2}x_1 . \label{y8tE2pf9}
\end{equation} Hence, by~\eqref{ulzzcUwf} and~\eqref{y8tE2pf9}, we obtain \begin{align}
(-\Delta)^s w +cw &=(-\Delta)^sv +cv +(-\Delta)^s(w-v) \nonumber \\
&\geqslant -C \bigg ( 1 + \tau d^{-n-2} +(\theta d)^{-n-2s-2} \bigg )x_1 \nonumber \\
&\geqslant -C \bigg ( (\theta d)^{-n-2s-2} + \tau d^{-n-2} \bigg )x_1\label{ayFE7nQK}
\end{align} in \( B_{\theta d/4}^+(a)\).
Next, let us consider Case 1 and Case 2 separately.
\emph{Case 1:} Suppose that \(a\in B_\rho^+\) and let \(\tilde w (x) = w (a_1 x)\) and \(\tilde c (x) = a_1^{2s} c (a_1 x)\). Then from~\eqref{ayFE7nQK}, we have that \begin{align}
(-\Delta)^s \tilde w(x) +\tilde c(x) \tilde w(x) &\geqslant - Ca_1^{2s+1} \bigg ( (\theta d)^{-n-2s-2} + \tau d^{-n-2} \bigg )x_1 \label{zrghHZhX}
\end{align} for all \( x\in B_{\theta d/(4a_1)}^+(e_1)\).
As in the proof of \thref{guDQ7}, we wish to apply the rescaled version of the weak Harnack inequality to \(\tilde w\); however, we cannot immediately apply either \thref{YLj1r} or \thref{g9foAd2c} to~\eqref{zrghHZhX}. To resolve this, let us split into a further two cases: (i) \(a_1 \geqslant \theta d/16 \) and (ii) \(a_1< \theta d/16\).
\emph{Case 1(i):} If \(a_1 \geqslant \theta d/16 \) then \( B_{\theta d/(32a_1)}(e_1) \subset B_{\theta d/(4a_1)}^+(e_1)\) and for each \(x\in B_{\theta d/(32a_1)}(e_1)\) we have that~\(x_1<1+\theta d/(32a_1)\le1+1/4=5/4\). Therefore, from~\eqref{zrghHZhX}, we have that \begin{align*}
(-\Delta)^s \tilde w(x) +\tilde c(x) \tilde w(x) &\geqslant - Ca_1^{2s+1} \big ( (\theta d)^{-n-2s-2} + \tau d^{-n-2} \big )
\end{align*} for all \( x\in B_{\theta d/(32a_1)}(e_1)\).
On one hand, by~\thref{YLj1r} (used here with~$\rho:=\theta d/(32a_1)$), \begin{align*}&
\left( \frac{\theta d}{64} \right)^{-n} \int_{B_{\theta d/64}(a)} w (x) \dd x \\= &\; \left( \frac{\theta d}{32a_1} \right)^{-n} \int_{B_{\theta d/(64a_1)}(e_1)} \tilde w (x) \dd x \\
\leqslant&\; C \left( \tilde w(e_1) + a_1 (\theta d)^{-n-2} + \tau a_1 \theta^{2s} d^{-n+2s-2} \right) \\
\leqslant& \;C \tau d^{-n-2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 \bigg ) a_1 +Ca_1(\theta d)^{-n-2} + C\tau a_1 \theta^{2s} d^{-n+2s-2}
\end{align*} using also \thref{KaMmndyO} and that \(u(a) = \tau d^{-n-2} \zeta(a)\).
On the other hand, by the definition of~$U$ in~\eqref{ei395v7b865998cn754mx984zUUU},
\begin{align}
B_r(a) \setminus U &\subset \left\{ \frac{w}{\zeta} > \tau d^{-n-2} \left( \left(1 - \frac \theta 2 \right)^{-n-2} -\frac 1 2 \right) \right\} \cap B_r(a), \qquad \text{ for all } r\in\left(0,\frac12\theta d\right),\label{I9FvOxCz}
\end{align} and so \begin{align*}
(\theta d)^{-n} \int_{B_{\theta d/64}(a)} w(x) \dd x &\geqslant \tau \theta^{-n} d^{-2n-2} \bigg ( \bigg (1 - \frac \theta 2 \bigg )^{-n-2} -\frac 1 2 \bigg ) \int_{B_{\theta d/64}(a) \setminus U } \zeta (x) \dd x \\
&\geqslant C \tau \theta^{-n} d^{-2n-2} \int_{B_{\theta d/64}(a) \setminus U } \zeta (x) \dd x
\end{align*} for \(\theta\) sufficiently small. Moreover, using that \(x_1>a_1-\theta d/64>Ca_1\) and \thref{KaMmndyO}, we have that\begin{align*}
\int_{B_{\theta d/64}(a) \setminus U } \zeta (x) \dd x \geqslant C\int_{B_{\theta d/64}(a) \setminus U } x_1 \dd x \geqslant C
a_1\cdot \big \vert B_{\theta d/64}(a) \setminus U \big \vert .
\end{align*}
Thus,\begin{align*}
\big \vert B_{\theta d/64}(a) \setminus U \big \vert &\leqslant C (\theta d)^n \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 \bigg ) +\frac{C\theta^{-2} d^n} \tau + C (\theta d)^{n+2s} \\
&\leqslant C (\theta d)^n \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 +\theta^{2s} \bigg ) +\frac{C\theta^{-2} d^n} \tau ,
\end{align*} using also that \(d^{n+2s}<d^n\).
Finally, we can take \(\theta\) sufficiently small so that \begin{align*}
C (\theta d)^n \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 +\theta^{2s} \bigg ) \leqslant \frac 1 4 \big \vert B_{\theta d/64} \big \vert
\end{align*}
which gives
$$ \big \vert B_{\theta d/64}(a) \setminus U \big \vert\le \frac14 \big \vert B_{\theta d/64} \big \vert
+\frac{Cd^n} \tau.
$$
This concludes the proof of the claim in Case 1(i).
\emph{Case 1(ii):} Let \(a_1< \theta d/16\) and fix \(R:= \frac 1 {2} \big ( \sqrt{ ( \theta d/(4a_1))^2-1} +2 \big )\). Observe that \begin{align*}
2 < R< \sqrt{ \left(\frac{ \theta d}{4a_1}\right)^2-1}.
\end{align*} Hence, \(e_1 \in B_{R/2}^+\). Moreover, if \(x\in B_R^+\) then \begin{align*}
\vert x -e_1\vert^2 <1+R^2< ( \theta d/(4a_1))^2,
\end{align*} so \(B_R^+ \subset B_{\theta d/(4a_1)}^+(e_1)\).
Thus, applying \thref{g9foAd2c} to the equation in~\eqref{zrghHZhX} in \(B_R^+\), we obtain \begin{align*}&
a_1^{-n-1}\bigg ( \frac R 2\bigg )^{-n-2} \int_{B_{a_1R/2}^+} x_1 w(x) \dd x \\
=&\; \bigg ( \frac R 2\bigg )^{-n-2} \int_{B_{R/2}^+} x_1 \tilde w(x) \dd x \\
\leqslant&\; C \left( \inf_{B_{R/2}^+} \frac{\tilde w(x)}{x_1} + a_1^{2s+1}R^{2s} (\theta d)^{-n-2s-2} + \tau d^{-n-2}a_1^{2s+1}R^{2s} \right) \\
\leqslant\;& C \tau d^{-n-2} \left( \left(1- \frac \theta 2 \right)^{-n-2}-1 \right)\zeta(a)
+ C a_1^{2s+1}R^{2s} (\theta d)^{-n-2s-2} +C \tau d^{-n-2}a_1^{2s+1}R^{2s}
\end{align*} using that \(e_1 \in B_{R/2}^+\).
Since \(R \leqslant C \theta d/a_1\) and \(\zeta(a)\leqslant Ca_1\) by \thref{KaMmndyO}, it follows that
\begin{eqnarray*}&&
a_1^{-1} \bigg ( \frac R {2a_1} \bigg )^{-n} \int_{B_{a_1R/2}^+} x_1 w(x) \dd x \\
&&\qquad\leqslant C \tau d^{-n-2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 \bigg )a_1 + C a_1 (\theta d)^{-n-2} +C \tau a_1 \theta^{2s} d^{-n+2s-2} .
\end{eqnarray*}
On the other hand, we claim that
\begin{equation}\label{skweogtry76t678tutoy4554yb76i78io896}
B_{(\theta d)/64}^+(a) \setminus U \subset B_{a_1R/2}^+ .\end{equation} Indeed,
if~$x\in B_{(\theta d)/64}^+ \setminus U $ then
$$ |x|\le |x-a|+|a|\le \frac{\theta d}{64} +a_1\le a_1\left(\frac{\theta d}{64a_1}+1\right).
$$
Furthermore,
\begin{eqnarray*}
&&R\ge \frac 1 {2} \sqrt{ \left(\frac{\theta d}{4a_1}\right)^2-\left(\frac{\theta d}{16a_1}\right)^2} +1\ge
\frac{\sqrt{15}\theta d}{32a_1}+1\ge\frac{3\theta d}{32a_1}+1\\&&\qquad
=2\left(\frac{3\theta d}{64a_1}+\frac12\right)=2\left(\frac{\theta d}{64a_1}+\frac{\theta d}{32a_1}+\frac12\right)\ge
2\left(\frac{\theta d}{64a_1}+1\right).
\end{eqnarray*}
{F}rom these observations we obtain that if~$x\in B_{(\theta d)/64}^+ \setminus U $ then~$|x|\le a_1R/2$,
which proves~\eqref{skweogtry76t678tutoy4554yb76i78io896}.
Hence, by~\eqref{I9FvOxCz} (used with~$r:=\theta d/64$), \eqref{skweogtry76t678tutoy4554yb76i78io896}, and \thref{KaMmndyO}, we have that \begin{align*}
a_1^{-1} \bigg ( \frac R {2a_1} \bigg )^{-n} \int_{B_{a_1R/2}^+} x_1 w(x) \dd x &\geqslant a_1^{-n-1} R^{-n-2} \tau d^{-n-2} \bigg ( \bigg (1 - \frac \theta 2 \bigg )^{-n-2} -\frac 1 2 \bigg ) \int_{B_{ (\theta d)/64}^+(a) \setminus U} x_1 \zeta(x) \dd x \\
&\geqslant C \tau a_1 \theta^{-n-2} d^{-2n-4} \int_{B_{ (\theta d)/64}^+(a) \setminus U } x_1^2 \dd x
\end{align*} for \(\theta\) sufficiently small. Thus, \begin{align*}
\int_{B_{( \theta d)/64}^+(a) \setminus U } x_1^2 \dd x &\leqslant C (\theta d)^{n+2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 \bigg ) + \frac{C d^{n+2}}\tau +C( \theta d)^{n+2s+2}\\
&\leqslant C (\theta d)^{n+2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 +\theta^{2s} \bigg ) + \frac{C d^{n+2}}\tau
\end{align*} using that \(d^{n+2s+2}<d^{n+2}\).
Recalling formula~\eqref{sdwet68980poiuytrkjhgfdmnbvcx23456789}
and taking~\(\theta\) sufficiently small, we obtain that \begin{align*}
C(\theta d)^{n+2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 +\theta^{2s} \bigg ) \leqslant \frac 1 4 \int_{B_{( \theta d)/64}^+ (a)} x_1^2 \dd x .
\end{align*} which concludes the proof in Case 1(b).
\emph{Case 2:} In this case, we can directly apply \thref{g9foAd2c} to~\eqref{ayFE7nQK}. When we do this we find that \begin{align*}&
\bigg ( \frac{\theta d }{64} \bigg )^{-n-2} \int_{B_{\theta d/64}^+(a)} x_1 w (x)\dd x \\
\leqslant&\, C \bigg ( \partial_1 w (0) + (r\theta)^{-n-2} + \tau \theta^{2s} d^{-n+2s-2} \bigg )\\
=&\, C\tau d^{-n-2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 \bigg )\partial_1\zeta(0) + C (r\theta)^{-n-2} + C\tau \theta^{2s} d^{-n+2s-2}.
\end{align*} On the other hand,~\eqref{I9FvOxCz} is still valid in Case 2 so \begin{align*}
(\theta d )^{-n-2} \int_{B_{\theta d/64}^+(a)} x_1 w (x)\dd x &\geqslant \tau \theta^{-n-2} d^{-2n-4} \bigg ( \bigg (1 - \frac \theta 2 \bigg )^{-n-2} -\frac 1 2 \bigg ) \int_{B_{\theta d/64}^+(a)\setminus U } x_1 \zeta (x) \dd y \\
&\geqslant C \tau \theta^{-n-2} d^{-2n-4} \int_{B_{\theta d/64}^+(a)\setminus U } x_1^2\dd x
\end{align*} using \thref{KaMmndyO} and taking \(\theta\) sufficiently small. Thus, \begin{align*}
\int_{B_{\theta d/64}^+(a)\setminus U } x_1^2\dd x &\leqslant C (\theta d)^{n+2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 \bigg ) + \frac{ C d^{n+2}}\tau + C (\theta d)^{n+2s+2}\\
&\leqslant C (\theta d)^{n+2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 +\theta^{2s} \bigg ) + \frac{ C d^{n+2}}\tau
\end{align*} using that \(d^{n+2s+2}<d^{n+2}\). Then, by~\eqref{sdwet68980poiuytrkjhgfdmnbvcx23456789}, we may choose \(\theta\) sufficiently small so that \begin{align*}
C (\theta d)^{n+2} \bigg ( \bigg (1- \frac \theta 2 \bigg )^{-n-2}-1 +\theta^{2s} \bigg ) \leqslant \frac 1 4 \int_{B_{\theta d/64}^+(a)} x_1^2\dd x
\end{align*} which concludes the proof in Case 2.
The proof of \thref{EP5Elxbz}
is thereby complete.
\end{proof}
\section{Proof of \protect\thref{Hvmln}} \label{lXIUl}
In this short section, we give the proof of \thref{Hvmln}. This follows from Theorems~\ref{DYcYH} and~\ref{C35ZH} along with a standard covering argument which we include here for completeness.
\begin{proof}[Proof of \thref{Hvmln}]
Recall that~\(\Omega^+=\Omega \cap \mathbb{R}^n_+\) and let \(B_{2R}(y)\) be a ball such that \(B_{2R}(y)\Subset\Omega^+\). We will first prove that there exists a constant \(C=C(n,s,R,y)\) such that \begin{align}
\sup_{B_R(y)} \frac{u(x)}{x_1} &\leqslant C \inf_{B_R(y)} \frac{u(x)}{x_1} . \label{c4oOr}
\end{align}Indeed, if \(\tilde u (x) := u (y_1x+(0,y'))\) and~\(\tilde c (x):= y_1^{2s} c(y_1x+(0,y'))\) then \begin{align*}
(-\Delta)^s \tilde u (x)+\tilde c \tilde u = 0, \qquad \text{in } B_{2R/y_1} (e_1).
\end{align*} By \thref{DYcYH}, \begin{align*}
\sup_{B_R(y)} \frac{u(x)}{x_1} \leqslant C \sup_{B_R(y)} u = C\sup_{B_{R/y_1}(e_1)} \tilde u & \leqslant C \inf_{B_{R/y_1}(e_1)} \tilde u =C\inf_{B_R(y)} u \leqslant C\inf_{B_R(y)} \frac{u(x)}{x_1}
\end{align*} using that \(B_R(y) \Subset\mathbb{R}^n_+\).
Next, let \(\{a^{(k)}\}_{k=1}^\infty,\{b^{(k)}\}_{k=1}^\infty \subset \tilde \Omega^+\) be such that \begin{align*}
\frac{u(a^{(k)})}{a^{(k)}_1} \to \sup_{\Omega'}\frac{u(x)}{x_1} \quad \text{and} \quad \frac{u(b^{(k)})}{b^{(k)}_1} \to \inf_{\Omega'}\frac{u(x)}{x_1}
\end{align*} as \(k\to \infty\). After possibly passing to a subsequence, there exist~\(a\), \(b\in \overline{\tilde \Omega^+}\) such that~\(a^{(k)}\to a\) and~\(b^{(k)}\to b\). Let \(\gamma \subset \tilde \Omega^+\) be a curve connecting \(a\) and \(b\). By the Heine-Borel Theorem, there exists a finite collection of balls \(\{B^{(k)}\}_{k=1}^M\) with centres in \(\tilde \Omega \cap \{x_1=0\}\) such that \begin{align*}
\tilde \Omega \cap \{x_1=0\} \subset \bigcup_{k=1}^M B^{(k)} \Subset\Omega .
\end{align*} Moreover, if \( \tilde \gamma := \gamma \setminus \bigcup_{k=1}^M B^{(k)}\) then there exists a further collection of balls \(\{B^{(k)}\}_{k=M+1}^N\) with centres in \(\tilde \gamma\) and radii equal to \( \frac 12 \operatorname{dist}(\tilde \gamma, \partial( \Omega^+))\) such that \begin{align*}
\tilde \gamma \subset \bigcup_{k=M+1}^N B^{(k)} \Subset \Omega^+.
\end{align*} By construction, \(\gamma\) is covered by \(\{B^{(k)}\}_{k=1}^n\). Thus, iteratively applying \thref{C35ZH} (after translating and rescaling) to each \(B^{(k)}\), \(k=1,\dots, M\), and~\eqref{c4oOr} to each \(B^{(k)}\), \(k=M+1,\dots, N\), we obtain the result.
\end{proof}
|
\section{Introduction}
The global minimization of a non-convex function is one of the most challenging problems in modern optimization. There are few global optimization methods which provide reasonable convergence guarantees, the most famous are probably the simulated annealing, whose premises are found in \cite{metropolis53}, or the moment method \cite{lasserre01}, and their many variants. On the other hand, metaheuristics methods are numerous and have some notable empirical success: they orchestrate interactions between local and global strategies, combining random and deterministic procedures, and often ending up with methods using optimizing agents. Some examples of metaheuristics are inspired by analogies with biology, as evolutionary algorithms \cite{Fogel:2000}, ethology (e.g. ant colonies \cite{DorigoBlum:2005}), or particle swarms, see e.g., \cite{KennedyEberhart:1995}. The goal of this paper is to introduce a new family of swarm methods through gradient descent in the Monge-Kantorovich space and give general guarantees for their convergence to global minimizers.
Let us be more specific and consider the problem of solving
\begin{equation}\quad \min_M U,\tag{$\P$}\end{equation}
where $U:M\to \mathbb R$ is a differentiable function defined on a compact Riemannian manifold $M$. In order to introduce our swarm methods, we need first some considerations on simulated annealing.
\paragraph{Three views on simulated annealing} Our starting point is indeed the famous si\-mu\-lated annealing method. In its time-continuous form, and when the state space is flat $M$ (e.g., a flat torus), it is a solution $X\df(X_t)_{t\geq 0}$ of the time-inhomogeneous Langevin-like stochastic differential equation
\begin{equation}\label{langevin}
dX_t=-\beta_t\nabla U(X_t)\,dt+\sqrt{2}dW_t,\tag{a}
\end{equation}
where $(W_t)_{t\geq 0}$ is a Brownian motion and $\beta_t\to +\infty$ \footnote{Often called the inverse temperature} is a time-dependent parameter tuned so that the expectation of $U(X_t)$ tends to $\min_M U$. The formulation \eqref{langevin} can be extended to any compact Riemannian manifold $M$, but this requires more involved notations, see e.g.\
Ikeda and Watanabe \cite{MR1011252} or Emery \cite{MR1030543}.
\par
The intuitive interpretation is quite natural, the method combines local gradient search with a vanishing Brownian exploration of the feasible set $M$.
Although the method is often used as a heuristics, its proof has been made rigorous in various frameworks via different approaches, the two main ones being based on large deviations, see e.g.\ the compendium by Azencott et al.\ \cite{zbMATH00053884}, and on functional inequalities, cf.\ Holley, Kusuoka and Stroock \cite{MR995752}.
Key to the foundational approach of \cite{MR995752} , is the establishment of a generalized log-Sobolev inequality followed by hypercontractivity arguments.
This approach was then simplified by Miclo \cite{zbMATH04157608}, via the identification of the relative entropy as a convenient Lyapunov function.
In order to explain the role of the log-Sobolev inequality, the relative entropy, the mechanisms behind the convergence properties, and understand the scope of the method, we view simulated annealing along three complementary angles:
\begin{itemize}
\item[(a)] the SDE form of the algorithm: the overdamped Langevin dynamics (a) above,
\item[(b)] the PDE counterpart of (a) which describes the time evolution of the density $t\mapsto\rho(t)$ of $X_t$. It assumes the form of a Fokker-Planck equation,
\begin{equation}\label{FokkerPlanck}\tag{b}
\frac{{\rm d}}{{\rm d}t} \rho =\beta_t \,\mathrm{div}(\rho\nabla U)+\Delta \rho,\quad t\geq 0.
\end{equation}
\item[(c)] Otto's formalism \cite{jko}
allows to interpret the latter as a gradient-like system in the space of probabilities on $M$ endowed with Monge-Kantorovich metric:
\begin{equation*}
\tag{c}
\frac{{\rm d}}{{\rm d}t}\,\rho(t)=-{\rm grad}_\M \, \mathcal U_{\beta_t}[\rho(t)],
\end{equation*}
\end{itemize}
with $\displaystyle \mathcal U_{\beta}[\rho]=\beta\int_M U\rho\,d\ell+\int_{M} \rho\log \rho\,d\ell,$
where $\ell$ is the Riemannian measure of $M$. This quantity is also known as the relative entropy of $\rho$ with respect to the Gibbs measure whose density is proportional to $\exp(-\beta U)$, which served as a Lyapunov function in \cite{zbMATH04157608}.
\smallskip
This triple perspective, mainly due to \cite{jko}, is not new and has known a recent success in sampling \cite{dalalyan17,durmus17}, optimization \cite{Ma19,Ma21} and machine learning \cite{Raginsky17,bartlett}.
\smallskip
Depending on the form we adopt to study the dynamics, subsequent results or developments may be considerably easier to understand. Indeed, while (a) classically provides operational algorithms through discretization, (b) offers a trac\-ta\-ble version amenable to classical PDE analysis methods as Lyapunov methods. As for the last angle, (c), it confers a sharp geometrical content to the method and allows to interpret the essential tools of convergence through classical intuitive geometric ideas. An essential fact about (c) is that functional inequalities, as the log-Sobolev inequality, may be seen as \L ojasiewicz gradient inequalities. In our case it means that there exists an exponent $\gamma\in (0,1)$ such that the slope of $(\mathcal U_\beta-\min\mathcal U_\beta)^\gamma$ is bounded away from zero (save at the stationary measure). This reparametrization sharpens the energy while leaving unchanged level sets: this allows for a direct convergence analysis of the gradient method (c), see \cite{blanchet18} and references therein for further insights. In the simulated annealing case the log-Sobolev inequality of Holley, Kusuoka and Stroock \cite{MR995752} turns out to be an instance of such an inequality, see \cite{blanchet18}.
\medskip
\paragraph{Swarm gradient dynamics} The triple-perspective (a)-(b)-(c) we used to describe the strategy of simulated annealing can be generalized to a much larger framework. For this, we adopt the angle (c) under which we observe that it is natural to consider more general convex functions $\varphi$ than $\mathbb R_+\ni r\mapsto r\ln(r)$ in the Boltzmann entropy. Referring to the results in \cite{ags}, we may indeed use a whole family of convex functionals
\begin{equation*}
\mathcal H[\rho]=\int_M \varphi(\rho)\,d\ell,
\end{equation*}
leading to a penalized cost $$ \mathcal U_\beta[\rho]=\beta\int_M U\rho\,d\ell+\mathcal H[\rho]$$
and to the triplet of ``equivalent'' minimizing dynamics modeled on (c), (b), (a),
\begin{align}\label{gradphi}
&\frac{{\rm d}}{{\rm d}t} \rho (t) = -{\rm grad}_\M \,\mathcal U_{\beta_t}[\rho(t)]\\
\label{pdephi}
& \frac{{\rm d}}{{\rm d}t} \rho = \beta_t\mathrm{div}(\rho\nabla U)+ \mathrm{div}\rho \nabla \varphi'(\rho)\\
& dX_t =-\beta_t\nabla U(X_t)\,dt+\sqrt{2}\alpha(\rho)dW_t,
\end{align}
with $\lim_{t\to\infty}\beta_t=+\infty$, $\varphi,\alpha$ are some positive functions to be specified, and $\rho$ is the law of $X$. The fact that a particle interacts with its law may be considered as a swarm effect, this is why we call these dynamics {\em swarm gradient dynamics}\footnote{Since $\beta$ is variable, they are actually time-dependent swarm dynamics.}, see Section \ref{s:swarm} for more insight.
Principles and other considerations behind the above dynamics are described in Sections \ref{s:globdyn} and \ref{s:swarm}.
\medskip
\paragraph{The key to convergence: functional inequalities} The central question is that of the convergence properties to a global minimizer. In particular, an essential question is: what are assumptions ensuring that the global minimization $\min_M U$ problem is solved by the above?
\smallskip
In simulated annealing, the essential tool for convergence is the log-Sobolev inequality of Holley, Kusuoka and Stroock \cite{MR995752}. We also recalled that this inequality can be advantageously thought as a \L ojasiewicz inequality when considering the problem along the gradient system angle (c). We follow therefore the same protocol but in a reverse way: we formally write \L ojasiewicz inequalities using the Monge-Kantorovich formalism, which reveals in turn the functional inequalities we would like to have at our disposal. This leads us to consider:
\begin{equation}\label{ineqf}
\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu_\beta)\vert^2\rho\, d\ell\geq c(\beta)\,\Omega\left(\int_M\varphi(\rho)-\varphi(\mu_\beta)-\varphi'(\mu)(\rho-\mu_\beta)\, d\ell\right)
\end{equation}
where $c,\Omega:(0,+\infty)\to \mathbb R_+$ are positive functions having specific properties and where $\mu_\beta$ is the unique stationary measure of $\mathcal U_\beta$.
We are at the heart of this paper and our central result: proving such a functional inequality under adequate assumptions. Our result holds in compact one-dimensional manifolds for power-like potential function $\varphi$. As a consequence we obtain a full convergence result of our global methods on compact one-dimensional manifolds. We also evidence the general mechanisms of global convergence and for completeness we sketch the form that operational algorithms could take.
\smallskip
Apart from the interest of our work for optimization, we believe that it raises important questions and hopes on the validity domain of the family of inequalities in \eqref{ineqf}. Positive outcomes would lead to new results in optimization and in other fields.
\paragraph{Related works} The quantitative comparison between entropy-type functional and its time derivative, often called the entropy production or dissipation, dates back to \cite{zbMATH04157608}, where it was exploited through the logarithmic Sobolev inequality of \cite{MR995752}. Using Otto's formalism this can be in turn reinterpreted as an approach \`a la \L ojasiewicz \cite{blanchet18}.
Equation \eqref{pdephi} may be seen as a formal generalization of porous media equation and fast diffusion equations -- which correspond to the case when $\varphi$ is a potential. These have been studied by several authors using the Monge-Kantorovich framework, see, e.g.\, Otto \cite{otto01},
and {Carrillo}, {McCann} and {Villani} \cite{zbMATH02109873,zbMATH05009584}.
Their asymptotic analysis is through the Bakry-Emery method \cite{MR889476}: it consists in the second-order time differentiation of the entropy. Contrary to \cite{MR995752,zbMATH04157608} and our current approach, this approach requires convexity which makes it unsuitable for general global minimization.
The article of Iacobelli, Patacchini and Santambrogio \cite{iacobelli:hal-01860062} is also connected to our approach since they consider ultrafast diffusion equations which corresponds to a negative exponent $m$ in \eqref{varphim}. However, the hypocoercive bounds they obtain do not seem well-suited to extensions to time-inhomogeneous situations, since they do not lead to a differential inequality satisfied by the entropy-like functionals.
The uniqueness of the stationary measure, i.e. of the minimizer of $\mathcal U_\beta$, is not a new result, it can be found in, e.g.,\
{Carrillo}, {J\"ungel}, {Markowich}, {Toscani} and {Unterreiter} \cite{zbMATH01661610}. For the sake of completeness, we provide a proof in the next section.
Let us conclude by mentioning a few works using non-linear diffusion where non-linearities generally affects the drift coefficient but not the diffusion term as here. In Eberle, Guillin and Zimmer \cite{eberle:hal-01334806}, {Carrillo}, the authors use coupling techniques, {Gvalani}, {Pavliotis} and {Schlichting} \cite{zbMATH07169279} treat the case of interaction potentials while Delarue and Tse \cite{delarue2021uniform} consider chaos propagation.
\section{Presentation of the problem}\label{LeProbleme}
\subsection{A family of relaxations in the probability space $\mathcal P(M)$}
Consider the non-convex minimization problem:
\begin{equation}
\begin{array}{c}
\hbox{Find a global minimizer of $U:M\to\mathbb R$ on a compact Riemannian manifold $M$.
\end{array}
\tag{$\mathcal P$}\end{equation}
Denote by $d$ the distance on $M$, and $\ell$ the natural Riemannian measure. Up to a normalization factor, assume $\ell(M)=1$. Let $\P(M)$ be the space of probability measures on $M$ equipped with the Monge-Kantorovich distance defined through
\begin{equation*}
\mathcal{W}^2_2(\mu,\nu)=\inf \left\{\int_M d^2(x,y)\, p(dx,dy):\mbox{$p$ is a coupling of $\mu$ and $\nu$ on $M^2$}\right\},\label{monge}
\end{equation*}
for any $\mu,\nu$ in $\P(M)$. The extreme values of $U$ play a special role in our approach, one defines
\begin{equation}\label{osc}
\mathrm{osc}(U):=\max_M U-\min_M U,
\end{equation}
which we may assume positive --since otherwise the problem would be trivial.
\smallskip
We make the following regularity assumptions:
\smallskip
\noindent
{\bf Assumption (A).} The manifold $M$ and the function $U:M\to [0,+\infty[$ are of class $C^2$.\\
The $C^2$ regularity assumptions are simple means to obtain existence results for the gradient evolution in the Monge-Kantorovich space as in \cite{ags,ferreira18}
\smallskip
\noindent
We embed our problem in $\P(M)$ and consider $\mathcal U[\rho]:=\int_MU\rho$ so that
$$\min_{\P(M)} \mathcal U=\min U.$$
Let $\beta>0$, we introduce a penalized relaxation of $U$ in the metric space $(\P(M),\mathcal{W}_2)$ through
\begin{equation}\label{penalty}
\mathcal U_\beta[\rho]=\beta \mathcal U[\rho] + \mathcal H[\rho]=\beta \int_M U\rho\,d\ell + \mathcal H[\rho]
\end{equation}
where
\begin{equation*}
\displaystyle\mathcal H[\rho]=
\left\{\begin{aligned}
&\int_{M}\varphi(\rho) \, d\ell &&\mbox{ if $\rho$ is absolutely continuous w.r.t the Riemannian measure}\\
& +\infty &&\mbox{ otherwise,}
\end{aligned}\right.
\end{equation*}
with $\varphi:[0,+\infty)\to \mathbb R_+$ is strictly convex and $C^2$ on $(0,+\infty)$. Up to a multiplicative factor, the first term in \eqref{penalty} is the classical relaxation of $U$ within the probability space over $M$. On the other hand, as in simulated annealing, the second term acts as a penalization forcing the minimizer to be unique and to have a density with respect to $\ell$ (see Lemma~\ref{varphiprime}).
\begin{remark} {\rm (a) (Power-like penalizations) A strong focus will be put on the class of power-like functions. For any $m\in(0,+\infty)\setminus\{1\}$, define the convex function $\varphi_m\,:\,\mathbb R_+\rightarrow \mathbb R_+$ via
\bqn{varphim}
\forall\ r\geq 0,\qquad
\varphi_m(r)&\df&
\f{r^m-1-m(r-1)}{m(m-1)} \end{eqnarray}
Let us observe that $\varphi_m$ is a strictly convex function, $C^2$ on $(0,+\infty)$ and such that $\varphi_m(1)=0$, $\varphi'_m(1)=0$ and $\varphi''_m(1)=1$ for every admissible $m$.
By the Taylor-Lagrange formula, we deduce that $\varphi_m$ is always positive, except at 1.
The convex function $\varphi_1\,:\,\mathbb R_+\rightarrow \mathbb R_+$ defined by $\varphi_1(r):=r\ln(r)-(r-1)$ is recovered as the limit of $\varphi_m$ when $m$ goes to $1$ and corresponds to the Boltzmann entropy. Hereafter, we will in particular consider functions $\varphi$ that are constructed by gluing together two different functions $\varphi_m$ at $1$.\\
(b) (Regularity of the penalization)
Observe as well, from \cite[Theorem~9.3.9, p.212]{ags} and \cite[Proposition~9.3.2, p.210]{ags}, that the function $\mathcal H$ is lower semi-continuous for the Monge-Kantorovich distance, and geodesically convex in $\mathcal P(M)$. }
\end{remark}
\medskip
The penalization approach we adopt is through the one-parameter family of problems
\begin{equation}
\qquad {\rm val} (\mathcal P_{\beta}):=\inf_{{\cal P}(M)}\mathcal U_\beta \tag{$\mathcal P_{\beta}$}
\end{equation}
where the parameter $\beta>0$ is the inverse of a penalization parameter or the inverse of ``the temperature'' according to the simulated annealing literature. It ultimately tends to $\infty$, and one has an elementary but important fact:
\begin{proposition}[A global optimization principle] Assume {\rm (A)} and that $\varphi:[0,+\infty)\to \mathbb R_+$ is strictly convex and $C^2$ on $(0,+\infty)$. Then
\begin{itemize}
\item[(i)] $\displaystyle \lim_{\beta\to+\infty}\frac1\beta \,{\rm val} (\mathcal P_{\beta})\, = \,\min_M U,$
\item[(ii)] if $\mu_\beta$ is a sequence of solutions to $(\P_\beta)$, the weak* limit points of $\mu_\beta$ have a support concentrated on the set of minimizers of $U$
\end{itemize}
\end{proposition}
\begin{proof}
As a first observation, it is clear that
$$
\min_{\P(M)}\int_MU\rho=\min_M U.
$$
Fix $\epsilon>0$ and choose $\rho_\epsilon$ to be an $\epsilon$-minimizer of $\mathcal U[\rho]=\int_MU\rho$. The above observation yields $\mathcal U[\rho_\epsilon]\leq\min _M U+\epsilon$. Since ${\rm dom}\, \mathcal H$ contains smooth densities, one can also assume that $\mathcal H(\rho_\epsilon)$ is finite. Take $\beta>0$ and
let $\mu_{\beta,\epsilon}$ be an $\epsilon$-solution to $(\P_\beta)$, that is
$$\frac1\beta\mathcal U_\beta [\mu_{\beta,\epsilon}]=\mathcal U[\mu_{\beta,\epsilon}]+1/\beta \mathcal H[\mu_{\beta,\epsilon}]\leq\mathcal U(\rho)+1/\beta \mathcal H[\rho] +\epsilon $$ for all $\rho$. Thus choosing $\rho=\rho_\epsilon$ yields
$$\frac1\beta\mathcal U_\beta[\mu_{\beta,\epsilon}]\leq \min_M U+1/\beta \mathcal H[\rho_\epsilon]+2\epsilon.$$
Letting $\beta$ goes to infinity yields $$\limsup_{\beta\to\infty}\frac1\beta\mathcal U_\beta[\mu_{\beta,\epsilon}]\leq \min_M U+2\epsilon.$$ Whence $\displaystyle \limsup_{\beta\to\infty,\epsilon\to0}\frac1\beta\mathcal U_\beta[\mu_{\beta,\epsilon}]\leq \min_M U$. Since $\mathcal U_\beta\geq \min_M U$ by positivity of $\varphi$, (i) follows readily.
\noindent
Let us prove (ii). Let $\rho$ be a limit point of $\mu_{\beta}$ for the weak* topology. Since $U$ is continuous, its mixed extension $\mathcal U$
is continuous for the weak* topology and thus $\lim_{\beta\to \infty} \mathcal U[\mu_\beta]= \mathcal U[\rho]$. On the other hand, by positivity of $\varphi$ one has
$\mathcal U\leq \frac1\beta\mathcal U_\beta$, thus (i) gives $\limsup_{\beta\to \infty} \mathcal U[\mu_\beta]\leq \min_M \mathcal U$ whence $\mathcal U[\rho]\leq \min_M U$ and (ii) follows.\end{proof}
\medskip
\noindent
Observe that
\begin{equation}
\mathrm{gap} (\beta)=\frac1\beta\inf_{\P(M)} \mathcal U_\beta-\min_M U\label{app}
\end{equation}
tends to zero when $\beta$ tends to $+\infty$ by the previous result. The quantity $\mathrm{gap} (\beta)$ measures the approximation abilities of the problem $(\P_\beta)$ with respect to the initial problem $\min_M U$.
\subsection{Variational considerations and stationary measures}\label{s2.2}
Let us analyze the first-order conditions for the above problem $(\mathcal P_{\beta})$ through the lenses of the Monge-Kantorovich metric. We shall use freely the definition of Monge-Kantorovich subgradients and related objects. As they are only central to our understanding but not to our proofs, we refer to \cite{ags} for details. The subgradient of $\mathcal U_\beta$ with respect to the Monge-Kantorovich metric has a domain contained in $L^1(M)$ and is formally given by
\begin{eqnarray*}
{\rm grad}_\M\; \mathcal U_\beta\,[\rho]&=&-\mathrm{div}(\rho(\beta \nabla U+\nabla \varphi'(\rho)))\end{eqnarray*}
for any admissible $\rho$ in $L^1(M)$.\\
Stationary solutions of \eqref{pdephi}, with $\beta_t\equiv\beta$, are thus probability densities $\mu$ solution to
\bqn{dive}
\mathrm{div}(\mu(\beta\nabla U+\nabla \varphi'(\mu)))&=&0,
\end{eqnarray}
which is to be understood in the standard weak sense. By integration by parts, we have for $f \in C^2(M)$,
\begin{eqnarray*}
\int_M \mathrm{div}(\mu\nabla \varphi'(\mu)) f\, d\ell
&=&-\int_M \lt\langle \mu \nabla\varphi'(\mu),\nabla f\rt\rangle\, d\ell\\
&=&-\int_M \mu\varphi''(\mu)\lt\langle \nabla \mu,\nabla f\rt\rangle\, d\ell\\
&=&-\int_M \lt\langle \nabla(\mu\alpha(\mu)),\nabla f\rt\rangle\, d\ell\\
&=&\int_M \mu\alpha(\mu) \triangle f\, d\ell\\
&=&\int_{\{\mu>0\}} \alpha(\mu)\triangle f\, \mu d\ell
\end{eqnarray*}
where the function $\alpha\,:\,(0,+\infty)\rightarrow\mathbb R_+$ is given by
\begin{eqnarray*}
\forall\ r> 0,\qquad \alpha(r)&\df &\f1r\int_0^r s\varphi''(s)\, ds.
\end{eqnarray*}
On the other hand, we have
\begin{eqnarray*}
\int_M \mathrm{div}(\mu \beta\nabla U) f\, d\ell=-\int_M \lt\langle \beta \nabla U, \nabla f\rt\rangle \mu\,d\ell.
\end{eqnarray*}
Finally, $\mu$ is a stationary solution to \eqref{dive} if
\begin{eqnarray*}
\forall\ f\in\mathcal{C}^2(M),\qquad \int_{\{\mu>0\}} L_\mu[f]\, \mu d\ell&=&0\end{eqnarray*}
with
\begin{eqnarray*}
L_\mu[f]&=&\alpha(\mu)\triangle f-\lt\langle \beta \nabla U,\nabla f\rt\rangle.\end{eqnarray*}
\begin{remark}[Infinitesimal generator]{\rm
Observe that the choice $\varphi=\varphi_1$ leads to $\alpha(r)=1$, that is to
$$
L_\mu[f]=\triangle f-\lt\langle \beta \nabla U,\nabla f\rt\rangle.
$$
This is the infinitesimal generator of the classical (overdamped) Langevin SDE. Up to a multiplicative constant, this is the only choice for the operator $L_\mu$ to be independent of $\mu$.}
\end{remark}
Let us apply formally the above relationship to the function $f=\beta U +\varphi'(\mu)$, assuming here that the stationary density $\mu$ is smooth enough. We obtain\footnote{One may observe that the first term in \eqref{grad} is the squared norm of the Monge-Kantorovich gradient of $\mathcal U_\beta$ evaluated at $\mu$.}
\begin{align}\label{grad}
\int_{\{\mu>0\}} |\nabla (\beta U +\varphi'(\mu))|^2\mu\,d\ell &=-\int_{\{\mu>0\}} L_\mu[\beta U +\varphi'(\mu)]\, \mu d\ell\\\notag
&=0,
\end{align}
where the last equality comes from the stationarity of $\mu$. Therefore, $\beta U + \varphi'(\mu)$ is constant on every connected component of the set $\{\mu>0\}$. We analyze this condition in the next paragraph.
\subsection{Uniqueness of the stationary density}
The main ingredient to study the uniqueness of the stationary density is the relation
\bqn{cO}
\beta U+\varphi'(\mu)&=&c,
\end{eqnarray}
where the constant $c$ depends on the considered connected component of the support~$\{\mu>0\}$.
\smallskip
Define $I\df\varphi'((0,+\infty))$ and denote by $\psi\,:\, I\rightarrow (0,+\infty)$ the inverse of $\varphi'$.
\smallskip
\begin{lem}[Existence and uniqueness of the minimizer of $(\P_\beta)$]\label{varphiprime}
Assume $\varphi'(0)=-\infty$, then there exists an unique stationary density $\mu_\beta$ solution to \eqref{dive}. Moreover,
\begin{itemize}
\item[(i)]
$\mu_\beta$ is positive everywhere on $M$ and is characterized by the relation
\bqn{rhopsicU}
\mu_\beta&=&\psi(c^*-\beta U),\end{eqnarray}
where $c^*$ is a normalization parameter characterized by the condition\\
$$\int_M \psi(c^*-\beta U)\, d\ell=1.$$
\item[(ii)] $\mu_\beta$ is the global minimizer of $\mathcal U_\beta$.
\end{itemize}
\end{lem}
\begin{proof}
Let us start with (i) and by showing that a stationary density $\mu$ is everywhere positive. Towards a contradiction, assume that the set $\{\mu=0\}$ is non-empty and let $M_1$ be a connected component of the open set $\{\mu>0\}$ with $\partial M_1\neq \emptyset$. Let $(x_n)_{n\in\mathbb{N}}$ be a sequence of elements of $M_1$ converging to a point of the boundary $\partial M_1$. According to \eqref{cO}, we have for every $n\in\mathbb{N}$
\begin{eqnarray*}
\beta U(x_n)+\varphi'(\mu(x_n))&=&c_{M_1}\end{eqnarray*}
where the left-hand side term converges to $-\infty$ which is absurd.
Therefore, $M_1=M$ and equation \eqref{cO} is valid everywhere on $M$. Set $c\df c_{M_1}$.
Being strictly convex, the function $\varphi'$ is one-to-one and onto between $(0,+\infty)$, and by definition of $\psi$,
equation \eqref{cO} rewrites:
\bqn{rhopaO}
\mu&=& \psi(c^*-\beta U)\end{eqnarray}
\par
\noindent
As $\mu$ is a density function, we must have
\bqn{norma}
\int_M \psi(c^*-\beta U)\, d\ell&=& 1.\end{eqnarray}
Since $\psi$ is strictly increasing and satisfies $\lim_{\inf I} \psi=\lim_{-\infty} \psi=0$ and $\lim_{\sup I} \psi=+\infty$, there is a unique value $c^*\in\mathbb R$ that satisfies equation \eqref{norma}.
We have
$$
\mu=\psi(c^*-\beta U)$$
which ends not only the proof of the uniqueness of the stationary density $\mu$ but also gives its existence and its explicit form. To see (ii) and that $\mu$ is the global minimizer, we observe that
\begin{eqnarray*}
\mathcal U_{\beta}[\rho]-\mathcal U_\beta[\mu]&=&\int_M(\varphi(\rho)-\varphi(\mu))\, d\ell+\int_M \beta U\,(\rho-\mu) \,d\ell\\
&=& \int_M(\varphi(\rho)-\varphi(\mu))\, d\ell+\int_M (c^*-\varphi'(\mu))\,(\rho-\mu) \,d\ell\\
&=& \int_M\left( \varphi(\rho)-\varphi(\mu)-\varphi'(\mu)\right)\,(\rho-\mu) \,d\ell,
\end{eqnarray*}
which is positive whenever $\rho\neq\mu$ by strict convexity of $\varphi$.
\end{proof}
\smallskip
When no confusion can occur we simply write $\mu$ for $\mu_\beta$. We gather the assumptions we need regarding $\varphi$ within
\medskip
\noindent
{\bf Assumption (B).} $\varphi:[0,+\infty)\to\mathbb R$ is convex, twice differentiable on $(0,+\infty)$ with $\varphi''>0$, and satisfies
$\varphi'(0)=-\infty$. \\
\smallskip
\subsection{Global minimization dynamics} \label{s:globdyn}
\paragraph{Minimizing dynamics} We are now in position to provide dynamical systems meant to solve the problem $(\P)$. Inspired by Holley, Kusuoka and Stroock's approach to simulated annealing \cite{MR995752}, as it was simplified in \cite{zbMATH04157608}, and using as well the gradient view provided by Otto's formalism, we consider, formally, the gradient system
\begin{equation}\label{grad-sys}
\quad\frac{{\rm d}}{{\rm d}t} \rho(t)=-{\rm grad}_\M \,\mathcal U_{\beta_t}\,[\rho(t)] \mbox{ a.e. on $\mathbb R_+$,} \quad \rho(0)=\rho_0,
\end{equation}
where the term $\beta_t$ is a $C^1$ positive time-varying parameter and where we use Newton's notation, $\frac{{\rm d}}{{\rm d}t} \rho$ here, for time derivatives. The initial distribution $\rho_0$ is chosen in the domain of $\mathcal U_\beta$, that is in the domain of $\mathcal H$.
The time dependent density $\rho$ turns out to satisfy the following partial differential equation
\bqn{evol0}
\qquad\frac{{\rm d}}{{\rm d}t} \rho&=&\mathrm{div}(\rho \,(\beta_t\nabla U))+\mathrm{div}(\rho(\nabla \varphi'(\rho))), \quad \rho(0)=\rho_0.\end{eqnarray}
The time-varying parameter $\beta_t$ is traditionally interpreted as an inverse of a temperature which typically cools down, i.e., \begin{equation}\label{infty}\lim_{t\to \infty}\beta_t=+\infty\end{equation} Here we also interpret this parameter as the inverse of a penalty term echoing the static formula \eqref{penalty}.
\par
\begin{remark}{\rm (a) (Simulated annealing) When $\varphi=\varphi_1$, \eqref{evol0} boils down to the famous simulated annealing dynamics
\bqn{siman}
\frac{{\rm d}}{{\rm d}t} \rho &=&\beta_t\mathrm{div}(\rho\nabla U)+\Delta\rho\end{eqnarray}
which, by a famous ``nonconvex'' extension of the log-Sobolev inequality, due to Holley, Kusuoka and Stroock \cite{MR995752}, is known to generate measures concentrating on the set of global minimizers of $U$ whenever the temperature schedule is finely tuned.\\
(b) (Porous media) Taking $U$ constant and $\varphi=\varphi_m$ in \eqref{evol0} with $m >0$, the dynamic cor\-res\-ponds to the porous media equation
$$
\frac{{\rm d}}{{\rm d}t} \rho=m\Delta \left( \rho^m \right).
$$
The case $m>1$ refers to the slow diffusion case while the case $m<1$, for which $\varphi'(0)=-\infty$ refers to the fast diffusion situation, (see Vazquez \cite{vazquez17} and Otto \cite{otto01})}.
\end{remark}
\paragraph{Existence results and evolution equations} Following the pioneering work of \cite{ags},
the non-autonomous theory for Monge-Kantorovich gradient flows has recently been developed in \cite{ferreira18}. In the line of \cite[Theorem 4.4, Theorem 5.4]{ferreira18} and the existence results of \cite{iacobelli:hal-01860062}, we {\em assume} that \eqref{grad-sys} and \eqref{evol0}, have a common unique solution curve $t\mapsto\rho(t)$ in $(\P(M),\mathcal{W}_2)$, which satisfies in addition
\begin{align}
& t\mapsto \mathcal U_{\beta_t}[\rho(t)] \mbox{ and $t\mapsto\rho(t)$ are absolutely continuous,}\label{ac0}\\
& \frac{{\rm d}}{{\rm d}t}\,\mathcal U_{\beta_t}[\rho(t)]=\int_M(\varphi'(\rho)+ \beta_t U)\frac{{\rm d}}{{\rm d}t} {\rho}\, d\ell+\frac{{\rm d}}{{\rm d}t} {\beta_t}\int_MU\rho\,d\ell,\label{ac1}
\end{align}
where the time derivatives are taken for almost all times.
\paragraph{Functional inequalities}
Under hypothesis {\bf (A), (B)} and some extra-assumptions on $\varphi$ related to the geometry of the penalized cost, we intend to prove that the dynamics \eqref{grad-sys}-\eqref{evol0} has global optimizing properties, in the sense that the global cost
\begin{align}
\mathcal U_{\beta_t}[\rho]=\int_M\left(\varphi(\rho)+ \beta_t U\rho\right) \,d\ell
\end{align}
evaluated along the trajectory $t\mapsto\rho(t)$ given by \eqref{grad-sys}-\eqref{evol0} should converge to the value of $(\mathcal P)$, i.e.,
$$\lim_{t\to+\infty}\mathcal U_{\beta_t}[\rho]=\mbox{val} ( \mathcal P)=\min_M U.$$
As it is customary in the analysis of PDEs the key to convergence is given by ``entropy-energy'' or ``entropy-production'' functional inequalities. In the ``gradient or in the optimization world'' , these can often be seen as \L ojasiewicz type inequalities, see \cite{blanchet18} and references therein. They connect the cost $\mathcal U_\beta$ to the norm of its gradient $\|{\rm grad}_\M\mathcal U_\beta\|$ and to the constant $\beta$:
\begin{equation}
\label{lojalike}\|{\rm grad}_\M\mathcal U_\beta\|^2\geq c(\beta)\,\Omega\Big( \mathcal U_\beta(\rho)-\min \mathcal U_\beta\Big),
\end{equation}
where $c,\Omega:(0,+\infty)\to\mathbb R_+$ are positive functions, with $\Omega$ being increasing and null at zero.
\noindent
Reexpressing $\mathcal U_\beta$ by means of its stationary density \eqref{rhopsicU} gives
\begin{eqnarray*}
\mathcal U_{\beta}(\rho)-\mathcal U_{\beta}(\mu)&=&\int_M\varphi(\rho)-\varphi(\mu)\, d\ell+\int_M \beta U\,(\rho-\mu) \,d\ell\\
&=& \int_M\varphi(\rho)-\varphi(\mu)\, d\ell+\int_M (c^*-\varphi'(\mu))\,(\rho-\mu) \,d\ell\\
&=& \int_M\left[\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)\,(\rho-\mu)\right] \,d\ell.
\end{eqnarray*}
Because $\varphi$ is convex, we obtain $\mathcal U_{\beta}(\rho) \ge \mathcal U_{\beta}(\mu)$ thus $\mathcal U_{\beta}(\mu)=\min\mathcal U_\beta$.\\
As a consequence, inequality \eqref{lojalike} writes
\begin{equation}\label{KL}
\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell\geq c(\beta)\, \Omega\left(\int_M\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)\, d\ell\right)
\end{equation}
where $c,\Omega:(0,+\infty)\to \mathbb R_+$ are positive functions.
A typical example is given by the log-Sobolev inequality of Holley, Kusuoka and Stroock which can be written as
\begin{equation}\label{logsob}
\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell\geq C_{\rm \tiny{HKS}}(\beta)\left(\int_M\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)\, d\ell\right)
\end{equation}
where $\varphi(r)=\varphi_1(r)=r\ln(r)-r+1$ and
\begin{eqnarray*}
\lim_{\beta\rightarrow+\infty}\f1\beta\ln(C_{\rm \tiny{HKS}}(\beta))&\geq &-\mathrm{osc}(U)\end{eqnarray*}
(see \cite{MR995752} for the precise description of the l.h.s.\ in terms of the landscape of $U$).
\par
When $U$ is convex and $\varphi$ is power-like, one can also recover Gagliardo-Nirenberg inequalities of \cite{delpino02}, see \cite{blanchet18} for connections with \L ojasiewicz inequalities.
\medskip
\paragraph{Convergence mechanisms for a fixed penalization parameter} As previously mentioned, we adapt the approach of \cite{zbMATH04157608} developed in the Boltzmann entropy case ($\varphi=\varphi_1$) to our generalized class of relaxations.
\smallskip
Let us provide a first account of the general method through the constant parameter case.
For $\rho\in \P(M)$ having a density with respect to $\ell$, set
\begin{align}
& \mathcal I[\rho]=\mathcal U_\beta(\rho)-\min_{\P(M)}\mathcal U_\beta\\
& \mathcal J[\rho]=\int_\mathbb{T}\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell
\end{align}
where the quantities may take infinite values and where $\mu$ is the unique stationary density (see Lemma \ref{varphiprime}), so that $\mathcal I[\mu]=0$.
\medskip
At this stage, we do not assume that $\beta>0$ depends on time.
\medskip
By time differentiation, using \eqref{ac0}-\eqref{ac1} and the evolution equation, we obtain
\begin{eqnarray*}
\frac{{\rm d}}{{\rm d}t}\, \mathcal I[\rho(t)]&=&\int_M\varphi'(\rho)\frac{{\rm d}}{{\rm d}t} {\rho}\, d\ell+\int \beta U\frac{{\rm d}}{{\rm d}t} {\rho}\, d\ell\\
&=&\int_M(\varphi'(\rho)+\beta U)\frac{{\rm d}}{{\rm d}t} {\rho}\, d\ell\\
&=&\int_M(\varphi'(\rho)+\beta U)\mathrm{div}(\rho(\beta \nabla U+\nabla \varphi'(\rho)))\, d\ell\\
&=&-\int_M\nabla (\varphi'(\rho)+\beta U)(\beta \nabla U+\nabla \varphi'(\rho)))\, \rho d\ell\\
&=&-\int_M\vert \nabla \varphi'(\rho)+\beta\nabla U\vert^2\, d\rho.
\end{eqnarray*}
Whence, if we have some inequality \`a la \L ojasiewicz like \eqref{KL} (as for instance the log-Sobolev inequality of Holley, Kusuoka and Stroock \cite{MR995752} when $\varphi=\varphi_1$), we derive a differential inequality for the one-variable function $\mathcal I[\rho]$,
\begin{equation} \label{liap}
\frac{{\rm d}}{{\rm d}t}\, \mathcal I[\rho(t)] \: \leq \: -c(\beta)\,\Omega(\mathcal I[\rho(t)])
\end{equation}
This implies in turn that $\mathcal I[\rho(t)]$ converges. If this limit was not zero, the fact that $\Omega$ is positive out of $0$ would imply, through \eqref{liap}, that the decrease-rate would be perpetually lower than a negative constant, which is absurd. This allows to prove that $\mathcal I[\rho(t)]$ tends to zero as $t\to\infty$. We thus have proved the first part of:
\begin{theorem}[Convergence with a non-vanishing penalty parameter] Assume that $\bf (A), (B)$ are satisfied, and that there exist $c>0$, $\Omega:\mathbb R_+\to\mathbb R_+$ increasing, such that an inequality of the type
\begin{eqnarray}\label{fineq}
\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell & \geq & c\,\Omega\left(\int_M\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)\, d\ell\right),
\end{eqnarray}
holds true whenever $\rho$ is measurable and the left hand side is finite. Then
\begin{enumerate}[(i)]
\item $\mathcal U_\beta[\rho(t)]\to \min \mathcal U_\beta$.
\item If moreover $\Omega(s)=\Theta(s^{2\theta})$ at $0$, with $\theta\in (0,1)$, then $\rho(t)$ tends to $\mu_\beta$ for the Monge-Kantorovich metric, i.e., for the weak* topology.
\end{enumerate}
\end{theorem}
\begin{proof} Item (i) is already proved. For (ii), from $\Omega(s)=\Theta(s^{2\theta})$ holds, we deduce that the lower semicontinuous function $\mathcal U_\beta$ satisfies a \L ojasiewicz inequality as in \cite[Theorem 2]{blanchet18}, so that one may assert that the curve $t\mapsto\rho(t)$ has a finite Monge-Kantorovich length; convergence rates depending on $\theta$ are also available. \end{proof}
\medskip
What are the conditions for the above inequality \eqref{fineq} to be valid, is a delicate open question. The subject of the next section, and the central result of this paper, is to establish such inequalities for one dimensional compact manifolds and a family of power-like potentials~$\varphi$.
\section{A functional inequality on the circle}
\subsection{Main theorem}
For $m\in (0,1/2)$, set
\bqn{varphi}
\forall\ r\geq 0,\qquad
\varphi_{m,2}(r)&\df&\left\{\begin{array}{ll}
\varphi_{m}(r)&\hbox{ if $r\in(0,1]$,}\\[3mm]
\varphi_{2}(r)=\frac{(r-1)^2}{2}&\hbox{ if $r\in(1,+\infty)$.}
\end{array}
\right.\end{eqnarray}
The function $\varphi_{m,2}$
is convex on $[0,+\infty)$ and $\mathcal{C}^2$ on $(0,+\infty)$. The latter property is a consequence of the fact that
\begin{eqnarray*}
\varphi_m(1)&=&0\\
\varphi'_m(1)&=&0\\
\varphi''_m(1)&=&1.\end{eqnarray*}
Observe also that $\varphi_{m,2}'$ is concave on $(0,+\infty)$ with $\varphi_{m,2}'(0)=-\infty$, so that Lemma \ref{varphiprime} applies. Let us recall that the unique solution of $(\P_\beta)$ is denoted by $\mu=\mu_\beta$ and that $\psi_{m,2}=[\varphi_{m,2}']^{-1}$.
\medskip
This section is devoted to the proof of:
\begin{theorem}[A new functional inequality on the circle]\label{funcineq}
Assume that $M$ is the circle $\mathbb{T}\df\mathbb R/(L\mathbb{Z})$ of perimeter $L>0$ endowed with its usual Riemannian structure.
Then there exists a constant $c(\beta)$, depending on $\mu_{\min}$ and $L$,
such that
for any measurable density $\rho$ on $\mathbb{T}$
\begin{eqnarray*}
\int_\mathbb{T}\vert \nabla \varphi_{m,2}'(\rho)-\nabla\varphi_{m,2}'(\mu)\vert^2\rho\, d\ell \geq c(\beta)\,\Omega\left(\int_\mathbb{T}\varphi_{m,2}(\rho)-\varphi_{m,2}(\mu)-\varphi_{m,2}'(\mu)(\rho-\mu)\, d\ell\right)
\end{eqnarray*}
where
\begin{eqnarray*}
c(\beta) & = & \displaystyle O\left(\beta^{\frac{-3(2-m)}{1-2m}}\right)\\
& &\notag \\
\Omega(r)&=&
\left\{\begin{aligned}
&r^{\f32}&& \hbox{ if $r\in[0,1)$}\\
\\
& r^{\f{1-2m}{2(1-m)}} && \hbox{ if $r\geq 1$}.
\end{aligned}\right.
\end{eqnarray*}
\end{theorem}
\begin{corollary}[An inequality \`a la Talagrand]\label{c:tal}
Under the assumptions of the previous theorem,
for any measurable density $\rho$ on $\mathbb{T}$,
\begin{eqnarray*}
\int_\mathbb{T}\varphi_{m,2}(\rho)-\varphi_{m,2}(\mu)-\varphi_{m,2}'(\mu)(\rho-\mu)\, d\ell \geq d(\beta)\,\omega\Big({\mathcal W}_2(\rho,\mu)\Big)
\end{eqnarray*}
where $d(\beta)= \displaystyle O\left(\beta^{\frac{-3(2-m)}{1-2m}}\right)$, and
\begin{eqnarray}
\omega(r)&=&
\left\{\begin{aligned}
&8/5 \; r^{\f58}&& \hbox{ if $r\in[0,1)$}\\
\\
& 4(1-m)/(3-2m)\; r^{\f{3-2m}{4(1-m)}} && \hbox{ if $r\geq 1$}.
\end{aligned}\right.
\end{eqnarray}
\end{corollary}
\bigskip
The rest of this section is devoted to the proof of this theorem (the corollary will follow easily using a generalization of Otto-Villani theorem \cite{blanchet18}). Most of the intermediary results we provide are valid for a general compact $C^2$ manifold, thus, unless otherwise stated, we assume for the moment that $M$ is arbitrary.
In the remaining subsections of the present section, for the sake of simplicity, we shall often write $\varphi=\varphi_{m,2}$ and $\psi_{m,2}=\psi$.
\subsection{Some estimates}
Let us define the positive quantities
\bqn{boundsformu}
\mu_{\min}\df \min_M\mu\,\hbox{ and }\,
\mu_{\max}\df\max_M\mu.
\end{eqnarray}
\begin{pro}[Bounds for the stationary measure]\label{controlwedge}
We have, for any $0<m<1$,
\begin{eqnarray*}
(1+(1-m)\beta \mathrm{osc}(U))^{\frac{1}{m-1}} \leq \ \mu_{\min}\ \leq \ \mu_{\max}\ \leq \beta\mathrm{osc}(U)+1.
\end{eqnarray*}
where we recall that $\mathrm{osc}(U)=\max_MU-\min_MU$.
\end{pro}
\begin{proof}
The stationary measure $\mu$ satisfies for every $x \in M$,
$\mu(x)=\psi(c^*-\beta U(x))$ for some real constant $c^*$ and with $\beta >0$ (recall \eqref{rhopaO}). Because $\psi$ is nondecreasing , we have
$$
\forall x \in M,\, \psi(c^*-\beta \max_MU)\le \mu(x) \le \psi(c^*-\beta \min_M U).
$$
Integrating with respect to the probability measure $\ell$, we obtain
$$
\psi(c^*-\beta \max_MU)\le 1 \le \psi(c^*-\beta \min_M U).
$$
Because $\varphi^\prime$ is nondecreasing and $\varphi^\prime(1)=0$, we obtain the following bounds for the constant~$c^*$
$$
\beta \min_M U \le c^* \le \beta \max_M U.
$$
Finally, for every $x \in M$, $\psi(-\beta\mathrm{osc}(U))\le \mu(x) \le \psi(\beta \mathrm{osc}(U))$. Because $\beta \mathrm{osc}(U) \ge 0$, we have
for any $m\in(0,1)$,
$$
(1-(m-1)\beta \mathrm{osc}(U))^{\frac{1}{m-1}} \leq \ \mu_{\min}\ \leq \ \mu_{\max}\ \leq 1+\beta\mathrm{osc}(U),
$$
which ends the proof.\end{proof}
\medskip
The following formal observation is essential. When the potential function $\varphi$ is the entropy function $\varphi_1$, the density of $\rho$ with respect to $\mu$ plays a pivotal role in the establishment of the log-Sobolev inequality \eqref{logsob}, see \cite{MR995752}. In our case,
the counterpart is the function
\begin{eqnarray*}
\psi(\varphi'(\rho)-\varphi'(\mu)).\end{eqnarray*}
It is also convenient to use the quantity
\bqn{g}
g:=\ \varphi'(\rho)-\varphi'(\mu)\end{eqnarray}
so that
\bqn{rho}
\rho&=&\psi(g+\varphi'(\mu)).\end{eqnarray}
\par
\smallskip
\paragraph{An upper bound for the reduced cost $\mathcal I[\rho]$}
This necessitates three steps.
\begin{lem}\label{step1}
For any measurable density $\rho$, we have
\begin{eqnarray*}
\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)&\leq &
\varphi(\psi(g+\varphi'(\mu_{\max})))-\varphi(\mu_{\max})-\varphi'(\mu_{\max})(\psi(g+\varphi'(\mu_{\max}))-\mu_{\max})\end{eqnarray*}
\end{lem}
\begin{proof}
By definition of $g$, we have
\begin{eqnarray*}
\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)&= &
\varphi(\psi(g+\varphi'(\mu)))-\varphi(\mu)-\varphi'(\mu)(\psi(g+\varphi'(\mu))-\mu).\end{eqnarray*}
Fix $x\in M$ and consider the function $F$ defined on $(0,+\infty)$ by
\begin{eqnarray*}
\forall\ r>0,\qquad
F(r)&\df& \varphi(\psi(g(x)+\varphi'(r)))-\varphi(r)-\varphi'(r)(\psi(g(x)+\varphi'(r))-r)\end{eqnarray*}
To prove the result, it is sufficient to show that $F$ is nondecreasing .
For $r>0$, we compute
\begin{eqnarray*}
F'(r)&=&\varphi'(\psi(g(x)+\varphi'(r)))\psi'(g(x)+\varphi'(r)))\varphi''(r)-\varphi'(r)-\varphi''(r)(\psi(g(x)+\varphi'(r))-r)
\\&&
-\varphi'(r)[\psi'(g(x)+\varphi'(r)))\varphi''(r)-1]\\
&=&\left[\varphi'(\psi(g(x)+\varphi'(r)))\psi'(g(x)+\varphi'(r))-(\psi(g(x)+\varphi'(r))-r)-\varphi'(r)\psi'(g(x)+\varphi'(r))\right]\varphi''(r)\\
&=&\left[(g(x)+\varphi'(r))\psi'(g(x)+\varphi'(r))-(\psi(g(x)+\varphi'(r))-r)-\varphi'(r)\psi'(g(x)+\varphi'(r))\right]\varphi''(r)\\
&=&\left[g(x)\psi'(g(x)+\varphi'(r))-(\psi(g(x)+\varphi'(r))-r)\right]\varphi''(r)\\
&=&\left[g(x)\psi'(\varphi'(s))-(s-r)\right]\varphi''(r)
\end{eqnarray*}
where we set
$s\df \psi(g(x)+\varphi'(r))$
in the last equality. Because $\psi=(\varphi')^{-1}$, we have
\begin{eqnarray*}
\psi'(\varphi'(s))&=&\f1{\varphi''(s)}\end{eqnarray*}
and we get for $r>0$,
\begin{eqnarray*}
F'(r)&=&[g(x)-\varphi''(s)(s-r)]\f{\varphi''(r)}{\varphi''(s)}\\
&=&[\varphi'(s)-\varphi'(r)-\varphi''(s)(s-r)]\f{\varphi''(r)}{\varphi''(s)}.
\end{eqnarray*}
Because, the function $\varphi$ is convex and $\varphi'$ is concave, we have ${\varphi''(r)}/{\varphi''(s)}$ is positive and the quantity
$\varphi'(s)-\varphi'(r)-\varphi''(s)(s-r)$ is non-negative, so that $F'\geq 0$ on $(0,+\infty)$
and $F$ is thus nondecreasing .\end{proof}
\medskip
We define further
$
\xi_{\max}(s)\df\psi(s+\varphi'(\mu_{\max}))
$ for any real number $s$
and set $\rho_{\max}\df\xi_{\max}(g)$. Therefore, Lemma \ref{step1} can be rewritten
\bqn{plusjoli}
\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)\le
\varphi(\rho_{\max})-\varphi(\mu_{\max})-\varphi'(\mu_{\max})(\rho_{\max}-\mu_{\max}).
\end{eqnarray}
We are in position to give an upper bound for $\mathcal I[\rho]$.
\smallskip
\begin{lem}[An upper bound for the reduced cost $\mathcal I$]\label{step2}
For every $\rho \in \P(M)$ such that $g\in L^2(\ell)$, we have
$$
\mathcal I[\rho]\le \int_M g^2(x)\,\ell(dx).
$$
\end{lem}
\begin{proof}
Because $\varphi$ is convex, we have
\begin{eqnarray*}
\varphi(\mu)-\varphi(\rho)-\varphi'(\rho)(\mu-\rho)&\ge& 0
\end{eqnarray*}
and
\begin{eqnarray*}
\varphi(\mu_{\max})-\varphi(\rho_{\max})-\varphi'(\rho_{\max})(\mu_{\max}-\rho_{\max})&\ge& 0.
\end{eqnarray*}
Adding the latter positive quantity to the right-hand-side of equation \eqref{plusjoli} gives
\begin{eqnarray*}
0 \le \varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)&\le&
(\varphi'(\rho_{\max})-\varphi'(\mu_{\max}))(\rho_{\max}-\mu_{\max})\\
&=& g(\xi_{\max}(g)-\xi_{\max}(0)).
\end{eqnarray*}
Recalling that $\varphi$ has been constructed by gluing $\varphi_m$ and $\varphi_2$ at 1 (see equation \eqref{varphi}), we have that $\psi=(\varphi')^{-1}$ is convex and increasing with $0\le \psi' \le 1$. Therefore, $\xi_{\max}$ is convex and we have
$$
\psi'(\varphi'(\mu_{\max}))g \le \xi_{\max}(g)-\xi_{\max}(0) \le \psi'(g+\varphi'(\mu_{\max}))g.
$$
Whence,
$$
|\xi_{\max}(g)-\xi_{\max}(0)|\le |g|,
$$
which gives the desired result.\end{proof}
\paragraph{A lower bound for the squared Monge-Kantorovich gradient $\mathcal J[\rho]$} Once more several steps are necessary to obtain a bound.
Let us define the function $\theta:\mathbb R\to\mathbb R$ via
\bqn{theta}
\forall\ r\in\mathbb R,\qquad \theta(r)&=&\int_{0}^r \sqrt{\psi(s+ \varphi'(\mu_{\min})) }\, ds.\end{eqnarray}
We observe first that:
\begin{lem}[Lower bound for $\theta$]\label{secondstep}
Assume $0 < m < \frac{1}{2}$. For any $r\in \mathbb R$,
$$
|\theta(r)|\ge \frac{2}{3} \left(\min\left(\frac{1}{(C_0+(1-m))^{\frac{3}{2}-\eta}},\sqrt{\psi^\prime(\varphi'(\mu_{\min}))}\right)\right)\min(|r|^{3/2},|r|^\eta),
$$ where $C_0=1-(1-m)\varphi^\prime (\mu_{\min}) $ and $\eta:=\frac{1-2m}{2(1-m)} \in (0,1/2)$.
\end{lem}
\begin{proof}
Assume first that $r>0$. Because $\psi$ is convex, we have for every $s$, $$\psi( s+\varphi'(\mu_{\min})) \ge \mu_{\min} +\psi^\prime(\varphi'(\mu_{\min}))s.$$ But $ \mu_{\min}>0$ thus
\begin{align*}
\theta(r)&\ge\sqrt{\psi^\prime(\varphi'(\mu_{\min}))}\int_0^r \sqrt{s}\,ds\\
&=\frac{2}{3}\sqrt{\psi^\prime(\varphi'(\mu_{\min}))} r^{3/2}.
\end{align*}
Now, assume that $r<0$. By a change of variables, $t=-s$ and $\tau=-r>0$, we get
$$
-\theta(r)=\int_{0}^\tau \sqrt{\psi( \varphi'(\mu_{\min})-t) }\, dt.
$$
Remembering that $\psi$ is both convex and positive, we get, for all $t$,
$$
\psi( \varphi'(\mu_{\min})-t)\ge \psi^\prime( \varphi'(\mu_{\min})-\tau)(\tau-t).
$$
We deduce that
$$
-\theta(r) \ge \frac{2}{3} \sqrt{\psi^\prime( \varphi'(\mu_{\min})-\tau) } r^{3/2}.
$$
We shall now use the explicit form of the derivative of $\psi$ given by
\bqn{deriveepsi}
\psi'(\tau)&\df&\left\{\begin{array}{ll}
(1+(m-1)\tau)^{\frac{2-m}{m-1}}&\hbox{if $\tau\in(-\infty,0)$}\\[3mm]
1&\hbox{if $\tau\in(0,+\infty)$}
\end{array}
\right.
\end{eqnarray}
By definition of $\mu_{\min}$, we have $\mu_{\min} <1$ whenever $U$ is not constant over $M$. Therefore $\varphi'(\mu_{\min}) \le 0 <\tau$, so that
$$
\sqrt{\psi^\prime( \varphi'(\mu_{\min})-\tau) }=\left( 1-(1-m) \varphi'(\mu_{\min})+(1-m)\tau \right)^{\frac{2-m}{2(m-1)}}.
$$
Set
$$
C_0=1-(1-m)\varphi^\prime (\mu_{\min}),\,\,\,\, \alpha=\frac{2-m}{2(1-m)} \in (1,3/2)$$ and
$$ \eta=3/2-\alpha=(3(1-m)-(2-m))/(2(1-m))=(1-2m)/2(1-m).
$$
Therefore,
\begin{eqnarray*}
-\theta(r)&\ge& \frac{2}{3} \frac{\tau^{3/2}}{(C_0+(1-m)\tau)^\alpha}\\
&=& \frac{2}{3}\left( \frac{\tau^{3/2}}{(C_0+(1-m)\tau)^\alpha}\ind{0<\tau\le1} + \frac{\tau^{3/2}}{(C_0+(1-m)\tau)^\alpha}\ind{\tau>1}\right)\\
&\ge& \frac{2}{3}\left( \frac{\tau^{3/2}}{(C_0+(1-m))^\alpha}\ind{0<\tau\le1} + \frac{\tau^{3/2}}{(C_0\tau+(1-m)\tau)^\alpha}\ind{\tau>1}\right)\\
&\ge& \frac{2}{3} \frac{1}{(C_0+(1-m))^\alpha} \min(\tau^{3/2},\tau^\eta).
\end{eqnarray*}
\end{proof}
\begin{remark}[On constants]\label{borneC1}{\rm
For reasons that will appear later during the study of our global optimization method, it is useful to have a compact expression for the inverse of the constant $2/(3(C_0+(1-m))^\alpha)$ appearing in Lemma \ref{secondstep}. Using the equality $
\psi^\prime(\varphi'(\mu_{\min}))=1/\varphi^{\prime\prime}(\mu_{\min})$,
this inverse writes
$$
C_1(\mu_{\min}):=\f32\max \left[(1+(1-m)(1-\varphi'(\mu_{\min})))^{{\frac{2-m}{2(1-m)}}},\sqrt{\varphi^{\prime\prime}(\mu_{\min})} \right]>1.
$$
Observe that, as a function of $\mu_{\min}$, $C_1$ is decreasing and therefore is bounded above by
\begin{eqnarray*}
\f32\max\Big\{ \left[1+(1-m)\left[1-\varphi'\left(\left[1+(1-m)\beta \mathrm{osc}(U)\right]^{1/(m-1)}\right)\right]\right]^{\frac{2-m}{2(1-m)}}\,;\qquad&&
\\
\sqrt{\varphi^{\prime\prime}((1+(1-m)\beta\mathrm{osc}(U))^{1/(m-1)}})\Big\}, &&
\end{eqnarray*}
according to Proposition \ref{controlwedge}. When $\beta$ goes to $+\infty$, this bound behaves as $\displaystyle O\left(\beta^{\frac{2-m}{2(1-m)}}\right)$. Finally note that
\begin{equation}\label{c1}
C_1(\mu_{\min})|\theta(r)|\ge \min(|r|^{3/2},|r|^\eta)\mbox{ for all $r$.}
\end{equation}
}
\end{remark}
\vspace{0.5cm}
\smallskip
We now turn to the desired lower bound for the squared Monge-Kantorovich gradient $\mathcal J[\rho]$.
\begin{lem}[Lower bound for the squared Monge-Kantorovich gradient]
We have
\begin{eqnarray*}
\mathcal J[\rho]=\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell
&\geq & \int_M \vert \nabla \theta(g)\vert^2\, d\ell.
\end{eqnarray*}
\end{lem}
\begin{proof}
Taking into account that both $\varphi'$ and $\psi$ are nondecreasing functions, we get
\begin{eqnarray*}
\rho=\psi(g+\varphi'(\mu))&\geq &\psi( g+\varphi'(\mu_{\min}))\\
&=& (\theta'(g))^2.\end{eqnarray*}
It ensues
\begin{eqnarray*}\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell
&=&\int_M\vert \nabla g\vert^2\rho\, d\ell\\
&\geq &\int_M\vert \nabla g\vert^2(\theta'(g))^2\, d\ell\\
&=&\int_M\vert \theta'(g)\nabla g\vert^2\, d\ell\\
&=&\int_M \vert \nabla \theta(g)\vert^2\, d\ell\end{eqnarray*}
\end{proof}
\subsection{Proof of Theorem \ref{funcineq} and its corollary}
Assume for the moment that $M$ is arbitrary.
\smallskip
\noindent
In the previous section, we have proved two inequalities:
$$
\mathcal I[\rho]\le \int_M g^2 \,d\ell
\mbox{ and }
\mathcal J[\rho]\ge \int_M \vert \nabla \theta(g) \vert^2\, d\ell.
$$
To reach a conclusion, it suffices to relate the quantities $$\int_M g^2 \,d\ell \mbox{ and } \int_M \vert \nabla \theta(g) \vert^2\, d\ell.$$
\medskip
\noindent
Since $\eta\in(0,1/2)$, Lemma \ref{secondstep} and \eqref{c1} (see Remark \ref{borneC1}) gives
$$
C_1(\mu_{\min})|\theta(r)|\geq|r|^{3/2},\mbox{ when $|r|<1$ and }C_1(\mu_{\min})|\theta(r)|\geq |r|^\eta\mbox{ otherwise.}
$$ When $\beta$ is large enough $C_1(\mu_{\min})\geq 1$, we may thus write
$$
g^2 \le C_2(\mu_{\min})\max(|h|^{\f43},|h|^{\f2\eta}),
$$
where we have set $h:=\theta(g)$ and $C_2=C^{\f2\eta}_1(\mu_{\min})$. Whence, taking the supremum and integrating yields
$$
\int_M g^2\,d\ell \le C_2(\mu_{\min}) \max(||h||_\infty^{\f43},||h||_\infty^{\f2\eta}).
$$
Recall that the function $\Omega:\mathbb R_+\to\mathbb R_+$ is such that
\begin{eqnarray*}
\Omega(r)&\df&
\left\{\begin{array}{ll}
r^{\f32}&\hbox{ if $r\in[0,1)$}\\[2mm]
r^\eta&\hbox{ if $r\geq 1$.}
\end{array} \right.\end{eqnarray*}
\noindent
Let us prove that the increasing function $\Omega$ satisfies the inequality
\bqn{propOmega}
\forall x>1,\,\forall y >0,\quad \Omega(xy) \le x^{\frac{3}{2}}\Omega(y).
\end{eqnarray}
Let us consider two cases:\\
{\it First case: $y\ge 1$}. Because $x>1$, this implies $xy>1$. Thus,
$$
\Omega(xy)=(xy)^\eta=x^\eta\Omega(y)\le x^{\f32}\Omega(y).
$$
{\it Second case: $y < 1$}. When $xy>1$, the inequality follows as above.
On the other hand, if $xy<1$, we have
$$
\Omega(xy)=(xy)^{\f32}=x^{\f32}\Omega(y).
$$
Therefore, by using the fact that $\Omega$ is increasing and satisfies \eqref{propOmega}, we get
$$
\Omega(\mathcal I[\rho])\le \Omega\left(\int_M g^2\,d\ell \right)\le (C_2(\mu_{\min}))^{\f32}||h||^2_\infty.
$$
To end the proof, it remains to compare
$$
||h||^2_\infty \hbox{ and }\int_M \vert \nabla h \vert^2\, d\ell.
$$
\smallskip
It is only at this point that we use the assumption about the dimension of $M$.
\smallskip
Let us start first by an observation regarding regularity and prove that if $\mathcal J$ is finite then $\rho$ must be continuous. Observe first that for $\gamma>0$ and any measurable $\rho$, we have
\begin{eqnarray}
\mathcal J(\rho)& = & \int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell\\
& = & \int_M\vert \nabla \varphi'(\rho)\vert^2 \rho \,d\ell+ \int_M\vert \nabla \varphi'(\mu)\vert^2\,d\ell - 2\int_M\nabla \varphi'(\mu)\nabla\varphi'(\rho)\rho\, d\ell\\
&\geq & \frac12\int_M\vert \nabla \varphi'(\rho)\vert^2 \rho\,d\ell - \int_M\vert \nabla \varphi'(\mu)\vert^2\rho\,d\ell
\end{eqnarray}
where we have used $2\vert \nabla \varphi'(\mu) \nabla\varphi'(\mu) \vert \leq 2\vert \nabla \varphi'(\mu)\vert^2 +\f12 \vert \nabla \varphi'(\rho)\vert^2$. Setting
$$v(r)=\int_1^r\sqrt{s}\varphi'(s)\,ds, \mbox{ for $r>0$}$$
we see that $$\int_M\vert \nabla \varphi'(\rho)\vert^2 \rho\,d\ell=\int_M |\nabla v(\rho)|^2\,d\ell.$$ Thus the finiteness of $\mathcal J[\rho]$ implies that $\int_M |\nabla v(\rho)|^2\,d\ell$ is finite too.
\smallskip
Assuming now that $\dim M=1$, standard results ensures that $v(\rho)$ is absolutely continuous and thus so is $\rho$
(so that we have furthermore $\int (v(\rho))^2\, d\ell<+\infty$ and $v(\rho)$ belongs to the Sobolev space $W^{1,2}(M)$).
Observe that we also obtain that $\rho$ is positive everywhere. Since the function $\rho-\mu$ is continuous and satisfies
$\int_M \rho-\mu\, d\ell =0$, there exists at least one point $x_0$ in $M$, such that $\rho(x_0)-\mu(x_0)=0$.
It follows from \eqref{g} that $g(x_0)=0$ and from \eqref{theta} that $h(x_0)=0$ (where $h=\theta(g)$). For any $x\in \mathbb R/(L\mathbb{Z})$, denote by $[x_0,x]$ the shortest segment in $\mathbb R/(L\mathbb{Z})$ with boundary points $\{x_0,x\}$. Since $h$ is absolutely continuous:
\begin{eqnarray*}
h^2(x)&=&\left(\int_{[x_0,x]}h'(y)\, \ell(dy)\right)^2\\
&\leq & \ell([x_0,x]) \int_{[x_0,x]}(h'(y))^2\, \ell(dy)\\
&\leq & \f{L}2\int_M \vert \nabla h\vert^2\, d\ell.
\end{eqnarray*}
Gathering the previous results gives
$$
\mathcal J[\rho] \ge c(\beta) \Omega(\mathcal I[\rho]),
$$
with
\begin{equation}\label{constant}
c(\beta)=\f2L C_2(\mu_{\min})^{-\f{3}2}
\end{equation}
Since $ 2/\eta=4(1-m)/(1-2m)$ and for $\beta$ large enough, $C_1=\displaystyle O\left(\beta^{\frac{2-m}{2(1-m)}}\right)$, one has $C_2=\displaystyle O\left(\beta^{\frac{2(2-m)}{1-2m}}\right)$ and
$$c(\beta)=\displaystyle O\left(\beta^{\frac{-3(2-m)}{1-2m}}\right).$$
Let us conclude with the proof of Corollary~\ref{c:tal}. The key is to observe that $1/\sqrt{\Omega}$ is integrable at $0$, so that the inequality is a \L ojasiewicz gradient inequality. It suffices to use the generalization of Otto-Villani theorem provided in \cite[Theorem 1(i)]{blanchet18}.
\section{Time-dependent swarm gradient methods}
Time dependence is key to obtain convergence to the actual global minimum: the penalty schedule $\beta$ is tuned so that the exploratory forces embodied in $\mathcal H$ are sufficiently active at the beginning of the process while progressively loosing their influence on the dynamics as global goals have been achieved. In this second phase, as the diffusion process generated by $\mathcal H$ fades away, the gradient dynamics of $U$ dominates and somehow terminates the process. As in the famous simulated annealing method, the presence of a functional inequality is fundamental for the dynamical system to converge.
\medskip
\noindent
In the remainder, unless otherwise stated we use a general potential $\varphi$.
\subsection{Main convergence results}
\paragraph{Convergence under a functional inequality}
The following general theorem shows the global optimization properties of \eqref{evol0}, provided that a functional inequality is available (as in simulated annealing or as in Theorem \ref{t:verrou} below).
\begin{theorem}[Global optimization under a weak functional inequality]\label{t:conv} Assume the set of hypothesis A, B are met, and that $\beta,U,$ satisfy a functional inequality of the type:
\begin{equation*}
\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\mu)\vert^2\rho\, d\ell\geq c(\beta)\,\Omega\left(\int_M\varphi(\rho)-\varphi(\mu)-\varphi'(\mu)(\rho-\mu)\, d\ell\right)
\end{equation*}
where $c,\Omega:(0,+\infty)\to\mathbb R$ are positive with $\Omega$ being nondecreasing. If the penalization schedule $t\mapsto\beta(t)$ satisfies
\begin{align}
\lim_{t\rightarrow+\infty} \dot{\beta}(t)/c(\beta (t))=0 \label{1}\\
\int_1^{+\infty}c(\beta(t))\, dt =+\infty\label{2}
\end{align}
then
\begin{equation}
\mathcal U[\rho(t)]-\min_M U= \int_M U\rho(t)-\min_M U\: \leq \: \mathrm{gap}(\beta_t) +o( \beta_t^{-1}),
\end{equation}
where the quantity $\mathrm{gap}(\beta_t)\to 0$ was defined in \eqref{app}.
\end{theorem}
\begin{remark}[Existence of a schedule]\label{r:schedule}{\rm Assume that the function $c$ satisfies for large $\beta>0$,
\begin{eqnarray*}
c(\beta)&=& O(\beta^{-\gamma})\end{eqnarray*}
for some exponent $\gamma>0$ (as it is the case in Proposition \ref{funcineq}).
Then for any $\alpha\in(0,1/(1+\gamma))$, any penalization schedule $t\mapsto \beta(t)$
such that for large enough $t>0$, $\beta(t)= t^{\alpha}$
satisfies the assumptions of the convergence theorem, as it is readily checked.}
\end{remark}
\paragraph{Proof of Theorem \ref{t:conv}}
\begin{proof}
Recall that the evolution equation \eqref{evol0} writes
$$
\forall\ t\geq 0,\qquad\frac{{\rm d}}{{\rm d}t} \rho(t)=\mathrm{div}(\rho(\beta_t\nabla U+\nabla \varphi'(\rho))).
$$
In view of \eqref{rhopsicU}, the curve of stationary measures $\nu_t\df\mu_{\beta_t}$, satisfies, for each fixed $t$
\begin{equation}\label{opt}
\beta_t\nabla U+\nabla \varphi'(\nu_t )\;=\;0.
\end{equation}
The functionals $\mathcal I$ and $\mathcal J$ used in Section~\ref{s:globdyn} are adapted to the time-inhomogeneous case as follows
$$
\mathcal I[t,\rho]\df\int_M\varphi(\rho)\, d\ell+\int_M \beta_t U\,\rho \,d\ell -\left(\int_M\varphi(\nu_t )\, d\ell+\int_M \beta_t U\,\nu_t \,d\ell\right),
$$
and
$$
\mathcal J[t,\rho]=\int_M\vert \nabla \varphi'(\rho)-\nabla\varphi'(\nu_t )\vert^2\rho\, d\ell,
$$
for any admissible $\rho$ and time $t\geq0$.
Using regularity results \eqref{ac0}-\eqref{ac1}, the differentiation of the objective along the evolution curve $t\mapsto\rho$ of \eqref{evol0} yields
\begin{eqnarray*}
\frac{{\rm d}}{{\rm d}t} \mathcal I[t,\rho(t)]&=&\int\varphi'(\rho)\frac{{\rm d}}{{\rm d}t} {\rho}\, d\ell+\int \beta_t U\frac{{\rm d}}{{\rm d}t} {\rho}\, d\ell+ \frac{{\rm d}}{{\rm d}t} {\beta_t}\int_M U(\rho-\nu_t )\,d \ell \\
&=&\int(\varphi'(\rho)+\beta_tU)\mathrm{div}(\rho(\beta_t \nabla U+\nabla \varphi'(\rho)))\, d\ell + \dot{\beta_t}\int_M U(\rho-\nu_t )\,d \ell \\
&=&-\int\nabla (\varphi'(\rho)+\beta_t U)(\beta_t \nabla U+\nabla \varphi'(\rho)))\, \rho d\ell + \dot{\beta_t}\int_M U(\rho-\nu_t )\,d \ell \\
&=&-\int_M\vert \nabla \varphi'(\rho)+\beta_t\nabla U\vert^2\, \rho d \ell+ \dot{\beta_t}\int_M U(\rho-\nu_t )\,d \ell\\
&=&-\mathcal J[t,\rho(t)]+ \dot{\beta_t}\int_M U(\rho-\nu_t )\,d \ell,
\end{eqnarray*}
where we have used \eqref{opt} in the first equality, the evolution equation in the second-one, and finally integration by parts.
Let us set $v(t)=\mathcal I(t,\rho)$. Using the functional inequality gives
$$
\dot v(t) \le -c(\beta_t) \Omega( v(t))+\dot{\beta_t}\int_M U(\rho-\nu_t )\,d \ell.
$$
To give an upper bound of the second term, we write
$$
\lt\vert\dot{\beta_t}\int_M U(\rho-\nu_t )\,d \ell\rt\vert=\lt\vert\dot{\beta_t}\left(\int_M U\rho\,d \ell-\int_MU\nu_t \,d \ell\right)\rt\vert \le \mathrm{osc}(U)\lt\vert\dot{\beta_t}\rt\vert,
$$
where the last inequality uses that $\rho$ and $\nu_t $ are probability densities on $M$.
Finally, we obtain
$$
\dot v(t) \le -c(\beta_t) \Omega( v(t))+\mathrm{osc}(U)\lt\vert\dot{\beta_t}\rt\vert,
$$
\medskip
We are thus led to consider differential inequalities of the type
\begin{eqnarray*}\label{}
\dot{v}&\leq & -c(\beta)\Omega(v)+\delta\lt\vert\dot{\beta}\rt\vert\end{eqnarray*}
where $v\,:\, \mathbb R_+\rightarrow \mathbb R_+$ is a non-negative function, $d>0$ is a positive constant and $\Omega$ is an nondecreasing function
taking positive values on $(0,+\infty)$.
\smallskip
Our goal is now to give conditions on the inverse temperature scheme $\beta\,:\, \mathbb R_+\rightarrow(0,+\infty)$ ensuring that $v$ converges to zero for large times:
\begin{pro}[Schedule conditions]\label{pro}
Assume that for large times, \eqref{1} and \eqref{2} hold,
then
\begin{eqnarray*}
\lim_{t\rightarrow+\infty} v(t)&=&0.\end{eqnarray*}
\end{pro}
\smallskip
\noindent
The proof of Proposition \ref{pro} is based on the two following observations.
\begin{lem}\label{liminfv}
Under the assumptions of Proposition \ref{pro}, we have
\begin{eqnarray*}
\liminf_{t\rightarrow+\infty} v(t)&=&0.\end{eqnarray*}
\end{lem}
\begin{proof}
Towards a contradiction assume that there exist $\epsilon>0$ and $T_0\geq 0$ such that
$v(t)\geq \epsilon$ for all $t\geq T_0$.
Then, $\Omega$ being nondecreasing,
\begin{eqnarray*}
\forall\ t\geq T_0,\qquad \dot{v}(t)&\leq & -c(\beta(t))\Omega(\epsilon)+\delta\lt\vert\dot{\beta}(t)\rt\vert.
\end{eqnarray*}
From the first condition \eqref{1}, there exists $T_1\geq T_0$ such that
\begin{eqnarray*}
\forall\ t\geq T_1,\qquad \lt\vert\dot{\beta}(t)\rt\vert&\leq &\f{1}{2}c(\beta(t))\Omega(\epsilon)\end{eqnarray*}
and thus
\begin{eqnarray*}
\forall\ t\geq T_1,\qquad \dot{v}(t)&\leq & -\f{1}{2}c(\beta(t))\Omega(\epsilon).
\end{eqnarray*}
This implies
\begin{eqnarray*}
\forall\ t\geq T_1,\qquad
v(t)&\leq &v(T_1) -\f{1}{2}\Omega(\epsilon)\int_{T_1}^tc(\beta(s))\, ds,\end{eqnarray*}
so letting $t$ go to infinity, due to the second condition \eqref{2}, $\lim_{t\rightarrow+\infty} v(t)\,=\,-\infty$, in contradiction with the non-negativity of $v$.
\end{proof}
\par
Our second ingredient is:
\begin{lem}\label{downward}
Under the assumptions of Proposition \ref{pro}, fix $\epsilon>0$. Then there exists
$T\geq 0$, such that
\begin{eqnarray*}
\forall\ t\geq T,\quad \big( v(t)=\epsilon&\Rightarrow& \dot{v}(t)<0\big).
\end{eqnarray*}
\end{lem}
\begin{proof}
Indeed, consider \begin{eqnarray*}
T&\df& \inf\left\{\tau\geq 0\,:\, \forall\ t\geq \tau,\ -\f{1}{2}c(\beta(t))\Omega(\epsilon)+\delta\lt\vert\dot{\beta}(t)\rt\vert\leq 0\right\}\end{eqnarray*}
which is finite by assumption \eqref{1}.
Proceeding as in Lemma~\ref{liminfv}, for any $t\geq T$ such that $v(t)=\epsilon$, we may assert that
\begin{eqnarray*}
\dot{v}(t)&\leq &-\f{1}{2}c(\beta(t))\Omega(\epsilon)\,<\,0.\end{eqnarray*}
\end{proof} \par
\smallskip
The proof of Proposition \ref{pro} is as follows:
by Lemma~\ref{liminfv}, for any arbitrary small level $\epsilon>0$,
$v$ will always go below $\epsilon$ at some point, but by Lemma~\ref{downward} there exists a time $T$ after which it will no longer be able to cross it upward. Whence $v$ must be below $\epsilon$ for large times. This concludes the proof of Proposition~\ref{pro}.
\smallskip
Let us now come back to the proof of Theorem~\ref{t:conv}: since $v(t)$ tends to zero, we have for all $\epsilon>0$ a $T_0$ such that
$$
\mathcal I[t,\rho]\df\int_M\varphi(\rho)\, d\ell+\int_M \beta_t U\,\rho \,d\ell -\left(\int_M\varphi(\nu_t )\, d\ell+\int_M \beta_t U\,\nu_t \,d\ell\right)\leq \epsilon
$$
that is
\begin{eqnarray*}
\int_M U\,\rho \,d\ell -\int_M U\,\nu_t \,d\ell & \leq & \beta_t^{-1} \left( \int_M\varphi(\nu_t )\, d\ell-\int_M\varphi(\rho)\, d\ell +\epsilon\right)
\end{eqnarray*}
which implies in turn
\begin{eqnarray*}
\int_M U\,\rho \,d\ell -\min_M U & \leq & \beta_t^{-1} \left( \int_M\varphi(\nu_t )\, d\ell+\epsilon\right) +\left(\int_M U\,\nu_t \,d\ell-\min_M U\right)\\
& \leq & \mathrm{gap}(\beta_t) +\epsilon \beta_t^{-1}.
\end{eqnarray*}
The value of $\epsilon$ being arbitrary, the results follows.\end{proof}
\paragraph{One-dimensional global optimization}
We may now combine Theorem~\ref{t:conv} with the functional inequality we obtained previously in Theorem~\ref{funcineq}.
\begin{theorem}[Global optimization by swarm gradient in dimension 1]\label{t:verrou}
Assume that $M$ is the circle $\mathbb{T}\df\mathbb R/(L\mathbb{Z})$ endowed with its usual Riemannian structure, that the function $\varphi$ is as in \eqref{varphi}, i.e., $\varphi=\varphi_{m,2}$ and that the schedule is given by
$$\beta(t)=kt^{1/\gamma}, \mbox{ with $k>0$ and $\gamma=\frac{3(2-m)}{1-2m}\in [6,+\infty)$.}$$
Then
\begin{equation}
\lim_{t\to\infty}\mathcal U[\rho(t)]=\min_M U
\end{equation}
\end{theorem}
\begin{proof}
Theorem~\ref{funcineq} provides a functional inequality as required by Theorem~\ref{t:conv}. From this inequality we may assume that $c(\beta)=\kappa \beta^{-\gamma}$ for some $\kappa>0$.
Now, using Remark~\ref{r:schedule}, we choose $\beta(t)=c^{-1}(t)=\kappa^{-1}t^{\frac{1}{\gamma}}$. This is the choice made by assumption provided that $k=\kappa^{-1}$. One easily checks that the assumptions of Theorem~\ref{t:conv} are satisfied, hence the result.
\end{proof}
\subsection{A stochastic view on the dynamics: particles swarm optimization}\label{s:swarm}
In this last section, we sketch a possible discrete approach of the swarm gradient dynamics; it will be properly developed in future works.
It relies on the diffusion process associated with the evolution equation
\bqn{evol}
\forall\ t\geq 0,\qquad\frac{{\rm d}}{{\rm d}t} \rho &=&\mathrm{div}(\rho (\beta_t\nabla U+\nabla \varphi'(\rho))).
\end{eqnarray}
To evidence this link, let us use the formal integration by parts presented in Section \ref{s2.2} to obtain at any time $t\geq 0$ and for every $ f\in\mathcal{C}^\infty(M)$,
\begin{eqnarray*}
\int_{\{\rho_t>0\}} L_{t,\rho}[f]\, \rho_t d\ell&=&\int_M f \mathrm{div}(\rho(\beta_t\nabla U+\nabla \varphi'(\rho)))\, d\ell
\end{eqnarray*}
where
\bqn{generator}
L_{t,\rho}[f]&=&\alpha(\rho)\triangle f-\lt\langle\beta_t \nabla U,\nabla f\rt\rangle
\end{eqnarray}
with $\alpha\,:\,(0,+\infty)\rightarrow\mathbb R_+$ given by
\begin{eqnarray*}
\forall\ r> 0,\qquad \alpha(r)&\df &\f1r\int_0^r s\varphi''(s)\, ds.
\end{eqnarray*}
\par
It is then natural to associate to equation \eqref{evol} a Markov process $(X_t)_{t\geq 0}$ whose infinitesimal generator is given by \eqref{generator} and whose law has density $\rho(t)$ for all $t\geq0$.
Due to the dependence of the evolution on the density, such a Markov process is said to be non-linear.
When $M$ is a flat torus of dimension $d$,
consider the stochastic differential equation
\bqn{SDE}
X_t=X_0-\int_0^t \beta_s\nabla U(X_s)\,ds+\int_0^t \sqrt{\alpha(\rho(X_s))} \,dB_s,
\end{eqnarray}
where $(B_s)_{s\geq 0}$ is a Brownian motion of dimension $d$. If $(X,\rho)$ is a solution of the stochastic differential equation \eqref{SDE} then an application of the It\^o formula shows that $\rho$ is a solution of equation \eqref{evol}.
This observation can be extended to any compact Riemannian manifold.\par
Two particular choices for which the existence of $(X,\rho)$ has been established in the literature are
\begin{itemize}
\item[(i)] $\varphi(\rho)=\rho\log(\rho)$, we have $\alpha\equiv 1$, so $L_{t,\rho}$ does not depend on $\rho$ and is the Langevin generator $\triangle\cdot -\beta_t\lt\langle \nabla U,\nabla \cdot\rt\rangle$. The existence and uniqueness of a strong solution to the extension of \eqref{SDE} on a compact Riemannian is well-known, as soon as $\nabla U$ is Lipschitz, which is true if $U$ is smooth.
\item[(ii)] $\varphi(\rho)=\rho^m$ which corresponds to the equation of porous media.
On $\mathbb R$, the existence and uniqueness of a strong solution to \eqref{SDE} is proven by Benachour, Chassaing, Roynette and Vallois \cite{ASNSP_1996_4_23_4_793_0}
(see {Belaribi} and {Russo} \cite{zbMATH06098167} for more general functions $\varphi$).
\end{itemize}
\par
The main difficulty in investigating the existence and uniqueness of \eqref{SDE} is the presence of the density $\rho_s$. It can be relaxed if it could be replaced by integrals of smooth functions with respect to $\rho_s(x)\, \ell(dx)$.
It leads us to consider convolutions of $\rho_s$ with respect to some smooth kernels, as those traditionally used in statistics for density estimation, cf.\ e.g.\ Silverman \cite{zbMATH04001209}. In our geometric setting, it seems natural to resort to the heat kernel (associated to the Laplace-Beltrami operator).
To avoid the introduction of further notation, let us just consider the case of a flat torus $M=(\mathbb R/\mathbb{Z})^d$, endowed with its usual Riemannian structure.
Let $K\,:\, (\mathbb R/\mathbb{Z})^d\rightarrow \mathbb R_+$ be a smooth function with support in $[-1/4,1/4]^d$ (seen as a subset of $(\mathbb R/\mathbb{Z})^d$) and such that $\int_{M}K\, d\ell =1$.
For any probability density $\rho$ on $M$ and $h\in(0,1)$, which is a bandwidth parameter, set
\bqn{rhoh}
\forall\ x\in M,\qquad \rho_h(x)&\df& h^{-d}\int_M K((x-y)/h)\, \rho(y) \ell(dy).\end{eqnarray}
\par
When $h$ is small, $\rho_h$ is an approximation of $\rho$.
Replacing in \eqref{SDE}, $\rho_s$ by $\rho_{h,s}$, we end up with the stochastic differential equation
\bqn{SDE2}
X_t=X_0-\int_0^t \beta_s\nabla U(X_s)\,ds+\int_0^t \sqrt{\alpha(\rho_{h}(X_s))} \,dB_s,
\end{eqnarray}
which is simpler to investigate, taking into account the usual mean field theory (see for instance Del Moral \cite{MR3060209}).
\par
Furthermore, \eqref{SDE} admits natural particles approximations.
Indeed, \eqref{rhoh} can be extended to any probability measure $\pi$ on $M$, by replacing $\rho(y) \ell(dy)$ by $\pi(dy)$.
In particular, it makes sense for the empirical measure of $N$ particles, which is a crucial feature for particle approximations.
More precisely, consider
a system of $N$ particles, $X_1, X_2, ..., X_N$ whose joint evolution is described by the stochastic differential equations,
\bqn{Xn}
\forall\ n\in\llbracket N\rrbracket,\qquad dX_n(t)&=&-\beta_t\nabla U(X_n(t))+\sqrt{\alpha(\rho_{N,h}(X_n(t)))} \,dB_n(t)\end{eqnarray}
where
the $(B_n(t))_{t\geq 0}$, for $n\in\llbracket N\rrbracket$, are independent Brownian motions of dimension $d$,
and where
\begin{eqnarray*}
\forall\ x\in M,\qquad \rho_{N,h}(x)&\df& \frac{1}{N}\sum_{n=1}^N \frac{1}{h}K\left( \frac{x-X_n(t)}{h}\right)\end{eqnarray*}
namely $\rho_{N,h}$ is given by \eqref{rhoh} when $\rho$ is replaced by the empirical measure
\begin{eqnarray*}
\rho_{N}(t)&\df& \f1N\sum_{n=1}^N\delta_{X_n(t)}\end{eqnarray*}
\par
Resorting to mean field theory and chaos propagation, when $N$ tends to infinity, the random evolution $(\rho_{N}(t))_{t\geq 0}$ converges in probability toward the dynamical system
$(\rho(t))_{t\geq 0}$ (for related initial conditions), the law of $(X_1(t))_{t\geq 0}$ converges toward that of $(X_t)_{t\geq 0}$, solution of \eqref{SDE2}, and the trajectories of any fixed finite number of particles become asymptotically independent.
\par
Putting together these observations, we believe that in addition to $(\beta_t)_{t\geq 0}$, schemes $ (h_t)_{t\geq 0}$ and $(N_t)_{t\geq 0}$ can be found, with $\lim_{t\rightarrow+\infty} h_t=0$ and
$\lim_{t\rightarrow+\infty} N_t=\infty$, so that for large times $t$, the corresponding empirical measures $\rho_{N_t}$ concentrate around the set of global minima of $U$.
There are several ways to increase the number of particles, the most natural one might be to duplicate some of the current particles --but it is also possible to make them appear in independent random positions.\par
This procedure would provide a new stochastic algorithm for global minimization. At least up to the simulation of particle systems such as \eqref{Xn}, but this can be done through traditional Euler-Maruyama scheme.
\par
\paragraph{Acknowledgements.}
JB acknowledges the support of ANR-3IA Artificial and Natural Intelligence Toulouse Institute, and thank Air Force Office of Scientific Research, Air Force Material Command, USAF, under grant numbers FA9550-19-1-7026, and ANR MasDol. JB, LM, SV acknowledges the support of ANR Chess, grant ANR-17-EURE-0010 and TSE-P.
\bibliographystyle{siam}
|
\section{Introduction}
\label{sec:introduction}
Optimal transport and Wasserstein distance~\cite{Villani-09,peyre2020computational} have become popular tools in machine learning and data science. For example, optimal transport has been utilized in generative modeling tasks to generate realistic images~\cite{arjovsky2017wasserstein,tolstikhin2018wasserstein}, in domain adaptation applications to transfer knowledge from source to target domains ~\cite{courty2017joint,bhushan2018deepjdot}, in clustering applications to capture the heterogeneity of data~\cite{ho2017multilevel}, and in other applications~\cite{le2021lamda, xu2021vocabulary, yang2020predicting}. Despite having appealing performance, Wasserstein distance has been known to suffer from high computational complexity, namely, its computational complexity is at the order of $\mathcal{O}(m^3 \log m)$~\cite{pele2009} when the probability measures have at most $m$ supports. In addition, Wasserstein distance also suffers from the curse of dimensionality, namely, its sample complexity is at the order of $\mathcal{O}(n^{-1/d})$~\cite{Fournier_2015} where $n$ is the sample size. A popular line of works to improve the speed of computation and the sample complexity of the Wasserstein distance is by adding an entropic regularization term to the Wasserstein distance~\cite{cuturi2013sinkhorn}. This variant is known as entropic regularized optimal transport (or equivalently entropic regularized Wasserstein). By using the entropic version, we can approximate the value of Wasserstein distance with the computational complexities being at the order of $\mathcal{O}(n^2)$~\cite{altschuler2017near, lin2019efficient, Lin-2019-Efficiency, Lin-2020-Revisiting} (up to some polynomial orders of approximation errors). Furthermore, the sample complexity of the entropic version had also been shown to be at the order of $\mathcal{O}(n^{-1/2})$~\cite{Mena_2019}, which indicates that it does not suffer from the curse of dimensionality.
\vspace{0.5 em}
\noindent
Another useful line of works to improve both the computational and sample complexities of the Wasserstein distance is based on the closed-form solution of optimal transport in one dimension. A notable distance along this direction is sliced Wasserstein (SW) distance~\cite{bonneel2015sliced}. Due to the fast computational complexity $\mathcal{O}(m \log_2 m)$ and no curse of dimensionality $\mathcal{O}(n^{-1/2})$, the sliced Wasserstein has been applied successfully in several applications, such as generative modeling~\cite{wu2019sliced,deshpande2018generative,kolouri2018sliced}, domain adaptation~\cite{lee2019sliced}, and clustering~\cite{kolouri2018slicedgmm}. The sliced Wasserstein is defined between two probability measures that have supports belonging to a vector space, e.g, $\mathbb{R}^d$. As defined in~\cite{bonneel2015sliced}, the sliced Wasserstein is written as the expectation of one-dimensional Wasserstein distance between two projected measures over the uniform distribution on the unit sphere. Due to the intractability of the expectation, Monte Carlo samples from the uniform distribution over the unit sphere are used to approximate the sliced Wasserstein distance. The number of samples is often called the number of projections and it is denoted as $L$. On the computational side, the computation of sliced Wasserstein can be decomposed into two steps. In the first step, $L$ projecting directions are first sampled and then are stacked as a matrix (the projection matrix). After that, the projection matrix is multiplied by the two data matrices resulting in two matrices that represent $L$ one-dimensional projected probability measures. In the second step, $L$ one-dimensional Wasserstein distances are computed between the two corresponding projected measures with the same projecting direction. Finally, the average of those distances is yielded as the value of the sliced Wasserstein.
\vspace{0.5em}
\noindent
Despite being applied widely in tasks that deal with probability measures over images~\cite{wu2019sliced,deshpande2018generative}, the conventional formulation of sliced Wasserstein is not well-defined to the nature of images. In particular, an image is not a vector but is a tensor. Therefore, a probability measure over images should be defined over the space of tensors instead of images. The conventional formulation leads to an extra step in using the sliced Wasserstein on the domain of images which is vectorization. Namely, all images (supports of two probability measures) are transformed into vectors by a deterministic one-one mapping which is the "reshape" operator. This extra step does not keep the spatial structures of the supports, which are crucial information of images. Furthermore, the vectorization step also poses certain challenges to design efficient ways of projecting (slicing) samples to one dimension based on prior knowledge about the domain of samples. Finally, prior empirical investigations indicate that there are several slices in the conventional Wasserstein collapsing the two probability measures to the Dirac Delta at zero~\cite{deshpande2018generative,deshpande2019max,kolouri2019generalized}. Therefore, these slices do not contribute to the overall discrepancy. These works suggest that the space of projecting directions in the conventional sliced Wasserstein (the unit hyper-sphere) is potentially not optimal, at least for images.
\vspace{0.5em}
\noindent
\textbf{Contribution.} To address these issues of the sliced Wasserstein over images, we propose to replace the conventional formulation of the sliced Wasserstein with a new formulation that is defined on the space of probability measures over tensors. Moreover, we also propose a novel slicing process by changing the conventional matrix multiplication to the convolution operators~\cite{fukushima1982neocognitron,goodfellow2016deep}. In summary, our main contributions are two-fold:
\begin{enumerate}
\item We leverage the benefits of the convolution operators on images, including their efficient parameter sharing and memory saving as well as their superior performance in several tasks on images~\cite{krizhevsky2012imagenet,he2016deep}, to introduce efficient slicing methods on sliced Wasserstein, named \emph{convolution slicers}. With the convolution slicers, we derive a novel variant of sliced Wasserstein, named \emph{convolution sliced Wasserstein} (CSW). We investigate the metricity of CSW, its sample and computational complexities, and its connection to other variants of sliced Wasserstein.
\item We then illustrate the favorable performance of CSW in comparing probability measures over images. In particular, we show that CSW provides an almost identical discrepancy between MNIST's digits compared to that of the SW while having much less slicing memory. Furthermore, we compare SW and CSW in training deep generative models on standard benchmark images datasets, including CIFAR10, CelebA, STL10, and CeleA-HQ. By considering quality of the trained generative models, training speed, and training memory of CSW and SW, we observe that the CSW has more favorable performance than the vanilla SW.
\end{enumerate}
\vspace{0.5 em}
\noindent
\textbf{Organization.} The remainder of the paper is organized as follows. We first provide background about Wasserstein distance, the conventional slicing process in the sliced Wasserstein distance, and the convolution operator in Section~\ref{sec:background}. In Section~\ref{sec:csw}, we propose the convolution slicing and the convolution sliced Wasserstein, and analyze some of its theoretical properties. The discussion on related works is given in Section~\ref{sec:relatedworks}. Section~\ref{sec:experiments} contains the application of CSW to generative models, qualitative experimental results, and quantitative experimental results on standard benchmarks. We conclude the paper In Section~\ref{sec:conclusion}. Finally, we defer the proofs of key results and extra materials in the Appendices.
\vspace{0.5 em}
\noindent
\textbf{Notation.} For any $d \geq 2$, $\sphere{d}:=\{\theta \in \mathbb{R}^{d}\mid ||\theta||_2^2 =1\}$ denotes the $d$ dimensional unit hyper-sphere in $\mathcal{L}_2$ norm, and $\mathcal{U}(\sphere{d})$ is the uniform measure over $\sphere{d}$. Moreover, $\delta$ denotes the Dirac delta function. For $p\geq 1$, $\mathcal{P}_p(\mathbb{R}^d)$ is the set of all probability measures on $\mathbb{R}^d$ that have finite $p$-moments. For $\mu,\nu \in \mathcal{P}_p(\mathbb{R}^d)$, $\Pi(\mu,\nu):=\{\pi \in \mathcal{P}_p(\mathbb{R}^d \times \mathbb{R}^d) \mid \int_{\mathbb{R}^d} \pi(x,y) dx = \nu, \int_{\mathbb{R}^d} \pi(x,y) dy = \mu \}$ is the set of transportation plans between $\mu$ and $\nu$. For $m\geq 1$, we denotes $\mu^{\otimes m }$ as the product measure which has the support is $m$ random variables follows $\mu$. For a vector $X \in \mathbb{R}^{dm}$, $X:=(x_1,\ldots,x_m)$, $P_{X}$ denotes the empirical measures $\frac{1}{m} \sum_{i=1}^m \delta_{x_i}$. For any two sequences $a_{n}$ and $b_{n}$, the notation $a_{n} = \mathcal{O}(b_{n})$ means that $a_{n} \leq C b_{n}$ for all $n \geq 1$ where $C$ is some universal constant.
\section{Background}
\label{sec:background}
In this section, we first review the definitions of the Wasserstein distance, the conventional slicing, and the sliced Wasserstein distance, and discuss its limitation. We then review the convolution and the padding operators on images.
\vspace{0.5 em}
\noindent
\textbf{Sliced Wasserstein:} For any $p \geq 1$ and dimension $d' \geq 1$, we first define the Wasserstein-$p$ distance~\cite{Villani-09, peyre2019computational} between two probability measures $\mu \in \mathcal{P}_p(\mathbb{R}^{d'})$ and $\nu \in \mathcal{P}_p(\mathbb{R}^{d'})$, which is given by:
\begin{align}
\label{eq:Wasserstein}
\text{W}_p(\mu,\nu) : = \Big{(} \inf_{\pi \in \Pi(\mu,\nu)} \int_{\mathbb{R}^{d'} \times \mathbb{R}^{d'}} \| x - y\|_p^{p} d \pi(x,y) \Big{)}^{\frac{1}{p}}.
\end{align}
When $d'=1$, the Wasserstein distance has a closed form which is $W_p(\mu,\nu) =
( \int_0^1 |F_\mu^{-1}(z) - F_{\nu}^{-1}(z)|^{p} dz )^{1/p}$
where $F_{\mu}$ and $F_{\nu}$ are the cumulative
distribution function (CDF) of $\mu$ and $\nu$ respectively. Given this closed-form property of Wasserstein distance in one dimension, the sliced Wasserstein distance~\cite{bonneel2015sliced} between $\mu$ and $\nu$ had been introduced and admitted the following formulation:
\begin{align}
\label{eq:SW}
\text{SW}_p(\mu,\nu) : = \left(\int_{\mathbb{S}^{d-1}} \text{W}_p^p (\theta \sharp \mu,\theta \sharp \nu) d\theta \right)^{\frac{1}{p}},
\end{align}
where $\theta \sharp \mu$ is the push-forward probability measure of $\mu$ through the function $T_\theta: \mathbb{R}^{d'} \to \mathbb{R}$ with $T_\theta(x) = \theta^\top x$. For each $\theta \in \mathbb{S}^{d'- 1}$, $\text{W}_p^p (\theta \sharp \mu,\theta \sharp \nu)$ can be computed in linear time $\mathcal{O}(m \log_2 m)$ where $m$ is the number of supports of $\mu$ and $\nu$. However, the integration over the unit sphere in the sliced Wasserstein distance is intractable to compute. Therefore, Monte Carlo scheme is employed to approximate the integration, namely, $\theta_1,\ldots,\theta_L \sim \mathcal{U}(\sphere{d'})$ are drawn uniformly from the unit sphere and the approximation of the sliced Wasserstein distance is given by:
\begin{align}
\label{eq:hatSW}
\text{SW}_p (\mu,\nu) \approx \left(\frac{1}{L}\sum_{i=1}^L \text{W}_p^p (\theta_i \sharp \mu,\theta_i \sharp \nu) \right)^{\frac{1}{p}}.
\end{align}
In practice, the number of projections $L$ should be chosen to be sufficiently large compared to the dimension $d'$. It can be undesirable since the computational complexity of SW is linear with $L$.
\vspace{0.5 em}
\noindent
\textbf{Sliced Wasserstein on Images:} Now, we focus on two probability measures over images: $\mu,\nu \in \mathcal{P}_p(\mathbb{R}^{c \times d \times d})$ for number of channels $c \geq 1$ and dimension $d \geq 1$. In this case, the sliced Wasserstein between $\mu$ and $\nu$ is defined as:
\begin{align}
\label{eq:SWimage}
\text{SW}_p(\mu,\nu) = \text{SW}_p(\mathcal{R}\sharp \mu,\mathcal{R}\sharp \nu),
\end{align}
where $\mathcal{R}: \mathbb{R}^{c \times d \times d }\to \mathbb{R}^{cd^2}$ is a deterministic one-to-one "reshape" mapping.
\vspace{0.5 em}
\noindent
\textbf{The slicing process:} The slicing of sliced Wasserstein distance on probability measures over images consists of two steps: vectorization and projection. For better understanding, we visualize an example of projecting a probability measure $\mu \in \mathbb{R}^{c\times \times d}$ that has $n$ supports in Figure~\ref{fig:sw}. In short, supports of $\mu$ are transformed into vectors in $\mathbb{R}^{cd^2}$ and are stacked as a matrix of size $n \times cd^2$. A projection matrix of size $L\times cd^2$ is then sampled and has each column as a random vector following the uniform measure over the unit hyper-sphere. Finally, the multiplication of those two matrices returns $L$ projected probability measures of $n$ supports in one dimension.
\vspace{0.5 em}
\noindent
\textbf{Limitation of the conventional slicing:} First of all, images contain spatial relations across channels and local information. Therefore, transforming images into vectors makes it challenging to obtain that information. Second, vectorization leads to the usage of projecting directions from the unit hyper-sphere, which can have several directions that do not have good discriminative power. Finally, sampling projecting directions in high-dimension is also time-consuming and memory-consuming. As a consequence, avoiding the vectorization step can improve the efficiency of the whole process.
\begin{figure*}[!t]
\begin{center}
\begin{tabular}{c}
\widgraph{0.9\textwidth}{figures/SW.pdf}
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{The conventional slicing process of sliced Wasserstein distance. The images $X_{1}, \ldots, X_{n} \in \mathbb{R}^{c \times d \times d}$ are first flattened into vectors in $\mathbb{R}^{cd^2}$ and then the Radon transform is applied to these vectors to lead to sliced Wasserstein~(\ref{eq:SWimage}) on images.
}
}
\label{fig:sw}
\vskip -0.1in
\end{figure*}
\vspace{0.5 em}
\noindent
\textbf{Convolution operator:} We now define the convolution operator on tensors~\cite{fukushima1982neocognitron}, which will be used as an alternative way of projecting images to one dimension in the sliced Wasserstein. The definition of the convolution operator with stride and dilation is as follows.
\begin{definition} (Convolution)
\label{def:conv}
Given the number of channels $c\geq 1$, the dimension $d\geq 1$, the stride size $s\geq 1$, the dilation size $b\geq 1$, the size of kernel $k\geq 1$, the convolution of a tensor $X \in \mathbb{R}^{c \times d \times d}$ with a kernel size $K \in \mathbb{R}^{c \times k \times k}$ is
\begin{align}
X \stackrel{s,b}{*} K = Y, \quad Y \in \mathbb{R}^{1 \times d' \times d'}, \nonumber
\end{align}
where $d' = \frac{d-b(k-1)-1}{s}+1$. For $i=1,\ldots,d'$ and $j=1,\ldots,d'$, $Y_{1,i,j}$ is defined as:
\begin{align*}
Y_{1,i,j} = \sum_{h=1}^{c} \sum_{i'=0}^{k-1} \sum_{j'=0}^{k-1} X_{h,s(i-1)+bi'+1,s(j-1)+bj'+1}\cdot K_{h,i'+1,j'+1}.
\end{align*}
\end{definition}
\vspace{0.5em}
\noindent
From its definition, we can check that the computational complexity of the convolution operator is $\mathcal{O}\left( c\left(\frac{d-b(k-1)-1}{s}+1\right)^2 k^2 \right)$.
\section{Convolution Sliced Wasserstein}
\label{sec:csw}
In this section, we will define a convolution slicer that maps a tensor to a scalar by convolution operators. Moreover, we discuss the convolution slicer and some of its specific forms including the convolution-base slicer, the convolution-stride slicer, the convolution-dilation slicer, and their non-linear extensions. After that, we derive the convolution sliced Wasserstein (CSW), a family of variants of sliced Wasserstein, that utilizes a convolution slicer as the projecting method. Finally, we discuss some theoretical properties of CSW, namely, its metricity, its computational complexity, its sample complexity, and its connection to other variants of sliced Wasserstein.
\subsection{Convolution Slicer}
\label{subsec:cslicing}
We first start with the definition of the convolution slicer, which plays an important role in defining convolution sliced Wasserstein.
\begin{definition} (Convolution Slicer)
\label{def:cslicer}
For $N\geq 1$, given a sequence of kernels $K^{(1)} \in \mathbb{R}^{c^{(1)}\times d^{(1)} \times d^{(1)}},\ldots,$\\$K^{(N)} \in \mathbb{R}^{c^{(N)}\times d^{(N)} \times d^{(N)}}$, a \emph{convolution slicer} $\mathcal{S}(\cdot|K^{(1)},\ldots,K^{(N)})$ on $\mathbb{R}^{c \times d \times d}$ is a composition of $N$ convolution functions with kernels $K^{(1)},\ldots, K^{(N)}$ (with stride or dilation if needed) such that:
\begin{align*}
\mathcal{S}(X|K^{(1)},\ldots, K^{(N)}) \in \mathbb{R} \quad \forall X \in \mathbb{R}^{c \times d \times d}.
\end{align*}
\end{definition}
\vspace{0.5 em}
\noindent
As indicated in Definition~\ref{def:cslicer}, the idea of the convolution slicer is to progressively map a given data $X$ to a one-dimensional subspace through a sequence of convolution kernels, which capture spatial relations across channels as well as local information of the data. It is starkly different from the vectorization step in standard sliced Wasserstein on images~(\ref{eq:SWimage}).
\vspace{0.5 em}
\noindent
Now, we will specify three particular types of convolution slicers based on using linear function on the convolution operator, named convolution-base, convolution-stride, and convolution-dilation slicers. We first start with the definition of the convolution-base slicer.
\begin{definition}
\label{def:linearslicer}
(Convolution-base Slicer) Given $X \in \mathbb{R}^{c \times d \times d}$ ($d \geq 2$),
\begin{enumerate}
\item When $d$ is even, $N = [\log_2 d]$, sliced kernels are defined as $K^{(1)}\in \mathbb{R}^{c \times (2^{-1}d+1) \times (2^{-1}d+1)}$ and $K^{(h)}\in \mathbb{R}^{1 \times (2^{-h}d+1) \times (2^{-h}d+1)}$ for $h =2,\ldots,N-1$, and $K^{(N)}\in \mathbb{R}^{1 \times a \times a}$ where $a= \frac{d}{2^{N-1}}$.
Then, the \emph{convolution-base slicer} $\mathcal{CS}\text{-b}(X|K^{(1)},\ldots,K^{(N)})$ is defined as:
\begin{align}
\mathcal{CS}\text{-b}(X|K^{(1)},\ldots,K^{(N)}) = X^{(N)}, \quad X^{(h)} = \begin{cases} X &h=0\\ X^{(h-1)} \conv{1,1}K^{(h)} & 1 \leq h \leq N,
\end{cases}
\end{align}
\item When $d$ is odd, the \emph{convolution-base slicer} $\mathcal{CS}\text{-b}(X|K^{(1)},\ldots,K^{(N)})$ takes the form:
\begin{align}
\mathcal{CS}\text{-b}(X|K^{(1)},\ldots,K^{(N)}) = \mathcal{CS}\text{-b} (X \conv{1,1} K^{(1)}|K^{(2)},\ldots,K^{(N)}),
\end{align}
where $K^{(1)} \in \mathbb{R}^{c \times 2 \times 2}$ and $K^{(2)}, \ldots, K^{(N)}$ are the corresponding sliced kernels that are defined on the dimension $d-1$.
\end{enumerate}
\end{definition}
\vspace{0.5em}
\noindent
The idea of the convolution-base slicer in Definition~\ref{def:linearslicer} is to reduce the width and the height of the image by half after each convolution operator. If the width and the height of the image are odd, the first convolution operator is to reduce the size of the image by one via convolution with kernels of size $2\times 2$, and then the same procedure as that of the even case is applied. We would like to remark that the conventional slicing of sliced Wasserstein in Section~\ref{sec:background} is equivalent to a convolution-base slicer $\mathcal{S}(\cdot|K^{(1)})$ where $K^{(1)} \in \mathbb{R}^{c\times d\times d}$ that satisfies the constraint $\sum_{h=1}^c \sum_{i=1}^d\sum_{j=1}^d K^{(1)2}_{h,i,j}=1$.
\vspace{0.5 em}
\noindent
We now discuss the second variant of the convolution slicer, named convolution-stride slicer, where we further incorporate stride into the convolution operators. Its definition is as follows.
\begin{definition}
\label{def:csslicer}
(Convolution-stride Slicer) Given $X \in \mathbb{R}^{c \times d \times d}$ ($d \geq 2$),
\begin{enumerate}
\item When $d$ is even, $N = [\log_2 d]$, sliced kernels are defined as $K^{(1)}\in \mathbb{R}^{c \times 2 \times 2}$ and $K^{(h)}\in \mathbb{R}^{1 \times 2 \times 2}$ for $h =2,\ldots,N-1$, and $K^{(N)}\in \mathbb{R}^{1 \times a \times a}$ where $a= \frac{d}{2^{N-1}}$.
Then, the \emph{convolution-stride slicer} $\mathcal{CS}\text{-s}(X|K^{(1)},\ldots,K^{(N)})$ is defined as:
\begin{align}
\mathcal{CS}\text{-s}(X|K^{(1)},\ldots,K^{(N)}) = X^{(N)}, \quad X^{(h)} = \begin{cases} X &h=0\\ X^{(h-1)} \conv{2,1}K^{(h)} & 1 \leq h \leq N-1, \\
X^{(h-1)} \conv{1,1}K^{(h)} & h=N,
\end{cases}
\end{align}
\item When $d$ is odd, the \emph{convolution-stride slicer} $\mathcal{CS}\text{-s}(X|K^{(1)},\ldots,K^{(N)})$ takes the form:
\begin{align}
\mathcal{CS}\text{-s}(X|K^{(1)},\ldots,K^{(N)}) = \mathcal{CS}\text{-s} (X \conv{1,1} K^{(1)}|K^{(2)},\ldots,K^{(N)}),
\end{align}
where $K^{(1)} \in \mathbb{R}^{c \times 2 \times 2}$ and $K^{(2)}, \ldots, K^{(N)}$ are the corresponding sliced kernels that are defined on the dimension $d-1$.
\end{enumerate}
\end{definition}
\vspace{0.5em}
\noindent
Similar to the convolution-base slicer in Definition~\ref{def:linearslicer}, the convolution-stride slicer reduces the width and the height of the image by half after each convolution operator. We use the same procedure of reducing the height and the width of the image by one when the height and the width of the image are odd. The benefit of the convolution-stride slicer is that the size of its kernels does not depend on the width and the height of images as that of the convolution-base slicer. This difference improves the computational complexity and time complexity of the convolution-stride slicer over those of the convolution-base slicer (cf. Proposition~\ref{proposition:space_time_complexities}).
\vspace{0.5em}
\noindent
Now, we discuss the next variant of convolution slicer, named convolution-dilation slicer, where we include dilation with appropriate size into the convolution operators.
\begin{definition}
\label{def:csdslicer}
(Convolution-dilation Slicer) Given $X \in \mathbb{R}^{c \times d \times d}$ ($d \geq 2$),
\begin{enumerate}
\item When $d$ is even, $N = [\log_2 d]$, sliced kernels are defined as $K^{(1)}\in \mathbb{R}^{c \times 2 \times 2}$ and $K^{(h)}\in \mathbb{R}^{1 \times 2 \times 2}$ for $h =2,\ldots,N-1$, and $K^{(N)}\in \mathbb{R}^{1 \times a \times a}$ where $a= \frac{d}{2^{N-1}}$.
Then, the \emph{convolution-dilation slicer} $\mathcal{CS}\text{-d}(X|K^{(1)},\ldots,K^{(N)})$ is defined as:
\begin{align}
\mathcal{CS}\text{-d}(X|K^{(1)},\ldots,K^{(N)}) = X^{(N)}, \quad X^{(h)} = \begin{cases} X &h=0\\ X^{(h-1)} \conv{1,2}K^{(h)} & 1 \leq h \leq N-1, \\
X^{(h-1)} \conv{1,1}K^{(h)} & h=N,
\end{cases}
\end{align}
\item When $d$ is odd, the \emph{convolution-dilation slicer} $\mathcal{CS}\text{-d}(X|K^{(1)},\ldots,K^{(N)})$ takes the form:
\begin{align}
\mathcal{CS}\text{-d}(X|K^{(1)},\ldots,K^{(N)}) = \mathcal{CS}\text{-d} (X \conv{1,1} K^{(1)}|K^{(2)},\ldots,K^{(N)}),
\end{align}
where $K^{(1)} \in \mathbb{R}^{c \times 2 \times 2}$ and $K^{(2)}, \ldots, K^{(N)}$ are the corresponding sliced kernels that are defined on the dimension $d-1$.
\end{enumerate}
\end{definition}
\vspace{0.5em}
\noindent
As with the previous slicers, the convolution-dilation slicer also reduces the width and the height of the image by half after each convolution operator and it uses the same procedure for the odd dimension cases. The design of kernels' size of the convolution-dilation slicer is the same as that of the convolution-stride slicer. However, the convolution-dilation slicer has a bigger receptive field in each convolution operator which might be appealing when the information of the image is presented by a big block of pixels.
\vspace{0.5 em}
\noindent
\textbf{Computational and projection memories complexities of the convolution slicers:} We now establish the computational and projection memory complexities of convolution-base, convolution-stride, and convolution-dilation slicers in the following proposition. We would like to recall that the projection memory complexity is the memory that is needed to store a slice (convolution kernels).
\begin{proposition}
\label{proposition:space_time_complexities}
(a) When $d$ is even, the computational and projection memory complexities of convolution-base slicer are respectively at the order of $\mathcal{O}(cd^4)$ and $\mathcal{O}(c d^2)$. When $d$ is odd, these complexities are at the order of $\mathcal{O}(cd^2 + d^4)$ and $\mathcal{O}(c + d^2)$.
\noindent
(b) The computational and projection memory complexities of convolution-stride slicer are respectively at the order of $\mathcal{O}(cd^2)$ and $\mathcal{O}(c + [\log_{2} d])$.
\noindent
(c) The computational and projection memory complexities of convolution-dilation slicer are respectively at the order of $\mathcal{O}(cd^2)$ and $\mathcal{O}(c + [\log_{2} d])$.
\end{proposition}
\noindent
Proof of Proposition~\ref{proposition:space_time_complexities} is in Appendix~\ref{subsec:proof:proposition:space_time_complexities}. We recall that the computational complexity and the projection memory complexity of the conventional slicing in sliced Wasserstein are $\mathcal{O}(cd^2)$ and $\mathcal{O}(cd^2)$. We can observe that the convolution-base slicer has a worse computational complexity than the conventional slicing while having the same projection memory complexity. Since the size of kernels does not depend on the size of images, the convolution-stride slicer and the convolution-dilation slicer have the same computational complexity as the conventional slicing $\mathcal{O}(cd^2)$. However, their projection memory complexities are cheaper than conventional slicing, namely, $\mathcal{O}(c+ [\log_{2} d])$ compared to $\mathcal{O}(cd^2)$.
\begin{figure*}[!t]
\begin{center}
\begin{tabular}{c}
\widgraph{1\textwidth}{figures/CSW.pdf}
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{The convolution slicing process (using the convolution slicer). The images $X_{1}, \ldots, X_{n} \in \mathbb{R}^{c \times d \times d}$ are directly mapped to a scalar by a sequence of convolution functions which have kernels as random tensors. This slicing process leads to the convolution sliced Wasserstein~(\ref{eq:csw}) on images.
}
}
\label{fig:csw}
\vskip -0.1in
\end{figure*}
\vspace{0.5 em}
\noindent
\textbf{Non-linear convolution-base slicer: } The composition of convolution functions in the linear convolution slicer and its linear variants is still a linear function, which may not be effective when the data lie in a complex and highly non-linear low-dimensional subspace. A natural generalization of linear convolution slicers to enhance the ability of the slicers to capture the non-linearity of the data is to apply a non-linear activation function after convolution operators. This enables us to define a non-linear slicer in Definition~\ref{def:nonlinearslicer} in Appendix~\ref{sec:addmarterial}. The non-linear slicer can be seen as a defining function in generalized Radon Transform~\cite{radon20051} which was used previously in generalized sliced Wasserstein~\cite{kolouri2019generalized}.
\subsection{Convolution Sliced Wasserstein}
\label{subsec:csw}
With the definition of convolution slicers on hand, we now state our general definition of convolution sliced Wasserstein. An illustration of the convolution sliced Wasserstein is given in Figure~\ref{fig:csw}.
\begin{definition}
\label{def:csw} For any $p \geq 1$, the \emph{convolution sliced Wasserstein} (CSW) of order $p >0$ between two given probability measures $\mu, \nu \in \mathcal{P}_p(\mathbb{R}^{c \times d \times d})$ is given by:
\begin{align}
\label{eq:csw}
& \text{CSW}_p (\mu,\nu) : = \nonumber \\
& \left(\mathbb{E}_{K^{(1)}\sim \mathcal{U}(\mathcal{K}^{(1)}),\ldots, K^{(N)}\sim \mathcal{U}(\mathcal{K}^{(N)})} \left[W^p_p\left(\mathcal{S}(\cdot|K^{(1)}, \ldots, K^{(N)}) \sharp \mu, \mathcal{S}(\cdot|K^{(1)}, \ldots, K^{(N)})\sharp \nu\right)\right]\right)^{\frac{1}{p}},
\end{align}
where $\mathcal{S}(\cdot|K^{(1)}, \ldots, K^{(N)})$ is a convolution slicer with $K^{(i)} \in \mathbb{R}^{c^{(i)}\times k^{(i)} \times k^{(i)}}$ for any $i \in [N]$ and $\mathcal{U}(\mathcal{K}^{(i)})$ is the uniform distribution with the realizations being in the set $\mathcal{K}^{(i)}$ which is defined as $\mathcal{K}^{(i)}:=\left\{K^{(i)} \in \mathbb{R}^{c^{(i)}\times k^{(i)} \times k^{(i)}}| \sum_{h=1}^{c^{(i)}}\sum_{i'=1}^{k^{(i)}}\sum_{j'=1}^{k^{(i)}}K^{(i)2}_{h,i',j'} =1\right\}$, namely, the set $\mathcal{K}^{(i)}$ consists of tensors $K^{(i)}$ whose squared $\ell_{2}$ norm is 1.
\end{definition}
\noindent
When we specifically consider the convolution slicer as convolution-base slicer ($\mathcal{CS}\text{-b}$), convolution-stride slicer ($\mathcal{CS}\text{-s}$), and convolution-dilation slicer ($\mathcal{CS}\text{-d}$), we have the corresponding notions of convolution-base sliced Wasserstein (\text{CSW}-b), convolution-stride sliced Wasserstein (\text{CSW}-s), and convolution-dilation sliced Wasserstein (\text{CSW}-d).
\vspace{0.5em}
\noindent
\textbf{Monte Carlo estimation and implementation:} Similar to the conventional sliced Wasserstein, the expectation with respect to kernels $K^{(1)}, \ldots, K^{(N)}$ uniformly drawn from the sets $\mathcal{K}^{(1)}, \ldots, \mathcal{K}^{(N)}$ in the convolution sliced Wasserstein is intractable to compute. Therefore, we also make use of Monte Carlo method to approximate the expectation, which leads to the following approximation of the convolution sliced Wasserstein:
\begin{align}
\text{CSW}_p (\mu,\nu) \approx \frac{1}{L} \sum_{i = 1}^{L} W^p_p\left(\mathcal{S}(\cdot|K^{(1)}_i, \ldots, K^{(N)}_i) \sharp \mu, \mathcal{S}(\cdot|K^{(1)}_i, \ldots, K^{(N)}_i)\sharp \nu\right), \label{eq:Monte_Carlo_approx_CSW}
\end{align}
where $K^{(\ell)}_i$ are uniform samples from the sets $\mathcal{K}^{(\ell)}$ for any $\ell \in [N]$ and $i \in [L]$. Since each of the convolution slicer $\mathcal{S}(\cdot|K^{(1)}_i, \ldots, K^{(N)}_i)$ is in one dimension, we can utilize the closed-form expression of Wasserstein metric in one dimension to compute $W_p\left(\mathcal{S}(\cdot|K^{(1)}_i, \ldots, K^{(N)}_i) \sharp \mu, \mathcal{S}(\cdot|K^{(1)}_{i}, \ldots, K^{(N)}_{i})\sharp \nu\right)$ with a complexity of $\mathcal{O}(m \log_2 m)$ for each $i \in [L]$ where $m$ is the maximum number of supports of $\mu$ and $\nu$. Therefore, the total computational complexity of computing the Monte Carlo approximation~(\ref{eq:Monte_Carlo_approx_CSW}) is $\mathcal{O}(L m \log_2 m)$ when the probability measures $\mu$ and $\nu$ have at most $m$ supports. It is comparable to the computational complexity of sliced Wasserstein on images~(\ref{eq:SWimage}) where we directly vectorize the images and apply the Radon transform to these flatten images. Finally, for the implementation, we would like to remark that $L$ convolution slicers in equation~(\ref{eq:Monte_Carlo_approx_CSW}) can be computed \textit{independently} and \textit{parallelly} using the group convolution implementation which is supported in almost all libraries.
\vspace{0.5 em}
\noindent
\textbf{Properties of convolution sliced Wasserstein:} We first have the following result for the metricity of the convolution sliced Wasserstein.
\begin{theorem}
\label{theorem:metricity_convolution_sliced}
For any $p \geq 1$, the convolution sliced Wasserstein $\text{CSW}_p(.,.)$ is a pseudo-metric on the space of probability measures on $\mathbb{R}^{c \times d \times d}$, namely, it is symmetric, satisfies the triangle inequality, and $\text{CSW}_{p}(\mu, \nu) = 0 \notiff \mu = \nu$.
\end{theorem}
\noindent
Proof of Theorem~\ref{theorem:metricity_convolution_sliced} is in Appendix~\ref{subsec:proof:theorem:metricity_convolution_sliced}. Our next result establishes the connection between the convolution sliced Wasserstein and max-sliced Wasserstein and Wasserstein distances.
\begin{proposition}
\label{proposition:connection_sliced}
For any $p \geq 1$, we find that
\begin{align*}
\text{CSW}_p (\mu,\nu) \leq \text{max-SW}_p(\mu,\nu) \leq W_{p}(\mu, \nu),
\end{align*}
where $\text{max-SW}_p(\mu,\nu) : = \max_{\theta \in \mathbb{R}^{cd^2}: \|\theta\| \leq 1} \text{W}_p (\theta \sharp \mu,\theta \sharp \nu)$ is max-sliced Wasserstein metric of order $p$ between $\mu$ and $\nu$.
\end{proposition}
\vspace{0.5em}
\noindent
Proof of Proposition~\ref{proposition:connection_sliced} is in Appendix~\ref{subsec:proof:proposition:connection_sliced}. Given the bounds in Proposition~\ref{proposition:connection_sliced}, we demonstrate that the convolution sliced Wasserstein does not suffer from the curse of dimensionality for the inference purpose, namely, the sample complexity for the empirical distribution from i.i.d. samples to approximate their underlying distribution is at the order of $\mathcal{O}(n^{-1/2})$.
\begin{proposition}
\label{proposition:rate_convolution}
Assume that $P$ is a probability measure supported on compact set of $\mathbb{R}^{c \times d \times d}$. Let $X_{1}, X_{2}, \ldots, X_{n}$ be i.i.d. samples from $P$ and we denote $P_{n} = \frac{1}{n} \sum_{i = 1}^{n} \delta_{X_{i}}$ as the empirical measure of these data. Then, for any $p \geq 1$, there exists a universal constant $C > 0$ such that
\begin{align*}
\mathbb{E} [\text{CSW}_p (P_{n},P)] \leq C \sqrt{\frac{(cd^2 + 1) \log n}{n}},
\end{align*}
where the outer expectation is taken with respect to the data $X_{1}, X_{2}, \ldots, X_{n}$.
\end{proposition}
\noindent
Proof of Proposition~\ref{proposition:rate_convolution} is in Appendix~\ref{subsec:proof:proposition:rate_convolution}. The result of Proposition~\ref{proposition:rate_convolution} indicates that the sample complexity of the convolution sliced Wasserstein is comparable to that of the sliced Wasserstein on images~(\ref{eq:SWimage}), which is at the order of $\mathcal{O}(n^{-1/2})$~\cite{Bobkov_2019}, and better than that of the Wasserstein metric, which is at the order of $\mathcal{O}(n^{-1/(2cd^2)})$~\cite{Fournier_2015}.
\vspace{0.5 em}
\noindent
\textbf{Extension to non-linear convolution sliced Wasserstein:} In Appendix~\ref{sec:addmarterial}, we provide a non-linear version of the convolution sliced Wasserstein, named non-linear convolution sliced Wasserstein. The high-level idea of the non-linear version is to incorporate non-linear activation functions to the convolution-base, convolution-stride, and convolution-dilation slicers. The inclusion of non-linear activation functions is to enhance the ability of slicers to capture the non-linearity of the data. By plugging these non-linear convolution slicers into the general definition of the convolution sliced Wasserstein in Definition~\ref{def:csw}, we obtain the non-linear variants of convolution sliced Wasserstein. Details of these variants are in Appendix~\ref{sec:addmarterial}.
\section{Related Works}
\label{sec:relatedworks}
Sliced Wasserstein is used for the pooling mechanism for aggregating a set of features in~\cite{naderializadeh2021pooling}. Sliced Wasserstein gradient flows are investigated in~\cite{liutkus2019sliced,bonet2021sliced}. Variational inference based on sliced Wasserstein is carried out in~\cite{yi2021sliced}. Similarly, sliced Wasserstein is used for approximate Bayesian computation in~\cite{nadjahi2020approximate}. Statistical guarantees of training generative models with sliced Wasserstein is derived in~\cite{nadjahi2019asymptotic}. Other frameworks for generative modeling using sliced Wasserstein are sliced iterative normalizing flows~\cite{dai2021sliced} and run-sort-rerun for fine-tuning pre-trained model~\cite{lezama2021run}. Differentially private sliced Wasserstein is proposed in~\cite{rakotomamonjy2021differentially}. Approximating Wasserstein distance based on one-dimensional transportation plans from orthogonal projecting directions is introduced in~\cite{rowland2019orthogonal}. To reduce the projection complexity of sliced Wasserstein, a biased approximation based on the concentration of Gaussian projections is proposed in~\cite{nadjahi2021fast}. Augmenting probability measures to a higher-dimensional space for a better linear separation is used in augmented sliced Wasserstein~\cite{chen2022augmented}. Projected Robust Wasserstein (PRW) metrics that find the best orthogonal linear projecting operator onto $k>1$ dimensional space and Riemannian optimization techniques for solving it are proposed in~\cite{paty2019subspace, lin2020projection,huang2021riemannian}. Sliced Gromov Wasserstein, a fast sliced version of Gromov Wasserstein, is proposed in~\cite{titouan2019sliced}. The slicing technique is also be applied in approximating mutual information~\cite{goldfeld2021sliced}. We would like to recall that all the above works assume working with vector spaces and need to use vectorization when dealing with images.
\section{Experiments}
\label{sec:experiments}
In this section, we focus on comparing the sliced Wasserstein (SW) (with the conventional slicing), the convolution-base sliced Wasserstein (CSW-b), the convolution sliced Wasserstein with stride (CSW-s), and the convolution sliced Wassersstein with dilation (CSW-d) on tasks that involve probability measures over images. In particular, we first show the values of the SW and the CSW variants between probability measures over digits of the MNIST dataset~\cite{lecun1998gradient}. We then compare SW and several variants of CSW in training generative models on image datasets such as CIFAR10 (32x32)~\cite{krizhevsky2009learning}, STL10 (96x96)~\cite{coates2011analysis}, CelebA (64x64), and CelebA-HQ (128x128)~\cite{liu2015faceattributes}. We recall that the number of projections in SW and CSW's variants is denoted as $L$.
\subsection{Comparing Measures over MNIST's digits}
In the MNIST dataset, there are 60000 images of size $28 \times 28$ of digits from 0 to 9. We compute SW between two empirical probability measures over images of every two digits, e.g., 1 and 2, 1 and 3, and so on. To compare on the same digit, e.g, 1, we split images of the same digit into two disjoint sets and then compute the SW between the corresponding empirical probability measures.
\begin{table}[!h]
\centering
\caption{\footnotesize{Values of SW and CSW variants between probability measures over digits images on MNIST with $L=100$}}
\scalebox{0.65}{
\begin{tabular}{cc|c|c|c|c|c|c|c|c|c|c}
\toprule
&&\multicolumn{1}{c|}{0}&\multicolumn{1}{c|}{1}&\multicolumn{1}{c|}{2}&\multicolumn{1}{c|}{3}&\multicolumn{1}{c|}{4}&\multicolumn{1}{c|}{5}&\multicolumn{1}{c|}{6}&\multicolumn{1}{c|}{7}&\multicolumn{1}{c|}{8}&\multicolumn{1}{c}{9}
\\
\midrule
\multirow{4}{*}{0}
&SW&0.58$\pm$0.01&23.19$\pm$0.88&15.81$\pm$0.88&15.31$\pm$0.83&17.25$\pm$0.57&12.45$\pm$0.91&16.44$\pm$0.8&17.71$\pm$0.71&15.8$\pm$1.12&18.14$\pm$0.94 \\
&CSW-b&0.83$\pm$0.03&32.33$\pm$3.02&24.86$\pm$2.11&25.73$\pm$2.43&24.71$\pm$2.55&18.6$\pm$1.76&21.86$\pm$1.71&25.6$\pm$1.72&27.24$\pm$2.36&24.93$\pm$0.92 \\
&CSW-s&0.59$\pm$0.04&24.13$\pm$2.36&16.95$\pm$1.21&15.21$\pm$2.02&19.2$\pm$1.33&13.33$\pm$1.85&18.0$\pm$1.57&18.04$\pm$2.21&15.51$\pm$2.21&17.99$\pm$2.64 \\
&CSW-d&0.59$\pm$0.01&22.65$\pm$1.47&16.15$\pm$1.28&16.79$\pm$0.79&17.91$\pm$0.65&12.6$\pm$1.28&17.81$\pm$1.28&18.53$\pm$1.54&14.85$\pm$1.76&16.93$\pm$0.97 \\
\midrule
\multirow{4}{*}{1}
&SW&22.36$\pm$0.92&0.45$\pm$0.0&16.48$\pm$1.24&16.26$\pm$0.48&16.58$\pm$0.79&15.53$\pm$0.37&16.95$\pm$1.04&15.71$\pm$0.8&14.59$\pm$0.45&15.82$\pm$0.67 \\
&CSW-b&34.71$\pm$1.82&0.65$\pm$0.02&24.19$\pm$2.05&25.62$\pm$1.61&27.75$\pm$1.6&23.7$\pm$1.92&28.07$\pm$0.58&27.05$\pm$2.75&23.84$\pm$1.37&25.44$\pm$0.93 \\
&CSW-s&22.59$\pm$3.07&0.45$\pm$0.03&16.04$\pm$1.25&17.2$\pm$0.8&16.25$\pm$1.13&15.7$\pm$1.3&17.37$\pm$1.37&15.87$\pm$0.76&15.85$\pm$0.96&17.08$\pm$0.96 \\
&CSW-d&23.48$\pm$1.47&0.46$\pm$0.01&16.41$\pm$0.73&16.39$\pm$0.74&16.93$\pm$0.99&15.01$\pm$0.74&16.85$\pm$1.02&16.48$\pm$0.99&15.22$\pm$0.78&15.76$\pm$0.8 \\
\midrule
\multirow{4}{*}{2}
&SW&16.03$\pm$0.84&16.4$\pm$0.29&0.62$\pm$0.02&12.9$\pm$0.53&12.98$\pm$1.39&12.83$\pm$0.39&11.11$\pm$0.31&16.41$\pm$0.54&11.35$\pm$0.79&14.61$\pm$0.75 \\
&CSW-b&24.7$\pm$0.84&24.57$\pm$1.05&0.89$\pm$0.05&19.56$\pm$1.07&19.09$\pm$0.48&20.65$\pm$1.91&17.95$\pm$0.94&20.9$\pm$1.96&16.98$\pm$1.21&18.81$\pm$0.66 \\
&CSW-s&16.38$\pm$1.76&16.3$\pm$0.87&0.64$\pm$0.03&11.92$\pm$0.89&14.81$\pm$2.17&11.42$\pm$1.09&11.3$\pm$0.85&15.27$\pm$1.29&10.58$\pm$1.38&14.84$\pm$2.31 \\
&CSW-d&16.22$\pm$0.98&17.09$\pm$0.93&0.6$\pm$0.01&13.22$\pm$0.37&13.81$\pm$0.73&11.92$\pm$0.5&12.13$\pm$1.0&16.3$\pm$0.93&11.82$\pm$1.26&15.26$\pm$1.45 \\
\midrule
\multirow{4}{*}{3}
&SW&15.89$\pm$0.82&15.7$\pm$0.63&12.6$\pm$0.96&0.57$\pm$0.01&15.04$\pm$0.93&8.89$\pm$0.57&14.96$\pm$1.34&14.8$\pm$0.46&9.85$\pm$0.62&13.52$\pm$0.77 \\
&CSW-b&26.62$\pm$1.65&25.43$\pm$3.13&18.57$\pm$1.66&0.87$\pm$0.05&22.38$\pm$2.45&14.11$\pm$1.52&23.83$\pm$2.36&24.15$\pm$1.44&17.0$\pm$1.84&19.68$\pm$1.21 \\
&CSW-s&16.71$\pm$1.88&16.25$\pm$1.41&12.31$\pm$1.55&0.6$\pm$0.01&13.7$\pm$0.91&8.97$\pm$1.41&15.69$\pm$1.04&14.94$\pm$1.41&10.91$\pm$0.63&14.07$\pm$1.26 \\
&CSW-d&15.23$\pm$1.83&16.37$\pm$1.05&13.19$\pm$0.79&0.58$\pm$0.02&15.0$\pm$0.91&9.21$\pm$0.61&16.14$\pm$0.32&15.64$\pm$1.24&11.1$\pm$0.76&13.93$\pm$0.6 \\
\midrule
\multirow{4}{*}{4}
&SW&17.02$\pm$1.0&16.82$\pm$0.86&12.61$\pm$0.55&14.75$\pm$0.99&0.58$\pm$0.01&11.39$\pm$0.44&12.07$\pm$0.51&10.51$\pm$0.56&12.43$\pm$0.78&6.78$\pm$0.47 \\
&CSW-b&26.86$\pm$2.04&26.44$\pm$1.75&18.91$\pm$2.74&22.08$\pm$1.47&0.83$\pm$0.03&18.51$\pm$1.15&18.49$\pm$1.35&18.95$\pm$1.67&17.29$\pm$2.19&10.54$\pm$0.69 \\
&CSW-s&16.2$\pm$2.1&15.65$\pm$1.16&13.94$\pm$1.92&15.23$\pm$1.32&0.58$\pm$0.03&11.29$\pm$2.18&12.33$\pm$1.05&11.07$\pm$0.9&12.39$\pm$1.71&7.84$\pm$0.79 \\
&CSW-d&17.34$\pm$1.77&17.28$\pm$1.27&13.08$\pm$1.54&15.3$\pm$0.67&0.57$\pm$0.01&12.0$\pm$0.52&13.23$\pm$0.44&11.98$\pm$0.71&11.39$\pm$0.75&7.26$\pm$0.51 \\
\midrule
\multirow{4}{*}{5}
&SW&11.77$\pm$0.36&14.55$\pm$0.93&12.64$\pm$0.47&8.7$\pm$0.71&10.68$\pm$1.3&0.64$\pm$0.01&11.83$\pm$0.83&12.54$\pm$0.2&8.99$\pm$0.78&10.4$\pm$0.75 \\
&CSW-b&20.55$\pm$1.98&25.31$\pm$2.14&19.68$\pm$0.92&13.55$\pm$1.5&18.43$\pm$1.22&0.91$\pm$0.02&16.55$\pm$1.0&17.45$\pm$0.8&14.4$\pm$1.07&15.85$\pm$1.21 \\
&CSW-s&13.04$\pm$0.61&15.15$\pm$1.18&12.65$\pm$0.94&8.27$\pm$1.01&11.83$\pm$0.85&0.62$\pm$0.01&12.04$\pm$1.0&12.36$\pm$1.48&8.64$\pm$0.55&10.8$\pm$0.97 \\
&CSW-d&11.79$\pm$1.28&15.31$\pm$1.15&13.54$\pm$1.22&8.82$\pm$1.07&12.33$\pm$0.75&0.62$\pm$0.04&12.45$\pm$0.79&13.02$\pm$0.81&9.18$\pm$0.54&10.73$\pm$0.85 \\
\midrule
\multirow{4}{*}{6}
&SW&15.97$\pm$0.87&16.84$\pm$1.4&11.52$\pm$0.53&15.56$\pm$0.66&12.09$\pm$0.63&11.98$\pm$0.82&0.65$\pm$0.01&16.69$\pm$1.63&12.52$\pm$0.42&13.84$\pm$0.93 \\
&CSW-b&25.66$\pm$2.37&26.39$\pm$0.68&15.93$\pm$0.91&22.98$\pm$3.47&18.8$\pm$1.9&17.0$\pm$1.66&0.91$\pm$0.02&23.31$\pm$2.45&17.62$\pm$0.99&18.73$\pm$0.84 \\
&CSW-s&17.84$\pm$1.85&17.61$\pm$1.92&11.49$\pm$0.42&14.07$\pm$1.43&12.25$\pm$1.23&11.74$\pm$0.35&0.66$\pm$0.01&15.71$\pm$1.03&13.33$\pm$0.68&12.55$\pm$1.4 \\
&CSW-d&16.95$\pm$1.45&17.15$\pm$1.12&11.47$\pm$0.79&15.71$\pm$1.24&11.91$\pm$0.37&12.63$\pm$0.94&0.67$\pm$0.02&16.36$\pm$1.29&13.15$\pm$1.0&14.35$\pm$0.92 \\
\midrule
\multirow{4}{*}{7}
&SW&17.55$\pm$1.35&16.65$\pm$0.79&15.3$\pm$0.83&15.47$\pm$0.73&11.39$\pm$0.77&12.4$\pm$0.54&16.04$\pm$1.19&0.61$\pm$0.01&13.66$\pm$1.12&8.16$\pm$0.06 \\
&CSW-b&27.36$\pm$2.07&28.35$\pm$1.32&22.24$\pm$1.59&23.56$\pm$1.2&18.46$\pm$2.75&19.32$\pm$1.68&25.38$\pm$1.94&0.94$\pm$0.04&22.63$\pm$1.67&14.71$\pm$0.52 \\
&CSW-s&16.74$\pm$2.14&15.81$\pm$1.23&17.72$\pm$2.26&14.75$\pm$0.83&13.6$\pm$2.24&13.45$\pm$1.94&15.37$\pm$1.44&0.64$\pm$0.05&12.92$\pm$0.77&8.95$\pm$1.3 \\
&CSW-d&18.21$\pm$1.44&16.31$\pm$1.55&16.3$\pm$1.05&14.97$\pm$0.76&11.45$\pm$0.35&12.82$\pm$1.54&16.9$\pm$0.95&0.69$\pm$0.04&13.3$\pm$0.59&8.72$\pm$0.48 \\
\midrule
\multirow{4}{*}{8}
&SW&16.16$\pm$1.14&15.09$\pm$0.96&11.02$\pm$0.54&10.02$\pm$0.79&11.45$\pm$0.69&8.46$\pm$0.75&13.41$\pm$0.29&14.33$\pm$1.27&0.65$\pm$0.02&10.62$\pm$0.35 \\
&CSW-b&26.49$\pm$2.12&21.76$\pm$0.63&15.73$\pm$1.07&17.16$\pm$1.58&18.25$\pm$1.36&14.5$\pm$0.94&18.87$\pm$1.68&21.36$\pm$1.76&0.97$\pm$0.04&15.85$\pm$0.81 \\
&CSW-s&17.19$\pm$1.17&14.26$\pm$1.07&11.01$\pm$0.79&10.32$\pm$1.02&11.86$\pm$1.4&8.75$\pm$0.63&13.23$\pm$0.96&13.72$\pm$1.3&0.66$\pm$0.04&10.65$\pm$1.02 \\
&CSW-d&15.42$\pm$1.31&15.41$\pm$0.87&11.06$\pm$0.43&10.56$\pm$0.44&12.51$\pm$1.74&8.98$\pm$0.61&13.87$\pm$1.29&14.77$\pm$0.67&0.65$\pm$0.03&11.09$\pm$1.06 \\
\midrule
\multirow{4}{*}{9}
&SW&17.94$\pm$1.19&15.68$\pm$0.64&13.83$\pm$1.05&12.72$\pm$0.48&7.37$\pm$0.66&10.62$\pm$0.92&13.54$\pm$0.48&8.24$\pm$0.31&10.66$\pm$0.38&0.59$\pm$0.02 \\
&CSW-b&26.67$\pm$3.65&26.0$\pm$1.95&20.52$\pm$1.24&19.68$\pm$1.14&10.39$\pm$0.42&16.36$\pm$2.03&19.24$\pm$0.99&14.95$\pm$1.29&15.71$\pm$1.44&0.84$\pm$0.04 \\
&CSW-s&16.73$\pm$1.84&16.04$\pm$1.28&14.31$\pm$1.66&13.22$\pm$1.43&7.42$\pm$0.45&10.32$\pm$0.65&13.74$\pm$2.08&8.64$\pm$0.8&10.52$\pm$1.33&0.6$\pm$0.02 \\
&CSW-d&17.58$\pm$1.17&15.43$\pm$1.09&13.98$\pm$0.51&13.55$\pm$1.43&7.18$\pm$0.37&10.89$\pm$0.58&13.94$\pm$1.11&8.58$\pm$0.42&11.68$\pm$0.62&0.6$\pm$0.02 \\
\bottomrule
\end{tabular}
}
\label{tab:MNIST_L100}
\end{table}
\vspace{0.5em}
\noindent
\textbf{Meaningful measures of discrepancy:} We approximate the SW and the CSW's variants by a finite number of projections, namely, $L=1$, $L=10$, and $L=100$. We show the mean of approximated values of $L=100$ over 5 different runs and the corresponding standard deviation in Table~\ref{tab:MNIST_L100}. According to the table, we observe that SW and CSW's variants can preserve discrepancy between digits well. In particular, the discrepancies between probability measures of the same digit are relatively small compared to the discrepancies between probability measures of different digits. Moreover, we see that the values of CSW-s and CSW-d are closed to the values of SW on the same pairs of digits. We also show similar tables for $L=1$ and $L=10$ in Tables~\ref{tab:MNIST_L1}-\ref{tab:MNIST_L10} in Appendix~\ref{sec:add_experiments}. From these tables, we observe that the number of projections can affect the stability of both SW and CSW's variants.
\begin{table}[!t]
\centering
\caption{\footnotesize{Summary of FID and IS scores of methods on CIFAR10 (32x32), CelebA (64x64), STL10 (96x96), and CelebA-HQ (128x128).}}
\scalebox{0.9}{
\begin{tabular}{l|cc|c|cc|c}
\toprule
\multirow{2}{*}{Method}& \multicolumn{2}{c|}{CIFAR10 (32x32)}&\multicolumn{1}{c|}{CelebA (64x64)}&\multicolumn{2}{c|}{STL10 (96x96)}&\multicolumn{1}{c|}{CelebA-HQ (128x128)}\\
\cmidrule{2-7}
& FID ($\downarrow$) &IS ($\uparrow$)& FID ($\downarrow$) & FID ($\downarrow$) &IS ($\uparrow$)& FID ($\downarrow$) \\
\midrule
SW (L=1)&87.97&3.59&128.81&170.96&3.68&\textbf{275.44}\\
CSW-b (L=1)&84.38&4.28&85.83&173.33&\textbf{3.89}&315.91 \\
CSW-s (L=1)&80.10&4.31&\textbf{66.52}&\textbf{168.9}3&3.75 &303.57\\
CSW-d (L=1)&\textbf{63.94}&\textbf{4.89} &89.37&212.61&2.48&321.06\\
\midrule
SW (L=100)&53.67&5.74&20.08&100.35&8.14&51.80\\
CSW-b (L=100)&49.78&5.78 &18.96&\textbf{91.75}&8.11 &53.05\\
CSW-s (L=100)&\textbf{43.88}&\textbf{6.13} & \textbf{13.76}&97.08&\textbf{8.20}&\textbf{32.94}\\
CSW-d (L=100)&47.16&5.90&14.96&102.58&7.53&41.01 \\
\midrule
SW (L=1000)&43.11&6.09&14.92 &84.78&9.06& 28.19\\
CSW-b (L=1000)&43.17&6.07&14.75&86.98&9.11 &29.69\\
CSW-s (L=1000)&\textbf{35.40}&\textbf{6.64}&\textbf{12.55} &\textbf{77.24}&9.31&\textbf{22.25}\\
CSW-d (L=1000)&41.34&6.33 &13.24&83.36&\textbf{9.42} & 25.93\\
\bottomrule
\end{tabular}
}
\label{tab:summary}
\end{table}
\begin{figure*}[!t]
\begin{center}
\begin{tabular}{ccc}
\widgraph{0.3\textwidth}{CIFAR/fid_cifar.pdf}
&
\widgraph{0.3\textwidth}{CIFAR/is_cifar.pdf}
&
\widgraph{0.3\textwidth}{CelebA/fid_celeba.pdf}
\\
\widgraph{0.3\textwidth}{STL/fid_stl.pdf}
&
\widgraph{0.3\textwidth}{STL/is_stl.pdf}
&
\widgraph{0.3\textwidth}{CelebAHQ/fid_celebahq.pdf}
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{FID scores and IS scores over epochs of different training losses on datasets. We observe that CSW's variants usually help the generative models converge faster.
}
}
\label{fig:iterCIFAR}
\vskip -0.1in
\end{figure*}
\vspace{0.5em}
\noindent
\textbf{Projection memory for slicers:} For SW, the conventional slicing requires $L\cdot 784$ float variables for $L$ projecting directions of $28\cdot 28$ dimension. On the other hand, CSW only needs $L \cdot 338$ float variables since each projecting direction is represented as three kernels $K^{(1)} \in \mathbb{R}^{15\times 15}$, $K^{(2)} \in \mathbb{R}^{8\times 8}$, and $K^{(3)} \in \mathbb{R}^{7\times 7}$. More importantly, CSW-s and CSW-d require only $L\cdot 57$ float variables since they are represented by three kernels $K^{(1)} \in \mathbb{R}^{2\times 2}$, $K^{(2)} \in \mathbb{R}^{2\times 2}$, and $K^{(3)} \in \mathbb{R}^{7\times 7}$. From this experiment, we can see that using the whole unit-hypersphere as the space of projecting directions can be sub-optimal when dealing with images.
\subsection{Generative Models}
We follow the framework of sliced Wasserstein generator in~\cite{deshpande2018generative}. In particular, we parameterize the model distribution $p_\phi (x) \in \mathcal{P}(\mathbb{R}^{c \times d \times d})$ and $p_\phi (x) = G_\phi \sharp \epsilon$ where $\epsilon$ is the standard multivariate Gaussian of 128 dimension and $G_\phi$ is a neural network with Resnet architecture~\cite{he2016deep}. Since the ground truth metric between
images is unknown, we need a discriminator as a type of ground metric learning. We denote the discriminator as a function $T_{\beta_2} \circ T_{\beta_1}$ where $T_{\beta_1}:\mathbb{R}^{c \times d \times d } \to \mathbb{R}^{c' \times d' \times d'}$ and $T_{\beta_2}: \mathbb{R}^{c' \times d' \times d'} \to \mathbb{R}$. In greater detail, $T_{\beta_1}$ maps the original images to their corresponding features maps and $T_{\beta_2}$ maps their features maps to their corresponding discriminative scores. Let the data distribution is $\mu$, our training objectives are:
\begin{align*}
&\min_{\beta_1,\beta_2} \left(\mathbb{E}_{x \sim \mu} [\min (0,-1+ T_{\beta_2}(T_{\beta_1}(x)))] + \mathbb{E}_{z \sim \epsilon} [\min(0, -1-T_{\beta_2}(T_{\beta_1} (G_\phi(z))))] \right), \\
&\min_{\phi} \mathbb{E}_{X \sim \mu^{\otimes m}, Y \sim \epsilon^{\otimes m}} \mathcal{D}(T_{\beta_1} \sharp P_X, T_{\beta_1}\sharp G_\phi \sharp P_Y),
\end{align*}
where $m \geq 1$ is the mini-batch size and $\mathcal{D}(\cdot,\cdot)$ is the SW or CSW's variants.
\begin{table}[!t]
\centering
\caption{\footnotesize{Computational time and memory of methods (reported in the number of iterations per a second and megabytes (MB).}}
\scalebox{0.8}{
\begin{tabular}{l|cc|cc|cc|cc}
\toprule
\multirow{2}{*}{Method}& \multicolumn{2}{c|}{CIFAR10 (32x32)}&\multicolumn{2}{c|}{CelebA (64x64)}&\multicolumn{2}{c|}{STL10 (96x96)}&\multicolumn{2}{c}{CelebA-HQ} (128x128)\\
\cmidrule{2-9}
& Iters/s ($\uparrow$) &Mem ($\downarrow$)& Iters/s ($\uparrow$) &Mem ($\downarrow$)& Iters/s ($\uparrow$) &Mem ($\downarrow$)& Iters/s ($\uparrow$) &Mem ($\downarrow$)\\
\midrule
SW (L=1)& 18.98&2071& 6.21& 8003 & 9.59& 4596 & 10.35& 4109\\
SW (L=100)& 18.53&2080&6.16& 8015 & 9.47&4601 &10.22 & 4117\\
SW (L=1000)& 18.15&2169&6.10 & 8102 &9.13 & 4647 &10.17 &4202\\
\midrule
CSW-b (L=1)& 18.43&2070 & 6.21& 8003& 9.56& 4596 & 10.33 & 4109\\
CSW-b (L=100)& 18.35&2077& 6.15& 8009 &9.40 & 4598 &10.19 & 4110\\
CSW-b (L=1000)& 18.06&2117 & 6.10&8049 &9.07 &4613 & 10.12& 4134\\
\midrule
CSW-s (d) (L=1)& 18.69&2070 &6.21&8003 &9.56 & 4596 &10.33 & 4109\\
CSW-s (d) (L=100)& 18.50&2073&6.16 & 8005 &9.41 & 4597& 10.20&4109\\
CSW-s (d) (L=1000)& 18.10&2098 &6.10 &8029 &9.10 & 4603& 10.12 &4114\\
\bottomrule
\end{tabular}
}
\label{tab:timeandmem}
\end{table}
\begin{figure*}[!h]
\begin{center}
\begin{tabular}{ccc}
\widgraph{0.26\textwidth}{CIFAR/sw_1.png}
&
\widgraph{0.26\textwidth}{CIFAR/sw_100.png}
&
\widgraph{0.26\textwidth}{CIFAR/sw_1000.png}
\\
SW ($L=1$) & SW ($L=100$) & SW ($L=1000$)
\\
\widgraph{0.26\textwidth}{CIFAR/csws_1.png}
&
\widgraph{0.26\textwidth}{CIFAR/csws_100.png}
&
\widgraph{0.26\textwidth}{CIFAR/csws_1000.png}
\\
CSW-s ($L=1$) & CSW-s ($L=100$) & CSW-s ($L=1000$)
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{Random generated images of SW and CSW-s on CIFAR10.
}
}
\label{fig:cifar}
\vskip -0.1in
\end{figure*}
\vspace{0.5em}
\noindent
We train the above model on standard benchmarks such as CIFAR10 (32x32)~\cite{krizhevsky2009learning}, STL10 (96x96)~\cite{coates2011analysis}, CelebA (64x64), and CelebAHQ (128x128)~\cite{liu2015faceattributes}. To compare models, we use the FID score~\cite{heusel2017gans} and the Inception score (IS)~\cite{salimans2016improved}. The detailed settings about architectures, hyperparameters, and evaluation of FID and IS are given in Appendix~\ref{sec:settings}. We first show the FID scores and IS scores of generative models trained by SW and CSW's variants with the number of projections $L \in \{1,100,100\}$ in Table~\ref{tab:summary}. In the table, we report the performance of models at the last training epoch. We do not report the IS scores on CelebA and CelebA-HQ since the IS scores are not suitable for face images. We then demonstrate the FID scores and IS scores across training epochs in Figure~\ref{fig:iterCIFAR} for investigating the convergence of generative models trained by SW and CSW's variants. After that, we report the training time and training memory of SW and CSW variants in Table~\ref{tab:timeandmem}. Finally, we show some randomly generated images from SW's models and CSW-s' models in Figures~\ref{fig:cifar}-\ref{fig:celebahq}. Additional experimental results with CSW-b and CSW-d are given in Appendix~\ref{sec:add_experiments}.
\vspace{0.5em}
\noindent
\textbf{Summary of FID scores and IS scores:} According to Table~\ref{tab:summary}, on CIFAR10, CSW-d gives the lowest values of FID scores and IS scores when $L=1$ while CSW-s gives the lowest FID scores when $L=100$ and $L=1000$. Compared to CSW-s, CSW-d and CSW yield higher FID scores and lower IS scores. However, CSW-d and CSW are still better than SW. On CelebA, CSW-s perform the best in all settings. On STL10, CSW's variants are also better than the vanilla SW, however, it is hard to figure out which is the best variant. On CelebA-HQ, SW gives the lowest FID score when $L=1$. In contrast, when $L=100$ and $L=1000$, CSW-s is the best choice for training the generative model. Since the FID scores of $L=1$ are very high on CelebA-HQ and STL10, the scores are not very meaningful for comparing SW and CSW's variants. For all models, increasing $L$ leads to better generative quality. Overall, we observe that CSW's variants enhance the performance of the generative models well.
\begin{figure*}[!t]
\begin{center}
\begin{tabular}{ccc}
\widgraph{0.26\textwidth}{CelebA/sw_1.png}
&
\widgraph{0.26\textwidth}{CelebA/sw_100.png}
&
\widgraph{0.26\textwidth}{CelebA/sw_1000.png}
\\
SW ($L=1$) & SW ($L=100$) & SW ($L=1000$)
\\
\widgraph{0.26\textwidth}{CelebA/csws_1.png}
&
\widgraph{0.26\textwidth}{CelebA/csws_100.png}
&
\widgraph{0.26\textwidth}{CelebA/csws_1000.png}
\\
CSW-s ($L=1$) & CSW-s ($L=100$) & CSW-s ($L=1000$)
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{Random generated images of SW and CSW-s on CelebA.
}
}
\label{fig:celeba}
\vskip -0.1in
\end{figure*}
\vspace{0.5em}
\noindent
\textbf{FID scores and IS scores across epochs:} From Figure~\ref{fig:iterCIFAR}, we observe that CSW's variants help the generative models converge faster than SW when $L=100$ and $L=1000$. Increasing the number of projections from $100$ to $1000$, the generative models from both SW and CSW's variants become better. Overall, CSW-s is the best option for training generative models among CSW's variants since its FID curves are the lowest and its IS curves are the highest.
\vspace{0.5em}
\noindent
\textbf{Training time and training memory:} We report in Table~\ref{tab:timeandmem} the training speed in the number of iterations per second and the training memory in megabytes (MBs). We would like to recall that the time complexity and the projection memory complexity of CSW-s and CSW-d are the same. Therefore, we measure the training time and the training memory of CSW-s as the result for both CSW-s and CSW-d. We can see that increasing the number of projections $L$ costs more memory and also slows down the training speed. However, the rate of increasing memory of CSW is smaller than SW. For CSW-s and CSW-d, the extent of saving memory is even better. As an example, $L=1000$ in CSW-s and CSW-d costs less memory than SW with $L=100$ while the performance is better (see Table~\ref{tab:summary}). In terms of training time, CSW-s and CSW-d are comparable to SW and they can be computed faster than CSW. We refer the readers to Section~\ref{sec:csw} for a detailed discussion about the computational time and projection memory complexity of CSW's variants.
\begin{figure*}[!t]
\begin{center}
\begin{tabular}{ccc}
\widgraph{0.26\textwidth}{STL/sw_1.png}
&
\widgraph{0.26\textwidth}{STL/sw_100.png}
&
\widgraph{0.26\textwidth}{STL/sw_1000.png}
\\
SW ($L=1$) & SW ($L=100$) & SW ($L=1000$)
\\
\widgraph{0.26\textwidth}{STL/csws_1.png}
&
\widgraph{0.26\textwidth}{STL/csws_100.png}
&
\widgraph{0.26\textwidth}{STL/csws_1000.png}
\\
CSW-s ($L=1$) & CSW-s ($L=100$) & CSW-s ($L=1000$)
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{Random generated images of SW and CSW-s on STL10.
}
}
\label{fig:stl}
\vskip -0.1in
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\begin{tabular}{ccc}
\widgraph{0.3\textwidth}{CelebAHQ/sw_1.png}
&
\widgraph{0.3\textwidth}{CelebAHQ/sw_100.png}
&
\widgraph{0.3\textwidth}{CelebAHQ/sw_1000.png}
\\
SW ($L=1$) & SW ($L=100$) & SW ($L=1000$)
\\
\widgraph{0.3\textwidth}{CelebAHQ/csws_1.png}
&
\widgraph{0.3\textwidth}{CelebAHQ/csws_100.png}
&
\widgraph{0.3\textwidth}{CelebAHQ/csws_1000.png}
\\
CSW-s ($L=1$) & CSW-s ($L=100$) & CSW-s ($L=1000$)
\end{tabular}
\end{center}
\vskip -0.2in
\caption{
\footnotesize{Random generated images of SW and CSW-s on CelebA-HQ.
}
}
\label{fig:celebahq}
\vskip -0.1in
\end{figure*}
\vspace{0.5em}
\noindent
\textbf{Generated images:} We show randomly generated images on CIFAR10, CelebA, STL10, and CelebA-HQ in Figures~\ref{fig:cifar}-\ref{fig:celebahq} as qualitative comparison between SW and CSW-s. From the figures, we can see that generated images of CSW-s is more realistic than ones of SW. The difference is visually clear when the number of projections $L$ is small e.g., $L=1$ and $L=100$. When $L=1000$, we can still figure out that CSW-s is better than SW by looking at the sharpness of the generated images. Also, we can visually observe the improvement of SW and CSW-s when increasing the number of projections. In summary, the qualitative results are consistent with the quantitative results (FID scores and IS scores) in Table~\ref{tab:summary}. For the generated images of CSW-b and CSW-d, we show them in Figures~\ref{fig:cifar_appendix}-\ref{fig:celebahq_appendix}.
\vspace{0.5em}
\noindent
\textbf{Non-linear convolution sliced Wasserstein:} We also compare non-linear extensions of SW and CSW variants in training generative models on CIFAR10 in Appendix~\ref{sec:add_experiments}. For details of non-linear extensions, we refer to Appendix~\ref{sec:addmarterial}. From experiments, we observe that convolution can also improve the performance of sliced Wasserstein in non-linear projecting cases. Compared to linear versions, non-linear versions can enhance the quality of the generative model or yield comparable results.
\section{Conclusion}
\label{sec:conclusion}
We have addressed the issue of the conventional slicing process of sliced Wasserstein when working with probability measures over images. In particular, sliced Wasserstein is defined on probability measures over vectors which leads to the step of vectorization for images. As a result, the conventional slicing process cannot exploit the spatial structure of data for designing the space of projecting directions and projecting operators. To address the issue, we propose a new slicing process by using the convolution operator which has been shown to be efficient on images. Moreover, we investigate the computational complexity and projection memory complexity of the new slicing technique. We show that convolution slicing is comparable to conventional slicing in terms of computational complexity while being better in terms of projection memory complexity. By utilizing the new slicing technique, we derive a novel family of sliced Wassersein variants, named convolution sliced Wasserstein. We investigate the properties of the convolution sliced Wasserstein including its metricity, its computational and sample complexities, and its connection to other variants of sliced Wasserstein in literature. Finally, we carry out extensive experiments in comparing digits images and training generative models on standard benchmark datasets to demonstrate the favorable performance of the convolution sliced Wasserstein.
|
\section{The Model}
\subsection{The Transformation}
\begin{figure}
\center{\includegraphics[width=0.45\textwidth]{fig1.eps}
\caption{(Color online) Shown is a schematic diagram of the transformation.}
\label{fig1}}
\end{figure}
We begin the non-relativistic limit of in an orthorhombically anisotropic medium
\begin{eqnarray}
H_0 &=&\Bigl[\sum_{\mu=1}^3 \frac{1}{\sqrt{2m_{\mu}}} \sigma_{\mu}(-i\partial_{\mu} + e A_{\mu})\Bigr]^2,
\end{eqnarray}
where ${\bm B}(\bm{r})={\bm\nabla}\times{\bm A}(\bm{r})$ is the magnetic induction, ${\bm{A}}$ is the vector potential, $e$ is the absolute value of the
electron charge, the $\sigma_{\mu}$ are the Pauli matrices, ${\partial_{\mu}\equiv\partial/ \partial x_{\mu}}$, $x_{\mu}$ is $x$, $y$, $z$ for $\mu=1, 2,
3$, respectively, and we set $\hbar=k_B=c=1$. At $B_{c2}$, the supercooling field, or the border between the superconducting and normal state of a
non-magnetic metal, we assume that the transition is a second-order thermodynamic transition, and that first-order transitions to some other state, such as the FFLO state\cite{Fulde1964,Larkin1964,Gruenberg1966,Matsuda2007,Agosta2017}, do not occur in the ${\bm B}(T)$ range of the experiments. In addition, we assume that a competing CDW or SDW state is either absent or does not contribute to the ${\bm B}_{c2}=\mu_0{\bm H}_{c2}$ measurements. Although such competing order parameters have apparently distorted the results of phase-sensitive $c$-axis twist experiments in underdoped or optimally-doped Bi2212 \cite{Korea,Harvard}, $B_{c2}(T)$ curves in that material can only be measured close to $T_c$ \cite{Naughton}, and the transitions are so broad at lower $T$ values as to make meaningful estimates of ${\bm B}_{c2}(T)$ almost impossible\cite{Layered}. We therefore may assume $\bm{B}$ to be constant throughout the superconductor, writing
\begin{eqnarray}
{\bm{B}}& =&B(\sin\theta\cos\phi, \sin\theta\sin\phi,\cos\theta)\nonumber
\end{eqnarray}
in spherical coordinates.
Expanding the squared quantities in $H_0$, we obtain \cite{Aiying2021}
\begin{eqnarray}
H_0&=&\sum_{\mu}\Bigl[\frac{(-i\partial_{\mu} + eA_{\mu})^2}{2m_{\mu}}+\frac{e}{2 m_g} \sigma_{\mu}\sqrt{\overline{m}_{\mu}} B_{\mu}\Bigr],
\end{eqnarray}
where $m_g=(m_1m_2m_3)^{1/3}$ is the geometric mean effective mass, $\overline{m}_{\mu}=m_{\mu}/m_g$, and $\sigma_{\mu}$ is the matrix representing the $\mu^{\rm th}$ component of the electron spin and $\tilde{\bm r}=\alpha{\bm r}''$. We emphasize that the effective masses $m_{\mu}$ appear in the orbital part of the electron motion, and are not necessarily related to the rest mass $m$ of an electron in vacuum.
We then make the Klemm-Clem transformations\cite{Layered,Klemm1980}, and Fig.1 gives the schematic diagram of the transformation, after which
$H_{0}$ reduces to
\begin{eqnarray}
\tilde{H}_{0}&=&\frac{1}{2m_g}(-i{\bm\nabla}^{\prime\prime} + e{\bm A}^{\prime\prime})^2 + \alpha \frac{e}{2 m_g} \sigma_z\tilde{B}_z,\\
\alpha(\theta,\phi)\!\!&=&\!\!\![\overline{m}_1\sin^2\theta\cos^2\phi+\overline{m}_2\sin^2\theta\sin^2\phi+\overline{m}_3\cos^2\theta]^{1/2},\label{alpha}\nonumber\\
\end{eqnarray}
where $\tilde{B}_z=B_z''/\alpha(\theta,\phi)$.
We note that the transformed Zeeman energy is proportional to $\alpha(\theta,\phi)$, the direction of the magnetic field ${\bm B}({\theta,\phi})$ has changed to $({\theta^{\prime},\phi^{\prime}})$, the Laudau-level orbits scale equivalently with the transformed Fermi Surface, and the geometric mean effective mass is $m_g$.
We have quantized the rotated spins along $z^{\prime\prime}$ axis, the direction of the transformed $\tilde{\bm B}$. Thus, in second quantization, the full
Hamiltonian $H'=H-\mu N$ may now be written in transformed real space as
\begin{eqnarray}
H'&=&\sum_{\sigma=\pm}\int d^3{\bm r}^{\prime\prime} \biggl[\psi^{\dag}_{\sigma}({\bm r}^{\prime\prime})
\Bigl[\frac{1}{2m_g}\Bigl(-i{\bm\nabla}^{\prime\prime}+e{\bm A}^{\prime\prime}(\tilde{\bm r})\Bigr)^2\nonumber\\
& &\hskip15pt-\mu+ \alpha \frac{e}{2 m_g}\sigma_z \tilde{B}_z\Bigr]\psi_{\sigma}({\bm r}^{\prime\prime})\nonumber\\
&&\!+\frac{1}{2}\!\int\!d^3{\bm r_1}^{\prime\prime} \psi^{\dag}_{\sigma}({\bm r_2}^{\prime\prime}) \psi^{\dag}_{-\sigma}({\bm r}_{1}^{\prime\prime})
V({\bm r}_{2}^{\prime\prime}-{\bm r}_{1}^{\prime\prime}) \psi_{-\sigma}({\bm r}_{1}^{\prime\prime})\psi_{\sigma}({\bm r}_{2}^{\prime\prime})\biggr],\nonumber\\
\end{eqnarray}
where $\mu$ is the chemical potential, $\tilde{\bm r}=\alpha{\bm r}''$, and the untransformed pairing interactions in reciprocal space may be written as\cite{ScharnbergKlemm1980}
\begin{eqnarray}
V(\hat{\bm k},\hat{\bm k}')&=&\left\{\begin{array}{lc} V_0,& s\\
V_0(\hat{k}_x^2-\hat{k}_y^2)(\hat{k}_x^{'2}-\hat{k}_y^{'2}),&d_{x^2-y^2}\end{array}\right.\label{interaction}
\end{eqnarray}
We note that the anisotropic scale transformation also causes the magnitude and direction of $\bm{k}(\theta_k,\phi_k)$ to be changed into $k^{\prime}(\theta^{\prime},\phi^{\prime})$, since $k_{\mu}'=\frac{k_{\mu}}{\sqrt{\bar{m}_{\mu}}}$, and the overall magnitude of the wave vector changes according to $k'=\beta(\theta_{k},\phi_{k}) k$, where
$\beta(\theta_k,\phi_k)$ is given by
\begin{eqnarray}
\beta=&\Bigl[\frac{\sin^2 \theta_{k} \cos^{2}\phi_{k}}{\overline{m}_{1}}+\frac{\sin^2\theta_{k}\cos^2\phi_{k}}{\overline{m}_{2}}
+\frac{\cos^2\theta_{k}}{\overline{m}_{3}}\Bigr]^{1/2},\label{beta}
\end{eqnarray}
and the angle relationship is
\begin{eqnarray}
&&\sin\theta_{k^{\prime}}=\frac{\beta(\frac{\pi}{2},\phi_k)}{\beta(\theta_{k},\phi_k)} \sin\theta_k;
\cos\theta_{k^{\prime}}=\frac{1/\sqrt{\overline{m}_3}}{\beta(\theta_{k},\phi_k)} \cos\theta_k; \nonumber\\
&&\sin\phi_{k^{\prime}}=\frac{1/\sqrt{\overline{m}_2}}{\beta(\frac{\pi}{2},\phi_k)} \sin\phi_k;
\cos\phi_{k^{\prime}}=\frac{\beta(1/\sqrt{\overline{m}_1}}{\beta(\frac{\pi}{2},\phi_k)} \cos\theta_k; \nonumber\\
\end{eqnarray}
However, at the chemical potential (or $E_F$), we have $(k')^2=2m_g\mu$, so in integrations about the Fermi energy, we may safely approximate
\begin{eqnarray}
\beta^2k^2\Big|_{\mu}&\approx&2m_g\mu.
\end{eqnarray}
\subsection{Gap Function}
All the representation of physical quantities are shown in new rectangular coordinate system with double prime, so it is convenient to omit the double prime in the following. We define the superconducting order parameter to be
\begin{eqnarray}
\Delta_{\sigma,\sigma'}(\bm{r}, \bm{r}')&=& V(\bm{r}, \bm{r}')F_{\sigma,\sigma'}(\bm{r}, \bm{r}',0^{+}),
\end{eqnarray}
in which $V(\bm{r}, \bm{r}')$ is the pairing interaction in weak-coupling theory. It may be written as $V(\bm{r}, \bm{r}')= \lambda \delta(\bm{r}-\bm{r}')$, and then $\Delta_{\sigma,\sigma'}(\bm{r},\bm{r}')=\Delta(\bm{r})\delta(\bm{r}-\bm{r}')\delta_{\sigma,-\sigma'}$, for traditional $s$-wave pairing. For
non-$s$-wave superconductors, it is easier to define the gap function in wave vector space\cite{ScharnbergKlemm1980}.
Using Gor'kov's description of weakly coupled superconductors, we can derive the fully transformed equations of motion for the normal and anomalous
temperature Green functions of anisotropic superconductors\cite{AGD},
\begin{eqnarray}
&&\Bigl[i\omega_n-\frac{({\bm\nabla}/i - e{\bm A})^2}{2m_g}+E_{F} - \frac{e}{2 m_g}\alpha\sigma_z\tilde{B}_z\Bigr]G_{\sigma,
\sigma'}({\bm{r}},{\bm{r}'},\omega_{n}) \nonumber \\
&&+\sum_{\rho}\int d^{3} {\bm \xi}\Delta_{\sigma, \rho}({\bm{r}},{\bm \xi})F_{\rho,
\sigma'}^{\dagger}({\bm\xi},{\bm{r}'},\omega_n)=\delta_{\sigma,\sigma'}\delta({\bm{r}}-{\bm{r}'}),\nonumber\\
\end{eqnarray}
\begin{eqnarray}
&&\Bigl[-i\omega_n-\frac{(i{\bm\nabla} - e{\bm A})^2}{2m_g}+E_{F} - \alpha\frac{e}{2 mg} \sigma_z\tilde{B}_z\Bigr]F_{\sigma,
\sigma'}^{\dagger}({\bm{r}},{\bm{r}'},\omega_{n}) \nonumber \\
&&-\sum_{\rho}\int d^{3} {\bm\xi} \Delta_{\sigma, \rho}^{*}({\bm{r}}, {\bm{\xi}})G_{\rho, \sigma'}({\bm{\xi}},{\bm{r}'}, \omega_n)=0,
\end{eqnarray}
where $\omega_n=(2n+1)\pi T$ are the Matsubara frequencies.
For a charged anisotropic superfluid in a magnetic field. We use the method introduced by Scharnberg and Klemm in their study of $p$-wave
superconductors to obtain the gap function before the transformations, \cite{ScharnbergKlemm1980}
\begin{eqnarray}
\Delta_{\sigma, \sigma'}(\bm{R},\bm{k}_{F})\!&=&\!2\pi TN(0)\sum_{\omega_{n}}\int\frac{d\Omega_{\bm{k}'}}{4\pi} V(\bm{k}_{F} -
\bm{k}'_{F})\nonumber\\
&&\times\int_{0}^{\infty}d\xi e^{-2|\omega_{n}|\xi}\cos[\frac{e}{2 m_g}\alpha(\sigma_z - \sigma'_{z})B_{z}\xi] \nonumber \\
&&\times e^{-i{\rm sgn}(\omega_{n})\xi \bm{v}_{F}({\bm{\hat{k}}}_{F}^{'})\cdot{\bm\Pi}({\bm
R})}\Delta_{\sigma,\sigma^{'}}(\bm{R},\bm{k}_{F}'),\nonumber\\
\end{eqnarray}
where ${\bm\Pi}({\bm R})={\bm\nabla}_{\bm{R}}/i+2e{\bm A}({\bm R})$, $\bm{R}$ denotes the center of mass position of the paired electrons, and $N(0)$ is
the electronic density of states.
\subsection{Particular Models}
The most notable quality of an anisotropic superconductor, especially for a heavy fermion superconductor (HFS), is its anisotropic gap. Such as anisotropic
$p$-wave superconductors, one may have a nodeless gap, or a gap with either planar nodes (such as a polar state), or point nodes (such as an axial state),
where the gap vanishes on the Fermi surface. We consider $s$-wave and $d_{x^2-y^2}$-wave superconductors with ellipsoidally anisotropic masses and
Zeeman energies in this section.
For a conventional $s$-wave superconductor, the gap function $\Delta_s(\bm{R},\bm{\hat{k}}_{F})=\Delta_s(\bm{R})$ and the interaction
$V(\hat{\bm{k}},\hat{\bm{k}}^{'})=V_{0}$ is a constant. But for a $d_{x^2-y^2}$ superconductor, the interaction in Eq. (\ref{interaction}) leads to
\begin{eqnarray}
\Delta_d(\bm{R},\bm{\hat{k}}_{F})&=&\Delta_d(\bm{R})(\hat{\bm{k}}_{x}^{2}-\hat{\bm{k}}_{y}^{2}),
\end{eqnarray}
in which $\hat{\bm{k}}_F$ denotes the unit vector of the wave vector ${\bm k}_F$ on the Fermi surface, and the $\hat{k}_i$ are its components.
We then transform the unit wave vector components, $\hat{\bm{k}}_i(\theta_{k}, \phi_{k})$ into the $\hat{{\bm{k}}}^{\prime\prime}_i(\theta_{k^{\prime}},\phi_{k^{\prime}})=\alpha\hat{\tilde{k}}_i(\theta_{k^{\prime}},\phi_{k^{\prime}})$, obtaining
\begin{eqnarray}
\hat{k}_{x}&=&
\sqrt{\overline{m}_1}\alpha\beta(\cos\theta'\cos\phi'\hat{\tilde{k}}_x
-\sin\phi'\hat{\tilde{k}}_y+\sin\theta'\cos\phi'\hat{\tilde{k}}_z),\nonumber\\
\hat{k}_{y}&=&\sqrt{\overline{m}_2}\alpha\beta(\cos\theta'\sin\phi'\hat{\tilde{k}}_x
+\cos\phi'\hat{\tilde{k}}_y+\sin\theta'\sin\phi'\hat{\tilde{k}}_z), \nonumber \\
\hat{k}_{z}&=&\sqrt{\overline{m}_3}\alpha\beta( - \sin\theta'\hat{\tilde{k}}_x
+\cos\theta'\hat{\tilde{k}}_z),
\end{eqnarray}
and the $d_{x^2-y^2}$ pairing interaction transforms according to
\begin{eqnarray}
V(\hat{\bm k},\hat{\bm k}^{'})&=&V_0f({\bm k})f({\bm k}')\\
f({\bm k})&=&(\hat{k}_{x}^{2}-\hat{k}_{y}^{2})\rightarrow\alpha^2\beta^2f(\tilde{\bm k}),
\end{eqnarray}
where
\begin{eqnarray}
f(\tilde{\bm k})\!\!&=&\!\overline{m}_{1}\Bigl(\cos\theta'\cos\phi'\hat{\tilde{k}}_{x} - \sin\phi'\hat{\tilde{k}}_{y}
+\sin\theta'\cos\phi'\hat{\tilde{k}}_{z}\Bigr)^{2}\nonumber\\
&&\!\!-\overline{m}_2\Bigl(\cos\theta'\sin\phi'\hat{\tilde{k}}_x+\cos\phi'\hat{\tilde{k}}_y+\sin\theta'\sin\phi'\hat{\tilde{k}}_z\Bigr)^2\>\>\nonumber\\
\end{eqnarray}
It is then elementary to transform the gap function
\begin{eqnarray}
\Delta_d(\bm{R},\bm{\hat{k}}_{F})&=&\Delta_d(\bm{R})f({\bm k})\\
&\rightarrow&\Delta_d(\tilde{\bm{R}})\alpha^2\beta^{2}f(\tilde{\bm k}),
\end{eqnarray}
where $f(\tilde{\bm k})$ is given by Eq. (18).
\begin{figure*}
\begin{center}
\includegraphics[width=0.3\linewidth]{fig2a.eps}
\includegraphics[width=0.3\linewidth]{fig2b.eps}
\includegraphics[width=0.3\linewidth]{fig2c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) Plots of the reduced upper critical field $b_{c2}(t)$ defined in Eq. (\ref{bc2}) for an $s$-wave superconductor with an anisotropic FS characterized by different effective masses and a correspondingly anisotropic Zeeman energy. The isotropic $(\overline{m}_1=\overline{m}_2=\overline{m}_3=1)$ case is identical to the solid black curve. For an $s$-wave OP with an anisotropic effective mass characterized by $\overline{m}_1=0.5$, $\overline{m}_2=2.0$, and $\overline{m}_3=1.0$, $b_{c2}(t)$ curves for the field along the $\hat{x}$ (dashed red), $\hat{y}$ (dotted blue), and $\hat{\bm z}$ (dash dot magenta) directions, respectively, are also shown in (a). Plots of the reduced $b_{c2}$ for a highly anisotropic effective mass with strong spin-orbit interactions for an $s$-wave OP are shown in (b)and (c). In both figures, ${\bm b}_{c2}$ along the $x$ direction is in solid red, along the $y$ direction is in dashed blue (just above the horizontal axis), and along the $z$ direction is in dotted magenta shown in the insets. In (b), $\overline{m}_1=1.0\times10^{-4}$, $\overline{m}_2=1.0$, and $\overline{m}_3=1.0\times10^4$, in (c), $\overline{m}_1=\overline{m}_2=0.0050$, and $\overline{m}_3=4.0\times10^4$ }
\label{fig2}
\end{figure*}
For respective $s$-wave and $d_{x^{2}-y^{2}}$-wave order parameters, we have
\begin{eqnarray}
\Delta_s({\bm{R}})&=&2\pi TN(0)\sum_{\omega_{n}}\int\frac{d\Omega_{\bm{k}'}}{4\pi} V_{0} \int_{0}^{\infty}d\xi
e^{-2|\omega_{n}|\xi} \nonumber\\
& &\!\!\times \cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)e^{-i{\rm sgn}(\omega_{n})\xi{\bm v}_{F}({\bm{\hat{k}}}_{F}^{'})\cdot{\bm\Pi}({\bm R})}
\Delta_s(\tilde{\bm{R}}),\nonumber\\
\Delta_d({\bm{R}})&=&2\pi TN(0)\sum_{\omega_{n}}\int\frac{d\Omega_{\bm{k}'}}{4\pi} V_{0}\alpha^2\beta^2f^2(\tilde{\bm
k}')\int_{0}^{\infty} d\xi \nonumber\\
&&\!\!\times
e^{-2|\omega_{n}|\xi}\cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)e^{-i{\rm sgn}(\omega_{n})\xi \bm{v}_{F}({\bm{\hat{k}}}_{F}^{'})\cdot{\bm\Pi}({\bm
R})}\nonumber\\
& &\times \Delta_d({\bm{R}}),
\end{eqnarray}
where
\begin{eqnarray}
{\bm\Pi}({\bm R})&=&-i\alpha{\bm\nabla}_{{\bm R}}+2e{\bm A}({\bm R}).
\end{eqnarray}
As shown previously\cite{ScharnbergKlemm1980,Loerscher}, the gap functions $\Delta_i({{\bm R}})$ may be expanded
\begin{eqnarray}
\Delta_i({\bm{R}})&=&\sum_{n}a_{n,i}|n({\bm R})>,
\end{eqnarray}
for $i=s,d$
in terms of the complete orthonormal set of one-dimensional harmonic oscillator eigenfunctions $|n\rangle$, where
$<n'|n>=\delta_{n,n'}$,
we respectively obtain a closed expression for the $a_{n,s}$ and a recursion relation for the $a_{n,d}$,
\begin{eqnarray}
a_{n,s}&=&a_{n,s}2\pi TN(0)\sum_{\omega_{m}}\int\frac{d\Omega_{\bm{k}^{'}}}{4\pi} V_{0}\int_{0}^{\infty} d\xi
e^{-2|\omega_{m}|\xi}\nonumber\\
& &\times \cos(2\frac{2}{2 m_g}\alpha\tilde{B}\xi)e^{-\frac{1}{2}|\eta|^2}\sum_{n'=0}^{\infty}L_{n'}^{0}(|\eta|^2),\\
a_{n,d}&=&2\pi TN(0)\sum_{\omega_{m}}\int\frac{d\Omega_{\bm{k}^{'}}}{4\pi} V_{0}\beta^2f^2({\bm k}') \nonumber\\
&&\times\sum_{n^{'}=0}^{\infty}\int_{0}^{\infty}\frac{d\xi}{\alpha} e^{-2|\omega_{m}|\xi}
\cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)\nonumber\\
& &\times a_{n',d}<n({\bm R})| e^{-i{\rm sgn}(\omega_{m})\xi \bm{v}_{F}({\bm{\hat{k}}}_{F}^{'})\cdot{\bm\Pi}({\bm R})}|n'({\bm
R})>,\nonumber\label{and}\\
\end{eqnarray}
where
\begin{eqnarray}
\eta&=&-i{\rm sgn}(\omega_m)\xi\sin\theta_ke^{i\phi_k}\sqrt{e\alpha\tilde{B}},\label{eta}
\end{eqnarray}
the $L_n^0(x)$ are the Laguerre polynomials, and the matrix elements in Eq. (\ref{and}) are evaluated exactly as by L{\"o}rscher {\it et
al.}\cite{Loerscher}, although the integrations over $f^2(\tilde{\bm k}')$ are more complicated than in that $p$-wave case with broken symmetry.
From Eq. (\ref{and}), we derived a recursion relation for the $a_{n,d}$ corresponding to the $d_{x^2-y^2}$-wave order parameter,
\begin{eqnarray}
0&=&C_{n,n-4}a_{n-4,d}+C_{n,n-2}a_{n-2,d}+C_{n,n}a_{n,d}\nonumber\\
& &+C_{n,n+2}a_{n+2,d}+C_{n,n+4}a_{n+4,d},
\end{eqnarray}
where the coefficients $C_{n,n'}$ are given in the Appendix.
\section{Numerical Results}
\begin{figure*}
\begin{center}
\includegraphics[width=0.3\linewidth]{fig3a.eps}
\includegraphics[width=0.3\linewidth]{fig3b.eps}
\includegraphics[width=0.3\linewidth]{fig3c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) The reduced upper critical field $b_{c2}(t)$ for a $d_{x^2-y^2}$-wave superconductor with the field along $\hat{\bm x}$ (solid
red), $\hat{\bm y}$ (dash blue), and $\hat{\bm z}$ (dot magenta) directions, all for which $\overline{m}_{1}\overline{m}_{2}\overline{m}_{3}=1$.
(a) $\overline{m}_1=\overline{m}_2=\overline{m}_3=1$. (b) $\overline{m}_1=\overline{m}_2=1.26$, $\overline{m}_3=0.63$. (c) $\overline{m}_1=0.5$,
$\overline{m}_2=2.0$, $\overline{m}_3=1.0$.}
\label{fig:3}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.3\linewidth]{fig4a.eps}
\includegraphics[width=0.3\linewidth]{fig4b.eps}
\includegraphics[width=0.3\linewidth]{fig4c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) The reduced $b_{c2}(t)$ for $d_{x^2-y^2}$-wave superconductor without the Zeeman energy. Figures (a), (b), and (c)
correspond to the same sets of $\overline{m}_i$ values as in Fig. 3, and the field directions are also indicated exactly as in Fig. 3.}
\label{fig:4}
\end{figure*}
Here we present our numerical results for the reduced upper critical induction $b_{c2}$ and reduced temperature $t$ defined by
\begin{eqnarray}
b_{c2}(\theta,\phi,t)&=&2e\tilde{B}_{c2}(\theta,\phi,t)v_{F}^2/(2\pi T_{c})^2,\label{bc2}\nonumber\\
t&=&T/T_c
\end{eqnarray}
for $s$-wave and $d_{x^2-y^2}$-wave superconductors, and discuss the influences of the anisotropic effective mass, Zeeman energy, and the order parameter (OP) symmetry.
First, we present our results for $B_{c2}$ of an $s$-wave superconductor with an anisotropic effective mass and Zeeman tensor, but which has an isotropic
superconducting gap on the basis of Section \uppercase\expandafter{\romannumeral2}. In Fig. 2(a), $b_{c2}$ for an $s$-wave superconductor is plotted versus $t$ for a spherical FS and for a FS with ellipsoidal anisotropy. Clearly, $b_{c2}$ is independent of the direction of the $s$-wave OP on an isotropic FS, as shown by the solid black line in Fig. 2(a). However, that is not the case for an $s$-wave OP on an anisotropic FS, and one example is presented in Fig. 2(a) for $\overline{m}_{1}=0.5, \overline{m}_{2}=2.0, \overline{m}_{3}=1.0$. The reduced upper critical field $b_{c2}$ depends upon the field direction, and $b_{c2}$ is largest in the direction corresponding to the smallest effective mass. Actually, the difference arises from the first step of Klemm-Clem transformation, in which ${\bm B}^{\prime}(\theta^{\prime},\phi^{\prime})$ is $\alpha(\theta,\phi)$ times the magnitude of ${\bm B}(\theta,\phi)$, but the gap function is a constant in wave vector space. So it is clear that when the FS is isotropic, $\alpha=1$, and $b_{c2}$ for an isotropic $s$-wave superconductor is independent of the field direction.
We note that the $b_{c2}$ as shown in Fig. 2(a) is larger in the direction with of the smaller effective mass. We therefore investigated the effects of the much stronger effective mass anisotropy but also with an
isotropic $s$-wave pairing interaction. In Figs. 2(b) and 2(c), The $b_{c2}$ for the fields along the $x$, $y$, and $z$ directions are
depicted respectively in solid red, dashed blue, and dotted black curves. In Fig. 2(b), the reduced effective mass values are $\overline{m}_1=10^{-4}$,
$\overline{m}_2=1.0$, and $\overline{m}_3=10^4$, and $\overline{m}_1=\overline{m}_2=0.005$, $\overline{m}_3=4.0\times10^4$ in Fig. 2(c). For the field along the largest effective mass direction, $b_{c2}(t)$ is so small that we displayed the results in the figure insets.
Helfand and Werthamer\cite{Helfand1966} investigated the upper critical field for an $s$-wave superconductor on an isotropic FS in then clean limit, and found that its slope at $T_c$ was given by $B_{c2}=0.73|d H_{c2}/d T|_{T=T_c} T_c$. We also obtained the same value, for which the slope for the reduced upper critical field $b_{c2}$ at $t=1$ is $-1.426$ for an isotropic FS. We also found that the slope of $B_{c2}(t)$ at $T_c$ for an anisotropic FS is $B_{c2}(t)=\frac{b_{c2} \alpha}{12/7 \zeta(3)}|d B_{c2}/d T|_{T=T_c} T_c $. There is a chance that if this slope were to persist to low temperatures, $B_{c2}(0)$ could exceed the conventional Pauli limiting $B_p ({\rm T})=1.76 T_c ({\rm K})$, obtained without any effective mass anisotropy, provided that $\frac{b_{c2} \alpha}{12/7 \zeta(3)} $ is greater than $1.76$, where $\alpha$ is given by Eq. (\ref{alpha}).
\begin{figure*}
\begin{center}
\includegraphics[width=0.3\linewidth]{fig5a.eps}
\includegraphics[width=0.3\linewidth]{fig5b.eps}
\includegraphics[width=0.3\linewidth]{fig5c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) In-plane $b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor. (a) and (b) show $b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor in the $xy$-plane including the Zeeman energy for an isotropic FS at $t=0$, and $t=0.9$, respectively. (c) shows the results for $b_{c2}(\pi/2,\phi,0)$ for a $d_{x^2-y^2}$-wave superconductor with an isotropic FS but without the Zeeman energy at $t=0$.}
\label{fig:5}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.24\linewidth]{fig6a.eps}
\includegraphics[width=0.24\linewidth]{fig6b.eps}
\includegraphics[width=0.24\linewidth]{fig6c.eps}
\includegraphics[width=0.24\linewidth]{fig6d.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) Here we present some details of the region of rapid change in the symmetry of $b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave
order parameter with an isotropic FS without the Zeeman interaction. From left to right, $t=0.415, 0.419, 0.421$ and $0.423$.}
\label{fig:6}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.32\textwidth]{fig7a.eps}
\includegraphics[width=0.32\textwidth]{fig7b.eps}
\includegraphics[width=0.32\textwidth]{fig7c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) Azimuthal dependence at $t=0$ of $b_{c2}(\pi/2,\phi,0)$ for a $d_{x^2-y^2}$-wave superconductor with decreasing
$\overline{m}_1/\overline{m}_2$ at fixed $\overline{m}_1 \overline{m}_2 \overline{m}_3=1$ and $\overline{m}_2=1$. From left to right: $\overline{m}_1=0.9$, 0.8, and 0.7. In all three of these
figures, $b_{c2}(\pi/2,\phi,0)$ has a dumbbell shape exhibiting $C_2$ symmetry, but the directions of the maxima rotate from the $x$ axis for
$\overline{m}_1=0.7$ to increasingly off that axis as $\overline{m}_1$ increases.}
\label{fig:7}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.32\linewidth]{fig8a.eps}
\includegraphics[width=0.32\linewidth]{fig8b.eps}
\includegraphics[width=0.32\linewidth]{fig8c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) Shown are plots of the polar ($\theta$) angular dependence of $b_{c2}(\theta,\phi,t)/b_{c2}(0,0,0)$ normalized to its
$\theta=0^{\circ}$ value for a $d_{x^2-y^2}$-wave superconductor as $t\rightarrow0$, for $\overline{m}_1=\overline{m}_2$,
and $\overline{m}_1/\overline{m}_3 =0.6, 0.8, 1.0, 1.6$ and $1.8$ respectively. For each tetragonal effective mass anisotropy pictured, $b_{c2}(\theta,\phi,t)$ is a
monotonically increasing or decreasing function of $\theta$ at fixed $t$ and $\phi$.}
\label{fig:8}
\end{figure*}
However, the situation is considerably more complex for a $d_{x^{2}-y^{2}}$-wave OP than for an $s$-wave OP, especially for an anisotropic FS. Firstly, after the transformations, the interaction and the
$d_{x^{2}-y^{2}}$ OP both depend upon the directions in the pairing plane, as can be seen in Eqs. (17) and (20). In Fig. 3, we exhibit $b_{c2}(t)$ for a $d_{x^{2}-y^{2}}$-wave superconductor with the field along the $x$, $y$, and $z$ directions for three different cases, each with $\overline{m}_1\overline{m}_2\overline{m}_3=1$. In Fig. 3(a), the FS is isotropic, $b_{c2}(t)$ is identical for the field along the $x$- and $y$-axes, and is larger than for the field along the $z$-axis. The strong anisotropy in slope just below $t=1$ is solely due to the OP anisotropy, as the FS is isotropic. In fact, $b_{c2}$ for a $d_{x^2-y^2}$-wave superconductor with an anisotropic effective mass becomes more complicated and difficult to describe, as shown in Figs. 3 (b) and (c), where $b_{c2}(\theta,\phi)(t)$ crossing points exist. In Fig. 3(b), $b_{c2}(t)$ along the $x$ and $y$ directions are identical, due to the equal effective masses in the $xy$ plane, but cross $b_{c2}(t)$ along the $z$ direction. In Fig. 3(c), effective mass anisotropies on the order 2 can cause dramatically different $b_{c2}(t)$ curves for the three orthorhombic directions. We note that these curves include the relativistically consistent anisotropic Zeeman energy, and the slopes at $T_c$ are sufficiently large as to approach the conventional Pauli limit.
In Fig. 4, we show the $b_{c2}(t)$ curves for the $d_{x^2-y^2}$-wave superconductor with the same FS anisotropy parameters, but without the self-consistent treatment of the Zeeman interaction, as has often been incorrectly assumed in the literature. In these figures, $g=0$ for all field directions, which is inconsistent with the anisotropic Dirac equation used to generate the non-relativistic limit of the Hamiltonian \cite{Aiying2021}. In these figures, there are no crossing points of $b_{c2}(t)$ for the different field directions.
Further study shows that the such crossing points seem to disappear comparing with the results in Figs. 4 (b) (c) without the Zeeman energy, but to strengthen this argument requires further study. We only show a few examples in Figs. 4(b) and (c), for which the crossing points vanish. The crossing point will exist for a pancake FS ($\bar{m}_1=\bar{m}_{2}\textgreater\bar{m}_3$) with the ratio $m_{1}/m_{3}$ between $1.6$ and $3.2$. Therefore, the Zeeman energy, which is intrinsically anisotropic, plays a crucial role in determining $b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor. At least, it suppresses the $B_{c2}(t)$ for $d_{x^2-y2}$-wave superconductor more strongly than for an $s$-wave superconductor. But there remains the possibility that the conventional Pauli limit could be exceeded in some field directions for a $d_{x^2-y^2}$-wave superconductor with a highly anisotropic FS.
It is interesting to investigate $b_{c2}(\pi/2,\phi,t)$ in the $xy$ plane. Not only have many researchers recently paid more attention to measuring
$b_{c2}$ in that plane\cite{Zuo2017,Vorontsov2010}, many superconductors, such as the cuprates, have been thought by many authors to exhibit $d_{x^2-y^2}$ OP
symmetry\cite{Shen2003,Tsuei2000}. Here we present results for $b_{c2}(\pi/2,\phi,t)$ in the $xy$ plane, with and without the Zeeman interaction. It is intriguing to find that
$b_{c2}(\pi/2,\phi,t)$ including the Zeeman energy at different $t$ values does not always follow the commonly held belief that the maxima $b_{c2}(\pi/2,\phi,t)$
always occurs along the antinodal directions of the of the OP. Instead, there is a $\pi/4$ azimuthal shift in these maxima if this slope were to persist to low $t$. That is, the maxima in
$b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor just below $t=1$ is indeed along the antinodal directions, but at low $t$, the maxima are along the nodal directions.
This change in symmetry is shown
in Figs. 5(a) and (b), which are the results for an isotropic FS at $t = 0$ and $t = 0.9$, respectively. Fig. 5(c) presents the results for $b_{c2}(\pi/2,\phi,t)$
without the Zeeman energy in the $xy$ plane at low $t$. The shape of $b_{c2}(\pi/2,\phi,t)$ in Fig. 5(c) without the Zeeman energy actually is similar to
that of Fig. 5 (b), but the four-fold anisotropy is greatly reduced in magnitude by the Zeeman energy in Fig. 5(b). This is easy to understand from Eq.
(21). The Zeeman energy suppresses the upper critical field for a $d_{x^2-y^2}$-wave superconductor and enters into the gap equation in the $\cos[(\sigma-\sigma^\prime)\frac{e}{2 m_g} \tilde{B}\xi]$ term, which can be expanded in powers of $\tilde{\bm B}$, yielding the factor $ 1-[(\sigma-\sigma^\prime)\frac{e}{2 m_g}\alpha\tilde{B}\xi]^2/2$, which is nearly unity near to $t=1$. However, at large $B_{c2}$ at low $t$, the $\pi/4$ phase shift demonstrates that the Zeeman energy is very important at low $t$.
Because $b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor with an isotropic FS with or without the Zeeman interaction is greatly different at $t=0$ and $t=0.9$, we explored the process of the $\pi/4$
phase change in its $\phi$-dependence, and found that it changes rapidly in the mid-$t$ range, as shown in Fig. 6 without the Zeeman interaction to emphasize the changes in the azimuthal dependence of $b{c2}(\pi/2,\phi,t)$ with decreasing $t$. As $t$ is lowered from 0.9,
$b_{c2}(\pi/2,\phi,t)$ initially increases very slightly along the antinodal directions down to 0.421, and then the intensity along the nodal directions
starts to grow quickly, finally taking over around $t=0.415$.
We also considered the effects of different effective masses on $b_{c2}(\pi/2,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor at low $t$. When the
effective mass anisotropy within the pairing $xy$ plane becomes strong, as in detwinned orthorhombic YBa$_2$Cu$_3$O$_{7-\delta}$, the in-plane azimuthal
angular dependence of $b_{c2}$ not only deviates strongly from the $C_4$ symmetry that the nominal $d_{x^2-y^2}$ Cooper pair obeys if the crystal were to
be tetragonal, its shape changes from the four-leaf clover of $C_4$ symmetry to that of a dumbbell with $C_2$ symmetry. The precise form of the in-plane azimuthal anisotropy of $b_{c2}(\pi/2,\phi,0)$ for an orthorhombic $d_{x^2-y^2}$-wave
superconductor changing with the $\overline{m}_1/\overline{m}_2$ effective mass anisotropy is pictured in Fig. 7. In this figure, we show the results of our calculations for a fixed geometric mean
$\overline{m}_1\overline{m}_2\overline{m}_3=1$ and $\overline{m}_2=1$, with $\overline{m}_1=0.9$, 0.8, and 0.7 in Figs. 7(a), (b), and (c), respectively. For the least in-plane anisotropy pictured in Fig. 7(a), the predicted $\pi/4$ rotation of the maxima in the four-fold azimuthal anisotropy shown in Figs. 5(a) and 6(a) is partly present for two of the maxima, but the figure shown in Fig. 7(a) has overall $C_2$ symmetry. At $\overline{m}_1/\overline{m}_2=0.7$, the azimuthal anisotropy of $b_{c2}(\pi/2,\phi,0)$ has rotated to the $x$ axis, and has the simpler dumbbell shape.
A peak in the polar dependence of $b_{c2}$ was predicted for $p$-wave superconductors with completely broken symmetry for anisotropies corresponding to
$\overline{m}_3/\alpha^2(\pi/2,\phi)>3$\cite{Loerscher}, we tried to study whether a peak in $b_{c2}(\theta, \phi,t)$ at some fixed $\phi$ and $t$ might occur for some effective mass anisotropy values. As shown in Fig. 8(a) and (c), for $m_{1}/m_{3}=1.6$ and 1.8, $b_{c2}(\theta,\phi,t)/b_{c2}(0,0,0)$ decreases monotonically with increasing $\theta$, but increases in Fig. 8(b), and for $\overline{m}_1/\overline{m}_3=0.6$, 0,8, and 1.0, the isotropic FS case, $b_{c2}(\theta,\phi,t)/b_{c2}(0,0,0)$ is a monotonically increasing function of $\theta$. For all of the cases plotted in Fig. 8, there is no peak in $b_{c2}(\theta,\phi,t)$ for $0 < \theta < \pi/2$, but it doesn't imply that $b_{c2}$ is a monotonic function of $\theta$ strictly. Recalling that there exits a crossing point in Fig. 3(b), we note that it doesn't stay monotonic near to the crossing point, but the change is slow and a little flat with increasing $\theta$.
\section{Conclusions}
In this paper we studied the temperature and angular dependence of the upper critical induction $B_{c2}(\theta,\phi,T)$ of $s$-wave and $d_{x^2-y^2}$-wave superconductors based
on the Gor'kov equation, which includes the anisotropic Zeeman energy arising from the ellipsoidal anisotropy of FS (with effective masses $m_i$ along the
three Cartesian coordinates for an orthorhomic crystal structure) that is treated self-consistently using the anisotropic Schr{\"o}diner-Pauli single particle Hamiltonian. While we
have derived the upper critical field for $s$-wave and $d$-wave superconductors with anisotropic effective masses, this work could easily be applied to
any anisotropic pairing function. It is analogous to the treatment of $p$-wave superconductors with completely broken symmetry and ellipsoidal effective
mass symmetry, except for the addition of the Pauli pair-breaking effects in singlet-spin superconductors.
For an $s$-wave superconductor, we find the reduced upper critical induction $b_{c2}(\theta,\phi,t)$ is angularly modulated by the universal orientation factor
$\alpha(\theta,\phi)$\cite{Klemm1980}, and the Pauli limiting can be exceeded by adjusting the effective mass, and the Zeeman energy almost can be neglected when comparing it with that for a $d_{x^2-y^2}$ OP. But it is not the case for $d_{x^{2}-y^{2}}$-wave superconductors, whose interaction and gap function depend upon the unit wave vector components in the $xy$ plane, which become more complicated after the transformation of the single particle Hamiltonian to isotropic form. $b_{c2}(\theta,\phi,t)$ for a $d_{x^2-y^2}$-wave superconductor not only changes along the axial $\phi$ direction with reduced temperature $t=T/T_c$, it shows an interesting $\phi$ dependence in the $xy$ plane. There is a $\pi/4$ shift in the $C_4$-symmetric calculated $b_{c2}(\pi/2,\phi,t)$ patterns between low and high $t$ values when the Zeeman interaction is included. The effective mass anisotropy in the plane of the pairing interaction is significant, as for untwinned orthorhombic YBa$_2$Cu$_3$O$_{7-\delta}$, the $\phi$ dependence of $b_{c2}(\pi/2,\phi,t)$ in $xy$ plane at low $t$ changes from exhibiting $C_4$ symmetry to $C_2$ symmetry. In addition, the variation of $b_{c2}(\theta,0,t)$ with $\theta$ at low $t$ is a monotonic function of $\theta$.
\begin{acknowledgments}
This work was supported by the National Natural Science Foundation of China through Grant no. 11874083. A. Z. acknowledges financial support from the China Scholarship Council. R. A. K. was partially supported by the U. S. Air Force Office of Scientific Research (AFOSR) LRIR \#18RQCOR100, and the AFRL/SFFP Summer Faculty Fellowship Program provided by AFRL/RQ at WPAFB.
\end{acknowledgments}
\section*{Appendix}
In the following, we present the forms of the recursion coefficients $C_{n,n'}$ for a $d_{x^2-y^2}$-wave superconductor. In these expressions,
$\eta$ is given by Eq. (\ref{eta}) in the text,
and the $L_n^k(x)$ are the associated Laguerre polynomials.
\begin{eqnarray}
C_{n,n}&=&\pi T\sum_{\omega_{m}}\int_0^{\pi}\sin\theta_{k}d\theta_k\int_0^{\infty}d\xi e^{-2|\omega_{m}|\xi}e^{-\frac{1}{2}|\eta|^2}\nonumber\\
& &\times\cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)
L_{n}^{0}(|\eta|^{2})\Bigl[\sin^{4}\theta_k\nonumber\\
& &\times(\frac{3A^2+3B^2}{8}+\frac{D^2+2AB}{8})+\sin^2\theta_k\cos^2\theta_k \nonumber\\
&&\frac{E^2+2AC+F^2+2BC}{2}+\cos^4\theta_k C^2\Bigr]- \frac{1}{N_{0}V_{0}},\nonumber\\
\end{eqnarray}
\begin{eqnarray}
C_{n,n+2}&=&\pi T\sum_{\omega_{m}}\int_0^{\pi}\sin\theta_{k}d\theta_k\int_0^{\infty}d\xi e^{-2|\omega_{m}|\xi}e^{-\frac{1}{2}|\eta|^2}\nonumber\\
& &\times\cos(2 \frac{e}{2 m_g}\alpha\tilde{B}\xi)
\frac{-|\eta|^2}{\sqrt{(n+2)(n+1)}}L_{n}^{2}(|\eta|^2)\Bigl[\sin^4\theta_k\nonumber\\
& &\times\Bigl(\frac{A^2-B^2}{4}-\frac{2AD+2BD}{8i}\Bigr) \nonumber\\
&&+\frac{1}{4}\sin^{2}2\theta_k\nonumber\\
& &\times\Bigl(\frac{E^2+2AC-F^2-2BC}{4}-\frac{2EF+2CD}{4i}\Bigr)\Bigr],\nonumber\\
\end{eqnarray}
\begin{eqnarray}
C_{n,n-2}&=&\pi T\sum_{\omega_{m}}\int_0^{\pi}\sin\theta_{k}d\theta_k\int_0^{\infty}d\xi e^{-2|\omega_{m}|\xi}e^{-\frac{1}{2}|\eta|^2}\nonumber\\
& &\times\cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)
\frac{-|\eta|^2}{\sqrt{n(n-1)}}L_{n-2}^{2}(|\eta|^2)\Bigl[\sin^4\theta_k\nonumber\\
& &\times\Bigl(\frac{A^2-B^2}{4}+ \frac{2AD+2BD}{8i}\Bigr) \nonumber\\
&&+\frac{1}{4}\sin^{2}2\theta_k\Bigl(\frac{E^2+2AC-F^2-2BC}{4}\nonumber\\
& &+\frac{2EF+2CD}{4i}\Bigr)\Bigr],
\end{eqnarray}
\begin{eqnarray}
C_{n,n+4}&=&\pi T\sum_{\omega_{m}}\int_0^{\pi}\sin\theta_{k}d\theta_k\int_0^{\infty}d\xi e^{-2|\omega_{m}|\xi}e^{-\frac{1}{2}|\eta|^2}\nonumber\\
& &\times\cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)
\frac{(-|\eta|^2)^2}{\sqrt{(n+4)(n+3)(n+2)(n+1)}}\nonumber\\
& &\times L_{n}^{4}(|\eta|^2)\Bigl[\sin^4\theta_k\Bigl(\frac{A^2+B^2-D^2-2AB}{16}\nonumber\\
& &-\frac{AD-BD}{8i}\Bigr)\Bigr],
\end{eqnarray}
\begin{eqnarray}
C_{n,n-4}&=&\pi T\sum_{\omega_{m}}\int_0^{\pi}\sin\theta_{k}d\theta_k\int_0^{\infty}d\xi e^{-2|\omega_{m}|\xi}e^{-\frac{1}{2}|\eta|^2}\nonumber\\
& &\times\cos(2\frac{e}{2 m_g}\alpha\tilde{B}\xi)
\frac{(-|\eta|^2)^2}{\sqrt{n(n-1)(n-2)(n-3)}}\nonumber\\
& &\times L_{n-4}^{4}(|\eta|^2)\Bigl[\sin^4\theta_k \nonumber\\
&&\times\Bigl(\frac{A^2+B^2-D^2-2AB}{16}+\frac{AD-BD}{8i}\Bigr)\Bigr],\nonumber\\
\end{eqnarray}
where
\begin{eqnarray}
A^2&=&\beta^4(\overline{m}_{1}\cos^2\phi'-\overline{m}_{2}\sin^2\phi')^2\cos^4\theta',\\
B^2&=&\beta^4(\overline{m}_{1}\sin^2\phi'-\overline{m}_{2}\cos^2\phi')^2\\
C^2&=&\beta^4(\overline{m}_{1}\cos^2\phi'-\overline{m}_{2}\sin^2\phi')^2\sin^4\theta',\\
D^2+2AB&=&\beta^4(\overline{m_{1}}+\overline{m}_{2})^2\cos^2\theta'\sin^{2}2\phi'\nonumber\\
& &+2\beta^4\cos^2\theta'(\overline{m}_{1}\cos^2\phi'-\overline{m}_{2}\sin^2\phi')\nonumber\\
& &\times(\overline{m}_{1}\sin^2\phi'-\overline{m}_{2}\cos^2\phi'),\\
E^{2}+2AC&=&\frac{3}{2}\beta^4\sin^{2}2\theta'(\overline{m}_{1}\cos^2\phi'\nonumber\\
& &-\overline{m}_{2}\sin^{2}\phi')^2,
\end{eqnarray}
\begin{eqnarray}
F^2+2BC&=&\beta^4\sin^2\theta'\sin^{2}2\phi'(\overline{m}_1+\overline{m}_2)^2\nonumber\\
& &+2\beta^4\sin^2\theta'\nonumber\\
& &\times\Bigl(\overline{m}_{1}\sin^2\phi'-\overline{m}_{2}\cos^2\phi'\Bigr)\nonumber\\
& &\times\Bigl(\overline{m}_{1}\cos^2\phi'-\overline{m}_{2}\sin^2\phi'\Bigr),\\
2AD&=&-2\beta^4\cos^3\theta'\sin 2\phi'(\overline{m}_{1}+\overline{m}_{2})\nonumber\\
& &\times\Bigl(\overline{m}_{1}\cos^2\phi'-\overline{m}_{2}\sin^2\phi'\Bigr),
\end{eqnarray}
\begin{eqnarray}
2BD&=&-2\beta^4\cos\theta'\sin 2\phi'(\overline{m}_{1}+\overline{m}_{2})\nonumber\\
& &\times\Bigl(\overline{m}_{1}\sin^2\phi'-\overline{m}_{2}\cos^2\phi'\Bigr),
\end{eqnarray}
\begin{eqnarray}
2EF+2CD&=&-3\beta^4\sin\theta'\sin 2\theta'\sin 2\phi'\nonumber\\
& &\times(\overline{m}_{1}+\overline{m}_{2})\nonumber\\
& &\times\Bigl(\overline{m}_{1}\cos^2\phi'-\overline{m}_{2}\sin^2\phi'\Bigr).
\end{eqnarray}
\begin {thebibliography}{99}
\bibitem{Sigrist1991} Sigrist M and Ueda K, {\it Phenomenological theory of unconventional superconductivity}, Rev. Mod. Phys. {\bf63}, 239 (1991).
\bibitem{Cao2018} Cao Y, Fatemi V, Fang S, Watanabe K, Taniguchi T, Kaxiras E and Jarillo-Herrero P, {\it Unconventional superconductivity in magic-angle graphene superlattices}, Nature {\bf 556}, 43 (2018).
\bibitem{Cao2020} Cao Y, Rodan-Legrain D, Rubies-Bigorda O, Park J M, Watanabe K, Taniguchi T and Jarillo-Herrero P, {\it Tunable correlated states and spin-polarized phases in twisted bilayer–bilayer graphene}, Nature,{\bf 583}, 215 (2020).
\bibitem{Drozdov} Drozdov A P, Eremets M I, Troyen I A, Ksenfontov V and Shylin S I, {\it Conventional superconductivity at 203 kelvin at high pressures in the sulfur hydride system}, Nature {\bf 525}, 73 (2015).
\bibitem{Pickard} Pickard C J, Errea I and Eremets M I, {\it Superconducting hydrides under pressure} Ann. Rev. Condens. Mat. Phys. {\bf11}, 57-76 (2020).
\bibitem{Eremets} Kong P P, Minkov V S, Kuzonikov M A, Drozdov A P, Besedin S P, Mozaffari S, Balicas L, Balakirev F F, Prapapenka V B, Chariton S, Knyazev D A, Greenberg E and Erements M I, {\it Superconductivity up to 243} K {\it in the yttrium-hydrogen system}, Nat. Commun. {\bf 12}, 5075 (2021).
\bibitem{Kogan2012} Kogan V G and Prozorov R, {\it Orbital upper critical field and its anisotropy of clean one- and two-band superconductors}, Rep. Prog. Phys. {\bf 75}, 114502 (2012).
\bibitem{Gurevich2003} Gurevich A, {\it Enhancement of the upper critical field by nonmagnetic impurities in dirty two-gap superconductors}, Phys. Rev. B {\bf67}, 184515 (2003).
\bibitem{Gurevich2010} Gurevich A, {\it Upper critical field and the Fulde-Ferrel-Larkin-Ovchinnikov transition in multiband superconductors}, Phys. Rev. B {\bf82}, 184504 (2010).
\bibitem{Norman2010} Norman M R, {\it Fermi-surface reconstruction and the origin of high-temperature superconductivity}, Physics {\bf 3}, 86 (2010).
\bibitem{Klemm2000} Klemm R A, {\it Striking similarities between the pseudogap phenomena in cuprates and in layered organic and dichalcogenide superconductors}, Physica C {\bf 341}, 839 (2000).
\bibitem{Klemm2015} Klemm R A, {\it Pristine and intercalated transition metal dichalcogenide superconductors}, Physica C {\bf 514}, 86 (2015).
\bibitem{Layered} Klemm R A, {\it Layered Superconductors Volume 1} (Oxford University Press, Oxford UK, 2012).
\bibitem{Klemm1975} Klemm R A, Luther A and Beasley M R , {\it Theory of the upper critical field in layered superconductors}, Phys. Rev. B {\bf12}, 877 (1975).
\bibitem{Wang2016} Wang J R, Liu G Z and Zhang C J, {\it Connection between in-plane upper critical field} H$_{c2}$ {\it and gap symmetry in layered} $d${\it-wave superconductors}, Phys. Rev. B {\bf94}, 014501 (2016).
\bibitem{Aiying2021}
Zhao A, Gu Q, Haugan T J and Klemm R A, {\it The Zeeman, spin-orbit, and quantum spin Hall interactions in anisotropic and low-dimensional conductors}, J. Phys.: Condens. Matter {\bf 33}, 085802 (2021).
\bibitem{Mineev} Mineev V P and Samokhin K V, {\it Introduction to Unconventional Superconductivity} (Gordon and Breach, New York, 1999).
\bibitem{Aoki} Aoki D and Flouquet J, {\it Ferromagnetism and superconductivity in uranium compounds}, J. Phys. Soc. Jpn. {\bf 81}, 011003 (2012).
\bibitem{ScharnbergKlemm1980} Scharnberg K and Klemm R A, {\it $p$-Wave superconductors in magnetic fields}, Phys. Rev. B {\bf 22}, 5233 (1980).
\bibitem{ScharnbergKlemm1985} Scharnberg K and Klemm R A, {\it Upper critical field in $p$-wave superconductors with broken symmetry}, Phys. Rev. Lett. {\bf 54}, 2445 (1985).
\bibitem{Loerscher} L{\"o}rscher C, Zhang J, Gu Q and Klemm R A, {\it Anomalous angular dependence of the upper critical induction of orthorhombic ferromagnetic superconductors with completely broken $p$-wave symmetry}, Phys. Rev. B {\bf 88}, 024504 (2013).
\bibitem{ZhangSr2RuO4} Zhang J, L{\"o}rscher C, Gu Q and Klemm R A, {\it Is the anisotropy of the upper critical field of} SrRuO$_4$ {\it consistent with a helical $p$-wave state?}, J. Phys.: Condens. Matter {\bf 26}, 252201 (2014).
\bibitem{Zhangfirstorder} Zhang J, L{\"o}rscher C, Gu Q and Klemm R A, {\it First-order chiral to non-chiral transition in the upper critical induction of the Scharnberg-Klemm $p$-wave pair state}, J. Phys.: Condens. Matter {\bf 26}, 252202 (2014).
\bibitem{Helfand1966} Helfand E and Werthamer N R, {\it Temperature and purity dependence of the superconducting critical field,} H$_{c2}$. $ II$, Phys. Rev. {\bf 147}, 288 (1966).
\bibitem{Clogston1962} Clogston A M, {\it Upper limit for the critical field in hard superconductors}, Phys. Rev. Lett. {\bf9}, 266 (1962).
\bibitem{Chandrasekhar1962} Chandrasekhar B S, {\it A note on the maximum critical field of high-field superconductors}, Appl. Phys. Lett. {\bf 1},7 (1962).
\bibitem{Maki1966} Maki K, {\it Effect of Pauli paramagnetism on magnetic properties of high-field superconductors}, Phys. Rev. {\bf 148}, 362(1966).
\bibitem{Lu2015} Lu J M, Zheliuk O, Leermakers I, Yuan N F, Zeitler U, Law K T and Ye J T, {\it Evidence for two-dimensional Ising superconductivity in gated} MoS$_{2}$, Science {\bf350}, 1353 (2015).
\bibitem{Werthamer1966} Werthamer N R, Helfand E and Hohenberg P C, {\it Temperature and purity dependence of the superconducting critical field} H$_{c2}$. $III$.{\it Electron spin and spin-orbit effects}, Phys. Rev. {\bf147}, 295 (1966).
\bibitem{Fulde1964} Fulde P and Ferrell R A, {\it Superconductivity in a strong spin-exchange field}, Phys. Rev. {\bf135}, A550 (1964).
\bibitem{Agosta2017} Agosta C C, Fortune N A, Hannahs S T, Gu S, Liang L, Park J-H and Schlueter J A, {\it Calorimetric measurements of magnetic-field-induced inhomogeneous superconductivity above the paramagnetic limit}, Phys. Rev. Lett. {\bf 118}, 267001 (2017).
\bibitem{Matsuda2007} Matsuda Y and Shimahara H, {\it Fulde-Ferrell-Larkin-Ovchinnikov state in heavy fermion superconductors}, J. Phys. Soc. Jpn. {\bf76}, 051005 (2007).
\bibitem{Gruenberg1966} Gruenberg L W and Gunther L, {\it Fulde-Ferrell effect in type-$II$ superconductors}, Phys. Rev. Lett. {\bf16}, 996 (1966).
\bibitem{Larkin1964} Larkin A I and Ovchinnikov Y N, {\it Nonuniform state of superconductors}, Zh. Eksp. Teor. Fiz. {\bf47}, 1136 (1964) [Sov. Phys. JETP {\bf20}, 762 (1965)].
\bibitem{Yip2014} Yip S, {\it Noncentrosymmetric superconductors}, Annu. Rev. Condens. Matter. Phys. {\bf5}, 15 (2015).
\bibitem{Kneidinger2015} Kneidinger F, Bauer E, Zeiringer I, Rogl P, Blass-Schenner C, Reith D and Podloucky P, {\it Superconductivity in non-centrosymmetric materials}, Physica C {\bf514}, 388(2015).
\bibitem{Goh2012} Goh S K, Mizukami Y, Shishido H, Watanabe D, Yasumoto S, Shimozawa M, Yamashita M, Terashima T, Yanase Y, Shibauchi T, Buzdin A I and Matsuda Y, {\it Anomalous upper critical field in} CeCoIn$_{5}$/YbCoIn$_{5}$ {\it superlattices with a Rashba-type
heavy fermion interface}, Phys. Rev. Lett. {\bf109}, 157006 (2012).
\bibitem{Shen2003} Damascelli A, Hussain Z and Shen Z X, {\it Angle-resolved photoemission studies of the cuprate superconductors}, Rev. Mod. Phys. {\bf75}, 473 (2003).
\bibitem{Tsuei2000} Tsuei C C and Kirtley J R, {\it Pairing symmetry in cuprate superconductors}, Rev. Mod. Phys. {\bf72}, 969 (2000).
\bibitem{MoessleKleiner} M{\"o\ss}le M and Kleiner R, {\it $c$-axis Josephson tunneling between} Bi$_{2}$Sr$_{2}$CaCu$_{2}$O$_{8+x}$ {\it and Pb}, Phys. Rev. B {\bf 59}, 4486 (1999).
\bibitem{Li1999} Li Q, Tsay Y N, Suenaga M, Klemm R A, Gu G D and Koshizuka N, Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ {\it Bicrystal $c$-axis twist Josephson junctions: A new phase-sensitive test of order parameter symmetry}, Phys. Rev. Lett. {\bf 83}, 4160 (1999).
\bibitem{Takano2002} Takano Y, Hatano T, Fukuyo A, Ishii A, Ohmori M, Arisawa S, Togano K and Tachiki M, {\it $d$-Like symmetry of the order parameter and intrinsic Josephson effects in} Bi$_{2}$Sr$_{2}$CaCu$_{2}$O$_{8+\delta}$ {\it cross-whisker junctions}, Phys. Rev. B {\bf 65}, 140513 (2002).
\bibitem{Takano2003} Takano Y {\it et al.}, {\it Cross-whisker intrinsic Josephson junction as a probe of symmetry of the superconducting order parameter}, J. Low Temp. Phys. {\bf 131}, 533 (2003).
\bibitem{Latyshev} Latyshev Y I, Orlov A P, Nikitina A M, Monceau P and Klemm R A, {\it $c$-Axis transport in naturally grown} Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ {\it cross-whisker junctions} Phys. Rev. B {\bf 70}, 094517 (2004).
\bibitem{Klemm2005} Klemm R A, {\it The phase-sensitive $c$-axis twist experiments on} Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ {\it and their implications}, Phil. Mag. {\bf 85}, 801-853 (2005).
\bibitem{Tsinghua2} Zhu Y, Liao M, Zhang Q, Xie H-Y, Meng F, Liu Y, Bai Z, Ji S, Zhang J, Jiang K, Zhong R, Schneeloch J, Gu G, Gu L, Ma X, Zhang D and Xue Q-K, {\it Presence of $s$-wave pairing in Josephson junctions made of twisted ultrathin} Bi$_2$Sr$_2$CaCu$_2$O$_{8+x}$ {\it flakes}, Phys. Rev. X {\bf11}, 031011 (2021).
\bibitem{Korea} Lee J, Lee W, Kim G-Y, Choi Y-B, Park J, Jang S, Gu G, Shoi S-Y, Cho G Y, Lee G-H and Lee H-J, {\it Twisted van der Waals Josephson junction based on a high} $T_c$ {\it superconductor}, Nano Lett. {\bf24} (21) 10469-10477 (2021).
\bibitem{Harvard} Zhao S Y F, Poccia N, Cui X, Volkov P A, Yoo H, Engelke R, Ronan Y, Zhong R, Gu G, Plugge S, Tummuru T, Franz M, Pixley J H and Kim P, {\it Emergent interfacial superconductivity between twisted cuprate superconductors} (unpublished) ArXiv:2108.13455v1.
\bibitem{Tonjes} Tonjes W C, Greanya W A, Liu R, Olson C G and Molini{\'e} P, {\it Charge-density wave mechanism in the} $2H$-NbSe$_2$ {\it family. Angle-resolved photoemission study}, Phys. Rev. B {\bf63}, 2351011 (2001).
\bibitem{Gamble} Gamble F R, DiSalvo F J, Klemm R A and Geballe T H, {\it Superconductivity in layered structure organometallic crystals}, Science {\bf 168}, 568 (1970).
\bibitem{Tsinghua} Zhong Y, Wang Y, Han S, Lv Y-F, Wang W-L, Zhang D, Ding H, Zhang Y-M, Wang L, He K, Zhong R, Schneeloch J, G G-D, Song C-L, Ma X-C and Xue Q-K, {\it Nodeless pairing in superconducting copper-oxide monolayer films on} Bi$_{2}$Sr$_2$CaCu$_{2}$O$_{8+\delta}$ , Sci. Bull. {\bf 61}, 1239 (2016).
\bibitem{Klemmprivate} The corresponding author communicated this privately and publically, respectively, to the corresponding authors of\cite{Korea,Harvard}, after those experiments were respectively published and publicized.
\bibitem{Kashiwagi}
Kashiwagi T, Sakamoto K, Kubo H, Shibano Y, Enomoto T, Kitamura T, Asamuma K, Yasui T, Watanabe C, Nakade K, Saiwai Y, Katsuragawa T, Tsuijimoto M, Yshizaki R, Yamamoto T, Minami H, Klemm R A and Kadowaki K, {\it A high $T_c$ intrinsic Josephson junction emitter tunable from 0.5 to 2.4 THz}, Appl. Phys. Lett. {\bf 107}, 082601 (2015).
\bibitem{Kleinerprivate}Reinhold Kleiner (private communication).
\bibitem{Naughton} Naughton M J, Yu R C, Davies P K, Fischer J E, Chamberlin R V, Wang Z Z, Jing T W, Ong N P and Chaikin P M, {\it Orientational anisotropy of the upper critical field in single-crystal} YBa$_{2}$Cu$_{3}$O$_{7}$ {\it and} Bi$_{2.2}$CaSr$_{1.9}$Cu$_{2}$O$_{8+x}$, Phys. Rev. B {\bf 38}, 9280 (1988).
\bibitem{Welp1} Welp U, Kwok W K, Crabtree G W, Vandervoort K G, and Liu J Z, {\it Magnetic measurements of the upper critical field of} YBa$_2$Cu$_3$O$_{7-\delta}$ {\it single crystals}, Phys. Rev. Lett. {\bf 62}, 1908 (1989).
\bibitem{Welp2} Welp U, Grimsditch M, You H, Kwok W K, Fang M M, Crabtree G W, and Liu J Z, {\it The upper critical field of untwinned} {\it crystals}, Physica C {\bf 161}. 1 (1989).
\bibitem{Miura}
Miura N, Nakagawa H, Sekitani T, Naito M, Soto H, and Enomoto Y, {\it High-magnetic-field study of high-} T$_{c}$ {\it cuprates}, Physica B {\bf 319}, 310 (2002).
\bibitem{CeCu2Si2}
Vierya H A, Oeschler N, Seiro S, Jeevan H S, Geibel C, Parker D, and Steglich F, {\it Determination of Gap Symmetry from Angle-Dependent} H$_{c2}$ {\it Measurements on} CeCu$_{2}$Si$_{2}$, Phys. Rev. Lett. {\bf 106}, 207001 (2011).
\bibitem{Klemm1980} Klemm R A and Clem J R, {\it Lower critical field of an anisotropic type-II superconductor}, Phys. Rev. B {\bf21}, 1868 (1980).
\bibitem{AGD} Abrikosov A A, Gor'kov L N, and Dzyaloshinskii I E, {\it Methods of Quantum Field Theory in Statistical Physics} (Dover Books on
Physics, New York, 1975).
\bibitem{Zuo2017} Zuo H, Bao J K, Wang J, Jin Z, Xia Z, Li L, Xu Z, Kang J, Zhu Z, and Cao G H, {\it Temperature and angular dependence of the upper critical field in} K$_{2}$Cr$_{3}$As$_{3}$ , Phys. Rev. B {\bf95}, 014502 (2017).
\bibitem{Vorontsov2010} Vorontsov A B and Vekhter I, {\it Vortex state in d-wave superconductors with strong paramagnetism: Transport and specific heat anisotropy}, Phys. Rev.B {\bf81}, 094527 (2010).
\end{thebibliography}
\end{document}
\begin{figure*}
\begin{center}
\includegraphics[width=0.32\textwidth]{fig7a.eps}
\includegraphics[width=0.32\textwidth]{fig7b.eps}
\includegraphics[width=0.32\textwidth]{fig7c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) Azimuthal dependence at $t=0$ of $b_{c2}(\pi/2,\phi,0)$ for a $d_{x^2-y^2}$-wave superconductor with decreasing
$\overline{m}_1/\overline{m}_2$ at fixed $\overline{m}_1 \overline{m}_2 \overline{m}_3=1$ and $\overline{m}_2=1$. From left to right: $\overline{m}_1=0.9$, 0.8, and 0.7. In all three of these
figures, $b_{c2}(\pi/2,\phi,0)$ has a dumbbell shape exhibiting $C_2$ symmetry, but the directions of the maxima rotate from the $x$ axis for
$\overline{m}_1=0.7$ to increasingly off that axis as $\overline{m}_1$ increases.}
\label{fig:7}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=0.32\linewidth]{fig8a.eps}
\includegraphics[width=0.32\linewidth]{fig8b.eps}
\includegraphics[width=0.32\linewidth]{fig8c.eps}
\vspace{-5mm}
\end{center}
\caption{(Color online) Shown are plots of the polar ($\theta$) angular dependence of $b_{c2}(\theta,0,0)/b_{c2}(0,0,0)$ normalized to its
$\theta=0^{\degree}$ value for a $d_{x^2-y^2}$-wave superconductor as $t\rightarrow0$, for $\overline{m}_1=\overline{m}_2$,
and $\overline{m}_1/\overline{m}_3 =0.6, 0.8, 1.0, 1.6$ and $1.8$ respectively. For each tetragonal effective mass anisotropy pictured, $b_{c2}(\theta,0,0)$ is a
monotonically increasing or decreasing function of $\theta$ at fixed $t$ and $\phi$.}
\label{fig:8}
\end{figure*}
{\it et al.}
\bibitem{Carbotte}
Prohammer M and Carbotte J R, {\it Upper critical field of $s$- and $d$-wave superconductors with anisotropic effective mass}, Phys. Rev. B {\bf 42}, 2032 (1990).
\bibitem{HallKlemm}
Hall B E and Klemm R A, {\it Microscopic model of the Knight shift in anisotropic and correlated metals}, J. Phys.: Condens. Matter {\bf 28}, 03LT01 (2016).
\bibitem{Klemm_Knight}
Klemm R A, {\it Towards a microscopic theory of the Knight shift in an anisotropic, multiband type-II superconductor}, Magnetochemistry {\bf 2018}, 4, 14.
|
\section{Introduction}
The popularity of techniques like Word2Vec \cite{mikolov2013distributed} and Wav2Vec \cite{baevski2020wav2vec} have shown the usefulness of embeddings, a learned low-level representation of high dimensional inputs, in solving human language processing problems. In recent years, with the push to address complex multimodal problems in an end-to-end manner, learning a common embedding space has been seen as a way of reducing the ``modality gap'' between speech and text data \cite{liu2020bridging}, not only with regards to bridging their continuous and discrete natures, but also in terms of widening data availability and complementing their features. This would enable for example, the combination of both acoustic and semantic information in a single model, which would be greatly benefit speech-based tasks such as speech translation (ST), automatic speech recognition (ASR), and spoken language understanding (SLU).
Indeed many state-of-the-art systems in these fields incorporate some form of a joint embedding as part of the model. Shared encoders have been employed together with distance-based losses to aggregate and learn from both text and speech data in several ST systems \cite{ye2021end,han2021learning,tang2021improving}. Dong et al. \cite{dong2021listen} instead utilized a separate pretrained BERT language model \cite{devlin-etal-2019-bert} to provide additional supervision to the speech encoder during training as a way of transferring semantic knowledge from a text source. Various models designed for SLU also take this approach, but treat the BERT supervision as a pretraining step \cite{Denisov_2020,chung2020splat,chuang2019speechbert}. Adversarial methods have also been used to bring text and speech representations closer \cite{chung2018unsupervised, alinejad2020effectively}.
Despite the resultant cross-modal embeddings proving their usefulness in pushing the state-of-the-art, there has been little work directly analysing their properties outside of evaluating them in the context of the downstream task. These same motivations inspired the body of work on probing \cite{conneau-etal-2018-cram,yossisenemb,47786}, that is checking for the retention of linguistic features in the embeddings via classifiers. However, existing probing literature mostly focus on pure text data. Outside of the speech-text domain, more of such research has been carried out for text and images \cite{aytar2017see, wang2019learning, radford2021learning, pmlr-v139-jia21b}. The CLIP \cite{radford2021learning} and ALIGN models \cite{pmlr-v139-jia21b} in particular demonstrated how a simple training task of matching images to captions, if scaled up, could produce embeddings that generalize well to different tasks and datasets.
In this paper, we study an analogous task for speech and text, which in the broader literature is referred to as semantic matching \cite{giunchiglia2007semantic}. The goal is to couple semantically-related objects from both domains by learning to map a speech utterance and its corresponding transcription to the same point in embedding space. Since the final objective is the embedding space itself, this allows us to gather some insight without the distraction of a further downstream task. Here, we employ a dual encoder arranged in a teacher-student setup, where the speech embedding space gradually adapts to the text embedding space. We investigate the extent of semantic alignment and knowledge transfer between a teacher text encoder and a student speech encoder directly through a retrieval accuracy metric and more generally via zero-shot classification of several speech datasets. Additionally, to ascertain the occurrence of transfer learning, we extend the probing analysis technique to cover both text and speech encoders. In general, we complement work related to the empirical analysis of embeddings and extend them to the multimodal context.
\section{Methods}
\subsection{Joint embedding model overview}
\begin{figure}[!htb]
\centering
\includegraphics[width=\linewidth]{JE_model.pdf}
\caption{Architecture for joint speech-text embedding training. A decoder is added when training for the multitask objective.}
\label{fig:model}
\end{figure}
As illustrated in Figure \ref{fig:model}, the overall model contains two pipelines that respectively transform speech and text features derived from sentence-level utterence and transcription pairs into a hidden representation. The text pipeline consists of a pretrained language model acting as the teacher whereas a transformer-based speech encoder \cite{wang2020fairseq} acts as the student. A projection head block was utilized after the speech encoder to maintain consistent dimensions between the speech and text embeddings. The final output of each model was mean pooled over the length of the sentence, and the resultant fixed-length vector normalized to unit length. To train for semantic matching, we minimize the L2 distance between the outputs. Teacher-student learning occurs throughout the training by backpropagating the loss only through the speech pipeline while keeping the weights of the text model fixed. This progressively brings the speech embeddings closer to the space defined by the text encoder to eventually construct a joint space.
To further enhance the representational strength of the speech embedding, we explored two techniques, namely pretraining and multitask training. For pretraining, the speech encoder was initialized with the encoder section of an encoder-decoder model trained for ASR. In contrast, the multitask setup trains the speech encoder for ASR from scratch simultaneous to minimizing the distance between embeddings. The overall loss is a combination of cross-entropy and the aforementioned L2 distance, weighted by $\gamma$ and $\beta$ respectively. The additional training schemes were found to be beneficial as preliminary experiments showed poor convergence for randomly-initialized models without additional supervision.
\subsection{Model analysis through zero-shot classification}
The joint embedding space was utilized for zero-shot classification as a way to study its potential for generalization to unseen distributions. Here, the task is to correctly classify speech utterances according to a set of labels. Labels are treated as ``natural language'' rather than categorical abstractions and directly fed to the text encoder in the same fashion as the text transcriptions used in training. The idea is the same as that of semantic matching; that an utterence and its label are semantically congruent and hence should lie in the same vicinity in embedding space. For classification, instead of directly learning a classifier over speech features, we seek the closest label embedding to a particular speech embedding in joint space, without requiring further training. In practice, we find the speech-label embedding pair with the highest cosine similarity for each speech input. In this paper, we only explore datasets containing labels that are direct transcriptions of the speech input. Presumably, the information required to carry out tasks with a higher level of abstraction (such as emotion classification) would not have been effectively captured by the current setup and is left for future work.
\subsection{Probing for learned language features}
Probing was used to verify the kind of language features captured by the models and if any transfer learning from the text to the speech encoder had taken place during the joint embedding training. Given sentence embeddings produced by the text model, we train a classifier on several language tasks. During this training, the text model itself is not updated, thus preserving its embedding space. If the classifier does well on a task, it implies that the embedding stores tangible information related to that task. To check for transfer learning, we make the assumption that the embedding space of the speech encoder is already coupled to the text encoder through the earlier joint embedding training and reuse the same classifiers to evaluate the speech embeddings.
\section{Experimental Setup}
\begin{table*}[tb]
\caption{Retrieval accuracy (\%) gauging how close speech-transcription embedding pairs are after joint embedding (JE) training. The last two rows refer to speech encoders in the randomly-initilaized and pretrained settings prior to the training, and serve as bottomline comparisons. Subsequent discussion will refer to the model variations by their designated letter.}
\label{table:accuracy}
\centering
\begin{tabular}{@{} cr|cccccccc @{}}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multirow{2}{*}{\textbf{speech encoder \& training conditions}} & \multicolumn{2}{c}{\textbf{libri-clean}} & \multicolumn{2}{c}{\textbf{libri-other}} & \multicolumn{2}{c}{\textbf{MuST-C}} & \multicolumn{2}{c}{\textbf{CoVoST}} \\
& &\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} &
\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} &
\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} &
\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} \\
\midrule
A & random & 30.23 & 12.02 & 15.89 & 5.78 & 1.97 & 0.90 & 0.01 & 0.01 \\
B & pretrained & 98.82 & 83.42 & 94.25 & 67.88 & 44.33 & 15.55 & \textbf{2.15} & \textbf{0.35} \\
C & pretrained, only projection head updated & 98.32 & 63.89 & 91.15 & 47.64 & 37.90 & 10.60 & 1.62 & 0.15 \\
D & random, multitask, $\gamma$=1, $\beta$=1 & 84.20 & 23.93 & 65.33 & 16.26 & 13.20 & 2.03 & 0.25 & 0.03 \\
E & random, multitask, $\gamma$=1, $\beta$=10 & 88.82 & 29.01 & 72.07 & 20.01 & 17.43 & 3.10 & 0.49 & 0.02 \\
F & pretrained, multitask, $\gamma$=1, $\beta$=100 & \textbf{99.12} & \textbf{91.91} & \textbf{95.75} & \textbf{79.58} & \textbf{50.71} & \textbf{21.04} & 1.55 & 0.25 \\
\hdashline
G & random, before JE training & 0.04 & 0.00 & 0.00 & 0.07 & 0.06 & 0.06 & 0.01 & 0.01 \\
H & pretrained, before JE training & 0.04 & 0.08 & 0.03 & 0.07 & 0.06 & 0.06 & 0.00 & 0.00 \\
\bottomrule
\end{tabular
\end{table*}
\subsection{Text encoder}
A pretrained RoBERTa large model \cite{liu2019roberta} from the Fairseq library \cite{ott2019fairseq} was adopted as the backbone of the text pipeline. RoBERTa improves on the masked language modeling objective of BERT \cite{devlin-etal-2019-bert} by tweaking several hyperparameters, removing the next-sentence prediction objective, and training on a much larger dataset, among other innovations. Before input into RoBERTa, byte pair encoding (BPE) was first applied to the text sequence at byte-level. RoBERTa large contains 24 transformer layers and an output embedding size of 1024.
\subsection{Speech encoder}
We used the \texttt{s2t\_transformer\_s} architecture from Fairseq, which consists of several convolutional neural network layers to downsample the input followed by 12 transformer layers, with four attention heads per layer and an output embedding size of 256. The post-encoder projection head contains a Gaussian Error Linear Unit (GELU) activation between two linear layers, followed by dropout and layernorm. The speech pipeline for the multitask objective additionally includes a six layer transformer decoder. To ascertain the impact of pretraining, we compared randomly initialized encoders against encoders pretrained for ASR in our experiments, for which the Fairseq checkpoint trained on Librispeech \cite{panayotov2015librispeech} was used. Raw audio was first converted into log-Mel filter bank spectrograms with 80 bins, frame length of 25ms, and shift of 10ms, then dynamically augmented following SpecAugment \cite{Park2019SpecAugmentAS} before input into the speech encoder.
\subsection{Joint speech-text embedding training}
The model was trained for the joint embedding objective on the Librispeech dataset using four NVIDIA V100 GPUs for 100 epochs with Adam optimizer, an inverse square root learning rate scheduler with a peak learning rate of $1\text{e-}3$, and 10000 warmup steps. For the multitask setup, early experiments showed that the contribution of the cross-entropy loss for ASR was about two orders magnitude larger than the L2 loss for embedding distance. The $\beta$ parameter was tweaked between 1-100 to increase the influence of the L2 loss. For all training variations, the checkpoint with the smallest validation loss was chosen for further analysis. In addition to the Librispeech test sets (libri-clean, libri-other), MuST-C (en-zh) \cite{di-gangi-etal-2019-must}, and CoVoST~2 (en) \cite{wang-etal-2020-covost} were also used for evaluation. For these, only the English speech segments and transcriptions were used.
\subsection{Details for zero-shot classification and probing tasks}
We tested for zero-shot classification on three datasets, namely AudioMNIST \cite{becker2018interpreting}, Speech Commands \cite{warden2018speech}, and TIMIT \cite{garofolo1993timit}. AudioMNIST contains 30000 audio recordings of spoken digits 0-9 with 50 repetitions per digit per speaker. For labels, digits were spelled out instead of using the numerical form (i.e. ``one" rather than ``1"). A subset of Speech Commands was used comprising 10 auxiliary words: ``Bed", ``Bird", ``Cat", ``Dog", ``Happy", ``House", ``Marvin", ``Sheila", ``Tree", and ``Wow". Each word is spoken once by each speaker, making up 20408 recordings in total. The SX subset of TIMIT was used to explore joint embeddings with multi-word sentences acting as labels. SX sentences were designed to provide a good coverage of pairs of phones. We randomly selected 10 SX sentences among TIMIT's TEST set for labels, with seven pronunciations each for classification.
The text datasets provided by the SentEval toolkit \cite{conneau-etal-2018-cram} were used for probing. These span 10 semantic and syntactic tasks, from word-level objectives, such as word constituents, surface information like sentence length, to grammatical structures like past and present tense. We refer the reader to the original paper for further details on each task. Each task is split into 100000 training, 10000 validation, and 10000 test examples. The classifier is a neural network comprising of two linear layers with dropout and a tanh activation. It was trained for 10 epochs and the checkpoint with the highest validation accuracy was chosen as the final model. To probe the speech encoder, the test sentences were first transformed into speech using an open-source text-to-speech system (TTS) by Silero \cite{Silero}, with the speaker model \texttt{lj\_16khz}.
\section{Results}
\subsection{Semantic matching using retrieval}
To assess the effectiveness of our setup for semantic matching, we measure the retrieval quality of the model given a speech or text input using cosine similarity as a proxy. This analysis gives us a reasonable impression of how close the final embeddings are. For a given speech input, a prediction is counted if the retrieved text embedding with the highest cosine similarity belongs to its associated transcription (S-T), and vice versa for a given text input (T-S). Table \ref{table:accuracy} shows the percentage accuracy of retrieval across various the test sets and models.
Prior to the joint embedding training, speech and text embeddings were far apart irrespective of whether the speech encoder was pretrained for ASR (models G, H). However, comparing the randomly initialized to the pretrained encoder after training (models A, B), shows the latter having a significant impact on the final closeness of the embeddings. We suspect given the pretrained ASR encoder providing much better audio features from the outset, the training procedure better focuses on distribution shift rather than feature extraction. This is evidenced by the relatively slight drop in scores when freezing the speech encoder and only updating the projector head (model C). Starting with a pretrained ASR model also provides better robustness on out-of-training data, although this is heavily influenced by the quality of ASR in relation to the particular test set. The large gap in the CoVoST and MuST-C accuracies, and in turn against the in-training Librispeech data, are reflected in their respective word error rates (WER): 104.7 (CoVoST), 75.5 (MuST-C), 9.1 (libri-other), and 4.5 (libri-clean) with the Fairseq Librispeech checkpoint used for pretraining
\comment{
\begin{figure}[!htb]
\centering
\includegraphics[width=0.85\linewidth]{wer_acc.pdf}
\caption{The effect of ASR model on S-T retrieval accuracy. The better the ASR model performance (smaller WER) on a particular test dataset the better the semantic matching strength with text embeddings. The WER was calculated with Model F from Table \ref{table:accuracy}.}
\label{fig:wer_acc}
\end{figure}
}
The multitask setups which train the speech encoder for ASR from scratch (models D, E), did not perform as well as the models initialized with pretrained checkpoints (models B, C, F), especially for S-T retrieval accuracy. The learned speech features conferred by the ASR pretraining were found to be essential in the initial stages of the joint embedding training, leading to smoother optimization and ``closer'' final embeddings. The best overall model, model F, combines both pretraining and multitask, showing the benefit of retaining ASR ability concurrent to minimizing the embedding distance.
\begin{figure*}[tb]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{tsne_roberta_asrpre_mnist_nl.png}
\label{fig:y equals x}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{tsne_roberta_asrpre_sc_nl.png}
\label{fig:three sin x}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{tsne_roberta_asrpre_timit_nl.png}
\label{fig:five over x}
\end{subfigure}
\caption{t-SNE plots of AudioMNIST (left), Speech Commands (centre), and TIMIT (right). Speech embeddings are marked by small circles while each label embedding is shown as a different, larger shape. Colour demarcates class for both modalities.}
\label{fig:tsne}
\end{figure*}
It is noteworthy that the T-S results were consistently higher than S-T. We suspect this asymmetry stems from the teacher-student learning where the text embeddings remain static but the speech embeddings get pulled apart to fit the space. The final position of the speech embedding plane may leave a particular point closer to its text counterpart compared to other points on the speech plane and hence be retrieved correctly given a text input as in T-S. Yet the same speech point may still be at a distance where it is closer to other unrelated points on the text embedding plane compared to its counterpart, which would make the corresponding S-T retrieval wrong. This phenomenon is exacerbated in training setups with relatively poorer convergence, causing S-T accuracy to deteriorate faster than T-S in those cases. We note that such asymmetry was not reported in works without a teacher-student model such as ALIGN \cite{pmlr-v139-jia21b}.
\subsection{Zero-shot speech classification}
The mechanism for retrieving transcriptions given a speech utterance as in S-T above was repurposed for classification by replacing the transcriptions with class labels. Table \ref{tab:zeroshot} summarizes the classification results for the top performing models B and F in the previous retrieval task, together with the untrained bottomline model H in Table \ref{table:accuracy}. The accuracy, in particular on AudioMNIST and Speech Commands, were only slightly better than the bottomline, which itself was close to random choice. However, scores were much higher on TIMIT. To provide more insight into the reasons behind this, speech and label embeddings were plotted using t-SNE in Figure \ref{fig:tsne}.
\begin{table}[htb]
\caption{Classification accuracy (\%) of Models B, F, and H from Tab. \ref{table:accuracy} on AudioMNIST (AM), Speech Commands (SC) and TIMIT. Each dataset contained 10 classes.}
\label{tab:zeroshot}
\centering
\begin{tabular}{@{} r|ccc @{}}
\toprule
\textbf{Dataset} & \multicolumn{1}{c}{\textbf{B}} & \multicolumn{1}{c}{\textbf{F}} & \multicolumn{1}{c}{\textbf{H}}\\
\midrule
AM & 12.59 & 15.88 & 10.28 \\
SC & 14.47 & 14.27 & 9.31 \\
TIMIT & 42.86 & 67.14 & 10.00 \\
\bottomrule
\end{tabular
\end{table}
Distinct phenomena were observed for the three datasets. For AudioMNIST, speech embeddings showed clear separation among the different classes (Fig.\ref{fig:tsne}, left). On the other hand, several label embeddings were found to overlap in the t-SNE plot, suggesting that text encoder RoBERTa was projecting them to a similar point in embedding space. This was a reasonable behaviour given that it was trained to produce contextualized word embeddings, and without any further context, individual digits may have been treated as semantically identical. Indeed, the more diverse set of words forming the Speech Commands labels were clearly separable (Fig.\ref{fig:tsne}, centre). On this dataset however, the speech encoder did not produce embeddings aligned to the classes, outside of some minor clusters. Since both encoders were trained on sentences rather than individual words, we expected them to do better on TIMIT, where each ``label'' was a full sentence (Fig.\ref{fig:tsne}, right). Both clustering of the speech input and alignment between the two modalities were seen, corroborating the better accuracy scores.
\subsection{Probing tasks}
To observe the extent of semantic knowledge transfer taking place, we compare the relative accuracy of RoBERTa against Models B, F, and bottomline H on 10 language probing tasks, shown in Table \ref{tab:prob}. Firstly, RoBERTa itself was poor at several tasks, especially the ones that required word-level knowledge such as word constituents (WC) and semantic odd man out (SOMO). We attribute this to the mean pooling carried out post encoder, which may have resulted in more ambiguous word representations. Coordination inversion (CoordInv), where the order of coordinate clauses were inverted half the time, may also have been impacted by mean pooling the original embeddings.
\begin{table}[htb]
\caption{Classification accuracy (\%) on 10 probing tasks for Models B, F, and H. RBT refers to RoBERTa.}
\label{tab:prob}
\centering
\begin{tabular}{@{} r|c|cccc @{}}
\toprule
\textbf{Task} & \textbf{Classes} & \multicolumn{1}{c}{\textbf{RBT}} & \multicolumn{1}{c}{\textbf{B}} & \multicolumn{1}{c}{\textbf{F}} & \multicolumn{1}{c}{\textbf{H}}\\
\midrule
SentLen & 6 & 36.65 & 38.19 & 37.82 & 16.19 \\
WC & 1000 & 0.10 & 0.10 & 0.00 & 0.10 \\
TreeDepth & 7 & 20.48 & 19.60 & 19.95 & 18.47 \\
TopConst & 20 & 29.25 & 22.68 & 21.57 & 5.14\\
BShift & 2 & 74.82 & 54.19 & 54.42 & 48.83\\
Tense & 2 & 81.89 & 75.41 & 77.59 & 50.01 \\
SubjNum & 2& 78.58& 73.86& 74.95& 49.96\\
ObjNum & 2& 76.26& 74.42& 75.66& 50.17\\
SOMO & 2& 60.16& 49.78& 50.56& 49.66\\
CoordInv & 2& 57.92& 50.45& 51.06& 51.19\\
\bottomrule
\end{tabular
\end{table}
Overall, we see some evidence of knowledge transfer of some language properties but not all of them. This is seen in tasks for which the speech encoder gets much closer to the RoBERTa score after training, including for sentence length (SentLen), top constituents (TopConst) which tests for sentence structure, past and present tense (Tense), and plurality of the subject (SubjNum) or object (ObjNum) in a sentence. Some tasks that do not carry over may have been due to the continuous nature of audio in contrast to discrete text. For example, bigram shift (BShift), which inverts two adjacent words at random, may have been much harder to detect directly from speech compared to text. Additionally, we acknowledge that some degree of error stems from the imperfect TTS conversion to create the dataset. For reference, the transformed SentLen test set had a WER of 26.78 with model F.
\comment{
\subsection{Analysis in relation to ASR}
\label{sec:asr}
\note{probably remove this section}
We noticed since RoBERTa text model trained on the same domain as text corpus it is evaluated on (e.g., MuST-C), it benefits the ASR model trained in multitask manner with joint embeddings and outperform regular ASR model (transformer\_s), see Tab.~\ref{tab:wer_asr}.
\begin{table}[htb]
\caption{Comparing Librispeech ASR checkpoint \cite{panayotov2015librispeech} against multitask-trained ASR with joint embeddings from Tab.~\ref{table:accuracy} \emph{(pretrained, multitask, $\gamma$=1, $\beta$=100)}. Word error rate (WER,\%).}
\label{tab:wer_asr}
\centering
\begin{tabular}{@{} r|cccc @{}}
\toprule
\textbf{Models} & \textbf{libri-clean} & \textbf{libri-other} & \textbf{must-c} & \textbf{covost}\\
\midrule
transformer\_s & 4.52 & 9.10 & 75.47 & 104.66 \\
multitask\_asr & 4.80 & 10.78 & 66.29 & 108.19 \\
\bottomrule
\end{tabular}
\end{table}
}
\section{Conclusions}
We studied some properties of a joint speech-text embedding space constructed through a semantic matching objective. The model was trained by minimizing the distance between the embedding outputs of a teacher RoBERTa text encoder and a student speech transformer encoder. We found that ASR pretraining of the speech encoder was essential for better semantic matching, measured in terms of retrieval accuracy. Performance could be further improved by combining it with a multitask objective incorporating both the joint embedding training and further ASR fine-tuning. Using the models for zero-shot classification was found to only work well if both speech and text embeddings manage to capture the properties of the classes. Probing showed that some but not all linguistic properties were transferred from the text to the speech model. In future work, we plan to investigate embeddings and training schemes that would enable semantic coupling between less constrained speech-text pairs compared to utterences and transcriptions, which may be applicable to a wider range of downstream tasks.
\clearpage
\bibliographystyle{IEEEtran}
\section{Introduction}
The popularity of techniques like Word2Vec \cite{mikolov2013distributed} and Wav2Vec \cite{baevski2020wav2vec} have shown the usefulness of embeddings, a learned low-level representation of high dimensional inputs, in solving human language processing problems. In recent years, with the push to address complex multimodal problems in an end-to-end manner, learning a common embedding space has been seen as a way of reducing the ``modality gap'' between speech and text data \cite{liu2020bridging}, not only with regards to bridging their continuous and discrete natures, but also in terms of widening data availability and complementing their features. This would enable for example, the combination of both acoustic and semantic information in a single model, which would be greatly benefit speech-based tasks such as speech translation (ST), automatic speech recognition (ASR), and spoken language understanding (SLU).
Indeed many state-of-the-art systems in these fields incorporate some form of a joint embedding as part of the model. Shared encoders have been employed together with distance-based losses to aggregate and learn from both text and speech data in several ST systems \cite{ye2021end,han2021learning,tang2021improving}. Dong et al. \cite{dong2021listen} instead utilized a separate pretrained BERT language model \cite{devlin-etal-2019-bert} to provide additional supervision to the speech encoder during training as a way of transferring semantic knowledge from a text source. Various models designed for SLU also take this approach, but treat the BERT supervision as a pretraining step \cite{Denisov_2020,chung2020splat,chuang2019speechbert}. Adversarial methods have also been used to bring text and speech representations closer \cite{chung2018unsupervised, alinejad2020effectively}.
Despite the resultant cross-modal embeddings proving their usefulness in pushing the state-of-the-art, there has been little work directly analysing their properties outside of evaluating them in the context of the downstream task. These same motivations inspired the body of work on probing \cite{conneau-etal-2018-cram,yossisenemb,47786}, that is checking for the retention of linguistic features in the embeddings via classifiers. However, existing probing literature mostly focus on pure text data. Outside of the speech-text domain, more of such research has been carried out for text and images \cite{aytar2017see, wang2019learning, radford2021learning, pmlr-v139-jia21b}. The CLIP \cite{radford2021learning} and ALIGN models \cite{pmlr-v139-jia21b} in particular demonstrated how a simple training task of matching images to captions, if scaled up, could produce embeddings that generalize well to different tasks and datasets.
In this paper, we study an analogous task for speech and text, which in the broader literature is referred to as semantic matching \cite{giunchiglia2007semantic}. The goal is to couple semantically-related objects from both domains by learning to map a speech utterance and its corresponding transcription to the same point in embedding space. Since the final objective is the embedding space itself, this allows us to gather some insight without the distraction of a further downstream task. Here, we employ a dual encoder arranged in a teacher-student setup, where the speech embedding space gradually adapts to the text embedding space. We investigate the extent of semantic alignment and knowledge transfer between a teacher text encoder and a student speech encoder directly through a retrieval accuracy metric and more generally via zero-shot classification of several speech datasets. Additionally, to ascertain the occurrence of transfer learning, we extend the probing analysis technique to cover both text and speech encoders. In general, we complement work related to the empirical analysis of embeddings and extend them to the multimodal context.
\section{Methods}
\subsection{Joint embedding model overview}
\begin{figure}[!htb]
\centering
\includegraphics[width=\linewidth]{JE_model.pdf}
\caption{Architecture for joint speech-text embedding training. A decoder is added when training for the multitask objective.}
\label{fig:model}
\end{figure}
As illustrated in Figure \ref{fig:model}, the overall model contains two pipelines that respectively transform speech and text features derived from sentence-level utterence and transcription pairs into a hidden representation. The text pipeline consists of a pretrained language model acting as the teacher whereas a transformer-based speech encoder \cite{wang2020fairseq} acts as the student. A projection head block was utilized after the speech encoder to maintain consistent dimensions between the speech and text embeddings. The final output of each model was mean pooled over the length of the sentence, and the resultant fixed-length vector normalized to unit length. To train for semantic matching, we minimize the L2 distance between the outputs. Teacher-student learning occurs throughout the training by backpropagating the loss only through the speech pipeline while keeping the weights of the text model fixed. This progressively brings the speech embeddings closer to the space defined by the text encoder to eventually construct a joint space.
To further enhance the representational strength of the speech embedding, we explored two techniques, namely pretraining and multitask training. For pretraining, the speech encoder was initialized with the encoder section of an encoder-decoder model trained for ASR. In contrast, the multitask setup trains the speech encoder for ASR from scratch simultaneous to minimizing the distance between embeddings. The overall loss is a combination of cross-entropy and the aforementioned L2 distance, weighted by $\gamma$ and $\beta$ respectively. The additional training schemes were found to be beneficial as preliminary experiments showed poor convergence for randomly-initialized models without additional supervision.
\subsection{Model analysis through zero-shot classification}
The joint embedding space was utilized for zero-shot classification as a way to study its potential for generalization to unseen distributions. Here, the task is to correctly classify speech utterances according to a set of labels. Labels are treated as ``natural language'' rather than categorical abstractions and directly fed to the text encoder in the same fashion as the text transcriptions used in training. The idea is the same as that of semantic matching; that an utterence and its label are semantically congruent and hence should lie in the same vicinity in embedding space. For classification, instead of directly learning a classifier over speech features, we seek the closest label embedding to a particular speech embedding in joint space, without requiring further training. In practice, we find the speech-label embedding pair with the highest cosine similarity for each speech input. In this paper, we only explore datasets containing labels that are direct transcriptions of the speech input. Presumably, the information required to carry out tasks with a higher level of abstraction (such as emotion classification) would not have been effectively captured by the current setup and is left for future work.
\subsection{Probing for learned language features}
Probing was used to verify the kind of language features captured by the models and if any transfer learning from the text to the speech encoder had taken place during the joint embedding training. Given sentence embeddings produced by the text model, we train a classifier on several language tasks. During this training, the text model itself is not updated, thus preserving its embedding space. If the classifier does well on a task, it implies that the embedding stores tangible information related to that task. To check for transfer learning, we make the assumption that the embedding space of the speech encoder is already coupled to the text encoder through the earlier joint embedding training and reuse the same classifiers to evaluate the speech embeddings.
\section{Experimental Setup}
\begin{table*}[tb]
\caption{Retrieval accuracy (\%) gauging how close speech-transcription embedding pairs are after joint embedding (JE) training. The last two rows refer to speech encoders in the randomly-initilaized and pretrained settings prior to the training, and serve as bottomline comparisons. Subsequent discussion will refer to the model variations by their designated letter.}
\label{table:accuracy}
\centering
\begin{tabular}{@{} cr|cccccccc @{}}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multirow{2}{*}{\textbf{speech encoder \& training conditions}} & \multicolumn{2}{c}{\textbf{libri-clean}} & \multicolumn{2}{c}{\textbf{libri-other}} & \multicolumn{2}{c}{\textbf{MuST-C}} & \multicolumn{2}{c}{\textbf{CoVoST}} \\
& &\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} &
\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} &
\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} &
\multicolumn{1}{c}{T-S} & \multicolumn{1}{c}{S-T} \\
\midrule
A & random & 30.23 & 12.02 & 15.89 & 5.78 & 1.97 & 0.90 & 0.01 & 0.01 \\
B & pretrained & 98.82 & 83.42 & 94.25 & 67.88 & 44.33 & 15.55 & \textbf{2.15} & \textbf{0.35} \\
C & pretrained, only projection head updated & 98.32 & 63.89 & 91.15 & 47.64 & 37.90 & 10.60 & 1.62 & 0.15 \\
D & random, multitask, $\gamma$=1, $\beta$=1 & 84.20 & 23.93 & 65.33 & 16.26 & 13.20 & 2.03 & 0.25 & 0.03 \\
E & random, multitask, $\gamma$=1, $\beta$=10 & 88.82 & 29.01 & 72.07 & 20.01 & 17.43 & 3.10 & 0.49 & 0.02 \\
F & pretrained, multitask, $\gamma$=1, $\beta$=100 & \textbf{99.12} & \textbf{91.91} & \textbf{95.75} & \textbf{79.58} & \textbf{50.71} & \textbf{21.04} & 1.55 & 0.25 \\
\hdashline
G & random, before JE training & 0.04 & 0.00 & 0.00 & 0.07 & 0.06 & 0.06 & 0.01 & 0.01 \\
H & pretrained, before JE training & 0.04 & 0.08 & 0.03 & 0.07 & 0.06 & 0.06 & 0.00 & 0.00 \\
\bottomrule
\end{tabular
\end{table*}
\subsection{Text encoder}
A pretrained RoBERTa large model \cite{liu2019roberta} from the Fairseq library \cite{ott2019fairseq} was adopted as the backbone of the text pipeline. RoBERTa improves on the masked language modeling objective of BERT \cite{devlin-etal-2019-bert} by tweaking several hyperparameters, removing the next-sentence prediction objective, and training on a much larger dataset, among other innovations. Before input into RoBERTa, byte pair encoding (BPE) was first applied to the text sequence at byte-level. RoBERTa large contains 24 transformer layers and an output embedding size of 1024.
\subsection{Speech encoder}
We used the \texttt{s2t\_transformer\_s} architecture from Fairseq, which consists of several convolutional neural network layers to downsample the input followed by 12 transformer layers, with four attention heads per layer and an output embedding size of 256. The post-encoder projection head contains a Gaussian Error Linear Unit (GELU) activation between two linear layers, followed by dropout and layernorm. The speech pipeline for the multitask objective additionally includes a six layer transformer decoder. To ascertain the impact of pretraining, we compared randomly initialized encoders against encoders pretrained for ASR in our experiments, for which the Fairseq checkpoint trained on Librispeech \cite{panayotov2015librispeech} was used. Raw audio was first converted into log-Mel filter bank spectrograms with 80 bins, frame length of 25ms, and shift of 10ms, then dynamically augmented following SpecAugment \cite{Park2019SpecAugmentAS} before input into the speech encoder.
\subsection{Joint speech-text embedding training}
The model was trained for the joint embedding objective on the Librispeech dataset using four NVIDIA V100 GPUs for 100 epochs with Adam optimizer, an inverse square root learning rate scheduler with a peak learning rate of $1\text{e-}3$, and 10000 warmup steps. For the multitask setup, early experiments showed that the contribution of the cross-entropy loss for ASR was about two orders magnitude larger than the L2 loss for embedding distance. The $\beta$ parameter was tweaked between 1-100 to increase the influence of the L2 loss. For all training variations, the checkpoint with the smallest validation loss was chosen for further analysis. In addition to the Librispeech test sets (libri-clean, libri-other), MuST-C (en-zh) \cite{di-gangi-etal-2019-must}, and CoVoST~2 (en) \cite{wang-etal-2020-covost} were also used for evaluation. For these, only the English speech segments and transcriptions were used.
\subsection{Details for zero-shot classification and probing tasks}
We tested for zero-shot classification on three datasets, namely AudioMNIST \cite{becker2018interpreting}, Speech Commands \cite{warden2018speech}, and TIMIT \cite{garofolo1993timit}. AudioMNIST contains 30000 audio recordings of spoken digits 0-9 with 50 repetitions per digit per speaker. For labels, digits were spelled out instead of using the numerical form (i.e. ``one" rather than ``1"). A subset of Speech Commands was used comprising 10 auxiliary words: ``Bed", ``Bird", ``Cat", ``Dog", ``Happy", ``House", ``Marvin", ``Sheila", ``Tree", and ``Wow". Each word is spoken once by each speaker, making up 20408 recordings in total. The SX subset of TIMIT was used to explore joint embeddings with multi-word sentences acting as labels. SX sentences were designed to provide a good coverage of pairs of phones. We randomly selected 10 SX sentences among TIMIT's TEST set for labels, with seven pronunciations each for classification.
The text datasets provided by the SentEval toolkit \cite{conneau-etal-2018-cram} were used for probing. These span 10 semantic and syntactic tasks, from word-level objectives, such as word constituents, surface information like sentence length, to grammatical structures like past and present tense. We refer the reader to the original paper for further details on each task. Each task is split into 100000 training, 10000 validation, and 10000 test examples. The classifier is a neural network comprising of two linear layers with dropout and a tanh activation. It was trained for 10 epochs and the checkpoint with the highest validation accuracy was chosen as the final model. To probe the speech encoder, the test sentences were first transformed into speech using an open-source text-to-speech system (TTS) by Silero \cite{Silero}, with the speaker model \texttt{lj\_16khz}.
\section{Results}
\subsection{Semantic matching using retrieval}
To assess the effectiveness of our setup for semantic matching, we measure the retrieval quality of the model given a speech or text input using cosine similarity as a proxy. This analysis gives us a reasonable impression of how close the final embeddings are. For a given speech input, a prediction is counted if the retrieved text embedding with the highest cosine similarity belongs to its associated transcription (S-T), and vice versa for a given text input (T-S). Table \ref{table:accuracy} shows the percentage accuracy of retrieval across various the test sets and models.
Prior to the joint embedding training, speech and text embeddings were far apart irrespective of whether the speech encoder was pretrained for ASR (models G, H). However, comparing the randomly initialized to the pretrained encoder after training (models A, B), shows the latter having a significant impact on the final closeness of the embeddings. We suspect given the pretrained ASR encoder providing much better audio features from the outset, the training procedure better focuses on distribution shift rather than feature extraction. This is evidenced by the relatively slight drop in scores when freezing the speech encoder and only updating the projector head (model C). Starting with a pretrained ASR model also provides better robustness on out-of-training data, although this is heavily influenced by the quality of ASR in relation to the particular test set. The large gap in the CoVoST and MuST-C accuracies, and in turn against the in-training Librispeech data, are reflected in their respective word error rates (WER): 104.7 (CoVoST), 75.5 (MuST-C), 9.1 (libri-other), and 4.5 (libri-clean) with the Fairseq Librispeech checkpoint used for pretraining
\comment{
\begin{figure}[!htb]
\centering
\includegraphics[width=0.85\linewidth]{wer_acc.pdf}
\caption{The effect of ASR model on S-T retrieval accuracy. The better the ASR model performance (smaller WER) on a particular test dataset the better the semantic matching strength with text embeddings. The WER was calculated with Model F from Table \ref{table:accuracy}.}
\label{fig:wer_acc}
\end{figure}
}
The multitask setups which train the speech encoder for ASR from scratch (models D, E), did not perform as well as the models initialized with pretrained checkpoints (models B, C, F), especially for S-T retrieval accuracy. The learned speech features conferred by the ASR pretraining were found to be essential in the initial stages of the joint embedding training, leading to smoother optimization and ``closer'' final embeddings. The best overall model, model F, combines both pretraining and multitask, showing the benefit of retaining ASR ability concurrent to minimizing the embedding distance.
\begin{figure*}[tb]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{tsne_roberta_asrpre_mnist_nl.png}
\label{fig:y equals x}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{tsne_roberta_asrpre_sc_nl.png}
\label{fig:three sin x}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{tsne_roberta_asrpre_timit_nl.png}
\label{fig:five over x}
\end{subfigure}
\caption{t-SNE plots of AudioMNIST (left), Speech Commands (centre), and TIMIT (right). Speech embeddings are marked by small circles while each label embedding is shown as a different, larger shape. Colour demarcates class for both modalities.}
\label{fig:tsne}
\end{figure*}
It is noteworthy that the T-S results were consistently higher than S-T. We suspect this asymmetry stems from the teacher-student learning where the text embeddings remain static but the speech embeddings get pulled apart to fit the space. The final position of the speech embedding plane may leave a particular point closer to its text counterpart compared to other points on the speech plane and hence be retrieved correctly given a text input as in T-S. Yet the same speech point may still be at a distance where it is closer to other unrelated points on the text embedding plane compared to its counterpart, which would make the corresponding S-T retrieval wrong. This phenomenon is exacerbated in training setups with relatively poorer convergence, causing S-T accuracy to deteriorate faster than T-S in those cases. We note that such asymmetry was not reported in works without a teacher-student model such as ALIGN \cite{pmlr-v139-jia21b}.
\subsection{Zero-shot speech classification}
The mechanism for retrieving transcriptions given a speech utterance as in S-T above was repurposed for classification by replacing the transcriptions with class labels. Table \ref{tab:zeroshot} summarizes the classification results for the top performing models B and F in the previous retrieval task, together with the untrained bottomline model H in Table \ref{table:accuracy}. The accuracy, in particular on AudioMNIST and Speech Commands, were only slightly better than the bottomline, which itself was close to random choice. However, scores were much higher on TIMIT. To provide more insight into the reasons behind this, speech and label embeddings were plotted using t-SNE in Figure \ref{fig:tsne}.
\begin{table}[htb]
\caption{Classification accuracy (\%) of Models B, F, and H from Tab. \ref{table:accuracy} on AudioMNIST (AM), Speech Commands (SC) and TIMIT. Each dataset contained 10 classes.}
\label{tab:zeroshot}
\centering
\begin{tabular}{@{} r|ccc @{}}
\toprule
\textbf{Dataset} & \multicolumn{1}{c}{\textbf{B}} & \multicolumn{1}{c}{\textbf{F}} & \multicolumn{1}{c}{\textbf{H}}\\
\midrule
AM & 12.59 & 15.88 & 10.28 \\
SC & 14.47 & 14.27 & 9.31 \\
TIMIT & 42.86 & 67.14 & 10.00 \\
\bottomrule
\end{tabular
\end{table}
Distinct phenomena were observed for the three datasets. For AudioMNIST, speech embeddings showed clear separation among the different classes (Fig.\ref{fig:tsne}, left). On the other hand, several label embeddings were found to overlap in the t-SNE plot, suggesting that text encoder RoBERTa was projecting them to a similar point in embedding space. This was a reasonable behaviour given that it was trained to produce contextualized word embeddings, and without any further context, individual digits may have been treated as semantically identical. Indeed, the more diverse set of words forming the Speech Commands labels were clearly separable (Fig.\ref{fig:tsne}, centre). On this dataset however, the speech encoder did not produce embeddings aligned to the classes, outside of some minor clusters. Since both encoders were trained on sentences rather than individual words, we expected them to do better on TIMIT, where each ``label'' was a full sentence (Fig.\ref{fig:tsne}, right). Both clustering of the speech input and alignment between the two modalities were seen, corroborating the better accuracy scores.
\subsection{Probing tasks}
To observe the extent of semantic knowledge transfer taking place, we compare the relative accuracy of RoBERTa against Models B, F, and bottomline H on 10 language probing tasks, shown in Table \ref{tab:prob}. Firstly, RoBERTa itself was poor at several tasks, especially the ones that required word-level knowledge such as word constituents (WC) and semantic odd man out (SOMO). We attribute this to the mean pooling carried out post encoder, which may have resulted in more ambiguous word representations. Coordination inversion (CoordInv), where the order of coordinate clauses were inverted half the time, may also have been impacted by mean pooling the original embeddings.
\begin{table}[htb]
\caption{Classification accuracy (\%) on 10 probing tasks for Models B, F, and H. RBT refers to RoBERTa.}
\label{tab:prob}
\centering
\begin{tabular}{@{} r|c|cccc @{}}
\toprule
\textbf{Task} & \textbf{Classes} & \multicolumn{1}{c}{\textbf{RBT}} & \multicolumn{1}{c}{\textbf{B}} & \multicolumn{1}{c}{\textbf{F}} & \multicolumn{1}{c}{\textbf{H}}\\
\midrule
SentLen & 6 & 36.65 & 38.19 & 37.82 & 16.19 \\
WC & 1000 & 0.10 & 0.10 & 0.00 & 0.10 \\
TreeDepth & 7 & 20.48 & 19.60 & 19.95 & 18.47 \\
TopConst & 20 & 29.25 & 22.68 & 21.57 & 5.14\\
BShift & 2 & 74.82 & 54.19 & 54.42 & 48.83\\
Tense & 2 & 81.89 & 75.41 & 77.59 & 50.01 \\
SubjNum & 2& 78.58& 73.86& 74.95& 49.96\\
ObjNum & 2& 76.26& 74.42& 75.66& 50.17\\
SOMO & 2& 60.16& 49.78& 50.56& 49.66\\
CoordInv & 2& 57.92& 50.45& 51.06& 51.19\\
\bottomrule
\end{tabular
\end{table}
Overall, we see some evidence of knowledge transfer of some language properties but not all of them. This is seen in tasks for which the speech encoder gets much closer to the RoBERTa score after training, including for sentence length (SentLen), top constituents (TopConst) which tests for sentence structure, past and present tense (Tense), and plurality of the subject (SubjNum) or object (ObjNum) in a sentence. Some tasks that do not carry over may have been due to the continuous nature of audio in contrast to discrete text. For example, bigram shift (BShift), which inverts two adjacent words at random, may have been much harder to detect directly from speech compared to text. Additionally, we acknowledge that some degree of error stems from the imperfect TTS conversion to create the dataset. For reference, the transformed SentLen test set had a WER of 26.78 with model F.
\comment{
\subsection{Analysis in relation to ASR}
\label{sec:asr}
\note{probably remove this section}
We noticed since RoBERTa text model trained on the same domain as text corpus it is evaluated on (e.g., MuST-C), it benefits the ASR model trained in multitask manner with joint embeddings and outperform regular ASR model (transformer\_s), see Tab.~\ref{tab:wer_asr}.
\begin{table}[htb]
\caption{Comparing Librispeech ASR checkpoint \cite{panayotov2015librispeech} against multitask-trained ASR with joint embeddings from Tab.~\ref{table:accuracy} \emph{(pretrained, multitask, $\gamma$=1, $\beta$=100)}. Word error rate (WER,\%).}
\label{tab:wer_asr}
\centering
\begin{tabular}{@{} r|cccc @{}}
\toprule
\textbf{Models} & \textbf{libri-clean} & \textbf{libri-other} & \textbf{must-c} & \textbf{covost}\\
\midrule
transformer\_s & 4.52 & 9.10 & 75.47 & 104.66 \\
multitask\_asr & 4.80 & 10.78 & 66.29 & 108.19 \\
\bottomrule
\end{tabular}
\end{table}
}
\section{Conclusions}
We studied some properties of a joint speech-text embedding space constructed through a semantic matching objective. The model was trained by minimizing the distance between the embedding outputs of a teacher RoBERTa text encoder and a student speech transformer encoder. We found that ASR pretraining of the speech encoder was essential for better semantic matching, measured in terms of retrieval accuracy. Performance could be further improved by combining it with a multitask objective incorporating both the joint embedding training and further ASR fine-tuning. Using the models for zero-shot classification was found to only work well if both speech and text embeddings manage to capture the properties of the classes. Probing showed that some but not all linguistic properties were transferred from the text to the speech model. In future work, we plan to investigate embeddings and training schemes that would enable semantic coupling between less constrained speech-text pairs compared to utterences and transcriptions, which may be applicable to a wider range of downstream tasks.
\clearpage
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
Stochastic programming (SP) is a mathematical framework to model the decision making in the presence of uncertainty \cite{birge2011introduction,shapiro2021lectures}. Two-stage SPs constitute a special class of this paradigm where partial decisions have to be made before the observation of the entire information, while the rest decisions are determined after the full information is revealed.
Most of the existing computational study of the continuous two-stage SPs is devoted to convex problems, especially linear problems \cite{ruszczynski1997decomposition,birge2011introduction, higle2013stochastic, shapiro2021lectures}. One of the most challenging features of the optimization models, nonconvexity, is usually abandoned and ignored.
However, there are many emerging applications in operations research and machine learning that call for complex nonlinear two-stage SP models and computational methods. Let us first introduce the mathematical formulation of such problems before discussing the applications.
The central optimization problem under consideration in this paper takes the following form:
\begin{equation}\label{eq:first_stage}
\displaystyle{
\operatornamewithlimits{\mbox{minimize}}_{x \, \in \, X}
} \;\, \zeta(x) \triangleq \varphi(x) + \mathbb{E}_{\tilde\xi}\left[ \, \psi(x;\tilde\xi) \, \right],
\end{equation}
where $\psi(x;\xi)$ is the second-stage recourse function that is given by:
\begin{equation}
\label{eq:second_stage}
\begin{array}{rl}
\psi(x;\xi) \, \triangleq \,
\displaystyle\operatornamewithlimits{minimum}_y \;\, & f(x,y;\xi)\\[0.1in]
\mbox{subject to} & G(x,y;\xi) \leq 0.
\end{array}
\end{equation}
In the above formulation, $X\subseteq \mathbb{R}^{n_1}$ is a compact and convex set, $\varphi:\mathbb{R}^{n_1} \to \mathbb{R}$ is a deterministic function that only depends on the first-stage decision $x$;
$\tilde\xi: \Omega \rightarrow \Xi$ is a random vector on a probability space $(\Omega, \mathcal F, \mathbb P)$ with $\Xi\subseteq\mathbb{R}^m$ being a measurable closed set; $\xi = \tilde\xi(\omega)$ for some $\omega \in \Omega$ represents a realization of the random vector $\tilde\xi$; and $f:\mathbb{R}^{n_1+n_2}\times \Xi\to\mathbb{R}$ and $G\triangleq (g_1, \ldots, g_\ell)^\top:\mathbb{R}^{n_1+n_2}\times \Xi\to\mathbb{R}^\ell$ are two Carath{\'e}odory functions (i.e., $f(\bullet,\bullet;\xi)$ and $G(\bullet,\bullet;\xi)$ are continuous for almost any $\xi\in\Xi$; $f(x,y;\bullet)$ and $G(x,y;\bullet)$ are measurable for any $(x,y) \in \mathbb{R}^{n_1 + n_2}$) that are jointly determined by the first-stage variable $x$ and the second-stage variable $y$.
Throughout this paper, we make the following assumptions.
\vskip 0.05in
\noindent
\begin{center}
\fbox{%
\parbox{0.98\textwidth}{
{\bf Assumption A}. The function $\varphi$ is convex on $X$ (for simplicity).\\[0.05in]
{\bf Assumption B}. For almost any $\xi\in \Xi$, the function $f(\bullet, \bullet; \xi)$ is concave-convex, i.e., $f(\bullet, y; \xi)$ is concave for any fixed $y$ and
$f(x,\bullet; \xi)$ is convex for any fixed $x$; and each $g_i(\bullet,\bullet; \xi)$ is jointly convex for $i=1, \ldots, \ell$.
}}
\end{center}
\vskip 0.05in
An example of $f(\bullet,\bullet;\xi)$ satisfying Assumption B is a bilinear function $x^\top D(\xi) y$ for some random matrix $D(\xi)\in \mathbb{R}^{n_1\times n_2}$.
The above settings notably extend the classical paradigm for continuous two-stage SPs \cite[Chapter 2.3]{shapiro2021lectures} in the following directions:\\[0.05in]
(i) The first-stage variable $x$ not only appears in the constraints of the second-stage problem, but also in the objective $f$. The recourse function $\psi(\bullet;\xi)$ is nonconvex since $f(\bullet,\bullet; \xi)$ is not jointly convex. This is fundamentally different from the
recent papers \cite{guigues2021inexact,guigues2021inexactmirror} that have assumed the joint convexity of $f(\bullet,\bullet;\xi)$ .\\[0.1in]
(ii) Both the objective function $f$ and the constraint map $G$ can be nonsmooth.\\[0.05in]
These two features together lead to {\sl a complex nonconvex and nonsmooth recourse function $\psi(\bullet;\xi)$} (see Figure \ref{fig:nonconvex} below), which constitutes the major challenge for designing rigorous and efficient numerical methods to solve problem \eqref{eq:first_stage}.
\begin{figure}[h]
\centering
\subfigure{
\begin{minipage}[t]{0.3\linewidth}
\centering
\includegraphics[width=1\textwidth]{Recourse_function1.png}
\end{minipage}
}
\qquad\qquad
\subfigure{
\begin{minipage}[t]{0.35\linewidth}
\centering
\includegraphics[width=1\textwidth]{Recourse_function2new.png}
\end{minipage}
}
\centering
\caption{The nonconvex nonsmooth recourse functions for fixed $\xi$'s. {\bf Left}: $x\in \mathbb{R}$; {\bf Right}: $x\in \mathbb{R}^2$.}
\label{fig:nonconvex}
\end{figure}
One important source of the nonconvex SP in \eqref{eq:first_stage} comes from the decision dependent/influenced uncertainty \cite{hellemo2018decision,nohadani2018optimization, jonsbraaten1998class, goel2004stochastic},
where the probability distribution $\mathbb{P}$ of the random vector $\tilde\xi$ is dependent on the first-stage variable $x$.
This is in contrast to the classical
SP paradigm under exogenous uncertainty, where the distribution of $\tilde\xi$ is not affected by the first-stage decisions.
There is growing interest in the endogenous uncertainty in recent literature of stochastic and robust programs \cite{goel2005lagrangean,goel2006class, tarhan2009stochastic,hellemo2018decision}.
A typical example where the random parameters can be altered by a decision is that the price (as a first-stage variable) may affect the distribution of the product demand.
Letting $p(x;\bullet)$ be the decision-dependent probability density function of the continuous random vector
$\tilde{\xi}$, we consider the corresponding two-stage SP model:
\[
\displaystyle{
\operatornamewithlimits{\mbox{minimize}}_{x \, \in \, X}
} \ \left\{\, \varphi(x) + \mathbb{E}_{\tilde\xi\sim \mathbb{P}(x)}[\psi(x;\tilde\xi)] = \varphi(x) + \displaystyle{
\int_{\Xi}
} \, \psi(x;\tilde\xi) \, p(x;\tilde\xi) \, \mbox{d}\tilde\xi\,\right\},
\]
where we adopt the push-in technique (see \cite{rubinstein1993discrete} and \cite[page 144]{pflug2012optimization}) to
transform it to the following formulation with decision-free uncertainty:
\[
\displaystyle{
\operatornamewithlimits{\mbox{minimize}}_{x \, \in \, X}
} \quad \varphi(x) + \displaystyle{
\int_{\Xi}
} \, \widehat{\psi}(x;\tilde\xi) \, \mbox{d}\tilde\xi, \;\, \mbox{where} \;\,
\widehat{\psi}(x;\tilde\xi) \, \triangleq \, p(x;\tilde\xi) \, \psi(x;\tilde\xi).
\]
However, this transformation makes the first-stage variable $x$ appear in the second-stage objective function. Even if the original recourse function $\psi(\bullet;\xi)$ is convex (such as in the two-stage linear SP model), the newly defined function $\widehat\psi(\bullet;\xi)$ is most likely nonconvex due to the multiplication of $\psi(\bullet;\xi)$ with the function $p(\bullet;\xi)$.
Another motivation to study \eqref{eq:first_stage} is from the stochastic interdiction problem \cite{cormican1998stochastic,hao2020piecewise}, where the defender may want to maximize the second-stage objective function instead of minimizing it. Even for the simple linear second-stage problem with only $x$ appearing in the constraints, the recourse function
\[
\begin{array}{rl}
\label{eq:second_stage defend}
\widetilde\psi(x;\xi) \, \triangleq \,
\displaystyle\operatornamewithlimits{maximum}_y &\;\, c(\xi)^\top y\\[0.05in]
\mbox{subject to} & \;\, T(\xi)x+W(\xi)y=h(\xi)
\end{array}
\]
is not convex in $x$. One may take the dual of the second-stage maximization problem so that the recourse is a parametrized minimization problem
\[
\begin{array}{rl}
\label{eq:second_stage reformulate}
\widetilde\psi(x;\xi) \, = \,
\displaystyle\operatornamewithlimits{minimum}_{\lambda} &\;\, \lambda^\top T(\xi) x - \lambda^\top h(\xi)\\[0.05in]
\mbox{subject to} & \;\, W(\xi)^\top \lambda + c(\xi) = 0.
\end{array}
\]
However, this dualization would bring a bilinear term $\lambda^\top T(\xi) x$ of the first-stage variable $x$ and the second-stage variable $\lambda$ to the objective function that necessities the concave-convex structure of $f(\bullet,\bullet;\xi)$ in Assumption B. To the best of our knowledge, there is no known rigorous decomposition method to solve a general nonconvex two-stage min-max stochastic programs even under the linear setting.
When the distribution of $\tilde\xi$ is taken as the empirical distribution of observed realizations $\xi^1,\ldots, \xi^S$, the simplest way to tackle the problem \eqref{eq:first_stage} is to simultaneously solve the first-stage variable $x$ and second-stage variables $y^1, \ldots, y^S$ (each $y^s$ is associated with one scenario $\xi^s$) via the sample average approximation \cite{shapiro2021lectures}:
\begin{equation}\label{determinstic}
\begin{array}{ll}
\displaystyle{
\operatornamewithlimits{\mbox{minimize}}_{x \, \in \, X, \;y^1, \ldots, y^S}
} & \;\, \varphi(x) + \displaystyle\frac{1}{S}\sum_{s=1}^S f(x,y^s;\xi^s)\\[0.15in]
\mbox{subject to} & \;\, G(x,y^s;\xi^s) \leq 0,\quad s=1, \ldots, S.
\end{array}
\end{equation}
However, this approach can be prohibitive when the number of scenarios $S$ is large since the dimension of the unknown variables is $n_1+n_2 S$. Even if $S$ is small or moderate, the above formulation may still be difficult to handle under our setting as the function $f(\bullet,\bullet;\xi)$ is not jointly convex (for example when $f(\bullet,\bullet;\xi)$ is bilinear). In fact, the nonconvexity of the recourse function also makes it challenging to apply the stochastic approximation method \cite{robbins1985stochastic,polyak1990new,nemirovski2009robust} to solve \eqref{eq:first_stage}, since it is not clear how to obtain a (Clarke) subdifferential of the nonconvex recourse function in \eqref{eq:first_stage}. Without strong assumptions like the uniqueness of the second-stage solutions, only a superset of the subdifferential $\partial \psi(\bullet;\xi)$ at given $x$ is computable \cite[Chapter 4]{bonnans2013perturbation}. When $f$ and $G$ are twice continuously differentiable, the authors in \cite{borges2020regularized} have adopted a smoothing method to deal with the possibly nonconvex recourse by adding the Tikhonov-regularized barrier of the inequality constraints to the second-stage objective function. For a special class of two-stage nonconvex quadratic SPs under the simplex constraint,
the paper \cite{bomze2022two} has derived upper and lower approximations of the objective values via copositive programs.
Notice that the constraints in \eqref{determinstic} are in fact block-wise separable in $y^1, \ldots, y^S$ so that there is a block-angular structure between the first- and second-stage variables.
Decomposition algorithms of two-stage SPs take advantage of this special structure to efficiently handle a large number of scenarios via solving $S$ numbers of low-dimensional subproblems \cite{ruszczynski1997decomposition}. Two classical decomposition algorithms for two-stage SPs are (augmented) Lagrangian decomposition and Benders decomposition.
{\it (Augmented) Lagrangian decompositions} (including the progressive hedging algorithm) copy the first-stage variable $S$ times and attach one to each scenario \cite{guignard2003lagrangean,rockafellar1991scenarios}. In order to force the non-anticipativity of the first-stage decision, one has to add equality constraints among all copies to ensure that $x$ is the same across different realizations of the uncertainty. However, there are two major bottlenecks to apply such kind of dual-based algorithms to solve the problem \eqref{eq:first_stage}. One, each subproblem pertaining to one pair of variables $(x^s, y^s)$ is still nonconvex if $f$ is not a jointly convex function, so that it is in general not easy to obtain its global optimal solution. Two, the convergence of these dual approaches is largely restricted to the convex problems or special integer problems \cite{rockafellar1991scenarios,chun1995scenario}. Although there are some recent advances for the convergence study of the progressive hedging algorithm for solving nonconvex SPs under the local convexity conditions \cite{rockafellar2019progressive,rockafellar2020augmented}, it is not clear whether the problem \eqref{eq:first_stage} satisfies those conditions without further assumptions on $f$ and $g$.
{\it Benders decomposition} (or L-shaped methods) \cite{benders1962partitioning,van1969shaped, wets1984large} alternatively updates the first-stage and second-stage variables, where
the second-stage subproblem can be solved in parallel to save the computational time and reduce storage burden.
In order to derive
valid inequalities of $x$ and add them to the first-stage master problem, one usually uses subgradient inequalities of the (convex) recourse function to generate a sequence of lower approximations. However, when the recourse function is associated with the complex nonconvex function in \eqref{eq:first_stage}, it is challenging to derive its lower bounds based on the computed second-stage solutions. In fact, for the recourse functions in Figure \ref{fig:nonconvex}, there seems not to exist a convex function that passes one of the downward cusps and at the same time approximates the original function from below.
In this paper, we tackle the two-stage stochastic programs \eqref{eq:first_stage} by a novel lifting technique that transforms the complex nonconvex and nondifferentiable recourse function \eqref{eq:second_stage} in the original space to a structured convex-concave function in a lifted space. The reveal of this latent structure enables us to construct convex surrogation of all recourse functions at the latest first-stage iterate, whose evaluations are decomposable across different scenarios. Such surrogate functions are then added to the master problem to generate the next first-stage iterate. We shall prove that repeating the above procedure, the sequence of the first-stage iterates converges to a properly defined stationary solution of \eqref{eq:first_stage}.
In order to further reduce the computational cost per step when the number of scenarios $S$ is large as well as to handle the case where $\tilde\xi$ is continuously distributed, we also propose a framework that incorporates sequential sampling into the surrogation algorithm.
The sequential sampling method gradually adds scenarios and generates cuts along the iterations, which has the advantage that one may obtain satisfactory descent progress in the early iterations with relatively small sample sizes to accelerate the overall procedure.
The paper is organized as follows. Section \ref{sec:pre} provides preliminary knowledge of the set-valued mappings. In Section \ref{sec:approx}, we discuss the implicitly convex-concave structure of the recourse function and derive its computationally tractable approximations. A decomposition algorithm for solving problem \eqref{eq:first_stage} with a fixed number of scenarios is proposed and analyzed in Section \ref{sec:algorithm}. To further handle the continuously distributed random vectors as well as to reduce the computational cost of the decomposition algorithm in its early stage, we provide an internal sampling version of the algorithm in Section \ref{sec:sampling} and show the almost surely convergence of the iterative sequence. In Section \ref{sec:exp}, we conduct extensive numerical experiments to show the effectiveness of our proposed frameworks. The paper ends with a concluding section.
\section{Preliminaries}
\label{sec:pre}
We first summarize the notation used throughout the paper. We write $\mathbb{Z}_+$ as the set of all nonnegative integers, and $\mathbb{R}^n$ as the $n$-dimensional Euclidean space equipped with the inner product $\langle x,y\rangle = x^\top y$ and the induced norm $\|x\| \triangleq \sqrt{x^\top x}$.
The symbol $\mathbb{B}(x, \delta)$ is used to denote the closed ball of radius $\delta>0$ centered at a vector $x \in \mathbb{R}^n$. Given a set $A\subseteq \mathbb{R}^n$, its diameter is defined as $R(A)\triangleq\sup\limits_{x,y\in A}\| x - y \|$, and the distance from a vector $x\in \mathbb{R}^n$ to $A$ is defined as $\operatorname{dist}(x, A) \triangleq \inf\limits_{y \in A} \| y - x \|$. The one-sided deviation of a set $A\subseteq \mathbb{R}^n$ from a set $C\subseteq \mathbb{R}^n$ and the two-sided Hausdorff distance between $A$ and $C$ are defined respectively as
\[
\mathbb{D}(A, C) \, \triangleq \, \sup\limits_{x \in A} \operatorname{dist}(x, C) \quad \mbox{and} \quad \mathbb H(A, C) \,\triangleq \, \max\{\, \mathbb{D}(A, C), \, \mathbb{D}(C, A) \}.
\]
We next introduce the concepts of generalized derivatives and subdifferentials for nonsmooth functions.
Interested readers are referred to the monographs \cite{clarke1990optimization, rockafellar2009variational, mordukhovich2006variational} for thorough discussions on these subjects. Consider a function $f: {\cal O} \rightarrow \mathbb{R}$ defined on an open set ${\cal O}\subseteq \mathbb{R}^n$. The classical one-sided directional derivative of $f$ at $\bar x\in {\cal O}$ along the direction $d\in\mathbb{R}^n$ is defined as
\[
f^\prime(\bar x; d) \, \triangleq \, \lim\limits_{t \downarrow 0} \, \frac {f(\bar x + t d) - f(\bar x)} {t}
\]
if this limit exists. The function $f$ is said to be directionally differentiable at $\bar x\in {\cal O}$ if it is directionally differentiable along any direction $d\in\mathbb{R}^n$. In contrast,
the Clarke directional derivative of $f$ at $\bar x\in {\cal O}$ along the direction $d\in\mathbb{R}^n$ is defined as
\[
f^\circ(\bar x; d) \, \triangleq \, \displaystyle\operatornamewithlimits{limsup}_{x\to \bar{x},\, t\downarrow 0} \, \frac {f(x + t d) - f(x)} {t},
\]
which is finite when $f$ is Lipschitz continuous near $\bar x$.
The Clarke subdifferential of $f$ at $\bar x$ is the set $\partial_C f(\bar x) \triangleq\{ v\in \mathbb{R}^n \mid f^\circ(\bar x;d) \geq v^\top d \text{ for all $d \in \mathbb{R}^n$} \}$, which coincides with the usual subdifferential in convex analysis for a convex function. If $f$ is strictly differentiable at $\bar{x}$, then $\partial_C f(\bar{x}) = \{\nabla f(\bar{x})\}$.
We say that $f$ is Clarke regular at $\bar x\in {\cal O}$ if $f$ is directionally differentiable at $\bar x$ and $f^\circ(\bar x;d)=f^\prime(\bar x;d)$ for all $d\in\mathbb{R}^n$. This Clarke regularity at $\bar{x}$ is equivalent to have
$
f(x) \geq f(\bar{x}) + \bar{v}^\top (x-\bar{x}) + o(\|x-\bar{x}\|)
$ for any $\bar{v}\in \partial_C f(\bar{x})$. Therefore, if a function fails to satisfy the Clarke regularity at $\bar{x}$ (for example at the downward cusp on the left panel of Figure \ref{fig:nonconvex}), there does not exist an approximate linear lower bound of the original function based on the Clarke subdifferentials with small $o$ error locally.
Let $X\subseteq \mathbb{R}^n$ be a closed convex set and $f:\mathbb{R}^n \to \mathbb{R}$ be a locally Lipschitz continuous function that is directionally differentiable.
We say $\bar x\in X$ is a
directional-stationary point of $f$ on $X$ if $f^\prime (\bar x; x-\bar x) \geq 0$ for all $x \in X$, and a
Clarke-stationary point if $f^\circ(\bar x; x-\bar x) \geq 0$ for all $x\in X$; the latter is equivalent to
$0 \in \partial_C f(\bar x)+\mathcal N_{X}(\bar x)$ with $\mathcal N_{X}(\bar x)$ being the normal cone of $X$.
Let $\mathcal F: \mathbb{R}^n \, {\rightrightarrows} \, \mathbb{R}^m$ be a set-valued mapping.
Its outer limit at $x\in \mathbb{R}^n$ is defined as
\[
\limsup\limits_{x \rightarrow \bar x} \mathcal F(x)
\triangleq \bigcup\limits_{x^\nu \rightarrow \bar x} \limsup\limits_{\nu \rightarrow \infty} \mathcal F(x^\nu)
= \big\{ u \mid \exists \; x^\nu \rightarrow \bar x, \exists \; u^\nu \rightarrow u \text{ with } u^\nu \in \mathcal F(x^\nu) \big\}.
\]
We say $\mathcal F$ is outer semicontinuous at $\bar x\in \mathbb{R}^n$ if $\limsup\limits_{x \rightarrow \bar x} \mathcal F(x) \subseteq \mathcal F(\bar x)$.
It is known that
the subdifferential mapping $\partial f$ of a proper closed convex function $f:\mathbb{R}^n\rightarrow\mathbb{R}\cup\{+\infty\}$ is outer semicontinuous; in particular, the normal cone of a closed convex set is outer semicontinuous \cite[Propositions 6.6 and 8.7]{rockafellar2009variational}.
\section{The implicit convexity-concavity of the recourse functions}
\label{sec:approx}
A key ingredient to design a decomposition method for solving the two-stage SP \eqref{eq:first_stage} is to derive a computational-friendly approximation of the nonconvex recourse function \eqref{eq:second_stage} at any given $x\in X$ and $\xi \in \Xi$. This is the main content of the present section.
Consider any fixed $\xi \in \Xi$ where Assumption B holds.
For simplicity, we omit $\xi$ in \eqref{eq:second_stage} throughout this section and rewrite the recourse function as, for $x\in \mathbb{R}^{n_1}$,
\begin{equation}\label{eq:value_function}
\begin{split}
\psi(x) \, \triangleq \,
\displaystyle\operatornamewithlimits{minimum}_{y} &\;\; f(x,y)\\
\mbox{subject to} & \;\; G(x,y)\leq 0.
\end{split}
\end{equation}
In the following, we show that the above function, although generally being nonconvex and nondifferentiable in $\mathbb{R}^{n_1}$, has a benign structure in a lifted space. Leveraging this structure, we then derive an approximate difference-of-convex decomposition of the recourse function that is computationally tractable. Such an approximation is the cornerstone of the decomposition methods to be presented in the next two sections.
\subsection{The implicit convexity-concavity of $\psi$}
As mentioned in the first section, the difficulty to design a decomposition method for solving \eqref{eq:first_stage} is due to the lack of a valid inequality of the recourse function, which is partially because
$x$ appears in both the objective and constraints of the parametric problem in \eqref{eq:value_function}. However, if either $x$ in the objective or in the constraints is fixed, the resulting functions are relatively easy to analyze. Specifically, for any fixed $\bar{x}\in \mathbb{R}^{n_1}$, consider the functions
\[
\psi_{\rm cvx}(x)
\triangleq
\left[ \begin{array}{rl}
\displaystyle\operatornamewithlimits{minimum}_{y} & f(\bar{x},y)\\[0.05in]
\mbox{subject to} & G(x,y)\leq 0
\end{array}\hspace{-0.05in} \right]\hspace{-0.05in}\;\;\;\mbox{and}\;\;\;
\psi_{\rm cve}(x)
\triangleq
\left[ \begin{array}{rl} \displaystyle\operatornamewithlimits{minimum}_{y} & f(x,y)\\[0.05in]
\mbox{subject to} & G(\bar{x},y)\leq 0
\end{array}\hspace{-0.05in}\right].
\]
The following structural properties of $\psi_{\rm cvx}$ and $\psi_{\rm cve}$ can be easily derived. We include the proof here for completeness.
\begin{lemma}\label{lemma:icc}
Let $\bar{x}\in \mathbb{R}^{n_1}$ be fixed.
Under Assumption B (for a fixed $\xi\in \Xi$), the function $\psi_{\rm cvx}$ is convex and the function $\psi_{\rm cve}$ is concave in $\mathbb{R}^{n_1}$.
\end{lemma}
\begin{proof}
Consider any $\lambda\in[0,1]$ and any $x^1,x^2\in \mathbb{R}^{n_1}$. We first show that
\begin{equation}\label{eq:proof lemma}
\psi_{\rm cvx} \big(\lambda x^1 + (1-\lambda) x^2\big) \leq \lambda \, \psi_{\rm cvx} (x^1) + (1-\lambda) \,\psi_{\rm cvx} (x^2).
\end{equation}
Let $\bar{y}^1$ and $\bar{y}^2$ be one of the optimal solutions of the optimization problems in defining $\psi_{\rm cvx}(x^1)$ and $\psi_{\rm cvx}(x^2)$ respectively. We assume such solutions exist since otherwise the right side of \eqref{eq:proof lemma} is $+\infty$ and the inequality holds trivially.
The joint convexity of each $g_j$ for $j=1, \cdots, \ell$ implies that
\[
G\big(\lambda x^1+(1-\lambda)x^2, \, \lambda \bar{y}^1+(1-\lambda)\bar{y}^2\big) \leq 0, \quad \forall \, j=1,\ldots,\ell,
\]
which yields the feasibility of $\lambda \bar{y}^1+(1-\lambda)\bar{y}^2$ for the minimization problem in \eqref{eq:lifting} when $x$ takes the value $\lambda x^1+(1-\lambda)x^2$. We can thus deduce that $\psi_{\rm cvx} \big(\lambda x^1 + (1-\lambda) x^2\big) \leq f \big(\bar{x}, \lambda \bar{y}^1 + (1-\lambda) \bar{y}^2\big)$. The convexity of $f(\bar{x},\bullet)$ further yields that
\begin{gather*}
f \big(\bar{x}, \lambda \bar{y}^1 + (1-\lambda) \bar{y}^2\big)
\leq \lambda f (\bar{x}, \bar{y}^1) + (1-\lambda) f (\bar{x}, \bar{y}^2) = \lambda \psi_{\rm cvx} (x^1) + (1-\lambda) \psi_{\rm cvx} (x^2),
\end{gather*}
which proves the inequality \eqref{eq:proof lemma}.
The concavity of $\psi_{\rm cve}$ is due to the fact that it is the minimum of concave functions $\{f(\bullet,y)\}$ for $y$ feasible to the problem in \eqref{eq:lifting}.
\end{proof}
Lemma \ref{lemma:icc} suggests that the recourse function \eqref{eq:value_function} has a hidden convex-concave structure. Indeed, such a function $\psi$ belongs to a special class of nonconvex functions coined {\sl implicitly convex-concave} (icc) functions that are formally
defined below.
\begin{definition}[{\cite[Definition 4.4.4]{cui2020modern}}]\label{definition:icc}
A function $\theta:\mathbb{R}^{n} \rightarrow \mathbb{R}\cup\{+\infty\}$ is said to be implicitly convex-concave if there exists a bivariate function $\overline\theta: \mathbb{R}^{n} \times \mathbb{R}^{n} \rightarrow \mathbb{R}\cup\{+\infty\}$ such that $\overline\theta(\bullet,z)$ is convex for any fixed $z\in \mathbb{R}^n$, $\overline\theta(x,\bullet)$ is concave for any fixed $x\in \mathbb{R}^n$, and $\theta(x) = \overline\theta(x,x)$.
\end{definition}
Th above concept is firstly introduced in \cite{liu2020two} to analyze the convergence property of a difference-of-convex algorithm to solve two-stage convex bi-parametric quadratic SPs. More properties of icc functions are studied in the recent monograph \cite{cui2020modern}. In fact, the term ``icc'' suggests that this class of functions is a generalization of the difference-of-convex (dc) functions, as the latter is ``explicitly convex-concave'', i.e., for any dc function $\theta(x) = \theta_1(x) - \theta_2(x)$ with both $\theta_1$ and $\theta_2$ convex, one can always associate it with the bivariate function $\overline{\theta}(x,y) = \theta_1(x) - \theta_2(y)$ to explicitly expose the convexity-concavity of $\theta$ in the lifted pair $(x,y)$.
Back to the recourse function \eqref{eq:value_function}, we consider its lifted bivariate counterpart in $\mathbb{R}^{n_1 + n_1}$:
\begin{equation}\label{eq:lifting}
\begin{split}
\overline\psi(x, z)
\, \triangleq \,
\displaystyle\operatornamewithlimits{minimum}_{y} & \;\; f(z,y)\\%[0.05in]
\mbox{subject to} & \;\; G(x,y)\leq 0.
\end{split}
\end{equation}
The following result is a direct consequence of Lemma \ref{lemma:icc}. No proof is needed.
\begin{proposition}\label{prop:icc}
Under Assumption B, the function $\psi$ in \eqref{eq:value_function} is an implicitly convex-concave function associated with the lifted function $\overline\psi$ in \eqref{eq:lifting}.
\end{proposition}
\begin{figure}
\centering
\includegraphics[scale=0.4]{Lifting_value_functions.png}
\centering
\caption{\rm \scriptsize An illustration of the implicitly convex-concave structure of the nonconvex recourse function.
{\bf Left}: the original 1-dimensional recourse function that is neither convex nor concave;
{\bf Middle}: the lifted bivariate counterpart in $\mathbb{R}^2$ that is convex in $x$ and concave in $z$; {\bf Right}: the exposure of the concave component in the $z$ coordinate and the convex component in the $x$ coordinate.}
\label{fig:lifting}
\end{figure}
One shall see from the subsequent sections that the derived implicitly convex-concave property of the recourse function is critical to study the two-stage SP \eqref{eq:first_stage}. On one hand, we can leverage this structure to construct an approximation of the nonconvex recourse function $\psi$ at any given $x$; on the other hand,
it enables us to define a stationary point of \eqref{eq:first_stage} that is provably computable by our later designed algorithms. To fulfill these tasks, we first derive a superset of the Clarke subdifferential of $\psi$.
To proceed, we denote $\partial_1 \overline\psi(x,z)$ as the subdifferential of the convex function $\overline\psi(\bullet,z)$ at $x$, and $\partial_2 (-\overline\psi)(x,z)$ as the subdifferential of the convex function $(-\overline\psi)(x, \bullet)$ at $z$. We also write $\overline{Y}(x,z)$ as the set of all optimal solutions of problem \eqref{eq:lifting}.
\begin{lemma}\label{lem:subdiff}
Assume that the objective function $f(\bullet,y)$ in \eqref{eq:value_function} is locally Lipschitz continuous at $x$ whose Lipschitz constant is independent of $y$; and
$\partial_1(-f)(x,y)$ is outer semicontinuous jointly in $(x,y)$.
The following two statements hold at all $(x,z)$ that satisfies $\overline{Y}(x,z)\neq \emptyset$.\\
(a) $\partial_2(-\overline\psi)(x,z) = \operatorname{conv}\Big\{ \partial_1 (-f)(z, y) \mid y\in \overline Y(x,z) \Big\}$, where $\operatorname{conv} S$ denotes the convex hull of a set $S$;\\
(b) $\partial_C \psi (x)\,\subseteq\, \partial_1 \;\overline\psi (x, x) - \partial_2 (-\overline\psi) (x, x)$.
\end{lemma}
\begin{proof}
Since $(-f)(\bullet,y)$ is convex for any fixed $y$, we have $(-f)^\circ(\bullet,y)(x;d) = (-f)^\prime(\bullet,y)(x;d)$ for any $x$ and any $d$. The statement (a) can then be obtained from \cite[Theorem 2.1]{clarke1975generalized}. Part (b) is a consequence of \cite[Proposition 4.4.26 (c)]{cui2020modern} on the relationship between the subdifferentials of an implicitly convex-concave function and its lifted counterpart.
\end{proof}
\subsection{The partial Moreau envelope}
Equipped with the lifted function $\overline\psi$, one may be able to construct computationally friendly surrogations of the recourse function in \eqref{eq:value_function} via a modification of the usual Moreau envelope. Let us first recall the definition of the classical Moreau envelope. Given a proper lower
semicontinuous function $\theta:\mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$ and a positive scalar $\gamma$, its {\sl Moreau envelope} is
\[
e_\gamma^{\rm ori} \,\theta(x) \, \triangleq \, \inf\limits_{y} \left\{ \theta(y) + \frac 1 {2\gamma} \|x-y\|^2\right\}, \quad x\in \mathbb{R}^n.
\]
We use the superscript ``ori'' to emphasize that this is the original definition of the Moreau envelope and is different from our later modification.
The function $f$ is said to be prox-bounded if there exists $\gamma>0$ such that $e_\gamma^{\rm ori}\, \theta(x) > -\infty$ for some $x\in \mathbb{R}^n$. It is known that a convex function is always prox-bounded and its Moreau-envelope is continuously differentiable (c.f.\ \cite[Theorem 2.26]{rockafellar2009variational}). In general, for any prox-bounded function $\theta$, the parametric functions
$e_\gamma^{\rm ori} \,\theta(x) \uparrow \theta(x)$ as $\gamma\downarrow 0$ for all $x\in \mathbb{R}^n$.
Therefore, one can view the Moreau envelope as a lower approximation of the original function.
However, if $\theta$ is nonconvex and nonsmooth, the function $e_\gamma^{\rm ori} \,\theta$ may be neither convex nor smooth. Nevertheless, for any $x\in \mathbb{R}^n$, it holds that (see, e.g., \cite{asplund1973differentiability,lucet2006fast})
\begin{equation}\label{eq:dc of moreau}
e_\gamma^{\rm ori}\, \theta(x) \, = \frac{1}{2\gamma}\|x\|^2 - \underbrace{\sup_{y} \left\{ -\theta(y) - \frac{1}{2\gamma}\|y\|^2 + \frac{1}{\gamma}y^\top x \right\}}_{\text{convex in $x$ even if $f$ is nonconvex}},
\end{equation}
which indicates that one can always obtain a dc decomposition of $e_\gamma^{\rm ori} \,\theta$ no matter $\theta$ is convex or not. The only trouble brought by the nonconvexity of $\theta$ is that the inner sup problem for $y$ may not be concave (especially if $\theta$ is not weakly convex), thus one may not be able to evaluate the subgradient of the second term at a given $x$ when using the dc algorithm to minimize the function $e_\gamma^{\rm ori} \,\theta$. Specifically, in the context of the recourse function \eqref{eq:value_function}, its associated Moreau envelope is
\[
\begin{array}{rl}
e_\gamma^{\rm ori}\, \psi (x)
\; = & \displaystyle\operatornamewithlimits{inf}_{z} \;\, \left\{\psi(z) + \frac 1 {2\gamma} \;\|x-z\|^2 \right\}\\[0.1in]
= &\displaystyle\frac{1}{2\gamma}\|x\|^2 - \left[
\begin{array}{ll}
\displaystyle\operatornamewithlimits{maximum}_{z,y}& -f(z, y) - \displaystyle\frac 1 {2\gamma} \|z\|^2 + \frac{1}{\gamma}z^\top y\\[0.12in]
\mbox{\rm subject to} & G(z,y)\leq 0
\end{array}
\right],
\end{array}
\]
where the inner max problem is not jointly concave in $(z,y)$ since $f$ is not assumed to be jointly convex. This issue motivates us to introduce the following new type of envelopes tailored to icc functions that is more computationally tractable:
\begin{equation}\label{eq:moreau}
e_\gamma \theta (z)
\triangleq \; \displaystyle\operatornamewithlimits{inf}_{x} \;\, \left\{\,\overline\theta(x,z) + \frac 1 {2\gamma} \;\|x-z\|^2 \,\right\},
\end{equation}
where $\theta:\mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$ is any icc function and $\overline\theta$ is its lifted counterpart as in Definition \ref{definition:icc}. When $\overline{\theta}(x,z)$ is independent of $z$ (so that this function only has the convex part), the above definition reduces to the usual Moreau envelope. Hence, we term the new regularization of $\theta$ in \eqref{eq:moreau} its {\sl partial Moreau envelope}. Similarly as in \eqref{eq:dc of moreau}, the newly defined partial Moreau has the following explicit dc decomposition
\begin{equation}\label{eq: decompose partial}
e_\gamma \theta (z) \,= \underbrace{\frac 1 {2\gamma} \|z\|^2}_{\text{strongly convex}} - \underbrace{\sup\limits_{x}\left\{ -\overline\theta(x, z) - \frac 1 {2\gamma}\|x\|^2 + \frac 1 \gamma z^\top x \right\}}_{\text{denoted as } g_\gamma(z),\; \text{convex}}.
\end{equation}
The optimal solution of the minimization problem in \eqref{eq:moreau} at a given $z\in \mathbb{R}^n$, which we follow the terminology in the literature
to call it the proximal point of $\overline\theta$ at $z$, is defined as
\[
P_{\gamma} \theta(z) \, \triangleq \, \underset{x}{\operatorname{argmin}}\left\{ \overline\theta(x,z) +\frac 1 {2\gamma} \|x-z\|^2\right\}.
\]
This mapping is always well defined and single-valued since the inner objective function is strongly convex in $x$.
Similar to the classical Moreau envelope, the partial Moreau envelope approximates the original function from below.
The following lemma provides the gap between the partial Moreau envelope and the original function under the uniform Lipschitz continuity of $\overline\theta(\bullet,z)$. The proof is adapted from {\cite[Proposition 3.4]{planiden2019proximal}} on a similar property regarding the Moreau envelope.
\begin{figure}
\centering
\includegraphics[scale=0.13]{PME_surrogation.png}
\caption{\rm \scriptsize The original recourse function, its partial Moreau envelopes and their majorizations with different $\gamma$'s. {\bf Left}: $\gamma=0.15$; {\bf Right}: $\gamma=0.02$.}
\vskip -0.3in
\label{fig:approx}
\end{figure}
\begin{lemma}\label{lem:gap_bound}
Consider an icc function $\theta: \mathbb{R}^n \to \mathbb{R}\cup\{+\infty\}$ and its lifted counterpart $\overline\theta :\mathbb{R}^n\times \mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$. Let $X\subseteq \operatorname{dom} \theta$.
For any $z\in X$, if the lifted recourse function $\overline\theta(\bullet,z)$ is Lipschitz continuous in $x$ for $z\in X$ with the Lipschitz constant $\kappa(z)$, i.e.,
\[
\left|\overline\theta(x_1,z)-\overline\theta(x_2,z)\right| \leq \kappa(z) \, \|x_1-x_2\|, \quad \forall \, x_1, x_2\in \mathbb{R}^n,\; z\in X,
\]
then $0\leq \theta(z) - e_{\gamma}\theta(z) \;\leq\; \gamma\kappa(z)^2/2$ for any $z\in X$.
\end{lemma}
\begin{proof}
For any $z\in X$, it holds that
\begin{align*}
0\leq\,\theta(z) - e_{\gamma}\theta(z)
&=\,\overline\theta(z,z) - \overline\theta\left(P_{\gamma}\theta(z),z\right) - \left\| P_{\gamma}\theta(z) - z \right\|^2/(2\gamma)\\
&\leq\,\kappa(z) \left\|P_{\gamma}\theta(z) - z\right\| - \left\| P_{\gamma}\theta(z) - z \right\|^2/(2\gamma)
\,\leq\,\gamma \kappa(z)^2/2,
\end{align*}
where the second inequality follows from the Lipschitz continuity of $\overline\theta(\bullet,z)$ and the last inequality uses the fact that $\max\limits_{t\ge 0}\left[\, \kappa(z) \,t- t^2/{(2\gamma)} \,\right] = \gamma \kappa(z)^2/2$.
\end{proof}
With $g_\gamma$ defined in \eqref{eq: decompose partial}, it can be shown that
(c.f. {\cite[Section 7.2.1]{cui2020modern}})
for any $z\in \mathbb{R}^n$,
\[
\partial g_\gamma(z)=\frac 1 \gamma P_{\gamma} \theta(z)+\partial_2(-\overline\theta)(P_{\gamma} \theta(z),z).
\]
One can then obtain the following convex majorization of $e_\gamma\theta(z)$ at any given point $z^\prime \in \mathbb{R}^n$ based on the subgradient inequality of the convex function $g_\gamma$:
\begin{equation}\label{eq:upper bound general}
e_\gamma\theta(z)\, \leq \, \widehat{e}_{\gamma}\theta(z;z^\prime)
\triangleq \frac 1 {2\gamma} \| z\|^2 -g_\gamma(z^\prime) -\big( P_{\gamma }\theta(z^\prime)/\gamma + c\big)^\top (z - z^\prime),\;\, \forall \;z\in \mathbb{R}^n,
\end{equation}
where $c\in \partial_2(-\overline\theta)\left(P_{\gamma }\theta(z^\prime),z^\prime\right)$. Figure \ref{fig:approx} shows the icc recourse function, its partial Moreau envelopes and their convex majorizations with two different values of $\gamma$'s.
\section{The decomposition algorithm and its convergence}
\label{sec:algorithm}
Based on the discussion in the last section, we are now ready to present the decomposition algorithm for solving the nonconvex two-stage SP \eqref{eq:first_stage} and analyze its convergence. In this section, we focus on the case where there are fixed scenarios $\{\xi^1,\ldots,\xi^S\}$, each realized with probability $1/S$. The
problem \eqref{eq:first_stage} then reduces to
\begin{equation}\label{eq:finite_scenarios}
\begin{array}{rl}
\displaystyle\operatornamewithlimits{minimize}_{x\in X\subseteq\mathbb{R}^{n_1}} \;\, \left\{\varphi(x)+\frac{1}{S}\sum\limits_{s=1}^S \psi(x; \xi^s)\right\},
\end{array}
\end{equation}
where each $\psi(x;\xi^s)$ is given by \eqref{eq:second_stage}. The above problem can be viewed as a sample average approximation of the two-stage SP \eqref{eq:first_stage}
under a prescribed sample size $S$. All the discussions in this section can be easily adapted to the case where the distribution of $\tilde\xi$ has finite support (but unequal probability mass for different $\xi^s$).
We will work on the internal sampling scheme for continuously distributed $\tilde\xi$ in the next section.
\subsection{The algorithmic framework}
Our goal is to
solve the nonconvex problem \eqref{eq:finite_scenarios} via a successive approximation scheme.
For any $\xi\in \Xi$, the partial Moreau envelope of the recourse function \eqref{eq:second_stage} associated with the bivariate function \eqref{eq:lifting} is
\begin{equation}\label{eq:subprob}
\begin{split}
e_{\gamma}\psi(z;\xi)\,
& \triangleq \; \operatornamewithlimits{minimum}_{x\in {\mathbb{R}}^{n_1}} \;\, \left\{\,\overline\psi(x, z;\xi) + \frac 1 {2\gamma} \|x-z\|^2\, \right\}\\
&=
\left[
\begin{array}{ll}
\displaystyle\operatornamewithlimits{minimum}_{x\in \mathbb{R}^{n_1}, \, y\in \mathbb{R}^{n_2}} &\;\, f(z, y; \xi) + \displaystyle\frac 1 {2\gamma} \|x - z\|^2\\[0.1in]
\mbox{subject to} &\;\, G(x,y;\xi) \leq 0
\end{array}
\right].
\end{split}
\end{equation}
We consider a double-loop algorithm where the outer loop updates the parameter $\gamma$ in the partial Moreau envelope and the inner loop solves the nonconvex problem $\displaystyle\operatornamewithlimits{minimize}_{x\in X}\, \left[\, \varphi(x) + \frac{1}{S} \sum_{s=1}^S e_{\gamma_\nu}\psi(x;\xi^s)\,\right]$ to stationarity for a fixed $\gamma_\nu$.
To solve the latter nonconvex problem during the $\nu$-th inner loop, we replace the Moreau-regularized recourse function $e_{\gamma_\nu}\psi(x;\xi)$ with its upper approximation constructed at the latest iterate $x_{\nu,i}$, where $i$ denotes the inner iterate index. For each $s$, let $\left(\,x^s_{\nu,i}, \, y^s_{\nu,i}\,\right)$ be one of the optimal solutions of \eqref{eq:subprob} at $(z,\xi) = (x_{\nu,i}, \xi^s)$, which can be computed by solving $s$ separable convex optimization problems (the convexity of this problem in $(x,y)$ can be seen from Assumption B in Section \ref{sec:introduction}). Notice that $x_{\nu,i}^s = P_{\gamma_\nu} \psi(x_{\nu,i};\xi^s)$.
One may then derive from \eqref{eq:upper bound general} the following
upper approximating function of $e_{\gamma_\nu}\psi(x;\xi^s)$:
\begin{equation}\label{eq:Moreau_surrogate}
\widehat{e}_{\gamma_{\nu}}\psi(x; \xi^{s}; x_{\nu,i})\triangleq \frac 1 {2\gamma_{\nu}} \|x\|^2 - g_{\gamma_{\nu}}(x_{\nu,i}; \xi^{s}) - \big( x^s_{\nu,i}/{\gamma_{\nu}} + c^s_{\nu,i}\big)^\top (x-x_{\nu,i}),
\end{equation}
where $
g_{\gamma_{\nu}}(x_{\nu,i}; \xi^{s}) \triangleq \displaystyle\frac 1 {2\gamma_{\nu}} \|x_{\nu,i}\|^2 - e_{\gamma_{\nu}}\psi(x_{\nu,i};\xi^{s})
$ and $c^s_{\nu,i} \in \partial_2\left(-\overline\psi\,\right)\left(x^s_{\nu,i}, x_{\nu,i}; \xi^{s}\right)$. Due to Lemma \ref{lem:subdiff}, under Assumption C to be imposed later, a particular way to choose $c^s_{\nu,i}$ is to take an element from $\partial_1 (-f)\left(x^s_{\nu,i}, y^s_{\nu,i}; \xi^{s}\right)$.
The resulting master problem to generate the next first-stage iterate $x_{\nu,i+1}$ is
\begin{equation}\label{eq:master}
\operatornamewithlimits{minimize}_{x\in X}\;\,
\left\{
\varphi(x)+
\frac{1}{S} \sum\limits_{s=1}^{S} \widehat{e}_{\gamma_{\nu}}\psi(x; \xi^{s}; x_{\nu,i})
\right\}.
\end{equation}
The inner iteration continues until the distance between two consecutive iterates $x_{\nu,i}$ and $x_{\nu,i+1}$ is sufficiently close.
We summarize the procedure of the decomposition algorithm below. When $S=1$, it reduces to the algorithm in \cite[Algorithm 7.2.1]{cui2020modern} to minimize an implicitly convex-concave function (without decomposition).
\begin{algorithm}[H]
\caption{A decomposition algorithm for the nonconvex two-stage SP \eqref{eq:finite_scenarios}}
\label{alg:finite_samples}
{\bf Input:} Initial point $x_0\in X$, and two scalar sequences $\{\gamma_{\nu}\}_{{\nu}\geq 0}\downarrow 0$ and $\{\varepsilon_\nu\}_{\nu\geq 0} \downarrow 0$.
\vskip 0.05in
{\bf Outer loop:} Set $\nu=0$.
\begin{algorithmic}[1]
\State Execute the inner loop with the initial point $x_\nu$, and parameters $\gamma_\nu$ and $\varepsilon_\nu$.
\State Set $\nu \leftarrow \nu+1$ and repeat step 1 until a prescribed stopping criterion is satisfied.
\end{algorithmic}
\vskip 0.02in
{\bf Inner loop:} Set $i=0$ and $x_{\nu,0}=x_\nu$.
\begin{algorithmic}[1]
\State Solve the subproblem \eqref{eq:subprob} at $(z,\xi) = (x_{\nu,i},\xi^s)$ for all $s$ to get solutions $\left(x^s_{\nu,i}, y^s_{\nu,i}\right)$.
\State Select $c^s_{\nu,i} \in \partial_1 (-f)\left(x^s_{\nu,i}, y^s_{\nu,i}; \xi^{s}\right) \subseteq \partial_2\left(-\overline\psi\,\right)\left(x^s_{\nu,i}, x_{\nu,i}; \xi^{s}\right)$ for each $s$.
\State Solve the master problem \eqref{eq:master} to obtain $x_{\nu,i+1}$.
\State Set $i \leftarrow i+1$ and repeat the above steps if $\|x_{\nu,i+1} - x_{\nu,i}\| > {\varepsilon_\nu}\gamma_\nu$. Otherwise, break the inner loop with $x_{\nu + 1} \triangleq x_{\nu,i+1}$.
\end{algorithmic}
\end{algorithm}
Notice that each $\widehat{e}_{\gamma_{\nu}}\psi(x; \xi^{s}; x_{\nu,i})$ is a strongly convex quadratic function in $x$. Therefore, the master problem \eqref{eq:master} is a strongly convex optimization problem with $n_1$ number of variables, which is usually easy to solve.
\subsection{Convergence Analysis}
\label{sec:convergence}
This subsection is about the global convergence of the sequence generated by Algorithm \ref{alg:finite_samples}.
We begin the analysis with
some technical assumptions. Since our focus of this section is to solve problem \eqref{eq:finite_scenarios} with fixed scenarios $\{\xi^1,\ldots,\xi^S\}$, one should interpret Assumptions B-F as requirements on all realizations $\{ \xi^1,\ldots,\xi^S \}$. The ``almost surely'' part will be used in the next section where a generally distributed $\tilde\xi$ is considered.
\vskip 0.05in
\noindent
\begin{center}
\fbox{%
\parbox{0.98\textwidth}{
{\bf Assumption C.} The recourse is relatively complete, i.e., $\psi(x;\xi)$ is finite for all feasible $x\in X$ (almost surely).\\[0.03in]
{\bf Assumption D.} The function $f(\bullet,y;\xi)$ is locally Lipschitz continuous in $x\in X$ with a uniform Lipschitz constant for all feasible second-stage solutions $y$ (almost surely). The partial subdifferential $\partial_1 (-f)(x,y;\xi)$ is outer semicontinuous jointly in $(x,y)$ (almost surely).\\[0.03in]
{\bf Assumption E.} There exists a measurable function $\kappa_1: \Xi \rightarrow \mathbb{R}_+$ such that $\kappa_1(\xi)<\infty$ (almost surely) and the following condition holds (almost surely):
\[
\left| \overline\psi(x_1, z; \xi) - \overline\psi(x_2, z; \xi) \right| \leq \kappa_1(\xi) \, \| x_1 - x_2 \|,\quad \forall\, x_1, x_2\in \mathbb{R}^{n_1}, z \in X.
\]
\vskip -0.07in
}}
\fbox{%
\parbox{0.98\textwidth}{
{\bf Assumption F.} The level set of the lifted function $\overline\psi(\bullet,z;\xi)$ is bounded for any $z\in X$ (almost surely).
}}
\end{center}
\vskip 0.05in
Some remarks are in order.
Assumption C is a frequently used assumption in the SP literature to ensure the continuity of $\psi(\bullet\,;\xi)$ on $X$ and $\overline\psi(\bullet, \bullet\,;\xi)$ on $\mathbb{R}^{n_1}\times X$ (almost surely).
Assumption D is used to characterize the partial subdifferential of $\overline\psi$ based on Lemma \ref{lem:subdiff}. Assumption E is a stochastic version of the Lipschitz condition in Lemma \ref{lem:gap_bound}. Since $X$ is a compact set, we assume without loss of generality that $\kappa_1(\xi)$ is independent of $z$.
This assumption also allows us to apply the law of large numbers (LLN) for the subdifferentials of a generally distributed $\xi$ in the next section.
It follows from Assumption F and the continuity of $\overline\psi(\bullet,z;\xi)$ that the latter function is bounded below for any $z\in X$ (almost surely), i.e.,
\begin{equation}\label{bounded below}
\displaystyle\inf_{x\in\mathbb{R}^{n_1}} \overline\psi(x,z;\xi) > -\infty \quad \forall\, z\in X \; \mbox{(almost surely)}.
\end{equation}
This condition is needed in Lemma \ref{lem:prox_converge1} to ensure the subsequential convergence of the proximal points $\{x^s_{\nu, i}\}_{i\geq 0}$ in the subproblems for each $s$.
For convenience, we denote
\begin{equation}\label{eq:zeta def}
\overline\zeta_{S}(x) \triangleq \varphi(x)+\frac{1}{S}\sum\limits_{s=1}^{S} \psi(x; \xi^s)\quad \mbox{and} \quad \widehat\zeta_{\,S,\gamma_\nu}(x) \triangleq \varphi(x)+\frac{1}{S}\sum\limits_{s=1}^{S} e_{\gamma_\nu}\psi(x; \xi^s).
\end{equation}
In the proposition below, we show that for any prescribed positive scalar $\varepsilon_\nu$, the $\nu$-th inner loop of Algorithm \ref{alg:finite_samples} terminates in finite steps.
\begin{proposition}\label{thm:innerconverge1}{\rm (Convergence of the inner loop for Algorithm \ref{alg:finite_samples})}
Suppose that Assumptions A-F hold. Then the following statements hold for any $\nu$-th inner loop.\\
(a)
$\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i+1}) \leq \widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i})- \displaystyle\frac 1 {2\gamma_\nu} \|x_{\nu,i} - x_{\nu,i+1}\|^2$ for any $i\geq 0$.\\
(b) $\displaystyle\lim_{i\to 0} \|x_{\nu,i} - x_{\nu, i+1}\| = 0$ and the stopping criterion $\|x_{\nu, i+1} - x_{\nu, i}\| \leq {\varepsilon_\nu} \gamma_\nu$ is achievable in finite steps, i.e.,
\[
i_{\nu}\triangleq\min\left\{\,i \in \mathbb{Z}_+ \left.\,\middle|\; \|x_{\nu,i+1} - x_{\nu,i}\| \leq {\varepsilon_\nu} \gamma_\nu \right.\right\} < +\infty.
\]
In addition, the output of the $\nu$-th inner loop $x_{\nu+1} = x_{\nu, i_\nu+1}$ satisfies
\begin{align*}
&\operatorname{dist}\left(0,
{\begin{array}{cc}
\displaystyle\frac 1 S\sum\limits_{s=1}^{S} \left[\partial_1 \overline\psi (x^s_{\nu,i_{\nu}}, x_{\nu, i_\nu};\xi^{s}) - \partial_2 (-\overline\psi) (x^s_{\nu,i_{\nu}},x_{\nu, i_\nu};\xi^{s})\right]\\
+ \; \partial\varphi(x_{\nu, i_\nu+1})+ \mathcal N_{X}(x_{\nu, i_\nu+1})
\end{array}}
\right) \leq \varepsilon_{\nu}.
\end{align*}
\end{proposition}
\begin{proof}
Consider the $\nu$-th inner loop of Algorithm \ref{alg:finite_samples}.
We first show that the sequence $\left\{\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i})\right\}_{i\geq 0}$ is nonincreasing. Since
the function $\widehat{e}_{\gamma_{\nu}}\psi(\,\bullet\,; \xi^{s}; x_{\nu,i})$ defined in \eqref{eq:Moreau_surrogate} is quadratic, we have, by writing $a_{\nu,i}^s \,\triangleq \, \nabla_x\,\widehat{e}_{\gamma_{\nu}}\psi(\,\bullet\,;\xi^{s}; x_{\nu,i})(x_{\nu,i+1})$,
\begin{align*}
\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i})
=
& \;\varphi(x_{\nu,i})+
\frac{1}{S} \sum\limits_{s=1}^{S} \widehat{e}_{\gamma_{\nu}}\psi(x_{\nu,i}; \xi^{s}; x_{\nu,i}) \\
= &\; \varphi(x_{\nu,i}) + \frac{1}{S} \sum\limits_{s=1}^{S}\widehat{e}_{\gamma_{\nu}}\psi(x_{\nu,i+1}; \xi^{s}; x_{\nu,i})
+\frac{1}{S}\sum\limits_{s=1}^{S}{(a_{\nu,i}^s)}^\top(x_{\nu,i}-x_{\nu,i+1})\\
&+ \frac 1 {2\gamma_\nu}\|x_{\nu,i}-x_{\nu,i+1}\|^2,
\end{align*}
where the first equality is because $\widehat{e}_{\gamma_{\nu}}\psi(x_{\nu,i};\xi^{s};x_{\nu,i}) = e_{\gamma_\nu}\psi(x_{\nu,i}; \xi^s)$.
Since $x_{\nu,i+1}$ is the optimal solution of the master problem \eqref{eq:master}, one may obtain that
\[
\left(b_{\nu,i}+\frac{1}{S}\sum\limits_{s=1}^S a^s_{\nu, i} \right)^\top(x_{\nu,i}-x_{\nu,i+1}) \geq 0 \quad\text{for some $b_{\nu, i}\in\partial\varphi(x_{\nu,i+1})$.}
\]
The convexity of $\varphi$ and the above inequalities imply that
\begin{equation}\label{eq:descent}
\begin{split}
\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i})\geq & \;\,\varphi(x_{\nu,i+1})+\frac{1}{S} \sum\limits_{s=1}^{S} \widehat{e}_{\gamma_{\nu}}\psi(x_{\nu,i+1};\xi^{s}; x_{\nu,i})+\frac 1 {2\gamma_\nu}\|x_{\nu,i}-x_{\nu,i+1}\|^2\\
\geq & \;\, \varphi(x_{\nu,i+1})+\frac{1}{S} \sum\limits_{s=1}^{S} e_{\gamma_{\nu}}\psi(x_{\nu,i+1};\xi^{s})+\frac 1 {2\gamma_\nu}\|x_{\nu,i}-x_{\nu,i+1}\|^2\\
= & \;\, \widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i+1}) + \frac 1 {2\gamma_\nu} \|x_{\nu,i} - x_{\nu,i+1}\|^2,
\end{split}
\end{equation}
where the second inequality uses the fact that $\widehat{e}_{\gamma}\,\psi(x;\xi^s; \bar{x}) \geq e_{\gamma}\psi(x;\xi^s)$ for any $x$, $\bar{x}$ and $\xi^s$. We thus prove the part (a). Using the compactness of $X$ and Assumption E, one may further derive from Lemma \ref{lem:gap_bound} that
\[
\inf\limits_{x \in X} \, \widehat\zeta_{\,S,\gamma_\nu}(x) \, \geq \, \left[\, \inf\limits_{x \in X}\, \varphi(x) \,\right] + \frac 1 {S} \sum\limits_{s=1}^{S} \left[\,\inf\limits_{x \in X}\psi(x; \xi^{s}) - \frac {\gamma_{\nu}} 2 \cdot\kappa_1(\xi^s)^2 \,\right]> -\infty.
\]
Hence, the sequence $\left\{ \widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i}) \right\}_{i\geq 0}$ is bounded below, which further yields that $\left\{\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu,i})\right\}_{i\geq 0}$ converges and $\|x_{\nu,i+1} - x_{\nu,i}\|$ converges to $0$ as $i\rightarrow \infty$. The latter convergence indicates that the inner iterations terminate after finite steps.
To show the rest statement of this proposition, we first obtain from the optimality condition of the master problem \eqref{eq:master} at $x_{\nu,i_\nu+1}$ that
\begin{equation}\label{eq:proof theorem1}
\begin{array}{rl}
0\;\in &\partial \left(\varphi(x_{\nu,i_\nu+1})+
\displaystyle\frac{1}{S} \sum\limits_{s=1}^{S} \widehat{e}_{\gamma_{\nu}}\psi (x_{\nu,i_\nu+1}; \xi^{s}; x_{\nu,i_\nu}) \right) + \mathcal N_{X}(x_{\nu,i_\nu+1})\\[0.15in]
= & \partial\varphi(x_{\nu,i_\nu+1}) + \displaystyle\frac 1 {S}\sum\limits_{s=1}^{S} \left( \frac {x_{\nu,i_\nu+1} - x^s_{\nu,i_\nu}} {\gamma_\nu} -c^s_{\nu,i_\nu}\right) + \mathcal N_{X}(x_{\nu,i_\nu+1}),
\end{array}
\end{equation}
where the equation is due to the sum rule of the subdifferentials for convex functions
\cite[Theorem 23.8]{rockafellar1997convex}.
From the optimality condition of the subproblem \eqref{eq:subprob}, we obtain
\begin{align*}
0\;&\in\;\partial_1 \;\overline\psi\left(x^s_{\nu,i_\nu}, x_{\nu,i_\nu};\xi^{s}\right) + \left(x^s_{\nu,i_\nu} - x_{\nu,i_\nu}\right)/{\gamma_{\nu}},\quad\forall \, s=1,\ldots,S.
\end{align*}
Taking the sum over \eqref{eq:proof theorem1} and the above inclusions from $s = 1, \ldots, S$, we get
\begin{align*}
0
&\in \frac 1 {S}\sum\limits_{s=1}^{S} \partial_1 \;\overline\psi\left(x^s_{\nu,i_\nu}, x_{\nu,i_\nu};\xi^{s}\right) + \left[ \partial \varphi(x_{\nu,i_{\nu+1}}) - \frac 1 S\sum\limits_{s=1}^S c^s_{\nu,i_\nu} + \mathcal N_{X}(x_{\nu,i_\nu+1}) \right] + (x_{\nu,i_\nu+1} - x_{\nu,i_\nu})/\gamma_{\nu}\\
& \subseteq \frac 1 S\sum\limits_{s=1}^{S} \left[\partial_1 \overline\psi (x^s_{\nu,i_{\nu}}, x_{\nu,i_{\nu}};\xi^{s}) - \partial_2 (-\overline\psi\,) (x^s_{\nu,i_{\nu}},x_{\nu,i_{\nu}};\xi^{s})\right]+ \partial\varphi(x_{\nu,i_\nu+1}) + (x_{\nu,i_\nu+1} - x_{\nu,i_\nu})/{\gamma_{\nu}}\\
& \quad + \mathcal N_{X}(x_{\nu,i_\nu+1}),
\end{align*}
where the last inclusion is due to the definition of $c^s_{\nu,i_\nu}$. Consequently, we derive
\begin{align*}
\operatorname{dist}\left(0,
{\begin{array}{cc}
\displaystyle\frac 1 S\sum\limits_{s=1}^{S} \left[\partial_1 \overline\psi (x^s_{\nu,i_{\nu}}, x_{\nu,i_{\nu}};\xi^{s}) - \partial_2 (-\overline\psi\,) (x^s_{\nu,i_{\nu}},x_{\nu,i_{\nu}};\xi^{s})\right]\\
+ \; \partial\varphi(x_{\nu,i_{\nu}+1})+ \mathcal N_{X}(x_{\nu,i_{\nu}+1})
\end{array}}
\right)
\leq \; \|x_{\nu,i_\nu+1} - x_{\nu,i_\nu}\|/{\gamma_{\nu}} \leq \varepsilon_{\nu},
\end{align*}
where the last inequality is due to the stopping rule of the inner loop.
\end{proof}
Let $\bar{x}$ be an accumulation point of the sequence $\{x_{\nu}\}_{\nu\geq 0}$, which must exist due to the compactness of $X$. We let $\{x_{\nu+1}\}_{\nu\in\alpha}$ be a convergent subsequence to $\bar{x}$, where $\alpha$ is a subset of $\mathbb N$.
In the following lemma, we show the convergence of the sequence $\{x^s_{\nu,i_\nu}\}_{\nu\in \alpha}$ for each $s$ with $x^s_{\nu,i_\nu}$ being the partial optimal solution of the subproblem \eqref{eq:subprob} at $(z,\xi) = (x_{\nu, i_\nu}, \xi^s)$.
\begin{lemma}\label{lem:prox_converge1}
Let Assumptions A-F hold and $\bar{x}$ be an accumulation point of the sequence $\{x_{\nu}\}_{\nu\geq 0}$ generated by Algorithm \ref{alg:finite_samples}. If the subsequence $\{x_{\nu+1}\}_{\nu\in\alpha}$ converges to $\bar{x}$, then $\{x_{\nu,i_\nu}\}_{\nu\in\alpha}$ converges to $\bar x$ and, for each $s=1,\ldots,S$, the corresponding subsequence $\{x^s_{\nu,i_\nu}\}_{\nu\in\alpha}$ also converges to $\bar{x}$.
\end{lemma}
\begin{proof}
Since $\lim\limits_{\nu \rightarrow \infty} \|x_{\nu+1} - x_{\nu,i_\nu}\| = \lim\limits_{\nu \rightarrow \infty} \|x_{\nu,{i_\nu+1}} - x_{\nu,i_\nu}\| = 0$ from Theorem \ref{thm:innerconverge1}, we know that $\{x_{\nu,i_\nu}\}_{\nu\in\alpha}$ also converges to $\bar x$. For each $s=1,\ldots,S$ and $\nu\in\alpha$, we have, by the definition of $x^s_{\nu,i_\nu} = P_{\gamma_\nu}\psi(x_{\nu,i_{\nu}};\xi^s)$,
\begin{align*}
\frac {1} {2 \gamma_{\nu}} \left\| x^s_{\nu,i_\nu} - x_{\nu,i_\nu}\right\|^2 + \inf_{x \in \mathbb{R}^{n_1}} \overline\psi (x, x_{\nu,i_\nu};\xi^{s})
\leq \; &\frac {1} {2 \gamma_{\nu}} \left\| x^s_{\nu,i_\nu} - x_{\nu,i_\nu}\right\|^2 + \overline\psi \left(x^s_{\nu,i_\nu}, x_{\nu,i_\nu};\xi^{s}\right)\\
= \; &e_{\gamma_{\nu}} \psi \left(x_{\nu,i_\nu}; \xi^{s}\right)
\leq \frac {1} {2 \gamma_{\nu}} \| \bar{x} - x_{\nu,i_\nu}\|^2 + \overline\psi (\bar{x}, x_{\nu,i_\nu};\xi^{s}).
\end{align*}
The above inequality yields that
\begin{align*}
\left\| x^s_{\nu,i_\nu} - x_{\nu,i_\nu} \right\|
&\leq \sqrt{\| \bar{x} - x_{\nu,i_\nu}\|^2 + 2\gamma_{\nu}\Big(\, \overline\psi (\bar{x}, x_{\nu,i_\nu};\xi^{s}) - \displaystyle\operatornamewithlimits{inf}_{x \in \mathbb{R}^{n_1}} \overline\psi (x, x_{\nu,i_\nu};\xi^{s}) \Big)}\\
&\leq \| \bar{x} - x_{\nu,i_\nu} \| + \sqrt{2\gamma_\nu \Big(\, \overline\psi (\bar{x}, x_{\nu,i_\nu};\xi^{s}) - \displaystyle\operatornamewithlimits{inf}_{x \in \mathbb{R}^{n_1}} \overline\psi (x, x_{\nu,i_\nu};\xi^{s}) \Big)}\\
& \quad \rightarrow 0 \;\,\text{as $\nu\rightarrow\infty$}\,,
\end{align*}
where the last step is because $\gamma_\nu \to 0$, $\overline{\psi}(\bar{x},\bullet;\xi^s)$ is continuous and thus bounded on $X$, and the inequality \eqref{bounded below}.
Hence, we derive that
\begin{align*}
\left\| x^s_{\nu,i_\nu} - \bar{x} \right\|
&\leq \left\| x^s_{\nu,i_\nu} - x_{\nu,i_\nu} \right\| + \left\| x_{\nu,i_\nu} - \bar{x} \right\| \;\, \rightarrow 0 \;\,\text{as $\nu\rightarrow\infty$}\,.
\end{align*}
The proof is completed.
\end{proof}
We are now ready to present the global convergence of the sequence generated by Theorem \ref{thm:consistency1}. We shall prove that every accumulation point $\bar{x}$ of $\{x_{\nu}\}_{\nu\geq 0}$ is a critical point of problem \eqref{eq:finite_scenarios} satisfying
\begin{equation}\label{eq: critical point}
0 \in \partial \varphi(\bar{x}) + \frac 1 S\sum\limits_{s=1}^{S} \left[\partial_1 \overline\psi\left(\bar{x}, \bar{x};\xi^{s}\right) - \partial_2 (-\overline\psi) \left(\bar{x},\bar{x};\xi^{s}\right) \right] + \mathcal N_{X}(\bar{x}).
\end{equation}
It has been shown in Lemma \ref{lem:subdiff} that under Assumption D, we have $\partial_C \psi(x;\xi)\subseteq \partial_1 \overline\psi\left(x, x;\xi\right) - \partial_2 (-\overline\psi)(x, x;\xi)$. Hence, the condition \eqref{eq: critical point} is weaker than the Clarke stationarity of problem \eqref{eq:finite_scenarios} pertaining to $0\in \partial \varphi(\bar{x}) + \displaystyle\frac 1 S\sum\limits_{s=1}^{S} \partial_C \psi(\bar{x};\xi^s) + \mathcal N_{X}(\bar{x})$. The term ``critical point'' is adapted from the result of the difference-of-convex (dc) algorithm to solve a dc problem $\displaystyle\operatornamewithlimits{minimize}_{x\in X}\,\left[\, \theta_1(x) - \theta_2(x)\,\right]$, where the accumulation point satisfies $0\in \partial \theta_1(x) -
\partial\theta_2(x) + \mathcal{N}_X(x)$ \cite{an2005dc}.
\begin{theorem}[Subsequential convergence for Algorithm \ref{alg:finite_samples}]\label{thm:consistency1}
Let Assumptions A-F hold and $\bar{x}$ be an accumulation point of the sequence $\big\{ x_{\nu}\big\}_{\nu \geq 0}$ generated by Algorithm \ref{alg:finite_samples}.
If the set-valued mapping $\partial_1\overline\psi\left(\bullet, \bullet;\xi^s\right) - \partial_2(-\overline\psi)\left(\bullet, \bullet;\xi^s\right)$ is outer semicontinuous at $(\bar{x},\bar{x})$ for each $s$, then $\bar{x}$ is a critical point of \eqref{eq:finite_scenarios} satisfying
\eqref{eq: critical point}.
\end{theorem}
\begin{proof}
Let $\bar{x}$ be the limit of
a convergent subsequence $\big\{x_{\nu+1}\big\}_{\nu\in \alpha}$.
By Theorem \ref{thm:innerconverge1} and Lemma \ref{lem:prox_converge1}, we know that the subsequences $\big\{x_{\nu,i_\nu}\big\}_{\nu\in \alpha}$ and $\big\{x^{s}_{\nu,i_\nu}\big\}_{\nu\in \alpha}$ for all $s$ converge to the same accumulation point $\bar{x}$ as $\nu\rightarrow\infty$. Using the triangle inequality of the distance function, we have
\begin{align*}
&\operatorname{dist} \left( 0,
\begin{array}{cc}
\displaystyle\frac 1 S\sum\limits_{s=1}^{S} \left[\partial_1 \overline\psi\left(\bar{x}, \bar{x};\xi^{s}\right) - \partial_2 (-\overline\psi) \left(\bar{x},\bar{x};\xi^{s}\right) \right]
+ \; \partial \varphi(\bar{x}) + \mathcal N_{X}(\bar{x})
\end{array}\right)\\
\leq &\; \operatorname{dist}\left(0,
{\begin{array}{lc}
\; \displaystyle\frac 1 S\sum\limits_{s=1}^{S} \left[\,\partial_1 \overline\psi (x^s_{\nu, i_{\nu}}, x_{\nu, i_{\nu}}; \xi^{s}) - \partial_2 (-\overline\psi) (x^s_{\nu, i_{\nu}}, x_{\nu, i_{\nu}}; \xi^{s})\,\right] \\[0.2in]
\; + \;\partial\varphi(x_{\nu,i_\nu+1}) + \mathcal N_{X}(x_{\nu,i_\nu+1})
\end{array}}
\right)\\[0.1in]
& \; + \; \frac 1 S\sum\limits_{s=1}^S \mathbb D\left(\,
\left[
\begin{array}{cc}
\partial_1\overline\psi(x^s_{\nu,i_\nu}, x_{\nu,i_\nu};\xi^{s} )\\[0.05in]
- \; \partial_2(-\overline\psi)(x^s_{\nu,i_\nu}, x_{\nu,i_\nu};\xi^{s} )
\end{array} \right],\;
\left[
\begin{array}{cc}
\partial_1\overline\psi\left(\bar{x}, \bar{x};\xi^{s}\right)\\[0.05in]
- \; \partial_2(-\overline\psi)\left(\bar{x}, \bar{x};\xi^{s}\right)
\end{array} \right] \,\right)\\[0.05in]
& \; + \; \mathbb D \left(\,\partial \varphi(x_{\nu,i_\nu+1}), \; \partial \varphi(\bar{x})\,\right)
+ \mathbb D\left(\, \mathcal N_{X}(x_{\nu,i_\nu+1}), \; \mathcal N_{X}(\bar{x}) \,\right),
\end{align*}
where the first term in the preceding inequality converges to $0$ by Theorem \ref{thm:innerconverge1};
the second term goes to $0$ due to the assumed outer semicontinuity of $\partial_1\overline\psi\left(\bullet, \bullet;\xi\right) - \partial_2(-\overline\psi)\left(\bullet, \bullet;\xi\right)$; and
the last two terms converge to $0$ due to the outer semicontinuity and compactness of $\partial \varphi$ and $\mathcal{N}_X$. The proof is thus completed.
\end{proof}
We remark that an obvious sufficient condition for $\partial_1\overline\psi\left(\bullet, \bullet;\xi^s\right) - \partial_2(-\overline\psi)\left(\bullet, \bullet;\xi^s\right)$ to be outer semicontinuous at $(\bar{x},\bar{x})$ is that both $\partial_1\overline\psi\left(\bullet, \bullet;\xi^s\right)$ and $\partial_2(-\overline\psi)\left(\bullet, \bullet;\xi^s\right)$ are outer semicontinuous at $(\bar{x},\bar{x})$. In fact, the set-valued mappings
$\partial_1\overline\psi\left(\bullet, \bar{x};\xi^s\right)$ and $\partial_2(-\overline\psi)\left(\bar{x}, \bullet;\xi^s\right)$ are known to be outer semicontinuous since they are the subdifferentials of convex functions.
In the following, we establish the convergence to a stronger type of stationarity under additional assumptions. Suppose that $f(\bullet,\bullet;\xi^s)$ and $G(\bullet, \bullet; \xi^s)$ are continuously differentiable for all $\xi^s$. At $x = \bar{x}$, we say $\bar{y}^s$ is an optimal solution of the convex second-stage problem with $\xi = \xi^s$ and $\bar{\lambda}^s$ being the corresponding multiplier if the following Karush-Kuhn-Tucker condition is satisfied:
\begin{equation}\label{eq:KKT}
0\in \nabla_y f(\bar{x},\bar{y}^{s};\xi^s) + \; \displaystyle\sum_{j=1}^\ell\bar{\lambda}_j^s \nabla_y g_j(\bar{x},\bar{y}^s;\xi^s)\,\;\mbox{and}\;\,
\bar{\lambda}^s \in N_{\mathbb{R}^\ell_-}\left(G(\bar{x},\bar{y}^s;\xi^s)\right).
\end{equation}
We use $Y(\bar{x}, \xi^s)$ and $M(\bar{x},\xi^s)$ to denote the set of all optimal solutions and multipliers satisfying the above condition, respectively. When $M(\bar{x},\xi^s)$ is nonempty, one may write the critical cone of the second-stage problem at $\bar{y}^s \in Y(\bar{x}, \xi^s)$ as
\begin{gather*}
C_{\bar{x}}(\bar{y}^s,\xi^s) \triangleq \left\{ d \in \mathbb{R}^{n_2} \,\middle| \begin{array}{cc} \nabla_y f(\bar{x},\bar{y}^s;\xi^s)^\top d=0,\\[0.05in]
\nabla_y\, g_j(\bar{x},\bar{y}^s;\xi^s)^\top d \in \mathcal T_{\mathbb{R}-}(g_j(\bar{x},y;\xi)),\; j=1, \cdots, \ell \end{array} \right\}.
\end{gather*}
where ${\cal T}_D(x)$ denotes the tangent cone of a closed convex set $D$.
In the following, we show that if the second-stage solutions $\{\bar{y}^s\}_{s=1}^S$ are unique at the accumulation point $\bar{x}$ for each $s$, then $\bar{x}$ in fact satisfies a stronger condition.
\begin{corollary}[Convergence to a directional stationary point]
\label{corollary:convergence}
Let $\bar{x}$ be an accumulation point of the sequence $\big\{ x_{\nu}\big\}_{\nu \geq 0}$ generated by Algorithm \ref{alg:finite_samples}. In addition to the assumptions in Theorem \ref{thm:consistency1}, if the following conditions are satisfied for each $s$:
\begin{enumerate}[leftmargin = 15pt, topsep = 2pt, itemsep = 2pt, label={(\alph*)}]
\item $f(\bullet,\bullet;\xi^s)$ and $G(\bullet,\bullet;\xi^s)$ are twice continuously differentiable;
\item the set of multipliers $M(\bar{x}, \xi^s)$ is nonempty and
there exists $\bar y^s\in Y(\bar{x}, \xi^s)$ satisfying the second order sufficient condition that for all $d\in C_{\bar{x}}( \bar y^s, \xi^s)\setminus \{0\}$,
\[
\sup_{(\mu,\lambda)\in M(\bar y^s,\xi^s)} d^\top \nabla^2_{yy}\left[\, f(\bar{x},\bar y^s;\xi^s) + \sum_{j=1}^\ell \lambda_j g_j(\bar{x},\bar y^s;\xi^s)\,\right] d > 0;
\]
\end{enumerate}
then $\bar{x}$ is a directional stationary point of problem \eqref{eq:finite_scenarios}, i.e.,
\[
\varphi^\prime(\bar{x};d) + \frac{1}{S}\sum_{s=1}^S \psi^\prime(\bar{x};d) \geq 0, \quad \forall\; d\in {\cal T}_X(\bar{x}).
\]
\end{corollary}
\begin{proof}
We first prove that $\partial_2 \overline\psi(\bar{x},\bar{x};\xi^s)$ is a singleton under given assumptions. It follows from Lemma \ref{lem:subdiff} that
\[
\partial_2(-\overline\psi)(\bar{x},\bar{x};\xi^s) = \operatorname{conv}\Big\{ -\nabla_x f(\bar{x},y;\xi^s) \mid y\in Y(\bar{x};\xi^s) \Big\}.
\]
Since the second order sufficient condition of the second-stage problem holds at $\bar{x}$ for any $\xi^s$, we have that $Y(\bar{x};\xi^s)$ is a singleton \cite[Example 13.25]{rockafellar2009variational}, which further implies that $ \partial_2(-\overline\psi)(\bar{x},\bar{x};\xi^s)$ is a singleton. The desired directional stationarity of $\bar{x}$ then follows from \cite[Proposition 6.1.11]{cui2020modern}.
\end{proof}
Theorem \ref{thm:consistency1} and Corollary \ref{corollary:convergence}
pertain to the subsequential convergence of the iterative sequence generated by Algorithm \ref{alg:finite_samples}. In the following, we show that the full sequence of the objective values along the iterations converges if the sequence of the Moreau parameters $\{\gamma_\nu\}_{\nu\geq 0}$ is summable. This result particularly indicates that although the sequence $\{x_{\nu}\}_{\nu \geq 0}$ may have multiple accumulation points, the objective values at the accumulation points are the same.
To proceed, we remind the readers of the definition of $\overline\zeta_{S}$ in \eqref{eq:zeta def}.
\begin{theorem}[Convergence of objective values for Algorithm \ref{alg:finite_samples}]
Suppose that Assumptions A-F hold. Let $\{ x_{\nu}\}_{\nu \geq 0}$ be the sequence generated by Algorithm \ref{alg:finite_samples} under the additional condition that $\sum\limits_{\nu=0}^\infty \gamma_\nu < \infty$. Then
$\lim\limits_{\nu\rightarrow\infty}\overline\zeta_S(x_\nu) = \overline\zeta_S(\bar{x})$,
where $\bar{x}$ is any accumulation point of the iterative sequence $\{x_{\nu}\}_{\nu \geq 0}$.
\end{theorem}
\begin{proof}
One may derive that
\begin{equation}\label{eq:value_diff}
\begin{split}
\overline\zeta_{S}(x_{\nu+1})-\overline\zeta_{S}(x_\nu)
= \; &\Big[\, \overline\zeta_{S}(x_{\nu+1})-\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu+1}) \,\Big]\\
&+\Big[\, \widehat\zeta_{\,S,\gamma_\nu}(x_{\nu+1})-\widehat\zeta_{\,S,\gamma_\nu}(x_\nu) \,\Big]+\Big[\, \widehat\zeta_{\,S,\gamma_\nu}(x_\nu)-\overline\zeta_{S}(x_\nu) \,\Big],
\end{split}
\end{equation}
where the first and last terms on the right side are gaps between the partial Moreau envelopes $\widehat\zeta_{\,S,\gamma_\nu}$ and original functions $\overline\zeta_{S}$ at $x_{\nu+1}$ and $x_\nu$, respectively. By Lemma \ref{lem:gap_bound} and Assumption E, we may obtain that
\begin{equation}\label{eq:PMEerror}
\overline\zeta_{S}(x_{\nu+1})-\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu+1}) \leq \frac 1 {S} \sum\limits_{s=1}^{S} {\gamma_\nu \, \kappa(\xi^s)^2}/2 \quad\text{and}\quad \widehat\zeta_{\,S,\gamma_\nu}(x_\nu)-\overline\zeta_{S}(x_\nu) \leq 0.
\end{equation}
Recall that $x_{\nu+1}=x_{\nu,i_\nu+1}$ and $x_\nu=x_{\nu,0}$. Then the second term on the right side of \eqref{eq:value_diff} can be bounded above based on Theorem \ref{thm:innerconverge1}(a) that
\begin{equation}\label{eq:Descenterror}
\widehat\zeta_{\,S,\gamma_\nu}(x_{\nu+1})-\widehat\zeta_{\,S,\gamma_\nu}(x_\nu) \leq -\frac 1 {2\gamma_\nu} \sum\limits_{i=0}^{i_\nu}\|x_{\nu,i+1}-x_{\nu,i}\|^2.
\end{equation}
Substituting \eqref{eq:PMEerror} and \eqref{eq:Descenterror} into the inequality \eqref{eq:value_diff}, we have
\[
\overline\zeta_{S}(x_{\nu+1})-\overline\zeta_{S}(x_\nu)\leq
- \frac 1 {2\gamma_\nu} \sum\limits_{i=0}^{i_\nu}\|x_{\nu,i+1}-x_{\nu,i}\|^2 + \frac {\gamma_\nu} 2 \left[\frac 1 {S} \sum\limits_{s=1}^{S}\kappa(\xi^s)^2 \right].
\]
Since $\displaystyle\sum\limits_{\nu=1}^\infty \frac {\gamma_\nu} 2 \left[\frac 1 {S} \sum\limits_{s=1}^{S}\kappa(\xi^s)^2 \right] < \infty$ and the sequence $\{ \overline\zeta_S(x_{\nu})\}_{\nu\geq 0}$ must be bounded below due to
Assumption C, one may easily obtain the convergence of $\left\{\overline\zeta_S(x_{\nu})\right\}_{\nu \geq 0}$ that is a so-called quasi-Fej{\'e}r monotone sequence; see, e.g., \cite[Lemma 3.1]{combettes2001quasi}.
For any convergent subsequence $\{x_{\nu+1}\}_{\nu\in\alpha}$ and its
limit $\bar{x}$, we have $\lim\limits_{\nu(\in \alpha)\rightarrow\infty}\overline\zeta_S(x_{\nu+1}) = \overline\zeta_S(\bar{x})$ by the continuity of $\psi(x;\xi^s)$ on $X$ for each $s$. Therefore, the full sequence $\left\{\overline\zeta_S(x_{\nu})\right\}_{\nu\geq 0}$ converges to $\overline\zeta_S(\bar{x})$ for any accumulation point $\bar{x}$.
\end{proof}
\section{A sampling-based decomposition algorithm}
\label{sec:sampling}
In this section, we consider a generally distributed random vector $\tilde\xi$ with a known distribution. Instead of the approach in the previous section that deals with a fixed batch of samples throughout the algorithm,
we incorporate the sampling strategy into the outer loop to progressively enlarge the problem size.
In general, there are two ways to do the sampling for solving SPs. One is to use the sample average approximation to select a subset of data before the execution of the numerical algorithm \cite{shapiro2003monte,shapiro2021lectures,homem2014monte}. The other is to adopt a sequential sampling technique \cite{higle2013stochastic,Homem-de-Mello03, pasupathy2021adaptive, RoysetPolak07} where scenarios are gradually added along the iterations. Our method falls into the latter category.
\begin{algorithm}[H]
\caption{A sampling-based decomposition algorithm for the SP \eqref{eq:first_stage}}
\label{alg:sampling_based}
{\bf Input:} Initial point $x_0\in X$, two positive scalar sequences $\{\gamma_{\nu}\}_{{\nu}=0}^\infty \downarrow 0$, $\{\varepsilon_\nu\}_{\nu=0}^\infty \downarrow 0$, and a sequence of incremental sample size $\{S_\nu\}_{\nu=0}^\infty$.
\vskip 0.1in
{\bf Outer loop:} Set $S_{-1}=0$ and $\nu=0$.
\begin{algorithmic}[1]
\State Generate i.i.d.\ samples $\left\{\xi^{S_{\nu-1} +\triangle}\right\}_{\triangle=1}^{S_{\nu} - S_{\nu-1}}$ from the distribution of $\tilde\xi$ that are independent of previous samples.
\State Execute the inner loop of Algorithm \ref{alg:finite_samples} with the initial point $x_\nu$, samples $\{\xi^s\}_{s=1}^{S_{\nu}}\triangleq\{\xi^s\}_{s=1}^{S_{\nu-1}}\bigcup\left\{\xi^{S_{\nu-1} +\triangle}\right\}_{\triangle=1}^{S_{\nu} - S_{\nu-1}}$, and parameters $\gamma_{\nu}$ and $\varepsilon_\nu$.
\State Set $\nu \leftarrow \nu+1$ and repeat step 1 until a prescribed stopping criterion is satisfied.
\end{algorithmic}
\end{algorithm}
We rely on the LLN for convex subdifferentials to establish the almost surely convergence of the iterative sequence generated by Algorithm \ref{alg:sampling_based}. To facilitate this tool, Lipschitz continuity of the original function is needed. In our setting, this requires that the lifted function $\overline\psi(\bullet,\bullet;\xi)$ is Lipschitz continuous. In addition to Assumption E on the Lipschitz continuity of $\overline\psi(\bullet,z;\xi)$, we further pose the following assumption on the Lipschitz continuity of $\overline\psi(x,\bullet;\xi)$.
\vskip 0.05in
\noindent
\begin{center}
\fbox{%
\parbox{0.98\textwidth}{
{\bf Assumption G.} There exists a measurable function $\kappa_2: \Xi \rightarrow \mathbb{R}_+$ such that $\mathbb{E}[\kappa_2(\xi)] < \infty$ and the following inequality holds for almost any $\xi\in \Xi$:
\[
\left| \overline\psi (x, x_1;\xi) - \overline\psi (x, x_2;\xi) \right| \leq \kappa_2(\xi) \, \| x_1 - x_2 \|,\quad\forall \, x\in\mathbb{R}^{n_1}, \;x_1, x_2 \in X.
\]
\vskip -0.05in
}}
\end{center}
\vskip 0.05in
For any $r>0$, we denote
\[
\partial_1^{\,r} \overline{\psi}(\bar{x}, \bar{x}; \xi) \triangleq \bigcup\limits_{x, x^\prime \in \mathbb{B}(\bar{x}, r)} \partial_1 \;\overline{\psi}(x, x^\prime; \xi),
\]
and similarly for $\partial_2^r$. We have the following result on the LLN for the subdifferentials of icc functions, which is a consequence of the LLN for random set-valued mappings \cite{artstein1975strong}. In fact, the result can be viewed as a pointwise version of \cite[Theorem 2]{shapiro2007uniform}.
\begin{lemma}
\label{lem:ULLN}
Suppose that Assumptions F and G hold. For any $x\in X$, if the set-valued mapping $\partial_1\overline\psi(\bullet, \bullet; \xi) - \partial_2(-\overline\psi)(\bullet, \bullet; \xi)$ is outer semicontinuous at $(x,x)$ almost surely, then for any $r > r^\prime \geq 0$, the following limit holds almost surely:
\[
\lim_{\nu\rightarrow\infty} \mathbb D \left(
\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu}
\begin{bmatrix} \hskip -0.2in \partial_1^{\,r^\prime} \,\overline\psi (x, x; \xi^{s}) \\[0.05in]
\; - \,\partial_2^{\,r^\prime} (-\overline\psi)(x, x; \xi^{s}) \, \end{bmatrix},\hskip -0.2in
\begin{array}{cc}\mathbb{E}_{\tilde\xi}\left[ \partial_1^{\,r} \, \overline\psi (x, x; \tilde\xi)\right] \\[0.05in]
\qquad -\mathbb{E}_{\tilde\xi}\left[ \partial_2^{\,r} (-\overline\psi)(x, x; \tilde\xi)\right] \end{array}\right)
= 0,
\]
where the expectation of a random set-valued mapping $\mathbb{E}_{\tilde\xi}\left[\mathcal A(x;\tilde\xi)\right]$ is defined as the set of $\mathbb{E}_{\tilde\xi}\left[a(x;\tilde\xi)\right]$ for all measurable selections $a(x;\xi) \in \mathcal A(x;\xi)$.
\end{lemma}
A noteworthy remark about the preceding lemma is that we can interchange the partial subdifferential and expectation in the right-hand side of the distance, i.e.,
\[
\mathbb{E}_{\tilde\xi}\left[\partial_1\overline\psi(x,x;\tilde\xi)\right]=\partial_1 \, \mathbb{E}_{\tilde\xi}\left[\, \overline\psi(x,x;\tilde\xi)\right].
\]
This is because a convex function is Clarke regular and the Clarke regularity ensures the interchangeability of the subdifferential and the expectation \cite[Proposition 2.3.6 and Theorem 2.7.2]{clarke1990optimization}. Next we prove a technical lemma that is similar to Lemma \ref{lem:prox_converge1}.
\begin{lemma}\label{lem:prox_converge2}
Let Assumptions A-G hold. If a subsequence $\{x_{\nu+1}\}_{\nu\in \alpha}$ generated by Algorithm \ref{alg:sampling_based} converges to a point $\bar{x}$, then the subsequence $\{x_{\nu, i_\nu}\}_{\nu\in \alpha}$ also converges to $\bar x$, and for any $r>0$ and all $\nu\in\alpha$ sufficiently large, the subsequence $\left\{x^s_{\nu,i_\nu}\right\}_{s=1}^{S_\nu} \subset \mathbb{B}(\bar{x},r)$ almost surely.
\end{lemma}
\begin{proof}
Following similar derivation in the proof of Lemma \ref{lem:prox_converge1}, we obtain the convergence of $\{x_{\nu, i_\nu}\}_{\nu\in \alpha}$ to $\bar x$ and the following inequalities
\begin{align*}
\left\| x^s_{\nu,i_\nu} - \bar{x} \right\|
&\leq \left\| x^s_{\nu,i_\nu} - x_{\nu,i_\nu} \right\| + \left\| x_{\nu,i_\nu} - \bar{x} \right\|\\
&\leq 2 \left\| x_{\nu,i_\nu} - \bar{x} \right\| + \sqrt{2\gamma_\nu \Big(\, \overline\psi (\bar{x}, x_{\nu,i_\nu};\xi^{s}) - \inf\limits_{x \in \mathbb{R}^{n_1}} \overline\psi (x, x_{\nu,i_\nu};\xi^{s}) \Big)}\\
&\leq 2 \left\| x_{\nu,i_\nu} - \bar{x} \right\| + \sqrt{2\gamma_\nu \, \kappa_1(\xi^s) \, R(X)},
\end{align*}
where $R(X)$ denotes the diameter of the compact set $X$, and the last inequality follows from \eqref{bounded below} due to Assumption F and the uniform Lipschitz continuity of $\overline\psi(\bullet,z;\xi)$ in Assumption E. Since $\kappa_1(\xi)<\infty$ almost surely and $\gamma_\nu \to 0$, the claimed result follows.
\end{proof}
Below is the main theorem of this section on the almost surely subsequential convergence of the iterative sequence generated by Algorithm \ref{alg:sampling_based}.
\begin{theorem}[Subsequential convergence of Algorithm \ref{alg:sampling_based}]\label{thm:consistency2}
Suppose that Assumptions A-G hold. Let $\{ x_{\nu}\}_{\nu \geq 0}$ be the sequence generated by Algorithm \ref{alg:sampling_based} and $\bar{x}$ be any accumulation point. If the set-valued mapping $\partial_1\overline\psi(\bullet, \bullet; \xi) - \partial_2(-\overline\psi)(\bullet, \bullet; \xi)$ is outer semicontinuous at $(\bar{x}, \bar{x})$ almost surely, then for any $r>0$, the following inclusion holds almost surely:
\[
0 \in \partial \varphi(\bar{x}) + \partial^{\,r}_1 \, \mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bar{x}, \bar{x}; \tilde\xi) \right] - \partial^{\,r}_2 \,\mathbb{E}_{\tilde\xi}\left[ (-\overline\psi)(\bar{x}, \bar{x}; \tilde\xi)\right] + \mathcal N_{X}(\bar{x}).
\]
In addition, if the set-valued mapping $\partial_1 \, \mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bullet, \bullet; \tilde\xi)\right] - \partial_2 \, \mathbb{E}_{\tilde\xi}\left[\, (-\overline\psi)(\bullet, \bullet; \tilde\xi) \right]$ is continuous at $(\bar{x}, \bar{x})$, then almost surely
\[
0 \in \partial \varphi(\bar{x}) + \partial_1 \, \mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bar{x}, \bar{x}; \tilde\xi) \right] - \partial_2\,\mathbb{E}_{\tilde\xi}\left[ (-\overline\psi)(\bar{x}, \bar{x}; \tilde\xi) \right] + \mathcal N_{X}(\bar{x}),
\]
i.e., every accumulation point $\bar{x}$ is a critical point of problem \eqref{eq:first_stage} almost surely.
\end{theorem}
\begin{proof}
Consider any subsequence $\{x_{\nu+1}\}_{\nu\in \alpha}$ that converges to $\bar{x}$. It follows from Lemma \ref{lem:prox_converge2} that for any $r^\prime>0$, the ball $\mathbb{B}(\bar{x},r^\prime)$ almost surely contains $x_{\nu,i_\nu}, x_{\nu,i_\nu+1},$ and the proximal points $\{x^s_{\nu,i_\nu}\}_{s=1}^{S_\nu}$ for all $\nu(\in\alpha)$ sufficiently large.
Consequently,
\begin{gather*}
\partial_1 \,\overline\psi ( x^{s}_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s})
\subseteq
\partial_1^{\,r^\prime} \, \overline\psi (\bar{x}, \bar{x}; \xi^{s}) \quad \text{almost surely for all $\nu\in\alpha$ sufficiently large.}
\end{gather*}
We then obtain from Lemma \ref{lem:ULLN} that the following limit holds almost surely:
\begin{equation}\label{eq:ULLN2}
\lim_{\nu\rightarrow \infty} \mathbb D \left(
\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \begin{bmatrix}\hskip -0.3in \partial_1 \,\overline\psi (x^s_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s}) \\[0.05in]
\;- \; \partial_2 (-\overline\psi)(x^s_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s})\end{bmatrix}
,\begin{array}{lc} \partial^{\,r}_1 \,\mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bar{x}, \bar{x}; \tilde\xi)\right] \\[0.05in] - \; \partial^{\,r}_2 \,\mathbb{E}_{\tilde\xi}\Big[ (-\overline\psi)(\bar{x}, \bar{x}; \tilde\xi)\Big]\end{array} \right)
= 0.
\end{equation}
Thus, for any $r > 0$, the following estimation follows almost surely:
\begin{align*}
&\operatorname{dist} \left( 0, \; \partial \varphi(\bar{x}) + \partial^{\,r}_1 \, \mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bar{x}, \bar{x}; \tilde\xi)\right] - \partial^{\,r}_2 \,\mathbb{E}_{\tilde\xi}\left[ (-\overline\psi)(\bar{x}, \bar{x}; \tilde\xi)\right] + \mathcal N_{X}(\bar{x}) \right)\\
&\leq
\underbrace{\operatorname{dist}\left(0,
{\begin{array}{cc}
\displaystyle\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu}\left[\partial_1 \, \overline\psi(x^s_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s}) - \partial_2 (-\overline\psi)(x^s_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s}) \right]\\
+ \; \partial\varphi(x_{\nu,i_\nu+1}) + \mathcal N_{X}(x_{\nu,i_\nu+1})
\end{array}}
\right)}_{\rightarrow 0 \text{ as }\nu \rightarrow \infty\text{ by Proposition }\rm\ref{thm:innerconverge1}
\text{\;(with sample size $S_\nu$ instead of $S$)}}\\[0.05in]
& + \underbrace{\mathbb D \left(
\displaystyle\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \begin{bmatrix}\hskip -0.3in \partial_1 \,\overline\psi (x^s_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s}) \\[0.05in]
\; - \; \partial_2 (-\overline\psi)(x^s_{\nu,i_\nu}, x_{\nu,i_\nu}; \xi^{s})\end{bmatrix}
,\begin{array}{lc} \partial^{\,r}_1 \,\mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bar{x}, \bar{x}; \tilde\xi)\right] \\ - \; \partial^{\,r}_2 \,\mathbb{E}_{\tilde\xi}\Big[ (-\overline\psi)(\bar{x}, \bar{x}; \tilde\xi)\Big]\end{array} \right)}_{\rightarrow 0 \text{ as }\nu \rightarrow \infty\text{ by }\eqref{eq:ULLN2}}\\[0.05in]
& +\underbrace{\mathbb D \left(\,\partial \varphi(x_{\nu,i_\nu+1}), \partial \varphi(\bar{x})\,\right) + \mathbb D\left( \mathcal N_{X}(x_{\nu,i_\nu+1}), \, \mathcal N_{X}(\bar{x}) \right)}_{\rightarrow 0 \text{ as }\nu \rightarrow \infty \; \text{due to the outer semicontinuity of $\partial \varphi$ and $\mathcal{N}_X$ }}.
\end{align*}
Finally, if the set-valued mapping $\partial_1 \, \mathbb{E}_{\tilde\xi}\left[\, \overline\psi(\bullet, \bullet;\tilde\xi)\right] - \partial_2\,\mathbb{E}_{\tilde\xi}\left[(-\overline\psi)(\bullet, \bullet;\tilde\xi)\right]$ is continuous at $(\bar{x}, \bar{x})$, one may adopt similar arguments as the proof of \cite[Theorem 3]{shapiro2007uniform} to derive the almost surely convergence to a critical point.
\end{proof}
One can further derive an analogous result of Corollary \ref{corollary:convergence} for the sequence generated by Algorithm \ref{alg:sampling_based} by strengthening the conditions (a) and (b) in the former corollary to almost any $\xi\in \Xi$ so that $\partial_2 \mathbb{E}_{\tilde\xi}\left[(-\overline\psi)(\bar{x}, \bar{x};\tilde\xi)\right]$ is a singleton. We omit the details here for brevity.
The last result of this section is the almost surely convergence of the objective values of $\{\overline\zeta_{S_\nu}(x_\nu)\}_{\nu\geq 0}$ under proper assumptions on the sample sizes $S_\nu$ and Moreau parameters $\gamma_\nu$. To proceed, we first
present a lemma on the convergence rate of the SAA in expectation. This result is obtained by using the Rademacher average of the random function $\psi(x;\tilde\xi)$, which has its source in \cite[Corollary 3.2]{ermoliev2013sample}; see also \cite[Theorem 10.1.5]{cui2020modern}.
\begin{lemma}\label{lem:ULLNrate}
Let $X$ be a compact set in $\mathbb{R}^n$ and $\tilde\xi: \Omega \rightarrow \Xi\subseteq\mathbb{R}^{m}$ be a random vector defined on a probability space $(\Omega, \mathcal F, \mathbb P)$. Let $\psi: X\times\Xi \rightarrow \mathbb{R}$ be a Carath{\'e}odory function.
Suppose that $\psi$ is uniformly bounded on $X\times\Xi$ and
Lipschitz continuous in $x$ with modulus independent of $\xi$.
Let $\{\xi^s\}_{s=1}^S$ be independent and identically distributed random vectors following the distribution of $\tilde\xi$. Then there exists a constant $C$ such that for any $\eta\in(0,1/2)$, we have
\[
\mathbb{E}\left[\, \sup\limits_{x\in X}\left|\frac 1 {S}\sum\limits_{s=1}^{S}\psi(x;\xi^s) - \mathbb{E}_{\tilde\xi}\left[ \psi(x;\tilde\xi)\right]\right| \,\right] \leq \frac {C \sqrt{1-2\eta}}{S^\eta}, \quad \forall\, S>0.
\]
\end{lemma}
We make a remark about the above lemma. For an icc function $\psi(x;\xi)$ associated with the lifted counterpart $\overline\psi(x,z;\xi)$, the uniform Lipschitz continuity of $\psi(\bullet;\xi)$ holds on $X$ when $\overline\psi(\bullet,z;\xi)$ is uniformly Lipschitz continuous over $(z;\xi) \in X\times\Xi$ and $\overline\psi(x,\bullet;\xi)$ is uniformly Lipschitz continuous over $(x;\xi) \in X\times\Xi$. Indeed, one can deduce this uniform Lipschitz continuity from Assumptions E and G with $\displaystyle\sup_{\xi\in \Xi} \left[\,\max(\kappa_1(\xi), \, \kappa_2(\xi) )\,\right]< \infty$. We are now ready to present the almost surely sequential convergence
of the objective values generated by the internal sampling scheme.
\begin{theorem}[Sequential convergence of objective values for Algorithm \ref{alg:sampling_based}]
Suppose that Assumptions A-G and conditions for the function $\psi$ in Lemma \ref{lem:ULLNrate} hold. Let $\{ x_{\nu}\}_{\nu \geq 0}$ be the sequence generated by Algorithm \ref{alg:sampling_based}. Assume that the parameter of the partial Moreau envelope $\gamma_\nu$ and the sample size $S_\nu$ satisfy
\[
\sum\limits_{\nu=1}^\infty \gamma_\nu < \infty, \qquad \sum\limits_{\nu=1}^\infty \frac {S_{\nu+1}-S_\nu} {S_{\nu+1} \, (S_\nu)^\eta} < \infty \quad \text{for some }\eta\in(0,1/2).
\]
Then
$\lim\limits_{\nu\rightarrow\infty}\overline\zeta_{S_\nu}(x_\nu) = \overline\zeta(\bar{x})$ almost surely,
where $\bar{x}$ is any accumulation point of the iterative sequence $\{x_{\nu}\}_{\nu \geq 0}$.
\end{theorem}
\begin{proof}
We first prove the almost sure convergence of $\left\{\overline\zeta_{S_\nu}(x_\nu)\right\}_{\nu \geq 0}$. We have
\begin{equation}\label{estimate2}
\begin{array}{ll}
\overline\zeta_{S_{\nu+1}}(x_{\nu+1})-\overline\zeta_{S_\nu}(x_\nu)
= &
\underbrace{\left[\, \overline\zeta_{S_{\nu+1}}(x_{\nu+1})-\overline\zeta_{S_\nu}(x_{\nu+1}) \right]}_{\triangleq \, R_{\nu,1}} +\underbrace{\Big[\, \overline\zeta_{S_\nu}(x_{\nu+1})-\widehat\zeta_{\,S_\nu,\gamma_\nu}(x_{\nu+1}) \Big]}_{\triangleq \, R_{\nu,2}}\\
&+\underbrace{\Big[\, \widehat\zeta_{\,S_\nu,\gamma_\nu}(x_{\nu+1})-\widehat\zeta_{\,S_\nu,\gamma_\nu}(x_\nu) \Big]}_{\triangleq \, R_{\nu,3}}+\underbrace{\Big[\, \widehat\zeta_{\,S_\nu,\gamma_\nu}(x_\nu)-\overline\zeta_{S_\nu}(x_\nu) \Big]}_{\triangleq \, R_{\nu,4}}.
\end{array}
\end{equation}
Using results of \eqref{eq:PMEerror} and \eqref{eq:Descenterror}, we obtain
\begin{align*}
R_{\nu,2} \leq \frac 1 {S_\nu} \sum\limits_{s=1}^{S_\nu} \frac {\gamma_\nu \, \kappa(\xi^s)^2} {2},\quad
R_{\nu,3} \leq -\frac 1 {2\gamma_\nu} \sum\limits_{i = 0}^{i_\nu}\|x_{\nu,i + 1}-x_{\nu,i}\|^2\quad \mbox{and}\quad
R_{\nu,4} \leq 0.
\end{align*}
Next we compute $R_{\nu,1}$ that is the error of the sample-augmentation. It holds that
\begin{align*}
R_{\nu,1}
&=\frac 1 {S_{\nu+1}}\left[\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)+\sum\limits_{s=S_\nu+1}^{S_{\nu+1}} \psi(x_{\nu+1};\xi^s)\right] - \frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)\\
&=\left(\frac {S_\nu} {S_{\nu+1}}-1\right) \frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s) + \frac 1 {S_{\nu+1}} \sum\limits_{s=S_\nu+1}^{S_{\nu+1}} \psi(x_{\nu+1};\xi^s)\\
&=\left(\frac {S_\nu} {S_{\nu+1}}-1\right) \left[\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)-\frac 1 {S_{\nu+1}-S_\nu}\sum\limits_{s=S_\nu+1}^{S_{\nu+1}} \psi(x_{\nu+1};\xi^s)\right].
\end{align*}
Let $\mathcal F_\nu\triangleq\sigma(\xi^1,\xi^2,\ldots,\xi^{S_\nu})$ be a filtration, i.e., an increasing sequence of $\sigma-$fields generated by samples used in outer iterations. Obviously $x_{\nu+1}$ is adapted to $\mathcal F_\nu$ and $\{\xi^s\}_{s=S_\nu+1}^{S_{\nu+1}}$ are independent of $\mathcal F_\nu$. Therefore, by taking conditional expectation of $R_{\nu,1}$ given $\mathcal F_\nu$, we obtain
\[
\mathbb{E}[\,R_{\nu,1}\mid \mathcal F_\nu\,]=\left(\frac {S_\nu} {S_{\nu+1}}-1\right) \left[\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)-\mathbb{E}_{\tilde\xi}\left[\, \psi(x_{\nu+1};\tilde\xi) \right]\right],
\]
where $\{\xi^s\}_{s=1}^{S_\nu}$ and $\tilde\xi$ are independent and identically distributed. Based on the estimations of the terms $R_{\nu,1},R_{\nu,2},R_{\nu,3}$ and $R_{\nu,4}$, we have, by taking conditional expectation of \eqref{estimate2} given $\mathcal F_\nu$,
\begin{equation}\label{eq:superMG}
\begin{split}
&\mathbb{E}\Big[\, \overline\zeta_{S_{\nu+1}}(x_{\nu+1})\mid\mathcal F_\nu\Big] - \overline\zeta_{S_\nu}(x_\nu) - \frac 1 {2\gamma_\nu} \sum\limits_{i = 0}^{i_\nu}\|x_{\nu,i + 1}-x_{\nu,i}\|^2\\
\leq &\; \frac {S_{\nu+1}-S_\nu} {S_{\nu+1}} \left|\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)-\mathbb{E}_{\tilde\xi}\left[\, \psi(x_{\nu+1};\tilde\xi)\right]\right| + \frac 1 {S_\nu} \sum\limits_{s=1}^{S_\nu} \frac {\gamma_\nu \, \kappa(\xi^s)^2} 2.
\end{split}
\end{equation}
In order to show the almost sure convergence of $\left\{ \overline\zeta_{s_\nu}(x_\nu) \right\}_{\nu \geq 0}$, we need to verify that the right side of the preceding inequality is summable over $\nu$ almost surely and the sequence $\left\{\overline\zeta_{S_\nu}(x_\nu)\right\}_{\nu \geq 0}$ is bounded below almost surely. We have
\begin{align*}
&\mathbb{E} \left[\, \sum\limits_{\nu=1}^\infty \left(\frac {S_{\nu+1}-S_\nu} {S_{\nu+1}}\right) \left|\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)-\mathbb{E}_{\tilde\xi}\left[\psi(x_{\nu+1};\tilde\xi)\right]\right| \,\right]\\
= \;&\sum\limits_{\nu=1}^\infty \frac {S_{\nu+1}-S_\nu} {S_{\nu+1}} \, \mathbb{E} \left[\, \left|\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)-\mathbb{E}_{\tilde\xi}\left[\, \psi(x_{\nu+1};\tilde\xi)\right]\right| \,\right]\\
\leq \;& \sum\limits_{\nu=1}^\infty \frac {S_{\nu+1}-S_\nu} {S_{\nu+1}} \; \mathbb{E} \left[\, \sup_{x\in X}\left|\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x;\xi^s)-\mathbb{E}_{\tilde\xi}\left[\, \psi(x;\tilde\xi)\right]\right| \,\right]\\
\leq \;&\sum\limits_{\nu=1}^\infty \frac {(S_{\nu+1}-S_\nu) \, C \sqrt{1-2\eta}} {S_{\nu+1} \, (S_\nu)^{\eta}} < \infty, \; \text{for some $\eta\in(0,1/2)$, by Lemma \ref{lem:ULLNrate}}.
\end{align*}
Hence, we derive that
\[
\sum\limits_{\nu=1}^\infty \left(\frac {S_{\nu+1}-S_\nu} {S_{\nu+1}}\right) \left|\frac 1 {S_\nu}\sum\limits_{s=1}^{S_\nu} \psi(x_{\nu+1};\xi^s)-\mathbb{E}_{\tilde\xi}\left[\psi(x_{\nu+1};\tilde\xi)\right]\right| < \infty \quad \text{almost surely.}
\]
Since $\{\gamma_\nu\}_{\nu=0}^\infty$ is assumed to be summable, we can obtain
\[
\mathbb{E}\left[\, \sum\limits_{\nu=0}^\infty \left(\frac 1 {S_\nu} \sum\limits_{s=1}^{S_\nu} \frac {\gamma_\nu \, \kappa(\xi^s)^2} 2\right) \,\right] = \sum\limits_{\nu=0}^\infty\mathbb{E} \left[\frac 1 {S_\nu} \sum\limits_{s=1}^{S_\nu} \frac {\gamma_\nu \, \kappa(\xi^s)^2} 2\right] = \sum\limits_{\nu=0}^\infty \frac {\gamma_\nu \, \mathbb{E}_{\tilde\xi}\left[\kappa(\tilde\xi)^2\right]} 2 < \infty.
\]
Consequently, $\displaystyle\sum\limits_{\nu=1}^\infty \left(\frac 1 {S_\nu} \sum\limits_{s=1}^{S_\nu} \frac {\gamma_\nu \, \kappa(\xi^s)^2} 2 \right) < \infty$ almost surely. We have thus proved that the right side of \eqref{eq:superMG} is summable over $\nu$ almost surely. Next, we show that $\left\{\overline\zeta_{S_\nu} (x_\nu)\right\}_{\nu \geq 0}$ is bounded below almost surely. To see this, note that
\[
\sup_{x\in X}\left|\overline\zeta_{S_\nu} (x)\right| \leq \sup_{x\in X}\left| \overline\zeta_{S_\nu} (x) - \zeta(x)\right| + \sup_{x\in X} \left|\zeta(x)\right|,
\]
where the first term converges to $0$ almost surely by the uniform LLN (c.f. \cite[Theorem 9.60]{shapiro2021lectures}) and the second one is bounded due to the continuity of $\zeta(x) = \varphi(x) + \mathbb{E}_{\xi}\left[ \, \psi(x;\xi) \, \right]$ on the compact set $X$. Therefore, there exists a constant $M$ such that $\overline\zeta_{S_\nu}(x_\nu)$ is bounded below by $M$ almost surely for any $\nu$. Applying Robbins-Siegmund nonnegative almost supermartingale convergence lemma (c.f.\ \cite[Theorem 1]{robbins1971convergence}), we have $\displaystyle\sum\limits_{\nu=1}^\infty \displaystyle\frac 1 {2\gamma_\nu} \sum\limits_{i = 0}^{i_\nu}\|x_{\nu,i + 1}-x_{\nu,i}\|^2 < \infty$ almost surely and the sequence $\left\{\overline\zeta_{S_\nu}(x_\nu)\right\}_{\nu \geq 0}$ converges almost surely. Finally,
let $\bar{x}$ be the limit of a convergent subsequence $\{x_{\nu}\}_{\nu\in\alpha}$. Using the uniform convergence of $\overline\zeta_{S_\nu}$ to $\zeta$ and the continuity of $\zeta$ on the compact set $X$, it follows from \cite[Proposition 5.1]{shapiro2021lectures} that $\overline\zeta_{S_{\nu}}(x_{\nu})$ converges to $\zeta(\bar{x})$ almost surely as $\nu(\in\alpha)\rightarrow\infty$. This argument, together with the convergence of the full sequence $\left\{ \overline\zeta_{S_\nu}(x_\nu) \right\}_{\nu \geq 0}$, completes the proof of this theorem.
\end{proof}
\section{Numerical experiments}
\label{sec:exp}
In this section, we present the numerical results of the proposed decomposition algorithm where the objective of the second-stage problem is a bilinear function in the first- and second-stage variables. The test example comes from the power system planning problem that is originated in \cite{louveaux1988optimal} and expanded in \cite{hellemo2018decision}. The uncertain information (scenarios are indexed by $s$) includes the demand $d_{js}$, the price of electricity $\pi_{js}$ in the location $j\in\mathcal J$, and the unit production cost $q_{is}$ of the power plant $i \in \mathcal I$. The probability of the $s$-th scenario is the linear combination of $|\mathcal G|$ given different distributions (each with probability $p_{sg}$) whose weights $\{x_g\}_{g\in\mathcal G}$ are determined as parts of
the first-stage decisions. The capacity of each power plant $\{x_i\}_{i\in\mathcal I}$ also needs to be determined in the first stage. The second-stage decisions are the production $z_{ijs}$ from the power plant $i$ to the location $j$ for each scenario $s$.
For each $s$, we can write the recourse function as
\[
\begin{array}{ll}
\psi(\{x_i\}_{i\in\mathcal I}, \{x_g\}_{g\in\mathcal G};d_s, \pi_s, q_s) = \displaystyle\operatornamewithlimits{minimum}_{l_z \leq z_s \leq u_z} \;\, S \sum\limits_{g\in\mathcal G} p_{sg}\,x_g \sum\limits_{i\in\mathcal I} \sum\limits_{j\in\mathcal J} (q_{is}-\pi_{js})\,z_{ijs}\\[0.2in]
\hskip 1.75in \mbox{subject to} \; \displaystyle\sum\limits_{j\in\mathcal J} z_{ijs}\leq x_{i},\; i\in\mathcal I;
\; \sum\limits_{i\in\mathcal I} z_{ijs} = d_{js},\; j\in\mathcal J,
\end{array}
\]
where the objective function linearly depends on the first-stage variables $\{x_g\}_{g\in {\cal G}}$.
The overall deterministic equivalent formulation to minimize the total cost is
\begin{equation*}\label{testprob}
\begin{split}
\displaystyle\operatornamewithlimits{minimize}_{l_x \leq x \leq u_x, l_z \leq z_s \leq u_z} &\;\, \sum\limits_{i\in\mathcal I} c_i \,x_i + \sum\limits_{g\in\mathcal G} c_g \,x_g + \sum\limits_{s=1}^S \left(\sum\limits_{g\in\mathcal G} p_{sg}\,x_g\right) \left[\sum\limits_{i\in\mathcal I} \sum\limits_{j\in\mathcal J} (q_{is}-\pi_{js})\,z_{ijs}\right]\\
\mbox{subject to} & \;\, \sum\limits_{i\in\mathcal I} c_{i}\,x_i+\sum\limits_{g\in\mathcal G} c_g\,x_g\leq B\;\,\text{(budget constraint)},\qquad\sum\limits_{g\in\mathcal G} x_g = 1,\\
& \;\, \sum\limits_{j\in\mathcal J} z_{ijs}\leq x_{i},\quad i\in\mathcal I,\;s=1,\ldots,S\quad\text{(capacity constraints)},\\
& \;\, \sum\limits_{i\in\mathcal I} z_{ijs} = d_{js},\quad j\in\mathcal J,\;s=1,\ldots,S\quad\text{(demand constraints)}.
\end{split}
\end{equation*}
In our experiments,
we set $|\mathcal I| = |\mathcal G| = 5$ and $|\mathcal J| = 8$. The box constraints of $x$ and $z_s$ are $[8, 15]^5 \times [0,1]^5$ and $[0,5]^{5 \times 8}$ for each $s = 1,2,\cdots,S$. The unit costs in the first-stage $\{c_i\}_{i\in\mathcal I}\cup\{c_g\}_{g\in\mathcal G}$ are independently generated from a uniform distribution on $[0, 5]$. For each scenario, $\{q_{is}\}_{i\in\mathcal I}, \{\pi_{js}\}_{j\in\mathcal J}$ and $\{d_{js}\}_{j\in\mathcal J}$ are generated from truncated normal distributions $\mathcal N(1,5^2)$ on $[2,4], [3,5]$ and $[2,5]$, respectively. To construct a set of probabilities $\displaystyle\cup_{s=1}^S\{p_{sg}\}_{g\in\mathcal G}$ satisfying $\sum_{s=1}^S p_{sg} = 1$, we first randomly generate $S\times|\mathcal G|$ values from uniform distributions on $[0,1]$, and then group every $|\mathcal G|$ values and normalizing them such that the sum of values in each group is $1$.
All the experiments are conducted in \textsc{Matlab} 2021b on a Intel Xeon workstation with sixteen 3.70 GHz processors and 128 GB of RAM.
\subsection{Fixed scenarios}
Since the text example with fixed scenarios is in fact a large-scale nonconvex quadratic problem, it can also be directly solved by off-the-shelf nonlinear programming solvers. We compare the performance of our proposed decomposition algorithm based on the partial Moreau envelope (DPME) with the interior-point-based solvers
Knitro \cite{byrd2006k} and IPOPT \cite{wachter2006implementation}, both of which run with linear solver MUMPS 5.4.1. The absolute and relative feasibility and optimality errors are computed according to the termination criteria of Knitro\footnote{Knitro user guide: https://www.artelys.com/docs/knitro/2\_userGuide/termination.html}.
The quantities $\text{KKT}_{\rm abs}$ and $\text{KKT}_{\rm rel}$ are defined as the max of absolute and relative feasibility and optimality errors, respectively. The initial points are chosen to be the same for all algorithms. Although this may not necessarily force all algorithms to converge to the same objective values, we do observe such a phenomenon in the experiments.
Further implementation details of these algorithms are provided below.\\[0.05in]
\underline{{\bf Knitro} (version 13.0.0)}: ``{\sl knitro\_qp}" function is called in our numerical experiments to solve nonconvex quadratic programs from the \textsc{Matlab} environment. We set ``hessopt = 0" to compute the exact Hessian in the interior point method instead of using the (L)BFGS approximations, as we have observed that the former choice is faster for all the problems tested here. We directly set ``convex = 0" to declare our problems are nonconvex so that the solver does not need to spend time on checking the convexity of the problems. We report the results based on three different settings:
\begin{enumerate}[leftmargin=*]
\item Knitro-direct: Set ``$\text{algorithm} = 1$" so that the direct solver is used to solve linear equations. For the termination options, the KKT relative and absolute tolerance is set to be $10^{-4}$ and $10^{-2}$, respectively.
, i.e., ``$\text{feastol} = \text{opttol} = 10^{-4}$'' and ``$\text{feastol\_abs} = \text{opttol\_abs} = 10^{-2}$''.
\item Knitro-CG-1: Set ``$\text{algorithm} = 2$" so that the KKT system is solved using a projected conjugate gradient method. Stopping criteria are the same as above.
\item Knitro-CG-2: All are the same as Knitro-CG-1 except that the KKT relative and absolute tolerance is set to be $10^{-6}$ and $10^{-3}$, respectively.
\end{enumerate}
\vskip 0.05in
\underline{{\bf IPOPT} (version 3.14.4)}:
Due to different scaling strategies and reformulations, the termination criteria of Knitro and IPOPT are not directly comparable.
We set ``ipopt.tol = $5 \times 10^{-2}$" in our experiments as we find the computed solutions based on this tolerance are about the same quality as those provided by Knitro. We also set ``ipopt.hessian\_constant = `yes' " to use exact Hessian in the interior point method, and have not adopted the (L)BFGS method for the same reason as mentioned above.\\[0.05in]
\underline{{\bf DPME}}: Each master problem for the first stage and the subproblem of the second stage are convex quadratic programs, which we have called Gurobi to solve. We compute the absolute (denoted as $\text{Feasabs}_\nu$) and relative (denoted as $\text{Feasrel}_\nu$) feasibility errors using the same way as Knitro. Let the overall objective value at the $\nu$-th outer loop be $\text{Obj}_\nu$. We terminate our algorithm if all of the conditions below are satisfied and then compute the absolute and relative KKT errors (as defined in Knitro) at the last iteration point:
\begin{equation}\label{eq:stopping}
\text{Feasabs}_\nu \leq 10^{-2}, \;\, \text{Feasrel}_\nu \leq 10^{-4} \;\, \mbox{and }\frac{|\text{Obj}_{\nu - 1} - \text{Obj}_\nu|}{\max\{1, |\text{Obj}_{\nu - 1}|\}} \leq 10^{-4}.
\end{equation}
Table \ref{tab:fixed} and Figure \ref{fig:fixed} summarize the performance of different algorithms when the number of scenarios $S$ varies from $1,000$ to $120,000$ over $100$ independent replications (the sizes of the deterministic equivalent problems are listed in Table \ref{table:problem size}). For each algorithm, we report the mean and the standard deviation of the total iteration numbers, the absolute and relative KKT errors, objective values, and the wall-clock time. We also conduct experiments for $S = 500,000$ over $10$ independent replications to demonstrate the scalability of our decomposition algorithm and put the results in Table \ref{tab:fixed}. One may find that for small-sized problems (such as when $S < 10,000$), the interior point method that is implemented by both Knitro and IPOPT can solve the problem faster than our DPME, which may be due to two reasons: one is that the gain of the decomposition cannot compensate for the overhead of the communication between the master problem and the subproblems; the other is that we have not used the second order information as in the interior point method. However, for the cases where $S$ is large (such as $S \geq 10,000$), the DPME is the fastest method and attains the smallest average objective values. While the computational time of IPOPT and Knitro-direct quickly exceeds the preset limit,
the computational time of the DPME scales approximately linearly in terms of the number of scenarios. Although the Knitro-CG-1 can solve all the problems within the time limit, it cannot produce solutions that have similar objective values with other methods. When we switch the solver to the Knitro-CG-2 setting, it could provide solutions with similar objective values as DPME but needs significantly longer computational time.
Therefore, we conclude that the DPME for fixed scenarios can significantly reduce the computational time for solving large-scale nonconvex two-stage SPs.
\begin{figure}
\centering
\begin{minipage}{0.55\textwidth}
\centering
\includegraphics[scale=0.3]{Experiment.png}
\vskip -0.2in
\captionof{figure}{\scriptsize Performance of all algorithms with different numbers of scenarios $S$ over $100$ independent replications. Shaded areas: the tubes between $10\%$ and $90\%$ quantiles of the running time; Solid lines: means of the running time.}
\label{fig:fixed}
\end{minipage}
\hskip 0.12in
\begin{minipage}{0.4\linewidth}
\centering
\footnotesize
\begin{tabular}{c c c}
\toprule
\multirow{2}{*}{$S$} &\multicolumn{2}{c}{problem sizes}\\[0.05in]
\cline{2-3} \\[-0.08in]
&{rows} &{columns}\\[0.03in]
\toprule
1,000 &93,022 &40,010\\[0.03in]
5,000 &465,022 &200,010\\[0.03in]
10,000 &930,022 &400,010\\[0.03in]
30,000 &2,790,022 &1,200,010\\[0.03in]
80,000 &7,440,022 &3,200,010\\[0.03in]
120,000 &11,160,022 &4,800,010\\[0.03in]
500,000 &46,500,022 &20,000,010\\[0.03in]
\toprule
\end{tabular}
\captionof{table}{\scriptsize Dimensions of test problems. ``Rows" stands for the number of constraints; ``Columns" stands for the number of variables.}
\label{table:problem size}
\end{minipage}
\vskip -0.2in
\end{figure}
\begin{sidewaystable}
\footnotesize
\setlength\tabcolsep{3pt}
\begin{center}
\begin{tabular}{c c c c c c}
\toprule
\multirow{2}{*}{$S$} &{iterations} &{KKT$_\text{abs}$} &{KKT$_\text{rel}$} &{objective values} &{time (in seconds)}\\
&{a $ | $ b $ | $ c $ | $ d $ | $ e} &{a $ | $ b $ | $ c $ | $ d $ | $ e} &{a $ | $ b $ | $ c $ | $ d $ | $ e} &{a $ | $ b $ | $ c $ | $ d $ | $ e} &{a $ | $ b $ | $ c $ | $ d $ | $ e}\\
\toprule
\multirow{2}*{1,000}
& 8$|$ 17$|$ 28$|$ 29$|$ 12
& 3.0-4$|$ 7.4-4$|$ 4.0-6$|$ 7.4-5$|$ 1.4-3
& 5.6-5$|$ 1.4-4$|$ 6.0-7$|$ 1.4-5$|$ 1.2-4
&\multirow{2}{*}{93.056$|$ 93.055$|$ {\bf 93.053}$|$ 93.055$|$ {\bf 93.053}}
& {\bf 1}$|$ 2$|$ 3$|$ 13$|$ 3\\
& (2$|$ 3$|$ 5$|$ 3$|$ 11)
& (1.6-4$|$ 1.1-3$|$ 1.5-6$|$ 1.1-5$|$ 1.2-3)
& (2.8-5$|$ 2.2-4$|$ 2.5-7$|$ 2.5-6$|$ 2.2-4)
&& (0$|$ 0$|$ 1$|$ 1$|$ 2)\\ \hline
\multirow{2}*{5,000}
& 9$|$ 16$|$ 42$|$ 31$|$ 11
& 3.7-4$|$ 2.4-4$|$ 4.0-6$|$ 5.6-5$|$ 1.2-3
& 6.9-5$|$ 4.9-5$|$ 7.0-7$|$ 1.1-5$|$ 8.9-5
&\multirow{2}{*}{93.041$|$ 93.140$|$ {\bf 93.038}$|$ 93.040$|$ {\bf 93.038}}
& 24$|$ 8$|$ 21$|$ 70$|$ {\bf 7}\\
& (2$|$ 2$|$ 4$|$ 3$|$ 8)
& (1.5-4$|$ 1.0-4$|$ 1.6-6$|$ 2.2-5$|$ 2.4-4)
& (2.1-5$|$ 2.2-5$|$ 2.6-7$|$ 4.5-6$|$ 5.0-5)
&& (1$|$ 1$|$ 2$|$ 7$|$ 4)\\ \hline
\multirow{2}*{10,000}
& 10$|$ 16$|$ 43$|$ 32$|$ 11
& 4.1-4$|$ 2.3-4$|$ 4.0-6$|$ 1.2-4$|$ 1.3-3
& 7.5-5$|$ 4.9-5$|$ 6.0-7$|$ 2.1-5$|$ 1.0-4
&\multirow{2}{*}{93.049$|$ 93.463$|$ {\bf 93.046}$|$ 93.059$|$ {\bf 93.046}}
& 108$|$ 17$|$ 45$|$ 139$|$ {\bf 12}\\
& (3$|$ 1$|$ 3$|$ 3$|$ 9)
& (1.2-4$|$ 1.1-4$|$ 1.5-6$|$ 6.4-4$|$ 6.3-4)
& (1.9-5$|$ 2.2-5$|$ 2.5-7$|$ 1.1-4$|$ 1.2-4)
&& (5$|$ 1$|$ 6$|$ 15$|$ 9)\\ \hline
\multirow{2}*{30,000}
& 10$|$ 16$|$ 47$|$ 35$|$ 11
& 3.7-4$|$ 2.3-4$|$ 3.0-6$|$ 4.0-4$|$ 1.3-3
& 6.3-5$|$ 5.0-5$|$ 5.0-7$|$ 7.2-5$|$ 9.8-5
&\multirow{2}{*}{93.068$|$ 98.336$|$ 93.061$|$ 93.076$|$ {\bf 93.060}}
& 1600$|$ 76$|$ 178$|$ 424$|$ {\bf 37}\\
& (4$|$ 2$|$ 5$|$ 3$|$ 8)
& (1.5-4$|$ 8.3-5$|$ 1.4-6$|$ 2.1-3$|$ 5.4-4)
& (2.4-5$|$ 1.8-5$|$ 2.4-7$|$ 3.8-4$|$ 1.0-4)
&& (63$|$ 4$|$ 21$|$ 38$|$ 25)\\ \hline
\multirow{2}*{80,000}
& --$|$ 14$|$ 46$|$ 35$|$ 10
& --$|$ 2.2-4$|$ 3.0-6$|$ 7.7-5$|$ 1.3-3
& --$|$ 4.6-5$|$ 5.0-7$|$ 1.6-5$|$ 9.1-5
&\multirow{2}{*}{--$|$ 122.644$|$ {\bf 93.078}$|$ 93.127$|$ {\bf 93.078}}
& {\bf t}$|$ 376$|$ 1135$|$ 1305$|$ {\bf 152}\\
& (--$|$ 4$|$ 8$|$ 4$|$ 8)
& (--$|$ 6.6-5$|$ 1.3-6$|$ 3.8-4$|$ 3.2-4)
& (--$|$ 1.4-5$|$ 2.3-7$|$ 8.1-5$|$ 6.5-5)
&& (--$|$ 28$|$ 181$|$ 177$|$ 108)\\ \hline
\multirow{2}*{120,000}
& --$|$ 11$|$ --$|$ --$|$ 11
& --$|$ 2.4-4$|$ --$|$ --$|$ 1.3-3
& --$|$ 5.1-5$|$ --$|$ --$|$ 1.1-4
&\multirow{2}{*}{--$|$ 133.885$|$ --$|$ --$|$ {\bf 93.101}}
& {\bf t}$|$ 856$|$ {\bf t}$|$ {\bf t}$|$ {\bf 260}\\
& (--$|$ 3$|$ --$|$ --$|$ 10)
& (--$|$ 8.3-5$|$ --$|$ --$|$ 5.8-4)
& (--$|$ 1.7-5$|$ --$|$ --$|$ 1.3-4)
&& (--$|$ 40$|$ --$|$ --$|$ 207) \\ \hline
\multirow{2}*{500,000}
& --$|$ 11$|$ --$|$ --$|$ 14
& --$|$ 2.0-4$|$ --$|$ --$|$ 1.5-3
& --$|$ 4.1-5$|$ --$|$ --$|$ 1.3-4
&\multirow{2}{*}{--$|$ 141.974$|$ --$|$ --$|$ {\bf 96.485}}
& {\bf t}$|$ 17464$|$ {\bf t}$|$ {\bf t}$|$ {\bf 1609}\\
& (--$|$ 2$|$ --$|$ --$|$ 14)
& (--$|$ 8.8-7$|$ --$|$ --$|$ 8.4-4)
& (--$|$ 3.0-6$|$ --$|$ --$|$ 1.6-4)
&& (--$|$ 553$|$ --$|$ --$|$ 1575) \\
\toprule
\end{tabular}
\caption{\scriptsize The performance of Knitro-direct, Knitro-CG, IPOPT, and DPME. In the table, ``$S$'' is the number of scenarios; the numbers without parentheses are the means over $100$ replications and the numbers in parentheses are the standard deviations; ``t" means the method exceeds the preset time limit, which is $1,800$s for $S \leq 120,000$ and $18,000$s for $S = 500,000$; ``$a$" stands for Knitro-direct; ``$b$" stands for Knitro-CG-1; ``$c$" stands for Knitro-CG-2; ``$d$" stands for IPOPT and ``$e$" stands for DPME.}
\label{tab:fixed}
\begin{tabular}{ c c c c c c c}
\toprule
\multirow{2}{*}{$\sigma$} & \multirow{2}{*}{$\eta$} & {iterations} & \multirow{2}{*}{KKT$_\text{abs}$} & \multirow{2}{*}{KKT$_\text{rel}$} & \multirow{2}{*}{objective values} & \multirow{2}{*}{time (s)}\\
&& {outer $|$ total} &&&&\\
\toprule
\multirow{3}*{0.5}
& 100
& 2 (0) $|$ 6 (5)
& 4.3-3 (1.7-2)
& 1.1-3 (4.6-3)
& 96.224
& 27 (14)\\
& 200
& 2 (0) $|$ 6 (5)
& 3.0-3 (1.3-2)
& 8.4-4 (3.8-3)
& 96.221
& 29 (17)\\
& 400
& 2 (0) $|$ 7 (8)
& 2.2-3 (1.1-2)
& 5.1-4 (2.6-3)
& 96.219
& 36 (31)\\
& 800
& 2 (0) $|$ 6 (5)
& 7.3-4 (4.3-3)
& 1.6-4 (9.1-4)
& 96.216
& 31 (18)\\
& --
& 2 (0.3) $|$ 6 (5)
& 3.8-4 (2.2-3)
& 9.1-5 (5.1-4)
& 96.215
& 43 (28)\\ \hline
\multirow{3}*{2}
& 100
& 2 (0) $|$ 8 (11)
& 4.3-2 (1.1-1)
& 8.3-3 (2.3-2)
& 93.463
& 33 (30)\\
& 200
& 2 (0) $|$ 7 (8)
& 8.1-3 (3.7-2)
& 1.5-3 (6.7-3)
& 93.406
& 29 (25)\\
& 400
& 2 (0) $|$ 7 (8)
& 4.0-3 (1.6-2)
& 7.3-4 (2.8-3)
& 93.399
& 33 (29)\\
& 800
& 2 (0) $|$ 6 (5)
& 2.6-3 (1.3-2)
& 4.9-4 (2.4-3)
& 93.399
& 33 (22)\\
& 1600
& 2 (0) $|$ 6 (5)
& 1.9-3 (8.6-3)
& 3.5-4 (1.6-3)
& 93.397
& 33 (20)\\
& --
& 2 (0) $|$ 6 (5)
& 2.4-4 (6.1-4)
& 4.3-5 (1.1-4)
& 93.395
& 44 (30)\\ \hline
\multirow{3}*{5}
& 100
& 2 (0) $|$ 13 (16)
& 1.1-1 (2.0-1)
& 1.7-2 (3.0-2)
& 92.440
& 46 (46)\\
& 400
& 2 (0) $|$ 11 (15)
& 2.4-2 (5.6-2)
& 3.7-3 (9.2-3)
& 92.285
& 51 (56)\\
& 1600
& 2 (0) $|$ 10 (14)
& 1.6-2 (3.1-2)
& 2.4-3 (4.6-3)
& 92.267
& 47 (51)\\
& 3200
& 2 (0) $|$ 9 (14)
& 1.4-2 (3.0-2)
& 2.1-3 (4.4-3)
& 92.264
& 46 (54)\\
& 6400
& 2 (0) $|$ 9 (14)
& 1.3-2 (3.0-2)
& 1.9-3 (4.4-3)
& 92.262
& 48 (58)\\
& 12800
& 2 (0) $|$ 9 (14)
& 1.3-2 (2.9-2)
& 1.8-3 (4.3-3)
& 92.261
& 55 (73)\\
& --
& 2 (0) $|$ 9 (14)
& 1.3-2 (3.0-2)
& 1.9-3 (4.3-3)
& 92.262
& 65 (83)\\
\toprule
\end{tabular}
\caption{\footnotesize The performance of the sampling-based DPME. In the table, ``$\sigma$" is the variance of the normal distribution from which we generate the data; ``$\eta$'' represents the linear growth rate of sample size such that the number of scenarios used in $\nu$-th outer iteration $S_\nu = \eta \nu$; ``--'' in the column of $\eta$ stands for the benchmark of DPME using full scenarios.}
\label{tab:sampling}
\end{center}
\end{sidewaystable}
\subsection{Sampling-based decomposition}
We test the sampling-based DPME proposed in Algorithm \ref{alg:sampling_based} for the same test problem with the total number of scenarios $S = 50,000$. Instead of using all scenarios at each iteration, we gradually add them to reduce the computational cost especially at the early stage. In our experiments, the sample size $S_\nu$ is taken as $ \eta \nu$ for different positive parameter $\eta$.
In order to understand how the growth rate $\eta$ depends on the distributions of the random scenarios, we generate $\{q_{is}\}_{i\in\mathcal I},\{\pi_{js}\}_{j\in\mathcal J}$ and $\{d_{js}\}_{j\in\mathcal J}$ from truncated normal distributions $\mathcal N(1,\sigma^2)$ on $[2,13], [3,10]$ and $[2,5]$ with $\sigma \in \{0.5, 2, 5\}$, and varies the values of $\eta$.
The stopping criteria for the sampling-based DPME is the same as \eqref{eq:stopping}, where we check the violation of the KKT system for the deterministic equivalent problem formulated by all scenarios $S = 50,000$. However, unlike the case for fixed scenarios, we do not have all the second-stage solutions $\{z_s\}_{s=1}^S$ to compute the KKT residual since some samples may not have been used yet. To resolver this issue, we compute all $\{z_s\}_{s=1}^S$ at every $\nu$-th outer iteration, and then estimate the multipliers corresponding to the first-stage budget and box constraints by minimizing the current KKT residual.
In Table \ref{tab:sampling}, we summarize the performance of Algorithm \ref{alg:sampling_based} for different combinations of $(\sigma, \eta)$, where we also provide the results obtained from Algorithm \ref{alg:finite_samples} without sampling for benchmarks. It can be observed from the table that problems with larger variability may need faster growth rate of the batch size to retain the same level of solution quality. If the growth rate is properly chosen, the sampling-based DPME can outperform the fixed-scenario version in the computational time with comparable solution qualities.
\section{Conclusion}
Compared with the extensive literature on the algorithms for convex (especially linear) two-stage SPs, efficient computational algorithms for solving continuous nonconvex two-stage SPs have been much less explored.
In this paper, we have made a first attempt on developing the decomposition scheme for a special class of latter problems.
The key of the proposed algorithm is the derivation of successive strongly convex approximations of the nonconvex recourse functions. We hope the work done in paper could stimulate researchers' interests in a broader paradigm of SPs that goes beyond the classical convex settings. There are a lot of open questions that deserve future investigations, such as
how to combine the stochastic dual dynamic programming approach with the tools developed in the current paper to solve nonconvex multistage SPs,
as well as how to design rigorous stopping criteria for the general nonconvex SPs with continuous distributions.
|
\section{Introduction}
The flavour eigenstates of neutral mesons oscillate. This yields an
experimentally precisely measurable mass and width difference between the CP
eigenstates of the $B^0_{(s)}$-$\overline{B}_{(s)}^0$-system. The most recent
experimental measurements for the mass differences $\Delta m_d^\mathrm{exp}$ and
$\Delta m_s^\mathrm{exp}$ for the $B_d-\overline{B}_d$ and the $B_s-\overline{B}_s$
systems, respectively are~\cite{LHCb:2021moh,Amhis:2019ckw}
\begin{equation}
\begin{aligned}
\Delta m^\mathrm{exp}_d &= (0.5065 \pm 0.0019) \,\mathrm{ps}^{-1}\,, \\
\Delta m^\mathrm{exp}_s & = (17.7656 \pm 0.0057) \,\mathrm{ps}^{-1}\,. \\
\end{aligned}
\end{equation}
These oscillations are mediated through the quark flow box diagrams displayed in
figure~\ref{fig:boxdiags}, where $q=d$ corresponds to the case of
$B_d^0-\overline{B}_d^0$ and $q=s$ to $B_s^0-\overline{B}_s^0$ mixing. Due to
the hierarchy of the CKM matrix and the large top-quark mass, the short-distance
contribution is top and CKM enhanced, whilst the long-distance contribution is
CKM suppressed, so that the short-distance contribution dominates. This feature
makes lattice QCD computations of the corresponding matrix elements feasible (as
opposed to e.g. $D^0$-$\overline{D}^0$ mixing, where lattice QCD predictions are
limited to the sub-leading short distance contribution,
see for example ref.~\cite{Bazavov:2017weg}).
There are 5 independent, parity even, local, dimension 6 operators $\mc{O}_i$ in
an effective Hamiltonian for $\Delta B=2$. In the SM, only the operator
$\mc{O}_1$ contributes to the mass difference $\Delta m_q$, whilst the operators
$\mc{O}_1$, $\mc{O}_2$ and $\mc{O}_3$ contribute to the width difference. The
operators $\mc{O}_4$ and $\mc{O}_5$ do not contribute in the Standard Model, but
by virtue of being a loop mediated process, neutral meson mixing is a sensitive
probe for New Physics. To account for possible beyond the Standard Model
scenarios, a precise determination of the non-perturbative matrix elements of
all five operators is desirable.
The mass differences $\Delta m_q$ can be parameterised in terms of known
functions, CKM factors and non-perturbative matrix elements which can be
computed via lattice QCD computations. More precisely
\begin{equation}
\Delta m_q = \frac{G_F^2 m_W^2 m_{B_q}}{6\pi^2} \abs{V_{tb}V_{tq}^*}^2 S_0(x_t) \eta_{2B} f_{B_q}^2 \hat{B}^{(1)}_{B_q}\,,
\end{equation}
where $q=d$ or $s$, $x_t = m_t^2/M_W^2$, $S_0(x_t)$ is an Inami-Lim
function~\cite{Inami:1980fz} and $\eta_{2B}$ captures short-distance QCD
corrections. Furthermore $f_{B_q}$ is the decay constant of the $B_q$ meson and
$\hat{B}^{(1)}_{B_q}$ is the renormalisation group independent bag parameter for
the operator $\mc{O}^{(1)}$. Precise non-perturbative predictions of
$f_{B_{(q)}}^2 \hat{B}^{(1)}_{B_q}$ , attainable via lattice QCD simulations,
enable the extraction of the combination of CKM matrix elements
$\abs{V_{tb}V_{tq}^*}$. Additionally, it is convenient to define the $SU(3)$
breaking ratio $\xi$~\cite{Bernard:1998dg}, defined by
\begin{equation}
\xi^2 = \frac{f^2_{B_s} \hat{B}^{(1)}_{B_s}}{f^2_{B_d} \hat{B}^{(1)}_{B_d}}\,.
\end{equation}
By combining $\xi$ with the experimental measurements of $\Delta m_d$ and
$\Delta m_s$, one can extract the ratio $\abs{V_{td}/V_{ts}}$ from
\begin{equation}
\abs{\frac{V_{td}}{V_{ts}}}^2 = \left(\frac{\Delta m_d}{\Delta
m_s}\right)_\mathrm{exp} \frac{m_{B_s}}{m_{B_d}}
\left(\xi^2\right)_\mathrm{lat}\,.
\end{equation}
This is favourable since various systematic and part of the statistical
uncertainties cancel in such a ratio. The results for $\abs{V_{td}}$,
$\abs{V_{ts}}$, and their ratio provide constraints which enter the global fits
produced by the UT fit~\cite{UTfit:2006vpt} and CKM-fitter~\cite{Charles:2004jd}
groups.
\begin{figure}
\begin{minipage}{6in}
\centering
\includegraphics[align=c,width=.47\textwidth]{box1.pdf}
\hspace*{.2in}
\includegraphics[align=c,width=.47\textwidth]{box2.pdf}
\end{minipage}
\caption{Quark flow diagrams of the box diagrams mediating neutral meson
mixing for the case of $B^0_{(s)}-\overline{B}^0_{(s)}$ mesons.}
\label{fig:boxdiags}
\end{figure}
\section{Neutral meson mixing results from lattice QCD}
The large bottom quark mass poses significant challenges for the direct
simulation of heavy-light systems. Currently all lattice simulations employ one
of the following two approaches. In the first approach the heavy quark is
simulated via an effective action such as static quarks, the non-relativistic
QCD action~\cite{Lepage:1992tx,HPQCD:2011qwj} or the Fermilab
method~\cite{Sheikholeslami:1985ij,El-Khadra:1996wdx}. In the second approach
simulations take place at unphysically light heavy quark masses. The results are
then extrapolated to the physical $b$-quark
mass~\cite{ETM:2013jap,Boyle:2018knm}. In addition, all current lattice QCD
results for neutral meson mixing include heavier-than-physical pion mass
ensembles. One leading uncertainty in current lattice QCD predictions stems from
the chiral-continuum extrapolation. This extrapolation depends strongly on the
available gauge field ensembles at or near the physical pion mass and gauge
field ensembles with fine lattice spacings.
Figure~\ref{fig:landscape} presents an overview over the lattice spacing and
pion mass properties of the the gauge field ensembles that are used in the most
recent computations~\cite{FermilabLattice:2016ipl, Boyle:2018knm,
Dowdall:2019bea, Boyle:2021kqn}. For the discussion of older
results~\cite{Gamiz:2009ku,ETM:2013jap,Aoki:2014nga}, we refer the reader to the
recent update of the flavour lattice averaging group (FLAG)~\cite{Aoki:2021kgd}.
The Fermilab/MILC result~\cite{FermilabLattice:2016ipl} has been discussed in
previous iterations of this conference, so we only briefly summarise the
computation. The light quarks are discretised using the asqtad
action~\cite{Orginos:1999cr}, the $b$-quark is discretised using the Fermilab
method~\cite{Sheikholeslami:1985ij,El-Khadra:1996wdx}. The authors compute the
product $f_{B_q} \sqrt{\hat{B}^{(i)}_{B_q}}$ for $q=d,s$ and for
$i=1,...,5$. The bag parameters $\hat{B}^{(i)}_{B_q}$ are then found by taking
external input for $f_{B_q}$. The relative uncertainties for $f^2_{B_q}
\hat{B}_{B_q}^{(1)}$ are approximately 8 and 6\% for $q=d,s$, respectively. The
relative uncertainty of $\xi$ is approximately 1.5\%.
The recent HPQCD result~\cite{Dowdall:2019bea} uses the HISQ
action~\cite{Follana:2006rc} for the light quarks and the NRQCD action for the
$b$-quark~\cite{Lepage:1992tx,HPQCD:2011qwj}. The calculation provides values
for the bag parameters $\hat{B}_{B_q}^{(i)}$ for $i=1,...,5$ and $q=d,s$. The
decay constants are then taken as an external input to convert this to $f_{B_q}
\sqrt{\hat{B}_{B_q}^{(i)}}$. The inclusion of two ensembles at the physical pion
mass significantly reduces uncertainties associated to the chiral
extrapolation. The dominant uncertainty in this work arises from matching terms
of the order $\alpha_s^2$ and $\alpha_s\Lambda_\mathrm{QCD}/m_b$. The quoted
uncertainties for $\hat{B}^{(1)}_{B_q}$ are 5.0 and 4.3\% for $q=d,s$,
respectively, whilst the uncertainty of their ratio is 2.5\%. For $\xi$ an
uncertainty of 1.3\% is quoted.
The recent work by RBC/UKQCD~\cite{Boyle:2018knm} follows a different approach
in the treatment of the heavy quark. Here the (to a good approximation) chirally
symmetric domain wall fermion
action~\cite{Kaplan:1992bt,Blum:1996jf,Shamir:1993zy,Brower:2012vk} is used for
light and heavy quarks, albeit with different choices for the domain wall
parameters for the light and the heavy quarks~\cite{Boyle:2015kyy}. The heavy
quarks are simulated in the region from below the physical charm quark mass up
to approximately half the bottom quark mass. Ref.~\cite{Boyle:2018knm} provides
results for the ratios $\hat{B}_{B_s}/\hat{B}_{B_d}$, $f_{B_s}/f_{B_d}$ and
$\xi$. Because these ratios have a favourable behaviour as a function of the
heavy quark mass, they allow a controlled extrapolation to the physical
$b$-quark mass. As in the case of the HPQCD result, the inclusion of two
ensembles with physical pion masses removes most uncertainties associated to the
chiral extrapolation. The dominant uncertainty in this work is the extrapolation
to the physical $b$-quark mass. However, by addition of ensembles with smaller
lattice spacings, this uncertainty can be systematically improved upon. The
total uncertainty for the quantity $\xi$ is quoted at the percent level.
\begin{figure}
\centering
\includegraphics[width=.45\textwidth]{landscape_all_pub.pdf}
\includegraphics[width=.45\textwidth]{landscape_JLQCD-RBC-UKQCD_fixedaxes.pdf}
\caption{Gauge field ensembles as a function of the lattice spacing squared
($a^2$) and the pion mass squared $m_\pi^2$ used in recent computations of
neutral meson mixing parameters. The black star corresponds to the physical
point to which results need to be extrapolated. \emph{Left}: The ensembles
entering the FNAL/MILC analysis~\cite{FermilabLattice:2016ipl}, the HPQCD
analysis~\cite{Dowdall:2019bea} and the RBC/UKQCD
analysis~\cite{Boyle:2018knm} are shown as green triangles, blue diamond and
red squares, respectively. \emph{Right}: In Ref.~\cite{Boyle:2021kqn} the
RBC/UKQCD dataset (red squares) is supplemented with a recently generated
JLQCD dataset (magenta squares). The gauge fields ensembles used in
refs~\cite{FermilabLattice:2016ipl,Boyle:2018knm,Boyle:2021kqn} include the
dynamical effects of two degenerate light quarks and the strange quark
($N_f=2+1$), whilst the ensembles used in ref.~\cite{Dowdall:2019bea} also
includes dynamical charm quark effects ($N_f=2+1+1$).}
\label{fig:landscape}
\end{figure}
\section{Summary and future prospects}
The three computations~\cite{FermilabLattice:2016ipl, Dowdall:2019bea,
Boyle:2018knm} highlighted above are based on completely disjoint gauge field
ensembles and are therefore independent predictions. These results are highly
complementary as all the discretisation of the light and the heavy quark action
differ between all three works. When comparing the numerical values obtained for
the bag parameters $B_{B_q}^{(i)}$, the product $f^2_{B_q} \hat{B}_{B_q}^{(1)}$,
and $\xi$, agreement between computations from the different groups can be seen.
Combining the lattice results with the experimentally measured mass differences
allows the extraction of $\abs{V_{td}}$, $\abs{V_{ts}}$ and
$\abs{V_{td}/V_{ts}}$ at the 2.6\%, 2.2\% and 1.3\%-level. However, in the
$\abs{V_{td}}-\abs{V_{ts}}$ plane, an approximately 2 sigma deviation remains
between the determinations from the different lattice results (see for example
figure~7 of ref.~\cite{Wingate:2021ycr}). Whilst the recent results are a
significant improvement, the uncertainties on the CKM matrix elements are still
dominated by the theory inputs, so additional work is required to improve upon
their accuracy.
Recently, work in progress was reported for a joint effort between the RBC/UKQCD
and the JLQCD collaborations~\cite{Boyle:2021kqn}. This supplements the existing
dataset from ref.~\cite{Boyle:2018knm} with additional ensembles with finer
lattice spacings provided by the JLQCD collaboration. The bound of only
simulating up to approximately half the physical $b$-quark mass was set due to
the lattice spacing on the finest ensemble~\cite{Boyle:2017jwu}. This can be
significantly extended by including the JLQCD ensembles, enabling simulations
close to the physical $b$-quark mass. The choice of the chirally symmetric
all-domain-wall fermion set-up significantly simplifies the renormalisation
structure and allows to draw on the expertise of the similar neutral kaon mixing
programme~\cite{Boyle:2017skn,Boyle:2017ssm}. Using this the authors anticipate
results for the full operator basis for the $B^0_d$ and the $B^0_s$ system in
addition to the $SU(3)$-breaking ratios presented in ref.~\cite{Boyle:2021kqn}.
This review focusses on the determination of the dimension 6 operators from
lattice QCD. Beyond this, in a recent computation~\cite{Davies:2019gnp} the
HPQCD collaboration presented the first lattice QCD computation of the matrix
elements of the dimension 7 operators which contribute to the width difference
at next to leading order. Finally since the last iteration a new computation of
the dimension 6 operators from sum rules has appeared~\cite{King:2019lal}.
\section*{Acknowledgements \label{sec:acknowledgements}}
J.T.T. would like to thank his collaborators, in particular M.~Della~Morte,
F.~Erben, J.~Flynn and O.~Witzel for critical proofreading of the draft. The
project leading to this application has received funding from the European
Union's Horizon 2020 research and innovation programme under the Marie
Sk{\l}odowska-Curie grant agreement No 894103.
\bibliographystyle{JHEP-notitle}
|
\section{Introduction}
In recent years, $q$-matroids, the $q$-analogue of matroids, have been intensively studied due to their connection to linear rank metric codes. They were first studied by Jurrius and Pellikaan in \cite{JP18}, who showed that an $\mathbb F_{q^m}$-linear rank metric code induces a $q$-matroid. It was shown later on that matrix linear rank metric codes induce a $q$-polymatroid, a generalization of $q$-matroids (see \cite{ GLJ21qpolymatroids,gorla2020rank}).
Since then, many results on $q$-(poly)matroids, and how they relate to rank metric codes, have been established, see for example \cite{cyclicflats,BCIJ21,BCJ21,GLJ21,GLJ21qpolymatroids,gorla2020rank,JPV21}.
Because of their $q$-analogue nature, many of the newly discovered properties of $q$-matroids turn out to be analogues of well established matroid theory results. See \cite{crapo1970foundations, Greene,oxley,RotaMatroid} for more information on matroid theory.
It has therefore been of interest to determine which notions and properties of matroid theory can be generalized to $q$-matroids.
In \cite{BCJ21}, the authors show that, similarly to matroids, there exists a variety of cryptomorphic definitions for $q$-matroids. In this paper, we define $q$-matroids via a rank function on the lattice of subspace of a finite dimensional vector space over a finite field, and occasionally use the flat cryptomorphism.
In \cite{BCIJ21}, Byrne and co-authors, define the notion of a characteristic polynomial for $q$-polymatroids and use it to establish a $q$-analogue of the Assmus-Mattson Theorem. Furthermore they show that the characteristic polynomial of a $q$-polymatroid induced by a linear rank metric code determines the weight distribution of the code.
Maps between $q$-matroids are defined and studied in \cite{GLJ21}, which allows the authors to consider $q$-matroids from a category theory perspective. They introduce the notions of weak and strong maps, which respectively respect the rank structure and the flat structure of $q$-matroids. Although those maps are defined in an analogous way than the weak and strong maps between matroids (see \cite{Heunen2018TheCO, RotaMatroid}), substantial differences occur when comparing categories of $q$-matroids with categories of matroids. In fact, the authors show that, unlike for categories of matroids, a coproduct does not always exist in the category of $q$-matroids with strong maps but always exists when the morphisms are linear weak maps.
In \cite{JPV21}, Johnsen and co-authors make the connection between matroids and $q$-matroids more apparent by showing that a $q$-matroid induces a matroid, called the projectivization matroid. Furthermore, they show that the lattice of flats of the $q$-matroid is isomorphic to the lattice of flats of its projectivization matroid. This allows them to express the generalized rank weights of an $\mathbb F_{q^m}$-linear rank metric code in terms of the Betti numbers of the dual of the projectivization matroid.
In this paper, we further investigate the construction introduced in \cite{JPV21} and use the projectivization matroid as a tool to study
maps between $q$-matroids and the characteristic polynomial of a $q$-matroid. We define and study properties of the projectivization matroid in Section 3.
In Section 4, we show that the projectivization map from a vector space to its projective space is a functor from the category of $q$-matroids with weak (resp.\ strong) maps to the category of matroids with weak (resp.\ strong) maps. We use the relation between those categories to show that strong maps between $q$-matroids are weak maps.
Although Section 4 shows how the projectivization matroid can be used in a more category theory approach, results therein are not used in later sections. The reader interested in the relation between the characteristic polynomial of a $q$-matroid and that of its projectivization matroid may skip Section 4 on a first reading.
We then proceed in Section 5 to study the characteristic polynomial of $q$-matroids. We start by showing that the characteristic polynomial is identically $0$ if the $q$-matroid contains a loop, and is fully determined by the lattice of flats otherwise. We use this fact to show that the characteristic polynomial of a $q$-matroid is equal to that of its projectivization matroid. This in turn, allows us to find a recursive formula for the characteristic polynomial of a $q$-matroid in terms of the characteristic polynomial of its minors.
Finally, in Section 6, we consider the projectivization matroid of $q$-matroids induced by an $\mathbb F_{q^m}$-linear rank metric code. In \cite{ABNR21}, Alfarano et.\ al.\ associate a linear block code with the Hamming metric to an $\mathbb F_{q^m}$-linear rank metric code $\mathcal{C}$. This code, called a Hamming-metric code associated to $\mathcal{C}$, induces a matroid that turns out to be equivalent to the projectivization matroid of the $q$-matroid associated with $\mathcal{C}$. This connection allows us to prove a $q$-analogue of the critical theorem in terms of $\mathbb F_{q^m}$-linear rank metric codes and $q$-matroids.
\vspace{.2cm}
\underline{\textbf{Notation:}}
Throughout $\mathbb F_q$ denotes a finite field of order $q$. $E$ denotes a finite dimensional vector space over $\mathbb F_q$, and $\mathcal{L}(E)$ denotes the lattice of subspace of $E$. $S$ and $T$ are finite sets, $2^{S}$ is the power set of $S$ and $[n] := \{1, \ldots, n\}$ for $n \in \mathbb N_0$. Furthermore given a set $S$ and $A \subseteq S$, let $S-A := \{e \in S \, : \, e \notin A\}$.
Finally, $q$-matroids will be denoted by the script letters $\mathcal M$, $\mathcal N$, whereas matroids will be denoted by the capital letters $M, N$.
\section{Basic Notions of Matroids and $q$-Matroids
In this section we review well-known notions of matroids and $q$-matroids that will be used throughout the paper. For more details about matroids and $q$-matroids the reader may refer to \cite{BCJ21, JP18, oxley, RotaMatroid}.
\begin{definition}
A \emph{matroid} is an ordered pair $M = (S, r)$, where $S$ is a finite set and $r$ is a function $r : 2^S \rightarrow \mathbb N_0$ such that for all $A, B \in 2^{S}$ :
\begin{enumerate}
\item[(R1)] Boundedness: $0 \leq r(A) \leq |A|$.
\item[(R2)] Monotonicity: If $A \subseteq B$ then $r(A) \leq r(B)$.
\item[(R3)] Submodularity: $r(A \cup B) + r(A \cap B) \leq r(A) + r(B)$.
\end{enumerate}
$S$ is called the \emph{groundset} of $M$ and $r$ its \emph{rank function}.
\end{definition}
Throughout, identify $\{e\}$ with $e$ and $\{v\}$ with $v$. Two matroids $M = (S, r_M)$ and $N = (T, r_N)$ are \emph{equivalent}, denoted $M \cong N$, if there exists a bijection between the groundsets, $\psi : S \rightarrow T$, such that $r_M(A) = r_N(\psi(A))$ for all $A \subseteq S$.
Given a matroid $M = (S,r)$, $e \in S$ is a \emph{loop} of $M$ if $r(e) = 0$. $M$ is said to be \emph{loopless} if it does not contain any loops.
A subset $F \subseteq S$ is a \emph{flat} if $r(F \cup v) > r(F)$ for all $v \notin F$. It is well known that the collection of flats, denoted $\mathcal{F}_M$, forms a geometric lattice.
For any $F_1, F_2 \in \mathcal{F}_M$, the meet and join are defined as follow $F_1 \wedge F_2 := F_1 \cap F_2$ and $F_1 \vee F_2 := \mathrm cl_M(F_1 \cup F_2)$, where $\mathrm cl_M(A) = \{v \in S \, :\, r(A \cup v) = r(A)\} = \bigcap \{F \in \mathcal{F}_{M} \, : \, A \subseteq F\}$.
Given $F_1, F_2 \in \mathcal{F}_{M}$, we say $F_2$ \emph{covers} $F_1$ if for all $F \in \mathcal{F}_M$ such that $F_1 \subseteq F \subseteq F_2$ then $F = F_1$ or $F = F_2$.
When discussing $\mathcal{F}_M$, we interchangeably use the terms collection of flats and lattice of flats.
The flats of a matroid satisfy three axiomatic properties that fully determine the matroid.
\begin{proposition}\label{flats matroid}\cite[Sec. 1.4 Prob 11.]{oxley}
Let $M = (S, r_M)$ be a matroid and $\mathcal{F}_M$ its collection of flats. Then $\mathcal{F}_M$ satisfies the following:
\begin{enumerate}
\item[(F1)] $S \in \mathcal{F}_M$.
\item[(F2)] If $F_1, F_2 \in \mathcal{F}_M$ then $F_1 \cap F_2 \in \mathcal{F}_M$.
\item[(F3)] Let $F \in \mathcal{F}_M$ and $v \notin F$, then there exists a unique $F' \in \mathcal{F}_M$ covering $F$ such that $F \cup v \subseteq F'$.
\end{enumerate}
Furthermore, $r_M$ is uniquely determined by $\mathcal{F}_M$ and $r_M(A) = h(\mathrm cl_M(A))$ for $A \subseteq S$, where $h(F)$ denotes the height of $F$ in the lattice $\mathcal{F}_M$.
\end{proposition}
We now turn to $q$-matroids, which are defined in an analogous way. Recall that $E$ denotes a finite dimensional vector space over $\mathbb F_q$ and $\mathcal{L}(E)$ is the collection of subspace of $E$.
\begin{definition}
A \emph{$q$-matroid} is an ordered pair $\mathcal M = (E, \rho)$, where $\rho$ is a function $\rho : \mathcal{L}(E) \rightarrow \mathbb N_0$ such that for all $V, W \in \mathcal{L}(E)$ :
\begin{enumerate}
\item[(qR1)] Boundedness: $0 \leq \rho(V) \leq \dim(V)$.
\item[(qR2)] Monotonicity: If $V \subseteq W$ then $\rho(V) \leq \rho(W)$.
\item[(qR3)] Submodularity: $\rho(V + W) + \rho(V \cap W) \leq \rho(V) + \rho(W)$.
\end{enumerate}
$E$ is called the \emph{groundspace} of $\mathcal M$ and $\rho$ its \emph{rank function}.
\end{definition}
Two $q$-matroids $\mathcal M = (E_1, \rho_{\mathcal M})$ and $\mathcal N = (E_2, \rho_{\mathcal N})$ are \emph{equivalent}, denoted $\mathcal M \cong \mathcal N$, if there exists a linear isomorphism $\psi: E_1 \rightarrow E_2$ such that $\rho_{\mathcal M}(V) = \rho_{\mathcal N}(\psi(V))$ for all $V \leq E_1$.
Given a $q$-matroid $\mathcal M = (E, \rho)$ we say $\langle e \rangle$, where $e \in E - \{0\}$, is a \emph{loop} if $\rho(\langle e \rangle) = 0$, and $\mathcal M$ is \emph{loopless} if it does not contain any loops.
A subspace $F \leq E$ is a \emph{flat} of $\mathcal M$ if $\rho(F + \langle v \rangle) > \rho(F)$ for all $v \notin F$.
Furthermore the collection of flats of a $q$-matroid, denoted $\mathcal{F}_{\mathcal M}$, forms a geometric lattice as well. The meet and join operation are given by $F_1 \wedge F_2 := F_1 \cap F_2$ and $F_1 \vee F_2 := \mathrm cl_{\mathcal M}(F_1 + F_2)$, where $\mathrm cl_{\mathcal M}(V) := \{v \, : \, \rho(V + \langle v \rangle) = \rho(V)\} = \bigcap \{F \in \mathcal{F}_{\mathcal M} \, : \, V \leq F\}$.
The notion of cover for the lattice of flats of $q$-matroids is identical to that of matroids.
The collection of flats of a $q$-matroid also satisfies three axiomatic properties that fully determine the $q$-matroid.
\begin{proposition}\cite[Thm 48]{BCJ21}\label{flat $q$-matroid}
Let $\mathcal M = (E, \rho_{\mathcal M})$ be a $q$-matroid and $\mathcal{F}_{\mathcal M}$ be its collection of flats. Then $\mathcal{F}_{\mathcal M}$ satisfies the following:
\begin{enumerate}
\item[(qF1)] $E \in \mathcal{F}_{\mathcal M}$.
\item[(qF2)] If $F_1, F_2 \in \mathcal{F}_{\mathcal M}$ then $F_1 \cap F_2 \in \mathcal{F}_{\mathcal M}$.
\item[(qF3)] Let $F \in \mathcal{F}_{\mathcal M}$ and $e \notin F$, then there exists a unique $F' \in \mathcal{F}_{\mathcal M}$ covering $F$ such that $F + \langle e \rangle \leq F'$.
\end{enumerate}
Furthermore, $\rho_{\mathcal M}$ is uniquely determined by $\mathcal{F}_{\mathcal M}$ and $\rho_{\mathcal M}(V) = h(\mathrm cl_{\mathcal M}(V))$ for all $V \leq E$, where $h(F)$ denotes the height of $F$ in the lattice $\mathcal{F}_{\mathcal M}$. Thus, we may denote this $q$-matroid as $\mathcal M = (E, \mathcal{F}_{\mathcal M})$.
\end{proposition}
For both matroids and $q$-matroids, there exists a notion of duality, defined respectively with complements of sets and orthogonal spaces.
\begin{definition}\label{dual matroid}
Let $M = (S, r)$ be a matroid. The \emph{dual matroid} $M^* = (S, r^*)$ is defined via the rank function
$$r^*(A) = |A| - r(S) + r(S - A).$$
\end{definition}
Duality for $q$-matroids depends on a choice of non-degenerate symmetric bilinear form (NSBF).
Let $E$ be a vector space over $\mathbb F_q$ and $\langle \cdot , \cdot \rangle : E \times E \rightarrow \mathbb F_q$ be a NSBF. The orthogonal space of $V \leq E$ w.r.t $\langle \cdot , \cdot \rangle$ is the space
$V^{\perp} := \{ w \in E \, : \, \langle w, v \rangle = 0 \, \textup{ for all } v \in V\}.$
\begin{definition}\label{dual $q$-matroid}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and $\langle \cdot , \cdot \rangle$ be a NSBF on $E$. The \emph{dual $q$-matroid} $\mathcal M^* = (E, \rho^*)$, w.r.t the chosen NSBF, is defined via the rank function
$$\rho^*(V) = \dim(V) - \rho(E) + \rho(V^{\perp}).$$
\end{definition}
It was shown in \cite[Thm 2.8]{GLJ21qpolymatroids} that given two NSBFs $\langle \cdot , \cdot \rangle_1$, $\langle \cdot , \cdot \rangle_2$, the respective dual $q$-matroids $\mathcal M^{*_1}$ and $\mathcal M^{*_2}$ of $\mathcal M$ are equivalent. For both matroids and $q$-matroids, an element $v$ of the groundset, respectively groundspace, is a \emph{coloop} if $e$ is loop in the dual matroid, respectively dual $q$-matroid.
We now define the operations of deletion and contraction for matroids and $q$-matroids.
\begin{definition}
Let $M = (S,r)$ be a matroid and let $A \subseteq S$.
\begin{itemize}
\item The matroid $M\setminus A = (S - A, r_{M\setminus A}),$ where $r_{M\setminus A}(B) = r(B)$ for all $B \subseteq S-A$, is called the \emph{deletion} of $A$ from $M$.
\item The matroid $M/A = (S - A, r_{M/A}),$ where $r_{M/A}(B) = r(B \cup A) - r(A)$ for all $B \subseteq S-A$, is called the \emph{contraction} of $A$ from $M$.
\end{itemize}
\end{definition}
The following well-known facts about the deletion and contraction of matroids will be needed. Refer to \cite[Prop 3.1.25]{oxley} for a proof.
\begin{proposition}\label{del cont commute}
Let $M = (S, r)$ be a matroid. Let $A, B \subseteq S$ be disjoint sets. Then
\begin{itemize}
\item $(M \setminus A) \setminus B = M \setminus (A \cup B) = (M \setminus B) \setminus A,$
\item $(M/A)/B = M/(A \cup B) = (M / B)/ A,$
\item $(M \setminus A) / B = (M / B) \setminus A.$
\end{itemize}
\end{proposition}
To avoid the surplus of parenthesis, we omit them if there is no risk of confusion.
At this point we make a brief comment about notation. The notation $\setminus$ always denotes the deletion operation and set exclusion is denoted by the $-$ sign. However, the notation $/$ is used to denote both the contraction of ($q$-)matroids and quotient space (i.e $E/V$). The reader should therefore use context in order to differentiate between the latter two.
For $q$-matroids, the operations of deletion and contraction are defined in an analogous way.
\begin{definition}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid, let $V \leq E$ and fix a NSBF on $E$. Furthermore let $\pi : E \rightarrow E/V$ be the canonical projection.
\begin{itemize}
\item The $q$-matroid $M \setminus V = (V^{\perp}, \rho_{M\setminus V})$, where $\rho_{M\setminus V}(W) = \rho(W)$ for all $W \leq V^{\perp}$ is called the \emph{deletion} of $V$ from $\mathcal M$.
\item The $q$-matroid $M/V = (E/V, \rho_{M/V})$, where $\rho_{M/V}(W) = \rho(\pi^{-1}(W)) - \rho(V)$ for all $W \leq E/W$, is called the \emph{contraction} of $V$ from $\mathcal M$.
\end{itemize}
\end{definition}
It is worth noting that for both matroids and $q$-matroids, deletion and contraction are dual operations, i.e. $\mathcal M^* \setminus V \cong (\mathcal M / V)^*$ (equality rather than equivalence holds for matroids only). A proof of this fact for $q$-matroids can be found in \cite[Thm 5.3]{GLJ21qpolymatroids} and in \cite[Sect. 3]{oxley} for matroids.
A matroid $N$ (resp.\ $q$-matroid $\mathcal N$) is a \emph{minor} of $M$ (resp.\ $\mathcal M$) if it can be obtained from $M$ (resp.\ $\mathcal M$) by a sequence of deletion and contraction.
For both matroids and $q$-matroids, the flats of a contraction can be characterized in terms of the flats of the original ($q$-)matroid.
\begin{proposition}\label{contraction loop}
Let $M = (S, r_M)$ be a matroid, $\mathcal M = (E, \rho_{\mathcal M})$ be a $q$-matroid and $\mathcal{F}_M$, $\mathcal{F}_{\mathcal M}$ their respective lattice of flats. Let $A \subseteq S$, $V \leq E$ and consider $M/A$ and $\mathcal M/V$. Then
\begin{itemize}
\item[(1)] $\mathcal{F}_{M/A} = \{F \subseteq S - A \, : \, F \cup A \in \mathcal{F}_M\}$,
\item[(2)] $\mathcal{F}_{\mathcal M/V} = \{F \, : \, \pi^{-1}(F) \in \mathcal{F}_{\mathcal M}\}$, where $\pi : E \rightarrow E/V$.
\end{itemize}
Furthermore $A$ (resp.\ $V$) is a flat of $M$ (resp.\ $\mathcal M$) if and only if $M/A$ (resp.\ $\mathcal M/V$) is loopless.
\end{proposition}
\begin{proof}
(1) is shown in \cite[Prop 3.3.7]{oxley}. For (2), first let $F \in \mathcal{F}_{\mathcal M/V}$ and consider the space $W: = \pi^{-1}(F) \leq E$. Let $x \notin W$. Then
$\rho_{\mathcal M}(W \oplus \langle x \rangle) = \rho_{\mathcal M/V}( F \oplus \langle \pi(x) \rangle) + \rho_{\mathcal M}(V) > \rho_{\mathcal M/V} (F) + \rho_{\mathcal M}(V) = \rho_{\mathcal M}(W)$, where the inequality holds because $F \in \mathcal{F}_{\mathcal M/V}$ and $\pi(x) \notin F$. Since this is true for all $x \notin W$ then $W \in \mathcal{F}_{\mathcal M}$.
Now let $F \leq E/V$ such that $\pi^{-1}(F) \in \mathcal{F}_{\mathcal M}$. Let $\langle x \rangle \leq F/V$ such that $x \notin F$. Then $\rho_{\mathcal M/V}(F \oplus \langle x \rangle) = \rho_{\mathcal M}(\pi^{-1}(F \oplus \langle x \rangle)) - \rho_{\mathcal M}(V) = \rho_{\mathcal M}(\pi^{-1}(F) + \pi^{-1} (\langle x \rangle)) - \rho_{\mathcal M}(V) > \rho_{\mathcal M}(\pi^{-1}(F)) - \rho_{\mathcal M}(V) = \rho_{\mathcal M/V}(F)$. Once again, since this is true for all $x \notin F$ then $F \in \mathcal{F}_{\mathcal M/V}$.
We show the second part of the statement for matroids, and note the proof for $q$-matroid is analogous to it. Consider $M/A$ with $A \in \mathcal{F}_M$ and let $e \in S - A$. Then $r_{M/A}(e) = r_M(e \cup A) - r_M(A) > 0$ since $A$ is a flat. Since this holds for all $e \in S-A$, then $M/A$ is loopless. Now assume $A \notin \mathcal{F}_M$ then $A \subsetneq \mathrm cl_M(A)$ and let $e \in \mathrm cl_M(A) - A \subseteq S - A$. Then $r_{M/A}( e) = r_M(A \cup e) - r(A) = 0$ since $e \in \mathrm cl_M(A).$ Hence $M/A$ contains a loop.
\end{proof}
The last matroid operation we discuss is that of the single element extension by adjoining a loop, which we refer to as loop extension.
The loop extension will play an important role in section 4 when defining maps between matroids.
The reader can refer to \cite[Sect.~7.2]{oxley} and \cite[Chap.~8]{RotaMatroid} for proofs and a more detailed discussion of the single element extension.
\begin{proposition}\label{loop extension}
Let $M = (S, r)$ be a matroid and $\{o_M\}$ denotes a symbol disjoint from $S$. Let $S_o := S \cup \{o_M\}$ and $r_o : 2^{S_o} \rightarrow \mathbb N_0$ be such that $r_o(A) = r(A - \{o_M\})$, for all $A \subseteq S_o$. Then $M_o := (S_o, r_o)$ is a matroid, and $\{o_M\}$ is a loop in $M_o$. Furthermore $M_o$ is called a \emph{loop extension of $M$}.
\end{proposition}
The subscript of the added loop may be omitted if it is clear from context in which matroid the loop is contained. The next proposition relates the flats $\mathcal{F}_{M_o}$ and $\mathcal{F}_M$. Furthermore, we recall that two lattices are \emph{isomorphic} (denoted by $\cong$) if there exists an order preserving bijection between the lattices that preserves meets and joins.
\begin{proposition}\label{flat loop extension}
Let $M$ be a matroid, $M_o$ a loop extension of $M$, and $\mathcal{F}_M$, $\mathcal{F}_{M_o}$ their respective collection of flats. Then
$$\mathcal{F}_{M_o} = \{ F \cup \{o\} \, : \, F \in \mathcal{F}_M\}.$$
and $\mathcal{F}_M \cong \mathcal{F}_{M_o}$ as lattices.
\end{proposition}
\begin{remark}\label{loop/emptyset}
Note that $M_o \setminus \{o\} = M$. This deletion can be seen as identifying the element $\{o\}$ with the empty set of $M$, and does not change the overall structure of the matroid. \end{remark}
\section{The Projectivization Matroid
In \cite{JPV21}, Johnsen and co-authors showed that a $q$-matroid $\mathcal M$ with groundspace $E$ induces a matroid $P(\mathcal M)$ with groundset the projective space of $E$. This induced matroid, called the projectivization matroid of $\mathcal M$ turns out to be an interesting object to study. In fact, it was shown in that same paper, that the projectivization preserves the flat structure of $\mathcal M$. It therefore becomes a useful tool when studying properties of $q$-matroids that depend only on flats.
For completeness, we reintroduce the construction of the projectivization matroid.
The following notation will be used. Given a finite dimensional vector space $E$ over $\mathbb F_q$, let $\mathbb{P}E := \{ \langle v \rangle_{\mathbb F_q} \, : v \in E - \{0\}\}$ be the \emph{projective space of $E$}.
The map, $\hat{P} : (E - \{0\}) \rightarrow \mathbb{P}E \, , \, v \mapsto \langle v \rangle_{\mathbb F_q}$ induces a lattice map $P: \mathcal{L}(E) \rightarrow 2^{\mathbb{P}E}$, where $P(\{0\}) = \emptyset$ and $P(V) = \{\hat{P}(v) \, : \, v \in V -\{0\}\} = \{P(\langle v \rangle) \, : \, v \in V- \{0\}\}$
for $V \leq E$.
We call the lattice map $P$ the \emph{projectivization map}. Usually, $\hat{P}$ is called the projectivization map, however for our purposes, it is more convenient to consider the projectivization as a lattice map. Note that $P$ is inclusion preserving and that $P(V \cap W) = P(V) \cap P(W)$ for all $V, W \in \mathcal{L}(E)$.
For any $S \subseteq \mathbb{P}E$ let $P^{-1}(S) := \{v \in E \, : \, \hat{P}(v) \in S\} = \{v \in E \, : \, P(\langle v \rangle) \in S\}$. Note that $(P^{-1} \circ P)(V) = V$ for all $V \leq E$.
Finally let $\langle S \rangle := \langle P^{-1}(S) \rangle_{\mathbb F_q}$ for any $S \subseteq \mathbb{P}E$. We say $S \subseteq \mathbb{P}E$ \emph{contains a basis} of $E$ if $\langle S \rangle = E$.
We can now introduce the projectivization matroid.
\begin{theorem}(\cite[Def.14, Prop. 15]{JPV21})\label{$q$-matroid/matroid}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and let $r: 2^{\mathbb{P}E} \rightarrow \mathbb N_0$ such that for all $S \subseteq \mathbb{P}E, $
$$r(S) = \rho(\langle S \rangle).$$
Then $P(\mathcal M) := (\mathbb{P}E, r)$ is a matroid, and is called the \emph{projectivization matroid of $\mathcal M$}.
\end{theorem}
We now turn towards the relation between the flats of a $q$-matroid $\mathcal M$ and those of its projectivization matroid $P(\mathcal M)$. In the following result, the meet and join refers to those of the lattice of flats defined in Section 1.
\begin{lemma}\cite[Lem.16, Prop.21]{JPV21}\label{flats properties}
Let $\mathcal M$ be a $q$-matroid, $P(\mathcal M)$ its projectivization matroid, and $\mathcal{F}_{\mathcal M}$, $\mathcal{F}_{P(\mathcal M)}$ their respective lattice of flats. Furthermore, let $P(\mathcal{F}_{\mathcal M}):= \{P(F) \, : \, F \in \mathcal{F}_{\mathcal M}\}$. Then the following hold:
\begin{itemize}
\item[1)] $\mathcal{F}_{P(\mathcal M)} = P(\mathcal{F}_{\mathcal M})$.
\item[2)] $P(F_1 \vee F_2) = P(F_1) \vee P(F_2)$ and $P(F_1 \wedge F_2) = P(F_1) \wedge P(F_2)$, for all $F_1, F_2 \in \mathcal{F}_{\mathcal M}$.
\end{itemize}
Therefore $\mathcal{F}_{P(\mathcal M)} \cong \mathcal{F}_{\mathcal M}$ as lattices.
\end{lemma}
The next result shows when a matroid with groundset $\mathbb{P}E$ is the projectivization matroid of a $q$-matroid with groundspace $E$.
\begin{theorem}\label{flat iso}
Let $M = (\mathbb{P}E, r)$ be a matroid and $\mathcal{F}_M$ its lattice of flats. Furthermore let $P^{-1}(\mathcal{F}_M) := \{P^{-1}(F) \cup \{0\} \, : \, F \in \mathcal{F}_M\}$. If $P^{-1}(F) \cup \{0\}$ is a subspace of $E$ for all $F \in \mathcal{F}_M$, then $\mathcal M = (E, P^{-1}(\mathcal{F}_M))$ is $q$-matroid.
Furthermore $\mathcal{F}_M \cong \mathcal{F}_{\mathcal M}$.
\end{theorem}
\begin{proof}
We show $\mathcal{F}:= P^{-1}(\mathcal{F}_M)$ is a collection of flats of a $q$-matroid by showing it satisfies ($q$F1)-($q$F3) of Proposition \ref{flat $q$-matroid}. Throughout the proof we use the fact that $\mathcal{F}_M$ is the collection of flats of a matroid, and hence satisfies (F1)-(F3) of Proposition \ref{flats matroid}.
Since $\mathcal{F}_M$ satisfies (F1), $\mathbb{P}E \in \mathcal{F}_M$, and therefore $P^{-1}(\mathbb{P}E) \cup \{0\} = E \in \mathcal{F}$. This shows (qF1).
Let $V_1 := P^{-1}(F_1)\cup \{0\}, V_2 := P^{-1}(F_2) \cup \{0\} \in \mathcal{F}$. Since $F_1, F_2 \in \mathcal{F}_M$ then $F_1 \cap F_2 \in \mathcal{F}_M$.
Furthermore, $ P(V_1 \cap V_2) = P(V_1) \cap P(V_2) = F_1 \cap F_2 \in \mathcal{F}_M$. Hence $V_1 \cap V_2 = P^{-1}(F_1 \cap F_2) \cup \{0\} \in \mathcal{F}$, showing (qF2).
Finally for (qF3), let $V := P^{-1}(F)\cup \{0\} \in \mathcal{F}$ and $w \notin V$. Since $P$ is inclusion preserving $P(\langle w \rangle) \notin P(V) = F$. Hence there exists a unique flat $F' \in \mathcal{F}_M$ covering $F$ such that $F \cup P( \langle w \rangle) \subseteq F'$. Let $V' := P^{-1}(F') \cup \{0\}$. By definition $V' \in \mathcal{F}$ and since $V'$ is a subspace containing $V \cup w$ then $V \oplus \langle w \rangle \leq V'$. To show $V'$ covers $V$, assume there exists $W \in \mathcal{F}$ such that $V \lneq W \leq V'$. Applying $P$ and using the fact that $P$ is inclusion preserving, we get $F= P(V) \subsetneq P(W) \subseteq P(V') = F'$. However because $W \in \mathcal{F}$ then $P(W) \in \mathcal{F}_M$. But $F'$ covers $F$ hence we must have that $P(W) = F'$ and therefore $W = V'$. This implies $V'$ is a cover of $V$ and shows $\mathcal{F}$ is the collection of flats of a $q$-matroid.
Finally to show $\mathcal{F}_M$ and $\mathcal{F}$ are isomorphic as lattices, note that $P(\mathcal{F}) = \mathcal{F}_M$ hence by Theorem \ref{flats properties} the isomorphism follows.
\end{proof}
We now show that the lattice of flats of the $q$-matroid $\mathcal M$ contracted by a flat $F$ is isomorphic to the lattice of flats of $P(\mathcal M)/P(F)$.
\begin{theorem}\label{contraction flat iso}
Let $\mathcal M$ be a $q$-matroid, $P(\mathcal M)$ its projectivization matroid and $\mathcal{F}_{\mathcal M}, \mathcal{F}_{P(\mathcal M)}$ their respective lattice of flats. Then $\mathcal{F}_{\mathcal M/F} \cong \mathcal{F}_{P(\mathcal M)/P(F)}$ (as lattices) for any $F \in \mathcal{F}_{\mathcal M}$.
\end{theorem}
\begin{proof}
Throughout let $F'_1, F'_2 \in \mathcal{F}_{\mathcal M/F}$ and $V_i = \pi^{-1}(F_i')$, where $\pi : E \rightarrow E/F$. By Proposition \ref{contraction loop} and Lemma \ref{flats properties},
$F'_i \in \mathcal{F}_{\mathcal M /F} \Leftrightarrow V_i \in \mathcal{F}_{\mathcal M} \Leftrightarrow P(V_i) \in \mathcal{F}_{P(\mathcal M)} \Leftrightarrow P(V_i) - P(F) \in \mathcal{F}_{P(\mathcal M) / P(F)}$. Furthermore, $F_1' = F_2' \Leftrightarrow V_1 = V_2 \Leftrightarrow P(V_1) - P(F) = P(V_2) - P(F)$. Hence there is a one-to-one correspondence between $\mathcal{F}_{\mathcal M/F}$ and $\mathcal{F}_{P(\mathcal M)/P(F)}$ described by the map $\psi:\mathcal{F}_{\mathcal M/F} \rightarrow \mathcal{F}_{P(\mathcal M)/P(F)}$, where $\psi(F'_i) = P(V_i) - P(F)$. Since the lattices of flats are finite, to show $\psi$ is a lattice isomorphism, we need only to show $\psi$ preserves meets. Recall that the meet of flats in either lattice is the intersection of the flats.
\begin{align*}
\psi(F_1' \cap F_2') &= P(V_1 \cap V_2) - P(F) \\
&= (P(V_1) \cap P(V_2)) - P(F) \\
&= (P(V_1) - P(F)) \cap (P(V_2) - P(F))\\
&= \psi(F_1') \cap \psi(F_2'),
\end{align*}
which completes the proof
\end{proof}
The next few properties about projectivization matroids, although not difficult to prove, will be useful in following sections.
\begin{proposition}\label{loop/qloop}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and $P(\mathcal M) = (\mathbb{P}E, r)$ its projectivization matroid. Then $\mathcal M$ contains a loop if and only if $P(\mathcal M)$ contains a loop.
\end{proposition}
\begin{proof}
Let $\langle e \rangle \leq E$ be a 1-dimensional subspace. By definition, $r(P(\langle e \rangle)) = \rho (\langle e \rangle)$. Hence $\langle e \rangle$ is a loop in $\mathcal M$ if and only if $P(\langle e \rangle)$ is a loop in $P(\mathcal M)$.
\end{proof}
\begin{proposition}\label{spanning set project}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and $P(\mathcal M)= (\mathbb{P}E, r)$ its projectivization matroid. Let $A \subseteq \mathbb{P}E$ be such that $A$ contains a basis of $E$. Then $r(A) = r(\mathbb{P}E)$.
\end{proposition}
\begin{proof}
Since $A$ contains a basis of $E$ then $\langle A \rangle = E$. Hence $r(A) = \rho(\langle A \rangle) = \rho(E) = r(\mathbb{P}E)$.
\end{proof}
We conclude the section by studying the relation between minors of a $q$-matroid and minors of its projectivization matroid. To do so we introduce the following notation.
\begin{notation}\label{notation}
Let $V \leq E$.
\begin{itemize}
\item $\mathcal{Q}_V := \{ \langle w \rangle \in \mathbb{P}E \, : \, \langle w \rangle \not \leq V \} = \mathbb{P}E - P(V)$.
\item $\mathcal{Q}_{V}^{*A} := \mathcal{Q}_{V}- A$ for $A \subseteq \mathcal{Q}_{V}$.
\end{itemize}
\end{notation}
Note that $\mathbb{P}E - \mathcal{Q}_V = \mathbb P V$.
Furthermore for spaces $W,V \leq E$ such that $W \oplus V = E$, every element in $E/W$ can be written as $v + W$ for a unique $v \in V$. Thus the map $\psi: E/W \rightarrow V$, $v + W \mapsto v$ is a well-defined vector space isomorphism and induces a bijection on projective spaces. By slight abuse of notation we use $\psi$ as both the vector space isomorphism and the projective space bijection.
It can then easily be shown that $\langle \psi(A) \rangle = \psi(\langle A \rangle)$ for all $A \subseteq \mathbb P(E/W)$.
\begin{theorem}\label{contraction proj}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and let $W,V \leq E$ such that $W \oplus V = E$. Let $V^{\perp}$ be the orthogonal space of $V$ w.r.t a fix NSBF. Furthermore let $S = \{\langle w_1 \rangle , \cdots , \langle w_t \rangle \} \subseteq \mathbb{P}E$ such that $\{w_1, \cdots, w_t\}$ is a basis of $W$.Then
$$P(\mathcal M/W) \cong (P(\mathcal M) / S )\setminus \mathcal{Q}_{V}^{*S} $$
$$P(\mathcal M \setminus V^{\perp}) \cong P(\mathcal M) \setminus \mathcal{Q}_{V}.$$
\end{theorem}
\begin{proof}
Let $N := P(\mathcal M)/S \setminus \mathcal{Q}_{V}^{*S}$. Note $N$ has groundset $\mathbb{P}E - \mathcal{Q}_{V} = \mathbb P V$ whereas $P(\mathcal M/W)$ has groundset $\mathbb P(E/W)$. Let $\psi : \mathbb P(E/W) \rightarrow \mathbb P V$ be the bijection described previously.
To show $N \cong P(M/W)$, we must show $r_{P(\mathcal M/W)}(A) = r_N(\psi(A))$ for all $A \subseteq \mathbb P(E/W)$.
Let $\pi : E \rightarrow E/W$ be the canonical projection. Since $S$ is a basis of $W$ then $\pi^{-1}(\langle A \rangle) = \langle \psi(A) \rangle + W = \langle \psi(A) \cup S \rangle.$ Furthermore by Theorem \ref{$q$-matroid/matroid}, $r_{P(\mathcal M)}(S) = \rho(W)$. Hence we get:
\begin{align*}
r_{P(\mathcal M/W)}(A) &= \rho_{\mathcal M/W}(\langle A \rangle)\\
&= \rho_{\mathcal M}(\pi^{-1}(\langle A\rangle)) - \rho_{\mathcal M}(W)\\
&= \rho_{\mathcal M}(\langle \psi(A) \cup S \rangle) - \rho_{\mathcal M}(\langle S \rangle)\\
&= r_{P(\mathcal M)}( \psi(A) \cup S) - r_{P(\mathcal M)}(S) \\
&= r_{P(\mathcal M)/ S}(\psi(A))\\
&= r_{N}(\psi(A)),
\end{align*}
where the last equality holds because $\psi(A) \subseteq \mathbb P V$ which is the groundset of $N$.
Moving on to the second equivalence. Both matroid $P(\mathcal M \setminus V^{\perp})$ and $P(\mathcal M) \setminus \mathcal{Q}_{V}$ have groundset $\mathbb P V$. Hence we need only to show that the rank functions of both matroids are equal.
Let $A \subseteq \mathbb P V$.
\begin{align*}
r_{P(\mathcal M \setminus V^{\perp})}(A) &= \rho_{\mathcal M \setminus V^{\perp}}(\langle A \rangle)\\
&= \rho_{\mathcal M}(\langle A \rangle) \\
&= r_{P(\mathcal M)}(A) \\
&= r_{P(\mathcal M) \setminus \mathcal{Q}_{V}}(A),
\end{align*}
where the last equality follows because $A \subseteq \mathbb P V$ which is the groundset of $P(\mathcal M)\setminus\mathcal{Q}_V$.
\end{proof}
\section{Maps of Matroids and $q$-Matroids
Maps between matroids were first introduced to study matroids from a category theory approach. The reader may refer to \cite{Heunen2018TheCO,RotaMatroid} for more details. Maps between $q$-matroids were defined for the same purposes, but were only recently introduced in \cite{GLJ21}. In this section we focus on the relation between maps of $q$-matroids and maps of matroids, and show that the projectivization map is a functor from categories of $q$-matroids to categories of matroids.
This, in turn, provides a new approach to study maps between $q$-matroids.
For both matroids and $q$-matroids, there exists two main types of maps that preserve the matroid structure: weak and strong maps. To avoid confusion between maps of matroids and maps of $q$-matroids, we use the terms weak and strong maps for the former, and, $q$-weak and $q$-strong maps for the latter. We recall the definitions of weak and strong maps between matroids and some of their properties.
The reader should note that the loop extension (Definition \ref{loop extension}) is needed to define maps between matroids. As the next definition will show a map between matroids is a map defined on the groundset of the loop extension matroids.
By remark \ref{loop/emptyset}, the added loop can be seen as an element representing the empty set of the matroid. Hence, mapping an element to the added loop of the codomain can be seen as mapping an element to the empty set.
\begin{definition}\label{Strong Map 1}
Let $M = (S, r_M)$ and $N = (T , r_N)$ be matroids and $M_o$, $N_o$ be their respective loop extension matroids. A map $\sigma : M \rightarrow N$ is a map between the groundset of the loop extension matroids, i.e. $\sigma : S_o \rightarrow T_o$, such that $\sigma(o_M) = o_N$. Furthermore $\sigma$ is said to be:
\begin{itemize}
\item \emph{weak} if $r_{N_o}(\sigma (A)) \leq r_{M_o}(A)$ for all $A \subseteq S_o$.
\item \emph{strong} if $\sigma^{-1}(F) \in \mathcal{F}_{M_o}$ for all $F \in \mathcal{F}_{N_o}$.
\end{itemize}
\end{definition}
It is well known (see \cite[Chap.~8, Lem.~8.1.7]{RotaMatroid}) that strong maps are weak maps. Furthermore a map $\sigma: M \rightarrow N$ induces a map $\sigma^{\#} : \mathcal{F}_{M_o} \rightarrow \mathcal{F}_{N_o}$, where $\sigma^{\#}(F) = \mathrm cl_{N_o}(\sigma(F))$ for all $F \in \mathcal{F}_{M_o}$. Using Proposition \ref{flat loop extension}, one can alternatively define $\sigma^{\#} : \mathcal{F}_M \rightarrow \mathcal{F}_N$. As the following theorem shows, the induced map $\sigma^{\#}$ provides an alternative definition for strong maps.
\begin{theorem}\label{Strong Map 2}(\cite[Prop 8.1.3]{RotaMatroid})
A map $\sigma: M \rightarrow N$ is a strong map if and only if the following hold:
\begin{itemize}
\item[(1)] for all $F_1, F_2 \in \mathcal{F}_M$,
$$\sigma^{\#}(F_1 \vee F_2) = \sigma^{\#}(F_1) \vee \sigma^{\#}(F_2)$$
\item[(2)] $\sigma^{\#}$ sends atoms of $\mathcal{F}_{M}$ to atoms or to the zero of $\mathcal{F}_N$.
\end{itemize}
\end{theorem}
The main result of this section is the analogue of Theorem \ref{Strong Map 2} for $q$-matroids.
We turn to the definitions of maps between $q$-matroids, as introduced in \cite{GLJ21}. Similarly to matroids, maps between $q$-matroids are maps between groundspaces that send subspaces to subspaces.
\begin{definition}\label{L-maps}
A map $\sigma : E_1 \rightarrow E_2$ is an \emph{$\mathcal{L}$-map} if $\sigma(V) \in \mathcal{L}(E_2)$ for all $V \in \mathcal{L}(E_1)$. An $\mathcal{L}$-map $\sigma$ induces a map $\sigma_{\mathcal{L}} : \mathcal{L}(E_1) \rightarrow \mathcal{L}(E_2)$. Two $\mathcal{L}$-maps $\sigma, \psi$ from $E_1$ to $E_2$ are \emph{$\mathcal{L}$-equivalent}, denoted by $\sigma \sim_{\mathcal{L}} \psi$, if $\sigma_{\mathcal{L}} = \psi_{\mathcal{L}}$.
\end{definition}
We consider the following two types of maps between $q$-matroids.
\begin{definition}
Let $\mathcal M = (E_1, \rho_{\mathcal M})$ and $\mathcal N = (E_2, \rho_{\mathcal N})$ be $q$-matroids. A map $\sigma: \mathcal M \rightarrow \mathcal N$ is an $\mathcal{L}$-map between the groundspaces of $\mathcal M$ and $\mathcal N$, i.e $\sigma: E_1 \rightarrow E_2$. Furthermore $\sigma$ is said to be
\begin{itemize}
\item[(a)] \emph{$q$-weak} if $\rho_{\mathcal N}(\sigma(V)) \leq \rho_{\mathcal M}(V)$ for all $V \leq E_1$.
\item[(b)] \emph{$q$-strong} if $\sigma^{-1}(F) \in \mathcal{F}_{\mathcal M}$ for all $F \in \mathcal{F}_{\mathcal N}$.
\end{itemize}
\end{definition}
To study the relation between maps of matroids and maps of $q$-matroids we need the following notation.
Given a vector space $E$, define the \emph{extended projective space of $E$} as $\mathbb{P}_oE = \mathbb{P}E \cup \{o\}$, where $\{o\}$ is an arbitrary element disjoint from $\mathbb{P}E$. Let $P_o : E \rightarrow \mathbb{P}_oE$, where $P_o(0) = o$ and $P_o(v) = \hat{P}(v)$ for $v \neq 0$ and $\hat{P}: E - \{0\} \rightarrow \mathbb{P}E$ is as introduced in the previous section. We call $P_o$ the \emph{extended projectivization map}.
Note, unlike the projectivization map, we do not consider $P_o$ as lattice map but as a map between a vector space to its extended projective space.
Given a $q$-matroid $\mathcal M = (E, \rho)$ and the loop extension of its projectivization matroid $P(\mathcal M)_o = (\mathbb{P}_oE, r_o)$, the map $P_o$ can be viewed as a map between the groundspace $E$ to the groundset $\mathbb{P}_oE$ such that $\rho(V) = r_o(P_o(V))$ for all $V \leq E$. Furthermore for any $A \subseteq \mathbb{P}_oE$ let $\langle A \rangle := \langle P_o^{-1}(A) \rangle_{\mathbb F_q}$. It can easily be shown that $r_o(A) = \rho(\langle A \rangle )$.
Recall from Definition \ref{L-maps} that an $\mathcal{L}$-map $\sigma : E_1 \rightarrow E_2$ induces a map on the lattices of subspaces $\sigma_{\mathcal{L}} : \mathcal{L}(E_1) \rightarrow \mathcal{L}(E_2)$. By restricting $\sigma_{\mathcal{L}}$ to the $1$-dimensional spaces and the $0$ of $E_1$, $\sigma_{\mathcal{L}}$ can be viewed as map between the extended projective spaces $\mathbb{P}_oE_1$ and $\mathbb{P}_oE_2$, i.e $\sigma_{\mathcal{L}} : \mathbb{P}_oE_1 \rightarrow \mathbb{P}_oE_2$. As the next proposition shows, $\sigma$ and $\sigma_{\mathcal{L}} $ commute with the extended projectivization map.
\begin{proposition}\label{maps commuting}
Let $\sigma : E_1 \rightarrow E_2$ be an $\mathcal{L}$-map, $\sigma_{\mathcal{L}} : \mathbb{P}_oE_1 \rightarrow \mathbb{P}_oE_2$ its induced map on the extended projective spaces, and $P_o : E_i \rightarrow \mathbb{P}_oE_i$ be the extended projectivization map. Then
$$P_o \circ \sigma = \sigma_{\mathcal{L}} \circ P_o.$$
\end{proposition}
\begin{proof}
Let $v \in E_1$. Since $\sigma$ is an $\mathcal{L}$-map then $\langle \sigma(v) \rangle = \sigma(\langle v \rangle) = \sigma_{\mathcal{L}}(\langle v \rangle)$.
But note $\langle \sigma( v ) \rangle = P_o(\sigma(v))$ and $\langle v \rangle = P_o( v )$. Hence the wanted equality follows.
\end{proof}
We now consider the case when an $\mathcal{L}$-map $\sigma$ is a map between $q$-matroids. The induced map $\sigma_{\mathcal{L}}$ between the extended projective spaces turns out to be a map between projectivization matroids. Furthermore $\sigma$ being $q$-weak or $q$-strong is fully determined by whether $\sigma_{\mathcal{L}}$ is weak or strong, and vice versa.
\begin{theorem}\label{maps/$q$-maps}
Let $\mathcal M = (E_1, \rho_{\mathcal M})$, $\mathcal N = (E_2, \rho_{\mathcal N})$ be $q$-matroids and $P(\mathcal M) = (\mathbb{P}E_1, r_{P(\mathcal M)})$, $P(\mathcal N) = (\mathbb{P}E_2, r_{P(\mathcal N)})$ be their projectivization matroids. Let $\sigma : \mathcal M \rightarrow \mathcal N$ be an $\mathcal{L}$-map. Then $\sigma_{\mathcal{L}}: P(\mathcal M) \rightarrow P(\mathcal N)$ is a map between the projectivization matroids and the following holds:
\begin{itemize}
\item $\sigma$ is $q$-weak if and only if $\sigma_{\mathcal{L}}$ is weak
\item $\sigma$ is $q$-strong if and only if $\sigma_{\mathcal{L}}$ is strong.
\end{itemize}
\end{theorem}
\begin{proof}
To start note that $\sigma_{\mathcal{L}}$ is a map between the groundsets of the loop extension matroid $P(\mathcal M)_o$ and $P(\mathcal N)_o$ with $\sigma_{\mathcal{L}}(o_{P(\mathcal M)}) = o_{P(\mathcal N)}$. Thus $\sigma_{\mathcal{L}} : P(\mathcal M) \rightarrow P(\mathcal N)$ is well defined.
We first prove $\sigma$ is $q$-weak if and only if $\sigma_{\mathcal{L}}$ is weak. Assume $\sigma$ is weak. Let $A \subseteq \mathbb{P}_oE_1$ and $V := \langle A \rangle$. Both $P_o$ and $\sigma$ are inclusion preserving, hence, $(P_o \circ \sigma) (P_o^{-1}(A)) \subseteq (P_o \circ \sigma)(V)$.
Using Proposition \ref{maps commuting} on the first term of the previous inclusion gives us $(\sigma_{\mathcal{L}} \circ P_o)(P_o^{-1}(A)) = \sigma_{\mathcal{L}}(A) \subseteq (P_o \circ \sigma)(V)$. Furthermore, by the monotonicity property of the rank functions and because $\sigma$ is weak, we get
$$r_{P(\mathcal N)_o}(\sigma_{\mathcal{L}}(A)) \leq r_{P(\mathcal N)_o}(( P_o \circ \sigma)(V))) = \rho_{\mathcal N}(\sigma(V)) \leq \rho_{\mathcal M}(V) = r_{P(\mathcal M)_o}(A).$$
Because $A \subseteq \mathbb{P}_oE_1$ was arbitrarily chosen, then $\sigma_{\mathcal{L}}$ is weak.
Now assume $\sigma_{\mathcal{L}}$ is weak. Let $V \leq E_1$ and recall $\rho_{\mathcal M}(V) = r_{P(\mathcal M)_o}(P_o(V))$. Since $\sigma_{\mathcal{L}}$ is weak $r_{P(\mathcal N)_o}( (\sigma_{\mathcal{L}} \circ P_o)(V)) \leq r_{P(\mathcal M)_o}(P_o(V))$. Hence by Proposition \ref{maps commuting}, $r_{P(\mathcal N)_o}((P_o \circ \sigma)(V)) \leq r_{P(\mathcal M)_o}(P_o(V))$. This implies $\rho_{\mathcal N}(\sigma(V)) \leq \rho_{\mathcal M}(V)$ and shows $\sigma$ is $q$-weak.
\vspace{.2cm}
We now show that $\sigma$ is $q$-strong if and only if $\sigma_{\mathcal{L}}$ is strong.
From Proposition \ref{flat loop extension} and Lemma \ref{flats properties}, $F \in \mathcal{F}_{\mathcal M} \Leftrightarrow P(F) \in \mathcal{F}_{P(\mathcal M)} \Leftrightarrow P(F) \cup \{o\} \in \mathcal{F}_{P(\mathcal M)_o}$. A similar chain of equivalence holds for $\mathcal{F}_{\mathcal N}$ and $\mathcal{F}_{P(\mathcal N)_o}$. Furthermore all flats of $P(\mathcal N)_o$ are of the form $P_o(F) = P(F) \cup \{o\} $ for some flat in $\mathcal N$.
Therefore $\sigma_{\mathcal{L}}$ is strong iff $\sigma_{\mathcal{L}}^{-1} (P_o(F)) \in \mathcal{F}_{P(\mathcal M)_o}$ for all $P_o(F) \in \mathcal{F}_{P_o(\mathcal N)}$ iff $ (\sigma_{\mathcal{L}} \circ P_o)^{-1}(P_o(F)) \in \mathcal{F}_{\mathcal M}$ for all $P_o(F) \in \mathcal{F}_{P_o(\mathcal N)}$ iff $(P_o \circ \sigma)^{-1}(P_o(F)) = \sigma^{-1}(F) \in \mathcal{F}_{\mathcal M}$ for all $F \in \mathcal{F}_{\mathcal N}$ iff $\sigma$ is $q$-strong.
\end{proof}
From the above theorem it can easily be seen that the projectivization map is a functor from the category of $q$-matroids with $q$-weak (resp.\ $q$-strong) map to the category of matroids with weak (resp.\ strong) maps.
We now turn towards a proof of the analogue of Theorem \ref{Strong Map 2}. We first define the analogue of the map $\sigma^{\#}$.
\begin{definition}\label{$q$-flat map}
Let $\mathcal M$ and $\mathcal N$ be $q$-matroids with respective groundspaces $E_1, E_2$ and $\sigma : \mathcal M \rightarrow \mathcal N$ be an $\mathcal{L}$-map. Define $\sigma^{\#}: \mathcal{F}_{\mathcal M} \rightarrow \mathcal{F}_{\mathcal N}$ such that
$$\sigma^{\#}(F) = \mathrm cl_{\mathcal N}(\sigma(F)).$$
\end{definition}
The next useful Lemma shows that the induced maps $\sigma^{\#}$ and $\sigma_{\mathcal{L}}^{\#}$ commute with the extended projectivization map.
\begin{lemma}\label{commute flats}
Let $\mathcal M, \mathcal N$ be $q$-matroids, $\mathcal{F}_{\mathcal M} , \mathcal{F}_{\mathcal N}$ their lattices of flats and $P(\mathcal M), P(\mathcal N)$ their projectivization matroids. Furthermore let $\sigma:\mathcal M \rightarrow \mathcal N$ be an $\mathcal{L}$-map, $\sigma_{\mathcal{L}} : P(\mathcal M) \rightarrow P(\mathcal N)$ its induced map and $P_o: E_i \rightarrow \mathbb{P}_oE_i$ the extended projectivization map.
Then
$$P_o \circ \sigma^{\#} = \sigma_{\mathcal{L}}^{\#} \circ P_o$$
\end{lemma}
\begin{proof}
First recall, $F \in \mathcal{F}_{\mathcal M} \Leftrightarrow P_o(F) \in \mathcal{F}_{P(\mathcal M)_o}$. Let $F \in \mathcal{F}_M$, then $\sigma (F) \subseteq \sigma^{\#}(F)$ and since $P_o$ is inclusion preserving $(P_o \circ \sigma)(F) \subseteq (P_o \circ \sigma^{\#})(F)$.
By Proposition \ref{maps commuting}, the above containment implies $(\sigma_{\mathcal{L}} \circ P_o)(F) \subseteq (P_o \circ \sigma^{\#})(F)$.
Applying the closure operator of $P(\mathcal N)_o$, we get
$$(\sigma_{\mathcal{L}}^{\#} \circ P_o)(F) = \mathrm cl_{P(\mathcal N)_o}((\sigma_{\mathcal{L}} \circ P_o)(F)) \subseteq \mathrm cl_{P(\mathcal N)_o}((P_o \circ \sigma^{\#})(F)) = (P_o \circ \sigma^{\#})(F),$$
where the final equality holds because $\sigma^{\#}(F) \in \mathcal{F}_{\mathcal N}$ and therefore $(P_o \circ \sigma^{\#})(F) \in \mathcal{F}_{P(\mathcal N)_o}$.
Assume, for sake of contradiction, that $(\sigma_{\mathcal{L}}^{\#} \circ P_o)(F) \subsetneq (P_o \circ \sigma^{\#}) (F)$. Let $F' := (\sigma_{\mathcal{L}}^{\#} \circ P_o)(F)$. Then
$$(\sigma_{\mathcal{L}} \circ P_o)(F) \subseteq F' \subsetneq (P_o \circ \sigma^{\#}) (F).$$
By considering their preimage under $P_o$ and because $\sigma_{\mathcal{L}} \circ P_o = P_o \circ \sigma$, we get
$$\sigma(F) \subseteq P_o^{-1}(F') \subsetneq \sigma^{\#}(F).$$
However since $F' \in \mathcal{F}_{P(\mathcal N)_o}$ then $P_o^{-1}(F') \in \mathcal{F}_{\mathcal N}$. Therefore $P_o^{-1}(F')$ must contain $\mathrm cl_{\mathcal N}(\sigma(F)) = \sigma^{\#}(F)$, a contradiction. Hence $$(\sigma_{\mathcal{L}}^{\#} \circ P_o)(F) = (P_o \circ \sigma^{\#}) (F)$$
\end{proof}
In the statement of the previous Lemma, one can replace the extended projectivization map $P_o$ by the projectivization map $P$ introduced in the previous section. In fact, as previously remarked, the map $\sigma_{\mathcal{L}}^{\#}$ can be considered as map between the lattices of flats $\mathcal{F}_{P(\mathcal M)}$ to $\mathcal{F}_{P(\mathcal N)}$. Furthermore the projectivization map can also be restricted to a map between the lattice of flats of a $q$-matroid and its projectivization matroid. In the following Lemma, $P$ refers to the projectivization map restricted to the lattice of flats $\mathcal{F}_{\mathcal M}$ and $\mathcal{F}_{\mathcal N}$.
\begin{lemma}\label{commute flats 2}
Let the data be as in Lemma \ref{commute flats}, and let $P$ be the projectivization map on the lattices of flats $\mathcal{F}_{\mathcal M}$ and $\mathcal{F}_{\mathcal N}$. Then
$$P \circ \sigma^{\#} = \sigma_{\mathcal{L}}^{\#} \circ P.$$
\end{lemma}
\begin{proof}
Recall $\mathcal{F}_{P(\mathcal M)} = \{ F' - \{o\} \, : \, F' \in P(\mathcal M)_o\} = \{P_o(F) - \{o\} \, : \, F \in \mathcal{F}_{\mathcal M}\} = \{P(F) \, : \, F \in \mathcal{F}_{\mathcal M}\}$ and that the same holds for $P(\mathcal N)$. From the above chain of equality and Lemma \ref{commute flats}, equality follows straightforwardly.
\end{proof}
We now state and show the analogue of Theorem \ref{Strong Map 2} for $q$-strong maps.
\begin{theorem}\label{$q$-strong map 2}
Let $\mathcal M$, $\mathcal N$ be $q$-matroids. An $\mathcal{L}$-map $\sigma: \mathcal M \rightarrow \mathcal N$ is a $q$-strong map if and only if the following holds:
\begin{itemize}
\item[(1)] for all $F_1, F_2 \in \mathcal{F}_{\mathcal M}$,
$$\sigma^{\#}(F_1 \vee F_2) = \sigma^{\#}(F_1) \vee \sigma^{\#}(F_2)$$
\item[(2)] $\sigma^{\#}$ sends of $\mathcal{F}_{\mathcal M}$ atoms to atoms or to the zero of $\mathcal{F}_{\mathcal N}$.
\end{itemize}
\end{theorem}
\begin{proof}
($\Rightarrow$) Let $\sigma : \mathcal M \rightarrow \mathcal N$ be a $q$-strong map, which implies by Theorem \ref{maps/$q$-maps} that $\sigma_{\mathcal{L}}: P(\mathcal M) \rightarrow P(\mathcal N)$ is a strong map.
By Lemma \ref{flats properties} (1), $F \in \mathcal{F}_{\mathcal M} \Leftrightarrow P(F) \in \mathcal{F}_{P(\mathcal M)}$. Furthermore, from Theorem \ref{Strong Map 2} we obtain $\sigma_{\mathcal{L}}^{\#}(P(F_1) \vee P(F_2)) = \sigma_{\mathcal{L}}^{\#}(P(F_1)) \vee \sigma_{\mathcal{L}}^{\#}(P(F_2))$ for all $F_1, F_2 \in \mathcal{F}_{P(\mathcal M)}$.
By Lemma \ref{flats properties} (2), $P(F_1) \vee P(F_2) = P(F_1 \vee F_2)$, hence $\sigma_{\mathcal{L}}^{\#}(P(F_1 \vee F_2)) = \sigma_{\mathcal{L}}^{\#}(P(F_1)) \vee \sigma_{\mathcal{L}}^{\#}(P(F_2))$.
Applying Lemma \ref{commute flats 2} on the above equalities gives us
\begin{align*}
(P \circ \sigma^{\#})(F_1 \vee F_2) &= (P \circ \sigma^{\#})(F_1) \vee (P \circ \sigma^{\#})(F_2)\\
&= P ( \sigma^{\#}(F_1) \vee \sigma^{\#}(F_2)).
\end{align*}
Finally since $P$ is an isomorphism on the lattice of flat, the above equality implies
$$\sigma^{\#}(F_1 \vee F_2) = \sigma^{\#}(F_1) \vee \sigma^{\#}(F_2),$$
which shows $\sigma$ satisfies property (1) for all $F_1, F_2 \in \mathcal{F}_{\mathcal M}$.
To show $\sigma$ satisfies property (2), let $F \in \mathcal{F}_{\mathcal M}$ be an atom. Since $P$ is a lattice isomorphism then $P(F)$ is an atom of $\mathcal{F}_{P(\mathcal M)}$. Moreover $\sigma_{\mathcal{L}}$ is a strong map, hence by Theorem \ref{Strong Map 2}, $(\sigma_{\mathcal{L}}^{\#} \circ P)(F)$ must be an atom or the zero of $\mathcal{F}_{P(\mathcal N)}$. But by Lemma \ref{commute flats 2}, $(\sigma_{\mathcal{L}}^{\#} \circ P)(F) = (P \circ \sigma^{\#})(F)$, which implies $\sigma^{\#}(F)$ must be an atom or the zero of $\mathcal{F}_{\mathcal N}$ because, once again, $P$ is a lattice isomorphism. This concludes that $\sigma^{\#}$ satisfies the wanting properties.
($\Leftarrow$) Let $\sigma^{\#}$ satisfy properties (1) and (2).
We show that $\sigma$ is a $q$-strong map by showing that $\sigma_{\mathcal{L}}$ is a strong map. To do so we show that $\sigma_{\mathcal{L}}^{\#}$ satisfies Proposition \ref{Strong Map 2}.
Let $P(F_1), P(F_2) \in \mathcal{F}_{P(\mathcal M)}$. By Lemma \ref{flats properties} (2) $\sigma_{\mathcal{L}}^{\#}(P(F_1) \vee P(F_2)) = \sigma_{\mathcal{L}}^{\#}(P(F_1 \vee F_2)).$ Using Lemma \ref{commute flats 2} and the fact that $\sigma^{\#}$ satisfies property (1), we get
\begin{align*}
(\sigma_{\mathcal{L}}^{\#} \circ P)(F_1 \vee F_2) &= (P \circ \sigma^{\#})(F_1 \vee F_2)\\
&= P ( \sigma^{\#}(F_1) \vee \sigma^{\#}(F_2))\\
&= (P \circ \sigma^{\#})(F_1) \vee (P \circ \sigma^{\#})(F_2)\\
&= (\sigma_{\mathcal{L}}^{\#} \circ P)(F_1) \vee (\sigma_{\mathcal{L}}^{\#} \circ P)(F_2),
\end{align*}
where the second to last equality follows from Lemma \ref{flats properties} (2). Hence $\sigma_{\mathcal{L}}^{\#}$ satisfies property (1) of Prop \ref{Strong Map 2}.
Let $P(F) \in \mathcal{F}_{P(\mathcal M)}$ be an atom which, since $P$ is a lattice isomorphism, implies $F$ is an atom of $\mathcal{F}_{\mathcal M}$. Once again we use $(\sigma_{\mathcal{L}}^{\#} \circ P)(F) = (P \circ \sigma^{\#})(F).$
Because $\sigma^{\#}$ satisfies property (2) then $\sigma^{\#}(F)$ is an atom or the zero of $\mathcal{F}_{\mathcal N}$. Finally since $P$ is a lattice isomorphism between $\mathcal{F}_{\mathcal N}$ and $\mathcal{F}_{P(\mathcal N)}$ then $P(\sigma^{\#}(F)) = \sigma_{\mathcal{L}}^{\#}(P(F))$ must be an atom or the zero of $\mathcal{F}_{P(\mathcal N)}$ which show $\sigma_{\mathcal{L}}^{\#}$ satisfies property (2) of Proposition \ref{Strong Map 2}.
Therefore $\sigma_{\mathcal{L}}$ is a strong map and by Theorem \ref{maps/$q$-maps} we get that $\sigma$ is $q$-strong, concluding the proof.
\end{proof}
We conclude the section by showing that $q$-strong maps are $q$-weak maps.
\begin{corollary}
Let $\mathcal M = (E_1, \rho_{\mathcal M})$, $\mathcal N = (E_2, \rho_{\mathcal N})$ be $q$-matroids and $\sigma : \mathcal M \rightarrow \mathcal N$ be a $q$-strong map. Then $\sigma$ is a $q$-weak map.
\end{corollary}
\begin{proof}
Let $V \leq E_1$, such that $\rho_{\mathcal M}(V) = s$ and let $F := \mathrm cl_{\mathcal M}(V)$. Note that $V \subseteq F$ and $\rho_{\mathcal M}(F) = \rho_{\mathcal M}(V) =h(F)$, where $h$ is the height function of the lattice of flats as in Proposition \ref{flat $q$-matroid}.
Because $\mathcal{F}_{\mathcal M}$ is a geometric lattice, and $F$ has height $s$, then $F = \bigvee_{i=1}^s a_i$ where $a_i$ are atoms of $\mathcal{F}_{\mathcal M}$.
Since $\sigma$ is a $q$-strong map, by Theorem \ref{$q$-strong map 2} we get, $\sigma^{\#}(F) = \sigma^{\#}(\bigvee_{i=1}^s a_i) = \bigvee_{i=1}^s \sigma^{\#}(a_i)$.
Because $a_i$ are atoms of $\mathcal{F}_M$,by Theorem \ref{$q$-strong map 2}, $\sigma^{\#}(a_i)$ must be an atom or the zero of $\mathcal{F}_{\mathcal N}$ and that for all $1 \leq i \leq s$. Hence $\sigma^{\#}(F)$ is be the join of at most $s$-atoms, which implies $ \rho_{\mathcal N}(\sigma^{\#}(F))= h(\sigma^{\#}(F)) \leq s$.
Finally, since $V \leq F$ and $\sigma$ is an $\mathcal{L}$-map, then $\sigma(V) \leq \sigma(F) \leq \sigma^{\#}(F)$, where the last containment follows from the definition of $\sigma^{\#}$.
Therefore by the monotonicity property of $\rho_{\mathcal N}$ we get $\rho_{\mathcal N}(\sigma(V)) \leq \rho_{\mathcal N}(\sigma^{\#}(F)) \leq s = \rho_{\mathcal M}(V)$. Since this holds true for all $V \leq E_1$, $\sigma$ must be $q$-weak.
\end{proof}
\section{The Characteristic Polynomial
The characteristic polynomial is a useful invariant for both matroids and $q$-matroids. For the former it was intensively studied over the years, see for example \cite{welsh2010matroid, zaslavsky1987}. The latter was more recently introduced for $q$-polymatroids \cite{BCIJ21}, and was used to establish a weaker version of the Assmus-Mattson Theorem. However, in this paper, we are only interested in the characteristic polynomial of $q$-matroids.
Before defining the characteristic polynomial, we recall the definition of the M\"obius function which will often be used throughout the section.
\begin{definition}\label{Mobius function}
Let $(P , \leq )$ be a finite partially ordered set. The M\"obius function for $P$ is defined via the recursive formula
$$\mu_P(x,y) := \begin{cases} 1 & \textit{if } x = y, \\ - \sum_{x \leq z \lneq y} \mu_P(x,z) & \textit{if } x < y, \\
0 & \textup{otherwise.} \end{cases}$$
\end{definition}
We use the subscript of $\mu$ to distinguish between the M\"obius functions of different posets.
If the underlying poset is clear, the subscript may be omitted.
We now define the characteristic polynomial of a matroid.
\begin{definition}
Let $M = (S, r)$ be a matroid. The characteristic polynomial of $M$ is defined as follow:
$$\chi_M(x) = \sum_{A \subseteq S} (-1)^{|A|}x^{r(S) - r(A)}.$$
\end{definition}
It is well known that if a matroid $M$ contains a loop, its characteristic polynomial is identically 0. On the other hand if $M$ is loopless, then the characteristic polynomial of $M$ is fully determined by the lattice of flats.
Furthermore, one can recursively define the characteristic polynomial of a matroid in terms of the characteristic polynomial of its minors. We summarize this in the following theorem. Proofs can be found in \cite[Sec.3]{welsh2010matroid} and \cite[Sec.7.1]{zaslavsky1987}.
\begin{theorem}\label{char matroid}
Let $M = (S, r)$ be a matroid and $\mathcal{F}$ be its lattice of flats. If $M$ contains a loop then $\chi_M(x) = 0$. If $M$ has no loops, then
$$\chi_M(x) = \sum_{F \in \mathcal{F}} \mu_{\mathcal{F}}(0,F) x^{r(S) - r(F)}.$$
Furthermore for $e \in S$,
$$\chi_M(x) = \begin{cases} \chi_{M \setminus e}(x) \chi_{M/e}(x) & \textup{if } e \textup{ is a coloop,}\\
\chi_{M\setminus e}(x) - \chi_{M/e}(x) & \textup{otherwise.}
\end{cases}$$
\end{theorem}
Similarly to matroids, the characteristic polynomial of a $q$-matroid $\mathcal M$ is identically 0 if $\mathcal M$ contains a loop, and is fully determined by the lattice of flats otherwise. This was in fact shown by Whittle in \cite{whittle1992characteristic}, where the author generalized the result to any weighted lattice endowed with a closure operator. However for self containment purposes proofs of those facts will be included in this paper.
Furthermore, we use the projectivization matroid to find a recursive formula for the characteristic polynomial of $q$-matroids. The characteristic polynomial of a $q$-matroid is defined in the following way.
\begin{definition}\cite[Def. 22]{BCIJ21}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and $\mathcal{L}(E)$ be the subspace lattice of $E$. The \emph{characteristic polynomial} is defined as
$$\chi_{\mathcal M}(x):= \sum_{V \leq E} \mu_{\mathcal{L}(E)}(0, V)x^{\rho(E) - \rho(V)}.$$
\end{definition}
We state a few straightforward lemmas that will be useful later on.
\begin{lemma}\label{space of loops}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid, $\mathcal{F}_{\mathcal M}$ its lattice of flats, and $L := \{ e \in E \, | \, \rho(\langle e \rangle ) = 0\}$. Then $L$ is a subspace and $\rho(L) = 0$. It is called the \emph{the subspace of loops} of $\mathcal M$. Furthermore $L \leq F$ for all $F \in \mathcal{F}_{\mathcal M}$.
\end{lemma}
\begin{proof}
The first statement was proven in \cite[Lemma 11]{JP18}. For the second statement let $V \leq E$. By the monotonicity and submodularity properties of the rank function, $\rho(V) \leq \rho(V + L) \leq \rho(V) + \rho(L) - \rho(V \cap L) \leq \rho(V)$. Hence equality holds throughout and $L \leq \mathrm cl_{\mathcal M}(V)$. Since this is true for all $V \leq E$ then $L \leq F$ for all flats $F$ of $\mathcal M$.
\end{proof}
\begin{lemma}\label{Mobius flat decomp}
Let $\mathcal M= (E, \rho)$ be a $q$-matroid, $\mathcal{F}$ its lattice of flats and $\mathcal{L}$ the subspace lattice of $E$. Then
$$\chi_{\mathcal M}(x) = \sum_{F \in \mathcal{F}} \hspace{.1cm} \sum_{V \, : \, \mathrm cl_{\mathcal M}(V) = F} \mu_{\mathcal{L}}(0, V) x^{\rho(E) - \rho(F)}.$$
\end{lemma}
\begin{proof}
Let $V \leq E$, then $\mathrm cl_{\mathcal M}(V) \in \mathcal{F}$ and $\rho(V) = \rho(\mathrm cl_{\mathcal M}(V))$. Hence we get
\begin{align*}
\chi_M(x) &= \sum_{V \leq E} \mu_{\mathcal{L}}(0, V)x^{\rho(E) - \rho(V)}\\
&= \sum_{F \in \mathcal{F}}\hspace{.1cm} \sum_{V \, : \, \mathrm cl_{\mathcal M}(V) = F} \mu_{\mathcal{L}}(0, V) x^{\rho(E) - \rho(F)}. \quad \qedhere
\end{align*}
\end{proof}
We can now show that if a $q$-matroid contains a loop its characteristic polynomial is identically 0, whereas if it is loopless then the characteristic polynomial is determined by the lattice of flats.
\begin{theorem}\label{loop char 0}[see also \cite[Prop. 3.4]{whittle1992characteristic}]
Let $\mathcal M$ be a $q$-matroid such that $\mathcal M$ contains a loop then $\chi_{\mathcal M}(x) = 0$.
\end{theorem}
\begin{proof}
From Lemma \ref{Mobius flat decomp}, we know
$$\chi_{\mathcal M}(x) = \sum_{F \in \mathcal{F}} \sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}}(0, V) x^{\rho(E) - \rho(F)}.$$
We show that for all flat $F \in \mathcal{F}$,
\begin{equation}\label{fl}
\sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}}(0, V) = 0.
\end{equation}
We proceed by induction on the rank value of flats.
Let $F \in \mathcal{F}_{\mathcal M}$ such that $\rho(F) = 0$, i.e $F = \mathrm cl_{\mathcal M}(0)$. Since $\mathcal M$ contains a loop, $\{0\} \subsetneq F$. Hence by Definition \ref{Mobius function},
$$\sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}}(0, V) = \sum_{0 \leq V \leq F}\mu_{\mathcal{L}}(0, V) = 0.$$
Assume (\ref{fl}) holds for all $F \in \mathcal{F}$ such that $\rho(F) \leq k-1$.
Fix $F \in \mathcal{F}$ such that $\rho(F) = k$. Then
\begin{align*}
0 &= \sum_{V \leq F} \mu_{\mathcal{L}} (0, V) \\
&= \sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}} (0,V) + \sum_{F' \lneq F, F' \in \mathcal{F}} \hspace{.1cm} \sum_{V \, : \, \mathrm cl(V) = F'} \mu_{\mathcal{L}}(0,V)\\
&= \sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}} (0,V),
\end{align*}
where the last equality follows by induction hypothesis. Therefore (\ref{fl}) holds, and $\chi_{\mathcal M}(x) = 0.$
\end{proof}
\begin{theorem}\label{char flats}[see also \cite[Thm. 3.2]{whittle1992characteristic}]
Let $\mathcal M = (E, \rho)$ be a loopless $q$-matroid and $\mathcal{F}$ its lattice of flats. Then
$$\chi_{\mathcal M}(x) = \sum_{F \in \mathcal{F}} \mu_{\mathcal{F}}(0, F) x^{\rho(E) - \rho(F)}.$$
\end{theorem}
\begin{proof}
Let $\mathcal{L} = \mathcal{L}(E)$.
By Lemma \ref{Mobius flat decomp} we know
$$\chi_{\mathcal M}(x) = \sum_{F \in \mathcal{F}} \hspace{.1cm} \sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}}(0, V)x^{\rho(E) - \rho(F)}.$$
Hence, for all $F \in \mathcal{F}$ we must show
\begin{equation}\label{fl2}
\sum_{V \, : \, \mathrm cl(V) = F} \mu_{\mathcal{L}}(0, V) = \mu_{\mathcal{F}}(0,F).
\end{equation}
We once again proceed by induction on the rank of the flats of $\mathcal M$.
Since $\mathcal M$ is loopless, $\{0\} \in \mathcal{F}$. Let $F = \{0\}$, then (\ref{fl2}) follows trivially from Definition \ref{Mobius function}. Now assume (\ref{fl2}) holds true for all $F \in \mathcal{F}$ such that $\rho(F) \leq k-1$.
Fix a flat $F \in \mathcal{F}$ with $\rho(F) = k$. Then
\begin{align*}
\mu_{\mathcal{F}}(0,F) &= - \sum_{F' \lneq F, F' \in \mathcal{F}} \mu_{\mathcal{F}}(0, F')\\
&= -\sum_{F' \lneq F, F' \in \mathcal{F}} \sum_{V \, : \, \mathrm cl(V) = F'} \mu_{\mathcal{L}}(0,V) \\
&= - \sum_{V \, : \, \mathrm cl(V) \lneq F} \mu_{\mathcal{L}}(0,V) \\
&= \sum_{V \, : \, \mathrm cl(V) = F}\mu_{\mathcal{L}}(0,V),
\end{align*}
where the second equality follows from the induction hypothesis, and the last equality follows from Definition \ref{Mobius function}. This completes the proof.
\end{proof}
As the next theorem shows, defining the characteristic polynomial in terms of the lattice of flats of the $q$-matroid allows us to link the characteristic polynomial of a $q$-matroid with that of its projectivization matroid.
\begin{theorem}\label{char $q$-matroid/matroid}
Let $\mathcal M$ be a $q$-matroid and $P(\mathcal M)$ be its projectivization matroid. Then
$$\chi_{\mathcal M}(x) = \chi_{P(\mathcal M)}(x).$$
\end{theorem}
\begin{proof}
By Proposition \ref{loop/qloop} $\mathcal M$ contains a loop if and only if $P(\mathcal M)$ contains a loop. Furthermore, by Theorem \ref{flats properties}, we know $\mathcal{F}_{\mathcal M} \cong \mathcal{F}_{P(\mathcal M)}$ as lattices. Due to Propositions \ref{flats matroid} and \ref{flat $q$-matroid}, $\rho_{\mathcal M}(F) = h_{\mathcal{F}_{\mathcal M}}(F) = h_{\mathcal{F}_{P(\mathcal M)}}(P(F)) = r_{P(\mathcal M)}(P(F))$ and this for all flats $F \in \mathcal{F}_{\mathcal M}$ and $P(F) \in \mathcal{F}_{P(\mathcal M)}$. Therefore, the result follows directly from Theorems \ref{char matroid}, \ref{loop char 0} and \ref{char flats}.
\end{proof}
We furthermore get the following result when considering the contraction of $\mathcal M$ by a subspace $V \leq E$.
\begin{proposition}\label{char project equiv}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and $P(\mathcal M) = (\mathbb{P}E, r)$ its projectivization matroid. Then for all $V \leq E$,
$$\chi_{\mathcal M / V}(x) = \chi_{P(\mathcal M) / P(V)}(x).$$
\end{proposition}
\begin{proof}
Let $V \leq E$, then $V \in \mathcal{F}_{\mathcal M} \Leftrightarrow P(V) \in \mathcal{F}_{P(\mathcal M)}$. If $V \notin \mathcal{F}_{\mathcal M}$ then by Proposition \ref{contraction loop} $\mathcal M/V$ and $P(\mathcal M)/P(V)$ contain loops, therefore $\chi_{\mathcal M / V}(x) = 0 = \chi_{P(\mathcal M) / P(V)}(x)$.
If $V \in \mathcal{F}_{\mathcal M}$, by Theorem \ref{contraction flat iso}, $\mathcal{F}_{\mathcal M/F} \cong \mathcal{F}_{P(\mathcal M)/P(V)}$ as lattices and, by Proposition \ref{contraction loop}, both matroids are loopless. Hence, Theorems \ref{char matroid} and \ref{char flats} imply, $\chi_{\mathcal M /V}(x) = \chi_{P(\mathcal M) / P(V)}(x)$.
\end{proof}
The close connection between the characteristic polynomial of a $q$-matroid $\mathcal M$ and that of its projectivization matroid gives a new approach to study the former. In fact, we use this approach to find a recursive formula for the characteristic polynomial of a $q$-matroid in terms of the characteristic polynomial of its minors.
Because we will be using the recursive formula defined in Theorem \ref{char matroid} on the projectivization matroid $P(\mathcal M)$ and its minors, we need to pay a particular attention on whether $P(\mathcal M)$ and its minors contain coloops. We thus need the following few lemmas.
\begin{lemma}\label{coloop prop}
Let $M = (S, r)$ be a matroid and $\mathcal M = (E, \rho)$ be a $q$-matroid. Then
\begin{itemize}
\item[(a)] $w \in S$ is a coloop of $M \Longleftrightarrow r^*(w) = 0 \Longleftrightarrow r(S - w) = r(S) - 1$
\item[(b)] $\langle w \rangle \leq E$ is a coloop of $\mathcal M \Longleftrightarrow \rho^*(\langle w \rangle) = 0 \Longleftrightarrow \rho(\langle w \rangle^{\perp}) = \rho(E) - 1$
\end{itemize}
\end{lemma}
\begin{proof}
Statement (a) can be found in \cite[Section 1.6, Exercise 6]{oxley}. Statement (b) follows from Definition \ref{dual $q$-matroid}.
\end{proof}
Recall the notation $\mathcal{Q}_V = \{\langle w \rangle \, : \, \langle w \rangle \not \leq V\}$, and $\mathcal{Q}_V^{*e} = \mathcal{Q}_V - \{\langle e \rangle\}$ for $\langle e \rangle \in \mathcal{Q}_V$ introduced in Notation \ref{notation}.
To make the results and proofs easier to read, we may omit the brackets to denote 1-dimensional spaces and we let $v^{\perp} := \langle v \rangle^{\perp}$ for $v \in E$.
\begin{lemma}\label{coloop}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid, $P(\mathcal M) = (\mathbb{P}E, r)$ its projectivization matroid and $e, v \in \mathbb{P}E$ such that $\langle e \rangle \oplus \langle v \rangle^{\perp} = E$. Let $A \subsetneq \mathcal{Q}_{v^{\perp}}^{*e}$, then:
\begin{itemize}
\item[(a)] for all $w \in \mathcal{Q}_{v^{\perp}}^{*e} - A$, the element $w$ is not a coloop of the matroid $P(\mathcal M)\setminus A$.
\item[(b)] for all $w \in \mathcal{Q}_{v^{\perp}}^{*e}$ and $z \in \mathcal{Q}_{v^{\perp}} - (A \cup w)$, the element $z$ is not a coloop of $P(\mathcal M) \setminus A /w$.
\item[(c)] for all $w_1, w_2 \in \mathcal{Q}_{v^{\perp}} - A$, the matroid $P(\mathcal M) \setminus A/\{w_1, w_2\}$ contains a loop.
\item[(d)] $e$ is a coloop of the matroid $P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}$ if and only if $\langle v \rangle$ is a coloop of $\mathcal M$.
\end{itemize}
\end{lemma}
\begin{proof}
Throughout, let $\mathcal{Q} := \mathcal{Q}_{v^{\perp}}^{*e}$, $H:= v^{\perp}$ and $\{h_1, \cdots h_{n-1} \}$ be a basis of $H$. Furthermore let $A \subsetneq \mathcal{Q}$, $w \in \mathcal{Q} - A$ and consider the matroid $N := P(\mathcal M) \setminus A = (\mathbb{P}E - A, r_N) $. Note that $r(S) = r_N(S)$ for all $S \subseteq \mathbb{P}E - A$.
\vspace{.1cm}
Statement (a). Since $e \notin H$ and $\dim H = \dim E -1 $, the set $B:= \{h_1, \cdots, h_{n-1} , e\}$ is a basis of $E$, and $B \subseteq \mathbb{P}E - \mathcal{Q} \subseteq \mathbb{P}E - (A \cup w) \subseteq \mathbb{P}E - A$.
Hence by Proposition \ref{spanning set project}, $r(\mathbb{P}E - \mathcal{Q}) = r(\mathbb{P}E - (A \cup w) ) = r(\mathbb{P}E - A) = r(\mathbb{P}E)$. Because $r_N(S) = r(S)$ for all $S \subseteq \mathbb{P}E - A$, we have $r_N(\mathbb{P}E - (A \cup w) ) = r_N(\mathbb{P}E - A)$ so, by Lemma \ref{coloop prop}, $w$ is not a coloop of $N$, proving statement (a).
\vspace{.1cm}
Statement (b). Let $B:= \{h_1, \cdots , h_{n-1} , w\}$ which is a basis of $E$. We need only to show $r_{N/w}^*(z) \neq 0$ for all $z \in \mathcal{Q}_{v^{\perp}} - (A \cup w)$.
\begin{align*}
r_{N/w}^*(z) &= |z| + r_{N/w}(\mathbb{P}E - (A\cup w \cup z)) - r_{N/w}(\mathbb{P}E - (A\cup w)) \\
&= 1 + r_{N}((\mathbb{P}E - (A\cup w \cup z)) \cup w) - r_{N}(w) - r_{N}((\mathbb{P}E - (A \cup w)) \cup w) + r_{N}(w)\\
&= 1+ r(\mathbb{P}E - (A \cup z)) - r(\mathbb{P}E - A) \\
&= 1 + \rho(E) - \rho(E) = 1
\end{align*}
where the last equality follows from Proposition \ref{spanning set project} because $B$ is a subset $\mathbb{P}E - ( A \cup z)$ and $\mathbb{P}E - A$.
\vspace{.1cm}
Statement (c). Let $w_1, w_2 \in \mathcal{Q}_{v^{\perp}}-A$ and $W = \langle w_1, w_2 \rangle$. Clearly $\dim W = 2$ and $\dim (W \cap \langle v \rangle^{\perp}) = 1$. Hence there exists $z \in W$ such that $\langle z \rangle \notin \mathcal{Q}_{v^{\perp}}$. We show $z$ is a loop of $N / \{w_1, w_2\}$, i.e. $r_{N/\{w_1, w_2\}}(z) = 0$.
\begin{align*}
r_{N/\{w_1, w_2\}}(z) &= r(z \cup \{w_1,w_2\}) - r(\{w_1 , w_2\}) \\
&= \rho(\langle z, w_1,w_2 \rangle) - \rho(\langle w_1, w_2 \rangle) \\
&= \rho(W) - \rho(W) = 0.
\end{align*}
\vspace{.1cm}
Statement (d). Let $N' := P(\mathcal M) \setminus \mathcal{Q} = (\mathbb{P}E - \mathcal{Q}, r_{N'})$. By Lemma \ref{coloop prop} $\langle v \rangle$ is a coloop of $\mathcal M$ if and only if $\rho(H) = \rho(E) -1$.
Moreover, $\rho(H) = r(\mathbb P H) = r_{N'}(\mathbb P H)$ and $\rho(E) = r(\mathbb{P}E) = r(\mathbb{P}E - \mathcal{Q}) = r_{N'}(\mathbb{P}E - \mathcal{Q})$. Hence $\langle v \rangle$ is a coloop of $\mathcal M$ if and only if $r_{N'}(\mathbb P H) = r_{N'}(\mathbb P H \cup e) -1$ if and only if $e$ is a coloop of $N'$.
\end{proof}
With those results in place, we are now ready to consider the first step of our main theorem. For the next results we use the following notation. Given $\mathcal{Q}_{v^{\perp}}^{*e}$, fix an ordering of its elements. Define $\mathcal S_0 := \emptyset$ and $\mathcal S_i := \{w_1, \cdots , w_i\}$ where $w_j$ is the $j^{\textup{th}}$ element of $\mathcal{Q}_{v^{\perp}}^{*e}$. Note furthermore that $|\mathcal S_i| = i$ and $\mathcal S_{q^{n-1}-1} = \mathcal{Q}_{v^{\perp}}^{*e}$. Moreover, in the proofs of the remaining results in this section, Proposition \ref{del cont commute} and Theorem \ref{char matroid} may be used without mention.
\begin{proposition}\label{intermediate form}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid, $P(\mathcal M) = (\mathbb{P}E, r)$ its projectivization matroid, $e,v \in \mathbb{P}E$ such that $\langle e \rangle \oplus \langle v\rangle^{\perp} = E$.
Then
$$\chi_{\mathcal M}(x) = \begin{cases} \chi_{\mathcal M \setminus v}(x) \cdot \chi_{\mathcal M / e}(x) - \sum_{i=0}^{q^{n-1}-2} \chi_{P(\mathcal M) \setminus S_i / w_{i+1}} & \textup{ if } v \textup{ is a coloop of } \mathcal M \\
\chi_{\mathcal M \setminus v}(x) - \chi_{\mathcal M / e}(x) - \sum_{i=0}^{q^{n-1}-2} \chi_{P(\mathcal M) \setminus S_i / w_{i+1}} & \textup{otherwise}
\end{cases}$$
\end{proposition}
\begin{proof}
We first use an induction argument on $k := |\mathcal S_k|$ to show that for all $1 \leq k \leq q^{n-1} -1$,
\begin{equation}\label{int form induc}
\chi_{P(\mathcal M)}(x) = \chi_{P(\mathcal M)\setminus \mathcal S_{k}}(x) - \sum_{i=0}^{k-1} \chi_{P(\mathcal M)\setminus \mathcal S_i / w_{i+1}}(x).
\end{equation}
We prove the base case when $k=1$. By Lemma \ref{coloop} (a), $w_1$ is not a coloop of $P(\mathcal M)$ hence, by Theorem \ref{char matroid}, $\chi_{P(\mathcal M)}(x) = \chi_{P(\mathcal M) \setminus \mathcal S_1}(x) - \chi_{P(\mathcal M) \setminus \mathcal S_0 / w_1}(x)$, where recall $\mathcal S_0 = \emptyset$ and $\mathcal S_1 = \{w_1\}$.
Now assume (\ref{int form induc}) holds for $ k = q^{n-1} - 2$. Then
\begin{align*}
\chi_{P(\mathcal M)}(x) &= \chi_{P(\mathcal M) \setminus \mathcal S_k}(x) - \sum_{i=0}^{k-1} \chi_{P(\mathcal M)\setminus \mathcal S_i / w_{i+1}}(x)\\
&= \chi_{P(\mathcal M) \setminus (\mathcal S_k \cup w_{k+1})}(x) - \chi_{P(\mathcal M) \setminus \mathcal S_k / w_{k+1}}(x) - \sum_{i=0}^{k-1} \chi_{P(\mathcal M)\setminus \mathcal S_i / w_{i+1}}(x) \\
&= \chi_{P(\mathcal M) \setminus \mathcal S_{k+1}}(x) - \sum_{i=0}^{k} \chi_{P(\mathcal M)\setminus \mathcal S_i / w_{i+1}}(x).
\end{align*}
The second equality holds true by Theorem \ref{char matroid}, because $w_{k+1}$ is not a coloop of $P(\mathcal M)\setminus \mathcal S_{k}$ by \ref{coloop} (a). This establishes (\ref{int form induc}).
Because $\mathcal S_{q^{n-1} -1} = \mathcal{Q}_{v^{\perp}}^{*e}$, we conclude the proof by using Theorem \ref{char matroid} on $\chi_{P(\mathcal M) \setminus \mathcal S_{q^{n-1} -1}}(x)$ with the element $e \in \mathcal{Q}_{v^{\perp}} - \mathcal S_{q^{n-1} -1}$.
We therefore consider two cases: when $e$ is a coloop of $P(\mathcal M)\setminus \mathcal{Q}_{v^{\perp}}^{*e}$ or not. By Lemma \ref{coloop} (d), those two cases correspond exactly to when $\langle v \rangle$ is a coloop of $\mathcal M$ or not.
First assume $e$ is a coloop of $P(\mathcal M)\setminus \mathcal{Q}_{v^{\perp}}^{*e}$, and therefore $\langle v \rangle$ is a coloop of $\mathcal M$.
Then by Theorem \ref{char matroid}
$$\chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}}(x) = \chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}}(x) \chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}/e}(x).$$
By Theorem \ref{contraction proj}, $P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}} = P(\mathcal M \setminus v)$ and $P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}/e \cong P(\mathcal M / e)$ hence their respective characteristic polynomials are equal. Furthermore by Theorem \ref{char $q$-matroid/matroid}, $\chi_{P(\mathcal M \setminus v)}(x) = \chi_{\mathcal M \setminus v}(x)$ and $\chi_{P(\mathcal M/e)}(x) = \chi_{\mathcal M/e}(x)$. Therefore,
$\chi_{P(\mathcal M) \setminus \mathcal S_{q^{n-1}-1}}(x) = \chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}}(x) \cdot \chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}/e}(x) = \chi_{\mathcal M \setminus v}(x) \cdot \chi_{\mathcal M / e}(x)$, which, when substituted in (\ref{int form induc}) gives us the wanted equality.
If $e$ is a not coloop of $P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}$, and therefore $\langle v \rangle $ is not a coloop of $\mathcal M$. Then $$\chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}}(x) = \chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}}(x) - \chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*e}/e}(x).$$
Once again using Theorems \ref{contraction proj} and \ref{char $q$-matroid/matroid}, the wanted equality follows.
\end{proof}
At this point, note that the characteristic polynomial $\chi_{\mathcal M}(x)$ depends on both the characteristic polynomial of minors of the $q$-matroid $\mathcal M$ and the characteristic polynomial of minors of $P(\mathcal M)$. In the following Theorem, we rewrite all characteristic polynomials of minors of $P(\mathcal M)$ in terms of characteristic polynomials of minors of $\mathcal M$.
\begin{theorem}
Let $\mathcal M = (E, \rho)$ be a $q$-matroid and $e,v \in E$ such that $\langle e \rangle \oplus \langle v\rangle ^{\perp} = E$.
Then
$$\chi_{\mathcal M}(x) = \begin{cases}
\chi_{\mathcal M \setminus v}(x) \cdot \chi_{\mathcal M/e}(x) - \sum_{w \in \mathcal{Q}_{v^{\perp}}^{*e}} \chi_{\mathcal M/w} (x)& \textup{ if } v \textup{ is a coloop of } \mathcal M, \\
\chi_{\mathcal M \setminus v}(x) - \sum_{w \in \mathcal{Q}_{v^{\perp}}} \chi_{\mathcal M/w} (x) & \textup{otherwise.}
\end{cases}$$
\end{theorem}
\begin{proof}
Given the equation of Proposition \ref{intermediate form}, we show that
\begin{equation}\label{final form}
\sum_{i=0}^{q^{n-1}-2} \chi_{P(\mathcal M) \setminus \mathcal S_i / w_{i+1}}(x) = \sum_{w \in \mathcal{Q}_{v^{\perp}}^{*e}} \chi_{\mathcal M/ w } (x).
\end{equation}
Fix $0 \leq i \leq q^{n-1}-2$ and consider $\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x)$. For all $A \subseteq \mathcal{Q}_{v^{\perp}} - (S_i \cup w_{i+1})$, we show by induction on $|A|$ that
\begin{equation}\label{induc final form}
\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x) = \chi_{P(\mathcal M)\setminus (S_i \cup A) /w_{i+1}}(x).
\end{equation}
First let $|A| = 1$ and let $w \in A$.
By Lemma \ref{coloop} (b), $w$ is not a coloop of $P(\mathcal M)\setminus \mathcal S_i/w_{i+1}$. Therefore
$$\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x) = \chi_{P(\mathcal M)\setminus (S_i \cup w) /w_{i+1}}(x) - \chi_{P(\mathcal M)\setminus S_i /\{w_{i+1} , w\}}(x).$$
By Lemma \ref{coloop} (c), $P(\mathcal M) \setminus S_i \, / \{w_{i+1},w\}$ contains a loop which implies its characteristic polynomial is $0$. Hence $\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x) = \chi_{P(\mathcal M)\setminus (S_i \cup w) /w_{i+1}}(x)$.
Now let $|A| = k$ and let $w \in A$. Since $|A - w| = k-1$ by induction hypothesis we get $\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x) = \chi_{P(\mathcal M)\setminus (S_i \cup (A-w)) /w_{i+1}}(x)$. Once again by Lemma \ref{coloop} (b), $w$ is not a coloop of $P(\mathcal M)\setminus (\mathcal S_i \cup (A-w))/w_{i+1}$ hence $$ \chi_{P(\mathcal M)\setminus (S_i \cup (A-w)) /w_{i+1}}(x) = \chi_{P(\mathcal M)\setminus (S_i \cup A) /w_{i+1}}(x) - \chi_{P(\mathcal M)\setminus (S_i \cup A) /\{w_{i+1}, w\}}(x).$$
By Lemma \ref{coloop} (c) $P(\mathcal M)\setminus (S_i \cup A) /\{w_{i+1}, w\}$ contains a loop and therefore its characteristic polynomial is 0. This completes the proof of (\ref{induc final form}), which if $A = \mathcal{Q}_{v^{\perp}} - (S_i \cup w_{i+1})$ shows that
$$\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x) =\chi_{P(\mathcal M) \setminus \mathcal{Q}_{v^{\perp}}^{*w_{i+1}} /w_{i+1}}(x).$$
Finally by Theorems \ref{contraction proj} (where $S = \{w_{i+1}\}$) and \ref{char $q$-matroid/matroid} we get
\begin{equation}\label{equ rewrting sum}
\chi_{P(\mathcal M)\setminus S_i /w_{i+1}}(x) = \chi_{\mathcal M/w_{i+1}}(x).
\end{equation}
Since the above induction holds true for any $i$ chosen, then (\ref{equ rewrting sum}) holds for all $0 \leq i \leq q^{n-1}-2$ and
\begin{align*}
\sum_{i=0}^{q^{n-1}-2} \chi_{P(\mathcal M) \setminus \mathcal S_i / w_{i+1}}(x) &= \sum_{i=0}^{q^{n-1}-2} \chi_{\mathcal M/w_{i+1}}(x) \\
&= \sum_{w \in \mathcal{Q}_{v^{\perp}}^{*e}} \chi_{\mathcal M/w} (x).
\end{align*}
Substituting (\ref{final form}) into the equation of Proposition \ref{intermediate form} gives the desired result.
\end{proof}
\section{Rank Metric and Linear Block Codes
The study of matroids and $q$-matroids plays an important role in coding theory. In fact, it is well known that $\mathbb F_{q^m}$-linear rank metric codes induce $q$-matroids, and that linear block codes with the Hamming metric give rise to matroids. Furthermore many of the code invariants can be determined from the induced ($q$-)matroid.
In \cite{ABNR21}, Alfarano and co-authors showed that an $\mathbb F_{q^m}$-linear rank metric code $\mathcal{C}$ induces a linear block code that shares similar parameters. We show in this section how the projectivization matroid of a $q$-matroid relates to the matroid associated to that linear block code. Furthermore we use this relation and results from Section 5 to show the $q$-analogue of the critical theorem in terms of $q$-matroids and $\mathbb F_{q^m}$-linear rank metric codes.
We start the section by recalling some coding theory concepts. Throughout, let $\mathbb F_{q^m}$ be a field extension of $\mathbb F_q$ of degree $m$. Furthermore let $\Gamma$ be a basis of the vector space $\mathbb F_{q^m}$ over $\mathbb F_q$. For all $v = (v_1, \cdots v_n) \in \mathbb F_{q^m}^n$, let $\Gamma(v)$ be the $n \times m$ matrix such that the $i^{\textup{th}}$ row of $\Gamma(v)$ is the coordinate vector of $v_i$ with respect to the basis $\Gamma$. Let $\mathrm{rk}_{\mathbb F}(-)$, $\mathrm{colsp}_{\mathbb F}(-), \mathrm{rowsp}_{\mathbb F}(-)$ respectively denote the rank, column space and row space of a matrix over the field $\mathbb F$. Moreover, throughout the section, we fix the NSBF on $\mathbb F_q^n$ to be the standard dot product. We define the following two weight functions.
\begin{definition}
For all $v \in \mathbb F_{q^m}$, the \emph{Hamming weight} $\omega_H$, and the \emph{rank weight} $\omega_R$ of $v$ are defined as follow:
\begin{align*}
\omega_H(v) &:= \# \textup{non-zero entries of }v \\
\omega_R(v) &:= \mathrm{rk}_{\mathbb F_q} (\Gamma(v)).
\end{align*}
\end{definition}
It is well known that the rank weight is independent of the basis $\Gamma$ chosen.
Both weight functions induce a metric on $\mathbb F_{q^m}^n$, where $d_{\Delta}(v, w) = \omega_{\Delta}(v - w)$ for $v, w \in \mathbb F_{q^m}$ and $\Delta \in \{H, R\}$.
A \emph{linear block code} is a subspace of the metric space $(\mathbb F_{q^m}^n, d_H)$, and an \emph{$\mathbb F_{q^m}$-linear rank metric code} is a subspace of the metric space $(\mathbb F_{q^m}^n, d_R)$.
Given a code $\mathcal{C} \leq \mathbb F_{q^m}^n$, let $d_H(\mathcal{C})$, respectively $d_R(\mathcal{C})$, denote the minimum weight over all non zero elements of the code.
If $\dim \mathcal{C} = k$, then $d_{\Delta}(\mathcal{C}) \leq n - k +1$ for $\Delta \in \{H, R\}.$
For each metric, the above bound is called the \emph{Singleton bound}. $\mathcal{C}$ is said to be \emph{maximum distance separable}, respectively \emph{maximum rank distance} if the Hamming-metric, respectively the rank-metric, Singleton bound is achieved.
For both metrics, the weight distribution of the code is defined as follows.
\begin{definition}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be a code. For $\Delta \in \{H, R\}$, let
$$W_{\Delta}^{(i)}(\mathcal{C}) := \left| \{v \in \mathcal{C} \, : \omega_{\Delta}(v) = i\}\right|.$$
Furthermore let $W_{\Delta}(\mathcal{C}) := (W_{\Delta}^{(i)}(\mathcal{C}) \, : 0 \leq i \leq n)$. $W_{H}(\mathcal{C})$ is called the \emph{Hamming weight distribution} of $\mathcal{C}$ and $W_R(\mathcal{C})$ is called the \emph{rank weight distribution} of $\mathcal{C}$.
\end{definition}
We now introduce two notions of support for elements of $\mathbb F_{q^m}^n$.
\begin{definition}\label{support}
Let $v = (v_1 , \cdots , v_n) \in \mathbb F_{q^m}^n$ and $V \subseteq \mathbb F_{q^m}^n$.
\begin{align*}
S_{H}(v) &= \{ i \, : v_i \neq 0\} \quad \textup{and} \quad S_H(V) = \bigcup_{v \in V} S_H(v) \\
S_R(v) &= \mathrm{colsp}_{\mathbb F_q}(\Gamma(v)) \quad \textup{and} \quad S_R(V) = \sum_{v \in V} S_R(v)
\end{align*}
$S_H$, respectively $S_R$, are called the \emph{Hamming support} and \emph{rank support} of $v \in \mathbb F_{q^m}^n$ or $V \subseteq \mathbb F_{q^m}^n$.
\end{definition}
Once again, the rank support of an element $v \in \mathbb F_{q^m}^n$, or subset $V \subseteq \mathbb F_{q^m}$, is independent of the basis $\Gamma$ chosen. Furthermore, a linear block code, respectively $\mathbb F_{q^m}$-linear rank metric code, $\mathcal{C} \leq \mathbb F_{q^m}^n$ is said to be \emph{non-degenerate} if $S_H(\mathcal{C}) = [n]$, respectively $S_R(\mathcal{C}) = \mathbb F_q^n$.
Given a code $\mathcal{C} \leq \mathbb F_{q^m}^n$, it is of interest to consider the set of elements of $\mathcal{C}$ with a given support.
\begin{definition}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be a code, $A \subseteq [n]$ and $V \leq \mathbb F_q^n$. Let
\begin{align*}
\mathcal{C}_H(A) &:= \{v \in \mathcal{C} \, : \, S_H(v) = A\} \\
\mathcal{C}_R(V) &:= \{v \in \mathcal{C} \, : \, S_R(v) = V\}
\end{align*}
\end{definition}
We now make the connection between codes and ($q$-)matroids. Note that linear block codes and $\mathbb F_{q^m}$-linear rank metric codes can be represented via a generating metric $G \in \mathbb F_{q^m}^{k \times n}$, where $\mathcal{C} := \mathrm{rowsp}_{\mathbb F_{q^m}}(G)$.
Both a matroid and a $q$-matroid can be induced from the generating matrix. The following construction is well known for matroids (see \cite[Sec.6]{oxley}) and has been established in \cite[Sec. 5]{JP18} for $q$-matroids.
\begin{proposition}\label{rank representable}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be a code and $G \in \mathbb F_{q^m}^{k \times n}$ a generating matrix of $\mathcal{C}$. For $i \in [n]$, let $e_i \in \mathbb F_q^n$ denote the $i^{th}$ standard basis vector and for $V \leq \mathbb F_q^n$ let $Y_V \in \mathbb F_q^{n \times t}$ such that $\mathrm{colsp}_{\mathbb F_q}(Y_V) = V$. Define $r: [n] \rightarrow \mathbb N_0$ and $\rho: \mathcal{L}(\mathbb F_q^n) \rightarrow \mathbb N_0$ such that:
\begin{align*}
r(A) &= \mathrm{rk}_{\mathbb F_{q^m}} \left( G \cdot \begin{bmatrix} e_{i_1} &\cdots & e_{i_a} \end{bmatrix} \right) \textup{ for all } A \subseteq [n]\\
\rho(V) &= \mathrm{rk}_{\mathbb F_{q^m}}\left(G \cdot Y_V \right) \textup{ for all } V \leq \mathbb F_q^n.
\end{align*}
Then $M_{\mathcal{C}} := ([n], r)$ is a matroid and $\mathcal M_{\mathcal{C}} = (\mathbb F_q^n, \rho)$ is a $q$-matroid and are called the matroid (resp.\ $q$-matroid) associated with $\mathcal{C}$.
\end{proposition}
Note that neither $M_{\mathcal{C}}$ nor $\mathcal M_{\mathcal{C}}$ depend on the choice of generating matrix for $\mathcal{C}$.
Furthermore a matroid or $q$-matroid is $\mathbb F_{q^m}$-representable if it is induced by a linear block code or rank metric code $\mathcal{C} \leq \mathbb F_{q^m}^n$.
The ($q$-)matroid induced by a code is a useful tool to determine some of the code's invariants. In the rest of this section, we consider invariants of the code that are determined by the characteristic polynomial of the induce ($q$-)matroid.
We first recall the notion of weight enumerator of a ($q$-)matroid. The weight enumerator of the $q$-matroid was defined in \cite[Def. 43]{BCIJ21} and a similar concept was established in \cite{Greene} for matroids.
\begin{definition}\label{weight distribution matroid}
Let $M = (S, r)$ be a matroid and $\mathcal M = (E, \rho)$ be a $q$-matroid, with $|S| = n = \dim E$. Let
\begin{align*}
A_M^{(i)}(x) &= \sum_{A \subseteq S, |A| = i} \chi_{M/(S-A)}(x)\\
A_{\mathcal M}^{(i)}(x) &= \sum_{V \leq E, \dim V = i} \chi_{\mathcal M/V^{\perp}}(x)
\end{align*}
The \emph{weight enumerator} of the matroid, respectively $q$-matroid, is the list $A_M := (A_M^{(i)}(x) \, : \, 1 \leq i \leq n)$, respectively $A_{\mathcal M} := (A_{\mathcal M}^{(i)}(x) \, : \, 1 \leq i \leq n)$.
\end{definition}
Note in the above equations that if $S-A$ or $V^{\perp}$ are not flats of their respective matroid or $q$-matroid, then $\chi_{M/(S-A)}(x) = 0 = \chi_{\mathcal M/V^{\perp}}(x)$. Hence the summands of the weight enumerator can be restricted to the complement, respectively the orthogonal space, of flats. Thus we have the following result, which is well-known for matroids (see \cite[Prop 3.3]{Greene}) and was hinted at in \cite{BCIJ21} for $q$-matroids.
\begin{theorem}\label{weight enum flats}
Let $M = (S, r)$ be a matroid and $\mathcal M = (E, \rho)$ be a $q$-matroid, with $|S| = n = \dim E$. Then
\begin{align*}
A_M^{(i)}(x) &= \sum_{F \in \mathcal{F}_M, |F| =n- i} \chi_{M/F}(x),\\
A_{\mathcal M}^{(i)}(x) &= \sum_{F \in \mathcal{F}_{\mathcal M}, \dim F = n - i} \chi_{\mathcal M/F}(x).
\end{align*}
\end{theorem}
It was shown in \cite[Lem 49]{BCIJ21} that the weight enumerator of a representable $q$-matroid is closely related to the weight distribution of its associated code. For matroids a similar relation holds and was established in \cite[Prop 3.2]{Greene}.
\begin{theorem}\label{weight distribution equality}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be a code, $M$ and $\mathcal M$ be, respectively, the matroid and $q$-matroid induced by $\mathcal{C}$. Let $A \subseteq [n]$ and $V \leq \mathbb F_q^n$. Then
\begin{itemize}
\item[(1)] $\chi_{M/A}(q^m) = |\mathcal{C}_H([n]-A)|$ \quad and \quad $W_H^{(i)}(\mathcal{C}) = A_{M}^{(i)}(q^m)$.
\item[(2)]$ \chi_{\mathcal M/V}(q^m) = |\mathcal{C}_R(V^{\perp})|$ \quad and \quad $W_R^{(i)}(\mathcal{C}) = A_{\mathcal M}^{(i)}(q^m).$
\end{itemize}
\end{theorem}
\begin{remark}\label{support/flats}
The above Theorem together with Proposition \ref{weight enum flats}, tells us that if $|\mathcal{C}_H([n] -A)|$ and $|\mathcal{C}_R(V^{\perp})|$ are non-zero then $A$ and $V$ are flats in $M$ and $\mathcal M$, respectively.
\end{remark}
Because the weight enumerator of a ($q$-)matroid can be expressed in terms of flats, we get the following relation between the weight enumerator of a $q$-matroid and that of its projectivization matroid.
\begin{proposition}\label{weight enum q-matroid/proj}
Let $\mathcal M = (E, \rho)$ be a q-matroid and $P(\mathcal M) = (\mathbb{P}E , r)$ its projectivization matroid. Then
\begin{equation}
A_{P(\mathcal M)}^{(j)}(x) = \begin{cases} A_{\mathcal M}^{(i)}(x) & \textup{if } j = \frac{q^n - q^{n-i}}{q-1}, \\
0 & \textup{otherwise}.
\end{cases}
\end{equation}
\end{proposition}
\begin{proof}
First, recall by Lemma \ref{flats properties}, $\mathcal{F}_{P(\mathcal M)} = \{P(F) \,: \, F \in \mathcal{F}_{\mathcal M}\}$ and $\dim F =n- i \Leftrightarrow |P(F)| = \frac{q^{n-i} -1}{q-1} \Leftrightarrow |\mathbb{P}E - P(F)| = \frac{q^n - q^{n-i}}{q-1} $. Furthermore, by Proposition \ref{char project equiv}, if $X \subseteq \mathbb{P}E$ is not a flat, then $\chi_{P(\mathcal M)/X}(x) = 0$. So for all $1 \leq j \leq \frac{q^n-1}{q-1}$, such that $j \neq \frac{q^n - q^{n-i}}{q-1}$ for some $1 \leq i \leq n$, we get $A_{P(\mathcal M)}^{(j)}(x) = 0$.
Now assume $j = \frac{q^n-q^{n-i}}{q-1}$ for some $1 \leq i \leq n$. Then
\begin{align*}
A_{\mathcal M}^{(i)}(x) &= \sum_{F \in \mathcal{F}_{\mathcal M}, \dim F = n-i} \chi_{\mathcal M/F}(x) \\
&= \sum_{F \in \mathcal{F}_{\mathcal M}, \dim F = n-i} \chi_{P(\mathcal M)/P(F)}(x)\\
&= \sum_{P(F) \in \mathcal{F}_{P(\mathcal M)}, |P(F)| = \frac{q^{n-i}-1}{q-1}} \chi_{P(\mathcal M)/P(F)}(x)\\
&= A_{P(\mathcal M)}^{(j)}(x),
\end{align*}
where the second equality follows from Proposition \ref{char project equiv}.
\end{proof}
With the above setup, we now discuss the linear block code induced by an $\mathbb F_{q^m}$-linear rank metric code, as introduced in \cite{ABNR21}. In their paper, the authors use $q$-systems and projective systems to introduce the Hamming-metric code associated to a rank metric code.
We use a slightly different approach to introduce the associated Hamming metric code that does not require the previously stated notions.
\begin{definition}\label{Fq decomp}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be an $\mathbb F_{q^m}$-linear rank metric code and let $G \in \mathbb F_{q^m}^{k\times n}$ be a generating matrix of $\mathcal{C}$. Furthermore let $H \in \mathbb F_q^{n \times \frac{q^n-1}{q-1}}$ where each column of $H$ is a representative of a distinct element of $\mathbb{PF}_q^n$. We call the matrix $G^H := G \cdot H$ an \emph{$\mathbb F_q$-decomposition} of $G$ via $H$ and $\mathcal{C}^H:= \mathrm{rowsp}_{\mathbb F_{q^m}}(G^H)$ is called a \emph{Hamming-metric code associated to $\mathcal{C}$ via $H$}
\end{definition}
\begin{remark}
Given a non-degenerate $\mathbb F_{q^m}$-rank metric code $\mathcal{C}$, the code $\mathcal{C}^H$ of Definition \ref{Fq decomp} is a Hamming-metric code associated with $\mathcal{C}$ as in \cite[Def. 4.5]{ABNR21}. In fact, it easy to show the projective system induced by the columns of $G^H$ is a representative of the equivalence class $(\mathrm{Ext}^H \circ \Phi)([\mathcal{C}])$ as introduced in \cite{ABNR21}.
Furthermore note that unlike the construction in \cite{ABNR21}, Definition \ref{Fq decomp} does not depend on $q$-systems and projective systems hence we do not require $\mathcal{C}$ to be non-degenerate code.
Finally, as noted in \cite{ABNR21}, a Hamming metric code associated to an $\mathbb F_{q^m}$-linear rank metric code $\mathcal{C}$ is not unique. In our case, $\mathcal{C}^H$ depends on the choice of the matrix $H$ of the $\mathbb F_q$-decomposition. However, all Hamming-metric codes associated with $\mathcal{C}$ are monomially equivalent.
\end{remark}
Because $\mathcal{C}^H$ is a linear block code, it induces a matroid $M_{\mathcal{C}^H}$. It turns out that $M_{\mathcal{C}^H}$ is equivalent to the projectivization matroid $P(\mathcal M_{\mathcal{C}})$ of the $q$-matroid induced by $\mathcal{C}$.
\begin{theorem}\label{representable project}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be a rank metric code, $\mathcal M_{\mathcal{C}}$ its associated $q$-matroid, and $P(\mathcal M_{\mathcal{C}})$ its projectivization matroid. Furthermore let $\mathcal{C}^H$ be a Hamming-metric code associated to $\mathcal{C}$ via $H$ and $M_{\mathcal{C}^H}$ its induced matroid. Then
$$P(\mathcal M_{\mathcal{C}}) \cong M_{\mathcal{C}^H} \quad \textit{as matroids}.$$
\end{theorem}
\begin{proof}
Let $G \in \mathbb F_{q^m}^{k \times n}$ be a generator matrix of $\mathcal{C}$ and let $G^H = G \cdot H$ be a $\mathbb F_q$-decomposition of $G$ via $H$.
Let $h_i$ be the $i^{th}$ column of $H$, hence $\mathbb{PF}_q^n = \{ \langle h_i \rangle \, : \, i \in \left[ \frac{q^n-1}{q-1} \right]$. Define the bijection $\psi : \mathbb{PF}_q^n \rightarrow \left[\frac{q^n-1}{q-1} \right]$, where $\psi(\langle h_i \rangle) = i$.
Furthermore, let $A \subseteq \mathbb{PF}_q^n$, $\psi(A) = \{i_1, \cdots i_a\}$ and $Y_A := \begin{bmatrix} e_{i_1} & \cdots & e_{i_a} \end{bmatrix} \in \mathbb F_q^{\frac{q^n-1}{q-1} \times |A|}$, where $e_j$ is the $j^{\textup{th}}$ standard basis element of $\mathbb F_q^{\frac{q^n-1}{q-1}}$.
By Proposition \ref{rank representable} we get
\begin{align*}
r_{M_{\mathcal{C}^H}}(\psi(A)) &= \mathrm{rk}_{\mathbb F_{q^m}}(G^H \cdot \begin{bmatrix} e_{i_1}& \cdots & e_{i_a} \end{bmatrix})\\
&= \mathrm{rk}_{\mathbb F_{q^m}}(G \cdot H \cdot \begin{bmatrix} e_{i_1}& \cdots & e_{i_{a}} \end{bmatrix})\\
&=\mathrm{rk}_{\mathbb F_{q^m}}(G \cdot \begin{bmatrix} h_{i_1} & \cdots & h_{i_a} \end{bmatrix}) \\
&= \rho_{\mathcal M_{\mathcal{C}}}( \langle h_{i_1}, \cdots , h_{i_a} \rangle_{\mathbb F_q})\\
&=r_{P(\mathcal M_{\mathcal{C}})}(A),
\end{align*}
where the last equality follows from Theorem \ref{$q$-matroid/matroid}.
\end{proof}
\begin{remark}\label{relabel}
The above theorem allows us to relabel the groundset $\left[\frac{q^n-1}{q-1} \right]$ of the matroid $M_{\mathcal{C}^H}$ in terms of the elements of the projective space $\mathbb{PF}_q^n$. Precisely if $G \cdot H$ is the $\mathbb F_q$-decomposition associated to $\mathcal{C}^H$, relabel $i \in \left[\frac{q^n-1}{q-1} \right]$ by $\langle h_i \rangle \in \mathbb{PF}_q^n$, where $h_i$ is the $i^{\textup{th}}$ column of $H$.
\end{remark}
We get the following result as an immediate corollary of Theorem \ref{representable project}.
\begin{corollary}
If $\mathcal M$ is $\mathbb F_{q^m}$-representable then its projectivization matroid $P(\mathcal M)$ is $\mathbb F_{q^m}$-representable.
\end{corollary}
In \cite[Theorem 4.8]{ABNR21}, it was established that the rank-weight distribution of a rank metric code $\mathcal{C}$ is closely related to the Hamming-weight distribution of any Hamming-metric code associated to $\mathcal{C}$. By Theorem \ref{weight distribution equality}, Proposition \ref{weight enum q-matroid/proj} and Theorem \ref{representable project} we arrive at the same result from a purely matroid/$q$-matroid approach.
\begin{theorem}\cite[Thm 4.8]{ABNR21}
Let $\mathcal{C}$ be an $\mathbb F_{q^m}$-linear rank metric code and $\mathcal{C}^H$ be a Hamming-metric code associated to $\mathcal{C}$.Then
$$W_H^{(j)}(\mathcal{C}^H) = \begin{cases} W_R^{(i)}(\mathcal{C}) & \textup{if } j = \frac{q^n - q^{n-i}}{q-1},\\
0 & \textup{otherwise.}
\end{cases}$$
\end{theorem}
We conclude the paper by showing the $q$-analogue of the critical Theorem. The critical Theorem, introduced by Crapo and Rota \cite[Thm 1]{crapo1970foundations}, states that the characteristic polynomial of the matroid $M_{\mathcal{C}}$ induced by the linear block code $\mathcal{C}$ determines the number of multisets of codewords with a given support. It was nicely restated in \cite[Thm 2]{BRITZ200555} in terms of coding theory terminology (recall Definition \ref{support}).
\begin{theorem}\label{critical theorem}
Let $\mathcal{C} \leq \mathbb F_{q}^n$ be a linear block code and $M = ([n], r)$ its induced matroid. For all $A \subseteq [n]$, the number of ordered $t$-tuples $V = (v_1, \cdots , v_t)$, where $v_j \in \mathcal{C}$ for all $1 \leq j \leq t$, such that $S_H(V) = A$ is given by $\chi_{M/([n]-A)}(q^t)$.
\end{theorem}
For our last result, we show an analogous statement for $\mathbb F_{q^m}$-linear rank metric codes and $q$-matroids by using the projectivization matroid. It is worth mentioning that Alfarano and Byrne were able to show an analogue of the critical theorem for $q$-polymatroids and matrix rank metric codes by using a different approach involving the M\"obius inversion formula \cite{ABcrit} .
For the next results we make use of Remark \ref{relabel}, and relabel the elements of the groundset of the matroid induced by $\mathcal{C}^H$ in terms of elements of $\mathbb{PF}_q^n$. Following this relabeling, we can also describe the support of a codeword of $\mathcal{C}^H$ in terms of the elements of $\mathbb{PF}_q^n$. More precisely, if $\mathcal{C}^H$ is induced by the $\mathbb F_q$-decomposition $G \cdot H$, for any $v \in \mathcal{C}^H$, let $S_H(v) = \{ \langle h_i \rangle \in \mathbb{PF}_q^n \, : \, v_i \neq 0\}$, where $ h_i$ and $v_i$ are respectively the $i^{\textup{th}}$ column of $H$ and the $i^{\textup{th}}$ component of $v$. Furthermore, we need the following well-known result for which we include a proof for self-containment. For two vectors $v, w$ we let $v \cdot w$ denote the standard dot-product.
\begin{lemma}\label{dual support}
Let $v \in \mathbb F_{q^m}^n$, $S_R(v) = W \leq \mathbb F_q^n$ and $w \in \mathbb F_q^n$. Then $v \cdot w = 0$ if and only if $w \in W^{\perp}$.
\end{lemma}
\begin{proof}
Let $\Gamma := \{\gamma_1 \, \cdots, \gamma_m\}$ be a basis of $\mathbb F_{q^m}$ over $\mathbb F_q$, and let $Y:= \Gamma(v) \in \mathbb F_q^{n \times m}$, where $W:= \mathrm{colsp}_{\mathbb F_q}(Y)$. Then $v \cdot w = 0 \Leftrightarrow \sum_{i=1}^n v_iw_i = 0 \Leftrightarrow \sum_{i=1}^n \left( \sum_{j=1}^m \gamma_{j}v_{ij} \right)w_i = 0 \Leftrightarrow \sum_{j=1}^m \gamma_j \left(\sum_{i=1}^n v_{ij}w_i\right) = 0$. Since $\Gamma$ is a basis of $\mathbb F_{q^m}$ over $\mathbb F_q$ and $v_{ij}w_i \in \mathbb F_q$, the previous equality holds if and only if $\sum_{i=1}^n v_{ij}w_i = 0$ for all $1 \leq j \leq m$. But note $\sum_{i=1}^n v_{ij}w_i = v^{(j)} \cdot w$, where $v^{(j)}$ is the $j^{\textup{th}}$ column of $Y$. Hence $v \cdot w = 0 \Leftrightarrow v^{(j)} \cdot w = 0$ for all $1 \leq j \leq m \Leftrightarrow w \in \mathrm{colsp}_{\mathbb F_q}(Y)^{\perp} \Leftrightarrow w \in W^{\perp}$.
\end{proof}
The following Lemma relates the rank support of elements of the code $\mathcal{C}$ with the Hamming support of elements of the associated Hamming-metric code $\mathcal{C}^H$.
\begin{lemma}\label{rank sup/ ham sup}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be a rank metric code, $\mathcal{C}^H$ be a Hamming-metric code associated to $\mathcal{C}$ via $H$. Furthermore, let $V = \{v_1, \cdots , v_t\}$ be a subset of $\mathcal{C}$ and $V \cdot H:= \{v_1 \cdot H, \cdots, v_t \cdot H\}$ . Then
$$S_R(V) = W \Leftrightarrow S_H(V \cdot H) = \mathbb{PF}_q^n - P( W^{\perp}).$$
Moreover if $\mathcal M$ and $P(\mathcal M)$ are the q-matroid and projectivization matroid induced respectively by $\mathcal{C}$ and $\mathcal{C}^H$ then $W^{\perp}$ and $P( W^{\perp})$ are, respectively, flats of $\mathcal M$ and $P(\mathcal M)$.
\end{lemma}
\begin{proof}
Consider the subset $V := \{v_1, \cdots , v_t\} \subseteq \mathcal{C}$. By definition, $S_R(V) = \sum_{j=1}^t S_R(v_j) =: W$.
Let $W_j := S_R(v_j)$. For all $1 \leq j \leq t$, by Lemma \ref{dual support}, $v_j \cdot w = 0$ if and only if $w \in W_j^{\perp}$ . Hence for all columns $h_i$ of $H$, it follows that $v_j \cdot h_i = 0 $ if and only if $ h_i \in W_j^{\perp}$. By definition, this is true if and only if $S_H(v_j \cdot H) = \mathbb{PF}_q^n - P(W_j^{\perp})$.
Hence $S_H( V \cdot H) = \bigcup_{i=1}^t S_H(v_j \cdot H) = \bigcup_{j=1}^t (\mathbb{PF}_q^n - P(W_j^{\perp})),$ where the first equality follows by definition.
Therefore $S_R(V) = W \Leftrightarrow S_H(V \cdot H) = \bigcup_{j=1}^t (\mathbb{PF}_q^n - P(W_j^{\perp})) = \mathbb{PF}_q^n - (\bigcap_{j=1}^t P(W_j^{\perp})) = \mathbb{PF}_q^n - P(W^{\perp})$.
Finally, $W^{\perp}$ and $P(W^{\perp})$ are flats of $\mathcal M$ and $P(\mathcal M)$ respectively because of Remark \ref{support/flats}.
\end{proof}
We are now ready to show the critical theorem for $\mathbb F_{q^m}$-linear rank metric codes and $q$-matroids.
\begin{theorem}
Let $\mathcal{C} \leq \mathbb F_{q^m}^n$ be an $\mathbb F_{q^m}$-linear rank metric code and $\mathcal M$ its induced $q$-matroid. For all $W \leq \mathbb F_q^n$, the number of ordered $t$-tuples $V = (v_1, \cdots , v_t)$, where $v_j \in \mathcal{C}$ for all $1 \leq j \leq t$, such that $S_R(V) = W$ is given by $\chi_{\mathcal M/W^{\perp}}(q^{mt})$.
\end{theorem}
\begin{proof}
Let $\mathcal{C}^H$ be the Hamming-metric code associated with $\mathcal{C}$ via $H$ and $P(\mathcal M)$ be its associated matroid. Note that every element of $\mathcal{C}^H$ is of the form $v \cdot H$ for some $v \in \mathcal{C}$. Hence every tuple of elements of $\mathcal{C}^H$ is of the form $V \cdot H$ for some $V \subseteq \mathcal{C}$. Furthermore, since $H$ has full-row rank, there is a bijection between elements of $\mathcal{C}$ and $\mathcal{C}^H$ and hence a bijection between $t$-tuples $V \subseteq \mathcal{C}$ and $t$-tuples $V \cdot H \subseteq \mathcal{C}^H$. By Theorem \ref{critical theorem}, the number of $t$-tuple $V \cdot H \subseteq \mathcal{C}^H$ such that $S_H(V \cdot H) = \mathbb{P}E - P(W^{\perp})$ is given by $\chi_{P(\mathcal M) / P(W^{\perp})}(q^{mt})$. Moreover, by Proposition \ref{char project equiv}, $\chi_{P(\mathcal M)/P(W^{\perp})}(q^{mt}) = \chi_{\mathcal M / W^{\perp}}(q^{mt})$.
Finally, by Lemma \ref{rank sup/ ham sup}, $S_R(V \cdot H) = \mathbb{P}E - P(W^{\perp})$ if and only if $S_R(V) = W$ and therefore $\chi_{\mathcal M / W^{\perp}}(q^{mt})$ counts the number of $t$-tuples $V \subseteq \mathcal{C}$ such that $S_R(V) = W$.
\end{proof}
The above result, and its proof, shows a close connection between the critical theorem for matroids and that for $q$-matroids. Furthermore it can easily been seen from the above approach that the critical problem for $q$-matroids, that is finding the smallest power of $q$ that makes the characteristic polynomial of a $q$-matroid non-zero, is a specific case of the critical problem for matroids.
\section*{Further questions}
Here are a few questions that arise from our work:
\begin{itemize}
\item If a projectivization matroid $P(\mathcal M)$ is $\mathbb F_{q^m}$-representable then is the $q$-matroid $\mathcal M$ representable?
\item It is well known that the characteristic polynomial of a matroid can be derived from the Tutte polynomial of that matroid. Hence the characteristic polynomial of a $q$-matroid $\mathcal M$ can also be derived from the Tutte polynomial of the projectivization matroid $P(\mathcal M)$. Can invariants of $\mathbb F_{q^m}$-rank metric codes be determined from the Tutte polynomial of the projectivization matroid associated to the code.
\end{itemize}
\bibliographystyle{plain}
|
\section{Introduction}
Search for a theoretical framework incorporating general relativity and quantum theory has proven to be one of the most difficult undertakings in physics. A common assumption behind such a pursuit is that gravity itself must have a quantum nature. In fact, the possibility of schemes in which matter fields are treated in quantum terms, but gravity is treated classically, has been argued against on several grounds \cite{eppley1977necessity,page1981indirect}. However, those arguments have been found to be less convincing than intended \cite{huggett2001quantize,mattingly2006eppley,carlip2008quantum}.
Quite independently of whether gravity ought to have a quantum nature at the fundamental level, questions concerning the viability of semiclassical gravity as an approximated description, and its precise regime of applicability, are of clear practical and conceptual interest. It is clear that general relativity works exquisitely well in a large set of regimes, including the solar system, binary pulsars, black hole collisions, the generation of gravity waves and sectors of cosmology. All these applications rely on a classical treatment of matter, but we know that matter is better described via a quantum treatment. Therefore, there must be a regime for which semiclassical gravity is an excellent approximation. Moreover, there are scenarios in which semiclassical gravity is employed (mostly) without raising objections, such as in the Hawking evaporation of black holes (at least in the relatively early stages).
At any rate, as with all physical questions, the final verdict regarding the fundamental nature of gravity must depend on experimental evidence arising from situations where both quantum theory and gravitation play a relevant role. The standard expectation is that such situations only emerge in phenomena involving extremely high energies or when curvature values approach the Planck scale (i.e., $ R\sim 1/m_{p}^2$), both of which are currently well beyond our empirical reach. However, there have been recent proposals to look for a possible quantum behaviour of gravity directly in a tabletop experiment, \cite{bose2017spin,Marletto}. In the meantime, there have also been proposals suggesting that useful hints might be acquired by exploring gedankenexperiments, together with well-established physical principles. The general idea is to obtain clues regarding the nature of gravity, by thinking about gravitational fields associated with matter sources in states that require a quantum mechanical treatment, \cite{cecile2011role,zeh2011feynman}.
A concrete instance of such an approach has been explored in some detail in \cite{mari2016experiments,belenchia2018quantum,wald2020quantum}. The gedankenexperiment considered involves two observers, one in control of an object placed in a quantum superposition of two spatial locations, and the other deciding whether a second object is allowed to react to its (electromagnetic or gravitational) interaction with the first. The setup in such that an interaction between the objects would prevent the observability of an interference pattern for the first object. And since the decision to allow the objects to interact or not can be taken with spacelike separation from the first object, that would seem to allow for superluminal signaling.
In \cite{mari2016experiments,belenchia2018quantum,wald2020quantum}, however, it is argued that, if one attributes quantum properties to the mediating fields, then the possibility for superluminal signaling is fully avoided. In particular, the claim is that taking into account the quantization of radiation of the fields, and the existence of their vacuum quantum fluctuations, allows for the undesired conclusion to be evaded. Moreover, in the gravitational case, this conclusion is used to argue for the view that the gravitational field must be given a quantum description. It is worth noting that this last claim is very strong indeed, as it goes well beyond the ``mere'' notion that ``everything is quantum at the fundamental level''. In particular, it seems to immediately discard the possibility, as proposed in \cite{Jacobson}, for the gravitational field to be an emergent entity, unsuitable for a quantum description.\footnote{We do seem to frequently admit the existence of objects of that nature, such as an ecological system, a nation's economy or even something as simple as the heat flow in an ordinary material, described by a ``field" which one would not attempt to quantize.}
Considerations like this motivate us to reexamine the seemingly dramatic conclusions reached in \cite{mari2016experiments,belenchia2018quantum,wald2020quantum}. It is not that we necessarily disagree on there being a high likelihood for everything ultimately having a fundamental quantum nature. However, we find it quite surprising that, at this stage in the history of physics, and via a simple gedankenexperiment, such robust conclusions could be reached. With this in mind, we explore what we perceive as weak points in the arguments that have been put forward in the works mentioned above.
In particular, we first point out that, in order for the signaling protocol to get off the ground, one needs to \emph{presuppose} the quantum nature of the fields. Otherwise, the two objects would not get entangled and the interaction between them would not destroy the interference pattern. Therefore, the fact that considering the quantum nature of the fields might eliminate the possibility of signaling cannot be used to argue that gravity must be quantized. Second, we note that the discussions in \cite{mari2016experiments, belenchia2018quantum,wald2020quantum} do not consider certain decoherence effects, inherent in the experimental design, that could prevent the protocol from actually working. In particular, taking into account the trap used to determine whether the objects are allowed to react to the interaction or not could prevent control of the signal. Finally, by explicitly constructing a version of the gedankenexperiment in which the second observer controls many particles, instead of one, and where the quantization of the fields and vacuum fluctuations do not avoid signaling, we show that the solution for the apparent superluminal communication presented in \cite{belenchia2018quantum,wald2020quantum} is defective (or at least incomplete).
Our manuscript is organized as follows. In section \ref{GE} we describe the gedankenexperiment, as discussed in \cite{belenchia2018quantum,wald2020quantum} and review their proposal for avoiding the (apparent) superluminal signaling. Then, in section \ref{EV}, we present our evaluation of such a proposal, pointing out what we take to be important limitations. Finally, in section \ref{Co} we offer our conclusion.
\section{The gedankenexperiment and the solution proposed by Belenchia \emph{et al.}}
\label{GE}
In \cite{mari2016experiments}, Mari \emph{et al.} proposed and analyzed a gedankenexperiment in which the (electromagnetic or gravitational) interaction between two objects, one placed in a quantum superposition of two locations, apparently allows for superluminal communication. The experiment was later reanalyzed by Belenchia \emph{et al.} in \cite{belenchia2018quantum} (see also \cite{wald2020quantum}). Although the settings considered are slightly different, the basic idea, as well as the conclusions reached, have the same spirit: if one takes into account the quantization of the radiation emitted by the fields, as well as the existence of their vacuum fluctuations, then the apparent possibility of superluminal signaling is completely avoided. Here, we focus on the analysis developed in \cite{belenchia2018quantum,wald2020quantum}. In what follows we review the gedankenexperiment considered in such works, as well as the details of the solution advocated by them.
\subsection{The gedankenexperiment}
The gedankenexperiment considered in \cite{belenchia2018quantum,wald2020quantum} has two versions, one electromagnetic and one gravitational. Both contain two observers, Alice and Bob, separated by a distance $D$. Alice has control over a particle with spin, charge $q_{A}$ and mass $m_{A}$ and Bob over a particle with charge $q_{B}$ and mass $m_{B}$. In the electromagnetic case, all gravitational effects are ignored; in the gravitational one, the charges is set to zero. We work in units with $\hbar = c = 1$.
The experiment starts by assuming that, in the distant past, Alice's particle was sent through a Stern-Gerlach apparatus, leaving its state in the superposition $\frac{1}{\sqrt{2}} (\ket{L}_{A} \ket{\downarrow}_{A}+\ket{R}_{A} \ket{\uparrow}_{A})$ with distance $d$ between $\ket{L}_{A}$ and $\ket{R}_{A}$. This step is assumed adiabatic, with negligible radiation emitted. Bob's particle, on the other hand, is initially assumed to be held on a strong trap, so its interaction with Alice's particle is negligible. The experiment then proceeds as follows. At time $t=0$, Bob decides whether to release his particle from the trap or leave it there; we call $T_B$ the time at which Bob completes his experiment. Also at $t=0$, Alice starts an interference experiment with her particle, which ends at time $T_A$ (see Figure 1).
\begin{figure}[!ht]
\centerline{\includegraphics[scale=0.35]{figP}}
\caption{Spacetime diagram of the gedankenexperiment.}
\label{fig}
\end{figure}
Now, if Bob decides to release his particle, it will get entangled with the fields produced by the components $\ket{R}_{A}$ and $\ket{L}_{A}$ of Alice's particle, which would put Bob's particle in a spatial superposition with separation $\delta x$. If $\delta x$ is large enough, the states of that superposition would be nearly orthogonal, so Bob's particle would be nearly maximally entangled with Alice's, which would prevent her from observing interference. If, on the other hand, Bob does not release his particle, nothing would prevent Alice from observing interference. Then, Bob's decision to release or not the particle determines whether Alice observes interference or not. Therefore, if $T_{A}, T_{B} < D$, in which case the experiments of Alice and Bob would be spacelike separated, Alice and Bob would have access to a superluminal channel.\footnote{In \cite{belenchia2018quantum}, this result is presented as a tension between complementarity and causality: if complementarity is the case, superluminal signaling would be possible.}
\subsection{The solution proposed by Belenchia \emph{et al.}}
Below we review the solution proposed in \cite{belenchia2018quantum,wald2020quantum}, first for the electromagnetic case and then for the gravitational one.
\subsubsection{The electromagnetic case}
For the electromagnetic case, \cite{belenchia2018quantum,wald2020quantum} treat the particles with non-relativistic quantum mechanics and the electromagnetic field as a relativistic quantum field. Then, at $t=0$, they take the state of the whole system to be given by
\begin{equation}
\ket{\Psi}=\frac{1}{\sqrt{2}} \left[ \ket{L}_{A} \ket{\downarrow}_{A} \ket{\alpha_{L}}_{F} + \ket{R}_{A} \ket{\uparrow}_{A} \ket{\alpha_{R}}_{F} \right] \ket{\psi_{0}}_{B},
\end{equation}
with $\ket{\psi_{0}}$ the state of Bob's particle inside the trap and $\ket{\alpha_{L}}_{F}$ and $\ket{\alpha_{R}}_{F}$ the states of the electromagnetic field associated with each term of the spatial superposition of Alice's particle. They point out that, typically, it will be the case that $| \braket{\alpha_{L}|\alpha_{R}}_F | \ll 1$. Therefore, in this sense, Alice's particle will have decohered, even before Bob could make a decision. However, \cite{belenchia2018quantum,wald2020quantum} argues that this would be a case of what \cite{unruh2000false} calls ``false decoherence'': if Alice recombines her particle adiabatically, then the fields would ``follow'' the particle and would allow for a complete recombination.
Now, in order to determine the effects of Bob opening the trap on the decoherence of Alice's particle, \cite{belenchia2018quantum,wald2020quantum} claim that there are two properties of the quantum electromagnetic field that play a crucial role: vacuum fluctuations and the quantization of the field. Regarding the former, it is argued that, due to the inevitable vacuum fluctuations of the electromagnetic field, a charged particle cannot be localized to better than its charge-radius $q/m$. Since this would be true for Bob's particle, in order for him to be able to destroy Alice's coherence, he needs for the displacement of his particle to be larger than its delocalization, i.e.,
\begin{equation}
\label{dx}
\delta x > \frac{q_B}{m_B} .
\end{equation}
In order to estimate $\delta x$, it is argued that it will depend on the difference between the resultant electromagnetic fields of Alice's particle, $ E \sim \mathcal{D}_A/D^3$, with the corresponding ``electric dipole moment'' given by $\mathcal{D}_A = q_A d$. Since Bob's particle is released for a time $T_B$,
\begin{equation}
\delta_x \sim \frac{q_B}{m_B} \frac{\mathcal{D}_A}{D^3} T_B^2,
\end{equation}
so, in order to ensure Eq. (\ref{dx}), it is necessary that
\begin{equation}
\label{BC}
\frac{\mathcal{D}_A}{D^3}T_B^2>1.
\end{equation}
Regarding the existence of quantized electromagnetic radiation, it is argued that for Alice to be able to coherently recombine her particle, she must be able to do the recombination avoiding the emission of even one photon. To estimate the amount of radiation emitted, it is noted that the total radiated energy would be given by $E \sim \mathcal{D}_A^2/T_A$. As this energy is quantized in photons with frequency $\sim 1/T_A$, the number of radiated photons would be of order $ \left(\mathcal{D}_A/T_A\right)^{2}$. Thus, it is concluded that, in order for Alice to maintain the coherence of her particle when recombining, it is necessary to have
\begin{equation}
\label{AC}
\mathcal{D}_A<T_A.
\end{equation}
Now, for Bob to be able to influence Alice's interference experiment superluminally, we need $T_A < D$ and $T_B < D$. In that case, there are two options, $\mathcal{D}_A < T_A$ and $\mathcal{D}_A > T_A$. If $\mathcal{D}_A < T_A$, by Eq. (\ref{AC}), Alice can recombine her particle without emitting radiation but, because of Eq. (\ref{BC}), Bob is unable make a difference, so no signaling is possible. If, on the other hand, $\mathcal{D}_A > T_A$, then, by Eq. (\ref{AC}), Alice will not see interference independently of what Bob does, once more avoiding signaling.
\subsubsection{The gravitational case}
The treatment of the gravitational case in \cite{belenchia2018quantum,wald2020quantum} follows closely the electromagnetic one. The main difference being that, because of conservation of stress-energy, and taking into account the entanglement of Alice with her lab, it is argued that the effective mass dipole resulting from the superposition of Alice's particle is zero. Therefore, the gravitational effect on Bob's particle and the radiation emission are taken to be mediated by the quadrupole moment $\mathcal{Q}_{A}$.
The analysis then proceeds as follows. First it is argued that, due to the inevitable vacuum fluctuations of the gravitational field, Bob's particle cannot be localized to better than the Planck length $l_{p} \sim 10^{-35}m$, \cite{calmet2004minimum}. Consequently, in order for him to be able to destroy Alice's coherence, he needs for the displacement of his particle to be larger than its delocalization, i.e.,
\begin{equation}
\delta x > l_{p} .
\end{equation}
Since the separation of Bob's components during time $T_B$ is estimated to be given by
\begin{equation}
\delta_x \sim \frac{\mathcal{Q}_A}{D^4} T_B^2,
\end{equation}
it is concluded that Bob will be able to cause decoherence only when (setting $l_p = 1$)
\begin{equation}
\label{BG}
\frac{\mathcal{Q}_A}{D^4}T_B^2 >1.
\end{equation}
On the other hand, as in the electromagnetic case, it is argued that Alice cannot recombine her particle arbitrarily fast. Otherwise, her particle would radiate, which would cause decoherence. In this gravitational case, the energy radiated is taken to be given by $E \sim (\mathcal{Q}_A/T_A^3)^2 T_A$, with the corresponding number of gravitons of order $ (\mathcal{Q}_A/T_A^2)^2$. Therefore, in order to avoid the emission of even one graviton, the time that takes for her to do the experiment must satisfy
\begin{equation}
\label{AG}
\mathcal{Q}_{A}<T_{A}^{2}.
\end{equation}
Finally, in analogy with the electromagnetic case, Eqs. (\ref{BG}) and (\ref{AG}) are argued to prevent superluminal communication between Alice and Bob. For that we would need $T_A < D$ and $T_B < D$ but, if Eq. (\ref{AG}) holds, then $\mathcal{Q}_A<D^2$ so, from Eq. (\ref{BG}), that means that Bob cannot disrupt Alice's interference in time for signaling to occur.
With all this, in \cite{belenchia2018quantum,wald2020quantum} it is concluded that, by postulating the quantization and the existence of quantum vacuum fluctuations of the gravitational field, the worry of superluminal signaling is completely avoided. Such a conclusion is then read as providing support for the view that the gravitational field must be given a quantum field description.
\section{Our evaluation}
\label{EV}
In this section, we point out important limitations of the previous assessment. We start by pointing out that, for the signaling protocol to work, one has to \emph{presuppose} that the fields have a quantum nature. Therefore, the use of the protocol to argue that gravity must be quantized is, at best, circular. Next, we argue that there are certain decoherence effects not considered, produced by the presence of the trap and surrounding objects, that could completely derail the protocol. Finally, by considering a version of the gedankenexperiment with $N$ traps, instead of one, we show that, contrary to what is argued in \cite{belenchia2018quantum,wald2020quantum}, taking into account the quantization of the fields and their vacuum fluctuations does not help in avoiding superluminal signaling.
\subsection{Is gravity quantum?}
As we saw above, according to \cite{belenchia2018quantum,wald2020quantum}, consideration of the quantization and the existence of quantum vacuum fluctuations of the gravitational field explains away the alleged possibility of superluminal signaling in the protocol under discussion. Moreover, such a result is promoted as lending support to the idea that the gravitational field must be given a quantum description. There is, however, a difficulty with this logic.
The issue is that, in order for the signaling protocol to function, one needs to \emph{presuppose} that the field in question (gravitational in this case) has a quantum nature. To see this, we note that it is only because one assumes that the field generated by the superposition of Alice's particle is also described by a corresponding quantum superposition of states---that is, that the gravitational field gets entangled with Alice's particle---that Bob's particle also ends up entangled with Alice's particle. Otherwise, Bob's particle would not get entangled with Alice's and Bob would be unable to destroy Alice's coherence.
For instance, in some sort of semiclassical description, in which matter fields are quantum, gravity remains classical, and where the field produced by the superposition of Alice's particle is some sort of \emph{average} of the two terms, Bob's particle would simply respond to that single field and not get entangled with Alice's particle. That is, if the Einstein curvature tensor for the spacetime metric is sourced by the expectation value of the energy-momentum, rather that the individual localized components of the state of matter, no non-trivial entanglement of the sort assumed by the communication protocol will be ever generated.
One could argue that semiclassical frameworks of this sort are already ruled out by works such as \cite{eppley1977necessity,page1981indirect}. However, as we mentioned above, such arguments are not as conclusive as intended.\footnote{Even though existing arguments against semiclassical gravity are not conclusive, they do set constraints on semiclassical frameworks (see also \cite{Tim} for a general assessment of the issue).} In any case, this is supposed to be an argument against semiclassical gravity, so it cannot depend upon assuming that semiclassical gravity is not viable.
Now, given that the protocol presupposes a quantum description of the field, it is clear that then using such a quantum nature to avoid signaling should be seen more like a proof of self-consistency, and cannot in any way be used to argue for the quantization of the field in the first place. In other words, in the argument, the conclusion is already contained in the hypotheses, so the argument is, at best, circular.
\subsection{Decoherence effects}
In the gedankenexperiment under discussion, the influence of any object \emph{external} to the setup itself, is completely neglected. This idealization is assumed in the theoretical treatment of almost all physics experiments. Since it is impossible to completely isolate any experimental setup from the outside, strictly speaking, no theoretical description is completely faithful to the experimental situation it intends to reproduce. However, in most cases, the disregard of external influences is fully justified, either by a comparison of the sizes of the internal and external influences, or by a systematization of external agents.
A crucial component of the proposed gedankenexperiment is the test of coherence of Alice's particle. Such a procedure is extremely sensitive to all sorts of external influences. Even worse, external effects are indistinguishable from internal ones, so it is not clear that the standard practice of neglecting external influences, is valid in this case. In fact, even if complete isolation could be achieved in practice, the analysis disregards the effect that Bob's trap itself could have on the coherence of Alice's particle.
There is, in fact, a direct way in which the trap could play a role. Since Alice's particle is in a superposition of two different locations, each influencing Bob's particle with different forces, the trap must also enter into a superposition of two different states. Thus, at $t=0$ the state of the system must be
\begin{equation}
\ket{\psi}=\frac{1}{\sqrt{2}} \left[ \ket{L}_{A} \ket{\downarrow} \ket{\alpha_{L}}_{F} \ket{T_L}_B + \ket{R}_{A} \ket{\uparrow} \ket{\alpha_{R}}_{F} \ket{T_R}_B \right] ,
\end{equation}
where $\ket{T_L}_B$ is the state of the trap when Bob's particle interacts with Alice's $\ket{L}_A$ component and $\ket{T_R}_B$ the state when interacts with Alice's $\ket{R}_A$ component. It is possible, then, that even if Bob doesn't release his particle, the trap itself causes decoherence in Alice's system.
One might alternatively reformulate the above consideration in the following terms: the ability of Bob's particle to cause decoherence in Alice's experiment depends on the generation of significant entanglement between them, caused by the action of the field of Alice's particle. Therefore, this decoherence effect could happen with any other object under the influence of Alice's field. In particular, the trap---with its millions of constituent particles---could cause this effect, preventing the establishment of a communication channel, even if Bob doesn't release his particle.
Another issue worth examining is the claim in \cite{belenchia2018quantum,wald2020quantum} that, when Alice recombines her particle slowly, the fields undergo so-called ``false decoherence''. The idea is that, while it is asserted that, when Alice's particle is in a superposition, the fields corresponding to the two components of the superposition are nearly orthogonal, it is also argued that, when Alice's particle is adiabatically recombined, then the fields ``follow'' the particle, allowing for a perfect recombination. The problem is that the analysis cited to support such a claim, \cite{unruh2000false}, is performed in a non-relativistic setting, with fields propagating instantaneously. It seems far from clear, then, whether such an analysis extends to the relativistic scenario under discussion. In fact, it seems quite possible that, if the two components of the fields were nearly orthogonal at some point in time, then they would continue to be so on any spacial hypersurface containing the recombination of Alice's particle. After all, in regions sufficiently far away from the recombination event, the state of the field would not have had time to change as a result of the recombination, simply because such information is expected to travel causally. Any failure of the previous expectation would seem to offer, by itself, a path for superluminal signaling.
In fact, it seems to us that, the only way in which Alice would ever be able to fully recombine her particle, is when the fields corresponding to the two components of the superposition of her particle are \emph{never} orthogonal. If that is so, it would mean that, if Alice is able to recombine her particle, then Bob's particle would not be able get entangled with it, avoiding from the very beginning the possibility of superluminal signaling.
\subsection{Multiple traps}
To conclude our analysis, we present a variation of the gedankenexperiment in which we show that the solution offered in \cite{belenchia2018quantum,wald2020quantum} must be invalid, or at least incomplete. To do so, we grant all claims in such works, but consider a scenario in which Bob, instead of controlling one particle, controls $N$ particles. What we show is that, if Bob controls a large enough number of particles, then he can cause decoherence on Alice's experiment at time $T_B<D$, even if $\mathcal{D}_A<T_A<D$. If so, he would be able to send a superluminal bit of information to Alice---even if the quantization of the field and its fluctuations are considered.
According to the analysis in \cite{belenchia2018quantum,wald2020quantum}, the vacuum fluctuations of the fields induce a limit on the localization of a particle in them. Such limits are argued to be given by $\sigma=q/m$, in the electromagnetic case, and by $\sigma=l_P$ in the gravitational one. As a result of this, it is argued that, in order for Bob to be able decohere Alice's particle, the separation in his superposition, $\delta x$, must be larger than $\sigma$. Finally, it is claimed that, if Alice recombines her particle without emitting radiation, Bob will not be able to obtain a $\delta x$ large enough, in time to disrupt Alice's recombination.
Suppose, then, that Alice does recombine her particle without emitting radiation. In that case, when Bob releases his particle, he will obtain a $\delta x$ smaller than $\sigma$, so the inner product of the states $\ket{L}_B$ and $\ket{R}_B$ of Bob's superposition will satisfy
\begin{equation}
|\braket{L|R}_B|=1-\epsilon
\end{equation}
for some $\epsilon \ll 1$.
Suppose, now, that Bob has not one, but $N$ particles. If he decides to release them, the state of the particles becomes
\begin{equation}
\ket{\Psi_N}=\frac{1}{\sqrt{2}}(\ket{L}_A \ket{L_1}_B \ket{L_2}_B... \ket{L_N}_B+\ket{R}_A \ket{R_1}_B \ket{R_2}_B... \ket{R_N}_B) .
\end{equation}
Assuming the same conditions for the $N$ particles than we had for the single particle above, the inner product of the right and left components of Bob's particles then becomes
\begin{equation}
\Pi_i |\braket{L_i|R_i}_B|=(1-\epsilon)^N \approx 1-N\epsilon.
\end{equation}
Clearly, if $N$ is large enough, this inner product approaches zero, so Bob's particles become able to cause decoherence of Alice's particle. And this is so, even maintaining the experiments of Alice and Bob at a spacelike distance and with Alice avoiding her particle to radiate. That is, if $N$ is large enough, a superluminal communication channel between Alice and Bob would be possible. We conclude that, contrary to what is claimed in \cite{belenchia2018quantum,wald2020quantum}, consideration of quantization and vacuum fluctuations of the fields in not enough to forbid superluminal signaling.
\section{Conclusions}
\label{Co}
Recent works explore a gedankenexperiment in which the interaction of a particle with the field of a charged or massive object in a quantum superposition, seems to allow for superluminal communication. Building on a previous analysis in \cite{mari2016experiments}, \cite{belenchia2018quantum} and \cite{wald2020quantum} argue that, if one considers the quantization of the radiation of the fields in question, together with the presence of quantum vacuum fluctuations of such fields, then the alleged signaling disappears. Moreover, in the gravitational case, the result that quantization and vacuum fluctuations of the gravitational field are required to avoid signaling is promoted as an argument in favor of the necessity to quantize the gravitational field.
In this work, we have identified a number of limitations of the aforementioned conclusions. First, we point out that, in order for the proposed signaling protocol to function, it is necessary to \emph{presuppose} the quantum nature of the fields. Therefore, the conclusion that the attribution of quantum properties to the fields might remove the possibility of signaling cannot be used as an argument for the quantization of gravity.
Next, we note that the discussions in \cite{belenchia2018quantum,wald2020quantum} of the gedankenexperiment fail to take into consideration possible crucial decoherence effects caused by one of the components of the experiment. Moreover, we question some claims regarding the alleged ``false decoherence'' process, supposedly undergone by the quantum states of the fields involved in the experiment. Finally, by considering a version of the gedankenexperiment in which Bob has control over $N$ particles, instead of just one, we show that the considerations involving the quantization of the fields and vacuum fluctuations do not eliminate the possibility of superluminal signaling. We conclude that the assessment of the gedankenexperiment offered in \cite{belenchia2018quantum,wald2020quantum}, as well as the proposed solution, are defective, or at least incomplete.
\section*{Acknowledgments}
We acknowledge support from CONACYT grant 140630. DS acknowledges partial financial support from PAPIIT-DGAPA-UNAM project IG100120 and the grant FQXI-MGA-1920 from the Foundational Questions Institute and the Fetzer Franklin Fund, a donor advised by the Silicon Valley Community Foundation.
\bibliographystyle{ieeetr}
|
\section{Introduction}\label{section1}
Let $\mathbb{X}, \mathbb{Y}$, and $\mathbb{Z}$ be real finite dimensional Euclidean spaces with an inner product $\langle \cdot,\cdot \rangle $ and its induced norm $\|\cdot\|$.
Let $f_1 : \mathbb{X} \rightarrow ( - \infty, + \infty)$ and $h_1 : \mathbb{Y} \rightarrow ( - \infty, + \infty)$ are convex functions with Lipschitz continuous gradients; $f_2 : \mathbb{X} \rightarrow ( - \infty, + \infty]$ and $h_2 : \mathbb{Y} \rightarrow ( - \infty, + \infty]$ are closed proper convex (not necessarily smooth) functions.
We consider the following composite convex optimization problem
\begin{equation}\label{prob1}
\min_{x\in\mathbb{X},y\in\mathbb{Y}} \ \big\{ f_1(x)+f_2(x) + h_1(y)+h_2(y) \ | \ {\mathcal A}^*x + {\mathcal B}^*y = c \big\},
\end{equation}
where ${\mathcal A} : \mathbb{Z} \rightarrow \mathbb{X}$ and ${\mathcal B} : \mathbb{Z} \rightarrow \mathbb{Y}$ are linear operators with adjoints ${\mathcal A}^*$ and ${\mathcal B}^*$, respectively; and $c\in\mathbb{Z}$ is a given vector.
Let $\sigma\in( 0 , +\infty)$ be a penalty parameter, the augmented Lagrangian function associated with problem (\ref{prob1}) is defined by, for any $(x,y,z)\in\mathbb{X}\times\mathbb{Y}\times\mathbb{Z}$,
\begin{equation}\label{alf}
{\mathcal L}_{\sigma}(x,y;z) := f_1(x)+f_2(x) + h_1(y)+h_2(y) + \langle z, {\mathcal A}^*x + {\mathcal B}^*y - c\rangle +\frac{\sigma}{2}\|{\mathcal A}^*x + {\mathcal B}^*y - c\|^2,
\end{equation}
where $(x,y,z)\in\mathbb{X}\times\mathbb{Y}\times\mathbb{Z}$ and $z$ is a multiplier.
One attempt to solve (\ref{prob1}) is the standard alternating direction method of multipliers (ADMM), which alternately updates the variables $(x,y)$ and the multiplier $z$ from an initial point $(x^0,y^0,z^0)\in \dom(f)\times \dom(h)\times\mathbb{Z}$, that is,
\begin{equation}\label{cadmm}
\left\{
\begin{array}{l}
x^{k+1} = \argmin_{x\in\mathbb{X}}\big\{{\mathcal L}_{\sigma}(x,y^k;z^k)=:f_1(x)+f_2(x)+\frac{\sigma}{2}\|{\mathcal A}^*x + {\mathcal B}^*y^k - c+z^k/\sigma\|^2\big\},\\[3mm]
y^{k+1} = \argmin_{y\in\mathbb{Y}}\big\{{\mathcal L}_{\sigma}(x^{k+1},y;z^k)=:h_1(y)+h_2(y)+\frac{\sigma}{2}\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y - c+z^k/\sigma\|^2\big\}, \\[3mm]
z^{k+1} = z^k + \tau\sigma({\mathcal A}^*x^{k+1} + {\mathcal B}^*y^{k+1} - c),
\end{array}
\right.
\end{equation}
where $\tau\in(0, (1+\sqrt{5})/2)$ is a step-length. The convergence of the standard ADMM has long been established by Gabay \& Mercier\cite{GMC}, and Fortin \& Glowinski\cite{FGC}. And especially, Gabay \cite{ADMMDRS} showed that the standard ADMM with $\tau=1$ is exactly the Douglas-Rachford splitting method to the sum of two maximal monotone operators from the dual of (\ref{prob1}). And then, Eckstein \& Bertsekas \cite{GADMM} showed that itself is an instance of the proximal point algorithm \cite{ALMPPA} to a specially generated operator. For more work of ADMM, one refer to an important note \cite{noteADMM} and excellent survey \cite{GHC}.
To improve the performance of (\ref{cadmm}) in the case of $\tau = 1$, Eckstein \& Bertsekas \cite{GADMM} also proposed a generalized variant of ADMM. Subsequently, Chen \cite[Section 3.2]{CHEND} made an interesting observation and concluded that the generalized ADMM of Eckstein \& Bertsekas \cite{GADMM} is equivalent to the following iterative scheme from an initial point $\widetilde{\omega}^0:=(\widetilde{x}^0, \widetilde{y}^0, \widetilde{z}^0)\in \dom(f)\times \dom(h)\times\mathbb{Z}$,
\begin{equation}\label{gadmm}
\left\{
\begin{array}{l}
x^{k} = \argmin_{x\in\mathbb{X}}\Big\{{\mathcal L}_{\sigma}(x,\widetilde{y}^k;\widetilde{z}^k)=:f_1(x)+f_2(x)+\frac{\sigma}{2}\|{\mathcal A}^*x + {\mathcal B}^*\widetilde{y}^k - c+\widetilde{z}^k/\sigma\|^2\Big\},\\[3mm]
z^{k} = \widetilde{z}^k + \sigma({\mathcal A}^*x^{k} + {\mathcal B}^*\widetilde{y}^{k} - c),\\[3mm]
y^{k} = \argmin_{y\in\mathbb{Y}}\Big\{{\mathcal L}_{\sigma}(x^{k},y;z^k)=:h_1(y)+h_2(y)+\frac{\sigma}{2}\|{\mathcal A}^*x^k + {\mathcal B}^*y - c+z^k/\sigma\|^2\Big\}, \\[3mm]
\widetilde{\omega}^{k+1} = \widetilde{\omega}^k + \rho(\omega^k - \widetilde{\omega}^k),
\end{array}
\right.
\end{equation}
where $\omega^k = (x^k, y^k, z^k)$, $\widetilde{\omega}^k = (\widetilde{x}^k, \widetilde{y}^k, \widetilde{z}^k)$, and $\rho\in(0,2)$ is an uniform relaxation factor. Clearly, for $\rho=1$, the above generalized ADMM scheme is scheme (\ref{cadmm}) with $\tau=1$.
We see that the efficiency of the scheme (\ref{gadmm}) is mainly determined by the $x$- and $y$-subproblems. It was known that, if $f_1(x)$ and $h_1(y)$ are quadratic, and $f_2(x)$ and $g_2(y)$ are in the form of $f_2(x)=f_2(x_1)$ and $h_2(y)=h_2(y_1)$, the $x$- and $y$-subproblems can be solved efficiently if a couple of suitable proximal point terms are added, that is
\begin{align}
x^{k} = \argmin_{x\in\mathbb{X}}\Big\{f_1(x)+f_2(x_1)+\frac{\sigma}{2}\|{\mathcal A}^*x + {\mathcal B}^*\widetilde{y}^k - c+\widetilde{z}^k/\sigma\|^2+\frac12\|x-\widetilde{x}^k\|_{\mathcal S}^2\Big\},\label{xiaox}\\[2mm]
y^{k} = \argmin_{y\in\mathbb{Y}}\Big\{h_1(y)+h_2(y_1)+\frac{\sigma}{2}\|{\mathcal A}^*x^k + {\mathcal B}^*y - c+z^k/\sigma\|^2+\frac12\|y-\widetilde{y}^k\|_{\mathcal T}^2\Big\},\label{xiaoy}
\end{align}
where ${\mathcal S}$ and ${\mathcal T}$ are two self-adjoint and positive semidefinite linear operators. It was shown from Li et al. \cite{spADMM,bsGs} that, if ${\mathcal S}$ and ${\mathcal T}$ are chosen properly, both subproblems can be split into some smaller ones and then solved separately in a sGS manner. However, if $f_1(x)$ and $h_1(y)$ are non-quadratic, it seemly difficult to split into some small pieces so that the favorable structure 'nonsmooth+smooth' cannot be used any more.
We must emphasize that the approach of (\ref{xiaox}) and (\ref{xiaoy}) was firstly proposed by Xiao et al. \cite{spGADMM} which has been numerically demonstrated very efficient to solve doubly non-negative semidefinite programming problems with moderate accuracy.
In optimization literature, one popular way to approximate a continuously differentially convex function into a quadratic is the using of majorization \cite{COVA}, such as, Hong et al. \cite{HONG} and Cui et al. \cite{CUIJOTA}.
Particularly, Li et al. \cite{miPADMM} majorized the augmented Lagrangian function involved in (\ref{cadmm}) when $f_1(x)$ and $g_1(y)$ are non-quadratic to make the subproblems become a composite convex quadratic minimization.
Therefore, the resulting problems were solved efficiently with respect to each variable in a sGS order \cite{bsGs}. The attractive feather of using sGS is to make this iterative scheme fill into the framework of Fazel et al. \cite{FST}.
In a similar way, Qin et al. \cite{MGADMM} also used a majorization technique to the generalized ADMM of Eckstein \& Bertsekas \cite{GADMM} to make it more flexible and robust. Extensive numerical experiments demonstrated that the generalized ADMM with a suitable relaxation factor achieved better performance than the method of Li et al. \cite{miPADMM}.
Nevertheless, the performance of the semi-proximal generalized ADMM stated in (\ref{gadmm}), (\ref{xiaox}), and (\ref{xiaoy}) still has not been studied.
Hence, one natural question is how can we use the majorization technique so that the corresponding subproblems more amenable to efficient computations when $f_1(x)$ and $h_1(y)$ are non-quadratic.
The main purpose of this paper is to use a majorization technique to the augmented Lagrangian function (\ref{alf}) to make the resulting subproblems (\ref{xiaox}) and (\ref{xiaoy}) more easier solve, and hence, it will enhance the capabilities of the generalized ADMM of Xiao et al. \cite{spGADMM}. At the beginning, we must clarify that the reason why we focus on the generalized ADMM of Xiao et al. \cite{spGADMM} is due to the fact that this type of method is highly efficient than some state-of-the-art algorithms according to a series of numerical experiments.
Let $\mathbb{X}:=\mathbb{X}_1\times\ldots\times\mathbb{X}_s$ and $\mathbb{Y}:=\mathbb{Y}_1\times\ldots\times\mathbb{Y}_t$ with positive constants $s$ and $t$. At each iteration, we use majorized functions to replace $f_1(x)$ and $h_1(y)$ at the associated augmented Lagrangian function, and then both subproblems have ``nonsmooth+quadratic" structures. If $f_2(x):=f_2(x_1)$ and $h_2(y):=h_2(y_1)$ are simple functions, the $x$-subproblem ({\itshape resp.} $y$-subproblem) can be solved individually in the order of $x_s\rightarrow\ldots\rightarrow x_2\rightarrow x_1\rightarrow x_2\rightarrow\ldots\rightarrow x_s$ ({\itshape resp.} $y_t\rightarrow\ldots\rightarrow y_2\rightarrow y_1\rightarrow y_2\rightarrow\ldots\rightarrow y_t$) by making full use of the structures of $f_2(x_1)$ and $h_2(y_1)$.
Then, from the sGS decomposition theorem of Li et al. \cite{bsGs}, it is easy to show that this cycle is equivalent to adding proximal point terms with proper linear operators ${\mathcal S}$ and ${\mathcal T}$. We draw the difference in the iterative points between our method and the methods in \cite{MGADMM,miPADMM}, that is
$$
\begin{array}{rrcl}
[\text{Methods in \cite{miPADMM,MGADMM} }] &\ldots\rightarrow(x^k,y^k)&\longrightarrow&(x^{k+1},y^{k+1})\rightarrow\ldots\\[2mm]
[\text{Our method}] &\ldots\rightarrow(x^k,y^k)&\rightarrow(\widetilde{x}^{k+1},\widetilde{y}^{k+1})\rightarrow&(x^{k+1},y^{k+1})\rightarrow\ldots
\end{array}
$$
which shows that the new point $(x^{k+1},y^{k+1})$ is from the relaxation point $(\widetilde{x}^{k+1},\widetilde{y}^{k+1})$ but not the previous $(x^k,y^k)$ as in \cite{miPADMM,MGADMM}.
We must emphasize that the relaxation point $(\widetilde{x}^{k},\widetilde{y}^{k})$ will lead to more technical difficulties, so that the theoretical analysis can not be obtained by mimicking the aforementioned methods in \cite{miPADMM,MGADMM}.
Most of the remainder of this paper will focus on theoretical analysis, that is, we prove that the sequence $\{x^k,y^k\}$ generated by our method converges to a Karush-Kuhn-Tucker (KKT) point of problem (\ref{prob1}) under some technical conditions. Finally, we do numerical experiments on a class of simulated convex composite optimization problems. The numerical results illustrate that the proposed method performs better than the M-ADMM of Li et al. \cite{miPADMM} and the M-GADMM of Qin et al. \cite{MGADMM}.
The remaining parts of this paper are organized as follows. In section \ref{section3}, we propose the majorized and generalized ADMM to solve the composite convex problem (\ref{prob1}) in the case of $f_1(x)$ and $h_1(y)$ being non-quadratic and then it followed by some important properties. Then, we focus on the convergence analysis of the proposed algorithm in section \ref{section4}. In section \ref{section5}, we are devoted to implementation issue to show the potential numerical efficiency of our proposed algorithm. Finally, we conclude this paper with some remarks in section \ref{section6}.
\setcounter{equation}{0}
\section{A generalized ADMM with majorization}\label{section3}
At the beginning of this section, we give some preliminaries needed in the subsequent developments.
Let $f:\mathbb{X}\rightarrow(-\infty,+\infty]$ be a closed proper convex function. The effective domain of $f$, which is denoted by $\dom(f)$, is defined as $\dom(f):=\{x:f(x)<+\infty\}$. A vector $x^*$ is said to be a subgradient of $f$ at point $x$ if $f(z)\geq f(x)+\langle x^*,z-x\rangle$ for all $z\in\mathbb{X}$. The set of all subgradients of $f$ at $x$ is called the subdifferential of $f$ at $x$ and is denoted by $\partial f(x)$ or $\partial f$.
It is well-known from \cite{COVA}
that $\partial f$ is a maximal monotone operator. Because $f_1(x)$ and $h_1(y)$ are smooth convex functions with Lipschitz continuous gradients, we know that there exist self-adjoint and positive semidefinite linear operators such that $\widehat{\Sigma}_{f_1}\succeq \Sigma_{f_1}$ and $\widehat{\Sigma}_{h_1}\succeq \Sigma_{h_1}$, and for any $x, x' \in \mathbb{X}$ and any $y, y' \in \mathbb{Y}$, it holds that
\begin{equation}\label{f1cov}
\frac{1}{2}\|x-x'\|_{\Sigma_{f_1}}^2 \le f_1(x) - f_1(x') - \langle x-x',\nabla f_1(x') \rangle \le \displaystyle{\frac{1}{2}}\|x-x'\|_{\widehat{\Sigma}_{f_1}}^2,
\end{equation}
\begin{equation}\label{h1cov}
\frac{1}{2}\|y-y'\|_{\Sigma_{h_1}}^2 \le h_1(y) - h_1(y') - \langle y-y',\nabla h_1(y') \rangle \le \displaystyle{\frac{1}{2}}\|y-y'\|_{\widehat{\Sigma}_{h_1}}^2.
\end{equation}
Using the majorization technique, we construct the majorized functions for $f_1(x)$ and $h_1(y)$ as
\begin{equation}\label{hatf1}
\hat{f}_1(x,x') := f_1(x') + \langle x-x',\nabla f_1(x') \rangle + \displaystyle{\frac{1}{2}}\|x-x'\|_{\hat{\Sigma}_{f_1}}^2,
\end{equation}
\begin{equation}\label{hath1}
\hat{h}_1(y,y') := h_1(y') + \langle y-y',\nabla h_1(y') \rangle + \displaystyle{\frac{1}{2}}\|y-y'\|_{\hat{\Sigma}_{h_1}}^2.
\end{equation}
And then, using both majorized functions to replace the functions $f_1(x)$ and $h_1(y)$ in (\ref{alf}), we can get the majorized augmented Lagrangian function
\begin{equation}\label{majalf}
\hat{{\mathcal L}}_{\sigma}\big(x,y;(z,x',y')\big) := \hat{f}_1(x,x') + f_2(x) + \hat{h}_1(y,y') + h_2(y) + \langle z, {\mathcal A}^*x + {\mathcal B}^*y - c\rangle +\frac{\sigma}{2}\|{\mathcal A}^*x + {\mathcal B}^*y - c\|^2.
\end{equation}
Clearly, it is a composite convex quadratic function except for the terms $f_2(x)$ and $h_2(y)$.
For subsequent developments, we need the following constraint qualification.
\begin{assumption}\label{assum}
There exists $(x^0, y^0)\in ri(\dom(f_2) \times \dom (h_2)) \cap \Omega$, where $\Omega :=\{(x,y)\in \mathbb{X}\times\mathbb{Y}\mid {\mathcal A}^*x + {\mathcal B}^*y = c \}$.
\end{assumption}
Under Assumption (\ref{assum}), it is from \cite[Corollaries 28.2.2 and 28.3.1]{COVA}, we can get the optimality conditions of the problem (\ref{prob1}).
\begin{theorem}\label{optcond}
If the Assumption (\ref{assum}) hold, then $(\bar{x}, \bar{y})\in \mathbb{X}\times\mathbb{Y}$ is an optimal solution to problem (\ref{prob1}) if and only if there
exists a Lagrangian multiplier $\bar{z}\in\mathbb{Z}$ such that $(\bar{x},\bar{y},\bar{z})$ satisfies the following KKT system:
\begin{equation}\label{kktcond}
0\in \partial f_2(\bar{x}) + \nabla f_1(\bar{x}) + {\mathcal A}\bar{z},\quad 0\in \partial h_2(\bar{y}) + \nabla h_1(\bar{y}) + {\mathcal B}\bar{z},\quad {\mathcal A}^*\bar{x} + {\mathcal B}^*\bar{y} - c = 0,
\end{equation}
where $\bar z\in\mathbb{Z}$ is an optimal solution to the dual problem of (\ref{prob1}).
\end{theorem}
Because $f_2(x)$ and $h_2(y)$ are convex functions, the KKT system (\ref{kktcond}) is equivalent to finding a triple of points $(\bar{x},\bar{y},\bar{z})\in\mathbb{W} :=\mathbb{X}\times\mathbb{Y}\times\mathbb{Z}$ such that for any $(x, y)\in\mathbb{X}\times\mathbb{Y}$ the following inequality holds
\begin{equation}\label{covineq}
\big(f_2(x) + h_2(y)\big) - \big(f_2(\bar{x}) + h_2(\bar{y})\big) + \big\langle x-\bar{x},\nabla f_1(\bar{x})+A\bar{z}\big\rangle + \big\langle y-\bar{y}, \nabla h_1(\bar{y})+B\bar{z}\big\rangle \ge 0.
\end{equation}
The inequality will be used frequently in the theoretical analysis part.
In light of above preliminary results, we are ready to the construct our algorithm.
For convenience, we denote
$$
{\mathcal F}:=\widehat{\Sigma}_{f_1} + {\mathcal S} + \sigma {\mathcal A}\A^*
\quad \text{and} \quad
{\mathcal H}:= \widehat{\Sigma}_{h_1} + {\mathcal T} + \sigma {\mathcal B}\B^*.
$$
At the current iteration, if we use the majorized augmented Lagrangian function (\ref{majalf}) to replace its standard form (\ref{alf}), the $x$-subproblem in (\ref{xiaox}) will reduce to
\begin{align*}
x^{k} =& \argmin_{x}\Big\{\hat{{\mathcal L}}_{\sigma}(x,\widetilde{y}^k;(\widetilde{z}^k,\widetilde{x}^k,\widetilde{y}^k)) + \frac{1}{2}\|x - \widetilde{x}^k\|_{\mathcal S}^2\Big\}\\[3mm]
=&\argmin_{x}\Big\{f_2(x) + \frac{1}{2}\langle x, {\mathcal F} x\rangle + \Big\langle \nabla f_1(\widetilde{x}^k) + \sigma {\mathcal A}({\mathcal A}^*\widetilde{x}^{k} + {\mathcal B}^*\widetilde{y}^{k} - c + \sigma^{-1}\widetilde{z}^k) - {\mathcal F} \widetilde{x}^k, x\Big\rangle\Big\},
\end{align*}
and the $y$-subproblem in (\ref{xiaoy}) will reduce to
\begin{align*}
y^{k} =& \argmin_{y}\Big\{\hat{{\mathcal L}}_{\sigma}(x^k,y;(z^k,x^k,\widetilde{y}^k)) + \displaystyle{\frac{1}{2}}\|y - \widetilde{y}^k\|_{\mathcal T}^2\Big\}\\[3mm]
=&\argmin_{y}\Big\{h_2(y) + \displaystyle{\frac{1}{2}}\langle y, {\mathcal H} y\rangle + \Big\langle \nabla h_1(\widetilde{y}^k) + \sigma {\mathcal B}({\mathcal A}^*x^{k} + {\mathcal B}^*\widetilde{y}^{k} - c + \sigma^{-1}z^{k}) - {\mathcal H} \widetilde{y}^k, y\Big\rangle\Big\}.
\end{align*}
Clearly, both subproblems have the favorable structure of ``nonsmooth + quadratic" so that they can be solved efficiently if the self-adjoint operators ${\mathcal S}$ and ${\mathcal T}$ are chosen properly. Taking the $x$-subproblem as an example, in the case of $f_2(x)=f_2(x_1)$ being a simple function, we denote $Q_x:=\widehat{\Sigma}_{f_1} + \sigma {\mathcal A}\A^*$ and then decompose it into $Q_x=U_x+\Sigma_x+U^\top_x$, where $U_x$ is a strictly upper triangular submatrix and $\Sigma_x$ is the diagonal of $Q_x$. Let ${\mathcal S}:=U_x\Sigma_x^{-1}U_x^\top$, then the $x$-subproblem can be computed in a sGS order $x_s\rightarrow\ldots\rightarrow x_2\rightarrow x_1\rightarrow x_2\rightarrow\ldots\rightarrow x_s$, which indicates that the $x$-subproblem is split into a series of smaller problems with each $x_i$ and solved separately.
For more theoretical details on this iterative scheme, one may refer to the excellent papers of Li et al. \cite{spGADMM,miPADMM}.
In light of the above analyses, we list the generalized ADMM with majorization (abbr. G-ADMM-M) for solving problem (\ref{prob1}) as follows.
\begin{framed}
\noindent
{\bf Algorithm: G-ADMM-M (Generalized ADMM with Majorization).}
\vskip 1.0mm \hrule \vskip 1mm
\noindent
\begin{itemize}
\item[Step 0.] Choose $\sigma> 0$ and $\rho\in (0,2)$. Choose self-adjoint positive semidefinite linear operators ${\mathcal S}$ and ${\mathcal T}$ in $\mathbb{X}$ and $\mathbb{Y}$ such that ${\mathcal F}\succ 0$ and ${\mathcal H}\succ 0$.
Input an initial point $\widetilde{\omega}^0 := (\widetilde{x}^0,\widetilde{y}^0,\widetilde{z}^0)\in \dom(f_2)\times \dom( h_2)\times \mathbb{Z}$. Let $k:=0$.
\item[Step 1.] Compute
\begin{equation}\label{mipgadmm}
\left\{
\begin{array}{l}
x^{k} = \argmin_{x}\Big\{f_2(x) + \displaystyle{\frac{1}{2}}\langle x, {\mathcal F} x\rangle + \langle \nabla f_1(\widetilde{x}^k) + \sigma {\mathcal A}({\mathcal A}^*\widetilde{x}^{k} + {\mathcal B}^*\widetilde{y}^{k} - c + \sigma^{-1}\widetilde{z}^k) - {\mathcal F} \widetilde{x}^k, x\rangle\Big\},\\[3mm]
z^{k} = \widetilde{z}^k + \sigma({\mathcal A}^*x^{k} + {\mathcal B}^*\widetilde{y}^{k} - c),\\[3mm]
y^{k} = \argmin_{y}\Big\{h_2(y) + \displaystyle{\frac{1}{2}}\langle y, {\mathcal H} y\rangle + \langle \nabla h_1(\widetilde{y}^k) + \sigma {\mathcal B}({\mathcal A}^*x^{k} + {\mathcal B}^*\widetilde{y}^{k} - c + \sigma^{-1}z^{k}) - {\mathcal H} \widetilde{y}^k, y\rangle\Big\}.
\end{array}
\right.
\end{equation}
\item[Step 2.] Terminate if a termination criterion is satisfied. Output $(x^k,y^k,z^k)$. Otherwise, compute
\begin{equation}\label{omegak}
\widetilde{\omega}^{k+1} = \widetilde{\omega}^k + \rho(\omega^k - \widetilde{\omega}^k).
\end{equation}
Let $k:=k+1$ and go to Step 1.
\end{itemize}
\vspace{-.1cm}
\end{framed}
For the convergence analysis of the G-ADMM-M, we present a couple of useful lemmas. The results are well known in the field of numerical algebra, hence, we omit their proof here.
\begin{lemma}\label{salop1}
For any vectors $u$, $v$ in the same Euclidean vector space $\Im$, and any self-adjoint positive semidefinite linear operator ${\mathcal G}: \Im\rightarrow\Im$, it holds that
\begin{equation}\label{noneq}
\|u\|_{\mathcal G}^2 + \|v\|_{\mathcal G}^2 \ge \displaystyle{\frac{1}{2}}\|u - v\|_{\mathcal G}^2,
\end{equation}
and
\begin{equation}\label{ident1}
2\langle u, {\mathcal G} v\rangle = \|u\|_{\mathcal G}^2 + \|v\|_{\mathcal G}^2 - \|u - v\|_{\mathcal G}^2 = \|u + v\|_{\mathcal G}^2 - \|u\|_{\mathcal G}^2 - \|v\|_{\mathcal G}^2.
\end{equation}
\end{lemma}
\begin{lemma}\label{salop2}
For any vectors $u_1$, $u_2$, $v_1$, and $v_2$ in the same Euclidean vector space $\Im$, and any self-adjoint positive semidefinite linear operator ${\mathcal G}: \Im\rightarrow\Im$, we have the identity
\begin{equation}\label{ident2}
2\langle u_1 - u_2, {\mathcal G} (v_1 - v_2)\rangle = \|u_1 - v_2\|_{\mathcal G}^2 + \|u_2 - v_1\|_{\mathcal G}^2 - \|u_1 - v_1\|_{\mathcal G}^2 - \|u_2 - v_2\|_{\mathcal G}^2.
\end{equation}
\end{lemma}
Denote $\mathbb{W}^*$ be the set of solutions satisfy (\ref{kktcond}), which is nonempty under the Assumption \ref{assum}, i.e., the solution set of problem (\ref{prob1}) is nonempty. For $(\bar{x},\bar{y},\bar{z})\in\mathbb{W}^*$ and any $(x,y,z)\in\mathbb{W}$, we denote $x_e := x - \bar{x}$, $y_e := y - \bar{y}$, and $z_e := z -\bar{z}$ for convenience. Using these notations, we have the following three properties which will be used in our desiring convergence theorem analysis.
\begin{lemma}\label{lem3}
Suppose that Assumption \ref{assum} holds. Let $\{(x^k, y^k, z^k)\}$ be generated by Algorithm G-ADMM-M, and $(\bar{x},\bar{y},\bar{z})\in\mathbb{W}^*$. Then for any $\rho\in(0,2)$, $\sigma>0$ and $k\ge 0$, we have
\begin{equation}\label{ident3}
\begin{array}{l}
\Big\langle {\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k, z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\Big\rangle \\[3mm]
=\displaystyle{\frac{1}{2\sigma\rho}}\Big(\|z_e^{k+1} + \sigma(\rho - 1)A^*x_e^{k+1}\|^2 - \|z_e^{k} + \sigma(\rho - 1){\mathcal A}^*x_e^{k}\|^2\Big) + \displaystyle{\frac{\sigma\rho}{2}}\|{\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k\|^2.
\end{array}
\end{equation}
\end{lemma}
\begin{proof}
From the iterative scheme(\ref{mipgadmm} and \ref{omegak}) in Algorithm G-ADMM-M, we have
\begin{equation}\label{lagmtran}
\begin{aligned}
z^{k+1} &= \widetilde{z}^{k+1} + \sigma({\mathcal A}^*x^{k+1} + {\mathcal B}^*\widetilde{y}^{k+1} - c)\\[3mm]
&= z^k - (1-\rho)(z^k - \widetilde{z}^k) + \sigma\rho({\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c) + \sigma(1-\rho)({\mathcal A}^*x^{k+1} + {\mathcal B}^*\widetilde{y}^k - c)\\[3mm]
&= z^k + \sigma\rho({\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c) + \sigma(\rho-1)({\mathcal A}^*x^{k} - {\mathcal A}^*x^{k+1}),
\end{aligned}
\end{equation}
which indicates
$$
\big[z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\big] - \big[z_e^{k} + \sigma(\rho - 1){\mathcal A}^*x_e^{k}\big] = \sigma\rho({\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k).
$$
According to Lemma \ref{salop1}, we get
\begin{equation}\label{ident4}
\begin{array}{l}
2\sigma\rho\langle {\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k, z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\rangle \\[3mm]
=\sigma^2\rho^2\|{\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k\|^2 + \|z_e^{k+1} + \sigma(\rho - 1)A^*x_e^{k+1}\|^2 - \|z_e^{k} + \sigma(\rho - 1){\mathcal A}^*x_e^{k}\|^2.
\end{array}
\end{equation}
Combining with (\ref{ident4}), it yields the desired result (\ref{ident3}).
\end{proof}
\begin{lemma}\label{lem4}
Suppose that Assumption \ref{assum} holds. Let $\{(x^k, y^k, z^k)\}$ be generated by Algorithm G-ADMM-M, and $(\bar{x},\bar{y},\bar{z})\in\mathbb{W}^*$. Then for any $\rho\in(0,2)$, $\sigma>0$ and $k\ge 0$, we have
\begin{equation}\label{ident5}
\begin{array}{l}
\Big\langle {\mathcal B}^*y_e^k, z^{k} + \sigma({\mathcal A}^*x_e^k + {\mathcal B}^*y_e^k) - z^{k+1} - \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\Big\rangle \\[3mm]
=\displaystyle{\frac{\sigma(2-\rho)}{2}}\Big(\|{\mathcal A}^*x_e^{k} + {\mathcal B}^*y_e^k\|^2 - \|{\mathcal A}^*x_e^{k}\|^2\Big) - \displaystyle{\frac{\sigma\rho}{2}}\Big(\|{\mathcal A}^*x_e^{k+1} + B^*y_e^k\|^2 - \|{\mathcal A}^*x_e^{k+1}\|^2\Big).
\end{array}
\end{equation}
\end{lemma}
\begin{proof}
From (\ref{lagmtran}), we have
\begin{equation}\nonumber
\begin{array}{ll}
&z^{k} + \sigma({\mathcal A}^*x_e^k + {\mathcal B}^*y_e^k) - z^{k+1} - \sigma(\rho - 1){\mathcal A}^*x_e^{k+1} \\[3mm]
=&-\sigma\rho({\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k) + \sigma({\mathcal A}^*x_e^{k} + {\mathcal B}^*y_e^k) - \sigma(\rho-1){\mathcal A}^*x_e^{k}.
\end{array}
\end{equation}
Then we have
\begin{equation}\nonumber
\begin{array}{rl}
&\langle {\mathcal B}^*y_e^k, z^{k} + \sigma({\mathcal A}^*x_e^k + {\mathcal B}^*y_e^k) - z^{k+1} - \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\rangle \\[3mm]
=&-\sigma\rho\langle {\mathcal B}^*y_e^k, {\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k\rangle + \sigma\langle {\mathcal B}^*y_e^k, {\mathcal A}^*x_e^{k} + {\mathcal B}^*y_e^k\rangle - \sigma(\rho-1)\langle {\mathcal B}^*y_e^k, {\mathcal A}^*x_e^{k}\rangle.
\end{array}
\end{equation}
According to the equality (\ref{ident1}) in Lemma \ref{salop1}, we have
\begin{equation}\nonumber
\begin{array}{rl}
&-\sigma\rho\langle {\mathcal B}^* y_e^k, {\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k\rangle + \sigma\langle {\mathcal B}^* y_e^k, {\mathcal A}^*x_e^{k} + {\mathcal B}^*y_e^k\rangle - \sigma(\rho-1)\langle {\mathcal B} y_e^k, {\mathcal A}^*x_e^{k}\rangle\\[3mm]
= &- \frac{\sigma\rho}{2}\Big(\|{\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k\|^2 + \|B^*y_e^k\|^2 - \|{\mathcal A}^*x_e^{k+1}\|^2\Big) + \displaystyle{\frac{\sigma}{2}}\Big(\|{\mathcal A}^*x_e^{k} + {\mathcal B}^*y_e^k\|^2 + \|{\mathcal B}^*y_e^k\|^2 - \|{\mathcal A}^*x_e^{k}\|^2\Big) \\[3mm]
& - \frac{\sigma(\rho- 1)}{2}\Big(\|{\mathcal A}^*x_e^{k} + {\mathcal B}^*y_e^k\|^2 - \|{\mathcal B}^*y_e^k\|^2 - \|{\mathcal A}^*x_e^{k}\|^2\Big)\\[3mm]
= &\frac{\sigma(2-\rho)}{2}\Big(\|A^*x_e^{k} + B^*y_e^k\|^2 - \|A^*x_e^{k}\|^2\Big) - \displaystyle{\frac{\sigma\rho}{2}}\Big(\|A^*x_e^{k+1} + B^*y_e^k\|^2 - \|A^*x_e^{k+1}\|^2\Big).
\end{array}
\end{equation}
The proof is complete.
\end{proof}
For notational convenience, we define
\begin{equation}\label{psik}
\Psi_k(\bar{x},\bar{y},\bar{z}) := \frac{1}{\sigma\rho}\|z_e^k + \sigma(\rho-1)A^*x_e^{k}\|^2 + \sigma(2-\rho)\|A^*x_e^{k}\|^2 + \frac{1}{\rho}\|\widetilde{x}_e^{k+1}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2 + \frac{1}{\rho}\|\widetilde{y}_e^k\|_{\widehat{\Sigma}_{h_1}+{\mathcal T}}^2,
\end{equation}
and
\begin{equation}\label{deltak}
\begin{array}{l}
\delta_k := \|\widetilde{x}^{k+1} - x^{k+1}\|_{\frac{1}{2}\Sigma_{f_1}}^2 + \|\widetilde{y}^{k} - y^{k}\|_{\frac{1}{2}\Sigma_{h_1}+(2-\rho)(\widehat{\Sigma}_{h_1}+{\mathcal T})}^2 + (1-\lambda)(2-\rho)\|\widetilde{x}^{k} - x^{k}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2\\[3mm]
\qquad + \sigma(2\lambda-1)(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^{k} - c\|^2 + \frac{\sigma(1-\lambda)(2-\rho)}{2}\|{\mathcal B}^*(y^k-y^{k-1})\|^2\\[3mm]
\qquad + \frac{\lambda(2-\rho)^2}{\rho}\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}+\sigma {\mathcal A}\A^*}^2,
\end{array}
\end{equation}
where $\lambda>0$ will defined later.
Furthermore, we also define
\begin{equation}\label{thetak}
\theta_k := \|\widetilde{x}^{k+1} - x^{k+1}\|_{\frac{1}{2}\Sigma_{f_1}-\hat{\Sigma}_{f_1}+(2-\rho)(\widehat{\Sigma}_{f_1}+{\mathcal S})}^2 + \|\widetilde{y}^{k} - y^{k}\|_{\frac{1}{2}\Sigma_{h_1}-\hat{\Sigma}_{h_1}+(2-\rho)(\widehat{\Sigma}_{h_1}+{\mathcal T})}^2,
\end{equation}
\begin{equation}\label{etak}
\eta_k := \Big\langle {\mathcal A}^*x_e^{k+1}, z_e^{k+1} \Big\rangle + \Big\langle {\mathcal B}^*y_e^k, z_e^k + \sigma({\mathcal A}^*x^k + {\mathcal B}^*y^k - c)\Big\rangle,
\end{equation}
and
\begin{equation}\label{xik}
\xi_k := \|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2 + \|\widetilde{y}^{k} - y^{k}\|_{\widehat{\Sigma}_{h_1}}^2.
\end{equation}
For the global convergence for the Algorithm G-ADMM-M, the following inequality is essential.
\begin{lemma}\label{lem5}
Suppose that Assumption \ref{assum} holds. Let $\{(x^k, y^k, z^k)\}$ be generated by Algorithm G-ADMM-M, and $(\bar{x},\bar{y},\bar{z})\in\mathbb{W}^*$. Let $\Psi_k$ and $\theta_k$ be defined as in (\ref{psik}) and (\ref{thetak}). Then for any $\rho\in(0,2)$, $\sigma>0$ and $k\ge 0$, we have
\begin{equation}\label{pivineq}
\Psi_k(\bar{x},\bar{y},\bar{z}) - \Psi_{k+1}(\bar{x},\bar{y},\bar{z}) \ge \theta_k + \sigma(2-\rho)\|{\mathcal A}^*x^k + {\mathcal B}^*y^k - c\|^2.
\end{equation}
\end{lemma}
\begin{proof}
From the left inequality of (\ref{f1cov}), let $x':=\widetilde{x}^{k+1}$, we have
\begin{equation}\label{ineqcovl}
f_1(x) \ge f_1(\widetilde{x}^{k+1}) + \langle x-\widetilde{x}^{k+1},\nabla f_1(\widetilde{x}^{k+1}) \rangle + \displaystyle{\frac{1}{2}}\|x-\widetilde{x}^{k+1}\|_{\Sigma_{f_1}}^2,
\end{equation}
and then from the right inequality of (\ref{f1cov}), let $x:=x^{k+1}$, $x':=\widetilde{x}^{k+1}$, we have
\begin{equation}\label{ineqcovr}
f_1(x^{k+1}) \le f_1(\widetilde{x}^{k+1}) + \langle x^{k+1}-\widetilde{x}^{k+1},\nabla f_1(\widetilde{x}^{k+1}) \rangle + \displaystyle{\frac{1}{2}}\|x^{k+1}-\widetilde{x}^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2.
\end{equation}
Subtracting both sides of (\ref{ineqcovl}) and (\ref{ineqcovr}), we obtain
\begin{equation}\label{ineq1}
f_1(x) - f_1(x^{k+1}) - \langle x-x^{k+1},\nabla f_1(\widetilde{x}^{k+1}) \rangle \ge \displaystyle{\frac{1}{2}}\|x - \widetilde{x}^{k+1}\|_{\Sigma_{f_1}}^2 - \displaystyle{\frac{1}{2}}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2.
\end{equation}
Using the necessary optimality conditions of $x^{k+1}$ in (\ref{mipgadmm}), we have for any $x\in\mathbb{X}$ and $\xi\in\partial f_2(x^{k+1})$ that
\begin{equation}\label{varineq}
\Big\langle x - x^{k+1}, \xi + {\mathcal F} x^{k+1} + \nabla f_1(\widetilde{x}^{k+1}) + \sigma {\mathcal A}({\mathcal A}^*\widetilde{x}^{k+1} + {\mathcal B}^*\widetilde{y}^{k+1} - c + \sigma^{-1}\widetilde{z}^{k+1}) - {\mathcal F}\widetilde{x}^{k+1} \Big\rangle \ge 0.
\end{equation}
From the convexity of $f_2(x)$, we get
\begin{equation}\label{convf2}
\langle x - x^{k+1}, \xi\rangle \le f_2(x) - f_2(x^{k+1}).
\end{equation}
Then, substituting (\ref{convf2}) into (\ref{varineq}) and using the relaxation step (\ref{omegak}), it yields
\begin{align}\label{ineq2}
&f_2(x) - f_2(x^{k+1}) + \Big\langle x - x^{k+1}, \nabla f_1(\widetilde{x}^{k+1}) + \sigma {\mathcal A}({\mathcal A}^*\widetilde{x}^{k+1} + {\mathcal B}^*\widetilde{y}^{k+1} - c + \sigma^{-1}\widetilde{z}^{k+1}) + {\mathcal F}(x^{k+1} - \widetilde{x}^{k+1})\Big\rangle\nonumber\\[3mm]
=& f_2(x) - f_2(x^{k+1}) + \Big\langle x - x^{k+1}, \nabla f_1(\widetilde{x}^{k+1}) + {\mathcal A} z^{k+1} - (\widehat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - x^{k+1})\Big\rangle \ge 0.
\end{align}
Combining (\ref{ineq1}) and (\ref{ineq2}), we obtain
\begin{equation}\label{ineq3}
\begin{array}{l}
f_1(x) - f_1(x^{k+1}) + f_2(x) - f_2(x^{k+1}) + \langle x - x^{k+1}, {\mathcal A} z^{k+1} - (\widehat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - x^{k+1})\rangle\\[3mm]
\ge \frac{1}{2}\|x - \widetilde{x}^{k+1}\|_{\Sigma_{f_1}}^2 - \frac{1}{2}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2.
\end{array}
\end{equation}
Similarly, for any $y\in\mathbb{Y}$, we have
\begin{equation}\label{ineq4}
\begin{array}{l}
h_1(y) - h_1(y^{k}) + h_2(y) - h_2(y^{k}) + \Big\langle y - y^{k}, \sigma {\mathcal B}({\mathcal A}^*x^{k} + {\mathcal B}^*y^{k} - c + \sigma^{-1}z^{k}) - (\hat{\Sigma}_{h_1}+{\mathcal T})(\widetilde{y}^{k} - y^{k})\Big\rangle\\[3mm]
\ge \frac{1}{2}\|y - \widetilde{y}^{k}\|_{\Sigma_{h_1}}^2 - \frac{1}{2}\|\widetilde{y}^{k} - y^{k}\|_{\widehat{\Sigma}_{h_1}}^2.
\end{array}
\end{equation}
Let $x=\bar{x}$ and $y=\bar{y}$, and combine (\ref{ineq3}) with (\ref{ineq4}), we obtain
\begin{equation}\label{ineq5}
\begin{array}{l}
f_1(\bar{x}) - f_1(x^{k+1}) + h_1(\bar{y}) - h_1(y^{k}) + f_2(\bar{x}) - f_2(x^{k+1}) + h_2(\bar{y}) - h_2(y^{k})\\[3mm]
- \Big\langle x_e^{k+1}, {\mathcal A} z^{k+1} - (\hat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - x^{k+1})\Big\rangle - \Big\langle y_e^{k}, \sigma {\mathcal B}({\mathcal A}^*x^{k} + {\mathcal B}^*y^{k} - c + \sigma^{-1}z^{k}) - (\widehat{\Sigma}_{h_1}+{\mathcal T})(\widetilde{y}^{k} - y^{k})\Big\rangle\\[3mm]
\ge \displaystyle{\frac{1}{2}}\|\widetilde{x}_e^{k+1}\|_{\Sigma_{f_1}}^2 - \displaystyle{\frac{1}{2}}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2 + \displaystyle{\frac{1}{2}}\|\widetilde{y}_e^{k}\|_{\Sigma_{h_1}}^2 - \displaystyle{\frac{1}{2}}\|\widetilde{y}^{k} - y^{k}\|_{\widehat{\Sigma}_{h_1}}^2.
\end{array}
\end{equation}
From (\ref{f1cov}) and (\ref{h1cov}), we can easily get
\begin{equation}\label{ineq6}
f_1(x^{k+1}) - f_1(\bar{x}) \ge \langle x_e^{k+1},\nabla f_1(\bar{x}) \rangle + \displaystyle{\frac{1}{2}}\|x_e^{k+1}\|_{\Sigma_{f_1}}^2,
\end{equation}
and
\begin{equation}\label{ineq7}
h_1(y^{k}) - h_1(\bar{y}) \ge \langle y_e^{k},\nabla h_1(\bar{y}) \rangle + \displaystyle{\frac{1}{2}}\|y_e^{k}\|_{\Sigma_{h_1}}^2.
\end{equation}
According to (\ref{ineq5}), (\ref{ineq6}), (\ref{ineq7}) and the inequality (\ref{noneq}), we obtain
\begin{align}\label{ineq8}
&f_2(\bar{x}) - f_2(x^{k+1}) + h_2(\bar{y}) - h_2(y^{k}) - \langle x_e^{k+1}, \nabla f_1(\bar{x}) + A\bar{z}\rangle - \langle y_e^{k}, \nabla h_1(\bar{y}) + {\mathcal B}\bar{z}\rangle - \langle {\mathcal A}^*x_e^{k+1}, z_e^{k+1} \rangle\nonumber\\[3mm]
&- \Big\langle {\mathcal B}^*y_e^k, z_e^k + \sigma({\mathcal A}^*x^k + {\mathcal B}^*y^k - c)\Big\rangle + \Big\langle x_e^{k+1}, (\hat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - x^{k+1})\Big\rangle + \Big\langle y_e^{k}, (\hat{\Sigma}_{h_1}+{\mathcal T})(\widetilde{y}^{k} - y^{k})\Big\rangle\nonumber\\[3mm]
\ge& \frac{1}{2}\|\widetilde{x}_e^{k+1}\|_{\Sigma_{f_1}}^2 - \frac{1}{2}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\hat{\Sigma}_{f_1}}^2 + \displaystyle{\frac{1}{2}}\|\widetilde{y}_e^{k}\|_{\Sigma_{h_1}}^2 - \frac{1}{2}\|\widetilde{y}^{k} - y^{k}\|_{\hat{\Sigma}_{h_1}}^2 + \frac{1}{2}\|x_e^{k+1}\|_{\Sigma_{f_1}}^2 + \displaystyle{\frac{1}{2}}\|y_e^{k}\|_{\Sigma_{h_1}}^2\nonumber\\[3mm]
\ge& \displaystyle{\frac{1}{4}}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\Sigma_{f_1}}^2 + \displaystyle{\frac{1}{4}}\|\widetilde{y}^{k} - y^{k}\|_{\Sigma_{h_1}}^2 - \displaystyle{\frac{1}{2}}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2 - \displaystyle{\frac{1}{2}}\|\widetilde{y}^{k} - y^{k}\|_{\widehat{\Sigma}_{h_1}}^2.
\end{align}
Replace $x$ with $x^{k+1}$, and $y$ with $y^k$ in the inequality (\ref{covineq}), we get
$$
f_2(\bar{x}) - f_2(x^{k+1}) + h_2(\bar{y}) - h_2(y^{k}) - \langle x_e^{k+1}, \nabla f_1(\bar{x}) + A\bar{z}\rangle - \langle y_e^{k}, \nabla h_1(\bar{y}) + B\bar{z}\rangle \le 0.
$$
Substituting this inequality and (\ref{etak}) into (\ref{ineq8}), it yields
\begin{align}\label{ineq9}
&-\eta^k + \Big\langle x_e^{k+1}, (\hat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - x^{k+1})\Big\rangle + \Big\langle y_e^{k}, (\hat{\Sigma}_{h_1}+{\mathcal T})(\widetilde{y}^{k} - y^{k})\Big\rangle\nonumber\\[3mm]
\ge& \frac{1}{4}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\Sigma_{f_1}}^2 + \frac{1}{4}\|\widetilde{y}^{k} - y^{k}\|_{\Sigma_{h_1}}^2 - \frac{1}{2}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2 - \frac{1}{2}\|\widetilde{y}^{k} - y^{k}\|_{\widehat{\Sigma}_{h_1}}^2.
\end{align}
Using Lemmas \ref{lem3} and \ref{lem4}, we have
\begin{equation}\label{etakconv}
\begin{array}{l}
\eta^k = \Big\langle {\mathcal A}^*x_e^{k+1}, z_e^{k+1} \Big\rangle + \Big\langle {\mathcal B}^*y_e^k, z_e^k + \sigma({\mathcal A}^*x^k + {\mathcal B}^*y^k - c)\Big\rangle\\[3mm]
\quad = \Big\langle {\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k, z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\Big\rangle - \Big\langle {\mathcal B}^*y_e^k, z_e^{k+1} + \sigma(\rho - 1)A^*x_e^{k+1}\Big\rangle\\[3mm]
\qquad - \Big\langle {\mathcal A}^*x_e^{k+1}, \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\Big\rangle + \Big\langle {\mathcal B}^*y_e^k, z_e^k + \sigma({\mathcal A}^*x^k + {\mathcal B}^*y^k - c)\Big\rangle\\[3mm]
\quad = \Big\langle {\mathcal A}^*x_e^{k+1} + {\mathcal B}^*y_e^k, z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\Big\rangle + \Big\langle {\mathcal B}^*y_e^k, z_e^k + \sigma({\mathcal A}^*x^k + {\mathcal B}^*y^k - c) - z_e^{k+1}\\[3mm]
\qquad - \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\Big\rangle - \sigma(\rho - 1)\|{\mathcal A}^*x_e^{k+1}\|^2\\[3mm]
\quad = \displaystyle{\frac{1}{2\sigma\rho}}\Big(\|z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\|^2 - \|z_e^{k} + \sigma(\rho - 1){\mathcal A}^*x_e^{k}\|^2\Big)\\[3mm] \qquad+ \displaystyle{\frac{\sigma(2-\rho)}{2}}\Big(\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^k - c\|^2
+ \|{\mathcal A}^*x_e^{k+1}\|^2 - \|{\mathcal A}^*x_e^{k}\|^2\Big).\\[3mm]
\end{array}
\end{equation}
From (\ref{omegak}) and (\ref{ident2}), we have
\begin{equation}\label{xekconv}
\begin{array}{l}
\Big\langle x_e^{k+1}, (\widehat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - x^{k+1})\Big\rangle\\[3mm]
= \displaystyle{\frac{1}{\rho}}\Big\langle x^{k+1} - \bar{x}, (\widehat{\Sigma}_{f_1}+{\mathcal S})(\widetilde{x}^{k+1} - \widetilde{x}^{k+2})\Big\rangle\\[3mm]
= \displaystyle{\frac{1}{2\rho}}\Big(\|\widetilde{x}_e^{k+1}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2 - \|\widetilde{x}_e^{k+2}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2 - \rho(2-\rho)\|x^{k+1} - \widetilde{x}^{k+1}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2\Big).
\end{array}
\end{equation}
In a similar way, we get
\begin{equation}\label{yekconv}
\begin{array}{l}
\Big\langle y_e^{k}, (\hat{\Sigma}_{h_1}+{\mathcal T})(\widetilde{y}^{k} - y^{k})\Big\rangle\\[3mm]
= \displaystyle{\frac{1}{2\rho}}\Big(\|\widetilde{y}_e^{k}\|_{\widehat{\Sigma}_{h_1}+{\mathcal T}}^2 - \|\widetilde{y}_e^{k+1}\|_{\widehat{\Sigma}_{h_1}+{\mathcal T}}^2 - \rho(2-\rho)\|y^{k} - \widetilde{y}^{k}\|_{\widehat{\Sigma}_{h_1}+{\mathcal T}}^2\Big).
\end{array}
\end{equation}
Combining (\ref{ineq9}), (\ref{etakconv}), (\ref{xekconv}) with (\ref{yekconv}), we obtain
\begin{align}\label{ineqconv}
&\frac{1}{2\sigma\rho}\Big(\|z_e^{k} + \sigma(\rho - 1){\mathcal A}^*x_e^{k}\|^2 - \|z_e^{k+1} + \sigma(\rho - 1){\mathcal A}^*x_e^{k+1}\|^2\Big) - \frac{\sigma(2-\rho)}{2}\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^k - c\|^2\nonumber\\[3mm]
&+ \frac{\sigma(2-\rho)}{2}\Big(\|A^*x_e^{k}\|^2 - \|A^*x_e^{k+1}\|^2\Big) - \frac{2-\rho}{2}\|x^{k+1} - \widetilde{x}^{k+1}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2 \nonumber\\[3mm]
&+ \frac{1}{2\rho}\Big(\|\widetilde{x}_e^{k+1}\|_{\hat{\Sigma}_{f_1}+{\mathcal S}}^2- \|\widetilde{x}_e^{k+2}\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2\Big) - \frac{2-\rho}{2}\|y^{k} - \widetilde{y}^{k}\|_{\widehat{\Sigma}_{h_1}+{\mathcal T}}^2 + \frac{1}{2\rho}\Big(\|\widetilde{y}_e^{k}\|_{\hat{\Sigma}_{h_1}+{\mathcal T}}^2 - \|\widetilde{y}_e^{k+1}\|_{\widehat{\Sigma}_{h_1}+{\mathcal T}}^2\Big)\nonumber\\[3mm]
\ge &
\frac{1}{4}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\Sigma_{f_1}}^2 + \frac{1}{4}\|\widetilde{y}^{k} - y^{k}\|_{\Sigma_{h_1}}^2 - \frac{1}{2}\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1}}^2 - \frac{1}{2}\|\widetilde{y}^{k} - y^{k}\|_{\widehat{\Sigma}_{h_1}}^2.
\end{align}
Substitutiing (\ref{psik}) and (\ref{thetak}) into (\ref{ineqconv}) and noting the definition of $\Psi_k$ in (\ref{psik}), we can get the conclusion (\ref{pivineq}) immediately.
\end{proof}
\newpage
\setcounter{equation}{0}
\section{Convergence analysis}\label{section4}
Based on the lemmas proved in the previous section, we can establish the convergence of G-ADMM-M for solving problem (\ref{prob1}). For this purpose, we firstly give some useful lemmas.
\begin{lemma}\label{lem6}
Suppose that the Assumption (\ref{assum}) hold. Let the sequence $\{(x^k,y^k,z^k)\}$ be generated by Algorithm G-ADMM-M and $(\bar{x},\bar{y},\bar{z})\in \mathbb{W}^*$. Then, for any $\sigma > 0$, $\rho\in(0,2)$, and $k > 1$, we have
\begin{align}\label{conneq}
\sigma(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^k - c\|^2 \ge& \sigma(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2 + \frac{(2-\rho)^2}{\rho}\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}+\sigma AA^*}\nonumber\\[3mm]
&- (2-\rho)\|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2.
\end{align}
\end{lemma}
\begin{proof}
Note that
\begin{align*}
&\sigma(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^k - c\|^2\\[3mm]
= &\sigma(2-\rho)\|A^*x^{k+1} + {\mathcal B}^*y^k - c + A^*(x^k - x^{k+1})\|^2\\[3mm]
= &\sigma(2-\rho)\|A^*x^{k+1} + {\mathcal B}^*y^k - c\|^2 + \sigma(2-\rho)\|A^*(x^k - x^{k+1})\|^2\\[3mm]
&+ 2\sigma(2-\rho)\Big\langle {\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c, {\mathcal A}^*(x^k - x^{k+1})\Big\rangle.
\end{align*}
From (\ref{lagmtran}), we have
\begin{align*}
&2\sigma(2-\rho)\Big\langle {\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c, A^*(x^k - x^{k+1})\Big\rangle\\[3mm]
=& \frac{2(2-\rho)}{\rho}\Big\langle z_e^{k+1} - z_e^k + \sigma(\rho-1)({\mathcal A}^*x_e^{k+1} - {\mathcal A}^*x_e^k), {\mathcal A}^*(x^k - x^{k+1})\Big\rangle\\[3mm]
= &\frac{2(2-\rho)}{\rho}\Big\langle z_e^{k+1} - z_e^k, {\mathcal A}^*(x^k - x^{k+1})\Big\rangle - \frac{2\sigma(\rho-1)(2-\rho)}{\rho}\|{\mathcal A}^*(x^k - x^{k+1})\|^2.
\end{align*}
Then we have
\begin{align}\label{eqcon}
&\sigma(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^k - c\|^2\nonumber\\[3mm]
= &\sigma(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2 +\frac{(2-\rho)^2}{\rho}\|x^{k+1} - x^{k}\|_{\sigma {\mathcal A}\A^*}^2
+ \frac{2(2-\rho)}{\rho}\Big\langle z_e^{k+1} - z_e^k, {\mathcal A}^*(x^k - x^{k+1})\Big\rangle.
\end{align}
According to the first-order optimality condition of the x-subproblem in (\ref{mipgadmm}), we know that
\begin{equation}\label{foocx}
\begin{array}{l}
-\nabla f_1(\widetilde{x}^k) - (\widehat{\Sigma}_{f_1} + {\mathcal S})(x^k - \widetilde{x}^k) - {\mathcal A} z^k \in \partial f_2(x^k),\\[3mm]
-\nabla f_1(\widetilde{x}^{k+1}) - (\widehat{\Sigma}_{f_1} + {\mathcal S})(x^{k+1} - \widetilde{x}^{k+1}) - {\mathcal A} z^{k+1} \in \partial f_2(x^{k+1}).
\end{array}
\end{equation}
Hence, from the convexity of $f_2(x)$, we have
\begin{equation}\nonumber
\Big\langle x^k - x^{k+1}, \nabla f_1(\widetilde{x}^{k+1}) - \nabla f_1(\widetilde{x}^k) + (\widehat{\Sigma}_{f_1} + {\mathcal S})[(x^{k+1} - x^k) - (\widetilde{x}^{k+1} - \widetilde{x}^k)] + {\mathcal A}(z^{k+1} - z^k)\Big\rangle\ge 0.
\end{equation}
Then, it yields that
\begin{align}\label{neqsubxz}
&\langle z^{k+1} - z^k, {\mathcal A}^*(x^k - x^{k+1})\rangle\nonumber\\[3mm]
=&\langle {\mathcal A}(z^{k+1} - z^k), x^k - x^{k+1}\rangle\nonumber\\[3mm]
\ge& \Big\langle x^{k+1} - x^k, \nabla f_1(\widetilde{x}^{k+1}) - \nabla f_1(\widetilde{x}^k)\Big\rangle - \Big\langle x^{k+1} - x^k, (\widehat{\Sigma}_{f_1} + {\mathcal S})(\widetilde{x}^{k+1} - \widetilde{x}^k)\Big\rangle\nonumber\\[3mm]
&+ \|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2.
\end{align}
Noting that $\nabla f_1$ is assumed be
globally Lipschitz continuous, then, from Clarkes Mean-Value Theorem \cite[Proposition 2.6.5]{OPNSA}, there exists a self-adjoint and positive semidefinite linear operator $\Gamma^k\in\partial^2(f_1(\gamma^k))$ with $\gamma^k$ being a point at the segment between $x^k$ and $x^{k+1}$, such that
\begin{equation}\nonumber
\nabla f_1(\widetilde{x}^{k+1}) - \nabla f_1(\widetilde{x}^k) = \Gamma^k(\widetilde{x}^{k+1} - \widetilde{x}^{k}).
\end{equation}
Then, according to (\ref{omegak}) and the left equality in (\ref{ident1}), the inequality (\ref{neqsubxz}) can be reorganized as
\begin{equation}\label{neqsubxzd}
\begin{array}{l}
\Big\langle z^{k+1} - z^k, {\mathcal A}^*(x^k - x^{k+1})\Big\rangle\\[3mm]
\ge \Big\langle x^{k+1} - x^k, (\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k)(\widetilde{x}^{k} - \widetilde{x}^{k+1})\Big\rangle + \|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2\\[3mm]
= \rho \Big\langle x^{k+1} - x^k, (\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k)(\widetilde{x}^{k} - x^{k})\Big\rangle + \|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2\\[3mm]
= \displaystyle{\frac{\rho}{2}}\Big[\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2 + \|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2 - \|\widetilde{x}^k - x^{k+1}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2\Big] + \|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2.
\end{array}
\end{equation}
Because $\widehat{\Sigma}_{f_1}\succeq \Gamma^k\succeq 0$ and $\frac{1}{2}\widehat{\Sigma}_{f_1} + {\mathcal S}\succeq 0$, we have
$$
\widehat{\Sigma}_{f_1} + {\mathcal S} - \frac{1}{2}\Gamma^k = \frac{1}{2}\widehat{\Sigma}_{f_1} + {\mathcal S} + \frac{1}{2}(\widehat{\Sigma}_{f_1} - \Gamma^k)\succeq 0.
$$
Then, according to (\ref{noneq}), the inequality (\ref{neqsubxzd}) can be rewritten as
\begin{equation}\label{neqsubxzdef}
\begin{array}{l}
\Big\langle z^{k+1} - z^k, A^*(x^k - x^{k+1})\Big\rangle\\[3mm]
\ge \displaystyle{\frac{\rho}{2}}\Big[\|x^{k+1} - x^k\|_{\hat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2 + \|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2 - \|\widetilde{x}^k - x^{k+1}\|_{\hat{\Sigma}_{f_1} + {\mathcal S} - \displaystyle{\frac{1}{2}}\Gamma^k}^2\Big]\\[3mm]
\quad + \|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2\\[3mm]
\ge \displaystyle{\frac{\rho}{2}}\Big[\|x^{k+1} - x^k\|_{\hat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2 + \|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S} - \Gamma^k}^2 - 2\|x^{k+1} - x^{k}\|_{\hat{\Sigma}_{f_1} + {\mathcal S} - \frac{1}{2}\Gamma^k}^2\\[3mm]
\quad - 2\|\widetilde{x}^k - x^{k}\|_{\hat{\Sigma}_{f_1} + {\mathcal S} - \frac{1}{2}\Gamma^k}^2\Big] + \|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2\\[3mm]
= \displaystyle{\frac{2-\rho}{2}}\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 - \displaystyle{\frac{\rho}{2}}\|\widetilde{x}^k - x^{k}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2.
\end{array}
\end{equation}
Substituting the inequality (\ref{neqsubxzdef}) into (\ref{eqcon}), it is easy to see that our conclusion (\ref{conneq}) holds.
\end{proof}
In light of above lemmas, we are ready to establish the convergence of the G-ADMM-M algorithm under some technical conditions to solve
the problem (\ref{prob1}).
\begin{theorem}\label{convmipgadmm}
Suppose that the Assumption (\ref{assum}) holds. Let the sequence $\{(x^k,y^k,z^k)\}$ be generated by Algorithm G-ADMM-M, and that $(\bar{x},\bar{y},\bar{z})\in \mathbb{W}^*$ is a solution of problem (\ref{prob1}). Let $\Psi_k$, $\delta_k$, $\theta_k$ and $\xi_k$ are defined by (\ref{psik}), (\ref{deltak}), (\ref{thetak}) and (\ref{xik}), respectively. For any $\sigma > 0$, $\rho\in(0,2)$, $\lambda\in(0,1]$, and $k > 1$, it holds that
\begin{equation}\label{convneq}
\begin{array}{l}
\Big(\Psi_k(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^k - x^{k}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2\Big)\\[3mm]
- \Big(\Psi_{k+1}(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^{k} - c\|^2\Big)\\[3mm]
\ge \delta_k - \xi_k.
\end{array}
\end{equation}
Moreover, if $\lambda\in(\frac{1}{2},1]$, we also assume that
\begin{equation}\label{pdefneq}
\widehat{\Sigma}_{h_1} + {\mathcal T} \succ 0, \quad {\mathcal F} \succ 0, \quad {\mathcal H} \succ 0,
\end{equation}
\begin{equation}\label{sumxik}
\sum_{k=1}^{\infty}\xi_k < +\infty,
\end{equation}
Then the sequence $\{(x^k,y^k)\}$ converges to an optimal solution of the primal problem (\ref{prob1})) and that $\{z^k\}$ converges to an optimal solution of the corresponding dual problem.
\end{theorem}
\begin{proof}
At the beginning, it holds that
\begin{equation}\label{constieq}
\begin{array}{rl}
&\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k} - c\|^2\\[3mm]
= &\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c + {\mathcal B}^*(y^k - y^{k-1})\|^2\\[3mm]
=&\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2 + \|{\mathcal B}^*(y^k - y^{k-1})\|^2 + 2 \Big\langle {\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c, {\mathcal B}^*(y^k - y^{k-1})\Big\rangle\\[3mm]
\ge& - \|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2 + \frac{1}{2}\|{\mathcal B}^*(y^k - y^{k-1})\|^2.
\end{array}
\end{equation}
According to (\ref{noneq}),(\ref{conneq}) and the inequality (\ref{constieq}) above, we have, for any $\sigma > 0$, $\rho\in(0,2)$, $\lambda\in(0,1]$, and $k > 1$, that
\begin{align}\label{thetaconstieq}
&\theta_k + \sigma(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k} - c\|^2\nonumber\\[3mm]
= &\theta_k + (1-\lambda)\sigma(2-\rho)\|{\mathcal A}^*x^{k} + B^*y^{k} - c\|^2 + \lambda\sigma(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k} - c\|^2\nonumber\\[3mm]
\ge& \theta_k + \lambda\Big(\sigma(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2 + \frac{(2-\rho)^2}{\rho}\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}+\sigma AA^*}\nonumber\\[3mm]
&- (2-\rho)\|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2\Big) - (1-\lambda)\sigma(2-\rho)\Big(\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2 - \frac{1}{2}\|{\mathcal B}^*(y^k - y^{k-1})\|^2\Big)\nonumber\\[3mm]
=& \|\widetilde{x}^{k+1} - x^{k+1}\|_{\frac{1}{2}\Sigma_{f_1}-\widehat{\Sigma}_{f_1}+(2-\rho)(\widehat{\Sigma}_{f_1}+{\mathcal S})}^2 + \|\widetilde{y}^{k} - y^{k}\|_{\frac{1}{2}\Sigma_{h_1}-\widehat{\Sigma}_{h_1}+(2-\rho)(\widehat{\Sigma}_{h_1}+{\mathcal T})}^2\nonumber\\[3mm]
&+ (1-\lambda)\sigma(2-\rho)\Big(\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2 - \|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2\Big) + (2\lambda-1)\sigma(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2\nonumber\\[3mm]
&+ \frac{\lambda(2-\rho)^2}{\rho}\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}+\sigma {\mathcal A}\A^*} - (2-\rho)\|\widetilde{x}^k - x^k\|_{\hat{\Sigma}_{f_1}+{\mathcal S}}^2 +(1-\lambda)(2-\rho)\|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2\nonumber\\[3mm]
& + \frac{\sigma(1-\lambda)(2-\rho)}{2}\|B^*(y^k - y^{k-1})\|^2\nonumber\\[3mm]
= &\|\widetilde{x}^{k+1} - x^{k+1}\|_{(2-\rho)(\widehat{\Sigma}_{f_1}+{\mathcal S})}^2 - (2-\rho)\|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\Big(\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2\nonumber\\[3mm]
&- \|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2\Big) + \|\widetilde{x}^{k+1} - x^{k+1}\|_{\frac{1}{2}\Sigma_{f_1}-\widehat{\Sigma}_{f_1}}^2 + \|\widetilde{y}^{k} - y^{k}\|_{\frac{1}{2}\Sigma_{h_1}-\widehat{\Sigma}_{h_1}+(2-\rho)(\widehat{\Sigma}_{h_1}+{\mathcal T})}^2\nonumber\\[3mm]
&+ (1-\lambda)(2-\rho)\|\widetilde{x}^k - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}}^2 + (2\lambda-1)\sigma(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^k - c\|^2\nonumber\\[3mm]
& + \frac{\sigma(1-\lambda)(2-\rho)}{2}\|{\mathcal B}^*(y^k - y^{k-1})\|^2 + \frac{\lambda(2-\rho)^2}{\rho}\|x^{k+1} - x^k\|_{\widehat{\Sigma}_{f_1}+{\mathcal S}+\sigma {\mathcal A}\A^*}.
\end{align}
Using Lemma \ref{lem5}, and substituting (\ref{thetaconstieq}) into (\ref{pivineq}), it can get the inequality (\ref{convneq}) easily.
Noting that $\delta_k \ge 0$ in the case of $\lambda\in(\frac{1}{2},1]$, then, using the conditions (\ref{pdefneq}) and (\ref{sumxik}), we have the following inequality from (\ref{convneq}), that is,
\begin{equation}\label{psikbneq}
\begin{array}{l}
0 \le \Psi_{k+1}(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^{k} - c\|^2\\[3mm]
\quad \le \Psi_k(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^k - x^{k}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k} + {\mathcal B}^*y^{k-1} - c\|^2 + \xi_k\\[3mm]
\quad \le \Psi_1(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^1 - x^1\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{1} + {\mathcal B}^*y^{0} - c\|^2 + \sum_{j=1}^k\xi_j\\[3mm]
\quad \le +\infty,
\end{array}
\end{equation}
which means that the sequences of $\{\Psi_k(\bar{x},\bar{y},\bar{z})\}$, $\{\|\widetilde{x}^k - x^{k}\|_{\hat{\Sigma}_{f_1} + {\mathcal S}}^2\}$ and $\{\|A^*x^{k+1} + B^*y^{k} - c\|^2\}$ are bounded. According to the definition of $\Psi_k$, we know that the sequences $\{\|z_e^k + \sigma(\rho-1){\mathcal A}^*x_e^{k}\|^2\}, \{\|{\mathcal A}^*x_e^{k}\|^2\}$, $\{\|\widetilde{x}_e^{k+1}\|_{\hat{\Sigma}_{f_1}+{\mathcal S}}^2\}$ and $\{\|\widetilde{y}_e^k\|_{\hat{\Sigma}_{h_1}+{\mathcal T}}^2\}$ are also bounded. Moreover, from (\ref{noneq}), we get
\begin{equation}\label{xekxkneq}
\frac{1}{2}\|x_e^{k}\|_{\hat{\Sigma}_{f_1} + {\mathcal S}}^2 \le \|\widetilde{x}^k - x^{k}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \|\widetilde{x}_e^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2.
\end{equation}
Also note that $\{\|A^*x_e^{k}\|^2\}$ is bounded, we have that the sequence of $\{\|x_e^{k}\|_{\hat{\Sigma}_{f_1} + {\mathcal S}+\sigma AA^*}^2\}$ is bounded. Then, the sequence $\{\|x^k\|^2\}$ is bounded because of the condition ${\mathcal F}\succ 0$. Similarly, we can get that the sequence $\{\|z^k\|^2\}$ is also bounded. Now, we prove that the sequence $\{\|y^k\|^2\}$ is bounded, too. For any $k\ge1$, from (\ref{convneq}), we have
\begin{equation}\nonumber
\begin{array}{l}
\sum_{j=1}^k\delta_j \le \Big(\Psi_1(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^1 - x^1\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{1} + {\mathcal B}^*y^{0} - c\|^2\Big)\\[3mm]
-\Big(\Psi_{k+1}(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^{k} - c\|^2\Big) + \sum_{j=1}^k\xi_j\\[3mm]
\le +\infty.
\end{array}
\end{equation}
Hence, it gets that $\lim_{k\rightarrow+\infty}\delta_k=0$. By the definition of $\delta_k$, when $k\rightarrow+\infty$, we have
\begin{equation}\label{deltaktend0}
\begin{array}{l}
\|\widetilde{x}^{k+1} - x^{k+1}\|_{\frac{1}{2}\Sigma_{f_1}}^2\rightarrow 0,\quad \|\widetilde{y}^{k} - y^{k}\|_{\frac{1}{2}\Sigma_{h_1}+(2-\rho)(\hat{\Sigma}_{h_1}+{\mathcal T})}^2\rightarrow 0,\quad\|\widetilde{x}^{k} - x^{k}\|_{\hat{\Sigma}_{f_1}+{\mathcal S}}^2\rightarrow 0,\\[3mm]
\|A^*x^{k+1} + B^*y^{k} - c\|^2\rightarrow 0,\quad \|B^*(y^k-y^{k-1})\|^2\rightarrow 0,\quad \|x^{k+1} - x^k\|_{\hat{\Sigma}_{f_1}+{\mathcal S}+\sigma AA^*}^2\rightarrow 0.\\[3mm]
\end{array}
\end{equation}
Then, from $\|\widetilde{y}^{k} - y^{k}\|_{\frac{1}{2}\Sigma_{h_1}+(2-\rho)(\widehat{\Sigma}_{h_1}+{\mathcal T})}^2\rightarrow 0$ and $\frac{1}{2}\Sigma_{h_1}+(2-\rho)(\widehat{\Sigma}_{h_1}+{\mathcal T})\succ 0$, we can easily know that $\|\widetilde{y}^{k} - y^{k}\|\rightarrow 0$. So, we know that the sequence $\{\|\widetilde{y}^{k} - y^{k}\|\}$ is bounded, and the sequence $\{\|\widetilde{y}_e^k\|\}$ is bounded too, because of $\hat{\Sigma}_{h_1}+{\mathcal T}\succ 0$. Moreover, from (\ref{noneq}), we have
\begin{equation}\nonumber
\frac{1}{2}\|y_e^{k}\|^2 \le \|\widetilde{y}^k - y^{k}\|^2 + \|\widetilde{y}_e^k\|^2.
\end{equation}
Hence, the sequence $\{\|y^k\|^2\}$ is bounded. Consequently, the sequence $\{(x^k,y^k,z^k)\}$ is bounded, and then there exists at least one subsequence $\{(x^{k_i},y^{k_i},z^{k_i})\}$ such that it converges to a cluster point, say $(x^{*},y^{*},z^{*})$.
In the following, we prove that $(x^{*},y^{*})$ is an optimal solution of the problem (\ref{prob1}) and $z^*$ is the corresponding Lagrange multiplier.
By using the first-order optimality condition of the $x$- and $y$-subproblem in (\ref{mipgadmm}) on the subsequence $\{(x^{k_i},y^{k_i},z^{k_i})\}$, we know that
\begin{equation}\label{fooconki}
\left\{
\begin{array}{l}
0 \in \partial f_2(x^{k_i}) + {\mathcal F} (x^{k_i} - \widetilde{x}^{k_i}) + \nabla f_1(\widetilde{x}^{k_i}) + \sigma {\mathcal A}({\mathcal A}^*\widetilde{x}^{k_i} + {\mathcal B}^*\widetilde{y}^{k_i} - c) + {\mathcal A}\widetilde{z}^{k_i},\\[3mm]
0 \in \partial h_2(y^{k_i}) + {\mathcal H} (y^{k_i} - \widetilde{y}^{k_i}) + \nabla h_1(\widetilde{y}^{k_i}) + \sigma {\mathcal B}({\mathcal A}^*x^{k_i} + {\mathcal B}^*\widetilde{y}^{k_i} - c) + Bz^{k_i}.
\end{array}
\right.
\end{equation}
From (\ref{deltaktend0}) and (\ref{fooconki}), we have
\begin{equation}\nonumber
\left\{
\begin{array}{l}
0 \in \partial f_2(x^{*}) + \nabla f_1({x}^{*}) + {\mathcal A}{z}^{*},\\[3mm]
0 \in \partial h_2(y^{*}) + \nabla h_1({y}^{*}) + {\mathcal B} z^{*},\\[3mm]
{\mathcal A}^*x^{*} + {\mathcal B}^*{y}^{*} - c = 0.
\end{array}
\right.
\end{equation}
which implies that $(x^*,y^*)$ is an optimal solution to problem (\ref{prob1}) and $z^*$ is the corresponding Lagrange multiplier.
Finally, we show that $(x^*,y^*,z^*)$ is the unique limit point of the sequence $\{(x^k,y^k,z^k)\}$. Without loss of generality, let $(\bar{x},\bar{y},\bar{z}):=(x^*,y^*,z^*)$, from the definition of $\Psi_k$ and the fact that the subsequence $\{(x^{k_i},y^{k_i},z^{k_i})\}\rightarrow (x^*,y^*,z^*)$ as $i\rightarrow\infty$, we have
$$
\lim_{i\rightarrow+\infty}\Big\{\Psi_{k_i}(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^{k_i} - x^{k_i}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k_i} + {\mathcal B}^*y^{k_i-1} - c\|^2\Big\} = 0.
$$
From $(\ref{psikbneq})$, for any $k>k_i$, we have
\begin{equation}\nonumber
\begin{array}{l}
0 \le \Psi_{k+1}(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^{k+1} - x^{k+1}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k+1} + {\mathcal B}^*y^{k} - c\|^2\\[3mm]
\quad \le \Psi_{k_i}(\bar{x},\bar{y},\bar{z}) + (2-\rho)\|\widetilde{x}^{k_i} - x^{k_i}\|_{\widehat{\Sigma}_{f_1} + {\mathcal S}}^2 + \sigma(1-\lambda)(2-\rho)\|{\mathcal A}^*x^{k_i} + {\mathcal B}^*y^{k_i-1} - c\|^2 + \sum_{j=k_i}^k\xi_j.\\[3mm]
\end{array}
\end{equation}
Obviously, combining this inequality with (\ref{deltaktend0}), we get
\begin{equation}\nonumbe
\lim_{k\rightarrow+\infty}\Psi_{k}(\bar{x},\bar{y},\bar{z}) = 0,
\end{equation}
which means that
\begin{subequations}
\begin{equation}\label{subzklim}
\lim_{k\rightarrow+\infty}\|z_e^k + \sigma(\rho-1){\mathcal A}^*x_e^k\|^2 = 0
\end{equation}
\begin{equation}\label{subAkxklim}
\lim_{k\rightarrow+\infty}\|{\mathcal A}^*x_e^k\|^2 = 0
\end{equation}
\begin{equation}\label{subwxklim}
\lim_{k\rightarrow+\infty}\|\widetilde{x}_e^k\|_{\hat{\Sigma}_{f_1} + {\mathcal S}}^2 = 0
\end{equation}
\begin{equation}\label{subwyklim}
\lim_{k\rightarrow+\infty}\|\widetilde{y}_e^k\|_{\hat{\Sigma}_{h_1} + {\mathcal T}}^2 = 0
\end{equation}
\end{subequations}
From (\ref{xekxkneq}), (\ref{deltaktend0}) and (\ref{subwxklim}), we know that $\lim_{k\rightarrow+\infty}\|x_e^k\|_{\hat{\Sigma}_{f_1} + {\mathcal S}}^2 = 0$. Then, from (\ref{subAkxklim}), we get
$$
\lim_{k\rightarrow+\infty}\|x_e^k\|_{\widehat{\Sigma}_{f_1} + {\mathcal S} + \sigma {\mathcal A}\A^*}^2 = 0,
$$
Because of $\widehat{\Sigma}_{f_1} + {\mathcal S} + \sigma {\mathcal A}\A^*\succ 0$, we know that $\lim_{k\rightarrow+\infty}x^k = \bar{x}$. Homoplastically, we get $\lim_{k\rightarrow+\infty}y^k = \bar{y}$.
In a similar way, from (\ref{noneq}), we get
\begin{equation}\nonumbe
\frac{1}{2}\|z_e^{k}\|^2 \le \|z_e^k + \sigma(\rho-1){\mathcal A}^*x_e^k\|^2 + \|\sigma(\rho-1){\mathcal A}^*x_e^k\|^2,
\end{equation}
then, from (\ref{subzklim}) and $\lim_{k\rightarrow+\infty}x_e^k = 0$, we get $\lim_{k\rightarrow+\infty}z_e^k = 0$, and hence, $\lim_{k\rightarrow+\infty}z^k = \bar{z}$. Then, we know that the sequence $\{(x^k,y^k,z^k)\}$ converges to $(\bar{x},\bar{y},\bar{z})$. The poof is complete.
\end{proof}
\setcounter{equation}{0}
\section{Numerical experiments}\label{section5}
In this section, we evaluate the feasibility and efficiency of the G-ADMM-M algorithm by using some simulated convex composite optimization
problems. All experiments are performed under Microsoft Windows 10 and MATLAB R2020b, and run on a Lenovo Laptop with an Intel Core i7-1165G7 CPU at 2.80 GHz and 8 GB of memory.
In this test, we focus on the following convex composite quadratic optimization problem which has been
considered in \cite{miPADMM} and \cite{MGADMM}, that is
\begin{equation}\label{numexprb1}
\min_{x\in{\mathcal R}^m,y\in{\mathcal R}^n} \ \Big\{\frac{1}{2}\langle y, {\mathcal Q} y \rangle - \langle b, y\rangle + \frac{\chi}{2}\|\Pi_{{\mathcal R}_{+}^m}({\mathcal D} (d-{\mathcal H} y))\|^2 + \mu\|y\|_1 + \delta_{{\mathcal R}_{+}^m}(x) \ : \ {\mathcal H} y + x = c \Big\}
\end{equation}
where ${\mathcal Q}$ is an $n\times n$ symmetric and positive semidefinite matrix (may not be positive definite); $\chi$ is a nonnegative penalty parameter; $\Pi_{{\mathcal R}_{+}^m}(\cdot)$ denotes the projection onto ${\mathcal R}_{+}^m$;
${\mathcal H}\in{\mathcal R}^{m \times n}$, $b\in{\mathcal R}^n$, $c\in{\mathcal R}^m$, $\mu > 0$ and $d\le c$ are given data; ${\mathcal D}$ is a positive definite diagonal matrix chosen to normalize each row of ${\mathcal H}$ to have an unit norm.
Observing that problem (\ref{numexprb1}) can be expressed in the from of (\ref{prob1}) if it takes
\begin{equation}\nonumber
\left\{
\begin{array}{l}
h_1(y) = \frac{1}{2}\langle y, {\mathcal Q} y \rangle - \langle b, y\rangle + \frac{\chi}{2}\|\Pi_{{\mathcal R}_{+}^m}({\mathcal D} (d-{\mathcal H} y))\|^2, \ h_2(y) = \mu\|y\|_1,\\[3mm]
f_1(x) \equiv 0, \ f_2(x) = \delta_{{\mathcal R}_{+}^m}(x), \ {\mathcal A}^* = I, \ {\mathcal B}^* = {\mathcal H}.
\end{array}
\right.
\end{equation}
It is not hard to deduce that the KKT system for problem (\ref{numexprb1}) is the following form:
\begin{equation}\label{numexpkkt}
\nabla h_1(y) + {\mathcal H}^* z + v = 0, \ v\in\partial\mu\|y\|_1, \ {\mathcal H} y+x-c=0, \ x \ge 0, \ z \ge 0, \ x \circ z=0.
\end{equation}
where $z$ and $v$ are the corresponding dual variables, $'\circ'$ denotes the elementwise product.
In the following, for performance comparisons, we also employ the M-GADMM of Qin et al. \cite{MGADMM} with $\rho = 1.9$ and M-ADMM of Li et al. \cite{miPADMM} with the step-length parameters $\tau = 1.618$ to solve the problem (\ref{numexprb1}), respectively.
In our algorithm G-ADMM-M, we set the parameter $\mu=5\sqrt{n}$ and $d = c-5e$ where $e$ is a vector of all ones. Besides, we choose the zero vector as an initial point, and the penalty parameter is initialized as $\sigma = 0.8$. Based on the KKT system (\ref{numexpkkt}), we stop the iterative process based on the residual of the KKT system, that is,
\begin{equation}\nonumber
Res:= \max\Big\{\frac{\|{\mathcal H} y^{k+1} + x^{k+1} -c\|}{1+\|c\|}, \frac{\|\nabla h_1(y^{k+1}) + {\mathcal H}^*z^{k+1} + v^{k+1}\|}{1+\|b\|}\Big\}\le 10^{-5}.
\end{equation}
In the numerical experiments, we choose the same proximal terms and operators as in \cite{miPADMM} and \cite{MGADMM}, i.e.
\begin{equation}\nonumber
\begin{array}{l}
\widehat{\Sigma}_{h_1} = {\mathcal Q} + \chi {\mathcal H}^*{\mathcal D}^2{\mathcal H}, \ \Sigma_{h_1}={\mathcal Q}, \ \widehat{\Sigma}_{f_1} = \Sigma_{f_1}= 0,\\[3mm]
{\mathcal T} = \lambda_{\max}(\widehat{\Sigma}_{h_1} + \sigma {\mathcal H}^*{\mathcal H} ){\mathcal I} - (\widehat{\Sigma}_{h_1} + \sigma {\mathcal H}^*{\mathcal H}), \ {\mathcal S} = 0.
\end{array}
\end{equation}
With these choices, it is easy to verify that the conditions (\ref{pdefneq}) in Theorem (\ref{convmipgadmm}) are satisfied. Using these notations, we choose different pairs of $(m, n)$ to generate the data used in (\ref{numexprb1}) and the numerical results derived by each algorithm are listed in Tables (\ref{tab1} - \ref{tab2}), which includes the number of iterations (Iter), the
computing time (Time), the KKT residuals of the solution (Res). From these tables, we find that the G-ADMM-M is beter than M-GADMM and M-ADMM in terms of iteration number and computation time almost
when a similar accuracy solutions are achieved. In particular, we can see that our proposed algorithm is $15-35\%$ faster than the M-ADMM.
\begin{table}[ht]
\centering
\renewcommand\tabcolsep{4.5pt}
{\scriptsize
\caption{\small Numerical results of M-ADMM, M-GADMM and G-ADMM-M with $\chi=0$.}\vspace{-.1cm}
\begin{tabular}{lc|ccc|ccc|ccc}
\hline
\multirow{2}{*}{m}&
\multirow{2}{*}{n}&
\multicolumn{3}{c|}{M-ADMM} & \multicolumn{3}{c|}{M-GADMM} & \multicolumn{3}{c}{G-ADMM-M} \\
\cline{3-11}
& & Iter & Time & Res & Iter & Time & Res & Iter & Time & Res \\
\hline
200& 500& 5966& 1.44& 9.98e-06& 5032& 1.10& 9.98e-06& 4886& 1.16& 9.98e-06 \\
500& 200& 576& 0.18& 1.00e-05& 581& 0.17& 9.99e-06& 423& 0.14& 9.96e-06 \\
500& 500& 794& 1.83& 9.96e-06& 747& 1.59& 9.96e-06& 628& 1.33& 1.00e-05 \\
1000& 500& 628& 2.80& 9.94e-06& 579& 2.50& 9.98e-06& 444& 1.88& 9.99e-06 \\
500& 1000& 1354& 4.98& 9.97e-06& 1179& 4.27& 9.90e-06& 1121& 4.12& 9.97e-06 \\
1000& 1000& 758& 6.39& 9.89e-06& 696& 5.72& 9.98e-06& 558& 4.64& 9.92e-06 \\
2000& 1000& 630& 11.11& 9.90e-06& 608& 10.61& 9.93e-06& 439& 7.73& 9.96e-06 \\
1000& 2000& 3088& 46.15& 1.00e-05& 2573& 38.23& 1.00e-05& 2667& 40.05& 1.00e-05 \\
2000& 2000& 778& 24.39& 9.88e-06& 720& 22.18& 9.96e-06& 550& 17.31& 1.00e-05 \\
4000& 2000& 587& 41.89& 9.95e-06& 575& 39.78& 9.97e-06& 397& 28.32& 9.88e-06 \\
2000& 4000& 1816& 105.28& 1.00e-05& 1553& 89.32& 1.00e-05& 1531& 90.12& 9.99e-06 \\
4000& 4000& 708& 89.75& 9.97e-06& 683& 86.87& 9.99e-06& 524& 68.35& 9.99e-06 \\
4000& 8000& 1334& 320.39& 9.99e-06& 1103& 262.73& 9.99e-06& 1034& 254.14& 9.99e-06 \\
8000& 4000& 681& 194.44& 9.95e-06& 708& 197.55& 9.99e-06& 487& 141.90& 9.95e-06 \\
8000& 8000& 793& 460.87& 9.98e-06& 779& 415.07& 9.97e-06& 547& 306.90& 9.99e-06 \\
\hline
\end{tabular}\label{tab1}
}
\end{table}
\begin{table}[ht]
\centering
\renewcommand\tabcolsep{4.5pt}
{\scriptsize
\caption{\small Numerical results of M-ADMM, M-GADMM and G-ADMM-M with $\chi=2\mu$.}\vspace{-.1cm}
\begin{tabular}{lc|ccc|ccc|ccc}
\hline
\multirow{2}{*}{m}&
\multirow{2}{*}{n}&
\multicolumn{3}{c|}{M-ADMM} & \multicolumn{3}{c|}{M-GADMM} & \multicolumn{3}{c}{G-ADMM-M} \\
\cline{3-11}
& & Iter & Time & Res & Iter & Time & Res & Iter & Time & Res \\
\hline
200& 500& 6221& 1.44& 9.99e-06& 5221& 1.18& 9.99e-06& 5004& 1.28& 1.00e-05 \\
500& 200& 593& 0.19& 9.92e-06& 568& 0.21& 9.89e-06& 385& 0.13& 9.31e-06 \\
500& 500& 526& 1.08& 9.87e-06& 495& 0.97& 9.86e-06& 360& 0.88& 9.58e-06 \\
1000& 500& 458& 2.09& 1.00e-05& 445& 1.95& 9.93e-06& 296& 1.25& 9.96e-06 \\
500& 1000& 1715& 6.49& 1.00e-05& 1423& 5.21& 9.99e-06& 1519& 5.53& 9.99e-06 \\
1000& 1000& 430& 3.72& 9.96e-06& 409& 3.46& 9.94e-06& 303& 2.51& 9.65e-06 \\
2000& 1000& 445& 7.86& 9.86e-06& 441& 7.76& 9.91e-06& 279& 4.95& 9.90e-06 \\
1000& 2000& 1466& 21.58& 1.00e-05& 1249& 18.57& 9.97e-06& 1296& 19.57& 9.99e-06 \\
2000& 2000& 341& 10.87& 9.88e-06& 325& 10.39& 9.86e-06& 237& 7.73& 9.79e-06 \\
4000& 2000& 433& 30.80& 9.97e-06& 430& 30.06& 9.97e-06& 263& 19.15& 9.82e-06 \\
2000& 4000& 1276& 74.11& 9.99e-06& 1075& 62.47& 9.96e-06& 1104& 65.02& 9.96e-06 \\
4000& 4000& 312& 42.04& 9.96e-06& 306& 41.44& 9.94e-06& 211& 29.88& 9.99e-06 \\
4000& 8000& 1028& 253.05& 1.00e-05& 841& 203.23& 9.97e-06& 903& 224.51& 9.99e-06 \\
8000& 4000& 463& 135.72& 9.98e-06& 470& 134.67& 9.92e-06& 298& 90.87& 9.92e-06 \\
8000& 8000& 315& 188.71& 9.91e-06& 314& 183.85& 9.99e-06& 207& 136.79& 9.85e-06 \\
\hline
\end{tabular}\label{tab2}
}
\end{table}
\section{Concluding remarks}\label{section6}
We have known that the generalized ADMM of Xiao et al. \cite{spGADMM} has the ability to solve two-block linearly constrained composite convex programming problem, in which the smooth term located in each block must be `quadratic'.
This paper enhanced the capacity of this method and showed that if we used a majorized quadratic function to replace the non-quadratic smooth function, the resulting subproblem in the ADMM framework can be split some smaller problems if a symmetric Gauss-Seidel iteration is used.
The majority of this paper is on the theoretical analysis, that is, we proved that the sequence generated by the proposed method converges to one of a KKT points of the considered problem. Besides, we also did some simulated experiments to illustrate the effectiveness of the majorization technique and the progressiveness of the proposed method. At last, we must note that, further performance evaluation of the method using various problems deserves investigating. Of course, research on the proposed method's convergence rate and its iteration complexity are also interesting topics for further research.
\section*{Acknowledgments}
We would like to thank Z. Ma from Henan University for her valuable comments on the proof of this paper. The work of Y. Xiao is supported by the National Natural Science Foundation
of China (Grant No. 11971149).
|
\section{Introduction}
Estimating the Hausdorff dimension of limsup sets obtained from the contractions of the elements of a given family of sets is a natural question of metric approximation theory, which arises in many contexts. In this article, given a sequence of balls of $\mathbb{R}^d$, $\mathcal{B}= (B(x_n,r_n))_{n\in\mathbb{N}}$, we investigate in a very general frame the size properties of the limsup sets obtained from smaller sets, i.e sets of the form $\limsup_{n\to +\infty} U_n $, where $U_n \subset B_n$.
Let us recall that the historical example of Jarnik-Besicovitch's theorem deals with the case $U_n=B_n^\delta:= B(x_n,r_n^{\delta })$, where $\delta>1$, $x_n$ is a rational number $\frac{p}{q}$ and $r_n =\frac{1}{q^2}$. More generally in metric number theory, one often aims at computing the Hausdorff dimension of sets $\limsup_{n\to\infty} U_n$, where $(U_n)_{n\in\mathbb{N}}$ has some algebraic or dynamical meaning.
Generalizations of Jarnik-Besicovitch's Theorem often consider a given sequence $(x_n)_{n\in\mathbb{N}}$ of points in $\R^d$, as well as a sequence of radii $(r_n)_{n\in\mathbb{N}}$ for which the associated limsup set $E_{1}=\limsup_{n\rightarrow+\infty}B_n $ has a controlled size (in terms of Lebesgue measure or Hausdorff dimension for instance); then, given a sequence of sets $\mathcal{U}=(U_n)_{n\in\mathbb{N}}$ with, for every $n\in\mathbb{N}$, $U_n \subset B_n$, one estimates the Hausdorff dimension of the smaller limsup set $E(\mathcal{U})=\limsup_{n\rightarrow+\infty}U_n$. The classical case is when the set $U_n$ is a shrunk ball $B_n^{\delta}$, for some $\delta>1$, that is $E(\mathcal{U})$ is the limsup set of the $\delta$-contracted balls, but different shapes for $U_n$ have also been considered (rectangles or ellipsoïds rather than balls for instance). Such problems are studied for instance in \cite{Ja,BV,BS2,RE,KR} among many references.
The same question arises on any topological dynamical system $(X,T)$ endowed with some metric, when the sequence $(x_n)_{n\in\mathbb{N}}$ is the orbit $(T^{n}(x))_{n\in\mathbb N}$ of a well chosen point $x$. Some specific cases are for instance treated in \cite{HV,LS,PR}. In probability theory, the famous Dvoretzky covering problem consists in computing, when it is possible, the Hausdorff dimension of the limsup set associated with a sequence of random balls drawn independently and uniformly in a compact Baire space, see for instance \cite{FJS,EP,BS2}. {In analysis, the value of the pointwise regularity exponents of measures and functions at a given point $x$ often relies on the ability to understand how $x$ is close to remarkable points $x_n$. The reader may refer to \cite{Ja,BS3,BS4}.
{As mentioned above, in the largest part of the literature, a strong geometric measure theoretic condition is initially imposed on $\mathcal{B}$ to obtain results, for instance that the Lebesgue measure of $\limsup_{n\to +\infty} B(x_n,r_{ n})$ is full (cf \cite{BV}). But there are many situations in which the Lebesgue measure is not the relevant measure to work with (cf \cite{BS}).}
{Our purpose in this article is to obtain a general lower bound for the Hausdorff dimension $\limsup_{n\to +\infty} U_n$, where the sets $(U_n)$ are open sets in some balls $(B_n)$ satisfying the property called $\mu$-asymptotically covering property, where $\mu$ is a probability measure on $\R^d$. This property, introduced in \cite{ED2}, is proved to be almost equivalent to verifying that $\mu(\limsup_{n\rightarrow+\infty}B_n)=1$ (e.g \cite{ED2}).
The results presented here extend, for instance, both the results of Koivusalo-Rams stated in \cite{KR} and the result of Barral-Seuret (\cite{BS2}) which deals with balls and self-similar measures under the open set condition. It is worth noticing that the work of Koivusalo and Rams in \cite{KR} highlighted the importance of the Hausdorff content to compute Hausdorff dimension of limsup sets and this article makes further use of this fact.
An important advantage of the lower bound obtained in the present paper is that its value is tractable in many cases. For instance, as a first application, a ubiquity theorem is given in the case where $\mu$ is a self-similar measure (we do not require any condition on the possible overlaps associated with such a $\mu$).
Two other applications of our main result are treated in this article as well. The problem of self-similar shrinking targets is studied when the corresponding iterated function system (in short IFS) is dimension-regular and has similarity dimension less than $d$, meaning in particular that for every self-similar measure, the similarity dimension and the Hausdorff dimension coincide (see Section \ref{shrtag}, Definition \ref{dimreg}).
Another application in Diophantine approximation is given. Let $K_{1/3}^{(0)}$ the set of points of $[0,1]$ such that in their sequence of digits in basis $3$, the asymptotic frequency of appearance of the digit $1$ is infinitely many often close to $0$ (note that this set contains the middle-third Cantor set $K_{ 1/3}$ and $\dim_H (K_{1/3}^{(0)})=\dim_H (K_{1/3})$). We compute the Hausdorff dimension of points of $K_{1/3}^{(0)}$ well approximable by rational (see Theorem \ref{main} for a precise statement).
\section{Definitions and main statements}\label{sec1}
Let us start with some notations
Let $d$ $\in\mathbb{N}$. For $x\in\mathbb{R}^{d}$, $r>0$, $B(x,r)$ stands for the closed ball of ($\mathbb{R}^{d}$,$\vert\vert \ \ \vert\vert_{\infty}$) of center $x$ and radius $r$.
Given a ball $B$, $\vert B\vert$ stands for the diameter of $B$. For $t\geq 0$, $\delta\in\mathbb{R}$ and $B=B(x,r)$, $t B$ stand for $B(x,t r)$, i.e. the ball with same center as $B$ and radius multiplied by $t$, and the $\delta$-contracted ball $B^{\delta}$ is defined by $B^{\delta}=B(x ,r^{\delta})$.
\smallskip
Given a set $E\subset \mathbb{R}^d$, $\widering{E}$ stands for the interior of the $E$, $\overline{E}$ its closure and $\partial E$ its boundary, i.e, $\partial E =\overline{E}\setminus \widering{E}$. If $E$ is a Borel subset of $\R^d$, its Borel $\sigma$-algebra is denoted by $\mathcal B(E)$.
\smallskip
Given a topological space $X$, the Borel $\sigma$-algebra of $X$ is denoted $\mathcal{B}(X)$ and the space of probability measure on $\mathcal{B}(X)$ is denoted $\mathcal{M}(X).$
\sk
The $d$-dimensional Lebesgue measure on $(\mathbb R^d,\mathcal{B}(\mathbb{R}^d))$ is denoted by
$\mathcal{L}^d$.
\smallskip
For $\mu \in\mathcal{M}(\R^d)$, $\supp(\mu)=\left\{x\in[0,1]: \ \forall r>0, \ \mu(B(x,r))>0\right\}$ is the topological support of $\mu$.
\smallskip
Given $E\subset \mathbb{R}^d$, $\dim_{H}(E)$ and $\dim_{P}(E)$ denote respectively the Hausdorff and the packing dimension of $E$.
\smallskip
Now we recall some definitions.
\begin{definition}
\label{hausgau}
Let $\zeta :\mathbb{R}^{+}\mapsto\mathbb{R}^+$. Suppose that $\zeta$ is increasing in a neighborhood of $0$ and $\zeta (0)=0$. The Hausdorff outer measure at scale $t\in(0,+\infty]$ associated with $\zeta$ of a set $E$ is defined by
\begin{equation}
\label{gaug}
\mathcal{H}^{\zeta}_t (E)=\inf \left\{\sum_{n\in\mathbb{N}}\zeta (\vert B_n\vert) : \, \vert B_n \vert \leq t, \ B_n \text{ closed ball and } E\subset \bigcup_{n\in \mathbb{N}}B_n\right\}.
\end{equation}
The Hausdorff measure associated with $\zeta$ of a set $E$ is defined by
\begin{equation}
\mathcal{H}^{\zeta} (E)=\lim_{t\to 0^+}\mathcal{H}^{\zeta}_t (E).
\end{equation}
\end{definition}
For $t\in (0,+\infty]$, $s\geq 0$ and $\zeta:x\mapsto x^s$, one simply uses the usual notation $\mathcal{H}^{\zeta}_t (E)=\mathcal{H}^{s}_t (E)$ and $\mathcal{H}^{\zeta} (E)=\mathcal{H}^{s} (E)$, and these measures are called $s$-dimensional Hausdorff outer measure at scale $t\in(0,+\infty]$ and $s$-dimensional Hausdorff measure respectively. Thus,
\begin{equation}
\label{hcont}
\mathcal{H}^{s}_{t}(E)=\inf \left\{\sum_{n\in\mathbb{N}}\vert B_n\vert^s : \, \vert B_n \vert \leq t, \ B_n \text{ closed ball and } E\subset \bigcup_{n\in \mathbb{N}}B_n\right\}.
\end{equation}
The quantity $\mathcal{H}^{s}_{\infty}(E)$ (obtained for $t=+\infty$) is called the $s$-dimensional Hausdorff content of the set $E$.
\begin{definition}
\label{dim}
Let $\mu\in\mathcal{M}(\mathbb{R}^d)$.
For $x\in \supp(\mu)$, the lower and upper local dimensions of $\mu$ at $x$ are defined as
\begin{align*}
\underline\dim_{{\rm loc}}(\mu,x)=\liminf_{r\rightarrow 0^{+}}\frac{\log(\mu(B(x,r)))}{\log(r)}
\ \mbox{ and } \ \overline\dim_{{\rm loc}}(\mu,x)=\limsup_{r\rightarrow 0^{+}}\frac{\log (\mu(B(x,r)))}{\log(r)}.
\end{align*}
Then, the lower and upper Hausdorff dimensions of $\mu$ are respectively defined by
\begin{equation}
\label{dimmu}
\dimm(\mu)={\mathrm{ess\,inf}}_{\mu}(\underline\dim_{{\rm loc}}(\mu,x)) \ \ \mbox{ and } \ \ \overline{\dim}_P (\mu)={\mathrm{ess\,sup}}_{\mu}(\overline\dim_{{\rm loc}}(\mu,x)).
\end{equation}
\end{definition}
It is known (for more details see \cite{F}) that
\begin{equation*}
\begin{split}
\dimm(\mu)&=\inf\{\dim_{H}(E):\, E\in\mathcal{B}(\mathbb{R}^d),\, \mu(E)>0\} \\
\overline{\dim}_P (\mu)&=\inf\{\dim_P(E):\, E\in\mathcal{B}(\mathbb{R}^d),\, \mu(E)=1\}.
\end{split}
\end{equation*}
When $\underline \dim_H(\mu)=\overline \dim_P(\mu)$, this common value is simply denoted by $\dim(\mu)$ and~$\mu$ is said to be \textit{exact dimensional}.
\medskip
\subsection{The $\mu$-a.c property}
\mk
We fix a sequence of closed balls $\mathcal B=(B_n)_{n\in\mathbb N}$ such that $\lim_{n\to +\infty} |B_n| = 0$ (otherwise the situation is trivial for the questions we consider).
\mk
The main property (introduced in \cite{ED2}) used for the sequence of balls $\mathcal{B}$ is meant to ensure that any set can be covered efficiently by the limsup of the $B_n$'s, with respect to a measure $\mu$. This property is a general version of the key covering property used in the KGB Lemma of Beresnevitch and Velani, stated in \cite{BV}, using a Borel probability measure~$\mu$. Observe that such properties (like the KGB Lemma) are usually key (cf \cite{Ja,BV,BS} for instance) to prove ubiquity or mass transference results.
\begin{definition}
\label{ac}
Let $\mu\in \mathcal{M}(\R^d)$. The sequence $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ of balls of $\R^d$ is said to be $\mu$-asymptotically covering (in short, $\mu$-a.c) when there exists a constant $C>0$ such that for every open set $\Omega\subset \R^d $ and $g\in\mathbb{N}$, there is an integer $N_\Omega \in\mathbb{N}$ as well as $g\leq n_1 \leq ...\leq n_{N_\Omega}$ such that:
\begin{itemize}
\item [(i)]$\forall \, 1\leq i\leq N_\Omega$, $B_{n_i}\subset \Omega$;
\item [(ii)]$\forall \, 1\leq i\neq j\leq N_\Omega$, $B_{n_i}\cap B_{n_j}=\emptyset$;
\item [(iii)] also,
\begin{equation}
\label{majac}
\mu\Big(\bigcup_{i=1}^{N_\Omega}B_{n_i} \Big )\geq C\mu(\Omega).
\end{equation}
\end{itemize}
\end{definition}
In other words, for any open set $\Omega$ and any integer $g\ge 1$, there exits a finite set of disjoint balls of $\left\{B_n\right\}_{n\geq g}$ supporting a fixed proportion of $\mu(\Omega)$.
This notion of $\mu$-asymptotically covering is related to the way the balls of $\mathcal{B}$ are distributed according to the measure $\mu$. This property is a priori slightly stronger than having a $\limsup$ of full $\mu$-measure when $\mu$ is not doubling, as suggested by the following lemma proved in \cite{ED2}, and whose second item will be used to apply our main theorem to self-similar measures. However, it follows from the proof of \cite[Lemma 5]{BV} that these properties are equivalent when $\mu$ is doubling.
\begin{lemme}
\label{equiac}
Let $\mu\in\mathcal{M}(\R^d)$ and $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ be a sequence of balls of $\R^d$ with $\lim_{n\to +\infty} r_{n}= 0$.
\begin{enumerate}
\smallskip
\item
If $\mathcal{B} $ is $\mu$-a.c, then $\mu(\limsup_{n\rightarrow+\infty}B_n)=1.$
\smallskip
\item
If there exists $v<1$ such that $ \mu \big(\limsup_{n\rightarrow+\infty}(v B_n) \big)=1$, then $\mathcal{B} $ is $\mu$-a.c.
\end{enumerate}
\end{lemme}
\subsection{Essential content and statement of the main result}
\mk
The key geometric notion for the ubiquity theorem developed in this paper is the following.
\begin{definition}
\label{mucont}
{Let $\mu \in\mathcal{M}(\R^d)$, and $s\geq 0$.
The $s$-dimensional $\mu$-essential Hausdorff content at scale $t\in(0,+\infty]$ of a set $A\subset \mathcal B(\R^d)$ is defined as}
{\begin{equation}
\label{eqmucont}
\mathcal{H}^{\mu,s}_{t}(A)=\inf\left\{\mathcal{H}^{s}_{t}(E): \ E\subset A , \ \mu(E)=\mu(A)\right\}.
\end{equation}}
\end{definition}
One will almost exclusively look at these contents at scale $t=+\infty$ and one refers to $ \mathcal{H}^{\mu,s}_{\infty}(A)$ as the {\em $s$-dimensional $\mu$-essential Hausdorff content of $A$}. Basic properties of those quantities are studied in Section \ref{sec-mino}, and precise estimates of $ \mathcal{H}^{\mu,s}_{\infty}(A)$ are achieved for the Lebesgue measure and self-similar measures in Section \ref{sec-example}.
Note that in \cite[Theorem 3.1]{KR} the key underlying geometric notion used to handle the variety of shapes of the sets $(U_n)_{n\in\mathbb{N}}$ is the Hausdorff content. It is easily seen from \eqref{hcont} that the Hausdorff content also carries some ``high scale'' geometric information (because there is no restriction concerning the diameter of the balls $(B_n)$ in \eqref{hcont}). This will also be the case in this article to handle not only the shape of the sets $(U_n)_{n\in\mathbb{N}}$ but also the geometric behavior related to the measure $\mu$ at high scale in the sets $(U_n)_{n\in\mathbb{N}}.$
\medskip
The $s$-dimensional $\mu$-essential Hausdorff content is now used to associate a critical exponent to any sequence of open sets $(U_n)_{n\in\mathbb N}$ such that $U_n\subset B_n$ for all $n\in\mathbb N$. This exponent is involved in our lower bound estimate of $\dim_H (\limsup_{n\to +\infty} U_n)$.
\begin{definition}
\label{expoani}
Let $\mu\in\mathcal{M}(\R^d)$. If $B$ and $U$ are Borel subsets of $\R^d$, the {\em $\mu$-critical exponent} of $(B,U)$ is defined as
\begin{equation}
\label{defsndani}
s_\mu(B,U) =\sup\left\{s\geq 0 \ : \ \mathcal{H}^{\mu,s}_{\infty}\left (U\right)\geq \mu( B)\right\}.
\end{equation}
Let $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ be a sequence of closed balls, $\mathcal{U}= (U_n)_{n\in\mathbb{N}}$ a sequence of Borel subsets of $\R^d$, and $s\geq 0$.
Let
\begin{align}
\label{defNsdani}
\mathcal{N}_\mu(\mathcal{B},\mathcal{U},s) & =\left\{n\in\mathbb{N} \ : s_\mu(B_n,U_n) \geq s\right\}.
\end{align}
Then, define the $\mu$-critical exponent of $(\mathcal{B},\mathcal{U})$ as
\begin{align}
\label{defsmdani}
s(\mu,\mathcal{B},\mathcal{U}) & =\sup\left\{s \geq 0 : \ (B_n)_{n\in\mathcal{N}_\mu(\mathcal{B},\mathcal{U},s)} \mbox{ is } \mu\mbox{-a.c.} \right\}.
\end{align}
\end{definition}
It is worth noting that, for $s^{\prime}\leq s$, one has $ \mathcal{N}_\mu(\mathcal{B},\mathcal{U},s)\subset \mathcal{N}_\mu(\mathcal{B},\mathcal{U},s^{\prime}) $.
\medskip
The main result of this paper is the following.
\begin{theoreme}
\label{lowani}
Let $\mathcal{B} =(B_n)_{n\in \N} $ be a sequence of closed balls of $\R^d$ such that $\vert B_n \vert \to 0$ and $\mathcal{U}=(U_n)_{n\in\mathbb{N}}$ a sequence of open sets such that $U_n \subset B_n$ for all $n\in\mathbb{N}$.
Then, for every $\mu \in\mathcal{M}(\R^d)$ such that $\min\left\{s(\mu , \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}>0$ there exists a gauge function $\zeta:\R^+\to \R^+$ such that $\lim_{r\to 0^+} \frac{\log \zeta(r)}{\log r} =\min\left\{ s(\mu, \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}$ and
$$\mathcal{H}^\zeta( \limsup_{n\rightarrow +\infty}U_n) >0.$$
In particular, for every $\mu \in\mathcal{M}(\R^d)$, one has
\begin{equation}
\label{conc2ani}
\dim_{H}\left(\limsup_{n\rightarrow +\infty}U_n\right)\geq \min\left\{s(\mu , \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}.
\end{equation}
\end{theoreme}
\begin{remarque}
(1) It is easily verified that the lower-bound in Theorem \ref{lowani} equals $-\infty$ if the sequence $(B_n)_{n\in\mathbb{N}}$ is not assumed to be $\mu$-a.c. Consequently, for the previous result to give non trivial information one has to assume that $(B_n)_{n\in\mathbb{N}}$ is $\mu$-a.c. The question is then to give more explicit estimates of $s(\mu , \mathcal{B},\mathcal{U})$ depending on the specifities of $(\mu , \mathcal{B},\mathcal{U})$.
\mk
(2) It is proved in Section~\ref{sec-mino} that $s(\mu , \mathcal{B},\mathcal{U})\leq \overline{\dim}_H (\mu).$ This implies that for exact dimensional measures, $\min\left\{s(\mu , \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}=s(\mu , \mathcal{B},\mathcal{U}).$\mk
(3) The case where $\mu$ satisfies $\min\left\{s(\mu , \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}=0$ could also be treated, but although \eqref{conc2ani} is still obviously true, some distinction should further be made when investigating the existence of the gauge function. If $\mathcal{H}^{\mu,s}_{\infty}(U_n)=0$ for any $n\in\mathbb{N}$, the set $\limsup_{n\rightarrow+\infty} U_n$ could, for instance, be empty. On the other hand, if $(B_n)_{n\in\mathbb{N}}$ is $\mu$-a.c and $s_{\mu}(B_n ,U_n)>0$ for any $n\in\mathbb{N}$, a gauge function can be constructed in a similar way than in the proof of Theorem \ref{lowani}. However that the existence of such a gauge function in the case $\min\left\{s(\mu , \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}=0$, is of little interest for practical applications, and is not treated in this article.
\end{remarque}
A quite direct, but useful, corollary of Theorem \ref{lowani} is the following:
\begin{corollary}
\label{zzani}
Let $\mu\in\mathcal{M}(\R^d)$ and $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ be a $\mu$-a.c. sequence of closed balls of~$\R^d$.
Let $\mathcal{U}=(U_n)_{n\in\mathbb{N}}$ be a sequence of open sets such that
$U_n \subset B_n$ for all $n\in \mathbb N$, and $0\leq s\leq \dimm(\mu)$. If $ \limsup_{n\rightarrow+\infty}
\frac{\log\mathcal{H}^{\mu,s}_{\infty}(U_n )}{\log \mu( B_n )} \leq 1$, then
$s(\mu,\mathcal B,\mathcal U)\ge s$, so that
$$\dim_{H}(\limsup_{n\rightarrow +\infty}U_n)\geq s .
$$
\end{corollary}
In the classical case where the sets $U_n$ are shrunk balls of the form $B_n ^{\delta}$ (with $\delta\geq 1$), it is convenient to consider the following quantity:
\begin{definition}
\label{deftdelta}
Let $\mu \in\mathcal{M}(\R^d)$, $\varepsilon>0$ and $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ be a sequence of balls of $\R^d.$ For every $\delta \ge 1$, set
\begin{equation}
\label{tdeltainf}
t(\mu, \delta,\varepsilon ,\mathcal{B} )=\limsup_{n\rightarrow+\infty}\frac{\log(\mathcal{H}_{\infty}^{\mu,\dimm(\mu)-\varepsilon}(\widering{B}_n ^{\delta}))}{\log(\vert B_n ^{\delta} \vert)}.
\end{equation}
Then the $(\mu,\delta)$-exponent of the sequence $\mathcal{B} $ is defined as
\begin{equation}
\label{muexpo}
t(\mu,\delta,\mathcal{B} )=\lim_{\varepsilon\to 0}t(\mu,\delta,\varepsilon ,\mathcal{B}).
\end{equation}
\end{definition}
It follows from the definitions that $t(\mu,\delta,\mathcal{B} )$ exists as a limit, since $\ep\mapsto t(\mu, \delta,\varepsilon ,\mathcal{B} )$ is monotonic. Moreover, one has $\dimm(\mu)\le t(\mu,\delta,\mathcal{B})$ (see the proof of Corollary~\ref{minoeffec}).
Next result provides a more explicit lower bound estimate of the Hausdorff dimension of the limsup of $\delta$-contracted balls; it is a consequence of Corollary~\ref{zzani}.
\begin{corollary}
\label{minoeffec}
Let $\mu \in\mathcal{M}(\R^d)$ and $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ a $\mu$-a.c sequence of closed balls of~$\R^d$.
Suppose that $\dimm(\mu)>0$. For every $\delta\geq 1$, setting
$$
s_\delta= \frac{\dimm(\mu)}{\delta} \cdot \frac{\dimm(\mu)} {t(\mu,\delta,\mathcal{B})},
$$
one has
$$
s\big (\mu, (B_n)_{n\in\N},(\widering B_n ^{\delta})_{n\in\N}\big )\ge s_\delta,
$$
hence
$$\dim_H (\limsup_{n\rightarrow+\infty}B_n ^{\delta})\ge \dim_H (\limsup_{n\rightarrow+\infty}\widering B_n ^{\delta})\ge s_\delta.$$
\end{corollary}
\subsection{Application to self-similar measures}
Let us start by recalling the definition of a self-similar measure.
\begin{definition}
\label{def-ssmu} A self-similar IFS is a family $S=\left\{f_i\right\}_{i=1}^m$ of $m\ge 2$ contracting similarities of $\mathbb{R}^d$.
Let $(p_i)_{i=1,...,m}\in (0,1)^m$ be a positive probability vector, i.e. $p_1+\cdots +p_m=1$.
The self-similar measure $\mu$ associated with $ \left\{f_i\right\}_{i=1}^m$ and $(p_i)_{i=1}^m$ is the unique probability measure such that
\begin{equation}
\label{def-ssmu2}
\mu=\sum_{i=1}^m p_i \mu \circ f_i^{-1}.
\end{equation}
The topological support of $\mu$ is the attractor of $S$, that is the unique non-empty compact set $K\subset X$ such that $K=\bigcup_{i=1}^m f_i(K)$.
\end{definition}
The existence and uniqueness of $K$ and $\mu$ are standard results \cite{Hutchinson}. Recall that due to a result by Feng and Hu \cite{FH} any self-similar measure is exact dimensional.
The essential Hausdorff contents of a self-similar measure $\mu$ can be estimated quite precisely.
\begin{theoreme}
\label{contss}
Let $S$ be a self-similar IFS of $\R^d$. Let $K$ be the attractor of $S$. Let~$\mu$ be a self-similar measure associated with $S$. For any $0\leq s<\dim(\mu)$, there exists a constant $c=c(d,\mu,s)>0 $ depending on the dimension $d$, $\mu$ and $s$ only, such that for any ball $B=B(x,r)$ centered on $K$ and $r\leq 1$, any open set $\Omega$, one has
\begin{align}
\label{genhcontss}
&c(d,\mu,s)\vert B\vert ^{s}\leq\mathcal{H}^{\mu, s }_{\infty}(\widering{B})\leq \mathcal{H}^{ \mu, s}_{\infty}(B)\leq\vert B\vert ^{s}\text{ and } \nonumber\\
&c(d,\mu,s)\mathcal{H}^{s}_{\infty}(\Omega \cap K)\leq \mathcal{H}^{\mu,s}_{\infty}(\Omega)\leq \mathcal{H}^{s}_{\infty}(\Omega \cap K).
\end{align}
For any $s>\dim(\mu)$, $ \mathcal{H}^{\mu,s}_{\infty}(\Omega)=0.$
\end{theoreme}
\begin{remarque}
(1) The system $S$ is not assumed to verify any separation condition.\mk
(2) In the special case of the Lebesgue measure restricted to $[0,1]^d$ (or some cube $K$), \eqref{genhcontss} implies that for any $0 \leq s\leq d$, the Lebesgue-essential $s$ dimensional Hausdorff content is strongly equivalent to the usual $s$-dimensional Hausdorff content (it is even possible to take the constant $c(d,\mu,s)$ in \eqref{genhcontss}, independent of~$s$), so that Theorem \ref{contss} together with Theorem \ref{lowani} in the this special case implies the main theorem of Koivusalo and Rams, \cite[Theorem 3.2]{KR}, recalled below.
\end{remarque}
\begin{theoreme}[\cite{KR}]
\label{lowkr}
Let $(B_n)_{n\rightarrow+\infty}$ be a sequence of balls of $[0,1]^d$ verifying $\vert B_n \vert \to 0$ and $\mathcal{L}^d (\limsup_{n\rightarrow+\infty}B_n)=1$.
Let $(U_n)_{n\in \mathbb{N}}$ be a sequence of open sets satisfying $U_n \subset B_n$. For any $0\leq s \leq d$ such that, for all $n\in\mathbb{N}$ large enough, $\mathcal{H}^{s}_{\infty}(U_n)\geq \mathcal{L}^d(B_n)$, it holds that
\begin{equation}
\dim_H (\limsup_{n\rightarrow+\infty}U_n)\geq s.
\end{equation}
\end{theoreme}
As a consequence of Theorem \ref{contss} and Corollary \ref{zzani}, one gets
\begin{corollary}
\label{zzaniss}
Let $\mu\in\mathcal{M}(\R^d)$ be a self-similar measure and $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ be a $\mu$-a.c. sequence of closed balls of~$\R^d$ centered in $\supp (\mu)$.
Let $\mathcal{U}=(U_n)_{n\in\mathbb{N}}$ be a sequence of open sets such that
$U_n \subset B_n$ for all $n\in \mathbb N$, and $0\leq s\leq \dim(\mu)$. If, for $n\in\mathbb{N}$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\geq \mu(B_n) $, then
$$\dim_{H}(\limsup_{n\rightarrow +\infty}U_n)\geq s .
$$
\end{corollary}
One also emphasizes that in the case of a self-similar measure, conversely, any $s\geq 0$ such that $\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \mu(B_n)$ for every $n$ large enough is an upper-bound for $\dim_H (\limsup_{n\rightarrow+\infty}U_n)$ if $\mathcal{B}$ verifies that, for any $p \in\mathbb{N}$, the balls $B_n$ with $\vert B_n \vert \approx 2^{-p}$ does not overlap too much. More precisely, in the companion paper of the present article, \cite{ED2}, the following result is proved.
\begin{theoreme}[\cite{ED2}]
\label{majoss}
Let $\mu \in \mathcal{M}(\R^d)$ be a self-similar measure, $K$ its support and $(B_n)_{n\rightarrow+\infty}$ be a weakly redundant sequence of balls of $\R^d$ (see \cite[Definition 1.5]{BS2} ) verifying $\vert B_n \vert \to 0$ and, for any $n\in\mathbb{N}$, $B_n \cap K \neq \emptyset$. Let $(U_n)_{n\in \mathbb{N}}$ be a sequence of open sets satisfying $U_n \subset B_n$. For any $0\leq s <\dim (\mu)$ such that, for all large enough $n\in\mathbb{N}$, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \mu(B_n)$, it holds that
\begin{equation}\label{uppernound18}
\dim_H (\limsup_{n\rightarrow+\infty}U_n )\leq s.
\end{equation}
\end{theoreme}
Combining Theorem~\ref{lowani} and Corollary~\ref{minoeffec} with Theorem~\ref{contss} and Lemma~\ref{equiac} yield the following consequence for self-similar measures.
\begin{theoreme}
\label{prop-ss}
Let $S$ be a self-similar IFS of $\mathbb{R}^d$ with attractor $K$ and $\mu$ be a self-similar measure associated with $S$. Let $(B_{n})_{n\in\mathbb{N}}$ be a sequence of closed balls centered on $K$, such that $\lim_{n\to+\infty} |B_{n}| = 0$.
\begin{enumerate}
\item Suppose that $(B_n)_{n\in\N}$ is $\mu $-a.c. Then $
t\big (\mu,\delta ,(B_n)_{n\in\mathbb{N}}\big )\le \dim(\mu)$; consequently $s\big (\mu, (B_n)_{n\in\N},(\widering B_n ^{\delta})_{n\in\N}\big ) \ge \frac{\dim(\mu)}{\delta}$ and there exists a gauge function~$\zeta$ such that $\lim_{r\to 0^+}\frac{\log(\zeta(r))}{\log(r)}\ge \frac{\dim(\mu)}{\delta}$ and $\mathcal H^\zeta(\limsup_{n\to\infty} \widering B_n ^{\delta}) >0$. In particular
\begin{equation}\label{weakversion}
\dim_{H}(\limsup_{n\rightarrow+\infty}\widering B_n ^{\delta})\geq \frac{\dim(\mu)}{\delta} .
\end{equation}
\item Suppose that $\mu(\limsup_{n\rightarrow+\infty}B_n)=1$. Then, \eqref{weakversion} still holds but the existence of the gauge function is not ensured. Furthermore if $\mu$ is doubling, then $(B_n)_{n\in\N}$ is $\mu$-a.c, so that the conclusion of item (1) holds.
\end{enumerate}
\end{theoreme}
\begin{remarque}
Since no separation condition is assumed about the system $S$, Theorem \ref{prop-ss} implies \cite[Theorem 1.6]{BS2} in the special case where the sequence of measures $(\mu_p )_{p\in\mathbb{N}}$ is constant and equal to some self-similar measure with open set condition $\mu$ and the sequence of contraction ratio $(\delta_p)_{p\in\mathbb{N}}$ is constant as well.
\end{remarque}
Corollary~\ref{zzani} and Theorem~\ref{contss} also make it possible to deal with more general open sets $(U_n)_{n\ge 1}$ than the contracted balls $(\widering B_n^\delta)$, if one is able to compare efficiently the $s$-dimensional Hausdorff contents of the sets $U_n\cap K$ with a power of $|B_n|$. It is then convenient to assume that $K$ is the closure of its interior. Here is an example.
Let $1\leq \tau_1 \leq ...\leq \tau_d$ be $d$ real numbers and $\tau=(\tau_1,...,\tau_d)$. One starts by defining a family of rectangles of $\mathbb{R}^d$ associated with $\tau$.
\begin{definition}
Let $1\leq \tau_1 \leq ... \leq \tau_{d} $ and $\tau=(\tau_1,...,\tau_d)$. For any $x=(x_i)_{1\leq i\leq d}\in \R^d$ and $r>0$, the $\tau$-rectangle centered in $x$ and associated with $r$ is defined by
\begin{equation}
R_{\tau}(x,r)=\prod_{i=1}^d [x_i-\frac{1}{2}r^{\tau_i},x_i+\frac{1}{2}r^{\tau_i}].
\end{equation}
\end{definition}
\begin{theoreme}
\label{rectss}
Let $S$ be a self-similar IFS of $ \R^d$ such that the attractor $K$ is equal to the closure of its interior. Let $\mu$ be a self-similar measure associated with~$S$. Let $1\leq\tau_1\leq...\leq \tau_d$, $\tau=(\tau_1, ...,\tau_d)$ and $(B_n :=B(x_n,r_n))_{n\in\mathbb{N}}$ be a sequence of balls of $\R^d$ satisfying $r_n \to 0$ and $\mu(\limsup_{n\rightarrow+\infty}B_n)=1.$ Define $R_n =\widering R_{\tau}(x_n,r_n).$ Then
\begin{equation}
\dim_H (\limsup_{n\rightarrow+\infty}R_n)\geq \min_{1\leq i\leq d }\left\{\frac{\dim(\mu)+\sum_{1\leq j\leq i}\tau_i-\tau_j}{\tau_i}\right\}.
\end{equation}
\end{theoreme}
\begin{remarque}
\label{remani}
(1) Since $(\tau_1 ,... ,\tau_d)\mapsto \min_{1\leq i\leq d }\left\{\frac{\dim(\mu)+\sum_{1\leq j\leq i}\tau_i-\tau_j}{\tau_i}\right\}$ is continuous, the result stands for the sequence of closed rectangles as well.
(2) One may also apply any rotation to the shrunk rectangles, this wouldn't change the bound (since Hausdorff contents are invariant by rotation).\mk
(3) Theorem \ref{rectss} extends the results of \cite{ED1}, where the measure was quasi-Bernoulli or verifying the open set condition, and supported on $[0,1]^d$.\mk
(4) When $K$ is the closure of its interior is that it is easy to compute $\mathcal{H}^s _{\infty}(R_n \cap K).$ Without this assumption, the conclusion of Theorem \ref{rectss} fails. Indeed, in general, no formula involving only the dimension of the measure and the contraction ratio can be accurate. For instance, consider a self-similar measure in $\mathbb{R}^2$ carried by a line $D$ and a sequence $(B_n)_{n\in\mathbb{N}}$ of balls centered on the attractor $K$ and verifying $\mu(\limsup_{n\rightarrow+\infty}B_n)=1.$ Then, consider the sequence of rectangles $R_n$ with side-length $\vert B_n \vert^{\tau_1}\times \vert B_n \vert^{\tau_2}$, $1\leq \tau_1 \leq \tau_2$ and where the largest side (of side-length $\vert B_n \vert^{\tau_1}$) is in the direction of $D$. In this case, Theorem~\ref{prop-ss} yields the lower-bound $\dim_H (\limsup_{n\rightarrow+\infty}R_n)\geq \frac{\dim_H (\mu)}{\tau_1}$. Then if $\widehat{R}_n$ are the rectangles $R_n$ rotated by $\frac{\pi}{2}$, Theorem~\ref{prop-ss} gives that $\dim_H (\limsup_{n\rightarrow+\infty}\widehat{R}_n)\geq \frac{\dim_H (\mu)}{\tau_2}$. Moreover, under additional conditions, these lower bounds are equalities.
\end{remarque}
\subsection{Application to self-similar shrinking targets}
We deal with points well approximable by orbits under an IFS with no exact overlap and satisfying the condition introduced by Barral and Feng (\cite{BFmult}) of being \textit{dimension regular} with similarity dimension less than than $d$, after Hochman's work (\cite{Hoc}).
\label{sec-notaself}
\begin{definition}
\label{dimreg}
Let $S=\left\{f_1,...,f_m\right\}$ be a self-similar IFS of $\mathbb{R}^d$. Denote by $0<c_1,\ldots, c_m<1$ the contraction ratio of $f_1 ,\ldots,f_m$. The system~$S$ is said to be dimension regular if, for any probability vector $(p_1,...,p_m)$, the self-similar measure associated with $S$ and the probability vector $(p_1,...,p_m)$ verifies $$\dim (\mu)=\min\left\{d,\frac{\sum_{1\leq i\leq m}p_i \log (p_i)}{\sum_{1\leq i\leq m} p_i \log(c_i)}\right\}.$$
This in particular, implies that, denoting by $\dim_{sim} (K)$ the unique real number $s$ satisfying $\sum_{i=1}^m c_i ^{s}=1,$ one has $\dim_H (K)=\min\left\{\dim_{sim}(K),d\right\}.$
\end{definition}
Some notation useful when dealing with IFS are introduced now. Those notations will be used repeatedly throughout this article.
Let $S=\left\{f_1 ,...,f_m\right\}$ be a self-similar IFS, $0<c_1\ldots, c_m <1$ the associated contraction ratios, and $K$ the attractor of $S$. Let $(p_1 ,...,p_m)$ be a probability vector with positive entries, $\mu$ the self-similar associated with $S$ and $(p_1 ,...,p_m)$.
Let $\Lambda=\left\{1,...,m\right\}$ and $\Lambda^* =\bigcup_{k\geq 0} \Lambda ^{k}$.
For $k\geq 0$ and $\underline{i}:=(i_1 ,...,i_k)\in\Lambda^k$, define
\begin{equation*}
\begin{split}
c_{\underline{i}}&=c_{i_1}\cdots c_{i_k},\ f_{\underline{i}}=f_{i_1}\circ \cdots\circ f_{i_k}, \\
\Lambda ^{(k)}&=\left\{\underline{i}=(i_1 ,\ldots, i_s) \in\Lambda^* : \ c_{i_s}2^{-k}< c_{\underline{i}}\leq 2^{-k} \right\}.
\end{split}
\end{equation*}
\begin{theoreme}
\label{shrtag}
Let $S=\left\{f_1 ,...,f_m\right\}$ be a dimension regular self-similar IFS with contraction ratio $0<c_1,\ldots, c_m <1$ and such that the attractor $K$ verifies $\dim_{sim}(K)=\dim_H (K)$. For any $x\in K$, for any $\delta \geq 1$,
\begin{equation}
\dim_{H}(\limsup_{\underline{i}\in \Lambda^*}B(f_{\underline{i}}(x), c_{\underline{i}}^{\delta}))=\frac{\dim_H (K)}{\delta}.
\end{equation}
\end{theoreme}
This result extends some of the results obtained in \cite{AllenB} and \cite{Baker}, under the open set condition.
\subsection{A result motivated by a question of Mahler}$\ $
\label{sec-mahlint}
Let $\mathcal{Q}=\left\{B(\frac{p}{q},q^{-2})\right\}_{q\in\mathbb{N}^* ,\, 0\leq p\leq q}.$ Recall the following result in Diophantine approximation \cite{Jarnik}:
\begin{align}
\label{approxrat}
&\bullet \limsup_{B\in\mathcal{Q}}B=[0,1]. \nonumber\\
&\bullet \text{ For any }\delta \geq 1, \dim_H (\limsup_{B\in\mathcal{Q}}B^{\delta})=\frac{1}{\delta}.
\end{align}
Unlike in the case of the points in $[0,1]$, the approximation by rational numbers of elements of the middle third Cantor set $K_{1/3}$ set is not well understood yet. This question was raised by Mahler, and only some partial results are known (see \cite{BV},~\cite{BS2}). Here we consider the set $K_{1/3}^{(0)}$ of points in $[0,1]$ having an asymptotic lower frequency of appearance of the digit $1$ in basis 3 equal to $0$. This set contains $K_{1/3}$ and has the same Hausdorff dimension as $K_{1/3}$. We compute the Hausdorff dimension of sets of points in $K_{1/3}^{(0)}$ which are well approximable by rational numbers.
To describe more precisely the problem, let $S=\left\{f_1 ,f_2,f_3\right\}$ where $f_1$, $f_2$ and $f_3$ are the contracting affine maps of $\mathbb R$ defined by $f_0 (x)=\frac{1}{3}x$, $f_1 (x)=\frac{1}{3}x+\frac{1}{3}$ and $f_2 (x)=\frac{1}{3}x+\frac{2}{3}$. The attractor of $S$ is $[0,1]$. Let $\Lambda=\left\{0,1,2\right\}$.
The shift operation on the symbolic space $\Lambda^\mathbb{N}$ is defined by $\sigma$. The canonical projection from $\Lambda^\mathbb{N}$ to $[0,1]$ is the mapping
\begin{equation}
\label{canoproj}
\pi:x=(x_n)_{n\in\mathbb{N}}\mapsto \lim_{n\to +\infty} f_{(x_1 ,...,x_n )}(0).
\end{equation}
The set $K_{1/3}$ is the attractor of $\{f_0,f_2\}$ and also the image by canonical projection of $\{0,2\}^\mathbb{N}$.
\begin{definition}
\label{defphi}
Let $\phi:\Lambda^\mathbb{N} \to \left\{0,1\right\}$ defined by $\begin{cases}\phi(x)=1 \text{ if } x_1 =1 \\
\phi(x)=0 \text{ if } x_1 =0 \text{ or }2.\end{cases}$
and
\begin{align}
\label{defK}
K_{1/3}^{(0)}=\pi\left(\left\{x\in \Lambda^\mathbb{N}: \liminf_{k\to +\infty}\frac{S_k\phi(x)}{k}=0 \right\}\right )\nonumber,
\end{align}
where $(S_k\phi)_{k\in\mathbb N}$ stands for the sequence of Birkhoff sums of $\phi$.
\end{definition}
It is proved in \cite{FanFengWu} that $\dim_H K_{1/3}^{(0)}=\frac{\log 2}{\log 3}(=\dim_H K_{1/3})$.
Let us state the main results of this subsection.
\begin{theoreme}
\label{main}
For every $\delta\geq 1,$
\begin{equation}
\label{mainequa}
\dim_H\Big (\limsup_{B\in\mathcal{Q}}B^{\delta}\cap K_{1/3}^{(0)}\Big )=\min\left\{\frac{\log 2}{\log 3},\frac{1}{\delta}\right\}.
\end{equation}
\end{theoreme}
Observe that a saturation phenomenon occurs : $\dim_H(\limsup_{B\in\mathcal{Q}}B^{\delta}\cap K_{1/3}^{(0)})=\frac{\log 2}{\log 3}$ for $1\leq \delta \leq \frac{\log 3}{\log 2}$.
\bigskip
In Section \ref{sec-low}, the general ubiquity theorem, Theorem \ref{lowani}, is proved as well as Corollary \ref{minoeffec}. Section \ref{sec-parti} gives estimations of essential contents in the self-similar case and Theorem \ref{contss} is proved.
Section \ref{sec-example} gives three applications to the main result, Theorem \ref{lowani}. More precisely, the ubiquity theorems for self-similar measures, Theorem \ref{prop-ss} and Theorem \ref{rectss}, are proved in the first sub-section. The second sub-section treats the case of self-similar shrinking targets for dimension regular IFS with similarity dimension less than $d$, e.g, Theorem \ref{shrtag} is proved. In the last sub-section, one gives an application in Diophantine approximation, Theorem \ref{main} is proved.
\section{Proof of Theorem \ref{lowani}}
\label{sec-low}
\subsection{Preliminary facts}
We gather in this subsection a series of results on which we will base the proof of Theorem \ref{lowani}.
The following lemma, which is a version of Besicovitch covering Lemma, as well as the subsequent one, both established in \cite{ED2}, will be used several times.
\begin{lemme}
\label{besimodi}
For any $0<v\leq 1$ there exists $Q_{d,v} \in\mathbb{N}^{\star}$, a constant depending only on the dimension $d$ and $v$, such that for every bounded subset $E\subset \R^{d}$, for every set $\mathcal{F}=\left\{B(x, r_{(x)} ): x\in E, r_{(x)} >0 \right\}$, there exists $\mathcal{F}_1,...,\mathcal{F}_{Q_{d,v}}$ finite or countable sub-families of $\mathcal{F}$ such that:\medskip
\begin{itemize}
\item
$\forall 1\leq i\leq Q_{d,v}$, $L\neq L'\in\mathcal{F}_i$, one has $\frac{1}{v}L \cap \frac{1}{v}L'=\emptyset.$\medskip
\item
$E$ is covered by the families $\mathcal{F}_i$, i.e.
\begin{equation}\label{besi}
E\subset \bigcup_{1\leq i\leq Q_{d,v}}\bigcup_{L\in \mathcal{F}_i}L.
\end{equation}
\end{itemize}
\end{lemme}
\begin{lemme}[\cite{ED2}] \label{dimconst}
Let $0<v\leq 1$, $\mathcal{B} =(B_n )_{n\in\mathbb{N}}$ a family of balls, and $B$ a ball such that
\medskip
\begin{itemize}
\item [(i)]
$\forall \ n\geq 1$, $\vert B_n\vert \geq \frac{1}{2}\vert B\vert,$
\item [(ii)]
$\forall \ n_1 \neq n_2 \geq 1 $, $vB_{n_1}\cap vB_{n_2}=\emptyset$.
\end{itemize}
Then $B$ intersects less than $Q_{d,v}$ elements of $\mathcal{B}$, where $Q_{d,v}$ can be taken equal to the constant considered in Lemma~\ref{besimodi}.
\end{lemme}
The following lemma will also be useful later on and is also proved in \cite{ED2}.
\begin{lemme}
\label{distcov}
Let $\mathcal{L}$ be a family of pairwise disjoint balls satisfying $\sup_{L\in\mathcal{L}}\vert L\vert <+\infty.$ Then, for any $v\geq 1$, there exists sub-families $\mathcal{L}_1 ,...,\mathcal{L}_{Q_{d,v}}$ (where $Q_{d,v}$ is the constant of the same name in Lemma \ref{besimodi}) of $\mathcal{L}$ such that $\mathcal{L}=\bigcup_{1\leq i\leq Q_{d,v}}\mathcal{F}_i$ and for any $L\neq L^{\prime}\in\mathcal{L}_i$, $vL \cap vL^{\prime}=\emptyset.$
\end{lemme}
Recall the following version of Frostman Lemma, due to Carleson.
\begin{proposition}[\cite{Ca}]
\label{carl}
Let $s\geq 0$. There is a constant $\kappa_d >0$ depending only on the dimension $d$ such that for any bounded set $E\subset \R^d$ with $\mathcal{H}^{s}_{\infty}(E)>0$, there exists a probability measure supported by $E$, that we denote by $m^{s}_{E}$, such that
\begin{equation}
\label{majcarl}
\mbox{for every ball $ B(x,r)$,
} \ \ \ \ m_{E}^{s}(B(x,r))\leq \kappa_d \frac{r^s}{\mathcal{H}^{s}_{\infty}(E)}.
\end{equation}
\end{proposition}
For $s\geq 0$ and $E\subset \R^d$, a bounded subset such that $\mathcal{H}^{s}_{\infty}(E)>0$, $m^{s}_{E}$ will always denote such a measure associated with a (fixed) constant $\kappa_d$.
\sk
In the next two lemmas, the choice of the interval $[5,6]$ is convenient to take enough space between the shrunk balls involved in the construction elaborated in Section~\ref{construction}.
\begin{lemme}
\label{doub2}
Let $t\in (5,6)$, $m\in\mathcal M(\R^d)$, and $\ep >0$. Let $x\in \R^d$ be such that $\overline {\dim}_{\mathrm{loc}}(m, x) \leq \beta$. Let $C_{\beta,\varepsilon}= \frac{1}{2}6^{-\frac{\beta}{ 2 \ep }}$.
There exists an integer $n_{x}$ such that for every $ n\geq n_{x}$,
\begin{equation}
\label{eq55}
\frac{\# \left\{0\leq k\leq n-1 : \ m(B(x,t^{-k-1}))\geq C_{\beta,\varepsilon} m(B(x,t^{-k}))\right\}}{n} \geq 1- \ep.
\end{equation}
\end{lemme}
Previous lemma is a slight extension of result by K\"aenm\"aki \cite[Lemma 2.2]{C}, which shows such a property at $m$-almost every point (where one has necessarily $\overline {\dim}_{\mathrm{loc}}(m, x) \leq d)$, and uses $t$ integer (a choice that we could make).
Thus, points with a given local dimension with respect to a measure $m$ are for most scales ``locally doubling''.
\begin{proof}
Observe first that if for a constant $0<C\leq 1$ and some integer $n\in\mathbb{N}$ one has
$$\frac{\#\left\{1\leq k\leq n : \ m(B(x,t^{-k-1}))\geq Cm(B(x,t^{-k}))\right\}}{n} \leq 1-\varepsilon,$$
then there necessarily exist $N = \lfloor (n-1)\ep \rfloor $ integers $0 <k_1 <\cdots < W_{n}<n $ such that
for every $ 1\leq i\leq N$, $m(B(x,t^{-k_i-1}))\leq Cm(B(x,t^{-k_i}))$.
In particular, writing $k_{N+1}=n$ and $k_0 =0$, this implies that
\begin{align*}
m(B(x,t^{-n}))& =\prod_{i=0}^{N} \frac{m(B(x,t^{-k_{i+1} }))}{m(B(x,t^{-k_i}))}\leq \prod_{i=0}^{N} \frac{m(B(x,t^{-k_{i}-1 }))}{m(B(x,t^{-k_i}))}\leq C^{N} \\
& \leq C^{(n-1)\ep } \leq C^{ n \ep/2 } = (t^{-n})^{\ep \frac{-\log(C)}{2\log(t)}}.
\end{align*}
The inequality $C^{(n-1)\ep } \leq C^{ n \ep/2 } $ occurs when $n$ is large enough. Recalling that $\overline{\dim}_{\mathrm{loc}}(m,x)\leq\beta$, if this happens for infinitely many $n$, one should have
$$\beta\geq \limsup_{r\rightarrow 0^+}\frac{\log m(B(x,r))}{\log r}\geq \ep\frac{-\log(C)}{2\log(t)},$$
which is equivalent to $C\ge t^{-\frac{\beta}{2 \ep }}$.
Setting $C_{ \ep ,\beta}=\frac{1}{2}6^{-\frac{\beta}{ 2 \ep }}$, one concludes that there exists $n_{x}$ such that for every $ n\geq n_{x}$, one necessarily has
$$\frac{\#\left\{0\leq k\leq n-1 : \ m(B(x,t^{-k-1}))\geq C_{ \ep ,\beta } m(B(x,t^{-k}))\right\}}{n} \geq 1- \ep ,$$
hence the result.
\end{proof}
\begin{lemme}
\label{doub}
Let $m$ and $\mu$ be two elements of $\mathcal M(\zu)$, $\beta \geq 0$ and $\varepsilon >0$. For every $x\in \R^d$ verifying $\overline {\dim}_{\mathrm{loc}}(m, x)\leq \beta$, there exists $\rho_x >0$ and $t_x\in (5,6)$ so that for all $0<r\leq \rho_x$ there exists $r\leq r'\leq r^{1-\varepsilon }$ such that
\begin{equation}
\label{eqdoub}
m(B(x,r'/t_x))\geq C_{\beta,\frac{\ep}{2}}m(B(x,r')) \mbox{ and } \mu(\partial B(x,r'/t_x))=0.
\end{equation}
\end{lemme}
\begin{proof}
Consider $x\in \R^d$ such that $\overline{\dim}_{\mathrm{loc}}(m,x)\leq\beta.$
We apply Lemma \ref{doub2} to $x$ and the measure $m$, and for an arbitrary $t\in [5,6]$ and $\varepsilon^{\prime}=\frac{\varepsilon}{2}$: for $n\geq n_{x}$, there must be an integer $n'$ such that $n(1-\ep) \leq n'\leq n $ and $m(B(x,t^{-n'-1}))\geq C_{\beta,\frac{\varepsilon}{2}}m(B(x,t^{-n'})).$
Let $\rho_x=\min\left\{t^{-n_x-1 },t^{-\frac{1}{\varepsilon}}\right\}$. For $r\in(0,\rho_x]$, let $n$ be the integer such that $t^{-n-1}<r\leq t^{-n}$. The previous claim yields an integer $n'\in [n(1-\frac{\ep}{2}),n] $ such that $m(B(x,t^{-n'}))\geq C_{\beta,\frac{\ep}{2}} m(B(x,t^{-n'+1}))$. Also,
$$ r \leq r'=t^{-n'+1 }\leq t^{1-(1-\frac{\ep}{2} )n}= t^2 \cdot t^{-n-1}\cdot t^{ \frac{\ep}{2} n}\leq t^2 \cdot r \cdot r^{-\frac{\ep}{2} }\leq r^{1- \ep }.$$
Consequently,
$$
m(B(x,r'/t))\geq C_{\beta,\frac{\ep}{2}}m(B(x,r')).
$$
The desired conclusion holds if we choose $t_x\in (5,t)$ such that $\mu(\partial B(x,r'/t_x))=0$.
\end{proof}
The previous lemma will be used in the case $\beta=d$ in our proof the main theorem (see step 2 of the construction in Section~\ref{construction}).
\mk
Next, we introduce some some sets associated to a given element of $\mathcal M(\R^d)$, which will play a natural role in our construction.
\begin{definition}
\label{emudef}
{Let $\beta\geq\alpha\geq 0$ be real numbers, $m\in\mathcal{M}(\R^d)$, and $\varepsilon,\rho>0$ two positive real numbers. Then define }
\begin{equation}
\label{emuhat}
{\widetilde{E}_{m}^{[\alpha,\beta],\rho,\varepsilon}=\left\{x\in\mathbb{R}^d : \ \underline\dim_{{\rm loc}}(m,x)\in[\alpha,\beta] \ and \ \ \forall r\leq \rho, \ m(B(x,r))\leq r^{\underline{\dim}_{\mathrm{loc}} (m,x)-\varepsilon}\right\}}
\end{equation}
and
\begin{equation}
\label{emu}
{E_{m}^{[\alpha,\beta],\rho,\varepsilon}=\left\{ x\in\widetilde{E}_{m}^{[\alpha,\beta],\rho,\varepsilon}: \forall r\leq \rho, \ \frac{3}{4}m(B(x,r)) \leq m(B(x,r)\cap \widetilde{E}_{m}^{[\alpha,\beta],\rho,\varepsilon})\right\}}
\end{equation}
\end{definition}
{Notice that, for every $0 <\rho <\rho '$, one has $ E_{m}^{[\alpha,\beta],\rho',\varepsilon }\subset E_{m }^{[\alpha,\beta],\rho,\varepsilon }.$}
\begin{definition}
\label{emudef2}
{Let $\beta \geq\alpha\geq 0$ be real numbers, $m\in\mathcal{M}(\R^d)$, and $\varepsilon>0$. Define}
\begin{equation}
\label{emut}
{E_{m}^{[\alpha,\beta],\varepsilon}=\bigcup_{n\geq 1}E_{m}^{[\alpha,\beta],\frac{1}{n},\varepsilon}} .
\end{equation}
\end{definition}
\begin{proposition}
\label{mle}
{For every $m\in\mathcal{M}(\R^d)$, every $\beta\geq \alpha\geq 0$ and $\varepsilon>0$,}
\begin{equation}
{m(E_{m}^{[\alpha,\beta],\varepsilon})=m(\left\{x:\underline{\dim}_{\mathrm{loc}}(m,x)\in [\alpha,\beta]\right\})}.
\end{equation}
\end{proposition}
{Notice that, for every $0<\rho ' <\rho$, one has $E_{m }^{[\alpha,\beta],\rho,\varepsilon }\subset E_{m}^{[\alpha,\beta],\rho',\varepsilon }.$}
{ These sets play a key role in the proofs of Theorem \ref{lowani} .}
\begin{proof}
{One first recalls the following result. }
\begin{lemme}\cite{Be}
\label{densibesi}
{Let $m\in\mathcal{M}(\R^d)$ and $A$ be a Borel set with $m(A)>0.$ For every $r>0$, set}
\begin{equation}
\label{defniv}
{A(r) =\left\{x\in A \ : \ \forall \tilde{r}\leq r, \ m(B(x,\tilde{r})\cap A)\geq \frac{3}{4}m(B(x,\tilde{r}))\right\}}
\end{equation}
{Then }
\begin{equation}
{ m\left(\bigcup_{r>0}A(r)\right)=m(A).}
\end{equation}
\end{lemme}
Note that it is clear from Definition \ref{dim} that
$$\left\{x:\underline{\dim}_{\mathrm{loc}}(m,x)\in [\alpha,\beta]\right\}=\bigcup_{\rho>0} \widetilde{E}_{m}^{[\alpha,\beta],\rho,\varepsilon}.$$
{ Let $\ep'>0$. By Definition \ref{dim}, there exists $\rho_{\ep'}$ small enough so that}
\begin{equation}
\label{eqvo}
{m(\widetilde{E}_{m}^{[\alpha,\beta],\rho_{\ep'}, \varepsilon})\geq (1-\ep')m(\left\{x:\underline{\dim}_{\mathrm{loc}}(m,x)\in [\alpha,\beta]\right\}).}
\end{equation}
{By Lemma \ref{densibesi} (and the notations therein) applied to $\widetilde{E}_{m}^{[\alpha,\beta],\rho_{\ep'}, \varepsilon}$, there exists $\tilde{\rho}_{\ep'}$ such that}
\begin{equation}
\label{eqvo2}
{ m(\widetilde{E}_{m}^{[\alpha,\beta],\rho_{\ep'}, \varepsilon}(\tilde{\rho}_{\ep'}))\geq (1-\ep')m(\widetilde{E}_{m}^{[\alpha,\beta],\rho_{\ep'}, \varepsilon}) .}
\end{equation}
{Finally for $\rho =\min\left\{\rho_{\ep'},\tilde{\rho}_{\ep'}\right\}$, by Definition \ref{emudef} and \eqref{defniv}, one has $ (\widetilde{E}_{m}^{[\alpha,\beta],\rho_{\ep'}, \varepsilon})_{\tilde{\rho}_{\ep'}}\subset E_{m}^{[\alpha,\beta],\rho, \varepsilon}$, so that, by \eqref{eqvo} and \eqref{eqvo2}}
\begin{align*}
m(E_{m}^{[\alpha,\beta],\rho, \varepsilon})
&\geq m((\widetilde{E}_{m}^{[\alpha,\beta],\rho_{\ep'}, \varepsilon}(\tilde{\rho}_{\ep'}))
\geq (1-\ep')m(E_{m}^{[\alpha,\beta],\varepsilon})\\
&\geq (1-\varepsilon^{\prime})^2 m(\left\{x:\underline{\dim}_{\mathrm{loc}}(m,x)\in [\alpha,\beta]\right\}).
\end{align*}
In particular
$$m(\left\{x:\underline{\dim}_{\mathrm{loc}}(m,x)\in [\alpha,\beta]\right\})\geq m(E_{m}^{[\alpha,\beta],\varepsilon}) \geq (1-\varepsilon^{\prime})^2 m(\left\{x:\underline{\dim}_{\mathrm{loc}}(m,x)\in [\alpha,\beta]\right\}).$$
{Letting $\ep' \to 0$ proves the result.}
\end{proof}
\begin{corollary}
\label{mlecoro}
For every $m\in\mathcal{M}(\R^d)$, for $\alpha=\underline{\dim}_H (m)$ and $\beta=\overline{\dim}_H (m)$, for any $\varepsilon>0$, one has
\begin{equation}
m(E_{m}^{[\alpha,\beta],\varepsilon})=1.
\end{equation}
\end{corollary}
\subsection{Construction of the Cantor set and the measure}\label{construction}
Recall that $\mu$ is a probability measure on $\R^d$, and that $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ is a $\mu$-a.c sequence of balls of $ \R^d$ with $\lim_{n\to +\infty} r_{n}= 0.$ Fix $\mathcal{U}=(U_n)_{n\in\mathbb{N}}$ a sequence of open sets satisfying $U_n \subset B_n$ for every $n\in\mathbb{N}$.
Set $\alpha=\dimm(\mu)$, and assume that $\min\left\{ s(\mu,\mathcal{B},\mathcal{U}),\alpha\right\}>0$.
Our goal is to construct a gauge function $\zeta:\R^+\to \R^+$ such that $\lim_{r\to 0^+} \frac{\log \zeta(r)}{\log r} =\min\left\{ s(\mu, \mathcal{B},\mathcal{U}),\dimm(\mu)\right\}$ as well as $\eta\in \mathcal M(\R^d)$ supported on $\limsup_{n\to\infty} U_n$ such that for all $r\in(0,1]$ and $x\in\R^d$ one has $\eta(B(x,r))\le \zeta(2r)$.
Let $(\varepsilon_k)_{k\in\mathbb{N}}$ be a sequence decreasing to 0 and such that $\varepsilon_1<s(\mu,\mathcal{B},\mathcal{U})$. For $k\geq 0$, set
\begin{equation}
\label{defskani}
s_k=\min\left\{ s(\mu,\mathcal{B},\mathcal{U}),\alpha\right\}-\varepsilon_{k}.
\end{equation}
Along the construction of $\zeta$, we only use that $s_k < s(\mu,\mathcal{B},\mathcal{U})$ and the fact that $s_k < \alpha$ is used at the end of our analysis (see equation \eqref{res3ani}).
\subsubsection*{Step 1} We need the following lemma.
\begin{lemme} [\cite{ED2}]
\label{covO}
Let $\mu\in\mathcal{M}(\R^d)$ and $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ be a $\mu$-a.c sequence of balls of $ \R^d$ with $\lim_{n\to +\infty} r_{n}= 0$.
Then for every open set $\Omega$ and every integer $g\in\mathbb{N}$, there exists a subsequence $(B_{(n)}^{(\Omega)})\subset \left\{B_n \right\}_{n\geq g} $ such that:
\begin{enumerate}
\item
$\forall \, n\in\mathbb{N}$, $B_{(n)}^{(\Omega)}\subset \Omega,$
\sk
\item
$\forall \, 1\leq n_1\neq n_2$, $B_{(n_1)}^{(\Omega)}\cap B_{(n_2)}^{(\Omega)}=\emptyset$,
\sk
\item
$\mu\left (\bigcup_{n\geq 1}B_{(n)}^{(\Omega)}\right)= \mu(\Omega).$
\end{enumerate}
In addition, there exists an integer $N_{\Omega}$ such that for the balls $ (B_{(n)}^{(\Omega)})_{n=1,...,N_\Omega}$, the conditions (1) and (2) are realized, and (3) is replaced by $\mu\left (\bigcup_{n=1}^{N_\Omega} B_{(n)}^{(\Omega)}\right)\geq \frac{3}{4} \mu(\Omega).$
\end{lemme}
The last part of Lemma \ref{covO} simply follows from item (3) and the $\sigma$-additivity of $\mu$.
Using Lemma \ref{covO} with, $(B_n)_{n\in \mathcal{N}_\mu(\mathcal{B},\mathcal U,s_1)}$ (which is $\mu$-a.c since $ s_1 <s(\mu,\mathcal{B},\mathcal{U})$), $g=0$ and $\Omega = \R^d$, one finds integers $N_1 $ and $n_1 < ... < n_{N_1}\in\mathcal{N}_\mu(\mathcal{B},\mathcal U,s_1)$ such that :
\begin{itemize}
\item[$(i):$] $\forall \, 1\leq i\leq N_1$, $B_{n_i}\cap B_{n_j}=\emptyset$,
\item[$(ii):$] $\mu(\bigcup_{1\leq i\leq N_1}B_{n_i})\geq \frac{1}{2}$.
\end{itemize}
By Lemma \ref{distcov} applied to $\left\{B_{n_i}\right\}_{1\leq i\leq N_1}$ and $v=4$, the balls $\left\{B_{n_i}\right\}_{1\leq i\leq N_1}$ can be sorted in $Q_{d,4}$ families of balls $\mathcal{L}_1 ,...,\mathcal{L}_{Q_{d,4}}$ such that
\begin{itemize}
\item[•] for any $1\leq i\leq Q_{d,4}$, any $L\neq L^{\prime}\in\mathcal{L}_i$, $4L \cap 4L^{\prime}=\emptyset, $\mk
\item[•] $\bigcup_{1\leq i\leq Q_{d,4}}\mathcal{L}_i=\left\{B_{n_i}\right\}_{1\leq 1\leq N_1}.$
\end{itemize}
At least one of these families, $\mathcal{L}_{i_0 }$, must satisfy
$$
\mu\big(\bigcup_{L\in\mathcal{L}_{i_0}}L\big)\geq \frac{1}{2Q_{d,4}}.
$$
In particular, if one must rename the balls of the family $\mathcal{L}_{i_0}$, we can assume that the family $\left\{B_{n_i}\right\}_{1\leq i\leq N_1}$ satisfies
\begin{itemize}
\item[$(i^{'}):$] for any $1\leq i< j\leq N_1$, $4 B_{n_i}\cap 4B_{n_j}=\emptyset$\mk
\item[$(ii^{\prime}):$] and
\begin{equation}
\label{recouetape1}
\mu\big(\bigcup_{1\leq i\leq N_1}B_{n_i}\big)\geq \frac{1}{2Q_{d,4}}.
\end{equation}
\end{itemize}
Set $$\mathcal{W}_{1}=\left\{U_{n_i} \right\}_{1\leq i\leq N_1} \ \ \mbox{ and } \ \ W_{1}=\bigcup_{1\leq i\leq N_1}U_{n_i} .$$
Along the construction of the Cantor set, for every $U \in\mathcal{U}$, the ball of $\mathcal{B}$ naturally associated with $U$ will be denoted $B^{[U]}$ (that is $B^{[U_n]}=B_n$).
The pre-measure $\eta$ on the $\sigma$-algebra generated by the sets of $\mathcal{W}_{1} $ is defined by
\begin{equation}
\label{defeta0}
\mbox{ for every $U \in \mathcal{W}_1$, } \ \ \ \ \eta(U )=\frac{\mu(B^{[U]})}{\sum_{\widetilde{ U} \in \mathcal{W}_1}\mu(B^{[\widetilde{U}]})} .
\end{equation}
It is obvious that $\eta(\R^d) = \eta (W_{1}) =1$.
Recalling \eqref{defNsdani} and \eqref{defsmdani}, since $s_1<s(\mu , \mathcal{B},\mathcal{U})$, the sub-sequence $(B_n )_{n\in\mathcal{N}_{\mu}(\mathcal{B},\mathcal{U},s_1 )} $ is $\mu$-a.-c. Recall also that $\lim_{n\to +\infty}r_n=0$ and for every $n\in \N$, $\vert U_n \vert \leq r_n.$
So, for every $ n\in\mathcal{N}_{\mu}(\mathcal{B},\mathcal{U},s_1)$,
\begin{equation}
\label{eqmunani}
\mathcal{H}^{\mu,s_1}_{\infty}\left (U_n \right)\geq \mu( B_n ) \ \ \ \ \ \ \ \mbox{ and } \ \ \ \vert U_n \vert\leq r_n.
\end{equation}
In particular, by Definition \ref{mucont}, for every $n\in\mathcal{N}_{\mu}(\mathcal{B},\mathcal{U},s_1) $ for any set $E_n \subset U_n $ with $\mu(E_n)=\mu(U_n)$,
$$
\mu(B_n)\leq \mathcal{H}^{\mu,s_1}_{\infty}(U_n)\leq \mathcal{H}^{s_1} _{\infty}(E_n).
$$
By Lemma \ref{carl}, and the notations therein, one has
$$
m_{E_n}^{s_1} (U_n)=1 \leq \frac{\kappa_d \vert U_n \vert ^{ {s_1} }}{\mathcal{H}^{{s_1} }_{\infty}(E_n)}\leq \frac{\kappa_d \vert U_n \vert ^{ {s_1} }}{\mu(B_n)}.
$$
This implies that
\begin{equation}
\label{bonnemajoani}
\mu(B_n)\leq \kappa_d \vert U_n \vert ^{ {s_1} }.
\end{equation}
By equation \eqref{bonnemajoani}, recalling the fact that the sets $ \mathcal{W}_1 \subset \left\{U_n\right\}_{n\in\mathbb{N}}$ , one has for every $U\in \mathcal{W}_1 ,$
\begin{equation}
\label{eq12}
\eta(U )\leq \frac{\mu(B^{[U]})}{\frac{1}{2Q_{d,4}} }\leq 2Q_{d,4}\kappa_d |U | ^{s_1 }.
\end{equation}
\subsubsection*{Step 2} This step (and all the following steps) is split into two sub-steps. First, into each open set $U$ of $\mathcal{W}_1$, smaller intermediary balls are selected according to the $\mu$-essential content of $U$. Then in a second time, each intermediary ball will be covered by balls of the sequence $(B_n)_{n\in\mathbb{N}}$ according to the measure $\mu$ and, as in step 1, the sets $U_n$ associated with this covering will form the generation $\mathcal{W}_2$.
\sk
Let $g\in \N$ be such that for every $n\geq g$, $r_n\leq \frac{1}{3}\min(|U|:U\in \mathcal{W}_1)$.
As above, since $s_2<s(\mu , \mathcal{B},\mathcal{U})$, the sub-sequence $(B_n )_{n\in\mathcal{N}_{\mu}(\mathcal{B},\mathcal{U},s_2), n\geq g } $ is $\mu$-a.c. The same arguments as above yield for every $ n\in\mathcal{N}_{\mu}(\mathcal{B},\mathcal{U},s_2) $,
\begin{equation}
\label{eqmun2ani}
\mathcal{H}^{\mu,s_2}_{\infty}\left (U_n\right)\geq \mu( B_n ) \ \ \ \ \ \ \ \mbox{ and } \ \ \ \vert U_n \vert \leq r_n
\end{equation}
and
\begin{equation}
\label{bonnemajo2ani}
\mu(B_n)\leq \kappa_d \vert U_n \vert^{ {s_2} }.
\end{equation}
\subsubsection*{Covering with respect to the $\mu$-essential content}
Consider $U \in\mathcal{W}_{1}$. Set $\beta=\overline\dim_H(\mu)$. For $0\leq k\leq \lfloor \frac{\beta-\alpha}{\varepsilon_2}\rfloor+1$, define $\theta_k =\alpha+k\varepsilon_2 $.
Write
\begin{equation}
\label{defELani}
E_{U}= U\cap E_{\mu}^{[\alpha,\beta],\varepsilon_2} \cap \limsup_{n\to +\infty }B_n.
\end{equation}
Notice that by Proposition \ref{mle} and by item (1) of Lemma~\ref{equiac}, one has $\mu(E_{U})=\mu(U)$.
In addition, using the definition \eqref{eqmucont} of $ \mathcal{H}^{\mu,s_2}_{\infty}$, the fact that $E_U\subset U$ and $\mu(E_{U})=\mu(U)$, and finally \eqref{defsndani} applied with $B_n = B^{[U]}$, one gets
\begin{equation}
\label{eq21ani}
\mathcal{H}_{\infty}^{s_2}(E_U)\geq \mathcal{H}^{\mu,s_2}_{\infty}(U)\geq \mu(B^{[U]})>0.
\end{equation}
This allows us to apply Proposition \ref{carl}: there exists a Borel probability measure $m_{E_U}^{s_2}$ supported on $E_{U}$ such that for every ball $B:=B(x,r)$, one has $$m_{E_U}^{s_2}(B)\leq \kappa_d \frac{r^{s_2}}{ \mathcal{H}^{s_2}_{\infty}(E_U)}.$$
Also, since $m_{E_U}^{s_2}(E_U)=1$ and $E_U \subset E_{\mu}^{[\alpha,\beta],\varepsilon_2}$, and recalling \eqref{emut}, for any $0\leq k\leq \lfloor \frac{\beta-\alpha}{\varepsilon_2}\rfloor+1$, there exists $\rho_{k,\varepsilon_2}$ such that
$$m_{E_U}^{s_2}(E^{[\theta_k,\theta_{k+1}],\rho_{k,\varepsilon_2},\varepsilon_2}_{\mu})\geq \frac{1}{2}m_{E_U}^{s_2}(E^{[\theta_k,\theta_{k+1}],\varepsilon}_{\mu}).$$
Setting $\rho_U =\min_{0\leq k\leq \lfloor \frac{\beta-\alpha}{\varepsilon_2}\rfloor+1}\rho_{k,\varepsilon_2}$ one has, for any $0\leq k\leq \lfloor \frac{\beta-\alpha}{\varepsilon_2}\rfloor +1$,
\begin{equation}
\label{equarecti}
m_{E_U}^{s_2}(E^{[\theta_k,\theta_{k+1}],\rho_{U},\varepsilon_2}_{\mu})\geq \frac{1}{2}m_{E_U}^{s_2}(E^{[\theta_k,\theta_{k+1}],\varepsilon}_{\mu}).
\end{equation}
In particular,
\begin{equation}
m_{E_U}^{s_2}(E_{\mu}^{[\alpha,\beta],\rho_U ,\varepsilon_2})\geq \frac{1}{2}
\end{equation}
\mk
Let
\begin{equation}
\label{defSLani}
S_U := \bigcup_{0\leq k\leq \lfloor\frac{\beta-\alpha}{\varepsilon_2}\rfloor+1}E^{[\theta_k,\theta_{k+1}],\rho_{U},\varepsilon_2}_{\mu}\cap E_U \cap \left\{x\in \R^d : \ \overline{\dim}_{\mathrm{loc}}(m_{E_U}^{s_2} ,x)\leq d\right\}.
\end{equation}
Recalling that for every probability measure $m$, $m(\{x= \overline{\dim}_{\mathrm{loc}}(m ,x)\leq d\})=1$, one necessarily has $m_{E_U}^{s_2}(S_U) \geq 1/2$.
Let $x\in S_U $; consider $0\leq k_x\leq \lfloor \frac{\beta-\alpha}{\varepsilon_2}\rfloor+1$ such that $x\in E^{[\theta_{k_x},\theta_{k_x+1}],\rho_{U},\varepsilon_2}_{\mu}.$ Applying Lemma \ref{doub}, there exists $0<r_{x}<\min \big (\rho_x,\frac{1}{3}\min\left\{\vert V\vert:V\in\mathcal{W}_1 \right\}\big) $ and $t_x\in (5,6)$ such that:
\begin{align}
\label{recap-ani}
&10\, r_x <\rho_{U};\\
\label{recap0ani}
& B(x,r_x)\subset U \ \ \mbox{ and } \ \ \mu(\partial B(x,r_x/t_x))=0;\\
\label{recap1ani}
& r_x^{-\ep_2} \geq 5^{d} \frac{4 Q_{d,1}} {C_{\ep_3,d}} \frac{\eta(U)}{\mu(B^{[U]}) }\geq 5^{s_2} \frac{4 Q_{d,1}} {C_{\ep_2,d}} \frac{\eta(U)}{\mu(B^{[U]}) };\\
\label{recap1bisani}
& r_{x} ^{\theta_{k_x}+2\varepsilon_2}\leq\mu(B(x,r_{x}))\leq r_{x} ^{\theta_{k_x}-2\varepsilon_2};\\
\label{recap2ani}
& m_{E_U}^{s_2}(B(x,r_x/t_x))\geq C_{\ep_2,d} \cdot m_{E_U}^{s_2}(B(x,r_x)).
\end{align}
Note that in \eqref{recap1ani} the second inequality follows automatically from the first one since $s_2\le \alpha\le d$ and the constant $C_{\ep,d}$ is an increasing function of $\varepsilon$.
The family $\left\{B(x,r_x) : x\in S_U\right\}$ forms a covering of $S_U$. We apply Lemma~\ref{besimodi} with $v=1$ (i.e., the standard Besicovich covering Theorem) to this family to extract $Q_{d,1} $ subfamilies of balls, $\mathcal{G}^{U}_1 ,...,\mathcal{G}^{U}_{Q_{d,1} }$ such that:
\begin{itemize}
\sk
\item $\forall 1\leq i \leq Q_{d,1} $, $\forall B \neq B' \in\mathcal{G}^{U}_{i}$, one has $B \cap B' =\emptyset,$
\sk
\item $S_U \subset \bigcup_{i=1}^{Q_{d,1} }\bigcup_{B \in\mathcal{G}^{U}_i}B.$
\end{itemize}
In particular, $m_{E_U}^{s_2} \left(\bigcup_{i=1}^{ Q_{d,1} }\bigcup_{B\in\mathcal{G}^{U}_i}B \right) \geq m_{E_U}^{s_2}(S_U) \geq 1/2 $.
At least one of these families, say $\mathcal{G}^{U}_{i_0}$, verifies that
$$m_{E_U}^{s_2}\left (\bigcup_{B \in\mathcal{G}^{U }_{i_0}} B \right )\geq \frac{m_{E_U}^{s_2}(S_U)}{Q_{d,1}}\geq \frac{1}{2Q_{d,1}}.$$
Writing $ \mathcal{G}^{U}_{i_0}=\left\{B_{i_0,k}^{U}\right\}_{k\in\mathbb{N}}$, one can find an integer $N_U $ so large that
$$m_{E_U}^{s_2}\left(\bigcup_{1\leq k \leq N_U}B_{i_0,k}^{U}\right)\geq \frac{1}{4 Q_{d,1}}. $$
Remind that each $B_{i_0,k}^{U}$ is a ball $B(x,r_x)$ satisfying \eqref{recap0ani}, \eqref{recap1ani} and \eqref{recap2ani}.
Finally, setting $\mathcal{G}^{U}=\left\{B(x, r_x/t_x): B(x,r_x) \in \mathcal{F}^{U }_{i_0} \right\}$, one has by construction
\begin{equation}
\label{eq22ani}
m_{E_U }^{s_2}\left(\bigcup_{B\in \mathcal{G}^{U} }B \right) = \sum_{B\in \mathcal{G}^{U} } m_{E_U }^{s_2}(B) \geq \frac{C_{\ep_2,d}}{4 Q_{d,1}} .
\end{equation}
One then extends the pre-measure $\eta$ to the Borel $\sigma$-algebra generated by the balls of $\mathcal{G}^{U}$, by the formula
\begin{equation}
\label{etaSani}
\mbox{ for every $B \in \mathcal{G}^{U}$, } \ \ \ \eta (B )=\eta(U)\times \frac{m_{E_U}^{s_2}(B )} {\sum_{B' \in \mathcal{G}^{U}}m_{E_U}^{s_2}(B')}.
\end{equation}
By construction, this formula is consistent since $\eta(U) = \sum_{B\in \mathcal{G}^{U}} \eta (B)$.
\mk
Observe that by \eqref{majcarl}, \eqref{eq22ani} and \eqref{eq21ani}, one has for every $B \in \mathcal{G}^{U}$,
\begin{align}
\label{eq23ani}
\eta (B ) & \leq \eta(U) \kappa_d \frac{|B|^{s_2}}{\mathcal{H}^{s_2}_{\infty}(E_U) } \frac{4 Q_{d,1}}{C_{\ep_2,d}} \leq \frac{4 Q_{d,1} \kappa_d }{C_{\ep_2,d}} \frac{ \eta(U) } {\mu(B^{[U]})} |B|^{s_2} \leq |B|^{s_2-\ep_2} ,
\end{align}
where the second inequality of \eqref{recap1ani} was used.
\mk
This is achieved simultaneously for all $U\in\mathcal{W}_{1}$.
\mk
\subsubsection*{Covering with respect to $\mu$} Now, in order to build the second generation of the Cantor set $K$, we select balls of $\mathcal{B}$ that lie in the interior of these intermediate balls $B\in \mathcal{G}^{U}$.
Let $U\in\mathcal{W}_{1}$ and $B\in\mathcal{G}^{U}$ be one of these intermediary balls. Since $\mathcal{B}$ is $\mu$-a.c., the last part of Lemma \ref{covO} proves the existence of a finite family $ \mathcal{F}^{B}=\left\{U_{n_i}\right\}_{ 1\leq i \leq {N_B}}$ such that
\begin{itemize}
\sk
\item[($i_1$)] for every $ 1\leq i\leq {N_B }$, one has $B_{n_i}\subset \widering{B} $ and
\begin{equation}
\label{eq24ani}
\max\left\{ 2 Q_{d,4} \frac{\eta(B)}{\mu(B)},\frac{5^d 4Q_{d,1} \kappa_d }{C_{\varepsilon_3 ,d}}\right\} \leq {r_{n_i }^{-\varepsilon_2}} ,
\end{equation}
\sk
\item[($i_2$)] for every $ 1\leq i\neq j\leq {N_B}$, one has $B_{n_i}\cap B_{n_j}=\emptyset.$
\end{itemize}
In addition, recalling that $\mu(\partial B)=0$ by \eqref{recap0ani}, one has
\begin{equation*}
\mu(B_{n_i}) >0 \ \ \mbox{ and } \ \ \mu\Big (\bigcup_{1\leq i\leq {N_B}}B_{n_i} \Big)\geq \frac{3\mu(\widering{B})}{4}=\frac{3\mu(B)}{4}.
\end{equation*}
{Recall the definitions \eqref{emu} and \eqref{defSLani} of the sets $E_{\mu}^{[a,b],\rho_{U},\varepsilon_2}$ and $S_{U}$. By equations \eqref{recap-ani}-\eqref{recap2ani}, there exists $\alpha\leq a\leq \beta$ such that the center of $B$ belongs to $S_U\subset E_{\mu}^{[a,a+\varepsilon_2],\rho_U \varepsilon_2,}$ and $\vert B\vert \leq \rho_U$, hence one has}
$$ {\mu(B\cap \widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\rho_{U},\varepsilon_2})\geq \frac{3}{4}\mu(B).}$$
{By $(i_2)$, and recalling \eqref{emu}, one has
\begin{align*}
&{\mu\Big (\bigcup_{B_{n_i} : B_{n_i} \cap \widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\rho_U ,\varepsilon_2}\neq \emptyset } B_{n_i}\Big)}
\geq \mu\Big(\bigcup_{1\leq i\leq N_B}B_{n_i}\cap \widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\rho_U ,\varepsilon_2}\Big)\\
&=\mu\left(\bigcup_{1\leq i\leq N_B}B_{n_i}\right)+\mu\left(\widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\rho_U ,\varepsilon_2}\right)- \mu\left( \widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\rho_U ,\varepsilon_2 }\bigcup \bigcup_{1\leq i\leq N_B} \!\! B_{n_i} \right)\nonumber\\
& {\geq \frac{3}{4}\mu(B)+\frac{3}{4}\mu(B)-\mu(B)=\frac{1}{2}\mu(B).}\nonumber
\end{align*}
By a slight abuse of notations, up to an extraction, we still denote by $\left\{B_{n_i}\right\}_{1\leq i\leq N_B}$ the balls $B_{n_i}$ such that $B_{n_i} \cap \widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\varepsilon_2,\rho_U }\neq \emptyset $. {The last inequality implies that the family of balls $\left\{B_{n_i}\right\}_{1\leq i\leq N_B}$ can be chosen so that it verifies conditions $(i_1)$ and $(i_2)$, as well as the two following additional conditions:}
\begin{align*}
{(i_3 )} \ \ \ \ \ & \mu(B_{n_i}) >0 \ \ \mbox{ and } \ \ \mu\Big (\bigcup_{1\leq i\leq {N_B}}B_{n_i} \Big)\geq \frac{\mu(\widering{B})}{2}=\frac{\mu(B)}{2},\\
{(i_4)} \ \ \ \ & \mbox{for every $ 1\leq i\leq N_B$}, \ \ \ \ \ {B_{n_i} \cap \widetilde{ E}_{\mu}^{[a,a+\varepsilon_2],\rho_U ,\varepsilon_2}\neq \emptyset.}
\end{align*}
The obtained family is still denoted by $\mathcal{F}^{B}$.
Applying again Lemma \ref{distcov} to $\mathcal{F}^{B}$ with $v=4$, as in step one (see \eqref{recouetape1}, $(i^{\prime})$ and $(ii^{\prime})$), if one must consider a subfamily, one can assume that the family $\mathcal{F}^{B}$ satisfies $(i_1)$ and $(i_4)$ as well as the following condition $(i_2^{\prime})$ and $(i_3 ^{\prime})$:\mk
\begin{itemize}
\item[$(i_2^{\prime}):$] for every $ 1\leq i\neq j\leq {N_B}$, one has $ 4 B_{n_i}\cap 4 B_{n_j}=\emptyset.$ \mk
\item[$(i_3^{\prime}):$] $\mu(B_{n_i}) >0 \ \ \mbox{ and } \ \ \mu\Big (\bigcup_{1\leq i\leq {N_B}}B_{n_i} \Big)\geq \frac{\mu(\widering{B})}{2Q_{d,4}}=\frac{\mu(B)}{2Q_{d,4}},$
\end{itemize}
\sk
Finally one defines
$$\mathcal{W}_{2}=\bigcup_{U \in\mathcal{W}_1}\bigcup_{B\in\mathcal{F}^{1,U }}\mathcal{F}^{B} \ \ \ \mbox{ and } \ \ \ W_{2}=\bigcup_{L\in\mathcal{W}_{2}}L.$$
The pre-measure $\eta$ is then extended to the $\sigma$-algebra generated by the elements of $\mathcal{W}_{2}$ by setting for every $U \in \mathcal{W}_{1}$, every $B\in \mathcal{G}^{U}$ and $V \in \mathcal{F}^{B}$,
\begin{equation}
\label{etaB2ani}
\eta (V)=\eta (B)\times \frac{\mu(B^{[V]})}{\sum_{V^{\prime} \in \mathcal{F}^{B} }\mu(B^{[V^{\prime}]})} .
\end{equation}
By construction, one has $\sum _{V \in \mathcal{F}^{B}} \eta(V) = \eta(B)$. Also, \eqref{eq24ani},\eqref{etaB2ani}, and $(i_3 ^{\prime})$ imply
\begin{equation}
\label{eqmicani}
\frac{\eta (V)}{\mu(B^{[V]})} \leq 2Q_{d,4}\frac{\eta (B)}{\mu(B)}\leq \vert V\vert ^{-\varepsilon_2},
\end{equation}
so that by \eqref{bonnemajo2ani} and \eqref{eqmicani} one has
\begin{align}
\label{eq25ani}
\eta (V) &\leq 2Q_{d,4}\frac{\eta (B)}{\mu(B)} \times \mu(B^{[V]} )\leq {|B^{[V]}| ^{-\varepsilon_2}} |V|^{s_2} \leq |V|^{s_2-\ep_2} .
\end{align}
\subsubsection{Recurrence scheme and end of the construction}
Let $p\in\mathbb{N}^{*}$ be an integer, and set $\mathcal{W}_0=\R^d$. Suppose that sets of balls $\mathcal{W}_1$, ..., $\mathcal{W}_{p}$ as well as the measure $\eta $ are constructed such that :
\begin{enumerate}
\mk\item
for every $ 1\leq q\leq p$, $\mathcal{W}_q \subset \{U_{n} \}_{n\geq q}$, $\mathcal{W}_q\subset \mathcal{W}_{q-1}$, and $\eta$ is defined on the $\sigma$-algebra generated by the elements of $\bigcup_{q= 1}^p \mathcal{W}_{q}$.
\mk
\item
For every $ 1\leq q\leq p-1$, for every $U \in\mathcal{W}_{q}$, setting, as in step 2, $E_U =\limsup_{n\in\mathcal{N}_\mu(\mathcal{B},\mathcal{U},s_q)}B_n \cap U\cap E_{\mu}^{[\alpha,\beta],\varepsilon_q}$, then $\mathcal{H}^{s_q}_{\infty}(E_U )>0$. If $m_{E_U}^{s_q}$ stands for the measure associated with $E_U$ provided by Proposition \ref{carl}, there exists $\rho_U >0$ such that, for every $0\leq k\leq\lfloor \frac{\beta-\alpha}{\varepsilon_q}\rfloor+1$, setting $\theta_k=\theta_k^{(q)}=\alpha+k\varepsilon_q$, one has
$$m_{E_U}^{s_q} (E_U \cap E_{\mu}^{[\theta_k ,\theta_{k+1}],\rho_U ,\varepsilon_q})\geq \frac{1}{2}m_{E_U}^{s_q} (E_U \cap E_{\mu}^{[\theta_k ,\theta_{k+1}],\varepsilon_q}).$$
In particular,
$$m_{E_U}^{s_q} (E_U \cap \bigcup_{0\leq k\leq\lfloor \frac{\beta-\alpha}{\varepsilon_q}\rfloor+1} E_{\mu}^{[\theta_k ,\theta_{k+1}],\rho_U ,\varepsilon_q})\geq \frac{1}{2}.$$
%
\mk
\item
For every $ 1\leq q\leq p-1$, for every $ U\in\mathcal{W}_{q}$, there exists a finite family $\mathcal{G}^{U}$ of balls $B(x,r_x/t_x)$, where $x$, $r_x <\frac{1}{3}\min\left\{\vert \widetilde{U}\vert:\widetilde{U}\in\mathcal{W}_{q}\right\}$ and $t_x$ satisfy \eqref{recap-ani}, \eqref{recap0ani}, \eqref{recap1ani}, \eqref{recap2ani} and \eqref{eq22ani}. Also, if $B\neq B' \in \mathcal{G}^{U}$, $3B\cap 3B' =\emptyset$.
Also, for every $B\in \mathcal{G}^{U}$, \eqref{etaSani} and \eqref{eq23ani} hold true. Moreover $\mathcal{W}_{q+1}\subset \bigcup_{U\in \mathcal{W}_q} \mathcal{G}^{U}$.
\mk
\item
For every $ 1\leq q\leq p-1$, for every $ U\in\mathcal{W}_{q}$, for every $B\in \mathcal{G}^{U}$ there exists a family $\mathcal{F}^{B} \subset \left\{U_{n} \right\}_{n\geq q}$ of pairwise disjoint open sets such that :
\begin{itemize}
\item for every $\widetilde{U}\neq \widehat{U} \in\mathcal{F}^{B}$, one has
\begin{equation}
\label{disjani}
4B^{[\widetilde{U}]}\cap 4 B^{[\widehat{U}]} =\emptyset;
\end{equation}
\item for every $\widetilde{U}\in \mathcal{F}^{B}$, $\widetilde{U}\subset \widering{B}$, \eqref{etaB2ani} and \eqref{eq25ani} hold true, as well as
\begin{equation}
\label{majornani}
2Q_{d,4} \frac{\eta(B)}{\mu(B)} \leq |B^{[\widetilde{U}]}|^{-\varepsilon_{q+1}}
\end{equation}
%
and
\begin{equation}
\label{hypsupani}
{ B^{[\widetilde{U}]}\cap \widetilde{ E}_{\mu}^{[\theta_{k_B} ,\theta_{k_B+1}],\rho_{U},\varepsilon_{q+1}}\neq \emptyset;}
\end{equation}
\item the following inequality also holds true:
\begin{equation}
\label{eq31ani}
\mu\left (\bigcup_{\widetilde{U}\in \mathcal{F}^{B}} B^{[\widetilde{U}]}\right) \geq \frac{\mu(B)}{2Q_{d,4}}.
\end{equation}
\end{itemize}
\end{enumerate}
\mk
In item (3), the fact that $3B\cap 3B' =\emptyset$ just follows from the choice of $B(x,r_x/t_x)$ instead of simply $B(x,r_x)$.
The proof follows then exactly and rigorously the same lines as those of Step 2. We do not reproduce it here, the only differences are that $\mathcal{W}_1$, $\mathcal{W}_2$ and $s_2$ are replaced by $\mathcal{W}_p$, $\mathcal{W}_{p+1}$ and $s_{p+1}$.
\medskip
Finally, define the Cantor set
$$K=\bigcap_{p\geq 1}W_{p}= \bigcap_{p\geq 1} \bigcup_{V\in \mathcal{W}_p}B^{[V]}.
$$
Applying Caratheodory's extension Theorem to the pre-measure $\eta$ yields a probability outer-measure on $\R^d$ that we still denote by $\eta$, which is metric, so that Borel sets are $\eta$-measurable and its restriction to Borel sets belongs to $\mathcal M(\R^d)$. The so obtained measure $\eta$ is fully supported on $K$. Also, for every $p\geq 2$, for any $U\in\mathcal{W}_p$, $B\in \mathcal{G}^{U}$, and $\widetilde{U}\in \mathcal{F}^{B}$, the inequalities \eqref{etaSani}, \eqref{eq23ani},\eqref{etaB2ani} and \eqref{eq25ani} holds with $s_p$ and $\varepsilon_p$ instead of $s_2$ and~$\varepsilon_2$.
\subsubsection{Upper-bound for the mass of a ball}
One first recall the following lemma (see, e.g.,~\cite{BV}).
\begin{lemme}
\label{geo}
Let $A=B(x ,r)$ and $B=B(x^{\prime} ,r^{\prime})$ be two closed balls, $q\geq 3$ such that $A\cap B\neq \emptyset$ and $A\setminus (qB)\neq \emptyset$. Then $r^{\prime}\leq r$ and $qB\subset 5A.$
\end{lemme}
Define the gauge function $\zeta:\mathbb{R}^{+}\mapsto \mathbb{R}^{+}$ as follows:
\begin{itemize}
\item
if for some $p\geq 1$, $\frac{1}{3}\min\left\{\vert U \vert: \ U\in\mathcal{W}_{p+1}\right\}\leq r< \frac{1}{3}\min\left\{\vert U \vert: \ U\in\mathcal{W}_p\right\}$, then $\zeta(r)=2Q_{d,4}10^d r^{s_p-5\varepsilon_p }$,\mk
\item
if $r \geq \frac{1}{3}\min\left\{\vert U \vert: \ U\in\mathcal{W}_{1}\right\}$, $\zeta(r)=1$,\mk
\item
$\zeta(0)=0$.\mk
\end{itemize}
Since $\ep_p\to 0$, one checks that $\lim_{r\to 0^+}\frac{ \log(\zeta(r))}{\log(r)}=\min\left\{s(\mu,\mathcal{B},\mathcal{U}),\dimm(\mu)\right\}$.
\medskip
Let $A$ be a ball of radius $r$. If there exists $n\in\mathbb{N}$ such that $A$ does not intersect $K_n$ then $\eta(A)=\eta(A\cap K_n)=0.$ Suppose that for every $n\in\mathbb{N}$, $A$ intersects $K_n$. The goal is to prove that $\eta(A) \leq \zeta(|A|)$ when $|A|$ is small.
\medskip
Some cases must be distinguished.
First if for every $n\in\mathbb{N}$, $A$ intersects only one contracted set $V_n$ of $K_n$, then by \eqref{eq23ani}
$$\eta(A)\leq \eta(V_n)\leq \vert V_n \vert^{s_n-\varepsilon_n }\underset{n\rightarrow+\infty}{\rightarrow}0.$$
In the other case, there exists $p \in\mathbb{N}$ such that $A$ intersects only one element of $\mathcal{W}_{p}$, and at least two elements of $\mathcal{W}_{p +1}$. Denote by $U$ the unique element of $\mathcal{W}_p$ intersecting~$A$.
\begin{enumerate}
\item {\bf Case 1:} If $ |A| \geq |U| $, then by \eqref{eq25ani}
\begin{equation}
\label{res0ani}
\eta(A)\leq \eta(U)\leq |U| ^{s_p-\varepsilon_p}\leq \zeta( |A|).
\end{equation}
\item {\bf Case 2:} If $|A| < |U| $ and $A$ intersects at least two balls of $\mathcal{G}^{U}$: Observe that when $A$ intersects two balls $B$ and $B'$ of $\mathcal{G}^{U}$, since by item $(3)$ of the recurrence scheme $3B\cap 3B' =\emptyset$, one necessarily has (by Lemma \ref{geo}) $B\cup B' \subset 5A$. Hence, $\bigcup_{B\in \mathcal{G}^{U}: B\cap A\neq \emptyset}B \subset 5A$ and by \eqref{etaSani} and \eqref{eq22ani},
\begin{align*}
\eta(A) & = \eta(U )\times \frac{\sum_{B\in \mathcal{G}^{U}: B\cap A\neq \emptyset} m_{E_{U}}^{s_{p+1}}(B)}{{\sum_{B' \in \mathcal{G}^{U}}m_{E_U}^{s_{p+1}}(B')}} \leq \frac{4 Q_{d,1}} {C_{\ep_{p+1},d}} \eta(U) m_{E_{U}}^{s_{p+1}}(5A ) .
\end{align*}
Then, by \eqref{majcarl}, \eqref{eq21ani}, \eqref{eq24ani} and \eqref{eqmicani}
\begin{align}
\label{res1ani}
\eta(A) &\leq \frac{4 Q_{d,1}} {C_{\ep_{p+1},d}} \eta(U ) \kappa_d \frac{ (5|A|)^{s_{p+1}}}{ \mathcal{H}^{\mu,s_{p+1}}_{\infty}(E_U )}
\leq 5^{s_{p+1}} \frac{4 Q_{d,1}\kappa_d} {C_{\ep_{p+1},d}} \frac{ \eta(U ) } { \mu(B^{[U]}) } { |A|^{s_{p+1}}}\nonumber\\
&\leq |A|^{s_{p+1}} |U|^{-2\ep_{p}} \leq |A|^{s_{p+1}-2\ep_{p+1}}\leq \zeta(|A|),
\end{align}
where we used that$ |A|< |U|$, and the mappings $x\mapsto \vert U\vert ^{-x}$ and $x\mapsto x^{-\ep_{p+1}}$ are decreasing.
\mk
\item {\bf Case 3:} If $A$ intersects only one ball of $\mathcal{G}^{U}$: calling $B$ this particular ball and $r_B$ its radius (at this stage there should be no confusion with the radii of the terms of the sequence $(B_n)_{n\ge 1}$), two cases must again be distinguished:
\begin{enumerate}
\mk
\item {\bf Subcase 3.1: $|B|\leq |A|$:} by \eqref{eq23ani},
\begin{align}
\label{res2ani}
\eta(A) & \leq \eta(B) \leq |B|^{s_{p+1}-\ep_{p+1}} \leq |A|^{s_{p+1}-\ep_{p+1} } \leq\zeta(|A|).
\end{align}
\mk
\item {\bf Subcase 3.2: $|A|\leq |B|$:}
Denote by $k_B$ the integer such that its center belongs to $E_{\mu}^{[\theta_{k_B} ,\theta_{k_B+1}] ,\rho_{U},\varepsilon_{p+1}}$.
The ball $A$ must intersect at least two elements $V\neq V'$ of $\mathcal{W}_{p+1}$ (by definition of $p$). Note that those sets must belong to $\mathcal{F}^{B}$ (because $A$ intersects only $B$). Applying Lemma \ref{geo} to the ball $A$ with any of those ball $V\in\mathcal{F}_{p+1}$ , since $A\cap V \neq \emptyset$ and $A\setminus B^{[V]}\neq \emptyset$ (because $A$ intersects an other dilated ball, $B^{[V^{\prime}]}$ by hypothesis and two such balls verifies \eqref{disjani}), one has
\begin{equation}
\label{incani}
\bigcup_{V\cap A\neq \emptyset}B^{[V]} \subset 5A.
\end{equation}
Then, \eqref{etaB2ani} and \eqref{eq31ani} imply that
\begin{align}
\label{majoxcani}
\eta(A) & = \eta (B)\cdot \frac{ \sum_{V\in\mathcal{W}_{p +1}: V\cap A\neq \emptyset} \mu(B^{[V]})}{\sum_{V' \in \mathcal{F}^{B} }\mu(B^{[V^{\prime}]})} \leq 2Q_{d,4}\frac{\eta (B)}{\mu(B)} \mu(5A).
\end{align}
Recalling \eqref{incani}, the ball $5A$ contains some of the balls of $\mathcal{F}^{B}$: Hence, by \eqref{hypsupani}, $\widetilde{ E}_{\mu}^{[\theta_{k_B} ,\theta_{k_B+1}],\rho_{U},\varepsilon_{p+1}}\cap 5A \neq \emptyset$. Since $\vert A\vert \leq \vert B\vert$, by \eqref{recap-ani}, since $r_B< \frac{1}{10}\rho_{U}$, for any $x\in \widetilde{ E}_{\mu}^{[\theta_{k_B} ,\theta_{k_B+1}],\rho_{U},\varepsilon_{p+1}}\cap 5A$ one has
\begin{equation}
\label{majo3cani}
\mu(5A)\leq \mu(B(x,10r))\leq (10r)^{\theta_{k_B}-2\varepsilon_{p+1}}.
\end{equation}
Recalling \eqref{recap1bisani} (applied to the ball $B$), one has
\begin{equation}
\label{majozcani}
\mu(B)\geq (r_B)^{\theta_{k_B}+2\varepsilon_{p+1}}.
\end{equation}
Using \eqref{eq23ani} (applied to $B$) \eqref{majoxcani}, \eqref{majo3cani} and \eqref{majozcani}, one obtains
\begin{align*}
\eta(A) &\leq 2Q_{d,4}r_B^{s_{p+1}-\varepsilon_{p+1}} \frac{\big(10 r\big)^{\theta_{k_B}-2\varepsilon_{p+1}}}{r_B^{\theta_{k_B}+2\varepsilon_{p+1}}} \\
&=2Q_{d,4} 10^{\theta_{k_B}-2\varepsilon_{p+1}} \frac{r_B^{s_{p+1}-\theta_{k_B} -\frac{\varepsilon_{p+1}}{\delta}-2\varepsilon_{p+1}}}{r^{s_{p+1}-\theta_{k_B}-\varepsilon_{p+1}-2\varepsilon_{p+1}}} r^{s_{p+1}-\theta_{k_B}-\varepsilon_{p+1}-4\varepsilon_{p+1}}\\
&\leq 2Q_{d,4} 10^{\theta_{k_B}-2\varepsilon_{p+1}} r^{s_{p+1}-5\varepsilon_{p+1}}.
\end{align*}
Finally, recalling \eqref{defskani}, $s_{p+1} -5\varepsilon_{p+1} \leq \alpha\leq \theta_k $, and since $r_B\geq r$ and $s_p \leq s_{p+1}$, one gets
$$
\eta(A)\leq 2 Q_{d,4} 10^{\theta_k-\varepsilon_{p+1}}\big(r \big)^{s_{p+1}-5\varepsilon_{p+1}}\leq 2Q_{d,4} 10^{d-\varepsilon_{p+1}}\vert A\vert ^{s_{p}-5\varepsilon_{p}}\leq 2Q_{d,4} 10^{d-\varepsilon_{p+1}}\vert A\vert ^{s_{p}-5\varepsilon_{p}},
$$
hence
\begin{equation}
\label{res3ani}\eta(A) \leq \zeta(|A|).
\end{equation}
\end{enumerate}
\end{enumerate}
Since for any $p\in\mathbb{N}$ and any ball $A$ satisfying $\vert A \vert \leq \frac{1}{3}\min\left\{\vert U \vert:U\in\mathcal{W}_p\right\}$, if $A$ intersects at most one element of $\mathcal{W}_p$, the inequalities \eqref{res0ani}, \eqref{res1ani}, \eqref{res2ani}, \eqref{res3ani} proves that for any such ball, one has
$\eta(A)\leq \zeta(|A|)$.
Hence recalling Definition \ref{hausgau}, by the mass distribution principle, one deduces that $\mathcal{H}^{\zeta}(K)\geq 1 $, which concludes the proof of Theorem \ref{lowani}.
\subsection{Proof of Corollary~\ref{minoeffec}}
\label{sec-mino}
\subsubsection{Some basic properties about the $\mu$-essential Hausdorff content}
{In this sub-section, basic properties of the $\mu$-essential content are established.}
\mk
First, we work in this article with the $\vert \vert \cdot \vert \vert_{\infty}$ norm for convenience. Any other norm could have been chosen, the corresponding quantities would have been equivalent.
\mk
In \eqref{hcont}, only closed balls are considered. Choosing open balls does not change the value of \eqref{eqmucont} in Definition \ref{mucont}.
{The following propositions are directly derived from the properties of the standard Hausdorff measures.
\begin{proposition}
\label{propmuc}
{Let $\mu \in\mathcal{M}(\R^d)$, $s\geq 0$ and $A\subset \R^d $ be a Borel set. The $s$-dimensional $\mathcal{H}^{\mu,s}_{\infty}(\cdot)$ outer measure satisfies the following properties:}
\begin{enumerate}
\item { If $\vert A\vert\leq 1$, the mapping $s\geq 0\mapsto \mathcal{H}^{\mu,s}_{\infty}(A)$ is decreasing from $\mathcal{H}^{\mu,0}_{\infty}(A)=1$ to $\lim_{t\to +\infty}\mathcal{H}^{\mu,t}_{\infty}(A)=0$.\mk
\item $0\leq \mathcal{H}^{\mu,s}_{\infty}(A)\leq \min\left\{\vert A\vert ^s, \mathcal{H}^{s}_{\infty}(A)\right\}$.
\mk
\item For every subset $ B\subset A$ with $\mu(A)=\mu(B)$, $\mathcal{H}^{\mu,s}_{\infty}(A)=\mathcal{H}^{\mu,s}_{\infty}(B).$}
\mk
\item For every $\delta \geq 1$, $\mathcal{H}^{\mu,\frac{s}{\delta}}_{\infty}(A)\geq (\mathcal{H}^{\mu,s}_{\infty}(A))^{\frac{1}{\delta}}.$
\mk
\item For every $ s>\overline{\dim}_H (\mu)$, $\mathcal{H}^{\mu,s}_{\infty}(A)=0.$\mk
\end{enumerate}
\end{proposition}
\begin{proof}
{Items $(1)$, $(2)$, $(3)$ directly follow from the definition. Item $(4)$ is obtained by concavity of the mapping $x\mapsto |x|^{1/\delta}$.}
\mk
{(5)} By Definition \ref{dim}, for any $s>\overline{\dim}_H (\mu)$, there exists a set $E$ with $\dim_{H}(E)<s$ and $\mu(E)=1.$ Using item (2), one has then $0\leq\mathcal{H}^{\mu,s}_{\infty}(A)=\mathcal{H}^{\mu,s}_{\infty}(A\cap E)\leq \mathcal{H}^{s}_{\infty}(A\cap E)\leq \mathcal{H}^{s}(E)=0$.
\sk
\end{proof}
\subsubsection{Proof of Corollary \ref{minoeffec}}
One starts with a lemma, the proof of which can be found in \cite{ED2}.
\begin{lemme}
\label{gscainf}
Let $\mu \in\mathcal{M}(\R^d)$.
Let $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ be a $\mu $-a.c sequence of balls of $ \R^d$.
Then for every $\varepsilon>0$, there exists a $\mu $-a.c sub-sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ of $\mathcal{B} $ such that for every $n\in\mathbb{N}$, $\mu (B_{\phi(n)})\leq (r_{\phi(n)})^{\underline{\dim}_H (\mu )-\varepsilon}.$
\end{lemme}
\begin{proof}[Proof of Corollary~\ref{minoeffec}](1) Observe that item (2) of Proposition \ref{propmuc} implies that $ t(\mu,\delta,\varepsilon ,\mathcal{B}) \geq \dimm(\mu)-\ep$, and $ t(\mu,\delta ,\mathcal{B}) \geq \dimm(\mu)$.
Now choose $\varepsilon>0$ so small that $\frac{(1-\varepsilon)\dimm(\mu)}{\delta\cdot t(\mu, \delta,\varepsilon ,\mathcal{B} )}\leq 1$. Recalling Lemma \ref{gscainf}, up to an extraction, one can assume that for any $n\in\mathbb{N}$, $$\mu(B_n)\leq \vert B_n \vert ^{(1-\varepsilon^2).\dimm(\mu)}.$$
Due to \eqref{tdeltainf}, there exists $N_{\varepsilon}\in\mathbb{N}$ such that for any $n\geq N_{\varepsilon}$, $$\mathcal{H}^{\mu ,\dimm(\mu)-\varepsilon}_{\infty}(\widering B_n ^{\delta})\geq \vert B_n ^{\delta} \vert ^{(1+\varepsilon). t(\mu,\delta,\varepsilon ,\mathcal{B}) }.
$$
Then, Proposition \ref{propmuc} (4) implies that for every $n\geq N_{\varepsilon}$,
\begin{align*}
\mathcal{H}_{\infty}^{\mu, \frac{(1-\varepsilon)\dimm(\mu)\times(\dimm(\mu)-\varepsilon)}{\delta \cdot t(\mu,\delta,\varepsilon ,\mathcal{B}) }}(\widering B_n ^{\delta})&\geq(\mathcal{H}^{\mu,\dimm(\mu)-\varepsilon}_{\infty}(B_n ^{\delta}))^{\frac{(1-\varepsilon)\dimm(\mu)}{\delta\cdot t(\mu,\delta,\varepsilon ,\mathcal{B}) }}\\
&\geq \vert B_n ^{\delta} \vert ^{\frac{ (1+\varepsilon) t(\mu,\delta,\varepsilon ,\mathcal{B}) }{\delta \cdot t(\mu,\delta,\varepsilon ,\mathcal{B})}(1-\varepsilon).\dimm(\mu)}\\
&\geq \vert B_n \vert ^{ (1+\varepsilon)(1-\ep) \dimm(\mu)}
\geq \mu(B_n) .
\end{align*}
Thus, setting $s_{\delta,\varepsilon} = \frac{(1-\varepsilon)\dimm(\mu)\times(\dimm(\mu)-\varepsilon)}{\delta \cdot t(\mu,\delta,\varepsilon ,\mathcal{B}) }$, Corollary~\ref{zzani} yields
$$\dim_{H}(\limsup_{n\rightarrow+\infty}\widering B_n ^{\delta})\geq s_{\delta,\varepsilon}.$$
Since the result holds for any $\varepsilon>0$, one gets the desired conclusion.
\end{proof}
\section{Estimation of essential content for self-similar measures}
\label{sec-parti}
In this section one computes the Hausdorff content of balls in the case of the Lebesgue measure, and estimates it for any self-similar measure.
\subsection{Computation of essential content for the Lebesgue measure}
When the measure $\mu$ is the Lebesgue measure, the computations are quite easy.
\begin{proposition}
Let $B=B(x,r) $ be a ball in $\R^d$, and $\mathcal{L}^d$ be the $d-$dimensional Lebesgue measure. Then for any $0\leq s\leq d$,
$\mathcal{H}^{\mathcal{L}^d,s }_{\infty}(B)=\mathcal{H}^{\mathcal{L}^d ,s}_{\infty}(\widering{B})=r^s .$
\end{proposition}
\begin{proof}
One starts first by computing $\mathcal{H}^{\mathcal{L}^d ,d}_{\infty}(B)$.
Let $\ep>0$, and let $E\subset B$ be a Borel set with $\mathcal{L}^d (E)=\mathcal{L}^d (B)$. Notice first that since $B$ covers $E$, recalling that $\mathbb{R}^d$ is endowed with $\vert \vert \cdot \vert \vert_{\infty}$ one has $$\mathcal{H}^{\mathcal{L}^d,d}_{\infty}(E)\leq \mathcal{H}^{d}_{\infty}(B)\leq \vert B \vert^d.$$
Consider a sequence of balls $(L_n)_{n\in\mathbb{N}}$ such that $$\mathcal{H}^{d}_{\infty}(E)\leq\sum_{n\geq 0}\vert L_n \vert ^d \leq (1+\varepsilon)\mathcal{H}^{d}_{\infty}(E).$$
This implies
\begin{align*}
(1+\varepsilon)\vert B\vert ^d &\geq (1+\varepsilon)\mathcal{H}^{d}_{\infty}(B )\geq (1+\varepsilon)\mathcal{H}^{d}_{\infty}(E)\geq\sum_{n\geq 0}\vert L_n \vert^d \\
& \geq \sum_{n\geq 0}\mathcal{L}^d (L_n)\geq \mathcal{L}^d (E)=\mathcal{L}^d (B)=\vert B\vert ^d.
\end{align*}
Taking the infimum on the Borel sets $E \subset B$ such that $\mathcal{L}^d (E)=\mathcal{L}^d (A)$ gives
$$ \vert B\vert ^d \leq (1+\varepsilon)\mathcal{H}^{\mathcal{L}^d,d }_{\infty}(B) .$$
In particular,
$$ \frac{1}{1+\varepsilon}\vert B\vert ^d \leq \mathcal{H}^{\mathcal{L}^d ,d }_{\infty}(B)\leq \vert B\vert ^d .$$ Letting $\varepsilon \to 0$ shows that $\mathcal{H}^{\mathcal{L}^d,d }_{\infty}(B)= \vert B\vert ^d$.
This implies, with item $(4)$ of Proposition \ref{propmuc}, that for any $\delta \geq 1$,
$$\vert B\vert ^{\frac{d}{\delta}}\geq \mathcal{H}^{ \mathcal{L}^d,\frac{d}{\delta}}_{\infty}(B)\geq (\mathcal{H}^{\mathcal{L}^d,d}_{\infty}(B) )^{\frac{1}{\delta}}=\vert B\vert ^{\frac{d}{\delta}},$$
hence the result.
\end{proof}
\subsection{Proof of Theorem \ref{contss}}
\begin{proposition}
\label{autosim2}
Let $\mu$ be a self-similar measure. For any $0<\varepsilon\leq \dim(\mu)$, there exists a constant $\kappa( d,\mu,\varepsilon)\in(0,1)$ such that for any ball $B=B(x,r)$ with $x\in K$ (the attractor of the underlying IFS) and $r\leq 1$, one has $$\kappa( d,\mu,\varepsilon)\vert B\vert ^{\dim(\mu)-\varepsilon}\leq\mathcal{H}^{\mu, \dim(\mu)-\varepsilon }_{\infty}(\widering{B})\leq \mathcal{H}^{ \mu, \dim(\mu)-\varepsilon}_{\infty}(B)\leq\vert B\vert ^{\dim(\mu)-\varepsilon}.$$
In addition, for any $s>\dim( \mu)$ one has $\mathcal{H}^{\mu,s}_{\infty}(B)=0.$
\end{proposition}
\begin{proof}
Let $\left\{f_1 ,...,f_m\right\}$ the underlying IFS. Denote by $c_i$ the contraction ration of $f_i$, and $(p_1 ,...,p_m)$ the probability vector with positive entries associated with $\mu$ so that \eqref{def-ssmu2} is satisfied.
Set $\alpha=\dim(\mu)$ and $\Lambda=\left\{1,...,m\right\}$.
For $k\geq 0$ and $\underline{i}:=(i_1 ,...,i_k)\in\Lambda^k$, define
\begin{itemize}
\item $c_{\underline{i}}=c_{i_1}...c_{i_k}$, $f_{\underline{i}}=f_{i_1}\circ ... \circ f_{i_k}$ and $K_{\underline{i}}=f_{\underline{i}}(K),$ so that $\vert K_{\underline{i}}\vert = c_{\underline{i}}\vert K\vert$. \medskip
\item $\Lambda ^{(k)}=\left\{\underline{i}:=(i_i ,..., i_s) \ : \ c_{i_s}2^{-k}< c_{\underline{i}}\leq 2^{-k} \right\}$.
\end{itemize}
Note first that item (5) of Proposition \ref{propmuc} implies that for any $s>\dim(\mu)$, $\mathcal{H}^{\mu,s}_{\infty}(B)=0.$
Let us consider $0\leq s<\dim_H (\mu)$ and start by few remarks.
Recalling \eqref{emu} and Proposition \ref{mle}, let us fix $\rho_{\varepsilon}$ so that $\mu(E_{\mu} ^{[\alpha,\alpha], \rho_{\varepsilon},\varepsilon})\geq \frac{1}{2}$ and write $E=E_{\mu}^{[\alpha,\alpha] ,\rho_{\varepsilon}, \varepsilon}.$
Set $\Lambda ^{*}:=\bigcup_{k\geq 0}\Lambda ^k $, and for $\underline{i}\in\Lambda^*$, define $E_{\underline{i}}=f_{\underline{i}}(E)$ and $\mu_{\underline{i}}=\mu(f_{\underline{i}}^{-1})$. One has
\begin{align}
\nonumber E_{\underline{i}}
\nonumber &=\left\{f_{\underline{i}}(x)\in \R^d : \, x\in K \forall\, r\leq \rho_{\varepsilon}, \ \mu(B(x,r))\leq r^{\alpha -\varepsilon}\right\}\\
\nonumber &=\left\{f_{\underline{i}}(x):\, x\in K, \, \forall, c_{\underline{i}} r\leq c_{\underline{i}} \rho_{\varepsilon}, \ \mu(f_{\underline{i}} ^{-1}(B(f_{\underline{i}}(x),r c_{\underline{i}})))\leq \left(\frac{r c_{\underline{i}}}{c_{\underline{i}}}\right)^{\alpha -\varepsilon}\right\}\\
\label{eq77} &=\left\{y\in K_{\underline{i}} : \, \forall r'\leq c_{\underline{i}}\rho_{\varepsilon} , \ \mu_{\underline{i}}(B(y,r'))\leq \left(\frac{r'}{c_{\underline{i}}}\right)^{\alpha -\varepsilon}\right\},
\end{align}
Also,
$\mu_{\underline{i}}(E_{\underline{i}})=\mu(E)\geq \frac{1}{2}.$
One emphasizes that iterating the self-similarity equation gives
$$\mu=\sum_{\underline{i}^{\prime}\in\Lambda ^k}p_{\underline{i}^{\prime}}\mu_{\underline{i}^{\prime}}, $$
which implies that $\mu_{\underline{i}}$ is absolutely continuous with respect to $\mu$ (since all $p_{\underline{i}}$'s are strictly positive).
\sk
We are now ready to estimate the $\mu$-essential content of a ball $B$ centered on $K$.
\sk
Let $B=B(x,r)$ with $x\in K$ and $r\leq \min_{1\leq i\leq m}c_i$.
\sk
Since $x\in K$, there exists an $\underline{i }$ such that $\min_{1\leq j\leq m}c_j r\leq c_{\underline{i}}\vert K\vert\leq r$ and $K_{\underline{i}}\subset \widering{B}$.
By construction, $E_{\underline{i}} \subset \widering{B}$.
\sk
Consider a Borel set $A\subset B$ such that $\mu(A)=\mu(B).$ One aims at giving a lower-bound of the Hausdorff content of $A$ which does not depends on $A$.
\sk
Consider a sequence of balls $(L_n=B( z_n,\ell_n))_{n\geq 1}$ covering $A\cap E_{\underline{i}} $, such that $\ell_{n}<\rho_{\varepsilon} c_{\underline{i}}$ and $ z_n\in A\cap E_{\underline{i}} $.
Since $\mu_{\underline{i}}$ is absolutely continuous with respect to $\mu$, it holds that $\mu_{\underline{i}}(A)=1.$
By \eqref{eq77} applied to every $n\in\mathbb{N}$ , one has $\left(\frac{\vert L_n \vert}{c_{\underline{i}}}\right)^{\alpha-\varepsilon}\geq \mu_{\underline{i}}(L_n)$, so that
\begin{equation}
\label{eq78}
\sum_{n\in\mathbb{N}}| L_n| ^{\alpha-\varepsilon}\ \geq \sum_{n\in\mathbb{N}}c_{\underline{i}} ^{\alpha -\varepsilon}\mu_{\underline{i}}(L _n )\geq c_{\underline{i}} ^{\alpha -\varepsilon}\mu_{\underline{i}}\left(\bigcup_{n\in \N }L_n \right)\geq c_{\underline{i}} ^{\alpha -\varepsilon}\mu_{\underline{i}}(E_{\underline{i}})\geq \frac{1}{2}c_{\underline{i}} ^{\alpha -\varepsilon}.
\end{equation}
This series of inequalities holds for any sequence of balls $(L_n)_{n\in\mathbb{N}}$ with radius less than $\rho_{\varepsilon} c_{\underline{i}}$ centered on $A\cap E_{\underline{i}}$ and covering $A\cap E_{\underline{i}}$.
Now, assume that $(L_n)_{n\in\mathbb{N}}$ is a sequence of balls covering $A\cap E_{\underline{i}} $, which still verifies $\ell_{n}<\rho_{\varepsilon} c_{\underline{i}}$ but $z_n$ does not necessarily belongs to $A\cap E_{\underline{i}} $.
Let $n\in\mathbb{N}$. One constructs recursively a sequence of balls $(L_{n,j})_{1\leq j\leq J_n}$ \ such that the following properties hold for any $1\leq j\leq J_n$:
\begin{itemize}
\item[•] $L_{n,j}$ is centered on $A\cap E_{\underline{i}}\cap L_n$;\mk
\item[•] $A\cap E_{\underline{i}}\cap L_n \subset \bigcup_{1\leq j \leq J_n}L_{n,j}$;\mk
\item[•] for all $1\leq j\leq J_n,$ $\vert L_{n,j}\vert= \vert L_n \vert$;\mk
\item[•]the center of $L_{n,j}$ does not belong to any $L_{n,j^{\prime}}$ for $1\leq j^{\prime}\neq j\leq J_n$. \mk
\end{itemize}
To achieve this, simply consider $y_1\in A\cap E_{\underline{i}}\cap L_n$ and set $L_{1,n}=B(y_1,\ell_n).$ If $A\cap E_{\underline{i}}\cap L_n \nsubseteq L_{1,n}$, consider $y_2 \in A\cap E_{\underline{i}}\cap L_n \setminus L_{1,n}$ and set $L_{2,n}=B(y_2 ,\ell_n)$. If $A\cap E_{\underline{i}}\cap L_n \nsubseteq L_{1,n}\cup L_{2,n}$, consider $y_3 \in A\cap E_{\underline{i}}\cap L_n \setminus L_{1,n}\cup L_{2,n}$ and set $L_{3,n}=B(y_3 ,\ell_n)$, and so on...
Note that, for any $1\leq j\leq J_n$, any ball $L_{j,n}$ has radius $\ell_n$, intersects $L_n$ (which also has radius $\ell_n$) and, because $y_j \notin \bigcup_{1 \leq j^{\prime}\neq j\leq J_n}L_{j^{\prime},n}$, it holds that, for any $j\neq j^{\prime}$, $\frac{1}{3}L_{n,j}\cap \frac{1}{3}L_{n,j^{\prime}}=\emptyset$. By Lemma \ref{dimconst}, this implies that $J_n \leq Q_{d, \frac{1}{3}}.$
Hence, denoting by $(\widetilde L_n)_{n\in \N} $ the collection of the corresponding balls centered on $A\cap E_{\underline{i}} $ associated with all the balls $L_n$, one has by \eqref{eq78} applied to $(\widetilde L_n)_{n\in\mathbb N}$:
$$\sum_{n\in\mathbb{N}}|L_n| ^{\alpha-\varepsilon}\geq \frac{1}{Q_{d,\frac{1}{3}}} \sum_{n\in\mathbb{N}}|\widetilde L_n| ^{\alpha-\varepsilon} \geq \frac{1}{2Q_{d,\frac{1}{3}}} c_{\underline{i}} ^{\alpha -\varepsilon}.$$
Remark also that any ball of radius smaller that $r$ can be covered by at most $(\frac{2}{\rho_{\varepsilon}})^d$ balls of radius $r\rho_{\varepsilon}$.
This proves that, for any sequence of balls $\widehat{L}_n$ covering $A\cap E_{\underline{i}}$, since $c_{\underline{i}}\geq \frac{\min_{1\leq j\leq m}c_j }{\vert K\vert }\vert B\vert$, it holds that
\begin{equation}
\label{loceq}
\sum_{n\in\mathbb{N}}|\widehat{L}_n| ^{\alpha-\varepsilon} \geq \frac{\rho_{\varepsilon}^d}{2^{d+1}Q_{d,\frac{1}{3}}} c_{\underline{i}} ^{\alpha -\varepsilon}\geq \frac{\min_{1\leq j\leq m}c_j ^{\alpha-\varepsilon}\rho_{\varepsilon}^d}{\vert K\vert^{\alpha-\varepsilon} 2^{d+1}Q_{d,\frac{1}{3}}} \vert B\vert ^{\alpha -\varepsilon}.
\end{equation}
Recall \eqref{hcont}. Since \eqref{loceq} is valid for any covering $(\widehat{L}_n )_{n\in\mathbb{N}}$ of $A\cap E_{\underline{i}}$, one has
\begin{equation}
\vert B\vert^{\alpha-\varepsilon}\geq \mathcal{H}^{\alpha-\varepsilon}_{\infty}(A)\geq \mathcal{H}^{\alpha-\varepsilon}_{\infty}(A \cap E_{\underline{i}})\geq \frac{\min_{1\leq j\leq m}c_j ^{\alpha-\varepsilon}\rho_{\varepsilon}^d}{\vert K\vert^{\alpha-\varepsilon}2^{d+1}Q_{d,\frac{1}{3}}} \vert B\vert ^{\alpha -\varepsilon}.
\end{equation}
Taking the infimum over all the Borel sets $A\subset B$ satisfying $\mu(A)=\mu(B)$, one gets
$$\vert B\vert^{\alpha-\varepsilon}\geq \mathcal{H}^{\mu,\alpha-\varepsilon}_{\infty}(B)\geq \frac{\min_{1\leq j\leq m}c_j ^{\alpha-\varepsilon}\rho_{\varepsilon}^d}{\vert K\vert^{\alpha-\varepsilon} 2^{d+1}Q_{d,\frac{1}{3}}} \vert B\vert ^{\alpha -\varepsilon}. $$
The results stands for balls of diameter less than $\min_{1\leq j \leq m}c_j $. Then for any ball $B$ centered on $K$ with $\vert B\vert\leq 1$, remarking that
$$\vert B\vert^{\alpha-\varepsilon}\geq \mathcal{H}^{\mu,\alpha-\varepsilon}_{\infty}(B)\geq \mathcal{H}^{\mu,\alpha-\varepsilon}_{\infty}(\min_{1\leq j\leq m}c_j B)$$
and setting $\kappa(d, \mu, \varepsilon)= \frac{\min_{1\leq j\leq m}c_j ^{2(\alpha-\varepsilon)}\rho_{\varepsilon}^d}{\vert K\vert^{\alpha-\varepsilon} 2^{d+1}Q_{d,\frac{1}{3}}}$ yields the desired inequality.
\end{proof}
\begin{remark}
Note that in the proof of Proposition \ref{autosim2}, the estimate of $\mathcal{H}^{\mu,s}_{\infty}(B)$ for $s<\dim (\mu)$ only relies on the absolute continuity of $\mu(f_{\underline{i}}^{-1}(\cdot))$, for any $\underline{i}\in\Lambda^*$. In particular, the same estimates holds for any quasi-Bernoulli measures (which are proved to be exact-dimensional, see \cite{H}).
\end{remark}
This result in hand, one establishes the more general Theorem \ref{contss}.
\begin{proof}[Proof of Theorem \ref{contss}]
Note first, that by item $(5)$ of Proposition \ref{propmuc}, for any $s>\dim(\mu)$ and any set $E$, one has $ \mathcal{H}^{\mu,s}_{\infty}(E)=0.$
\sk
Let us fix $s<\dim (\mu)$ and set $\varepsilon=\dim(\mu)-s >0.$ Since $K\cap \Omega \subset \Omega$ and $\mu(K\cap \Omega)=\mu(\Omega)$, it holds that
$$\mathcal{H}^{\mu,s}_{\infty}(\Omega)\leq \mathcal{H}^{s}_{\infty}(\Omega \cap K).$$
It remains to show that there exists a constant $c( d,\mu,s)$ such that for any open set $\Omega$, the converse inequality
$$c(d,\mu,s)\mathcal{H}^{s}_{\infty}(\Omega \cap K)\leq \mathcal{H}^{\mu,s}_{\infty}(\Omega)$$
holds.
\sk
Let $E \subset \Omega$ be a Borel set such that $\mu(E)=\mu(\Omega)$ and
\begin{equation}
\label{equat0}
\mathcal{H}^{s}_{\infty}(E)\leq 2\mathcal{H}^{\mu,s}_{\infty}(\Omega).
\end{equation}
Let $\left\{L_n \right\}_{n\in\mathbb{N}}$ be a covering of $E$ by balls verifying
\begin{equation}
\label{equat1}
\mathcal{H}^{s}_{\infty}(E)\leq\sum_{n\geq 0}\vert L_n \vert^s \leq 2\mathcal{H}^{s}_{\infty}(E).
\end{equation}
The covering $(L_n)_{n\in\mathbb{N}}$ will be modified to get a covering $(\widetilde{L}_n)_{n\in\mathbb{N}}$ which verifies the following properties:
$\bullet$ $K\cap \Omega\subset \bigcup_{n\in\mathbb{N}}\widetilde{L}_n$;
\mk
$\bullet$ $\bigcup_{n\in\mathbb{N}}L_n \subset \bigcup_{n\in\mathbb{N}}\widetilde{L}_n$;
\mk
$\bullet $ one has $$\sum_{n\geq 0}\vert \widetilde{L}_n \vert^s \leq 8\cdot 2^s\frac{Q_{d,1}^2}{\kappa( d,\mu,\varepsilon)}\sum_{n\geq 0}\vert L_n \vert ^s,
$$
where $\kappa( d,\mu,\varepsilon)$ is the constant introduced in Proposition \ref{autosim2} and $Q_{d,1}$ is the constant arising in Proposition \ref{densibesi} applied with $v=1.$ Last item together with \eqref{equat0} and \eqref{equat1} then immediately imply that
$$\frac{\kappa( d,\mu,\varepsilon)}{8\cdot 2^s Q_{d,1}^2}\mathcal{H}^{s}_{\infty}(K\cap \Omega)\leq \mathcal{H}^{\mu,s}_{\infty}(\Omega),$$
and setting $c( d,\mu,s)=\frac{\kappa( d,\mu,\dim(\mu)-s)}{ 8\cdot2^s Q_{d,1}^2 }$ concludes the proof.
\mk
Let us start the construction of the sequence of balls $(\widetilde{L}_n)_{n\in\mathbb{N}}$. Let $X=( K \setminus \bigcup_{n\in\mathbb{N}}L_n )\cap \Omega$. For every $x\in X$, fix $ 0< r_x \leq 1$ such that $B(x,r_x)\subset \Omega$. One of the following alternatives must occur:
\medskip
\begin{enumerate}
\item for any ball $L_n$ such that $L_n \cap B(x,r_x)\neq \emptyset$, it holds that $\vert L_n \vert \leq r_x$, or \mk
\item there exists $n_x \in\mathbb{N}$ such that $L_{n_x}\cap B(x,r_x)\neq \emptyset$ and $\vert L_{n_x}\vert \geq r_x$.
\end{enumerate}
Consider the set $S_1$ be the set of points in $X$ for which the first alternative holds. By Lemma~\ref{besimodi} applied with $v=1$, it is possible to extract from the covering of $S_1$, $\left\{B(x,r_x),x\in S_1\right\}$, $Q_{d,1}$ families of pairwise disjoint balls, $\mathcal{F}_1 ,...,\mathcal{F}_{Q_{d,1}}$, such that
$$S_1 \subset \bigcup_{1\leq i\leq Q_{d,1}}\bigcup_{L\in\mathcal{F}_i}L.$$
Now, any ball $L_n$ intersecting a ball $L\in \bigcup_{1\leq i\leq Q_{d,1}}\mathcal{F}_i$ must satisfy $\vert L_n \vert \leq |L|.$ In particular, since for any $1\leq i\leq Q_{d,1}$ the balls of $\mathcal{F}_i$ are pairwise disjoint, applying Lemma \ref{dimconst} to the balls of $\mathcal{F}_i$ intersecting $L_n$ we get that $L_n$ intersects at most $Q_{d,1}$ balls of $\mathcal{F}_i$, hence at most $Q_{d,1}^2$ balls of $\bigcup_{1\leq i\leq Q_{d,1}}\mathcal{F}_i$.
\sk
Let $L \in \bigcup_{1\leq i\leq Q_{d,1}}\mathcal{F}_i$. One aims at replacing all the balls $L_n$ intersecting $L$ by the ball $2L$.
For any $1\leq i \leq Q_{d,1}$ and any ball $L\in\mathcal{F}_i$, denote by $\mathcal{G}_L$ the set of balls $L_n$ intersecting $L$. Since $E\subset\bigcup_{n\in\mathbb N} L_n$ and $\mu(E)=\mu(\Omega)$, one has $E\cap L\subset \bigcup_{B\in \mathcal{G}_L}B$ and $\mu(E\cap L)=\mu(L)$. By Definition \ref{mucont} and Proposition \ref{autosim2}, this implies that
\begin{equation}
\label{equat2}
\kappa( d,\mu,\varepsilon)\vert L\vert^s \leq\mathcal{H}^{\mu,s}_{\infty}(L)\leq \sum_{B\in\mathcal{G}_L}\mathcal{H}^{\mu,s}_{\infty}(B)\leq \sum_{B\in\mathcal{G}_L}\vert B\vert^s.
\end{equation}
Replace the balls of $\mathcal{G}_L$ by the ball $\widehat L=2L$ (recall that $\bigcup_{B\in\mathcal{G}_{L}}B \subset 2L$). The new sequence of balls so obtained by the previous construction applied to all the balls $L\in\bigcup_{\leq i\leq Q_{d,1}}\mathcal{F}_i$ is denoted by $(\widehat{L}_k)_{1\le k \le K}$, where $0\le K\le+\infty$.
It follows from the construction and \eqref{equat2} that $S_1 \subset \bigcup_{1\le k\le K}\widehat{L}_k$ and
\begin{equation}
\label{equat3}
\sum_{1\le k\le K}\Big (\frac{\vert \widehat{L}_k \vert}{2}\Big)^s \leq \frac{Q_{d,1}^2}{\kappa(d,\mu,\varepsilon)}\sum_{n\geq 0}\vert L_n \vert^s.
\end{equation}
On the other hand, since for any $x\in S_2 = X\setminus S_1$, there exists $n_x\in\mathbb{N}$ such that ${L}_{n_x}\cap B(x,r_x)\neq \emptyset$ and $r_x \leq \vert {L}_{n_x}\vert$, one has $S_2\subset \bigcup_{n\in\mathbb{N}}2{L}_n$, so that
$$
\Big (\bigcup_{n\in\mathbb N}L_n\Big )\cup \Big (K\cap \Omega \setminus \bigcup_{n\in\mathbb N}L_n\Big ) \subset \Big (\bigcup_{1\le k\le K}\widehat L_k\Big )\cup \Big ( \bigcup_{n\in\mathbb N}2{L}_n\Big ) .
$$
Putting the elements of $(\widehat{L}_k)_{1\le k \le K}$ and $(2L_n)_{n\ge 0}$ in a single sequence $(\widehat{ L}_n)_{n\ge 0}$, writing $(\widetilde{L}_n :=2\widehat{L}_n)_{n\in\mathbb{N}}$, by construction, $K\cap \Omega \subset \bigcup_{n\in\mathbb N}\widetilde{L}_n$ and due to \eqref{equat3}:
$$\mathcal{H}^ {s}_{\infty}(K\cap \Omega)\leq\sum_{n\in\mathbb N}\vert \widetilde{L}_n \vert^s \leq 2^s\Big (\frac{Q_{d,1}^2}{\kappa(d,\mu,\varepsilon)}+1\Big )\sum_{n\in\mathbb N}\vert L_n \vert^s \leq 8\cdot2^s \frac{Q_{d,1}^2}{\kappa( d,\mu,\varepsilon)}\mathcal{H}^{\mu,s}_{\infty}(\Omega) .$$
\end{proof}
\begin{remark}
The proof of Theorem \ref{contss} only uses Proposition \ref{prop-ss}. In particular, Theorem \ref{contss} holds for any measure $\mu \in \mathcal{M}(\mathbb{R}^d)$ supported on $K$ and verifying, for any $\underline{i}\in\Lambda^{*}$, $\mu(f_{\underline{i}}^{-1}(\cdot))$ is absolutely continuous with respect to $\mu$.
\end{remark}
\section{Applications of Theorem \ref{lowani}}
\label{sec-example}
\subsection{Ubiquity Theorems for self-similar measures}
\subsubsection{Proof of Theorem \ref{prop-ss}}
\mk
Let $\mu$ be a self-similar measure with support $K$, and
set $\alpha=\dim(\mu).$ Let $(B_{n}:=B(x_n,r_n))_{n\in\mathbb{N}}$ be a sequence of balls such that $x_n\in K$ for all $n\in\mathbb N$, $\lim_{n\to+\infty} r_{n} = 0$ and $\mu(\limsup_{n\rightarrow+\infty}B_n)=1.$
Fix $\varepsilon>0$, $v>1$ and $\delta \geq 1$ and set $\mathcal{B}_v =\left\{vB_n\right\}_{n\in\mathbb{N}}$.
Lemma~\ref{equiac} shows that $\mathcal{B}_v$ is $\mu$-a.c. Then, by Proposition \ref{autosim2}, for $n$ large enough, one has
\begin{align*}
\mathcal{H}_{\infty}^{\mu,\alpha-\varepsilon}(\widering{(v B_{n} ) ^{\delta}})\ge\kappa(d,\mu,\varepsilon)(vr_{n} )^{\delta(\alpha-\varepsilon)}\geq (v r_{n} )^{\delta(\alpha-\frac{\varepsilon}{2})}.
\end{align*}
Consequently, $$t(\mu,\delta,\varepsilon,\mathcal B_v)=\limsup_{n\rightarrow+\infty}\frac{\log \mathcal{H}_\infty^{\mu,\alpha-\varepsilon}(\widering{(v B_{n} )^{\delta}})}{\delta\log \vert v B_{n} \vert}\leq \alpha-\frac{\varepsilon}{2}$$
so $t(\mu,\delta,\varepsilon,\mathcal B_v)\leq \alpha.$
Due to Corollary~\ref{minoeffec}, one concludes that
$$ \dim_{H}(\limsup_{n\rightarrow+\infty}(vB_n) ^{\delta})\geq \frac{\alpha}{\delta}.$$
But for any $\ep'>0$, $\limsup_{n\rightarrow+\infty}(v B_n)^{\delta}\subset\limsup_{n\rightarrow+\infty}B_n ^{\delta -\ep'}$, so that
$$\dim_{H}(\limsup_{n\rightarrow+\infty}B_n ^{\delta -\ep'})\geq \frac{\alpha}{\delta}.$$
It follows that for any $\ep'>0$ and $\delta\geq 1$ one has
$$ \dim_{H}(\limsup_{n\rightarrow+\infty}B_n ^{\delta})\geq \frac{\alpha}{\delta+\ep'}.$$
Letting $\ep'\to 0$ proves that $\dim_{H}(\limsup_{n\rightarrow+\infty}B_n ^{\delta})\geq \frac{\dim(\mu)}{\delta}$, hence the result.
\begin{remarque}
\label{remarkballopti}
If the sequence of balls $(B_n)_{n\in\mathbb{N}}$ is not assumed to be $\mu$-a.c, but only to verify $\mu(\limsup_{n\rightarrow+\infty}B_n)=1$, then the same lower-bound estimate holds for $\dim_H (\limsup_{n\to\infty} B_n^\delta)$, but the existence of a gauge function as in Theorem \ref{lowani} does not hold in general.
\item[•] Let us also notice that the computation in the proof of Theorem \ref{prop-ss} actually shows that, under the assumption that $\lim_{n\to+\infty}\frac{\log \mu(B_n)}{\log \vert B_n \vert}=\dim (\mu)$, it holds that, for $n$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(B_n ^{\delta})\geq \mu(B_n) \Leftrightarrow s<\frac{\dim (\mu)}{\delta}.$
\end{remarque}
\subsubsection{Proof of Theorem \ref{rectss}}
Given $ \tau_1 =1\leq \tau_2 \leq ...\leq \tau_{d}$ and $s\ge 0$, set $\tau=(\tau_1,\dots,\tau_{d})$ and
$$
g_{\tau }(s)=\max_{1\leq k\leq d}\left\{s\tau_{k}-\sum_{1\leq i\leq k}\tau_k-\tau_i\right\}.
$$
We will need the following lemma (one refers to \cite{KR}, Proposition 2.1 for the proof, although it is stated in terms of singular values functions).
\begin{lemme}
\label{rectc}
Let $ \tau_1 =1\leq \tau_2 \leq ...\leq \tau_{d}$.
The are two positive constants $C_1$ and $C_2$ depending on $d$ only such that for all $s\ge 0$, $r>0$ and $x\in\mathbb{R}^d$ one has
$$
C_1 r^{g_{\tau}(s)}\leq\mathcal{H}^{s}_{\infty}(R_{\tau}(x,r))=\mathcal{H}^{s}_{\infty}(\widering R_{\tau}(x,r))\leq C_2 r^{g_{\tau}(s)}.
$$
\end{lemme}
Recall that $K$ is the closure of its interior, and note that since the weights $p_i $ are taken positive in Definition \ref{def-ssmu}, one must have $\mu(\widering{K})>0.$
Denote $\widetilde{\mu}=\frac{\mu(\cdot)}{\mu(\widering{K})}$ and $\alpha =\dim (\mu)=\dim (\widetilde{\mu}).$ It is easily verified that the computation made in the proof of Theorem \ref{contss} implies that, for any open set $\Omega \subset \widering{K}$, there exists a constant $c(d,\mu,s)$ given by Theorem \ref{contss}, so that
\begin{equation}
\label{contmut}
\begin{cases} c(\mu,d,s)\mathcal{H}^s _{\infty}(\Omega)\leq \mathcal{H}^{\tilde{\mu},s}_{\infty}(\Omega)\leq \mathcal{H}^{s}_{\infty}(\Omega)\text{ if } s<\alpha\\
\mathcal{H}^{\tilde{\mu},s}_{\infty}(\Omega)=0\text{ if } s>\alpha.
\end{cases}
\end{equation}
Also, $\widetilde{\mu}$ being absolutely continuous with respect to $\mu$, the sequence $(B_n)_{n\in\mathbb{N}}$ is $\widetilde{\mu}$-a.c. Furthermore, up to a $\widetilde{\mu}$-a.c extraction, we can assume that each ball $(B_n)_{n\in \mathbb{N}}$ is included in $\widering{K}$ (and we will do so).
Let $\varepsilon>0$. Set $\mathcal{R}=\left\{R_n\right\}_{n\geq 0}.$
By Lemma \ref{gscainf}, up to a $\widetilde{\mu}$-a.c extraction, one can assume that for every $n\in\mathbb{N}$, the ball $B_n$ satisfies
$$\widetilde{\mu}(B_n)\leq r_n ^{\alpha-\varepsilon}.$$
Setting $\tau^{\prime}=(\frac{\tau_i}{\tau_1})_{1\leq i\leq d}$, for all $0\le s<\alpha-\varepsilon$, one has $$g_{\tau^{\prime}}(s)=\max_{1\leq k\leq d}\left\{\frac{s\tau_k -\sum_{1\leq i\leq k}\tau_k-\tau_i}{\tau_1}\right\}.$$
From equation \eqref{contmut} and Lemma \ref{rectc}, one deduces that
\begin{equation}
C_1 c(d,\mu,s) r_n^{\tau_1 g_{\tau^{\prime}}(s)}\leq \mathcal{H}^{\tilde{\mu},s}_{\infty}(R_n).
\end{equation}
In particular, for any $s$ verifying
\begin{equation}
\label{equab}
\tau_1 g_{\tau^{\prime}}(s)\leq \alpha-\frac{\varepsilon}{2},
\end{equation}
if $r_n\le 1$ one has
$$
C_1 c(d,\mu,s) r_n ^{\alpha-\frac{1}{2}\varepsilon}\leq C_1 c(d,\mu,s) r_n^{ \tau_1 g_{\tau^{\prime}}(s)}\leq \mathcal{H}^{\tilde{\mu},s}_{\infty}(R_n).
$$
Since $r_n \to 0$, for $n$ large enough, this yields
\begin{equation}
\label{equaa}
\widetilde{\mu}(B_n)\leq r_n ^{\alpha-\varepsilon}\leq C_1 c(d,\mu,s) r_n^{\tau_1 g_{\tau^{\prime}}(s)}\leq \mathcal{H}^{\tilde{\mu},s}_{\infty}(R_n),
\end{equation}
hence $(B_n)_{n\in\mathcal{N}_{\tilde\mu}(\mathcal{B},\mathcal{R},s)}$ is $\widetilde \mu$-a.c., and $s(\widetilde{\mu}, \mathcal{R},\mathcal{B})\geq s$.
It remains to note that
\begin{align}
\label{shtrou}
\eqref{equab}&\Leftrightarrow \max_{1\leq k\leq d}\left\{\frac{s\tau_k -\sum_{1\leq i\leq k}\tau_k-\tau_i}{\tau_1}\right\}\leq \frac{\alpha-\frac{\varepsilon}{2}}{\tau_1}\nonumber \\
& \Leftrightarrow \forall 1\leq k\leq d, \ \frac{s\tau_k -\sum_{1\leq i\leq k}\tau_k-\tau_i}{\tau_1}\leq \frac{\alpha-\frac{\varepsilon}{2}}{\tau_1}\nonumber\\
&\Leftrightarrow \forall 1\leq k\leq d, \ s\leq \frac{\alpha-\frac{1}{2}\varepsilon+\sum_{1\leq i\leq k}\tau_k -\tau_i}{\tau_k}\nonumber\\
& \Leftrightarrow s\leq \min_{1\leq k\leq d}\left\{\frac{\alpha-\frac{1}{2}\varepsilon+\sum_{1\leq i\leq k}\tau_k -\tau_i}{\tau_k}\right\}.
\end{align}
Since $\varepsilon>0$ was arbitrary, this implies that $$s(\widetilde{\mu}, \mathcal{R},\mathcal{B})\geq \min_{1\leq k\leq d}\left\{\frac{\alpha+\sum_{1\leq i\leq k}\tau_k -\tau_i}{\tau_k}\right\},$$
and applying Theorem \ref{lowani} gives the desired lower bound estimate.
\begin{remarque}
\label{remarkrect}
Note that the estimates made in the proof of Theorem \ref{rectss}, together with Lemma~\ref{rectc}, can be used to show that, under the assumption that $\lim_{n\to\infty}\frac{\log \mu(B_n )}{\log \vert B_n \vert}=\dim (\mu)$, one has the following properties:
If $s <\min_{1\leq k\leq d}\left\{\frac{\dim (\mu)+\sum_{1\leq i\leq k}\tau_k -\tau_i}{\tau_k}\right\}$ then, for $n\in\mathbb{N}$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(R_n)\geq \mu (B_n)$. If $s>\min_{1\leq k\leq d}\left\{\frac{\dim (\mu)+\sum_{1\leq i\leq k}\tau_k -\tau_i}{\tau_k}\right\}$ then, for $n$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(R_n)\leq \mu (B_n).$
\end{remarque}
\subsection{Application to self-similar shrinking targets}
\begin{proof}
In this section, Theorem \ref{shrtag} is proved and we adopt the notation of the proof of Proposition \ref{autosim2}.
Set $s=\dim_H (K)$. Note that, for each $k\in\mathbb{N}$, the set $\left\{B(f_{\underline{i}}(x),2\vert K\vert c_{\underline{i}})\right\}_{\underline{i}\in \Lambda^{k}}$ covers $K.$ In particular, for any measure $\mu$ supported on $K$, the family $\left\{B(f_{\underline{i}}(x),2\vert K\vert c_{\underline{i}})\right\}_{\underline{i}\in \Sigma^*}$ is $\mu$-a.c.
Set $B_{\underline{i}}=B(f_{\underline{i}}(x),2\vert K\vert c_{\underline{i}}).$ One now focuses on proving that, for any $\delta\geq 1$, $\dim_H (\limsup_{\underline{i}\in\Sigma^*}B_{\underline{i}}^{\delta})=\frac{s}{\delta}.$ If this holds, since for any $\varepsilon>0$,
\begin{equation}
\label{enca1}
\limsup_{\underline{i}\in\Sigma^{*}}B_{\underline{i}}^{\delta+\varepsilon}\subset\limsup_{\underline{i}\in \Sigma^*}B(f_{\underline{i}}(x), c_{\underline{i}}^{\delta})\subset \limsup_{\underline{i}\in\Sigma^{*}}B_{\underline{i}}^{\delta},
\end{equation}
it also holds that $\dim_H (\limsup_{\underline{i}\in \Sigma^*}B(f_{\underline{i}}(x), c_{\underline{i}}^{\delta}))=\frac{s}{\delta}.$
Note that $s$ satisfies the equation $\sum_{1\leq i\leq m}c_i ^s =1.$ Let also be $\nu_s$, the measure on $(\Sigma, \mathcal{B}(\Sigma))$ associated with the probability vector $(p_i =c_i ^s )_{1\leq i\leq m}$ and $\mu_s$ its projection on $K$ by the canonical coding map.
Let $0<t<\min_{1\leq i\leq m}c_i$ and
$$\Lambda ^{\left(k\right)}_t=\left\{\underline{i}=\left(i_1 ,\ldots, i_{\ell}\right) \in\Lambda^* : \ c_{i_{\ell}}t^k< c_{\underline{i}}\leq t^k \right\}.
$$
If $\underline{i}\in \Lambda^{(k)}_t$, then for any $\ell \in \Lambda$, the word $\underline{i}\ell \notin \Lambda^{(k)}_t .$ This implies that for any $\underline{i}\neq \underline{j}\in \Lambda^{(k)}_t$, $[\underline{i}]\cap [\underline{j}]=\emptyset.$
Then for any $\delta \geq 1$, one has, for any $\varepsilon>0$,
\begin{align*}
\sum_{k\geq 0}\sum_{\underline{i}\in\Lambda^{\left(k\right)}_t} \left(\vert B_{\underline{i}}\vert^{\delta}\right)^{\frac{s+\varepsilon}{\delta}}&=\left(4\vert K\vert\right)^{s+\varepsilon}\sum_{k\geq 0}\sum_{\underline{i}\in\Lambda^{\left(k\right)}_t} c_{\underline{i}}^{s+\varepsilon}\\
&\leq \left(4\vert K\vert\right)^{s+\varepsilon}\sum_{k\geq 0}\sum_{\underline{i}\in\Lambda^{\left(k\right)}_t}t^{k\varepsilon}\nu_s\left([\underline{i}]\right).
\end{align*}
Since $\sum_{\underline{i}\in\Lambda^{\left(k\right)}_t}\nu_s\left(\underline{i}\right)\leq 1$, one obtains
\begin{align*}
\sum_{k\geq 0}\sum_{\underline{i}\in\Lambda^{\left(k\right)}_t} \left(\vert B_{\underline{i}}\vert^{\delta}\right)^{\frac{s+\varepsilon}{\delta}}\leq \left(4\vert K\vert\right)^{s+\varepsilon}\sum_{k\geq 0}t^{k\varepsilon}<+\infty.
\end{align*}
This shows that
\begin{equation}
\label{majoshri}
\displaystyle \dim_H \Big(\limsup_{\underline{i}\in \Lambda^*}B_{\underline{i}}^{\delta}\Big)\leq \frac{s}{\delta}.
\end{equation}
One now establishes the lower-bound estimate. By the dimension regularity assumption (see Definition \ref{dimreg}), $\dim_H (\mu_s)=s.$ Since $(B_{\underline{i}})_{\underline{i}\in\Sigma^*}$ is $\mu_s$-a.c, Theorem~\ref{prop-ss} yields $\dim_H (\limsup_{\underline{i}\in\Sigma^{*}}B_{\underline{i}}^{\delta})\geq \frac{s}{\delta}$.
\end{proof}
\subsection{Study of a problem related to a question of Mahler}
Let us first notice that by Theorem \ref{approxrat}, one has $$\dim_{H}\Big (\limsup_{B\in\mathcal Q}B^\delta\cap K_{1/3}^{(0)}\Big )\leq \min\left\{\frac{1}{\delta},\frac{\log 2}{\log 3}\right\}.$$
In particular, this proves that the expected upper-bound in Theorem \ref{main} stands.
\mk
Before showing that the lower-bound also holds, let us start with some facts and remarks.
\mk
\begin{remarque}
$\bullet$ One has $\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(K_{1/3})>0$ (this is well known and easily follows from the fact that $K_{1/3}$ carries an Alfhors regular measure of dimension $\frac{\log 2}{\log 3}$).
Moreover, for every $k\in\mathbb{N}$, setting $\mathcal{K}_k =\left\{ f_{\underline{i}}([0,1]) \right\}_{\underline{i}\in\Lambda^k}$, one has
\begin{equation}
\label{eqegal1}
1=\sum_{I \in\mathcal{K}_k}\vert I\vert^{\frac{\log 2}{\log 3}}.
\end{equation}
\mk
$\bullet$ For every $k\in\mathbb{N},$ let us define
\begin{equation}
\label{defomega}
\Omega_{k}=\bigcup_{I\in\mathcal{K}_k}\overset{\circ}{I}.
\end{equation}
Since $\mathcal{H}^{\frac{\log2}{\log 3}}_{\infty}(\bigcup_{I\in\mathcal{K}_k}I \setminus \Omega_k )=0$ (it is a finite set of points), it follows from \eqref{eqegal1} that
\begin{equation}
\label{contom}
C\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(\Omega_{k})\leq\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(K_{1/3})\leq \mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}\Big (\bigcup_{I\in\mathcal{K}_k}I\Big )=\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(\Omega_{k})\le 1,
\end{equation}
with $C=\mathcal{H}^{\frac{\log 2}{ \log 3}}_{\infty}(K_{1/3})>0$.
\mk
$\bullet$ If $n\in\mathbb{N}$ and $T\in\mathcal{T}_n =\left\{[\frac{k}{3^{n}},\frac{k+1}{3^n}[,0\leq k\leq 3^n-1 \right\}$ is a triadic interval of generation $n$, denote by $F_T$ the canonical homothetical mapping which sends $[0,1]$ to $ \overline{T}.$ For every $I\in\bigcup_{J\in\mathcal{K}_k}F_T (J)$, for all $n\leq k^{\prime}\leq n+k$ and all $x=(x_n)_{n\in\mathbb{N}}\in\Sigma$ such that $\pi (x)\in I$, one has
\begin{equation}
\label{equafre1}
S_{n+k^{\prime}}\phi(x)=S_{n}\phi (x)\times \frac{n}{n+k^{\prime}}.
\end{equation}
\end{remarque}
\mk
One are now ready to finish the proof of Theorem \ref{main}.
\mk
Let $(\varepsilon_{q} )_{q\in\mathbb{N}}$ be a positive sequence such that $\lim_{q\to \infty}\varepsilon_{q}= 0.$ One constructs a family $\{U_{p,q,\delta}\}_{\delta\ge 1,\, q\in\mathbb{N},\, 0\leq p\leq q}$ of open sets as follows: Let $\delta\geq 1$, $q\in\mathbb{N}^{*}$ and $0\leq p\leq q$. Consider $T$ a triadic interval of generation $n_q=\lfloor \log_{3} (q^{2\delta})\rfloor +1$ included in $B(\frac{p}{q},q^{-2\delta}).$ Let $N_{p,q,\delta}$ be large enough to ensure that for any $x\in\Sigma$ verifying $\pi(x)\in T$, one has
\begin{equation}
\label{equafre2}
S_{n_q}\phi(x)\times \frac{n_q}{n_q+N_{p,q,\delta}}\leq \varepsilon_{q}.
\end{equation}
Set
\begin{equation}
\label{defu}
U_{p,q,\delta}=F_{T}(\Omega_{N_{p,q,\delta}}).
\end{equation}
By \eqref{equafre1} and \eqref{equafre2}, for all $x\in U_{p,q, \delta}$ one has
$$
S_{n_q+N_{p,q}}\phi(x)\leq \varepsilon_{q}.
$$
This implies that $\bigcap_{Q\ge 1}\bigcup_{q\ge Q}\bigcup_{0\le p\le q} U_{p,q, \delta}\subset K_{1/3}^{(0)}\cap \bigcap_{Q\ge 1}\bigcup_{q\ge Q}\bigcup_{0\le p\le q}B(\frac{p}{q},q^{-2\delta})$.
Since $U_{p,q,\delta}$ is an homothetic copy of $\Omega_{N_{p,q,\delta}}$ (see \eqref{defu}), by \eqref{contom}, due to the choice of $n_q$ there exists $\widetilde{C}>0$ independent of $p$, $q$ and $\delta$ such that
\begin{equation}
\label{hcontc}
\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(U_{p,q, \delta})\geq\widetilde{C} q^{-2\delta \frac{\log(2)}{\log (3)}}.
\end{equation}
For $1\leq \delta\leq \frac{\log 3}{\log 2}$, it follows that
\begin{equation}
\label{hcont1}
\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(U_{p,q, \delta})\geq \widetilde{C} q^{-2}=\mathcal{L} \Big(B\Big(\frac{p}{q},q^{-2}\Big)\Big).
\end{equation}
For $\delta \geq \frac{\log 3}{\log 2}$, by concavity of $x\mapsto x^{\frac{\log 3}{\delta \log 2}},$
\begin{align}
\label{hcont2}
\mathcal{H}^{\frac{1}{\delta}}_{\infty}(U_{p,q, \delta})\geq (\mathcal{H}^{\frac{\log 2}{\log 3}}_{\infty}(U_{p,q, \delta}))^{\frac{\log 3}{\delta \log 2}}\geq \widetilde{C} (q^{-2\delta \frac{\log 2}{\log 3}})^{ \frac{\log 3}{\delta\log 2}}=\widetilde{C}\mathcal{L} \Big(B\Big(\frac{p}{q},q^{-2}\Big)\Big).
\end{align}
By Theorem \ref{lowani} (or by Rams-Koivusalo's Theorem \ref{lowkr}) applied to $\mathcal{Q}=(B(\frac{p}{q},\frac{1}{q^2}))_{q\in\mathbb{N}^* ,0\leq p\leq q}$, $\mathcal{U}=(U_{p,q,\delta})_{q\in\mathbb{N}^* ,0\leq p\leq q}$ and the Lebesgue measure, one gets
\begin{align*}
&\dim_H \Big (\bigcap_{Q\ge 1}\bigcup_{q\ge Q}\bigcup_{0\le p\le q} U_{p,q, \delta}\Big )\geq \frac{\log 2}{\log 3} &\text{ if }&1\leq \delta \leq \frac{\log 3}{\log 2}\\
&\dim_H \Big (\bigcap_{Q\ge 1}\bigcup_{q\ge Q}\bigcup_{0\le p\le q} U_{p,q, \delta}\Big )\geq \frac{1}{\delta}&\text{ if }&\delta\geq \frac{\log 3}{\log 2}.
\end{align*}
\bibliographystyle{plain}
|
\section{Introduction}
The hydrodynamic model was first derived by Bl{\o}tekj{\ae}r \cite{Bl70} for electrons in a semiconductor. After appropriate simplification the one-dimensional time-dependent system in the isentropic case reads:
\begin{equation}\label{1.1}
\left\{
\begin{split}
&n_t+(nu)_x=0,\\
&(nu)_t+\left(nu^2+p(n)\right)_x=nE-\frac{nu}{\tau },\\
&E_x=n-b(x),
\end{split}
\right.
\end{equation}
where $n(x,t)$, $u(x,t)$ and $E(x,t)$ denote the electron density, velocity, and electric field respectively. The given function $p=p(n)$ is the pressure-density relation on which a commonly used hypothesis is
\begin{equation*}
p(n)=Tn^\gamma,
\end{equation*}
where $T>0$ is Boltzmann's constant and $\gamma\ge1$ is the adiabatic exponent. The constant parameter $\tau>0$ is the momentum relaxation time. The given background density $b(x)>0$ is called the doping profile. The hydrodynamic model \eqref{1.1} is also called Euler-Poisson equations with semiconductor effect. For more details we refer to treatises \cite{MRS90,ZH16} and references therein.
In this paper, the main focus is on the isothermal steady-state flows satisfying equations
\begin{equation}\label{hu17}
\left\{
\begin{split}
&J\equiv\text{constant},\\
&\left(\frac{J^2}{n}+p(n)\right)_x=nE-\frac{J}{\tau},\\
&E_x=n-b(x),
\end{split}
\right.
\end{equation}
where $J=nu$ stands for the current density, and $p(n)=Tn$ corresponds to the isothermal ansatz. By the terminology from gas dynamics, we call $c:=\sqrt{P'(n)}=\sqrt T>0$ the speed of sound. The flow is referred to as subsonic, sonic or supersonic provided the velocity satisfies
\begin{equation}\label{1.5}
u<c,\quad u=c\quad\text{or}\quad u>c,\quad\text{respectively}.
\end{equation}
For convenience of notation, we introduce
\begin{equation}\label{1.2*}
\alpha=\frac{1}{\tau},~~\mbox{the reciprocal of the momentum relaxation time}.
\end{equation}
Without loss of generality, we set
\begin{equation}\label{z}
T=1 \mbox{ and } J=1,
\end{equation}
thus the system \eqref{hu17} is equivalently reduced to the system
\begin{equation}\label{1.6}
\left\{
\begin{split}
&\left(1-\frac{1}{n^2}\right)n_x=nE-\alpha,\\
&E_x=n-b(x).
\end{split}
\right.
\end{equation}
From \eqref{1.5} and \eqref{z}, it is easy to see that the flow is subsonic if $n>1$, sonic if $n=1$, or supersonic if $0<n<1$. By virtue of \eqref{1.2*}, we call the system \eqref{1.6} the Euler-Poisson equations with the semiconductor effect if $\alpha>0$, and without the semiconductor effect if $\alpha=0$, respectively. Throughout this paper, we are interested in the system \eqref{1.6} in the open interval $(0,1)$, which is subjected to the sonic boundary condition
\begin{equation}\label{1.7}
n(0)=n(1)=1.
\end{equation}
We also assume that the doping profile $b(x)$ is of class $C[0,1]$, satisfying the subsonic condition $b(x)>1$ on $[0,1]$. For simplicity of notation, its infimum and supremum over $[0,1]$ is denoted by
\[
\underline{b}:=\inf_{x\in[0,1]}b(x) \mbox{ and } \bar{b}:=\sup_{x\in[0,1]}b(x),
\]
respectively.
Over the past three decades, major advances in the mathematical theory of steady-state Euler-Poisson equations with/without the semiconductor effect have been made by many authors. In what follows, we just list several results which are closely linked to the present paper.
For the purely subsonic steady-state flows, in 1990, Degond
\textit{et al.}\cite{DM90} first proved the existence of the subsonic solution
to the one-dimensional steady-state Euler-Poisson with the
semiconductor effect when its boundary states belongs to the
subsonic region. Subsequently, Degond \textit{et al.}\cite{DM93} further
showed the existence and local uniqueness of irrotational subsonic
flows to the three-dimensional steady-state semiconductor
hydrodynamic model under a smallness assumptions on the data. Along
this line of research, the steady-state subsonic flows with and
without the semiconductor effect were investigated in various
physical boundary conditions and different dimensions
\cite{FI97,GS05,NS07,BDX14}. As for the purely supersonic
steady-state flows, Peng \textit{et al.}\cite{PV06} established the existence
and uniqueness of the supersonic solutions with the semiconductor
effect, which correspond to a large current density.
Note that the system \eqref{hu17} or \eqref{1.6} will be degenerate
at the sonic state, thus the study on the transonic solutions and
various steady states satisfying the sonic boundary condition
becomes very difficult. Ascher \textit{et al.}\cite{AMPS91} first examined the
existence of the transonic solution to the one-dimensional
isentropic Euler-Poisson equations without and with the
semiconductor effect when the doping profile is a supersonic
constant, and then Rosini \cite{Ro05} extended this work to the
non-isentropic case by the analysis of phase plane. When the doping
profile is non-constant, Gamba \cite{Ga92,GM96} investigated the
one-dimensional and two-dimensional transonic solutions with shocks,
respectively. However, these transonic solutions yield boundary
layers because they are constructed as the limits of vanishing
viscosity. Luo \textit{et al.}\cite{LX12,LRXX11} further considered the
one-dimensional Euler-Poisson equations without the semiconductor
effect, under the restriction that boundary data are far from the
sonic state and the doping profile is either a subsonic constant or
a supersonic constant, a comprehensive analysis on the structure and
classification of steady states was carried out in \cite{LX12} by
using the analysis of phase plane. Meanwhile, both structural and
dynamical stability of steady transonic shock solutions was obtained
in \cite{LRXX11}.
What if the sonic state appears in the solutions? As we have seen,
all the existing works introduced above cannot answer this question.
Even the works regarding transonic shocks cannot radically answer it
either because the two different phase states are connected by the
jump of shocks satisfying the Rankine-Hugoniot condition and
entropy condition, avoiding the degeneracy caused by the sonic
state. Recently, Li \textit{et al.}\cite{LMZZ17,LMZZ18} systematically
explored the critical case, that is, the one-dimensional
semiconductor Euler-Poisson equations with the sonic boundary
condition. The existence, nonexistence and classification of all
types of physical steady states to this critical boundary-value
problem was obtained for the subsonic doping profile in
\cite{LMZZ17} and supersonic doping profile in \cite{LMZZ18}. More
precisely, in \cite{LMZZ17}, the authors proved that the critical
boundary-value problem admits a unique subsonic solution, at least
one supersonic solution, infinitely many transonic shocks if
$\alpha\ll1$, and infinitely many transonic $C^1$-smooth solutions
if $\alpha\gg1$; in \cite{LMZZ18}, the authors showed the
nonexistence of all types of physical steady states to the critical
boundary-value problem assuming that the doping profile is small
enough and $\alpha\gg1$, and they also discussed the existence of
supersonic and transonic shock solutions under the hypothesis that
the doping profile is close to the sonic state and $\alpha\ll1$.
Inspired by the groundbreaking works \cite{LMZZ17,LMZZ18}, there is
a series of interesting generalizations into the transonic doping
profile case in \cite{CMZZ20}, the case of transonic
$C^\infty$-smooth steady states in \cite{WMZZ21}, the
multi-dimensional cases in \cite{CMZZ21,CMZZ22}, and even the
bipolar case \cite{MMZ20}.
Compared with the existence theory in the critical case,
there are rather few works on the study of stability (both structural
and dynamical stability) in the critical case due to the interior
or boundary degeneracy. To the best of our knowledge,
Feng \textit{et al.}\cite{FMZ22} first demonstrated the structural
stability of the $C^1$-smooth transonic steady states with
respect to the small perturbation of both the supersonic
constant doping profiles and non-degenerate boundary data.
In \cite{FMZ22}, for supersonic doping profile, the authors
also discussed the structural stability and linear dynamic
instability of the transonic steady states with the non-degenerate
boundary data under some appropriate hypotheses. Thus far, the most
difficult part of the critical case, namely, the problem about the
structural and dynamical stability of all types of physical solutions
with the sonic boundary condition for the non-constant subsonic
doping profile (for the existence theory, see \cite{LMZZ17}),
is still open. To thoroughly solve this problem is full of challenges,
owing to the boundary degeneracy. However, in the present paper,
we intend to shed new light on this problem.
The purpose of this paper is to show that interior subsonic
solutions to the system \eqref{1.6} with the sonic boundary
condition \eqref{1.7} are structurally stable if we propose
a monotonicity restriction on subsonic doping profiles (see Theorem \ref{T2}).
This paper is organized as follows. Some necessary preliminaries
and the main result are stated in Section \ref{s2}. The proof
of the main result, Theorem \ref{T2}, is given in Section \ref{s3}.
\section{Preliminaries and the main result}\label{s2}
In this section we shall present the main result.
Before proceeding, we first give the important preliminaries
from the foregoing research \cite{LMZZ17}. First of all,
we recall the definition of the interior subsonic solution.
\begin{defi}\label{def}
We say a pair of functions $(n,E)(x)$ is an interior subsonic
solution of the boundary value problem \eqref{1.6}\&\eqref{1.7}
provided \textup{(i)} $(n-1)^2\in H_0^1(0,1)$, \textup{(ii)} $n(x)>1$,
for all $x\in(0,1)$, \textup{(iii)} $n(0)=n(1)=1$, \textup{(iv)}
the following equality holds for all test functions $\varphi\in H_0^1(0,1)$,
\begin{equation}\label{1.9}
\int_0^1\left(\frac{1}{n}-\frac{1}{n^3}\right)n_x\varphi_xdx+\alpha\int_0^1\frac{\varphi _x}{n}dx+\int_0^1\left(n-b\right)\varphi dx=0,
\end{equation}
and \textup{(v)} $E(x)$ is given by
\begin{equation}\label{E}
E(x)=\alpha +\int_0^x\left(n(y)-b(y)\right)dy.
\end{equation}
\end{defi}
In addition, we continue to recall the existence and uniqueness
of interior subsonic solutions, which is excerpted from the first
part of Theorem 1.3 in \cite{LMZZ17}.
\begin{prop}[Existence theory in \cite{LMZZ17}]\label{T1}
Suppose that the doping profile $b\in L^\infty(0,1)$ is subsonic
such that $\underline{b}>1$. Then for any $\alpha\in[0,\infty)$
the boundary value problem \eqref{1.6}\&\eqref{1.7} admits a unique
interior subsonic solution $(n,E)\in C^{\frac{1}{2}}[0,1]\times H^1(0, 1)$
satisfying the boundedness
\begin{equation}\label{ulb}
1+m\sin(\pi x)\le n(x)\le\bar{b}, \quad x\in [0,1],
\end{equation}
and the boundary behavior at endpoints
\begin{equation}\label{E(1)}
E(0)=\alpha,\quad E(1)<\alpha,
\end{equation}
\begin{equation}\label{L14}
\left\{
\begin{split}
& C_1(1-x)^\frac{1}{2}\le n(x)-1\le C_2(1-x)^\frac{1}{2},\\
& -C_3(1-x)^{-\frac{1}{2}}\le n_x(x)\le-C_4(1-x)^{-\frac{1}{2}},
\end{split}
\right. \quad\text{for}\ x\ \text{near}\ 1,
\end{equation}
where $m=m(\alpha, \underline{b})>0$, $C_2>C_1>0$ and $C_3>C_4>0$ are
certain uniform estimate constants.
\end{prop}
\begin{rem}
Note that the degeneracy of the boundary value problem
\eqref{1.6}\&\eqref{1.7} occurs merely on the boundary.
Thus, if we assume that the doping profile has relatively
higher-order regularity, say $b\in C[0,1]$, then by virtue
of the standard theory for elliptic interior regularity and
Sobolev's embedding theorem, the corresponding interior
subsonic solution $(n,E)$ is actually of class
$\left(C^1(0,1)\cap C^{\frac{1}{2}}[0,1]\right) \times C^1[0,1]$.
This fact will be tacitly exploited hereafter.
\end{rem}
We are now in a position to formulate the main result in the present paper.
\begin{theo}[Structural stability of interior subsonic solutions]\label{T2}
Assume that $b_1, b_2\in C[0,1]$ are subsonic such that $b_1(x)\ge b_2(x)>1$
for all $x\in[0,1]$. For $i=1,2$, let $(n_i,E_i)(x)$ denote the
interior subsonic solution relative to the doping profile $b_i$,
respectively. Then the two interior subsonic solutions are
structurally stable to one another in the sense that
\begin{equation}\label{1.10**}
\|n_1-n_2\|_{C[0,1]}+\|(1-x)^{\frac{1}{2}}(n_1-n_2)_x\|_{C[0,1]}+\|E_1-E_2\|_{C^1[0,1]}\le C\|b_1-b_2\|_{C[0,1]},
\end{equation}
where $C>0$ is a certain constant independent of $\|b_1-b_2\|_{C[0,1]}$.
\end{theo}
We conclude this section with a brief sketch of the strategy that
underlies the proof of our main result. Due to the boundary
degeneracy of interior subsonic solutions, the study of their
globally structural stability over the entire interval $[0,1]$
becomes sophisticated and challenging. Therefore, we shall have
to divide the whole interval $[0,1]$ into three domains as follows:
\begin{equation*}
[0,1]=[0,\delta)\cup[\delta,1-\delta]\cup(1-\delta,\delta],
\end{equation*}
where the intrinsic segmentation constant $\delta>0$ would be
appropriately determined (see Lemma \ref{L5}), and we will also
have to establish structural stability estimates separately on
their respective domains in the following order: (i) near the
left endpoint $x=0$; (ii) near the right endpoint $x=1$; (iii)
on the middle domain. This strategy is feasible because we have
discovered the following facts:
\begin{enumerate}
\item the local singularity analysis reveals that the plausible
singularity at the left endpoint $x=0$ is removable (see Lemma \ref{L1});
based on this, we are able to establish the local structural stability
estimate on an inherent neighborhood $[0,\delta_0)$ by the
monotonicity argument. The main point is that both the radius
$\delta_0$ and the positive estimate constant are independent
of $\|b_1-b_2\|_{C[0,1]}$ (see Lemma \ref{L3}). This is the
reason why this type of neighborhood is referred to as ``to be intrinsic
or inherent''. This sort of tacit convention will be used throughout
the present paper.
\item the local weighted singularity analysis discloses that the
genuine singularity at the right endpoint $x=1$ can be well controlled
by the $(1-x)^{\frac{1}{2}}$-weight (see Lemma \ref{L4-1}); thus,
the monotonicity argument further ensures that the local weighted
structural stability holds on an intrinsic neighborhood
$(1-\delta_1,1]$ (see Lemma \ref{L4}).
\item the remaining part constitutes the middle domain, which is
regular as to the structural stability (see Lemma \ref{L5}).
\end{enumerate}
It is worth mentioning that the monotonicity argument has been playing
a crucial role in establishing structural stability estimates near
both endpoints. The principle behind the monotonicity argument
is given by Lemma \ref{Lcp}. From a technical point of view,
the monotonicity argument is useful, but only at the cost of
adding an extra restriction $b_1(x)\ge b_2(x)$ on $[0,1]$.
How to get rid of this restriction is a tough question, and
we will explore it in the future study.
\section{Proof of Theorem \ref{T2}}\label{s3}
This section is devoted to proving our main result. In order to make
the line of reasoning accessible to the reader,
the proof will be
divided into a sequence of lemmas.
We let $(n_i,E_i)(x)$ denote the interior subsonic solution
corresponding to the subsonic doping profile $b_i(x)>1$, satisfying
the sonic boundary value problem
\begin{equation}\label{1}
\left\{
\begin{split}
&\left(1-\frac{1}{n_i^2}\right)n_{ix}=n_iE_i-\alpha,\\
&E_{ix}=n_i-b_i(x), \quad x\in (0,1), \\
&n_i(0)=n_i(1)=1,
\end{split}
\right. \quad\text{for}\ i=1,2, \ \text{respectively}.
\end{equation}
First of all, we adapt the comparison principle in
\cite{LMZZ17}(Lemma 2.2, P4773) for use with two doping profiles
and their corresponding interior subsonic solutions, which is the
basis of the monotonicity argument in studying the structural
stability near endpoints.
\vspace{1.5mm}
\begin{lem}[Comparison principle]\label{Lcp}
Let the doping profiles $b_1, b_2\in C[0,1]$.
If $b_1(x)\ge b_2(x)>1$ on $[0,1]$. Then
\begin{equation}\label{1.10*}
n_1(x)\ge n_2(x),\ \text{on}\ [0,1].
\end{equation}
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
According to the relevant arguments from \cite{LMZZ17}(Equation (17), P4773),
for $i=1,2$, since $(n_i,E_i)$ is the interior subsonic solution,
thereby having the approximate solution sequence
$\{n_{ij}\}_{0<j<1}\subset C^1[0,1]$ satisfying the weak form
\begin{equation}\label{wf}
\int_0^1 A(n_{ij}, n_{ijx})\varphi_xdx+\int_0^1 (n_{ij}-b_i)\varphi dx=0,\quad\forall\varphi\in H_0^1(0,1),
\end{equation}
where
\begin{equation*}\label{1.12}
A(z,p):=\left(\frac{1}{z}-\frac{j^2}{z^3}\right)p+\alpha\frac{j}{z}.
\end{equation*}
Subtracting $\eqref{wf}|_{i=1}$ from $\eqref{wf}|_{i=2}$,
for all nonnegative test functions $\varphi\in H_0^1(0,1)$, we have
\begin{equation}\label{1.13}
\int_0^1 \left(A(n_{2j},n_{2jx})-A(n_{1j},n_{1jx})\right)\varphi_xdx+\int_0^1 \left(n_{2j}-n_{1j}\right)\varphi dx=\int_0^1 \left(b_2-b_1\right)\varphi dx\le0,
\end{equation}
where we have used the assumption that $b_1(x)\ge b_2(x)$ on $[0,1]$
in the last inequality. This is exactly the crucial Equation (19) in
\cite{LMZZ17}, the same result therefore applies to \eqref{1.13}
provided we simply imitate the remaining arguments in Lemma 2.2
of \cite{LMZZ17}. That is,
\begin{equation}\label{1.14}
n_{1j}(x)\ge n_{2j}(x),\ \text{on}\ [0,1], \ \text{for}\ 0<j<1.
\end{equation}
Now the monotonicity relation \eqref{Lcp} follows after a passage to
the limit as $j\to1^-$ on both sides of the inequality \eqref{1.14}.\qed
\vspace{3mm}
In addition, for $i=1,2$, we set about analyzing the boundary behavior of
the first-order derivative of $n_{i}(x)$ at the left endpoint $x=0$.
It seems plausible that the singularity should have appeared there,
as a matter of fact this ``fake'' singularity at $x=0$ is removable
because of $E_i(0)=\alpha$.
\vspace{1.5mm}
\begin{lem}\label{L1}
Suppose that $b_i, i=1,2$ satisfy the same conditions in Lemma \ref{Lcp},
and $\alpha\ge2\sqrt{2}\max\{\sqrt{b_1(0)-1}, \sqrt{b_2(0)-1}\}$. Then
\begin{equation}\label{3}
\lim_{x\to0^+}n_{ix}(x)=\frac{1}{4}\left(\alpha-\sqrt{\alpha^2-8\left(b_i(0)-1\right)}\right)=:A_i>0,\quad i=1,2.
\end{equation}
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
In much the same way as in \cite{LMZZ17}(Theorem 5.6, P4802),
owing to $n_i(0)=1$ and $E_i(0)=\alpha$, it is easy to see
that $\lim_{x\to0^+}n_{ix}(x)$ exists by the monotone convergence
argument. Then from the first equation of \eqref{1}, we have
\begin{equation*}
n _{ix}=\frac{ E_i n _i^2 }{ n _i + 1 } + \frac{\left(E_i-\alpha\right) n _i^2 }{ \left( n _i - 1 \right)\left( n _i + 1 \right)}, \quad\text{in}\ (0,1).
\end{equation*}
Noting that $n_i(0)=1$ and $E_i(0)=\alpha$, it follows from the
L'Hospital Rule that
\begin{align*}
A_i &=\lim_{x\to0^+}n_{ix}(x)=\lim_{x\to0^+}\frac{E_in_i^2}{n_i+1}+\lim_{x\to0^+}\frac{\left(E_i-\alpha\right)n_i^2}{\left(n_i-1\right)\left(n_i+1\right)} \\
&=\frac{\alpha}{2}+\frac{1}{2}\lim_{x\to0^+}\frac{\left(E_i-\alpha\right)_x}{\left(n_i-1\right)_x}=\frac{\alpha}{2}+\frac{1}{2}\lim_{x\to0^+}\frac{E_{ix}}{n_{ix}}\\
&=\frac{\alpha}{2}+\frac{1}{2}\lim_{x\to0^+}\frac{n_i(x)-b_i(x)}{n_{ix}}\\
&=\frac{\alpha}{2}+\frac{1-b_i(0)}{2A_i},
\end{align*}
which in turn implies that
\begin{equation*}
A_i=\frac{1}{4}\left(\alpha-\sqrt{\alpha^2-8\left(b_i(0)-1\right)}\right)=\frac{2\left(b_i(0)-1\right)}{\alpha+\sqrt{\alpha^2-8\left(b_i(0)-1\right)}}=O(\frac{1}{\alpha}),
\end{equation*}
or
\begin{equation*}
A_i=\frac{1}{4}\left(\alpha+\sqrt{\alpha^2-8\left(b_i(0)-1\right)}\right)=O(\alpha).
\end{equation*}
According to the local singularity analysis in \cite{LMZZ17}(Lemma 5.3, P4796),
we know that in a small neighborhood of $x=0$, the drastic change of the
density component $n_i(x)$ of the interior subsonic solution is impossible
when $\alpha$ is suitably large. Therefore, we have to choose the former
root as the limit value of $\lim_{x\to0^+}n_{ix}(x)$, and the latter one is
the extraneous root.\qed
\vspace{3mm}
Based on Proposition \ref{T1} and Lemmas \ref{Lcp}$\sim$\ref{L1}, we are
now preparing to establish the local structural stability of interior
subsonic solutions to the boundary value problem \eqref{1.6}\&\eqref{1.7}
on an intrinsic neighborhood of the left endpoint $x=0$.
\vspace{1.5mm}
\begin{lem}[Local structural stability estimate near $x=0$]\label{L3}
Under the same conditions in Lemma \ref{L1}. There exist two positive
constants $\delta_0\in(0,\frac{1}{2})$ and $C>0$ independent of
$\|b_1-b_2\|_{C[0, 1]}$ such that
\begin{equation}\label{key-1}
\|n_1-n_2\|_{C^1[0,\delta_0)}+\|E_1-E_2\|_{C^1[0,\delta_0)}\le C\| b_1-b_2\|_{C[0,1]}.
\end{equation}
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
Firstly, in light of Lemma \ref{Lcp}, it is clear that the following
monotonicity relation holds,
\begin{equation}\label{mr}
\frac{n_1^3}{n_1+1}\ge\frac{n_2^3}{n_2+1},\quad\forall x\in[0,1].
\end{equation}
Next, for simplicity, we set $\tilde{E_i}:=E_i-\frac{\alpha}{n_i}$.
Multiplying Equation $\eqref{1}_1$ by $\frac{n_i^2}{n_i^2-1}$, we have
\begin{equation}\label{h1}
n_{ix}=\frac{\tilde{E}_i n_i^3}{n_i^2-1}, \quad i=1,2.
\end{equation}
Taking the difference of Equations $\eqref{h1}|_{i=1}$ and $\eqref{h1}|_{i=2}$,
near $x=0$, we compute together with the monotonicity relation \eqref{mr} that
\begin{align}\label{33}
(n_1-n_2)_x=&\frac{\tilde{E}_1 n_1^3}{n_1^2-1}-\frac{\tilde{E}_2 n_2^3}{n_2^2-1}\\
=&\frac{n_1^3}{n_1+1}\frac{\tilde{E}_1}{n_1-1}-\frac{n_2^3}{n_2+1}\frac{\tilde{E}_1}{n_1-1}+\frac{n_2^3}{n_2+1}\frac{\tilde{E}_1}{n_1-1}-\frac{n_2^3}{n_2+1}\frac{\tilde{E}_2}{n_2 - 1}\notag\\
=&\frac{\tilde{E}_1}{n_1-1}\left(\frac{n_1^3}{n_1+1}-\frac{n_2^3}{n_2+1}\right)+\frac{n_2^3}{n_2+1}\left(\frac{\tilde{E}_1}{n_1-1}-\frac{\tilde{E}_2}{n_2-1}\right)\notag\\
\leq&M_0\alpha\left(\frac{n_1^3}{n_1+1}-\frac{n_2^3}{n_2+1}\right)+\frac{n_2^3}{n_2+1}M_0\|b_1-b_2\|_{C[0, 1]},\notag\\
\leq&C(n_1-n_2)+C\|b_1-b_2\|_{C[0, 1]},\quad x\in[0,\delta_0), \notag
\end{align}
where we have used the fact that there exist two positive constants
$\delta_0\in(0,\frac{1}{2})$ and $M_0>0$ independent of $\|b_1-b_2\|_{C[0, 1]}$ such that
\begin{equation}\label{0227}
\frac{\tilde{E}_1}{n_1- 1}(x)\leq M_0\alpha,\quad\text{and}\quad \left( \frac{\tilde{E}_1}{n_1- 1}-\frac{\tilde{E}_2}{n_2- 1}\right) (x)\leq M_0\|b_1-b_2\|_{C[0, 1]},\quad x\in[0,\delta_0).
\end{equation}
To prove that the crucial estimate \eqref{0227} on a certain intrinsic
neighborhood $[0,\delta_0)$ holds, we assume for the sake of contradiction
that for any $\delta\in(0,\frac{1}{2})$ and $M>0$, there exists
$x_\delta\in[0,\delta)$ such that
\begin{equation}\label{C0227}
\frac{\tilde{E}_1}{n_1-1}(x_\delta)>M\alpha, \quad\text{or}\quad \left( \frac{\tilde{E}_1}{n_1-1}-\frac{\tilde{E}_2}{n_2-1}\right) (x_\delta)> M\|b_1-b_2\|_{C[0, 1]}.
\end{equation}
Particularly, we take $\delta=\frac{1}{k}, k=3,4,5,\cdots$, for any $M>0$,
there is $x_k\in[0,\frac{1}{k})$ such that
\begin{equation*}
\frac{\tilde{E}_1}{n_1-1}(x_k)>M\alpha, \quad\text{or}\quad \left( \frac{\tilde{E}_1}{n_1-1}-\frac{\tilde{E}_2}{n_2-1}\right) (x_k)> M\|b_1-b_2\|_{C[0, 1]},
\end{equation*}
which implies that
\begin{equation}\label{0227-1}
\varliminf_{x_k\to0^+}\frac{\tilde{E}_1}{n_1-1}(x_k)\geq M\alpha,
\end{equation}
or
\begin{equation}\label{0227-2}
\varliminf_{x_k\to0^+}\left(\frac{\tilde{E}_1}{n_1-1}-\frac{\tilde{E}_2}{n_2-1}\right)(x_k)\geq M\|b_1-b_2\|_{C[0, 1]}.
\end{equation}
Combining the boundary behavior \eqref{E(1)}, the L'Hospital Rule,
Equation $\eqref{1}_2$ and Lemma \ref{L1}, we calculate
\begin{align}\label{0227-3}
\lim_{x\to0^+}\frac{\tilde{E}_i}{n_i-1}(x)=&\lim_{x\to0^+}\frac{E_i-\frac{\alpha}{n_i}}{n_i-1}(x)\\
=&\lim_{x\to0^+}\frac{E_i(x)-E_i(0)+\alpha-\frac{\alpha}{n_i(x)}}{n_i(x)-1} \notag\\
=&\lim_{x\to0^+}\frac{E_i(x)-E_i(0)}{n_i(x)-1}+\lim_{x\to0^+}\frac{\alpha}{n_i(x)}\notag\\
=&\lim_{x\to0^+}\frac{n_i-b_i}{n_{ix}}+\alpha \notag\\
=&\frac{1-b_i(0)}{A_i}+\alpha<\alpha, \quad i=1,2,\notag
\end{align}
and
\begin{align}\label{0227-4}
\lim_{x\to0^+}\left( \frac{\tilde{E}_1}{n_1-1}-\frac{\tilde{E}_2}{n_2-1}\right)(x)=&\frac{b_2(0)-1}{A_2}-\frac{b_1(0)-1}{A_1}\\
=&\frac{1}{2}\left(\sqrt{\alpha^2-8(b_2(0)-1)}-\sqrt{\alpha ^2-8(b_1(0)-1)}\right) \notag\\
=&\frac{2}{\sqrt{\alpha^2-8(\eta-1)}}\left(b_1(0)-b_2(0)\right) \notag\\
\leq&\tilde{C}_0\|b_1-b_2\|_{C[0, 1]},\notag
\end{align}
where $\eta\in(b_2(0), b_1(0))$. Furthermore, we note that the constant $M$
in \eqref{0227-1} and \eqref{0227-2} can be chosen arbitrarily. Consequently,
if we take $M=2$ in \eqref{0227-1}, together with \eqref{0227-3}, we obtain
the contradiction that $2\alpha<\alpha$; if we take $M=2\tilde{C}_0$ in
\eqref{0227-2}, combined with \eqref{0227-4}, we have the contradiction $2\leq1$.
Based on the local estimate \eqref{33}, we continue establishing the structural
stability locally on the intrinsic neighborhood $[0,\delta_0)$. To this end,
we multiply through the inequality \eqref{33} by $n_1-n_2$ and calculate
\begin{equation}\label{26}
\frac{d}{dx}( n_1 -n_2)^2(x)\le C( n_1-n_2)^2(x)+C\|b_1-b_2\|_{C[0,1]}^2,\quad x\in[0, \delta_0),
\end{equation}
where we have used Lemma \ref{Lcp} and Cauchy's inequality. By Gronwall's
inequality and the sonic boundary condition $n_1(0)=n_2(0)=1$, we get
\begin{equation}\label{27}
(n_1-n_2)^2(x)\le C\|b_1-b_2\|_{C[0, 1]}^2,\quad x\in[0,\delta_0),
\end{equation}
which in turn implies that
\begin{equation}\label{28}
|n_1-n_2|(x)+|\left(n_{1}-n_{2}\right)_x|(x)\le C\|b_1-b_2\|_{C[0, 1]},\quad x\in[0,\delta_0).
\end{equation}
with the aid of the foregoing local estimate \eqref{33} again.
Finally, from Equation \eqref{E} in Definition \ref{def}, we have
\begin{equation} \label{29}
E_i(x)=\alpha+\int_0^x\left(n_i(y)-b_i(y)\right)dy,\quad i=1,2.
\end{equation}
Taking the difference of $\eqref{29}|_{i=1}$ and $\eqref{29}|_{i=2}$, we
compute that
\begin{align}\label{30}
|E_1-E_2|(x)\le&\int_0^x|n_1-n_2|(y)dy+\int_0^x|b_1-b_2|(y)|dy\\
\le&C\|b_1-b_2\|_{C[0,1]},\quad x\in[0,\delta_0),\notag
\end{align}
and
\begin{align}\label{31}
|(E_1-E_2)_x|(x)=&|n_1-n_2-(b_1-b_2)|(x)\\
\le&C\|b_1-b_2\|_{C[0,1]}, \quad x\in[0,\delta_0).\notag
\end{align}
Hence, the local structural stability estimate \eqref{key-1} follows
immediately from Equations \eqref{28}, \eqref{30} and \eqref{31}.\qed
\vspace{3mm}
We now turn to analyzing the refined boundary behavior of the first-order
derivative of $n_i(x)$ at the right endpoint $x=1$. From the boundary estimate
displayed in the second line of $\eqref{L14}$, we know that
$\lim_{x\to1^-}n_{ix}(x)=-\infty$. This means the ``genuine''
singularity will occur at the right endpoint $x=1$. Inspired
by $\eqref{L14}$, we are able to implement the local ``weighted''
singularity analysis. The result is summarized as follows.
\vspace{1.5mm}
\begin{lem}\label{L4-1}
Assume that $b_i, i=1,2$ satisfy the same conditions in Lemma \ref{Lcp}. Then
\begin{equation}\label{(1)}
\lim_{x\to1^-}(1-x)^{\frac{1}{2}}n_{ix}(x)=-\frac{1}{2}\sqrt{\int_0^1(b_i-n_i)dx}=:B_i<0,\quad i=1,2.
\end{equation}
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
For $i=1,2$, from the boundary estimate \eqref{L14}, we know that the
coefficient $1-\frac{1}{n_i^2}$ in the degenerate principal part of
Equation $\eqref{1}_1$ is comparable to $(1-x)^{\frac{1}{2}}$ near
the right endpoint $x=1$. Thus the regularity theory of
boundary-degenerate elliptic equations in one dimension
(e.g. \cite{Sc81}) ensures that $(1-x)^\frac{1}{2}n_{ix}(x)$ is
continuous up to the right endpoint $x=1$.
We now proceed to calculate the exact limit value of
$\lim_{x\to1-}(1-x)^{\frac{1}{2}}n_{ix}(x)$. For convenience, we set
\begin{equation*}
B_i:=\lim_{x\to1-}(1-x)^{\frac{1}{2}}n_{ix}(x).
\end{equation*}
Thereupon, multiplying through Equation $\eqref{1}_1$ by $(1-x)^{\frac{1}{2}}\frac{n_i^2}{n_i^2-1}$, we have
\begin{equation*}
(1-x)^\frac{1}{2}n_{ix}=\frac{n_i^3}{n_i+1}\left(E_i-\frac{\alpha}{n_i}\right)\frac{(1-x)^\frac{1}{2}}{n_i-1}.
\end{equation*}
By virtue of the sonic boundary condition $n_i(1)=1$, the known
boundary behavior \eqref{E(1)} and the L'Hospital Rule, we compute
\begin{align}\label{(3)}
B_i=&\lim_{x\to1^-}(1-x)^\frac{1}{2}n_{ix}\\
=&\lim_{x\to1^-}\frac{n_i^3}{n_i+1}\lim_{x\to1^-}(E_i-\frac{\alpha}{n_i})\lim_{x\to1^-}\frac{(1-x)^\frac{1}{2}}{n_i-1}\notag\\
=&\frac{1}{2}(E_i(1)-\alpha)\lim_{x\to1^-}\frac{-\frac{1}{2}(1-x)^{-\frac{1}{2}}}{n_{ix}}\notag\\
=&\frac{1}{4}(\alpha-E_i(1))\lim_{x\to1^-}\frac{1}{n_{ix}(1-x)^\frac{1}{2}}\notag\\
=&\frac{1}{4}(E_i(0)-E_i(1))\frac{1}{B_i},\notag
\end{align}
which implies from Equation $\eqref{1}_2$ that
\begin{equation*}
B_i=-\frac{1}{2}\sqrt{E_i(0)-E_i(1)}=-\frac{1}{2}\sqrt{\int_0^1(b_i-n_i)dx}<0,
\end{equation*}
where the boundary estimate $\eqref{L14}_2$ has been employed to uniquely
determine the value of $B_i$, which is strictly negative.\qed
\vspace{3mm}
Proposition \ref{T1} alongside Lemmas \ref{Lcp} and \ref{L4-1} now enable us
to demonstrate the local weighted structural stability of interior subsonic
solutions to the boundary value problem \eqref{1.6}\&\eqref{1.7} on an
intrinsic neighborhood of the right endpoint $x=1$.
\vspace{1.5mm}
\begin{lem}[Local weighted structural stability estimate near $x=1$]\label{L4}
Under the same conditions in Lemma \ref{L1}. There exist two positive
constants $\delta_1\in(0,\frac{1}{2})$ and $C>0$ independent
of $\| b_1-b_2\|_{C[0, 1]}$ such that
\begin{align}\label{key-2}
&\left\|(1-x)^{-\frac{1}{2}}(n_1-n_2)\right\|_{C(1-\delta_1,1]}+\left\|(1-x)^\frac{1}{2}(n_{1x}-n_{2x})\right\|_{C(1-\delta_1,1]}\\
&+\|E_1-E_2\|_{C^1(1-\delta_1,1]}\le C\|b_1-b_2\|_{C[0,1]}.\notag
\end{align}
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
For $i=1,2$, from $\eqref{L14}_1$, we have known that $\frac{n_i-1}{(1-x)^{1/2}}$
possesses the uniform positive upper and lower bounds near $x=1$, and
so does its reciprocal $\frac{(1-x)^{1/2}}{n_i-1}$. This property will be
used repeatedly hereafter.
Owing to the fact that $n_{ix}(x)$ has the genuine singularity at $x=1$, we
are compelled to establish the structural stability estimate near $x=1$ only
in the weighted manner as follows.
Firstly, multiplying through Equation $\eqref{1}_1$ by $(1-x)^{\frac{1}{2}}\frac{n_i^2}{n_i^2-1}$ and taking the difference of resultant equations for $i=1,2$, we calculate that
\begin{align}\label{(4)}
&(1-x)^\frac{1}{2}(n_{1x}-n_{2x})\\
=&\frac{n_1^2}{n_1+1}(n_1E_1-\alpha)\frac{(1-x)^\frac{1}{2}}{n_1-1}-\frac{n_2^2}{n_2+1}(n_2E_2-\alpha)\frac{(1-x)^\frac{1}{2}}{n_2-1}\notag\\
=&\frac{n_1^2}{n_1+1}(n_1E_1-\alpha)\left(\frac{(1-x)^\frac{1}{2}}{n_1-1}-\frac{(1-x)^\frac{1}{2}}{n_2-1}\right)\notag\\
&+\left(\frac{n_1^2}{n_1+1}(n_1E_1-\alpha)-\frac{n_2^2}{n_2+1}(n_2E_2-\alpha)\right)\frac{(1-x)^\frac{1}{2}}{n_2-1}\notag\\
=&h(n_1,E_1)\frac{(1-x)^\frac{1}{2}}{n_1-1}\frac{(1-x)^\frac{1}{2}}{n_2-1}\frac{n_2-n_1}{(1-x)^\frac{1}{2}}+\left(h(n_1,E_1)-h(n_2,E_2)\right)\frac{(1-x)^\frac{1}{2}}{n_2-1}\notag\\
=:&I_1+I_2,\notag
\end{align}
where
\begin{equation*}
h(n_i,E_i):=\frac{n_i^2}{n_i+1}(n_iE_i-\alpha),\quad i=1,2.
\end{equation*}
In what follows, near $x=1$, we shall estimate $I_1$ and $I_2$,
respectively. But first, we claim that the following estimates
\begin{gather}
|E_i(x)|\le\alpha+2\bar{b}_i,\quad x\in[0,1],\quad i=1,2,\label{hu1}\\
|E_1(1)-E_2(1)|\le C\|b_1-b_2\|_{C[0,1]}\label{hu2}
\end{gather}
hold, where the estimate constant $C>0$ is independent
of $\|b_1-b_2\|_{C[0,1]}$, and the proof of which is deferred
to Lemma \ref{lem-hu} at the end of this paper.
As for $I_1$, it is clear from $\eqref{L14}_1$ and \eqref{hu1} that
\begin{equation}\label{(5)}
|I_1|=\left|h(n_1,E_1)\frac{(1-x)^\frac{1}{2}}{n_1-1}\frac{(1-x)^\frac{1}{2}}{n_2-1}\frac{n_2-n_1}{(1-x)^\frac{1}{2}}\right|\le C\frac{|n_1-n_2|}{(1-x)^\frac{1}{2}}.
\end{equation}
However, as far as $I_2$ is concerned, the situation becomes more
complicated because of the factor $h(n_1,E_1)-h(n_2,E_2)$. Next, we are
taking it step by step. Precisely, a straightforward computation gives
\begin{align}\label{(6)}
h(n_1,E_1)-h(n_2,E_2)&=\frac{n_1^2}{n_1+1}(n_1E_1-\alpha)-\frac{n_2^2}{n_2+1}(n_2E_2-\alpha)\\
&=\alpha\left(\frac{n_2^2}{n_2+1}-\frac{n_1^2}{n_1+1}\right)+\left(\frac{n_1^3E_1}{n_1+1}-\frac{n_2^3E_2}{n_2+1}\right)\notag\\
&=:R_1+R_2.\notag
\end{align}
From \eqref{ulb} and Lemma \ref{Lcp}, we know that
\begin{equation}\label{hu0}
1\le1+m(\alpha,\underline{b}_2)\sin(\pi x)\le n_2(x)\le n_1(x)\le\overline{b}_1,\quad x\in[0,1].
\end{equation}
Consequently, it follows from the mean-value theorem of differentials that
\begin{equation}\label{(7)}
|R_1|=\left|\alpha\left(\frac{n_2^2}{n_2+1}-\frac{n_1^2}{n_1+1}\right)\right| \le C|n_1-n_2|\le C\frac{|n_1-n_2|(x)}{(1-x)^\frac{1}{2}}.
\end{equation}
We now turn to estimating $R_2$ near $x=1$. Combining \eqref{hu0},
\eqref{hu1}, \eqref{hu2}, the mean-value theorem of differentials,
and the mean-value theorem of integrals, we have
\begin{align}\label{hu3}
|R_2|&=\left|\frac{n_1^3E_1}{n_1+1}-\frac{n_2^3E_2}{n_2+1}\right|\\
&=\left|E_1\left(\frac{n_1^3}{n_1+1}-\frac{n_2^3}{n_2+1}\right)+\frac{n_2^3}{n_2+1}(E_1-E_2)\right|\notag\\
&\le C|n_1-n_2|(x)+C\left|\left[\big(E_1(1)-E_2(1)\big)-\left(\int_x^1(n_1-n_2)-(b_1-b_2)dy\right)\right]\right|\notag\\
&\le C|n_1-n_2|(x)+C|E_1(1)-E_2(1)|+C\int_x^1|n_1-n_2|dy+C\|b_1-b_2\|_{C[0,1]}\notag\\
&\le C\|b_1-b_2\|_{C[0,1]}+C\left(\frac{|n_1-n_2|(x)}{(1-x)^{\frac{1}{2}}}+\frac{|n_1-n_2|(\xi)}{(1-\xi)^{\frac{1}{2}}}\right),\quad\exists\xi\in[x,1],\notag
\end{align}
where we have used the formula
\begin{equation}
E_i(x)=E_i(1)-\int_x^1(n_i-b_i)(y)dy,\quad i=1,2.
\end{equation}
Substituting \eqref{(7)} and \eqref{hu3} into \eqref{(6)}, we have
\begin{equation*}
|h(n_1,E_1)-h(n_2,E_2)|\le C\|b_1-b_2\|_{C[0,1]}+C\left(\frac{|n_1-n_2|(x)}{(1-x)^{\frac{1}{2}}}+\frac{|n_1-n_2|(\xi)}{(1-\xi)^{\frac{1}{2}}}\right),\quad\exists\xi\in[x,1],
\end{equation*}
which further implies that
\begin{align}\label{hu4}
|I_2|&=\left|\left(h(n_1,E_1)-h(n_2,E_2)\right)\frac{(1-x)^\frac{1}{2}}{n_2-1}\right|\\
&\le C\|b_1-b_2\|_{C[0,1]}+C\left(\frac{|n_1-n_2|(x)}{(1-x)^{\frac{1}{2}}}+\frac{|n_1-n_2|(\xi)}{(1-\xi)^{\frac{1}{2}}}\right),\quad\exists\xi\in[x,1].\notag
\end{align}
Inserting \eqref{(5)} and \eqref{hu4} into \eqref{(4)}, near $x=1$, we
obtain \begin{align}\label{(14)}
&(1-x)^\frac{1}{2}|(n_1-n_2)_x|(x)\\
\le&C\left(\frac{|n_1-n_2|(x)}{(1-x)^{\frac{1}{2}}}+\frac{|n_1-n_2|(\xi)}{(1-\xi)^{\frac{1}{2}}}\right)+C\|b_1-b_2\|_{C[0,1]},\quad\exists\xi\in[x,1].\notag
\end{align}
It is worth mentioning that the generic constant $C>0$ in \eqref{(14)} is
independent of $\|b_1-b_2\|_{C[0,1]}$. Moreover,
the term $\frac{|n_1-n_2|(x)}{(1-x)^{1/2}}+\frac{|n_1-n_2|(\xi)}{(1-\xi)^{1/2}}$
on the right-hand side of \eqref{(14)} can be bounded by an appropriate
constant multiple of $\|b_1-b_2\|_{C[0,1]}$ in an intrinsic neighborhood
of the right endpoint $x=1$. Precisely, we claim that there exist two
positive constants $0<\delta_1<\frac{1}{2}$ and $M_1>0$ independent
of $\|b_1-b_2\|_{C[0,1]}$ such that
\begin{equation}\label{(14*)}
\frac{|n_1-n_2|(x)}{(1-x)^\frac{1}{2}}\le M_1\|b_1-b_2\|_{C[0,1]}, \quad x\in(1-\delta_1,1].
\end{equation}
Aiming for a contradiction, suppose that for any $\delta\in (0,\frac{1}{2})$
and $M>0$, there is $x_\delta\in(1-\delta,1]$ such that
\begin{equation}
\frac{|n_1-n_2|(x_\delta)}{(1-x_\delta)^\frac{1}{2}}>M\|b_1-b_2\|_{C[0,1]}.
\end{equation}
By the arbitrariness, we could take $\delta=\frac{1}{k}, k=3,4,5,\cdots$, for
arbitrary $M>0$, there exists $x_k\in(1-\frac{1}{k},1]$ such that
\begin{equation}
\frac{|n_1-n_2|(x_k)}{(1-x_k)^\frac{1}{2}}>M\|b_1-b_2\|_{C[0,1]},
\end{equation}
which implies that
\begin{equation}\label{0227-5}
\varliminf_{x_k\to1^-}\frac{|n_1-n_2|(x_k)}{(1-x_k)^\frac{1}{2}}\ge M\|b_1-b_2\|_{C[0,1]}.
\end{equation}
Besides, combining Lemma \ref{Lcp}, the L'Hospital Rule and Lemma
\ref{L4-1}, we calculate that
\begin{align}\label{0227-6}
\lim_{x\to1^-}\frac{|n_1-n_2|(x)}{(1-x)^\frac{1}{2}}=&\lim_{x\to1^-}\frac{(n_1-n_2)(x)}{(1-x)^\frac{1}{2}}=\lim_{x\to1^-}\frac{(n_1-1)(x)}{(1-x)^\frac{1}{2}}-\lim_{x\to1^-}\frac{(n_2-1)(x)}{(1-x)^\frac{1}{2}}\\
=&-2\lim_{x\to1^-}(1-x)^{\frac{1}{2}}n_{1x}+2\lim_{x\to1^-}(1-x)^{\frac{1}{2}}n_{2x}\notag\\
=&\sqrt{\int_0^1(b_1-n_1)dx}-\sqrt{\int_0^1(b_2-n_2)dx}\notag\\
=&\frac{\int_0^1(b_1-b_2)dx-\int_0^1(n_1-n_2)dx}{\sqrt{\int_0^1(b_1-n_1)dx}+\sqrt{\int_0^1(b_2-n_2)dx}}\notag\\
\le&\frac{\int_0^1(b_1-b_2)dx}{\sqrt{\int_0^1(b_1-n_1)dx}+\sqrt{\int_0^1(b_2-n_2)dx}}\notag\\
\le&\tilde{C}_1\|b_1-b_2\|_{C[0,1]}.\notag
\end{align}
Moreover, we note that the constant $M>0$ in \eqref{0227-5} is arbitrary.
Therefore, together with \eqref{0227-6}, taking $M=2\tilde{C}_1$
in \eqref{0227-5} leads to the contradiction that $2\le1$.
Applying \eqref{(14*)} to \eqref{(14)}, we have
\begin{equation}\label{hu5}
(1-x)^\frac{1}{2}|(n_1-n_2)_x|(x)\le C\|b_1-b_2\|_{C[0,1]}, \quad x\in(1-\delta_1,1].
\end{equation}
Similarly to \eqref{hu3}, we are able to compute that
\begin{align}\label{(21)}
|E_1-E_2|(x)\le&C\|b_1-b_2\|_{C[0,1]}+\frac{|n_1-n_2|(\xi)}{(1-\xi)^{\frac{1}{2}}},\quad\exists\xi\in[x,1]\\
\le&C\|b_1-b_2\|_{C[0,1]},\quad x\in(1-\delta_1,1],\notag
\end{align}
and
\begin{align}\label{(22)}
|(E_1-E_2)_x|(x)=&|n_1-n_2-(b_1-b_2)|(x)\le|n_1-n_2|(x)+|b_1-b_2|(x)\\
\le&\frac{|n_1-n_2|(x)}{(1-x)^{\frac{1}{2}}}+\|b_1-b_2\|_{C[0,1]}\notag\\
\le&C\|b_1-b_2\|_{C[0,1]},\quad x\in(1-\delta_1,1].\notag
\end{align}
Finally, putting results \eqref{(14*)}, \eqref{hu5}, \eqref{(21)}
and \eqref{(22)} together, we obtain the desired local weighted
estimate \eqref{key-2}.\qed
\vspace{3mm}
Up to now, we have obtained two intrinsic small domains $[0,\delta_0)$
and $(1-\delta_1,1]$ distributed around the two endpoints $x=0$ and $x=1$,
respectively. This fact enables us to establish the structural stability
estimate on a certain regular domain $[\delta,1-\delta]$, where
$0<\delta:=\min\{\delta_0,\delta_1\}<1/2$.
\vspace{1.5mm}
\begin{lem}\label{L5}
Under the same conditions in Lemma \ref{L1}. Let $\delta:=\min\{\delta_0,\delta_1\}$. Then there is a positive constant $C>0$ independent of $\|b_1-b_2\|_{C[0,1]}$ such that
\begin{equation}\label{key-3}
\|n_1-n_2\|_{C^1[\delta,1-\delta]}+\|E_1-E_2\|_{C^1[\delta,1-\delta]}\le C\|b_1-b_2\|_{C[0,1]}.
\end{equation}
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
We are now able to work with Equations \eqref{1} on a regular closed
interval $[\delta,1-\delta]$ away from singularities, where $\delta$
has been defined in the hypothesis of the present lemma.
Firstly, we rewrite the estimate \eqref{hu0} on the regular interval as follows.
\begin{equation}\label{hu6}
1<l:=1+m(\alpha,\underline{b}_2)\sin(\pi\delta)\le n_2(x)\le n_1(x)\le\overline{b}_1,\quad x\in[\delta,1-\delta].
\end{equation}
Secondly, subtracting $\eqref{1}|_{i=2}$ from $\eqref{1}|_{i=1}$, for $x\in[\delta,1-\delta]$, we thus get
\begin{align}\label{41}
(n_1-n_2)_x=&\frac{n_1^3E_1-\alpha n_1^2}{n_1^2-1}-\frac{n_2^3E_2-\alpha n_2^2}{n_2^2-1}\\
=&E_1\left(f(n_1)-f(n_2)\right)+f(n_2)(E_1-E_2)-\alpha\left(g(n_1)-g(n_2)\right)\notag\\
=&\left(E_1 f'(\bar{\eta})-\alpha g'(\tilde{\eta})\right)(n_1-n_2)+f(n_2)(E_1-E_2),\quad\exists\bar{\eta},\tilde{\eta}\in(n_2,n_1),\notag
\end{align}
and
\begin{equation}\label{42}
(E_1-E_2)_x =(n _1-n _2)-(b_1-b_2),
\end{equation}
where
\begin{equation*}
f(n):=\frac{n^3}{n^2-1},\quad g(n):=\frac{n^2}{n^2-1}, \quad\forall n\in[l,\overline{b}_1],
\end{equation*}
and we have used the mean-value theorem of differentials in the third
line of Equation \eqref{41}.
Thirdly, multiplying through \eqref{41} by $n_1-n_2$, and using \eqref{hu1},
\eqref{hu6} and Cauchy's inequality together, we have
\begin{equation}\label{hu7}
\left((n_1-n_2)^2\right)_x\le C(\alpha,l,\bar{b}_1)\left((n_1-n_2)^2+(E_1-E_2)^2\right), \quad x\in[\delta,1-\delta].
\end{equation}
Similarly, multiplying through \eqref{42} by $E_1-E_2$, and employing
Cauchy's inequality, we obtain
\begin{equation}\label{hu8}
\left((E_1-E_2)^2\right)_x\le(n_1-n_2)^2+2(E_1-E_2)^2+\|b_1-b_2\|_{C[0,1]}^2, \quad x\in[\delta,1-\delta].
\end{equation}
And then, summing estimates \eqref{hu7} and \eqref{hu8} gives
\begin{align}\label{hu9}
&\left((n_1-n_2)^2+(E_1-E_2)^2\right)_x(x)\\
\le&C\left((n_1-n_2)^2+(E_1-E_2)^2\right)(x)+\|b_1-b_2\|_{C[0,1]}^2, \quad x\in[\delta,1-\delta].\notag
\end{align}
Applying the Gronwall inequality to \eqref{hu9}, we have
\begin{align}\label{hu10}
&\left((n_1-n_2)^2+(E_1-E_2)^2\right)(x)\\
\le&e^{\int_\delta^xCdy}\left[\left((n_1-n_2)^2+(E_1-E_2)^2\right)(\delta)+\int_\delta^x\|b_1-b_2\|_{C[0,1]}^2dy\right]\notag\\
\le&C\left[\left((n_1-n_2)^2+(E_1-E_2)^2\right)(\delta)+\|b_1-b_2\|_{C[0,1]}^2\right], \quad x\in[\delta,1-\delta].\notag
\end{align}
Noting that $\delta\le\delta_0$ and the continuity of the error
function pair $(n_1-n_2, E_1-E_2)(x)$ at $x=\delta_0$, from
Lemma \ref{L3} we see
\begin{equation}\label{hu11}
\left((n_1-n_2)^2+(E_1-E_2)^2\right)(\delta)\le C\|b_1-b_2\|_{C[0,1]}^2,
\end{equation}
which along with \eqref{hu10} implies
\begin{equation}\label{47}
|n_1-n_2|(x)+|E_1-E_2|(x)\le C\|b_1-b_2\|_{C[0,1]}, \quad x\in[\delta,1-\delta].
\end{equation}
Finally, from Equations \eqref{41}\&\eqref{42}, and the estimate \eqref{47},
we directly calculate
\begin{equation}\label{hu12}
|(n_1-n_2)_x|(x)+|(E_1-E_2)_x|(x)\le C\|b_1-b_2\|_{C[0,1]}, \quad x\in[\delta,1-\delta].
\end{equation}
Combining estimates \eqref{47} and \eqref{hu12} yields the desired structural
stability estimate \eqref{key-3} on the regular domain $[\delta,1-\delta]$.\qed
\vspace{3mm}
Last but not least, let us prove the estimates \eqref{hu1} and \eqref{hu2} in
the following lemma.
\vspace{1.5mm}
\begin{lem}\label{lem-hu}
Under the same conditions in Lemma \ref{L4}. Then there exists a
positive constant $C$ independent of $\|b_1-b_2\|_{C[0,1]}$ such that
estimates \eqref{hu1} and \eqref{hu2} hold, that is,
\begin{equation*}
|E_i(x)|\le\alpha+2\bar{b}_i,\quad x\in[0,1],\quad i=1,2,
\end{equation*}
and
\begin{equation*}
|E_1(1)-E_2(1)|\le C\|b_1-b_2\|_{C[0,1]},
\end{equation*}
respectively.
\end{lem}
\vspace{3mm}
\noindent\textbf{Proof.}
From Equation \eqref{E} in Definition \ref{def}, we have
\begin{equation} \label{hu13}
E_i(x)=\alpha+\int_0^x(n_i-b_i)(y)dy,\quad\forall x\in[0,1], \quad i=1,2.
\end{equation}
First of all, in light of the lower and upper bounds \eqref{ulb} of $n_i(x)$,
a straightforward computation gives
\begin{align}
|E_i(x)|=\left|\alpha+\int_0^x(n_i-b_i)dy\right|\le\alpha+\int_0^1(n_i+b_i)dy\le\alpha+2\bar{b}_i,\quad\forall x\in[0,1], \quad i=1,2.
\end{align}
Next, taking the value $x=1$ in Equation \eqref{hu13}, we have
\begin{equation}\label{hu14}
E_i(1)=\alpha+\int_0^1(n_i-b_i)(y)dy, \quad i=1,2.
\end{equation}
Furthermore, taking the difference of Equations $\eqref{hu14}|_{i=1}$
and $\eqref{hu14}|_{i=2}$, we calculate
\begin{align}\label{hu15}
|E_1(1)-E_2(1)|\le&\left|\int_0^1\Big[(n_1-b_1)-(n_2-b_2)\Big]dy\right|\\
\le&\int_0^1|n_1-n_2|(y)dy+\|b_1-b_2\|_{C[0,1]}\notag\\
=&|n_1-n_2|(\xi)+\|b_1-b_2\|_{C[0,1]}, \quad\exists\xi\in[0,1],\notag
\end{align}
where we have used the mean-value theorem of integrals in the last line.
Finally, we claim that there is a positive constant $C$ independent of
$\|b_1-b_2\|_{C[0,1]}$ such that
\begin{equation}\label{hu16}
|n_1-n_2|(\xi)\le C\|b_1-b_2\|_{C[0,1]},
\end{equation}
wherever the point $\xi$ is located in the whole interval $[0,1]$.
In fact, take $\delta$ the same as in Lemma \ref{L5}, and if
$\xi\in[0,1-\delta]$, it is clear from estimates \eqref{key-1}
and \eqref{key-3} that \eqref{hu16} is true; if $\xi\in(1-\delta,1]$,
the intrinsic local estimate \eqref{(14*)} guarantees that \eqref{hu16}
is true as well. Consequently, substituting \eqref{hu16} into \eqref{hu15},
we obtain the desired estimate \eqref{hu2}.\qed
\vspace{3mm}
We end this section by a summary for the proof of Theorem \ref{T2}
because the previous lemmas have already made the proof evident.
\vspace{3mm}
\noindent\textbf{Proof of Theorem \ref{T2}.}
Obviously, putting all the estimates we have established in
Lemmas \ref{L3}, \ref{L5} and \ref{L4} together, we have the
globally structural stability estimate \eqref{1.10**}.\qed
\vspace{5mm}
\noindent{\sc Acknowledgments:} This work was commenced while the
first two authors were visiting McGill University from 2021 to 2022.
They would like to express their gratitude to McGill University for
its hospitality. The research of Y. H. Feng was supported by China
Scholarship Council for the senior visiting scholar program
(202006545001). The research of H. Hu was partially supported by
National Natural Science Foundation of China (Grant No.11801039),
China Scholarship Council (No.202007535001) and Scientific Research
Program of Changchun University (No.ZKP202013). The research of M.
Mei was partially supported by NSERC grant RGPIN 354724-2016.
|
\section{Our Algorithm to Certify a Binary Monotone Function}\label{sec:bin-search}
We first restate the problem.
\begin{problem*}
Given query access to a monotone function $\myfunc{f}{\inbraces{0,1}^n}{\inbraces{0,1}}$ with certificate complexity $C(f)$ and an input $x^{\star}$, output a size-$C(f)$ certificate for $f$’s value on $x^{\star}$.
\end{problem*}
\subsection{Overview of Our Algorithm}
We informally describe our algorithm. Without loss of generality, we let $x^{\star}$ be such that $f(x^{\star})=1$. A valid certificate is any subset $S\subset S_{x^{\star}}$ of indices such that $f(x_S)=1$.
We add elements into our certificate $A$ one-by-one. To do this, we simply iterate the following steps until $A$ is a valid certificate:
\begin{enumerate}
\item Find the smallest $s\in S_{x^{\star}}$ such that including all $i\leq s \in S_{x^{\star}}$ in the certificate, along with elements already in $A$, yields a valid certificate.
\item Add $s$ to $A$.
\end{enumerate}
Because the function is monotone, $s$ can be found through binary search at each step. Moreover, observe that removing any one element from $A$ no longer yields a valid certificate; thus, as we will show in Lemma~\ref{lemma:cf_minimal_certificate}, the output certificate is length at most $C(f)$. This also implies the algorithm makes a total of $O(C(f)\cdot\log{n})$ queries.
\subsection{Formal Description of Our Algorithm}\label{algo}
We state our algorithm formally. In our algorithm description and analysis, we assume without loss of generality that $f(x^{\star})=1$. We can make this assumption since if $f(x^{\star})=0$, we can instead run the algorithm making queries to $g(x) \coloneqq 1-f(\mathbbm{1}^n-x)$, which is a monotone function with $g(x^{\star})=1$.
\begin{definition}[$\mathsf{search}$]
\label{def:search}
The procedure $\mathsf{search}(f,A,S)$ acts on a monotone function $f\in \inbraces{0,1}^n\to \inbraces{0,1}$, and two sets $A,S\subseteq [n]$. If $f(x_A)=1$ or $f(x_{A\cup S})=0$, it outputs $\mathsf{ERROR}$. Else, it outputs the smallest $s\in S$ for which $f(x_{A\cup([s]\cap S)})=1$. The function proceeds using binary search, which can be done because $f$ is monotone.
\end{definition}
\protocol{Algorithm to Certify a Binary Monotone Function Where $f(x^{\star}) = 1$}{s33d}{
\begin{enumerate}
\item \textbf{Input:} Query access to a function $\myfunc{f}{\inbraces{0,1}^n}{\inbraces{0,1}}$ and point $x^{\star}$ for which $f(x^{\star}) = 1$.
\item Initialize the sets $A\gets \emptyset$ and $S\gets S_{x^{\star}}$
\item\label{binary-step} Run the following procedure until $f(x_A)=1$:
\begin{enumerate}
\item\label{search-step} Set $s\gets \mathsf{search}(f,A,S)$.
\item Add $s$ to $A$.
\item Set $S \gets S\cap [s-1]$
\end{enumerate}
\item \textbf{Output:} $A$.\label{line:output}
\end{enumerate}
}
\subsection{Analysis}
\begin{theorem}\label{mainthm}
The Algorithm in \ref{algo} outputs a certificate of length at most $C(f)$ for $f$ on $x^{\star}$ and makes at most $O(C(f)\cdot \log{n})$ queries.
\end{theorem}
We break the proof down into a series of lemmas.
\begin{lemma}
\label{lemma:cf_minimal_certificate}
If $S$ is a minimal certificate, then $\abs{S} \le C(f)$.
\end{lemma}
\begin{proof}
Consider the shortest certificate $C$ for the input $x_S$. We must have $f(x_C)=f(x_S)$, and $|C|\leq C(f)$. The fact that $|C|<|S|$ and $f(x_C)=f(x_S)$ contradicts that $S$ is minimal.
\end{proof}
\begin{lemma}\label{no-error}
The Algorithm in \ref{algo} never outputs $\mathsf{ERROR}$.
\end{lemma}
\begin{proof}
If the algorithm outputs $\mathsf{ERROR}$, it must be in Step~\ref{search-step}. By definition, an error occurs if $f(x_A)=1$ or $f(x_{A\cup S})=0$. The former cannot be true because the algorithm checks this exact condition in Step~\ref{binary-step}. The latter cannot be true because:
\begin{itemize}
\item If this is the first iteration of Step~\ref{binary-step}, $A\cup S=S_{x^{\star}}$, which means $f(x_{A\cup S_{x^{\star}}})=f(x^{\star})=1$.
\item Else, in the previous iteration of Step~\ref{search-step} (let the values of $A,S,s$ at that step be $A', S', s'$ respectively), it must have been the case that $f(x_{A'\cup(S'\cap[s'])})=1$. Note that $A'\cup (S'\cap[s'])=A\cup(S'\cap[s'-1])=A\cup S$, and so $f(x_{A\cup S})=1$.
\end{itemize}
\end{proof}
\begin{lemma}
If the Algorithm in \ref{algo} terminates, it outputs a minimal certificate for $f$ on $x^{\star}$.
\end{lemma}
\begin{proof}
It must be the case that $f(A)=1$; otherwise, we could not have left Step~\ref{binary-step}. Consider any $s\in A$ and we will show that $f(A\setminus s)=0$.
At the iteration of Step~\ref{binary-step} where $s$ was added to $A$ (let the temporary certificate $A$ at the start of that step be $A_s$), it must be the case that $f(x_{(S\cap[s])\cup A_s})=1$ but $f(x_{(S\cap[s-1])\cup A_s})=0$. All future elements that are added to create the final certificate $A$ must be a subset of $S\cap [s-1]$ (where $S$ is being referenced from the current iteration of Step~\ref{binary-step}). Therefore, $A\setminus S \subseteq (S\cap [s-1])\cup A_s$, and therefore $f(A\setminus s)=0$.
\end{proof}
\begin{lemma}
The Algorithm in \ref{algo} terminates, making at most $O(C(f)\log{n})$ queries.
\end{lemma}
\begin{proof}
Observe that in every iteration of the main loop, we add exactly one element to $A$. By Lemma \ref{lemma:cf_minimal_certificate}, there are at most $C(f)$ coordinates in the output $A$. Hence, we run the main loop at most $C(f)$ times.
Next, $\mathsf{search}(f, A, S)$ is a binary search over a domain of size $\abs{S} \le \abs{S_{x^{\star}}} \le n$. Therefore, $\mathsf{search}(f, A, S)$ uses at most $\log n$ queries.
Finally, the check $f(x^{\star}_A) = 1$ costs $1$ query, and this runs at the beginning of every iteration of the loop. In total, we make at most $C(f) \cdot (\log n + 1)$ queries, as desired.
\end{proof}
Combining these lemmas finishes the proof of Theorem~\ref{mainthm}.
\section{Finding the Shortest Certificate for a Monotone Function}
In this section, we show that there exists a family of instances on which the problem of finding the shortest certificate for a binary-valued $f$ on an input $x^{\star}$ (denoted $k\coloneqq C(f,x^{\star})$) requires at least $\Omega\inparen{\binom{n}{k}}$ queries.
Notice that this result is essentially optimal: for any function $f$, any input $x^{\star}$ and $k=C(f,x^{\star})$, $O\inparen{\binom{n}{k}}$ suffice to find a size-$k$ certificate. Assuming $f(x^{\star})=1$, the algorithm can simply query $f(x_S)$ for all subsets $S$ of size $k$ and check if each one of them is a certificate.
\begin{definition}[$F_k$] We define the set of $k$-indicator functions, denoted $F_k$ as follows.
Let $\myfunc{f_P}{\inbraces{0,1}^n}{\inbraces{0,1}}$ for some $P\subset[n]$ be defined as follows:
\[
f_P(x)=\begin{cases}
0 & \abs{S_x}<k\\
1 & \abs{S_x}>k\\
0 & \abs{S_x}=k, P\neq S_x\\
1 & \abs{S_x}=k, P=S_x
\end{cases}
\]
Finally, let $F_k=\inbraces{f_P{\;\; : \;\;} \abs{P}=k}$.
\end{definition}
\begin{lemma}
\label{lemma:cert_complexity_fk}
Every function $f\in F_k$ has $C(f,\mathbbm{1}^n)=k$.
\end{lemma}
\begin{proof}
It is easy to see that every function in $F_k$ is monotone for all $k$.
Let $P \subset [n]$ be such that $f_P = f$. Observe that $\abs{P} = k$. Next, notice that $f(\mathbbm{1}^n_P) = f(\mathbbm{1}^n) = 1$. This implies that $C(f,\mathbbm{1}^n) \le \abs{P} = k$. Finally, consider any $S \subset [n]$ such that $\abs{S} < k$. Note that for $x = \mathbbm{1}^n_S$, we have $\abs{S_x} < k$, so $f_P(x) = 0$. Thus, we have $C(f,\mathbbm{1}^n) \ge k$, and we're done.
\end{proof}
\begin{theorem}
For any $k \in [n-1]$ and (randomized) algorithm $\mathcal{A}$, there exists a function $\myfunc{f}{\inbraces{0,1}^n}{\inbraces{0,1}}$ and input $x^{\star}$ with $C(f,x^{\star})=k$ such that $\mathcal{A}$ must make at least $\nfrac{1}{2}\cdot\binom{n}{k}$ queries to $f$ to find the size-$k$ certificate with probability $\ge\nfrac{1}{2}$.
\end{theorem}
\begin{proof}
Fix an arbitrary $k \in [n-1]$. We will show that some function $f\in F_k$ takes $\ge\nfrac{1}{2}\cdot\binom{n}{k}$ queries to certify on the input $\mathbbm{1}^n$. Note that $C(f,\mathbbm{1}^n) = k$. Additionally, observe that any randomized algorithm to find a certificate for $f(\mathbbm{1}^n)=1$ can be converted to one that only makes queries $x$ satisfying $\abs{S_x} = k$.
Let $X=\inbraces{x\in\inbraces{0,1}^n {\;\; : \;\;} \abs{S_x} = k}$. Notice that $\abs{X} = \binom{n}{k}$. Any randomized algorithm for finding the single $x \in X$ such that $f(x)=1$ can be viewed as one that samples a permutation from some distribution over permutations of $X$ and makes queries to $f$ in the order determined by the permutation until the algorithm encounters the $x\in X$ for which $f(x) = 1$. This is because query $i$ only depends on the values of the queries $1, \dots, i - 1$, and not their responses -- in particular, the responses to queries $1, \dots, i - 1$ are all $0$ if the algorithm has not terminated prior to issuing query $i$. With this interpretation in mind, fix some distribution of permutations of $X$; call this distribution $\mathcal{P}$.
For each $x\in X$, consider $\prvv{P\in \mathcal{P}}{P^{-1}(x)\le\nfrac{\abs{X}}{2}}$ where $P^{-1}(x)$ is the index of element $X$. Let $\mu(P)$ denote the probability that a random permutation drawn from $\mathcal{P}$ is $P$, and observe the following manipulations:
\begin{align*}
\sum_{x\in X} \prvv{P\in \mathcal{P}}{P^{-1}(x)\le\nfrac{\abs{X}}{2}} &= \sum_{x\in X} \sum_{P} \mu(P) \cdot \indicator{P^{-1}(x) \le \nfrac{\abs{X}}{2}}\\
&= \sum_{P} \mu(P) \cdot \sum_{x \in X} \indicator{P^{-1}(x) \le \nfrac{\abs{X}}{2}}\\
&= \sum_{P} \mu(P) \cdot \frac{1}{2}\cdot\abs{X} = \frac{1}{2}\cdot\abs{X}
\end{align*}
Thus, there exists at least one $x \in X$ for which $\prvv{P\in \mathcal{P}}{P^{-1}(x)\le \nfrac{\abs{X}}{2}} \le \nfrac{1}{2}$. It follows that the algorithm does not find a sized-$k$ subset of $\mathbbm{1}^n$ certifying $f(\mathbbm{1}^n)=1$ with probability $>\frac12$ without making at least $\frac12\binom{n}{k}$ queries.
\end{proof}
\section{Introduction}
Given a function $\myfunc{f}{\inbraces{0,1}^n}{\mathbb{D}}$ for some output domain $\mathbb{D}$ and an input $x^{\star}$, is there a short proof for why $f(x^{\star})$ takes on the value it does? This natural question motivates the notion of \emph{certificate complexity} in complexity theory. Loosely speaking, a certificate for $f(x^{\star}) = y$ is a subset of the bits of $x^{\star}$ that ``fixes'' the value of $f(x^{\star})$. In other words, every input $x$ that agrees with $x^{\star}$ on the bits in the certificate will satisfy $f(x) = f(x^{\star})$. Besides being a quantity of interest in complexity theory and in the analysis of Boolean functions, certificate complexity has a natural interpretation in the context of explainable AI. Here, the practitioner aims to find simple properties of a given input that explain a classifier's prediction on the input.
We formalize the notion of a certificate in Definition \ref{def:certificate}.
\begin{definition}[Certificate (see, e.g., \cite{arorabarak})]
\label{def:certificate}
Let $x\vert_S$ denote the substring of $x$ in the coordinates of $S$.
For a function $f:\inbraces{0,1}^n\to \mathbb{D}$ and an input $x^{\star}\in\inbraces{0,1}^n$, we say a set $S \subseteq [n]$ is a certificate if for all $y\in\inbraces{0,1}^n$ such that $x^{\star}\vert_S=y\vert_S$, we have $f(x^{\star})=f(y)$.
\end{definition}
We use Definition \ref{def:certificate} to define the certificate complexity of a function $f$.
\begin{definition}[Certificate complexity (see, e.g., \cite{arorabarak})]
\label{def:cert_complexity}
For any function $\myfunc{f}{\inbraces{0,1}^n}{\mathbb{D}}$ and $x \in \inbraces{0,1}^n$, we let $C(f,x)$ be the smallest integer such that there exists a $C(f,x)$-sized certificate for $f(x)=j$. We now let the certificate complexity of $f$ be $\max\limits_{x\in\inbraces{0,1}^n} C(f,x)$.
\end{definition}
A natural follow-up question from Definition \ref{def:cert_complexity} is whether a short certificate can be found in a given input if we know that all inputs have a short certificate. The following problem, posed and studied in the STOC '22 paper of \cite{BKLT22}, formalizes this question.
\begin{problem*}
Given queries to a function $f : \inbraces{0,1}^n \to \inbraces{0,1}$ with certificate complexity $C(f)$ and an input $x^{\star}$, output a size-$C(f)$ certificate for $f$’s value on $x^{\star}$.
\end{problem*}
The main result of \cite{BKLT22} is an algorithm for the case where $f$ is monotone and the output range $\mathbb{D}=\inbraces{0,1}$. The authors design a randomized algorithm that makes at most $O\inparen{C(f)^8\cdot\log{n}}$ queries using a novel connection to threshold phenomena. Furthermore, \cite{BKLT22} show that $\Omega(C(f)\cdot\log{n})$ queries for the certification problem are necessary in the worst-case. The authors identify closing this gap as a concrete direction for future work.
\subsection{Our Results}
Our main result is a simple, deterministic algorithm that makes $O(C(f)\cdot\log{n})$ queries to find a size-$C(f)$ certificate for any monotone binary-valued function $f$ and input $x^{\star}$. This completely resolves the aforementioned open question from \cite{BKLT22}. Formally, we have Theorem \ref{thm:main_result_intro}.
\begin{theorem}
\label{thm:main_result_intro}
Given query access to a monotone function $\myfunc{f}{\inbraces{0,1}^n}{\inbraces{0,1}}$ and an input $x^{\star}$, there exists an algorithm that makes $O(C(f)\cdot\log n)$ queries to $f$ and outputs a size-$C(f)$ subset $S$ corresponding to a subset of indices of $x^{\star}$ certifying the value of $f(x^{\star})$.
\end{theorem}
We can extend our result to obtain as a simple corollary an algorithm that finds a size-$2C(f)$ certificate for any monotone \textbf{real}-valued function $f$ and input $x^{\star}$. Specifically, we have Theorem \ref{thm:main_result_reals_intro}.
\begin{theorem}
\label{thm:main_result_reals_intro}
Given query access to a monotone function $\myfunc{f}{\inbraces{0,1}^n}{\mathbb{R}}$ and an input $x^{\star}$,there exists an algorithm that makes $O(C(f)\cdot \log n)$ queries to $f$ and outputs a size-$2C(f)$ subset $S$ corresponding to a subset of indices of $x^{\star}$ certifying the value of $f(x^{\star})$.
\end{theorem}
The careful reader might also wonder why we are only looking for a certificate of size-$C(f)$ on the input $x^{\star}$ -- by definition, the shortest certificate on a fixed input $x^{\star}$ is size-$C(f,x^{\star})$. We show that finding a certificate of length $C(f,x^{\star})$ may require far more queries than simply finding one of length $C(f)$. In particular, in the case where $C(f,x^{\star}) = \nfrac{n}{2}$, it may require exponentially many queries. Moreover, our result matches the trivial upper bound provided by an algorithm which simply queries all $C(f,x^{\star})$-size certificates. See Theorem \ref{thm:lower_bound_intro}.
\begin{theorem}
\label{thm:lower_bound_intro}
For any $k$, for any (randomized) algorithm that queries a given function, there exists a function $f$ and input $x^{\star}$ such that $k=C(f,x^{\star})$, and the algorithm must make at least $\frac12\binom{n}{k}$ queries to determine a certificate with probability $>\nfrac{1}{2}$.
\end{theorem}
\subsection{Related Work}
We derive our setting and problem statements from the work of \cite{BKLT22}. The authors of \cite{BKLT22} formally propose the problem of certifying a monotone function $f$ on an input $x^{\star}$ and provide an algorithm for doing so, as mentioned earlier. They also look at the certification question for a general (non-monotone) function $f$. Here, they show $\Omega\inparen{2^{C(f)}+C(f)\cdot\log{n}}$ queries are necessary, and $O\inparen{2^{C(f)}\cdot C(f)\cdot\log{n}}$ queries suffice with high probability. Closing this gap remains an interesting open direction.
Before the work of \cite{BKLT22}, Angluin (see \cite{angluin}) gave a local search algorithm that can be used to certify a monotone function $f$ on an input $x^{\star}$ with query complexity $O(n)$. See Appendix C in \cite{BKLT22} for a detailed exposition and proof of correctness of Angluin's algorithm.
\section{Technical Overview}
\subsection{Algorithm for Certifying a Binary Monotone Function}
\paragraph{Setup} We describe our algorithm for finding a $C(f)$-size certificate for a binary-valued monotone function $f$ on an input $x^*$. Without loss of generality, we let $f(x^*)=1$.\footnote{It may be worth reading the notation section of the preliminaries, though the rest of the formal definitions should not be necessary.}
Let $S_{x^*}$ be the set of indices on which $x^*$ is 1. Note that $S_{x^*}$ is a valid certificate, but not necessarily small enough. Our final certificate will be a subset of $S_{x^*}$; without loss of generality, we will let $S_{x^*}=[r]$ for some $r\geq 0$. A certificate $S$ is \emph{minimal} if $f(x_S)=1$, but by removing an element of $S$ to obtain $P\subset S$, we always have $f(x_P)=0$ ($S=\emptyset$ if $f(0^n)=1$).
The first observation is that \emph{any minimal certificate $S$ has length $\leq C(f)$}. This follows because on the input $x_S$, one must witness all of $S$ to see that the function value is $1$. Thus, our goal is to find any minimal certificate. If $f(x_{\emptyset})=1$, our minimal certificate is $\emptyset$.
\mnote{I do think the next section can probably be simplified significantly, esp since I didn't realize the phrasing "find the smallest index for which this holds" lmao}
\paragraph{Finding the Minimal Certificate} We now add elements into our minimal certificate. We want to find $a_1$ such that $f(x_{[a_1-1]})=0$ but $f(x_{[a_1]})=1$ (the set $[0]=\emptyset$ since we are 1-indexing). In other words, we want to find the value $a$ at which the function ``transitions'' from $0$ to $1$. There is exactly one such transition since $f$ is monotone (and therefore increasing with $a$. This can be found by binary searching. The value $a_1$ is now the first element of our certificate. If $f(x_{a_1})=1$, then $a_1$ is a minimal certificate, since $f(x_\emptyset)=0$.
Otherwise, we repeat the procedure to find $a_2<a_1$ such that $f(x_{[a_2-1]\cup a_1})=0$ but $f(x_{[a_2]\cup a_1})=1$.
This must exist because when $a_2=0$, $f(x_{[a_2-1]\cup a_1})=0$, but when $a_2=a_1-1$, we have $f(x_{[a_2-1]\cup a_1})=1$, so there must be a transition point.
Then $a_2$ becomes the second element of our certificate. Now if $f(x_{\{a_1,a_2\}})=1$, then $\inbraces{a_1,a_2}$ is a minimal certificate, since $f(x_{\{a_1\}})=0$ by assumption, and $f(x_{\{a_2\}})=0$, because of the stronger statement that $f(x_{[a_1-1]})=0$ (and the function is monotone). Otherwise, we add another element to our certificate similarly.
To add the general $(i+1)$th element to our certificate, find $a_{i+1}$ such that $f(x_{[a_{i+1}-1]\cup a_1\ldots a_i})=0$ but $f(x_{[a_{i+1}]\cup a_1\ldots a_i})=1$. If we then have $f_{\inbraces{a_1\ldots a_{i+1}}}=1$, we must have that $\inbraces{a_1\ldots a_{i+1}}$ is a minimal certificate. This is because for any $j\leq i+1$, we know that for $P=[j-1]\cup a_1\ldots a_{j-1}$, $x_{P}=0$ and $P$ is a superset of $\inbraces{a_1\ldots a_{j-1},a_{j+1}\ldots a_{i+1}}$, so the value on that set must also be $0$.
This procedure eventually terminates with a certificate; we can repeat unless $a_i=1$, and if that's the case we must have a certificate already. Note that adding each element takes up to $O(\log{n})$ queries. Since the certificate is minimal, the procedure took less than $C(f)$ steps, so the final query complexity is $O(C(f)\log{n})$.
\section{Preliminaries}
\paragraph{Notation} In this work, we use the following notation.
\begin{itemize}
\item We denote the set $\inbraces{x \in \mathbb{Z}_{\ge 0} {\;\; : \;\;} 1 \le x \le n}$ as $[n]$. In an abuse of notation, let $[0] = \emptyset$.
\item For a set $S \subseteq [n]$, we write $x_S$ to be the indicator vector for $S$; i.e., $x_S$ is such that $x_i = \indicator{i \in S}$, for all $i \in [n]$. Additionally, we write $x\vert_S$ to be the substring of $x$ in the coordinates of $S$. Specifically, we have $x\vert_S = \inbraces{(i, x_i) \text{ for all } i \in S}$.
\item Let $\mathbbm{1}^n$ denote the all-$1$s vector in $n$ dimensions.
\item For a vector $x\in \{0,1\}^n$, we denote $S_{x}$ to be $\inbraces{i {\;\; : \;\;} x_i=1}$.
\end{itemize}
In our work, it is helpful to distinguish a \emph{minimal} certificate from a general certificate.
\begin{definition}[Minimal Certificate]
\label{def:minimal_certificate}
For a given function $f$, we say a certificate $S \subseteq [n]$ is \emph{minimal} if for all $a\in S$, we have that $S \setminus a$ is not a certificate for $f(x)$.
If $f$ is monotone, this is equivalent to requiring that for all $A\subset S$, we have $f(x\vert_A)\neq f(x\vert_S)$.
\end{definition}
Finally, we note the information-theoretic lower bound from \cite{BKLT22} on the query complexity of any algorithm used to certify $f(x)$ for a monotone function $f$.
\begin{lemma}[Claim 1.2 in \cite{BKLT22}]
For any $c < 1$ and any $k\le l \le n^c$, let $\mathcal{A}$ be an algorithm which, given query access to a monotone function $\myfunc{f}{\inbraces{0,1}^n}{\inbraces{0,1}}$ with certificate complexity $\le k$ and an input $x^{\star}$, returns a size-$l$ certificate for $f$'s value on $x^{\star}$ with high probability. The query complexity of $\mathcal{A}$ must be $\Omega\inparen{k\log n}$.
\end{lemma}
\section{Extension to Real-Valued Functions}
In this section, we prove the following corollary of our main result wherein the output domain is $\mathbb{R}$ instead of $\inbraces{0,1}$.
\begin{corollary}
There exists an algorithm that, given an input $x^{\star}$ and query access to a monotone $\myfunc{f}{\inbraces{0,1}^n}{\mathbb{R}}$, makes $O(C(f) \cdot \log n)$ queries to $f$ and outputs a size-$2\cdot C(f)$ certificate for $f(x^{\star})$.
\end{corollary}
\subsection{Our Algorithm}\label{reals}
We begin with two necessary definitions.
\begin{definition}[$\mathsf{binary\_cert}(f,x^{\star})$]
The procedure $\mathsf{binary\_cert}(f,x^{\star})$ is given query access to function $f:\inbraces{0,1}^n\to\inbraces{0,1}$ and an input $x^{\star}$, runs our algorithm from Section~\ref{sec:bin-search}, and outputs a size-$C(f)$ certificate for $f(x^{\star})$.
\end{definition}
\begin{definition}[$g_{0,f,x^{\star}}(x)$, $g_{1,f,x^{\star}}(x)$]
Let $b\in \inbraces{0,1}$, $f:\inbraces{0,1}^n\to\mathbb{R}$ and $x^{\star}\in\{0,1\}^n$. The function $g_{b,f,x^{\star}}:\inbraces{0,1}^n\to\inbraces{0,1}$ is defined as follows:
\[
g_b(x)=\begin{cases}
0 & f(x)<f(x^{\star}) \\
1 & f(x)>f(x^{\star}) \\
b & f(x)=f(x^{\star})
\end{cases}
\]
We will abbreviate $g_{b,f,x^{\star}}:\inbraces{0,1}^n\to\inbraces{0,1}$ as $g_b$ when $f$ and $x^{\star}$ are clear.
\end{definition}
\protocol{Algorithm to Certify a Real-Valued Monotone Function}{reals}{
\begin{enumerate}
\item \textbf{Input:} $f, x^{\star}$.
\item Set $C_0 \gets \mathsf{binary\_cert}(g_0,x^{\star})$.
\item Set $C_1 \gets \mathsf{binary\_cert}(g_1,x^{\star})$.
\item \textbf{Output:} $C_0 \cup C_1$.
\end{enumerate}
}
\subsection{Analysis}
\begin{theorem}\label{thm:reals}
The Algorithm in \ref{reals} outputs a certificate for $f$ of length at most $2C(f)$ and makes at most $O(C(f)\log{n})$ queries.
\end{theorem}
We break the proof into a series of lemmas. Call the output $A$.
\begin{lemma}
$A$ is a valid certificate for $f$ on $x^{\star}$.
\end{lemma}
\begin{proof}
For any input $y$ such that $y\vert_A=x^{\star}\vert_A$, we must have $g_b(y)=g_b(x^{\star})$ for both $b=0,1$. Notice that both of the following hold:
\begin{align*}
g_0(y)=g_0(x^{\star}) &\text{ implying } f(y)\leq f(x^{\star}) \\
g_1(y)=g_1(x^{\star}) &\text{ implying } f(y)\geq f(x^{\star})
\end{align*}
Hence, we have $f(y)=f(x^{\star})$.
\end{proof}
\begin{lemma}
$\abs{A}\leq 2C(f)$.
\end{lemma}
\begin{proof}
It suffices to show that $C(f)\geq C(g_b)$ for $b\in \inbraces{0,1}$. We will show that any certificate $B$ for $f$ on $x$ is also a certificate for $g_b$. For all $y,y'$ with $y\vert_C=y'\vert_C$, we have $f(y) = f(y')$, but this implies by definition that $g_b(y)=g_b(y')$. Hence, $B$ is also a certificate for $g_b$.
\end{proof}
Combining these lemmas concludes the proof of Theorem~\ref{thm:reals}. |
\section{#1}\setcounter{equation}{0}}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\newcommand{\langle\!\langle}{\langle\!\langle}
\newcommand{\rangle\!\rangle}{\rangle\!\rangle}
\newcommand{\bd}[1]{\boldsymbol{#1}}
\newcommand{\mathbb{I}}{\mathbb{I}}
\newcommand{\mathbb{I}_{\mathcal{TH}}}{\mathbb{I}_{\mathcal{TH}}}
\newcommand{\mathbb{I}_{\mathcal{SH}}}{\mathbb{I}_{\mathcal{SH}}}
\newcommand{NS\textrm{-}NS}{NS\textrm{-}NS}
\newcommand{R\textrm{-}NS}{R\textrm{-}NS}
\newcommand{NS\textrm{-}R}{NS\textrm{-}R}
\newcommand{R\textrm{-}R}{R\textrm{-}R}
\newcommand{(s,\bar{s},t)}{(s,\bar{s},t)}
\newcommand{(s,\bar{s})}{(s,\bar{s})}
\newcommand{(\bd{\lambda}+\bar{\bd{\lambda}})}{(\bd{\lambda}+\bar{\bd{\lambda}})}
\newcommand{\longrightleftarrows}[1]{\mathrel{\substack{\xrightarrow{#1} \\[-.9ex] \xleftarrow{#1}}}}
\begin{document}
\baselineskip=17pt
\begin{titlepage}
\rightline{\tt YITP-22-32}
\begin{center}
\vskip 2.5cm
{\Large \bf {Open-closed homotopy algebra\\
in superstring field theory}}
\vskip 1.0cm
{\large {Hiroshi Kunitomo}}
\vskip 1.0cm
{\it {Center for Gravitational Physics and Quantum Information}},\\
{\it {Yukawa Institute for Theoretical Physics}}
{\it {Kyoto University}},\\
{\it {Kitashirakawa Oiwakecho, Sakyo-ku, Kyoto 606-8502, Japan}}\\
<EMAIL>
\vskip 2.0cm
{\bf Abstract}
\end{center}
\noindent
We construct open-closed superstring interactions based on the open-closed
homotopy algebra structure. It provides a classical open superstring field theory
on general closed-superstring-field backgrounds described by classical solutions
of the nonlinear equation of motion of the closed superstring field theory.
We also give the corresponding WZW-like action through the map connecting
the homotopy-based and WZW-like formulations.
\end{titlepage}
\tableofcontents
\newpage
\sectiono{Introduction}
It is known that several homotopy algebras are naturally realized as algebraic structures
in string field theories and play a significant role. This was first recognized in closed
bosonic string field theory \cite{Zwiebach:1992ie,Lada:1992wc},
where the $L_\infty$ structure determines the (classical) gauge-invariant action.
Open bosonic string field theory was first formulated as a cubic theory using the
(Witten's) associative product \cite{Witten:1985cc} but can be extended to that
with an $A_\infty$ structure more generally \cite{Nakatsu:2001da,Kajiura:2003ax}.
This is also deformed to the theory on general closed string backgrounds
\cite{Zwiebach:1990qj,Zwiebach:1997fe,Zwiebach:1992bw} based on
the open-closed homotopy algebra (OCHA) structure
\cite{Kajiura:2004xu,Kajiura:2005sn,ocha coalgebra}.
In the superstring field theories, the homotopy algebra structure is more important.
Since it seems inevitable to avoid associativity anomaly \cite{Wendt:1987zh}, the $A_\infty$
structure becomes essential to determine the gauge-invariant action in the open superstring
field theory \cite{Erler:2013xta,Erler:2016ybs,Kunitomo:2020xrl}.
The $L_\infty$ structure again plays the role of guiding principle to determine the action
with appropriate picture numbers in the heterotic and type II superstring field theories
\cite{Erler:2014eba,Jurco:2013qra,Kunitomo:2019glq,Kunitomo:2019kwk,Kunitomo:2021wiz}.
On the other hand, the current understanding is that there is no essential difference
between the theory of open string/closed string mixed system and the theory of purely
closed string. It merely describes the perturbation on the different backgrounds,
those with and without D-brane \cite{Polchinski:1995mt,Polchinski:1996fm}.
They should be derived from non-perturbatively formulated fundamental theory such as
string field theory, but it is not a priori clear which one should be considered more
fundamental. The closed string field theory is simpler, but the open-closed string
field theory has a larger symmetry structure, the OCHA structure\footnote{
In a formulation that introduces an auxiliary degree of freedom,
the open-closed superstring field theory has already been constructed
\cite{FarooghMoosavian:2019yke}.
It does not, however, decrease the worthwhile to construct the theory based on
the OCHA structure.}. The purpose of
this paper is to construct an open-closed string field theory realizing the OCHA
structure. The action obtained explains the classical open string field theory
on general closed-string backgrounds.
The paper is organized as follows. In section \ref{A}, we briefly review
the open superstring field theory with general $A_\infty$ structure.
After introducing some conventions and fundamental ingredients, we show how
we construct the open superstring field theory based on the $A_\infty$
structure. The superstring products with appropriate picture numbers
satisfying the $A_\infty$ relations can be
obtained by recursively solving the differential equations.
We similarly review the closed superstring field theory with the $L_\infty$ structure
in section \ref{B}.
We define the string products multiplying both open and closed string field
in section \ref{C} and show the relations they must satisfy to form
the OCHA. We also give the differential equations that the products
with OCHA structure should follow. They provide an action
of the open superstring field theory on the general closed superstring backgrounds.
In section \ref{D}, we obtain, as a byproduct, the corresponding WZW-like action through
the map connecting the homotopy-based and WZW-like formulations,
which is a generalization considered in \cite{Maccaferri:2021lau}.
Section \ref{E} is devoted to the summary and
discussion. Appendix \ref{App A} is added to make the paper self-contained.
We introduce two composite string fields, the pure-gauge open string field
and the associated open string field, which is nontrivial in the theory with
general $A_\infty$ structure.
\sectiono{Open superstring field theory with $A_\infty$-structure}\label{A}
We summarize in this section how the open superstring field theory is
constructed based on the $S_\infty$ algebra structure.
\subsection{Open superstring field}
The first-quantized Hilbert space
of open superstring is composed of two sectors:
$\mathcal{H}_o = \mathcal{H}_{NS} + \mathcal{H}_R$.
Correspondingly,
the open superstring field $\Psi$ has two components: $\Psi = \Psi_{NS} + \Psi_R$,
both of which are Grassmann odd and have ghost number $1$.
The component $\Psi_{NS}$ ($\Psi_R$) has picture number $-1$ ($-1/2$)
and represents space-time bosons (fermions).
We impose on it a constraint
\begin{equation}
\mathcal{P}^o_{XY}\Psi\ =\ \Psi,\qquad
\mathcal{P}^o_{XY}\ =\ \mathcal{G}^o(\mathcal{G}^o)^{-1},
\label{open constraint}
\end{equation}
with
\begin{equation}
\mathcal{G}^o\ =\ \pi^0 + X^o\pi^1,\qquad
(\mathcal{G}^o)^{-1}\ =\ \pi^0 + Y^o\pi^1,
\end{equation}
where $\pi^0$ and $\pi^1$ are the projection operators onto the NS and R components,
respectively: $\pi^0\Psi = \Psi_{NS}$ and $\pi^1\Psi = \Psi_R$. The picture changing operator
(PCO) of open superstring $X^o$ and its inverse $Y^o$ are defined by
\begin{equation}
X^o\ =\ -\delta(\beta_0)G + (\gamma_0\delta(\beta_0)+\delta(\beta_0)\gamma_0)b_0,\qquad
Y^o\ =\ - \frac{G}{L_0}\delta(\gamma_0).
\end{equation}
The PCO $X^o$ is BRST exact in the large Hilbert space:
\begin{equation}
X^o=[Q,\Xi^o],\qquad \Xi^o\ =\ \xi_0 + (\Theta(\beta_0)\eta\xi_0-\xi_0)P_{-3/2}
+ (\xi_0\eta\Theta(\beta_0)-\xi_0)P_{-1/2},
\end{equation}
where $P_{-3/2}$ ($P_{-1/2}$) is the projection operator onto the states with
picture number $-3/2$ ($-1/2$).
We call the Hilbert space restricted by the constraint (\ref{open constraint})
the restricted Hilbert space and denote $\mathcal{H}_o^{res}$.
Note that $\mathcal{G}^o$ and $(\mathcal{G}^o)^{-1}$ satisfy
\begin{equation}
\mathcal{G}^o(\mathcal{G}^o)^{-1}\mathcal{G}^0\ =\ \mathcal{G}^o,\qquad
(\mathcal{G}^o)^{-1}\mathcal{G}^o(\mathcal{G}^o)^{-1}\ =\ (\mathcal{G}^o)^{-1},\qquad
[Q,\mathcal{G}^o]\ =\ 0,
\label{G open}
\end{equation}
and thus $\mathcal{P}^o_{XY}$ is a projection operator that is compatible
with the BRST cohomology: $Q\mathcal{P}^o_{XY}=\mathcal{P}^o_{XY}Q\mathcal{P}^o_{XY}$.
The open superstring field satisfying (\ref{open constraint}) is expanded in the ghost zero-modes as
\begin{equation}
\Psi\ =\ (\phi_{NS} - c_0\psi_{NS}) + \left(\phi_R-\frac{1}{2}(\gamma_0+c_0G)\psi_R\right)\
\in \mathcal{H}_o^{res}.
\end{equation}
Natural symplectic form $\omega_s^o$ and $\Omega^o$ in $\mathcal{H}_o$ and $\mathcal{H}_o^{res}$,
respectively, are defined by using the BPZ inner product as
\begin{align}
\omega_s^o(\Psi_1,\Psi_2)\ =&\ (-1)^{\textrm{deg}(\Psi_1)}\langle\Psi_1|\Psi_2\rangle,
\label{symp open s}\\
\Omega^o(\Psi_1,\Psi_2)\ =&\ (-1)^{\textrm{deg}(\Psi_1)}\langle\Psi_1|(\mathcal{G}^o)^{-1}|\Psi_2\rangle,
\label{symp open res}
\end{align}
where deg$(\Psi)=1$ or $0$ if $\Psi$ is Grassmann even or odd, respectively.
We also use a natural symplectic form $\omega_l^o$ in the large Hilbert space $\mathcal{H}_l^o$,
which is similarly defined using the BPZ inner product in $\mathcal{H}_l^o$,
and related to $\omega_s^o$ as $\omega_l^o(\xi_0\Psi_1,\Psi_2)=\omega_s^o(\Psi_1,\Psi_2)$
if $\Psi_1,\Psi_2\in\mathcal{H}_s^o$.
\subsection{Interaction with $A_\infty$-structure}
Open superstring interactions are described by the string products $M_n$ mapping $n$ open
superstring fields to an open superstring field as
\begin{alignat}{3}
M_n\ :\qquad\qquad
&\ (\mathcal{H}_o^{res})^{\otimes n} &\qquad &\ \longrightarrow&\qquad &\ \mathcal{H}_o^{res},\qquad
(n\ge1),
\nonumber\\
&\ \ \ \ \text{\rotatebox[origin=c]{90}{$\in$}}&\qquad &\ &\qquad &\ \ \text{\rotatebox[origin=c]{90}{$\in$}}\\
\ \ \ \Psi_1&\otimes \ \cdots\otimes\Psi_n &\qquad &\ \longmapsto&\qquad M_n(\Psi_1&,\ \cdots, \Psi_n).
\nonumber
\end{alignat}
We identify the one-string product as the open superstring BRST operator: $M_1=Q_o$.
Note that the conditions
\begin{equation}
\mathcal{P}^o_{XY}M_n(\Psi_1,\cdots,\Psi_n)\ =\ M_n(\Psi_1,\cdots,\Psi_n)
\end{equation}
hold by definition. The multi-linear maps $M_n$ further satisfy the $A_\infty$ relations
\begin{equation}
\sum_{m=0}^n\sum_{k=0}^{n-m}
(-1)^{\epsilon(1,k)}
M_{n-m+1}(\Psi_1,\cdot,\Psi_k,
M_{m+1}(\Psi_{k+1},\cdots,\Psi_{k+m+1}),\Psi_{k+m+2},\cdots,\Psi_n)\ =\ 0,
\label{A infinity}
\end{equation}
where $\epsilon(1,k)=\sum_{i=1}^k\textrm{deg}(\Psi_i)$, and cyclicity with respect to
the symplectic form $\Omega^o$,
\begin{equation}
\Omega^o(\Psi_1,M_n(\Psi_2,\cdots,\Psi_{n+1}))\ =\
-(-1)^{\textrm{deg}(\Psi_1)}\Omega^o(M_n(\Psi_1,\cdots,\Psi_n),\Psi_{n+1}).
\label{cyclicity open}
\end{equation}
The linear maps satisfying (\ref{A infinity}) and (\ref{cyclicity open}) form
the cyclic $A_\infty$ algebra $(\mathcal{H}_o^{res},\Omega^o,\{M_m\})$.
Coalgebra representation allows us to describe these infinite number of relations
of maps $M_n$ concisely \cite{Erler:2015uba}. The set of maps $\{M_n\}$ are represented by
a degree-odd coderivation
$\bd{M}=\sum_{n=1}^\infty \bd{M}_{n}$
acting on the tensor algebra
$\mathcal{TH}_o=\sum_{n=0}^\infty(\mathcal{H}_o^{res})^{\otimes n}$
as
\begin{equation}
\bd{M}\ =\ \sum_{n=1}^\infty\ \bd{M}_n\
=\ \sum_{n=1}^\infty\sum_{k,l=0}^\infty\left(\mathbb{I}^{\otimes k}\otimes M_n\otimes \mathbb{I}^{\otimes l}
\right)\pi^o_{k+n+l},
\end{equation}
where $\pi_m^o$ is the projection operator onto
$(\mathcal{H}_o^{res})^{\otimes m}\subset\mathcal{TH}_o$.
Then the $A_\infty$ relations in Eq.~(\ref{A infinity}) is concisely written as\footnote{
In this paper, $[\ ,\ ]$ denotes the graded commutator.}
\begin{equation}
[\bd{M},\bd{M}]\ =\ 0.
\end{equation}
For open superstring field theory, the string interaction $M_n$ must be defined
for each combination of NS and R inputs
so that the picture number must be conserved\footnote{
Whether the output is NS or R string is determined by the space-time fermion number
conservation.}:
\begin{equation}
\bd{M}_{n+1}\ =\ \sum_{p+r=n}\bd{M}_{p+r+1}^{(p)}|_{2r},
\end{equation}
where $p$ is the picture number that the the map itself has and
$2r$ is the Ramond number ($=$ number of Ramond inputs $-$ number of Ramond output).
The action with $A_\infty$ structure is given by
\begin{equation}
I_o\ =\ \int_0^1dt\, \Omega^o\left(\Psi,\pi_1^o\bd{M}\left(\frac{1}{1-t\Psi}\right)\right),
\label{A action}
\end{equation}
where we introduce a real parameter $t\in[0,1]$ and the group-like element $\frac{1}{1-\Psi}$ defined by
\begin{equation}
\frac{1}{1-\Psi}\ =\ \mathbb{I}_{\mathcal{TH}_o}+\sum_{n=1}^\infty\Psi^{\otimes n}.
\end{equation}
Here, $\mathbb{I}_{\mathcal{TH}_o}$ is the identity in $\mathcal{TH}_o$ satisfying
$\mathbb{I}_{\mathcal{TH}_o}\otimes V = V = V\otimes\mathbb{I}_{\mathcal{TH}_o}$ for ${}^\forall V\in\mathcal{TH}_o$.
The arbitrary variation of $I_o$ is given by
\begin{equation}
\delta I_o\ =\ \Omega^o\left(\delta\Psi,\pi_1^o\bd{M}\left(\frac{1}{1-\Psi}\right)\right),
\end{equation}
where we used the cyclicity (\ref{cyclicity open}).
We can show that the action (\ref{A action}) is invariant under the gauge transformation
\begin{equation}
\delta_{\Lambda}\Psi\ =\ \pi_1^o\bd{M}\left(\frac{1}{1-\Psi}\otimes\Lambda\otimes\frac{1}{1-\Psi}\right),
\end{equation}
using the $A_\infty$ relation (\ref{A infinity}) and cyclicity (\ref{cyclicity open}):
\begin{align}
\delta_{\Lambda}I_o\ =&\
\Omega^o\left(\pi_1^o\bd{M}\left(\frac{1}{1-\Psi}\otimes\Lambda\otimes\frac{1}{1-\Psi}\right),
\pi_1^o\bd{M}\left(\frac{1}{1-\Psi}\right)\right)
\nonumber\\
=&\ \Omega^o\left(\Lambda, \pi_1^o\bd{M}\left(\frac{1}{1-\Psi}\otimes
\pi_1^o\bd{M}\left(\frac{1}{1-\Psi}\right)\otimes\frac{1}{1-\Psi}\right)\right)
\nonumber\\
=&\ \Omega^o\left(\Lambda,\pi_1^o\bd{MM}\left(\frac{1}{1-\Psi}\right)\right)\ =\ 0.
\end{align}
\subsection{Explicit construction of interactions}
The cyclic $A_\infty$ algebra $(\mathcal{H}_o^{res},\Omega^o,\bd{M})$ for open superstring field theory
is constructed in two steps. First, we consider a cyclic $A_\infty$ algebra
$(\mathcal{H}_l^o,\omega_l^o,\bd{Q}-\bd{\eta}+\bd{A})$.
A degree odd coderivation
\begin{equation}
\bd{A}=\sum_{p,r=0}^\infty\bd{A}_{p+r+1}^{(p)}|^{2r}\qquad
(\bd{A}_1^{(0)}\mid^0\equiv 0)
\label{A for A-inf}
\end{equation}
is defined respecting the cyclic Ramond number ($=$ number of Ramond inputs $+$
number of Ramond output) to make it easier to realize cyclicity\footnote{
Note that the cyclic Ramond number has the upper bound $p+2\ge r$. We consider
$\bd{A}_{p+r+1}^{(p)}\mid^{2r}\equiv 0 $ against the outside of this region.}.
This another $A_\infty$ algebra can be decomposed into two mutually commutative
$A_\infty$ algebras $(\mathcal{H}_l,\bd{\mathcal{D}})$ and $(\mathcal{H}_l,\bd{\mathcal{C}})$ with
\begin{equation}
\pi_1\bd{\mathcal{D}}\ =\ \pi_1\bd{Q} +\pi_1^0\bd{A},\qquad
\pi_1\bd{\mathcal{C}}\ =\ \pi_1\bd{\eta} -\pi_1^1\bd{A}
\end{equation}
depending on the picture number deficit of the output.
The $A_\infty$ relation $[\bd{A},\bd{A}]=0$ can also be decomposed as
\begin{subequations} \label{A inf bare}
\begin{align}
[\bd{Q},\bd{A}] + \frac{1}{2}[\bd{A},\bd{A}]^1\ =\ 0,\\
[\bd{\eta},\bd{A}] - \frac{1}{2}[\bd{A},\bd{A}]^2\ =\ 0,
\end{align}
\end{subequations}
where the bracket with subscript $[\cdot,\cdot]^{1\,\textrm{or}\,2}$ is defined by
projecting the intermediate state onto the NS or R state after taking the (graded) commutator.
The relation $[\cdot,\cdot]=[\cdot,\cdot]^1+[\cdot,\cdot]^2$ holds
since the intermediate state is either the NS state or R state.
If such $A_\infty$ algebras are obtained,
we can transform them by the cohomomorphism
\begin{equation}
\hat{\bd{F}}^{-1}\ =\ \pi_1\mathbb{I} - \Xi^o\pi_1^1\bd{A}
\label{cohomo A}
\end{equation}
to the cyclic $A_\infty$ algebra of interest $(\mathcal{H}_o^{res},\Omega^o,\bd{M})$
and a (trivial) $A_\infty$ algebra $(\mathcal{H}^o_l,\bd{\eta})$:
\begin{equation}
\pi_1\hat{\bd{F}}^{-1}\bd{\mathcal{D}}\hat{\bd{F}}\ =\ \pi_1\bd{Q} + \mathcal{G}^o\pi_1\bd{A}\hat{\bd{F}}\
\equiv\ \pi_1\bd{M},\qquad
\pi_1\hat{\bd{F}}^{-1}\bd{\mathcal{C}}\hat{\bd{F}}\ =\ \pi_1\bd{\eta}.
\end{equation}
We consider a generating function
\begin{equation}
\bd{A}(s,t)\ =\ \sum_{m,p,r=0}^\infty s^mt^p\bd{A}_{m+p+r+1}^{(p)}|^{2r}
\label{gen fun A}
\end{equation}
for constructing the $A_\infty$ algebra $(\mathcal{H}_l^o,\omega_l^o,\bd{Q}-\bd{\eta}+\bd{A})$
and extend the $A_\infty$ relation (\ref{A inf bare}) to
\begin{subequations}\label{ext A infinity rel}
\begin{align}
\bd{I}(s,t)\ \equiv\ [\bd{Q},\bd{A}(s,t)] + \frac{1}{2}[\bd{A}(s,t),\bd{A}(s,t)]_{\mathfrak{o}_1(s)}\ =&\ 0,\\
\bd{J}(s,t)\ \equiv\ [\bd{\eta},\bd{A}(s,t)] - \frac{1}{2}[\bd{A}(s,t),\bd{A}(s,t)]_{\mathfrak{o}_2(t)}\ =&\ 0,
\end{align}
\end{subequations}
by introducing parameters $s$ and $t$ counting the picture number deficit and the picture number,
respectively.
Here, in Eqs.~(\ref{ext A infinity rel}), $\mathfrak{o}_1(s)\ =\ \pi^0+s\pi^1,\ \mathfrak{o}_2(t)\ =\ t\pi^1$ and
the bracket with subscript $[\cdot,\cdot ]_{\mathcal{O}}$ is another simple notation for $[\cdot,\cdot]^{1,2}$
and is defined by inserting the operator $\mathcal{O}$ into the intermediate state after taking (graded) commutation relation,
\begin{equation}
\pi_1[\bd{D},\bd{D}']_{\mathcal{O}}\ =\ \sum_n\pi_1\Big( \bd{D}_n\big(\mathcal{O}\pi_1\bd{D}'\wedge\mathbb{I}_{n-1})
-(-1)^{DD'+\mathcal{O}(D+D')}\bd{D}'_n\big(\mathcal{O}\pi_1\bd{D}\wedge\mathbb{I}_{n-1}\big)\Big).
\end{equation}
At $(s,t)=(0,1)$, the generating function (\ref{gen fun A}) and the relations (\ref{ext A infinity rel})
reduce to $\bd{A}(0,1)=\bd{A}$ and the $A_\infty$ relation (\ref{A inf bare}), respectively.
Then, we can show that if $\bd{A}(s,t)$ satisfies the differential equations
\begin{subequations} \label{diff open}
\begin{align}
\partial_t\bd{A}(s,t)\ =&\ [\bd{Q},\bd{\mu}(s,t)]+[\bd{A}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_1(s)}\\
\partial_s\bd{A}(s,t)\ =&\ [\bd{\eta},\bd{\mu}(s,t)] -[\bd{A}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_2(t)},
\end{align}
\end{subequations}
with introducing the degree even coderivation
\begin{equation}
\bd{\mu}(s,t)\ =\ \sum_{m,p,r=0}^\infty s^mt^p\bd{\mu}^{(p+1)}_{m+p+r+2}|^{2r},
\end{equation}
the $t$ derivative of the left hand sides of the relations (\ref{ext A infinity rel}) become
\begin{align}
\partial_t\bd{I}(s,t)\ =&\ [\bd{I}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_1},\\
\partial_t\bd{J}(s,t)\ =&\ [\bd{J}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_1}
-[\bd{I}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_2} - \partial_s\bd{I}(s,t).
\end{align}
Thus, if
\begin{subequations} \label{A inf boson}
\begin{align}
\bd{I}(s,0)\ =&\ [\bd{Q},\bd{A}(s,0)] + \frac{1}{2}[\bd{A}(s,0),\bd{A}(s,0)]_{\mathfrak{o}_1}\ =\ 0,\\
\bd{J}(s,0)\ =&\ [\bd{\eta},\bd{A}(s,0)]\ =\ 0,
\end{align}
\end{subequations}
then $\bd{I}(s,t)=\bd{J}(s,t)=0$.
However, the relations (\ref{A inf boson}) are nothing less than those satisfied by
the \textit{geometric} string products
constructed similarly to those for the bosonic $A_\infty$ algebra, $\bd{Q} + \bd{M}_B(s)$,
without any insertion\footnote{
The parameter $s$ distinguishes the number of R strings associated with the interaction.}:
\begin{equation}
\bd{M}_B(s)\ \equiv\ \sum_{m,r=0}^\infty s^m(\bd{M}_B)_{m+r+1}\mid^{2r}.
\label{bosonic A}
\end{equation}
We can obtain the cyclic $A_\infty$ algebra, $(\mathcal{H}_l^o,\omega_l^o,\bd{Q}-\bd{\eta}+\bd{A})$
by recursively solving the differential equations (\ref{diff open})
with the initial condition $\bd{A}(s,0)=\bd{M_B}(s)$ so that $\bd{A}_n$ is cyclic with respect to $\omega_l^o$.
Finally, the cohomomorphism (\ref{cohomo A})
gives the cyclic $A_\infty$ algebra $(\mathcal{H}^{res}_o,\Omega^o,\bd{M})$.
\sectiono{Closed superstring field theory with $L_\infty$-structure}\label{B}
Similarly to the open superstring field theory, closed (type II) superstring field theory
is constructed based on the $L_\infty$ algebra structure. We next summarize it
in this section.
\subsection{Closed superstring field}
The first-quantized Hilbert space, $\mathcal{H}_c$, of type II (closed)
superstring is composed of four sectors:
$ \mathcal{H}_c = \mathcal{H}_{NS\textrm{-}NS} + \mathcal{H}_{R\textrm{-}NS} + \mathcal{H}_{NS\textrm{-}R} + \mathcal{H}_{R\textrm{-}R}$.
Correspondingly, the type II superstring field $\Phi$ has four components,
$\Phi = \Phi_{NS\textrm{-}NS} + \Phi_{R\textrm{-}NS} + \Phi_{NS\textrm{-}R} + \Phi_{R\textrm{-}R}$,
all of which are Grassmann even and have ghost number $2$.
The components $\Phi_{NS\textrm{-}NS}$ and $\Phi_{R\textrm{-}R}$ have picture numbers $(-1,-1)$ and $(-1/2,-1/2)$,
respectively and represent space-time bosons. The components $\Phi_{R\textrm{-}NS}$ and $\Phi_{NS\textrm{-}R}$
have picture numbers $(-1/2,-1)$ and $(-1,-1/2)$ and represent space-time fermions.
We impose it closed string constraints
\begin{equation}
b_0^-\Phi\ =\ L_0^-\Phi\ =\ 0,
\end{equation}
and also an extra constraint
\begin{equation}
\mathcal{P}_{XY}^c\Phi\ =\ \Phi,\qquad \mathcal{P}_{XY}^c\ =\ \mathcal{G}^c(\mathcal{G}^c)^{-1},
\label{closed constraint}
\end{equation}
with
\begin{align}
\mathcal{G}^c\ =&\
\pi^{(0,0)} + X^c \pi^{(1,0)} + \bar{X}^c \pi^{(0,1)} + X^c\bar{X}^c\pi^{(1,1)},\\
(\mathcal{G}^c)^{-1}\ =&\
\pi^{(0,0)} + Y^c \pi^{(1,0)} + \bar{Y}^c \pi^{(0,1)} + Y^c\bar{Y}^c\pi^{(1,1)},
\end{align}
where $\pi^{(0,0)}$, $\pi^{(1,0)}$, $\pi^{(0,1)}$, and $\pi^{(1,1)}$, are the projection
operators onto the NS-NS, R-NS, NS-R, and R-R components respectively:
$\pi^{(0,0)}\Phi=\Phi_{NS\textrm{-}NS}$, $\pi^{(1,0)}\Phi=\Phi_{R\textrm{-}NS}$,
$\pi^{(0,1)}\Phi=\Phi_{NS\textrm{-}R}$, and $\pi^{(1,1)}\Phi=\Phi_{R\textrm{-}R}$.
The PCO $X^c$ $(\bar{X}^c)$ and and its inverse $Y^c$ $(\bar{Y}^c)$ are defined by
\begin{align}
X^c\ =&\ -\delta(\beta_0)G + \frac{1}{2}(\gamma_0\delta(\beta_0)+\delta(\beta_0)\gamma_0)b^+_0,\qquad
Y^c\ =\ - 2\frac{G}{L^+_0}\delta(\gamma_0),\\
\bar{X}^c\ =&\ -\delta(\bar{\beta}_0)\bar{G} + \frac{1}{2}
(\bar{\gamma}_0\delta(\bar{\beta}_0)+\delta(\bar{\beta}_0)\bar{\gamma}_0)b^+_0,\qquad
\bar{Y}^c\ =\ - 2\frac{\bar{G}}{L^+_0}\delta(\bar{\gamma}_0).
\end{align}
The PCOs $X^c$ and $\bar{X}^c$ are BRST exact in the large Hilbert space:
\begin{alignat}{2}
X^c\ =&\ [Q,\Xi^c],\qquad&
\Xi^c\ =&\ \xi_0 + (\Theta(\beta_0)\eta\xi_0-\xi_0)P_{-3/2}+(\xi_0\eta\Theta(\beta_0)-\xi_0)P_{-1/2},\\
\bar{X}^c\ =&\ [Q,\bar{\Xi}^c],\qquad&
\bar{\Xi}^c\ =&\ \bar{\xi}_0 + (\Theta(\bar{\beta}_0)\bar{\eta}\bar{\xi}_0-\bar{\xi}_0)\bar{P}_{-3/2}
+(\bar{\xi}_0\bar{\eta}\Theta(\bar{\beta}_0)-\bar{\xi}_0)\bar{P}_{-1/2},
\end{alignat}
where $P_{-3/2}$ and $P_{-1/2}$ ($\bar{P}_{-3/2}$ and $\bar{P}_{-1/2}$)
are the projectors onto the states with the left-moving (right-moving)
picture numbers $-3/2$ and $-1/2$, respectively.
We denote the restricted Hilbert space of type II superstring as $\mathcal{H}_c^{res}$.
Similarly to the relations (\ref{G open}) for the open superstring,
$\mathcal{G}^c$ and $(\mathcal{G}^c)^{-1}$ satisfy the relations
\begin{align}
\mathcal{G}^c (\mathcal{G}^c)^{-1} \mathcal{G}^c \ =\ \mathcal{G}^c,\qquad
(\mathcal{G}^c)^{-1}\mathcal{G}^c (\mathcal{G}^c)^{-1}\ =\ (\mathcal{G}^c)^{-1},\qquad
[Q,\mathcal{G}^c]\ =\ 0,
\end{align}
and thus, $\mathcal{G}^c(\mathcal{G}^c)^{-1}$ is a projector that is compatible with
the BRST cohomology: $Q\mathcal{P}_{XY}^c=\mathcal{P}_{XY}^cQ\mathcal{P}_{XY}^c$.
The type II superstring field satisfying the constraint (\ref{closed constraint})
is expanded in the ghost zero-modes as
\begin{align}
\Phi\ =&\ (\phi_{NS\textrm{-}NS}-c_0^+\psi_{NS\textrm{-}NS})
+ \left(\phi_{R\textrm{-}R}-\frac{1}{2}(\gamma_0\bar{G}-\bar{\gamma}_0G
+2c_0^+G\bar{G})\psi_{R\textrm{-}R}\right)
\nonumber\\
&\
+ \left(\phi_{R\textrm{-}NS}-\frac{1}{2}(\gamma_0+2c_0^+G)\psi_{R\textrm{-}NS}\right)
+ \left(\phi_{NS\textrm{-}R}-\frac{1}{2}(\bar{\gamma}_0+2c_0^+\bar{G})\psi_{NS\textrm{-}R}\right)
\in\mathcal{H}_c^{res}.
\end{align}
Natural symplectic forms $\omega_s^c$ and $\Omega^c$ in $\mathcal{H}_c$ and $\mathcal{H}_c^{res}$,
respectively, are defined by using the BPZ inner product as
\begin{align}
\omega_s^c(\Phi_1,\Phi_2)\ =&\ (-1)^{\Phi_1}\langle\Phi_1|c_0^-|\Phi_2\rangle,\\
\Omega^c(\Phi_1,\Phi_2)\ =&\ (-1)^{\Phi_1}\langle\Phi_1|c_0^-(\mathcal{G}^c)^{-1}|\Phi_2\rangle.
\end{align}
Natural symplectic form $\omega_l^c$ in the large Hilbert space $\mathcal{H}_l^c$ is similarly
defined by using the BPZ inner product in $\mathcal{H}_l^c$, and related to $\omega_s^c$ as
$\omega_l^c(\xi_0\bar{\xi}_0\Phi_1,\Phi_2)=\omega_s^c(\Phi_1,\Phi_2)$ if $\Phi_1,\
\Phi_2\in\mathcal{H}_s^c$.
\subsection{Interaction with $L_\infty$-structure}
Type II superstring interactions are descried by the string products $L_n$ that map
$n$ closed superstring fields to a closed superstring field as
\begin{alignat}{3}\label{L maps}
L_n\ :\qquad\qquad
&\ (\mathcal{H}_c^{res})^{\wedge n} &\qquad &\ \longrightarrow&\qquad &\ \mathcal{H}_c^{res},\qquad
(n\ge1),
\nonumber\\
&\ \ \ \ \text{\rotatebox[origin=c]{90}{$\in$}}&\qquad &\ &\qquad &\ \ \text{\rotatebox[origin=c]{90}{$\in$}}\\
\ \ \ \Phi_1&\wedge \ \cdots\wedge\Phi_n &\qquad &\ \longmapsto&\qquad L_n(\Phi_1& ,\ \cdots, \Phi_n),
\nonumber
\end{alignat}
where $\Phi_1\wedge\cdots\wedge\Phi_n$ is the symmetrized tensor product defined by
\begin{equation}
\Phi_1\wedge\cdots\wedge\Phi_n\ =\ \sum_{\sigma}\Phi_{\sigma(1)}\otimes\cdots\otimes\Phi_{\sigma(n)},
\end{equation}
We identify the one-string product as the closed superstring BRST operator: $L_1=Q_c$.
By definition, these products must satisfy
\begin{align}
&\ b_0^+L_n(\Phi_1,\cdots,\Phi_n)\ =\ b_0^+L_n(\Phi_1,\cdots,\Phi_n)\ =\ 0,\\
&\ \mathcal{P}_{XY}^cL_n(\Phi_1,\cdots,\Phi_n)\ =\ L_n(\Phi_1,\cdots,\Phi_n).
\end{align}
We further impose the $L_\infty$ relations
\begin{equation}
\sum_{\sigma} \sum_{m=1}^n
(-1)^{\epsilon(\sigma)}
\frac{1}{m!(n-m)!}
L_{n-m+1}(L_m(\Phi_{\sigma(1)},\cdots,\Phi_{\sigma(m)}),\Phi_{\sigma(m+1)},\cdots,\Phi_{\sigma(n)})\ =\ 0,
\label{L infinity}
\end{equation}
and cyclicity
\begin{equation}
\Omega^c(\Phi_1,L_n(\Phi_2,\cdots,\Phi_{n+1}))\ =\
-(-1)^{|\Phi_1|}\Omega^c(L_n(\Phi_1,\cdots,\Phi_n),\Phi_{n+1}).
\label{cyclicity closed}
\end{equation}
The linear maps satisfying (\ref{L infinity}) and (\ref{cyclicity closed}) form
the cyclic $L_\infty$ algebra $(\mathcal{H}_c^{res},\Omega^c,\{L_m\})$.
The linear maps (\ref{L maps}) are also represented by a degree-odd coderivation
$\bd{L}=\sum_{n=1}^\infty\bd{L}_n$ acting on the symmetrized tensor algebra
$\mathcal{SH}_c=\sum_{n=0}^\infty(\mathcal{H}_c^{res})^{\wedge n}$ as
\begin{equation}
\bd{L}\ =\ \sum_{n=1}^\infty\ \bd{L}_n\
=\ \sum_{n=1}^\infty\sum_{m=0}^\infty
(L_n\wedge\mathbb{I}_m)\pi_{n+m}^c.
\end{equation}
with
$
\mathbb{I}_m\ =\ \frac{1}{m!}\, \mathbb{I}^{\wedge m
=\ \mathbb{I}^{\otimes m},
$
where $\pi_m^c$ is the projection operator onto
$(\mathcal{H}_c^{res})^{\wedge m}\subset\mathcal{SH}_c$.
Then, the $L_\infty$ relations in Eq.~(\ref{L infinity}) is written as
\begin{equation}
[\bd{L},\bd{L}]\ =\ 0.
\end{equation}
The string interaction $L_n$
of the type II superstring field theory must be defined
for each combination of NS-NS, R-NS, NS-R and R-R inputs so that the picture numbers
of left- and right-moving sectors must be conserved separately:
\begin{equation}
\bd{L}_{n+1}\ =\ \sum_{p+r=n}\sum_{\bar{p}+\bar{m}=n}
\bd{L}^{(p,\bar{p})}_{p+r+1,\bar{p}+\bar{m}+1}\mid_{(2r,2\bar{r})},
\end{equation}
where we used the diagonal matrix representation $\bd{L}_{n,m}=\delta_{n,m}\bd{L}_n$.
The superscript $p$ ($\bar{p}$) is the left-moving (right-moving)
picture number that the map itself has,
and the subscript $2r$ ($2\bar{r}$) is the left-moving (right-moving) Ramond number.
Introducing a real parameter $t\in[0,1]$,
the action with $L_\infty$ structure is given by
\begin{equation}
I_c\ =\ \int^1_0dt\,\Omega^c(\Phi,\pi_1^c\bd{L}(e^{\wedge t\Phi})),
\label{L action}
\end{equation}
with the group-like element
\begin{equation}
e^{\wedge\Phi}\ =\ \mathbb{I}_{\mathcal{SH}_c} + \sum_{n=1}^\infty\frac{1}{n!}\Phi^{\wedge n},
\end{equation}
where $\mathbb{I}_{\mathcal{SH}_c}$ is the identity in $\mathcal{SH}_c$ that satisfies
$\mathbb{I}_{\mathcal{SH}_c}\wedge V=V$ for ${}^{\forall}V\in\mathcal{SH}_c$.
The arbitrary variation of $I_c$ is given by
\begin{equation}
\delta I_c\ =\ \Omega^c(\delta\Phi,\pi_1\bd{L}(e^{\wedge\Phi})).
\end{equation}
We can show the action (\ref{L action}) is invariant under the gauge transformation
\begin{equation}
\delta_{\Lambda}\Phi\ =\ \pi_1^c\bd{L}(e^{\wedge\Phi}\wedge\Lambda),
\end{equation}
using the $L_\infty$ relation (\ref{L infinity}) and cyclicity (\ref{cyclicity closed}):
\begin{align}
\delta_{\Lambda} I_c\ =&\ \Omega^c\left(\pi_1^c\bd{L}(e^{\wedge\Phi}\wedge\Lambda),
\pi_1^c\bd{L}(e^{\wedge\Phi})\right)
\nonumber\\
=&\ \Omega^c\left(\Lambda,\pi_1^c\bd{L}\left(e^{\wedge\Phi}\wedge\pi_1^c\bd{L}(e^{\wedge\Phi})\right)\right)
\nonumber\\
=&\ \Omega^c\left(\Lambda,\pi_1^c\bd{L}^2(e^{\wedge\Phi})\right)\ =\ 0.
\end{align}
\subsection{Explicit construction of interactions}
The cyclic $L_\infty$ algebra $(\mathcal{H}_c^{res},\Omega^c,\bd{L})$ is constructed
in two steps. We consider first an $L_\infty$ algebra
$(\mathcal{H}_l^c,\bd{\mathcal{O}})$ with
\begin{equation}
\pi_1\bd{\mathcal{O}}\ =\ \pi_1(\bd{Q} - \bd{\eta} - \bar{\bd{\eta}}+\bd{B})
-\left(1-\frac{1}{2}(X+\bar{X})\right)\pi_1^{(1,1)}\bd{B}.
\end{equation}
introducing a degree odd coderivation
\begin{equation}
\bd{B}=\sum_{p,r=0}^\infty\sum_{\bar{p},\bar{r}=0}^\infty
\bd{B}_{p+r+1,\bar{p}+\bar{r}+1}^{(p,\bar{p})}\mid^{(2r,2\bar{r})}.
\label{B for L-inf}
\end{equation}
This $L_\infty$ algebra is equivalent to three mutually commutative $L_\infty$
algebras $(\mathcal{H}_l,\bd{\mathcal{D}})$,
$(\mathcal{H}_l,\bd{\mathcal{C}})$, and $(\mathcal{H}_l,\bar{\bd{\mathcal{C}}})$
with
\begin{align}
\pi_1\bd{\mathcal{D}}\ =&\ \pi_1\bd{Q}+\pi_1^{(0,0)}\bd{B},\\
\pi_1\bd{\mathcal{C}}\ =&\ \pi_1\bd{\eta}-\left(\pi_1^{(1,0)}
+\frac{1}{2}\bar{X}\pi_1^{(1,1)}\right)\bd{B},\qquad
\pi_1\bar{\bd{\mathcal{C}}}\ =\ \pi_1\bar{\bd{\eta}}-\left(\pi_1^{(0,1)}
+\frac{1}{2}X\pi_1^{(1,1)}\right)\bd{B}.
\end{align}
decomposed according to the picture number deficit.
Then, the $L_\infty$ relations are written as
\begin{subequations} \label{L inf bare}
\begin{align}
[\bd{Q},\bd{B}] + \frac{1}{2}[\bd{B},\bd{B}]^{11}\ =&\ 0,\\
[\bd{\eta},\bd{B}] -\frac{1}{2}[\bd{B},\bd{B}]^{21} -\frac{1}{4}[\bd{B},\bd{B}]_{\bar{X}}^{22}\ =&\ 0,\\
[\bd{\eta},\bd{B}] -\frac{1}{2}[\bd{B},\bd{B}]^{12} -\frac{1}{4}[\bd{B},\bd{B}]_{X}^{22}\ =&\ 0.
\end{align}
\end{subequations}
Here, the bracket $[\cdot,\cdot]^{11,21,12\,\textrm{or}\,22}$ is defined by projecting the intermediate
state of the (graded) commutator to the NS-NS, R-NS, NS-R, or R-R state. We also define the bracket
$[\cdot,\cdot]^{22}_{X\,\textrm{or}\,\bar{X}}$ by further inserting $X$ or $\bar{X}$ at
the intermediate R-R state.
If such $L_\infty$ algebras are found, we transform them by
cohomomorphism
\begin{equation}
\pi_1\hat{\bd{F}}^{-1}\ =\ \pi_1\mathbb{I}
-\left(\Xi\pi_1^{(1,0)}+\bar{\Xi}\pi^{(0,1)}_1
+\frac{1}{2}(\Xi\bar{X}+X\bar{\Xi})\pi_1^{(1,1)}\right)\bd{B}
\label{cohomo L}
\end{equation}
to the cyclic $L_\infty$ algebra $(\mathcal{H}_c^{res},\Omega^c,\bd{L})$
and two (trivial) $L_\infty$ algebras $(\mathcal{H}_l^c,\bd{\eta})$
and $(\mathcal{H}_l^c,\bar{\bd{\eta}})$ as
\begin{align}
\pi_1\hat{\bd{F}}^{-1}\bd{\mathcal{D}}\hat{\bd{F}}\ =&\
\pi_1\bd{Q}+\mathcal{G}^c\pi_1\bd{B}\hat{\bd{F}}\ \equiv \pi_1\bd{L},
\label{final L}\\
\pi_1\hat{\bd{F}}^{-1}\bd{\mathcal{C}}\hat{\bd{F}}\ =&\ \pi_1\bd{\eta},\qquad
\pi_1\hat{\bd{F}}^{-1}\bar{\bd{\mathcal{C}}}\hat{\bd{F}}\ =\ \pi_1\bar{\bd{\eta}}.
\end{align}
Note that the $L_\infty$ algebra $(\mathcal{H}_l^c,\bd{\mathcal{D}})$ is not cyclic
unlike the open superstring case. However, we can show,
in a similar way given in the Appendix C of Ref.~\cite{Kunitomo:2019glq},
that the $\bd{L}$ in (\ref{final L}) is cyclic with respect to $\Omega^c$
if $\bd{B}$ is cyclic with respect to $\omega_l^c$.
In the next step, we consider a generating function
\begin{equation}
\bd{B}(s,\bar{s},t)\ =\ \sum_{m,p,r=0}^\infty\sum_{\bar{m},\bar{p},\bar{r}=0}^\infty
s^m\bar{s}^{\bar{m}}t^{p+\bar{p}}\bd{B}^{(p,\bar{p})}_{m+p+r+1,\bar{m}+\bar{p}+\bar{r}+1}\mid^{(2r,2\bar{r})}
\label{gen fun L}
\end{equation}
and extend the $L_\infty$ relations (\ref{L inf bare}) to
\begin{subequations} \label{ext L infinity rel}
\begin{align}
\bd{I}(s,\bar{s},t)\ \equiv&\ [\bd{Q},\bd{B}(s,\bar{s},t)] + \frac{1}{2}[\bd{B}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\mathfrak{c}_1(s,\bar{s},t)}\ =\ 0,\\
\bd{J}(s,\bar{s},t)\ \equiv&\ [\bd{\eta},\bd{B}(s,\bar{s},t)] -\frac{1}{2}[\bd{B}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\mathfrak{c}_2(t)}\ =\ 0,\\
\bar{\bd{J}}(s,\bar{s},t)\ \equiv&\ [\bar{\bd{\eta}},\bd{B}(s,\bar{s},t)] -\frac{1}{2}[\bd{B}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\bar{\mathfrak{c}}_2(t)}\ =\ 0,
\end{align}
\end{subequations}
for constructing the $L_\infty$ algebra $(\mathcal{H}_l^c,\omega_l^c,\bd{\mathcal{O}})$.
The parameters $s$, $\bar{s}$, and $t$ counting the left-moving picture number deficit,
right-moving picture number deficit, and the total picture number, respectively.
The bracket with subscript is defined by inserting
\begin{align}
\mathfrak{c}_1(s,\bar{s},t)\ =&\ \pi^{(0,0)}
+s\pi^{(1,0)}+\bar{s}\pi^{(0,1)}
+\left(s\bar{s}+t(s\bar{X}+\bar{s}X)\right)\pi^{(1,1)},\\
\mathfrak{c}_2(t)\ =&\ t\pi^{(1,0)}+\frac{t^2}{2}\bar{X}\pi^{(1,1)},\qquad
\bar{\mathfrak{c}}_2(t)\ =\ t\pi^{(0,1)}+\frac{t^2}{2}X\pi^{(1,1)},
\end{align}
at the intermediate state.
At $(s,\bar{s},t)=(0,0,1)$, the generating function (\ref{gen fun L}) and the relations (\ref{ext L infinity rel})
reduce to $\bd{B}(0,0,1)=\bd{B}$ and the $L_\infty$ relation (\ref{L inf bare}), respectively.
We can show that if $\bd{B}(s,\bar{s},t)$ satisfies
the differential equations
\begin{subequations}\label{diff closed}
\begin{align}
\partial_t\bd{B}(s,\bar{s},t)\ =&\ [\bd{Q},(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]
\nonumber\\
&\
+ [\bd{B}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\mathfrak{c}_1(s,\bar{s},t)}
+\frac{1}{2}[\bd{B}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\mathfrak{d}(s,\bar{s})}\,\\
\partial_s\bd{B}(s,\bar{s},t)\ =&\ [\bd{\eta},\bd{\lambda}(s,\bar{s},t)]
- [\bd{B}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\mathfrak{c}_2(t)},\\
\partial_{\bar{s}}\bd{B}(s,\bar{s},t)\ =&\ [\bar{\bd{\eta}},\bd{\lambda}(s,\bar{s},t)]
- [\bd{B}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\bar{\mathfrak{c}}_2(t)},
\end{align}
\end{subequations}
with $\mathfrak{d}(s,\bar{s})=(s\bar{\Xi}+\bar{s}\Xi)\pi^{(1,1)}$ and
the degree even coderivations
\begin{align}
\bd{\lambda}(s,\bar{s},t)\ =&\
\sum_{m,p,r=0}^\infty\sum_{\bar{m},\bar{p},\bar{r}=0}^\infty s^m
\bar{s}^{\bar{m}}t^{p+\bar{p}}
\bd{\lambda}^{(p+1,\bar{p})}_{m+p+r+2,\bar{m}+\bar{p}+\bar{r}+1}\mid^{(2r,2\bar{r})},\\
\bar{\bd{\lambda}}(s,\bar{s},t)\ =&\ \sum_{m,p,r=0}^\infty\sum_{\bar{m},\bar{p},\bar{r}=0}^\infty
s^m\bar{s}^{\bar{m}}t^{p+\bar{p}}
\bar{\bd{\lambda}}^{(p,\bar{p}+1)}_{m+p+r+1,\bar{m}+\bar{p}+\bar{r}+2}\mid^{(2r,2\bar{r})},
\end{align}
the $t$ derivative of the left hand sides of the relations (\ref{ext L infinity rel}) become
\begin{subequations}
\begin{align}
\partial_t\bd{I}(s,\bar{s},t)\ =&\ [\bd{I}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\mathfrak{c}_1(s,\bar{s},t)}
+ [\bd{I}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\mathfrak{d}(s,\bar{s})},\\
\partial_t\bd{J}(s,\bar{s},t)\ =&\ [\bd{J}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\mathfrak{c_1}(s,\bar{s},t)}
+ [\bd{J}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\mathfrak{d}(s,\bar{s})}
\nonumber\\
&\
-\partial_s\bd{I}(s,\bar{s},t)
- [\bd{I}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\mathfrak{c}_2(t)},\\
\partial_t\bar{\bd{J}}(s,\bar{s},t)\ =&\ [\bar{\bd{J}}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\mathfrak{c_1}(s,\bar{s},t)}
+ [\bar{\bd{J}}(s,\bar{s},t),\bd{B}(s,\bar{s},t)]_{\mathfrak{d}(s,\bar{s})}
\nonumber\\
&\
-\partial_{\bar{s}}\bd{I}(s,\bar{s},t) - [\bd{I}(s,\bar{s},t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,\bar{s},t)]_{\bar{\mathfrak{c}}_2(t)}.
\end{align}
\end{subequations}
They imply if
\begin{subequations} \label{L inf boson}
\begin{align}
\bd{I}(s,\bar{s},0)\ =&\ [\bd{Q},\bd{B}(s,\bar{s},0)]
+ \frac{1}{2}[\bd{B}(s,\bar{s},0),\bd{B}(s,\bar{s},0)]_{\mathfrak{c}_1(s,\bar{s},0)}\ =\ 0,\\
\bd{J}(s,\bar{s},0)\ =&\ [\bd{\eta},\bd{B}(s,\bar{s},0)]\ =\ 0,\\
\bar{\bd{J}}(s,\bar{s},0)\ =&\ [\bar{\bd{\eta}},\bd{B}(s,\bar{s},0)]\ =\ 0,
\end{align}
\end{subequations}
then $\bd{I}(s,\bar{s},t)=\bd{J}(s,\bar{s},t)=\bar{\bd{J}}(s,\bar{s},t)=0$.
The relations (\ref{L inf boson}) are those satisfied by the geometric string product
without any insertion, which can be constructed similarly to that for the bosonic $L_\infty$ algebra,
$\bd{Q}+\bd{L}_B(s,\bar{s})$:
\begin{equation}
\bd{L}_B(s,\bar{s})\ \equiv\
\sum_{m,r=0}^\infty\sum_{\bar{m},\bar{r}=0}^\infty
s^m\bar{s}^{\bar{m}}(\bd{L}_B)_{m+r+1,\bar{m}+\bar{r}+1}\mid^{(2r,2\bar{r})}.
\label{bosonic L}
\end{equation}
We can obtain the $L_\infty$ algebra $(\mathcal{H}_l^c,\bd{\mathcal{O}})$ by solving
the differential equations (\ref{diff closed}) with the initial condition
$\bd{B}(s,\bar{s},0)=\bd{L}_B(s,\bar{s})$.
The cohomomorphism (\ref{cohomo L}) gives the cyclic $L_\infty$ algebra $(\mathcal{H}_c^{res},\Omega^c,\bd{L})$
if we choose the solution to be cyclic with respect to $\omega_l^c$.
\sectiono{Open-closed superstring field theory with OCHA-structure}\label{C}
Now, we are ready to discuss OCHA, the main subject of this paper.
In this section, we first see what OCHA is and how it is realized in the superstring
field theory and then give a prescription to construct them explicitly.
\subsection{Interaction with OCHA structure}
We define classical interactions among the open and closed (type II) superstrings mixed
system by the vertices described by the following two kinds of surfaces:
\begin{itemize}
\item A sphere with $n\, (\ge3)$ closed-superstring punctures.
\item A disk with $n\,(\ge0)$ closed-superstring punctures on the bulk
and $l+1\,(\ge1)$ open-superstring punctures on the boundary with $n+l\ge1$.
\end{itemize}
We can identify the former as the linear maps $\{L_n\}$ given in the previous section,
which form the cyclic $L_\infty$ algebra $(\mathcal{H}_c^{res},\Omega^c,\{L_n\})$.
The latter includes both the open-superstring interactions ($n=0$)
and interactions between open and closed superstrings ($n>0$) and is described by
the string products $N_{n,l}$ that maps
$n$ closed-superstring fields and $l$ open-superstring fields to an open-superstring field:
\begin{alignat}{3}\label{linear map n}
N_{n,l}\ :\qquad
&\ (\mathcal{H}_c^{res})^{\wedge n}\otimes(\mathcal{H}_o^{res})^{\oplus l}
&\ &\longrightarrow& &\ \mathcal{H}_o^{res},\quad
(n,l\ge0,\,n+l>0),
\nonumber\\
&\ \ \ \ \ \ \ \ \ \ \ \ \
\text{\rotatebox[origin=c]{90}{$\in$}} &\ & &\ &\ \text{\rotatebox[origin=c]{90}{$\in$}}\\
\ \ \ (\Phi_1&\wedge \ \cdots\wedge\Phi_n)\otimes(\Psi_1\otimes\cdots\otimes\Psi_l)
&\ &\longmapsto &\ N_{n,l}(\Phi_1&,\ \cdots, \Phi_n ; \Psi_1,\cdots,\Psi_l),
\nonumber
\end{alignat}
with the identification $N_{0,l}=M_l$.
By definition, the condition
\begin{equation}
\mathcal{P}_{XY}^oN_{n,l}(\Phi_1,\cdots,\Phi_n;\Psi_1,\cdots,\Psi_l)\ =\
N_{n,l}(\Phi_1,\cdots,\Phi_n;\Psi_1,\cdots,\Psi_l)
\end{equation}
holds. The linear maps $\{L_n,N_{n,l}\}$ satisfying the OCHA relation
\begin{align}\label{OCHA rel}
0\ =&\ \sum_{\sigma}\sum_{m=1}^n(-1)^{\epsilon(\sigma)}\frac{1}{m!(n-m)!}N_{n-m+1,l}(L_{m}(\Phi_{\sigma(1)},\cdots,
\Phi_{\sigma(m)}),\Phi_{\sigma(m+1)},\cdots,\Phi_{\sigma(n)};\Psi_1,\cdots,\Psi_l)
\nonumber
\\
&\
+ \sum_{\sigma}\sum_{m=0}^n\sum_{j=0}^{l}\sum_{i=0}^{l-j}(-1)^{\mu_{m,i}(\sigma)}
\frac{1}{m!(n-m)!}
N_{m,l-j+1}(\Phi_{\sigma(1)},\cdots,\Phi_{\sigma(m)};
\nonumber\\
&\hspace{20mm}
\Psi_1,\cdots,\Psi_i,
N_{n-m,j}(\Phi_{\sigma(m+1)},\cdots,\Phi_{\sigma(n)};\Psi_{i+1},\cdots,\Psi_{i+j}),\Psi_{i+j+1},\cdots,\Psi_l),
\end{align}
and the cyclicity condition
\begin{align}
&\ \Omega^o(\Psi_1,N_{n,l}(\Phi_1,\cdots,\Phi_n;\Psi_2,\cdots,\Psi_{l+1}))
\nonumber\\
&\hspace{2cm}
=\
-(-1)^{\textrm{deg}(\Psi_1)(|\Phi_1|+\cdots+|\Phi_l|+1)}
\Omega^o(N_{n,l}(\Phi_1,\cdots,\Phi_n;\Psi_1,\cdots,\Psi_l),\Psi_{l+1})
\label{cyclicity N}
\end{align}
form the cyclic OCHA $(\mathcal{H}_c\oplus\mathcal{H}_o,\Omega^o,\{L_n,N_{n,l}\})$.
Here, the sign factor $\mu_{m,i}(\sigma)$ in (\ref{OCHA rel}) is given by
\begin{equation}
\mu_{m,i}(\sigma)\ =\ \epsilon(\sigma) +
\sum_{j=1}^{m}|\Phi_{\sigma(j)}| +
\sum_{j=1}^{i}|\Psi_{j}|(1 +
\sum_{k=m+1}^{n}|\Phi_{\sigma(k)}|).
\end{equation}
Note that the OCHA relation (\ref{OCHA rel}) includes the $A_\infty$ relation
(\ref{A infinity}) as $n=0$, at which the cyclicity condition becomes the one
for the open superstring (\ref{cyclicity open}).
In other words, the purely open-superstring interactions $N_{0,l}$
form the cyclic $A_\infty$ algebra $(\mathcal{H}_o,\Omega^o,\{N_{0,l}\})$.
The linear maps (\ref{linear map n}) are also represented by a degree odd coderivation
\begin{equation}
\bd{N}=\sum_{\substack{n,l=0\\ n+l\ge1}}^\infty \bd{N}_{n,l}
\end{equation}
acting on $\mathcal{SH}_c^{res}\otimes\mathcal{TH}_o^{res}$ as
\begin{equation}
\bd{N}\ =\ \sum_{\substack{n,l=0\\n+l>0}}\bd{N}_{n,l}\
=\ \sum_{\substack{n,l=0\\n+l>0}}\sum_{m,j,k=0}^\infty
\bigg(\mathbb{I}_m \otimes \Big(\mathbb{I}^{\otimes j}\otimes N_{n,l}\otimes \mathbb{I}^{\otimes k}\Big)\bigg)
\pi_{m+n,j+k+l},
\label{coderivation N}
\end{equation}
where $\pi_{n,l}$ is the projector onto the subspace
$(\mathcal{H}_c^{res})^{\wedge n}\otimes (\mathcal{H}_o^{res})^{\otimes l}$.
By extending $\bd{L}$ to the coderivation acting on $\mathcal{SH}_c^{res}\otimes\mathcal{TH}_o^{res}$ as
\begin{equation}
\bd{L}\ =\ \sum_{n=1}^\infty\ \bd{L}_n\
=\ \sum_{n=1}^\infty\sum_{m,l=0}^\infty
\bigg(\Big(L_n\wedge\mathbb{I}_m\Big)\otimes \mathbb{I}^{\otimes l}\bigg)\pi_{n+m,l},
\end{equation}
we can consider the coderivation $\bd{L}+\bd{N}$.
The OCHA relation (\ref{OCHA rel}) can then be written as
\begin{equation}
[\bd{L},\bd{N}]+\frac{1}{2}[\bd{N},\bd{N}]\ =\ 0,
\label{sOCHA rel}
\end{equation}
which we can rewrite as
\begin{equation}
[\bd{L}+\bd{N},\bd{L}+\bd{N}]\ =\ 0,
\end{equation}
by combining with the $L_\infty$ relation $[\bd{L},\bd{L}]=0$.
For open-closed superstring field theory, the string interaction $N_{n,l}$ must be defined
for any combination of four sectors of closed superstring and two sectors of open superstring
so that the sum of three kinds (open, left-moving, and right-moving) of picture numbers are conserved:
\begin{equation}
\bd{N}\ =\ \sum_{\substack{n,m,r=0\\ n+m\ge1\\ 2n+m\ge r+1}}^\infty \bd{N}_{n;m}^{(2n+m-r-1)}\mid_{2r},
\end{equation}
where $2r$ is the \textit{total} Ramond number defined by
\begin{align}
&\ \textrm{total Ramond number}\ =\
\nonumber\\
&\hspace{5mm}
\textrm{\# of $R\textrm{-}NS$ inputs} + \textrm{\# of $NS\textrm{-}R$ input}
+ 2 (\textrm{\# of $R\textrm{-}R$ inputs})
+ \textrm{\# of open $R$ inputs}
\nonumber\\
&\hspace{5mm}
- \textrm{\# of open $R$ output}.
\nonumber
\end{align}
The action with OCHA structure is given by
\begin{equation}
I_{oc}\ =\ \int^1_0 dt\, \Omega^o\left(\Psi,\pi_1
\bd{N}\left(e^{\wedge{\bf\Phi}}\otimes\frac{1}{1-t\Psi}\right)\right),
\label{action oc}
\end{equation}
which describes the open superstring field theory on the closed-superstring background
\footnote{
We omitted here the terms corresponding to a disk with closed strings in the bulk and no open strings
on the boundary, which are included in the action proposed in Ref.~\cite{Zwiebach:1997fe}. These terms
give a constant determined by a closed string background but do not relevant to symmetry structure
of the theory \cite{Kajiura:2005sn}.}.
The open-superstring field $\Psi$ is dynamical and the closed-superstring field ${\bf\Phi}$ is
the background field satisfying the equation of motion
\begin{equation}
\pi_1\bd{L}(e^{\wedge{\bf\Phi}})\ =\ 0.
\label{closed eom}
\end{equation}
The arbitrary variation of $I_{oc}$ is given by
\begin{equation}
\delta I_{oc}\ =\ \Omega^o\left(\delta\Psi, \pi_1\bd{N}
\left(e^{\wedge{\bf\Phi}}\otimes
\frac{1}{1-\Psi}\right)\right).
\end{equation}
We can show that the action (\ref{action oc}) is invariant
under the gauge transformation
\begin{equation}
\delta_{\Lambda}\Psi\ =\ \pi_1\bd{N}\left(e^{\wedge{\bf \Phi}}\otimes
\left(\frac{1}{1-\Psi}\otimes
\Lambda\otimes\frac{1}{1-\Psi}\right)\right),
\label{gauge tf oc}
\end{equation}
using the relation (\ref{sOCHA rel}):
\begin{align}
\delta_{\Lambda}I_{oc}\ =&\
\Omega^o\left(\Lambda,\pi_1\bd{N}\left(e^{\wedge{\bf\Phi}}\otimes\left(
\frac{1}{1-\Psi}\otimes\pi_1\bd{N}\left(e^{\wedge{\bf\Phi}}\otimes\frac{1}{1-\Psi}\right)
\otimes\frac{1}{1-\Psi}\right)\right)\right)
\nonumber\\
=&\ \Omega^o\left(\Lambda, \pi_1\bd{NN}\left(e^{\wedge{\bf\Phi}}\otimes\frac{1}{1-\Psi}\right)
\right)
\nonumber\\
=&\ -\Omega^o\left(\Lambda,\pi_1\bd{NL}\left(e^{\wedge{\bf\Phi}}\otimes\frac{1}{1-\Psi}\right)\right)
\nonumber\\
=&\ -\Omega^o\left(\Lambda, \pi_1\bd{N}\left(\left(e^{\wedge{\bf\Phi}}\wedge\pi_1\bd{L}\left(
e^{\wedge{\bf\Phi}}\right)\right)\otimes\frac{1}{1-\Psi}
\right)\right)\ =\ 0.
\end{align}
The open-closed superstring interaction $\bd{N}$ deforms
by the background closed superstring field ${\bf\Phi}$ gives a weak $A_\infty$ algebra
$(\mathcal{H}_o^{res},\bd{M}({\bf\Phi}))$ with
\begin{equation}
\bd{M}({\bf\Phi})\ =\ \left(\bd{N}(e^{\bf\Phi}\otimes\mathbb{I})\right)\pi^o.
\end{equation}
Here, $\mathbb{I}$ is the identity map in $\mathcal{TH}_o^{res}$
and $\pi^o$ is the projector onto $\mathcal{TH}_o^{res}$
\subsection{Explicit construction of interactions}
Let us construct a cyclic OCHA
$(\mathcal{H}_c^{res}\oplus\mathcal{H}_o^{res},\Omega^c\oplus\Omega^o,\bd{L}+\bd{N})$.
We assume that the cyclic sub-$L_\infty$-algebra
$\bd{L}$ is already constructed in the way given in the previous subsection.
Similarly to the previous cases,
we can construct $\bd{N}$ satisfying the relation (\ref{sOCHA rel}) and the cyclicity
condition (\ref{cyclicity N}) in the following two steps. First consider a degree odd nilpotent
coderivation
\begin{equation}
\pi_1\bd{\mathcal{O}}\
=\ \pi_1\left(\bd{Q}-\bd{\eta}+\bd{A}+\bd{B}+\bd{C}\right)
-\left(1-\frac{1}{2}(X+\bar{X})\right)\pi_1^{(1,1)}\bd{B}
\label{OCHA O}
\end{equation}
satisfying $[\bd{\mathcal{O}},\bd{\mathcal{O}}]=0$,
or equivalently two mutually commutative coderivations
\begin{align}
\pi_1\bd{\mathcal{D}}\ =&\ \pi_1\bd{Q}+\pi_1^{(0,0)}\bd{B}+\pi_1^0(\bd{A}+\bd{C}),\\
\pi_1\bd{\mathcal{C}}\ =&\ \pi_1\bd{\eta}-\left(\pi_1^{(1,0)}+\pi_1^{(0,1)}+
\frac{1}{2}(X+\bar{X})\pi_1^{(1,1)}\right)\bd{B}-\pi_1^1(\bd{A}+\bd{C}),
\end{align}
satisfying $[\bd{\mathcal{D}},\bd{\mathcal{D}}]=[\bd{\mathcal{C}},\bd{\mathcal{C}}]
=[\bd{\mathcal{D}},\bd{\mathcal{C}}]=0$,
where $\bd{Q}$ acts as $\bd{Q}_c$ or $\bd{Q}_o$ on $\mathcal{H}_c$ or $\mathcal{H}_o$,
respectively, and similarly $\bd{\eta}$ acts as $\bd{\eta}+\bar{\bd{\eta}}$ or $\bd{\eta}$ on
on $\mathcal{H}_c$ or $\mathcal{H}_o$, respectively.
Degree odd coderivations $\bd{A}$ and $\bd{B}$ are those for constructing $A_\infty$
and $L_\infty$ algebras in (\ref{A for A-inf}) and (\ref{B for L-inf}), and $\bd{C}$ is the one
for constructing open-closed interaction
defined by respecting the cyclic Ramond number:
\begin{equation}
\bd{C}\ =\
\sum_{n,l,r=0}^\infty\bd{C}_{n+1,l}^{(2n+l+1-r)}\mid^{2r}.
\end{equation}
The OCHA relations can be written as
the $L_\infty$ relations (\ref{L inf bare}) and the relations
\begin{subequations} \label{OCHA bare}
\begin{align}
&\ [\bd{Q},\bd{C}] + [\bd{A},\bd{C}]^1
+ \frac{1}{2}[\bd{C},\bd{C}]^1 + [\bd{B},\bd{C}]^{11}\ =\ 0,\\
&\ [\bd{\eta},\bd{C}] - [\bd{A},\bd{C}]^2 - \frac{1}{2}[\bd{C},\bd{C}]^2
- [\bd{B},\bd{C}]^{21} - [\bd{B},\bd{C}]^{12}
- [\bd{B},\bd{C}]^{22}_{X+\bar{X}}\ =\ 0.
\end{align}
\end{subequations}
If we find such $\bd{A}$, $\bd{B}$, and $\bd{C}$, the cohomomorphism
\begin{align}
\pi_1\hat{\bd{F}}^{-1}\ =&\
\pi_1\mathbb{I} - \left(\Xi^c\pi_1^{(1,0)} + \bar{\Xi}^c\pi_1^{(0,1)}
+\frac{1}{2}(\bar{\Xi}^cX^c+\Xi^c\bar{X}^c)\pi_1^{(1,1)}
\right)\bd{B} - \Xi^o\pi_1^1(\bd{A}+\bd{C})
\label{cohomo OCHA}
\end{align}
transforms $\bd{\mathcal{D}}$ and $\bd{\mathcal{C}}$ to the ones we eventually construct as
\begin{align}
\pi_1\hat{\bd{F}}^{-1}\bd{\mathcal{D}}\hat{\bd{F}}\ =&\ \pi_1\bd{Q} + \mathcal{G}^c\pi_1\bd{B}\hat{\bd{F}}\
+ \mathcal{G}^0\pi_1(\bd{A}+\bd{C})\hat{\bd{F}}\
=\ \pi_1(\bd{L}+\bd{N}),\\
\pi_1\hat{\bd{F}}^{-1}\bd{\mathcal{C}}\hat{\bd{F}}\ =&\ \bd{\eta}.
\end{align}
We can construct $\bd{C}$ similarly to $\bd{A}$ and $\bd{B}$, which we already find.
Consider a generating functions\footnote{
We consider $\bd{C}_{n+1,l}^{(p)}\mid^{2r}\equiv0$ for $p<0$.
}
\begin{align}
\bd{A}(s,t)\ =&\ \sum_{m,p,r=0}^\infty
s^m t^p \bd{A}_{m+p+r+1}^{(p)}\mid^{2r},\\
\bd{B}(s,s,t)\ =&\ \sum_{m,p,r=0}^\infty\sum_{\bar{m},\bar{p},\bar{r}=0}^\infty
s^{m+\bar{m}}t^{p+\bar{p}}\bd{B}^{(p,\bar{p})}_{m+p+r+1,\bar{m}+\bar{p}+\bar{r}+1}\mid^{(2r,2\bar{r})},\\
\bd{C}(s,t)\ =&\
\sum_{n,l,r,m=0}^\infty
s^m t^{2n+l+1-r-m}\bd{C}_{n+1,l}^{(2n+l+1-r-m)}\mid^{2r},
\end{align}
and extend the relations (\ref{OCHA bare}) to
\begin{subequations} \label{ext OCHA rel}
\begin{align}
\bd{I}(s,t)\ \equiv&\ [\bd{Q},\bd{C}(s,t)]
+ [\bd{A}(s,t),\bd{C}(s,t)]_{\mathfrak{o}_1(s)}
+ \frac{1}{2}[\bd{C}(s,t),\bd{C}(s,t)]_{\mathfrak{o}_1(s)}
\nonumber\\
&\hspace{62mm}
+ [\bd{B}(s,s,t),\bd{C}(s,t)]_{\mathfrak{c}_1(s,s,t)}\ =\ 0,\\
\bd{J}(s,t)\ \equiv&\ [\bd{\eta},\bd{C}(s,t)]
- [\bd{A}(s,t),\bd{C}(s,t)]_{\mathfrak{o}_2(t)}
- \frac{1}{2}[\bd{C}(s,t),\bd{C}(s,t)]_{\mathfrak{o}_2(t)}
\nonumber\\
&\hspace{61mm}
- [\bd{B}(s,s,t),\bd{C}(s,t)]_{\mathfrak{c}_2(s)+\bar{\mathfrak{c}}_2(s)}\ =\ 0,
\end{align}
\end{subequations}
by introducing parameters $s$ and $t$.
Here, $\bd{A}(s,t)$ is the generating function of $A_\infty$ algebra in (\ref{gen fun A}),
$\bd{B}(s,s,t)$ is that of $L_\infty$ algebra
in (\ref{gen fun L}) with setting $s=\bar{s}$.
We can show that
if $\bd{A}(s,t)$ and $\bd{B}(s,\bar{s},t)$ satisfy
the differential equations (\ref{diff open}) and (\ref{diff closed}), respectively, and
$\bd{C}(s,t)$ satisfy
\begin{subequations}\label{diff OCHA}
\begin{align}
\partial_t\bd{C}(s,t)\ =&\ [\bd{Q},\bd{\nu}(s,t)]
\nonumber\\
&\
+ [\bd{A}(s,t),\bd{\nu}(s,t)]_{\mathfrak{o}_1(s)}
+ [\bd{C}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_1(s)}
+ [\bd{C}(s,t),\bd{\nu}(s,t)]_{\mathfrak{o}_1(s)}
\nonumber\\
&\
+ [\bd{B}(s,s,t),\bd{\nu}(s,t)]_{\mathfrak{c}_1(s,s,t)}
+ [\bd{C}(s,t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,s,t)]_{\mathfrak{c}_1(s,s,t)}
\nonumber\\
&\
+ [\bd{B}(s,s,t),\bd{C}(s,t)]_{\mathfrak{d}(s,s)},\\
\partial_s\bd{C}(s,t)\ =&\ [\bd{\eta},\bd{\nu}(s,t)]
\nonumber\\
&\
- [\bd{A}(s,t),\bd{\nu}(s,t)]_{\mathfrak{o}_2(t)}
- [\bd{C}(s,t),\bd{\mu}(s,t)]_{\mathfrak{o}_2(t)}
- [\bd{C}(s,t),\bd{\nu}(s,t)]_{\mathfrak{o}_2(t)}
\nonumber\\
&\
- [\bd{B}(s,s,t),\bd{\nu}(s,t)]_{\mathfrak{c}_2(t)+\bar{\mathfrak{c}}_2(t)}
- [\bd{C}(s,t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,s,t)]_{\mathfrak{c}_2(t)+\bar{\mathfrak{c}}_2(t)}
\end{align}
\end{subequations}
with degree even coderivation\footnote{
Here, $\bd{\nu}_{n+1,l}^{(p+1)}\mid^{2r}$ is considered to be 0 if $p<0$.},
\begin{equation}
\bd{\nu}(s,t)\ =\
\sum_{n,l,r,m=0}^\infty s^m t^{2n+l-r-m}\bd{\nu}_{n+1,l}^{(2n+l+1-r-m)}\mid^{2r},
\end{equation}
then, the $t$ derivative of the left hand sides of (\ref{ext OCHA rel}) become
\begin{subequations}
\begin{align}
\partial_t\bd{I}(s,t)\ =&\ [\bd{I}(s,t),(\bd{\mu}(s,t)+\bd{\nu}(s,t))]_{\mathfrak{o}_1(s)}
+ [\bd{I}(s,t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,s,t)]_{\mathfrak{c}_1(s,s,t)}
\nonumber\\
&\
+ [\bd{I}(s,t), \bd{B}(s,s,t)]_{\mathfrak{d}_1(s,s)},\\
\partial_t\bd{J}(s,t)\ =&\ [\bd{J}(s,t),(\bd{\mu}(s,t)+\bd{\nu}(s,t))]_{\mathfrak{o}_1(s)}
+ [\bd{J}(s,t),(\bd{\lambda}+\bar{\bd{\lambda}})(s,s,t)]_{\mathfrak{c}_1(s,s,t)}
\nonumber\\
&\
+ [\bd{J}(s,t),\bd{B}]_{\mathfrak{d}(s,s)}
- \partial_s\bd{I}(s,t)]
- [\bd{I}(s,t),(\bd{\mu}(s,t)+\bd{\nu}(s,t))]_{\mathfrak{o}_2(t)}.
\end{align}
\end{subequations}
Therefore, if the relations at $t=0$
\begin{subequations} \label{OCHA boson}
\begin{align}
\bd{I}(s,0)\ \equiv&\ [\bd{Q},\bd{C}(s,0)]
+ [\bd{A}(s,0),\bd{C}(s,0)]_{\mathfrak{o}_1(s)}
+ \frac{1}{2}[\bd{C}(s,0),\bd{C}(s,0)]_{\mathfrak{o}_1(s)}
\nonumber\\
&\
+ [\bd{B}(s,s,0),\bd{C}(s,0)]_{\mathfrak{c}_1(s,s,0)}\ =\ 0,\\
\bd{J}(s,0)\ \equiv&\ [\bd{\eta},\bd{C}(s,0)]\ =\ 0,
\end{align}
\end{subequations}
hold, then $\bd{I}(s,t)=\bd{J}(s,t)=0$ for any $t$.
We can easily find that the coderivation $\bd{C}(s,0)$ satisfying (\ref{OCHA boson})
has no picture number and is given by setting $\bd{C}(s,0)=\bd{N}_B(s)$ with
\begin{equation}
\bd{N}_B(s)\ =\ \sum_{n,l,r=0}^\infty
s^{2n+l+1-r}(\bd{N}_B)_{n+1,l}\mid^{2r},
\end{equation}
which can constructed similarly to those of the bosonic open-closed string field theory
\cite{Zwiebach:1997fe}. Therefore, we can obtain $\bd{C}(s,t)$ satisfying (\ref{ext OCHA rel})
by recursively solving the differential equations (\ref{diff OCHA}) to be cyclic with respect to $\omega_l^o$.
The cyclic OCHA $(\mathcal{H}^{res},\Omega,\bd{L}+\bd{N})$ is eventually constructed by
transforming using cohomomorphism (\ref{cohomo OCHA}).
\sectiono{Mapping to WZW-like action}\label{D}
The WZW-like formulation is the other complementary way to construct superstring field theories
using the large Hilbert space \cite{Berkovits:1995ab,Erler:2015rra,Kunitomo:2015usa,Berkovits:2004xh,Matsunaga:2016zsu,
Erler:2017onq,Kunitomo:2019glq,Kunitomo:2019kwk,Kunitomo:2021wiz}.
We can map
the action we constructed in the previous section to the WZW-like action as in the open, heterotic, and
type II superstring field theories
\cite{Erler:2015rra,Kunitomo:2019glq,Kunitomo:2019kwk,Kunitomo:2021wiz}.
Let us first focus on the NS $\oplus$ NS-NS sector, which we simply call the NS sector in this section.
The map between two formulations, the homotopy-based and WZW-like formulations, is given
by the cohomomorphism $\hat{\bd{g}}=\hat{\bd{g}}_c\otimes\hat{\bd{g}}_o$
\cite{Erler:2016ybs,Kunitomo:2019glq,Kunitomo:2021wiz} with
\begin{equation}
\hat{\bd{g}}_c\ =\ \vec{\mathcal{P}}\exp\left(
\int^1_0dt\,(\bd{\lambda}+\bar{\bd{\lambda}})^{NS}(0,t)\right),\qquad
\hat{\bd{g}}_o\ =\ \vec{\mathcal{P}}\exp\left(
\int^1_0dt\,\bd{\mu}^{NS}(0,t)\right),
\end{equation}
where
\begin{align}
(\bd{\lambda}+ \bar{\bd{\lambda}})^{NS}(s,t)\ =&\
\sum_{m,p=0}^\infty\sum_{\bar{m},\bar{p}=0}^\infty
s^{m+\bar{m}}t^{p+\bar{p}} \left(
\bd{\lambda}^{(p+1,\bar{p})}_{m+p+2,\bar{m}+\bar{p}+1}\mid^{(0,0)}
+ \bar{\bd{\lambda}}^{(p,\bar{p}+1)}_{m+p+1,\bar{m}+\bar{p}+2}\mid^{(0,0)}\right),\\
\bd{\mu}^{NS}(s,t)\ =&\ \sum_{m,p=0}^\infty s^m t^p \bd{\mu}_{m+p+2}^{(p+1)}\mid^{0}.
\end{align}
This cohomomorphism maps the string fields $(\Phi_{NS\textrm{-}NS},\Psi_{NS})$ to those
in the WZW-like formulation $(V_o,V_c)$ as
\begin{equation}
\pi_1^c\hat{\bd{g}}_c(e^{\wedge\Phi_{NS\textrm{-}NS}})\ =\ G_c(V_c),\qquad
\pi_1^o\hat{\bd{g}}_o\left(\frac{1}{1-\Psi_{NS}}\right)\ =\ G_o(V_o),
\end{equation}
where
\begin{equation}
G_c(V)\ =\
\eta\bar{\eta} V + \frac{1}{2}\left(
L^\eta_2(\eta\bar{\eta} V,\bar{\eta}V)+\eta L_2^{\bar{\eta}}(\eta\bar{\eta}V,V)
\right) + \cdots,
\end{equation}
is the pure-gauge string fields of type II superstring identically satisfying
\begin{equation}
\bd{L}_c^\eta(e^{\wedge G_c(V_c)})\ =\ 0,\qquad
\bd{L}_c^{\bar{\eta}}(e^{\wedge G_c(V_c)})\ =\ 0,
\end{equation}
with $(\bd{L}_c^\eta,\bd{L}_c^{\bar{\eta}})=(\hat{\bd{g}}_c\bd{\eta}_c\hat{\bd{g}}_c^{-1},
\hat{\bd{g}}_c\bar{\bd{\eta}}_c\hat{\bd{g}}_c^{-1})$ \cite{Matsunaga:2016zsu}.
The pure-gauge string field $G_o(V_o)$ of the open superstring is similarly defined by
a composite string field of $V_o$ identically satisfying the equation
\begin{equation}
\bd{L}_o^\eta\left(\frac{1}{1-G_o(V_o)}\right)\ =\ 0,
\end{equation}
with $\bd{L}_o^\eta=\hat{\bd{g}}_o\bd{\eta}_o\hat{\bd{g}}_o^{-1}$.
We give a prescription to obtain explicit form of
$G_o(V_o)$ in Appendix~\ref{App A}.
The (dynamical) equation of motion
of the open superstring
is mapped as
\begin{equation}
\pi_1\tilde{\bd{N}}_{NS}\left(e^{\wedge G_c(V_c)}\otimes\frac{1}{1-G_o(V_o)}\right)\ =\ 0,
\end{equation}
with
\begin{equation}
\tilde{\bd{N}}_{NS}\ =\ \hat{\bd{g}}\bd{N}_{NS}\hat{\bd{g}}^{-1}\ =\
\bd{Q}_o+\hat{\bd{g}}(\bd{N}_{NS}-\bd{M}_{NS})\hat{\bd{g}}^{-1},
\end{equation}
where
$V_c$
is a background field satisfying the equation of motion of the closed-superstring $Q_cG_c(V_c)=0$.
In order to give the WZW-like action deriving this equation of motion, we define the associated
string field as
\begin{equation}
B_d(V_o)\ =\ \pi_1^o\hat{\bd{g}}_o\bd{\xi_d}\left(\frac{1}{1-\Psi_{NS}}\right),
\end{equation}
where $d=t, \delta$ or $Q$ and $\bd{\xi_d}$ is the coderivation derived from
$\xi\partial_t$, $\xi\delta$ or $-\xi\pi_1\bd{M}_{NS}$, respectively. We can show that the relations
\begin{align}
&\ dG_o(V_o)\ =\ (-1)^d D_\eta B_d(V_o),\label{dG}\\
&\ D_\eta\left(\partial_tB_\delta(V_o)-\delta B_{\partial_t}(V_o)\right)\ =\ 0.
\end{align}
hold\footnote{Note that deg$(V_o)=1$ and deg$(B_d)=$deg$(d)+1$.}, where
$D_{\eta}$ is the nilpotent linear operator defined by
\begin{align}
D_\eta\varphi\
=&\ \pi_1^o\bd{L}_o^\eta\left(\frac{1}{1-G_o(V_o)}\otimes\varphi\otimes\frac{1}{1-G_o(V_o)}\right)
\nonumber\\
=&\ \pi_1^o\bd{L}^\eta\left(e^{\wedge G_c(V_c)}\otimes
\left(
\frac{1}{1-G_o(V_o)}\otimes\varphi\otimes\frac{1}{1-G_o(V_o)}\right)\right),
\end{align}
acting on an open superstring field $\varphi\in\mathcal{H}_{NS}$. The coderivation
$\bd{L}^{\eta}$ acts as $\bd{L}_c^{\eta}+\bd{L}_c^{\bar{\eta}}$ on $\mathcal{H}_c$ and as $\bd{L}_o^{\eta}$ on
$\mathcal{H}_o$.
Then, the WZW-like action for the NS sector is given by
\begin{equation}
I_{WZW}^{NS}\ =\ \int_0^1 dt\, \omega_l^o\left(B_{t}(t V_o),
\pi_1\tilde{\bd{N}}_{NS}\Big(e^{\wedge G_c(V_c)}\otimes\frac{1}{1-G_o(t V_o)}\Big)
\right),
\end{equation}
which is invariant under the gauge transformation
\begin{align}
B_\delta(V_o)\ =&\ \pi_1\tilde{\bd{N}}_{NS}\left(e^{\wedge G_c(V_c)}\otimes
\left(
\frac{1}{1-G_o(V_o)}\otimes\Lambda\otimes\frac{1}{1-G_o(V_o)}\right)\right)
+ D_\eta\Omega.
\end{align}
It is straightforward to extend these results of the NS sector to all the sectors.
Since $\hat{\bd{g}}_c$ and $\hat{\bd{g}}_o$ act as the identity operators outside the NS sector, we find that
\begin{align}
\pi_1^c\hat{\bd{g}}_c\left(e^{\wedge \Phi}\right)\ =&\ \pi_1^c\hat{\bd{g}}_c\left(e^{\wedge \Phi_{NS\textrm{-}NS}}\right)
+ \Phi_{R\textrm{-}NS} + \Phi_{NS\textrm{-}R} + \Phi_{R\textrm{-}R},\\
\pi_1^o\hat{\bd{g}}_o\left(\frac{1}{1-\Psi}\right)\ =&\
\pi_1^o\hat{\bd{g}}_o\left(\frac{1}{1-\Psi_{NS}}\right) + \Psi_{R},
\end{align}
and can identify the components $(\Phi_{R\textrm{-}NS},\Phi_{NS\textrm{-}R},\Phi_{R\textrm{-}R};\Psi_R)$ to those in the WZW-like formulation
$(\Psi_c, \bar{\Psi}_c,\Sigma_c;\Psi_o)$:
\begin{equation}
\Phi_{R\textrm{-}NS}\ =\ \Psi_c,\qquad \Phi_{NS\textrm{-}R}\ =\ \bar{\Psi}_c,\qquad
\Phi_{R\textrm{-}R}\ =\ \Sigma_c,\qquad \Psi_R\ =\ \Psi_o.
\end{equation}
Thus, these components are also annihilated by $\eta_c$ and $\bar{\eta}_c$ (or $\eta_o$)
and satisfy the constraint (\ref{closed constraint})
(or (\ref{open constraint})).
The WZW-like action of the open superstring field theory on the general closed-string backgrounds
is eventually written as
\begin{align}
I_{WZW}\ =&\ \int^1_0 dt\,\omega_l^o\Bigg(\mathcal{B}_{t}(\mathcal{V}_o(t)),
(\mathcal{G}^o)^{-1}\pi_1\tilde{\bd{N}}\bigg(e^{\wedge (G(V_c)+\Psi_c+\bar{\Psi}_c+\Sigma_c)}\otimes
\frac{1}{1-G_o(V_o(t))-\Psi_o(t)}\bigg)\Bigg),
\label{WZW}
\end{align}
where $\tilde{\bd{N}}=\hat{\bd{g}}\bd{N}\hat{\bd{g}}^{-1}$ and
\begin{equation}
\mathcal{B}_{t}(\mathcal{V}_o(t))\ =\ B_{t}(V_o(t))+\xi_0\partial_t\Psi_o(t).
\end{equation}
The closed superstring backgrounds $(V_c,\Psi_c,\bar{\Psi}_c,\Sigma_c)$ satisfy
\begin{equation}
\pi_1\tilde{\bd{L}}\left(e^{\wedge(G_c(V_c)+\Psi_c+\bar{\Psi}_c+\Sigma_c)}\right)\ =\ 0
\end{equation}
with $\tilde{\bd{L}}=\hat{\bd{g}}\bd{L}\hat{\bd{g}}^{-1}$.
Note that, since $\hat{\bd{g}}$ acts as the identity except on the NS sector,
$\tilde{\bd{N}}$ and $\tilde{\bd{L}}$ preserve the constraint (\ref{open constraint})
and (\ref{closed constraint}), respectively.
The WZW-like action (\ref{WZW}) is invariant under the gauge transformation
\begin{align}
\mathcal{B}_\delta(\mathcal{V}_o)\ =&\ \pi_1\tilde{\bd{N}}
\left(e^{\wedge(G_c(V_c)+\Psi_c+\bar{\Psi}_c+\Sigma_c)}\otimes
\left(
\frac{1}{1-G_o(V_o)-\Psi_o}\otimes
(\Lambda+\xi\lambda)\otimes
\frac{1}{1-G_o(V_o)-\Psi_o}\right)\right),
\end{align}
which is also obtained through the map $\hat{\bd{g}}$.
Here, $\Lambda$ and $\lambda$ are the gauge parameters in the NS and R sectors, respectively,
and $\lambda$ is annihilated by $\eta$ and satisfies the constraint (\ref{open constraint}).
\sectiono{Summary and discussion}\label{E}
In this paper, we constructed interactions for the open-closed superstring
field theory based on the OCHA structure. It provides the open-closed superstring
field theory on general closed-superstring backgrounds.
We also give a corresponding WZW-like action for open-closed superstring field
theory through a field redefinition.
Recently, the open string field theory deformed with a gauge invariant open-closed coupling
is studied \cite{Erbin:2020eyc,Koyama:2020qfb,Maccaferri:2021ulf,Maccaferri:2021ksp,Maccaferri:2021lau}.
The effective open superstring field theory is governed by a weak $A_\infty$ structure
which includes non-trivial tadpole term, destabilizing the initial perturbative vacuum.
It requires to shift the vacuum to a new equilibrium point.
The open-closed superstring field theory, given in this paper, provides a basis for such
an analysis on more general closed-superstring backgrounds described by classical solutions
of the nonlinear equation of motion of the closed superstring field theory.
In order to quantize the classical superstring field theory, we must
extend the classical action to the quantum master action satisfying the quantum BV equation.
Such an open-closed superstring field theory is recently given in Ref.~\cite{FarooghMoosavian:2019yke}
based on the formalism using the extra free field \cite{Sen:2015hha,Sen:2015uaa}.
It is interesting to give a quantum master action using the formulation based on the homotopy
algebra, which requires to extend the OCHA structure to the quantum OCHA structure
\cite{Munster:2011ij}.
The quantum open-closed superstring field theory is also practically useful
to study the string dynamics on the Ramond-Ramond backgrounds \cite{Cho:2018nfn},
the D-brane backgrounds \cite{Sen:2019qqg,Sen:2020cef,Sen:2020ruy,Sen:2020oqr,Sen:2021qdk},
and so on, which are difficult in the first-quantized formulation using the NSR formalism.
The (quantum) OCHA structure should shed new light on such nonperturbative studies.
\vspace{1cm}
\noindent
{\bf \large Acknowledgments}
The author would like to thank Jojiro Yoshinaka for pointing out an error in the draft.
This work is supported in part by JSPS Grant-in-Aid for Scientific
Research (C) Grant Number JP18K03645.
\vspace{1cm}
|
\section{Introduction}
Recently, deep convolution neural networks have made significant improvements in lots of fields, including but not limited to computer vision~\cite{krizhevsky2012imagenet,lin2017focal}, speech recognition, and natural language processing. Attracted by the excellent performance of CNN models, many people try to deploy CNNs to real-world applications. However, the superior performance of DNN usually requires powerful hardware with abundant computing and memory resources, for example, high-end graphics processing units (GPUs).
\begin{figure}[t]
\flushright
\includegraphics[width=3in]{figs/fig1_1.pdf}
\centering
\includegraphics[height=2.36in]{figs/fig1_2.pdf}
\centering
\includegraphics[width=3.15in]{figs/fig1_3.pdf}
\caption{Comparison between \textit{soft} and \textit{hard} threshold. The weight distributions are drawn from ResNet-18 on ImageNet. We take Layer1.0.conv1 as an example. (a) Our soft threshold ternarization. (b) Hard threshold ternarization which splits intervals with $\Delta$. The blue/red/green distribution comes from floating point weights quantized to -1/0/1, respectively. The blue/red/green lines below distributions denote positions of floating point weights that will be quantized to -1/0/1 when sorting weights from small to large.}
\label{fig:soft_threshold}
\end{figure}
Deploying deep neural networks to resource limited devices is still a challenging problem due to the requirements of abundant computing and memory resources. A variety of methods have been proposed to reduce the parameter size and accelerate the inference phase, such as compact model architecture design (in a handcrafted way or automatic search way)~\cite{howard2017mobilenets,zoph2016neural}, network pruning~\cite{han2015deep,he2019filter}, knowledge distilling~\cite{hinton2015distilling}, low-rank approximation~\cite{jaderberg2014speeding,jaderberg2014speeding}, etc. Besides, recent works~\cite{courbariaux2014low,gupta2015deep} show that full-precision weights and activations are not necessary for networks to achieve high performances. This discovery indicates that both weights and activations in neural networks can be quantized to
low-bit formats. In this way, both storage and computation resources can be saved.
The extreme cases of network quantization are binary and ternary quantization ($\{-1,1\}$ and $\{-1,0,1\}$). In computationally heavy convolutions, multiply-accumulate consumes most of the operation time. Through binarization or ternarization, multiply-accumulate can be replaced by cheap bitwise $xnor$ and $popcount$ operations~\cite{courbariaux2016binarized}.
Although binary networks can achieve high compression rate and computing efficiency, they inevitably suffer from a long training procedure and noticeable accuracy drops owing to poor representation ability.
In order to make a balance between efficiency and accuracy, ternary CNNs convert both weights and activations into ternary values. Theoretically, ternary CNNs have stronger representation ability than binary CNNs, and should have a better performance. Although many works~\cite{li2016ternary,zhu2016trained,wang2018two,wan2018tbn} provide tremendous efforts on ternary neural networks, we argue that there are two main issues in existing works:
1) All previous ternary networks use a \textit{hard} $\Delta$ to divide floating point weights into three intervals (see Eq.\ref{eq:wt} for more details), which introduces an additional constraint into ternary networks.
2) Although the theoretical optimal $\Delta$ can be calculated through optimization, calculating it at each training iteration is too time-consuming. Therefore, previous methods either use a hyper-parameter to estimate $\Delta$ or only calculate the exact $\Delta$ at the first training iteration then keep it for later iterations, which further limits ternary network accuracy. A brief summary: they introduce an additional constraint based on $\Delta$. Further more, a suboptimal $\Delta$ is used instead of the exact one.
In this paper, we propose Soft Threshold Ternary Networks (STTN), an innovative method to train ternary networks. STTN avoids the \textit{hard} threshold and enables the model to automatically determine which weights to be $-1/0/1$ in a \textit{soft} manner.
We convert training ternary networks into training equivalent binary networks (more details in section 3.2). STTN enjoys several benefits: 1) The constraint based on $\Delta$ is removed. 2) STTN is free of calculation of $\Delta$. 3) As shown in Figure~\ref{fig:soft_threshold} (a), weights are ternarized in a soft manner, which is why we name it \textit{soft} threshold.
Our contributions are summarized as follows.
\begin{itemize}
\item We divide previous ternary networks into two catalogues: Optimization-based method and Learning-based method. Analysis is given about issues in the existing ternary networks that prevent them from reaching high performance.
\item We propose STTN, an innovative way to train ternary networks by decomposing one ternary kernel into two binary kernels during training. No extra parameters or computations are needed during inference. Quantizing to -1/0/1 is determined by two binary variables rather than decided by a hard threshold.
\item We show that our proposed ternary training method provides competitive results on the image classification, i.e. CIFAR-10, CIFAR-100 and ImageNet datasets. Qualitative and quantitative experiments show that it outperforms previous ternary works.
\end{itemize}
\section{Related Work}
Low-bit quantization of deep neural networks has recently received increasing interest in deep learning communities.
By utilizing fixed-point weights and feature representations, not only the model size can be dramatically reduced, but also inference time can be saved.
The extreme cases of network quantization are binary neural networks (BNN) and ternary neural networks (TNN). The runtime speed of BNN/ TNN can be comparable with GPU by replacing expensive floating-point matrix multiplication with high computation efficiency POPCOUNT-XNOR operations~\cite{rastegari2016xnor}. And our study of ternary networks touches upon the following areas of research.
\subsection{Binary Networks}
BNN~\cite{courbariaux2016binarized} constrains both the weights and activations to either +1 or -1, which produces reasonable results on small datasets, such as MNIST and CIFAR-10. However, there is a significant accuracy drop on large scale classification datasets, such as ImageNet. Some improvements based on BNN have been investigated. For example,~\cite{darabi2018bnn+} introduces a regularization function that encourages training weights around binary values.~\cite{tang2017train} proposes to use a low initial learning rate. Nevertheless, there is still a non-negligible accuracy drop. To improve the quality of the binary feature representations, XNOR-Net~\cite{rastegari2016xnor} introduces scale factors for both weights and activations during binarization process. DoReFa-Net~\cite{zhou2016dorefa} further improves XNOR-Net by approximating the activations with more bits. Since ABC-Net~\cite{lin2017towards}, several works propose to decompose a single convolution layer into K binary convolution operations~\cite{gu2019bonn,zhu2019binary}. Although higher accuracy can be achieved, K$\times$ extra parameters and computations are needed for both training and inference time, which defeats the original purpose of binary networks. Different from those K$\times$ binary methods, we decompose a ternary convolution layer into two parallel binary convolution layers only during training. Once training finishes, we only store the ternary model and use the ternary model for inference.
\subsection{Ternary Networks}
To make a balance between efficiency and accuracy, ternary CNNs convert both weights and activations into ternary values. Cheap bitwise operations can still be used to save inference time in this case. TWN~\cite{li2016ternary} ternarize neural networks weights to be $\{+1, 0, -1\}$ by minimizing Euclidian distance between full precision weights and the ternary weights along with a scaling factor. TTQ~\cite{zhu2016trained} set scaling factors as learnable parameters to achieve higher performance. TSQ~\cite{wang2018two} train a full precision network at first, then initialize ternary weights with pretrained parameters through solving an optimization problem. RTN~\cite{li2019rtn} proposes to reparameter activation and weights by introducing extra learnable parameters. To tackle the challenge that the complexity of ternary inner product is the same as the standard 2-bit counterparts,
FATNN~\cite{chen2021fatnn} proposes a new ternary quantization pipeline, which reduces the complexity of TNNs by 2$\times$.
We will divide those methods into two catalogues and analyse the unresolved/overlooked issues in previous works.
We will review previous ternary works in detail in section 3.1.
\section{Methodology}
We first revisit the formulation of previous ternary networks. We divide them into two catalogues and show their common issues in calculating the appropriate $\Delta$.
We then present our novel Soft Threshold Ternary Networks in detail, including scaling coefficient constraint and backward approximation. We show that our method can avoid the previous issue in a simple but effective way.
\subsection{Review of Previous Ternary Networks}
\subsubsection{Problem Formulation}
Full precision convolution or inner-product can be formed as: $Y=\sigma(W^TX)$. Here $\sigma(\cdot)$ represents the nonlinear activation function, such as ReLU. $W\in\mathbb{R}^{n\times{chw}}$ and $X\in\mathbb{R}^{chw\times{HW}}$ are float point weights and inputs respectively, where $(n, c, h, w)$ are filter number, input channels, kernel height and kernel width, and $(H,W)$ are height and width of output feature maps. Ternary networks convert both weights and inputs into ternary values: $T\in\{+1, 0, -1\}^{n\times{chw}}$ and $X^t\in\{+1, 0, -1\}^{chw\times{HW}}$.
As for weights ternarization,
$\alpha$ is used to estimate the floating weight $W$ alone with ternary weight $T$.
\begin{equation}
W\approx\alpha T
\end{equation}
Previous works formulate ternarization as a weight approximation optimization problem:
\begin{equation}
\begin{cases}
\alpha^*,T^*= & \mathop{\arg\min}\limits_{\alpha,T}
J(\alpha, T) = \Vert W-\alpha T\Vert_2^2 \\
\qquad s.t. & \alpha\ge0, T_i \in \{-1, 0, 1\}, i = 1,2,...,n. \\
\end{cases}
\label{eq:opt1}
\end{equation}%
A \textit{hard} threshold $\Delta$ is then introduced by previous works to divide quantization intervals, which sets an additional constraint on ternary networks:
\begin{equation}
T_i=
\begin{cases}
+1,& \text{if $W_i>\Delta$}\\
0,& \text{if $\left|W_i\right|\le \Delta$}\\
-1,& \text{if $W_i<-\Delta$}
\end{cases}
\label{eq:wt}
\end{equation}%
As illustrated in Figure \ref{fig:soft_threshold}(b), $\Delta$ plays a role as hard threshold here, splitting the floating point weights into $\{-1,0,1\}$. We will show that the above constraint can be avoided in our method. According to methods of calculating $\alpha$ and $\Delta$, we divide previous ternary networks into 1) optimization-based ternary networks, and 2) learning-based ternary networks.
\subsubsection{Optimization-based Ternary Networks}
TWN~\cite{li2016ternary} and TSQ~\cite{wang2018two} formulate the ternarization as Eq.(\ref{eq:opt1}).
In TWN, $T$ is obtained according to threshold-based constraint Eq.(\ref{eq:wt}). They give exact optimal $\alpha^*$ and $\Delta^*$ as follows:
\begin{equation}
\resizebox{.91\linewidth}{!}{$
\alpha^*_{\Delta}= \frac{1}{|I_\Delta|}\sum\limits_{i\in I_\Delta}|W_i|;~~
\Delta^* = \mathop{\arg\max}\limits_{\Delta>0} \frac{1}{|I_\Delta|}(\sum\limits_{i\in I_\Delta} |W_i|)^2
$}
\label{eq:solution1}
\end{equation}%
where $I_\Delta=\{i\big||W_i|>\Delta\}$ and $|I_\Delta|$ denotes the number of elements in $I_\Delta$.
Note that $\Delta^*$ in Eq.(\ref{eq:solution1}) has no straightforward solution. It can be time and computing consuming if using discrete optimization to solve it. In TWN, they adopt an estimation $\Delta '\approx 0.7\cdot E(|W|) $ to approximate the optimal $\Delta^*$. Obviously there is a gap between $\Delta^*$ and $\Delta '$. What's more, the optimal scaling factor $\alpha^*$ in Eq.(\ref{eq:solution1}) will also be affected because of dependence on $\Delta$.
In TSQ, they try to obtain $T^*$ by directly solving Eq.(\ref{eq:opt1}). They give exact optimal $\alpha^*$ and $T^*$ as follows\footnote{$W^T$ denotes transpose of matrix $W$. $T$ denotes ternary weights.}:
\begin{equation}
\alpha^*= \frac{W^T T}{T^T T};~~~
T^* = \arg\max\limits_{T}\frac{(W^TT)^2}{T^T T}
\label{eq:solution2}
\end{equation}
Note that there is still no straightforward solution for $T^*$ in Eq.(\ref{eq:solution2}). TSQ proposes OTWA(Optimal Ternary Weights Approximation) to obtain $T^*$. However, sorting for $|W_i|$ is needed in their algorithm. Time complexity is higher than $O(N\log(N))$, where $N$ is the number of elements in kernel weights. To avoid time consuming sorting, they only calculate $\Delta$ at the first training iteration and then keep it for later iterations, which is suboptimal.
\subsubsection{Learning-based Ternary Networks}
TTQ~\cite{zhu2016trained} tries another way to obtain scaling factor $\alpha$ and threshold $\Delta$. They set $\alpha$ as a learnable parameter and set $\Delta$ as a hyper-parameter. Based on TTQ, RTN~\cite{li2019rtn} introduces extra float point scaling and offset parameters to reparameterize ternary weights.
Because the search space of hyper-parameter $\Delta$ is too large, both of TTQ and RTN set $\Delta$ the same across all layers, e.g. $\Delta=0.5$ in RTN. We argue that ternary methods with hyper-parameter are not reasonable as following:
\begin{figure}[t]
\includegraphics[width=3.33in]{figs/fig2.pdf}
\caption{Weight distribution of ResNet-18 on ImageNet. (a) Distributions from different layers. (b) Distributions from different kernels in the same layer. "Layer1.0.conv2[36]" denotes the 37th kernel in Layer1.0.conv2. The blue/red/green lines below each distribution denote positions of floating point weights that will be quantized to -1/0/1, which are obtained through our STTN.}
\label{fig:weight_distribution}
\end{figure}
In Figure~\ref{fig:weight_distribution}, we visualize the weight distribution of ResNet-18 on ImageNet from different layers and from different kernels in the same layer, respectively. From the figure, we can observe that the distributions are different from layer to layer. Even for kernels in the same layer, their weight distributions can be diverse. Therefore, 1) it is not reasonable for previous works to set a fixed threshold for all kernels (e.g. $\Delta$ = 0.5 in TTQ and RTN). 2) What's more, their ternarization also use hard threshold: once $\Delta$ is set, their quantized intervals are determined according to Eq.(\ref{eq:wt}).
So can we remove the hard $\Delta$ constraint Eq.(\ref{eq:wt}), solving ternary networks in another way?
\subsection{Soft Threshold Ternary Networks}
Due to the issues analyzed above, we propose our novel STTN. Our motivation is to enable the model to automatically determine which weights to be -1/0/1, avoiding the hard threshold $\Delta$.
We introduce our methods via convolution layers. The inner-product layer has a similar form. Concretely, at \textit{training} time, we replace the ternary convolution filter $T$ with two parallel binary convolution filters $B_1$ and $B_2$. They are both binary-valued and have the same shape with ternary filter: $n\times chw$. Due to the additivity of convolutions with the same kernel sizes, a new kernel can be obtained by:
\begin{equation}
T = B_1 + B_2
\end{equation}
A key requirement for $T$ to have ternary values is that those two binary filters should have the same scaling factors: $\alpha$. With $\alpha B_1,\alpha B_2 \in \{+\alpha,-\alpha \}^{n\times chw}$, the sum of $\alpha B_1$ and $\alpha B_2$ is ternary-valued, i.e. $\alpha T \in \{+2\alpha, 0, -2\alpha\}^{n\times{chw}}$. Note that we only decompose the filters at training time. After the training, the two trained parallel binary filters are added up to obtain the ternary filter. Thus there is no extra computation when deploying trained models to devices. An example is illustrated in Figure \ref{fig:add}.
Zeroes are introduced in ternary filters (white squares in Figure \ref{fig:add} at positions where two parallel filters have opposite values). And -1/1 is obtained at positions where two parallel filters have the same signs. In this way, ternary values are determined by the combination of two corresponding binary values.
It is obvious that the outputs of training-time model are equal to the outputs of inference-time model:
\begin{equation}
Y=\sigma ( \alpha B_1^T X + \alpha B_2^T X)
=\sigma(\alpha T^T X)
\end{equation}
In our proposed STTN, we convert training ternary networks into training equivalent binary networks. We abandon the constraint in Eq.(\ref{eq:wt}). Not only $\Delta$ can be avoided, but also quantization intervals can be divided in a soft manner.
\begin{figure}[t]
\includegraphics[height=1.5in]{figs/fig3.pdf}
\caption{We illustrate our method with a 2D convolution for simplicity. The top of the figure are two training-time binary convolution kernels. Their weights are binary-valued:$\{+\alpha, -\alpha\}$. The bottom of the figure is an inference-time ternary convolution kernel who takes the same input. The inference-time kernel can be obtained by easily adding the corresponding two binary kernels. The weights are ternary-valued: $\{+2\alpha, 0, -2\alpha\}$. The additivity of convolutions tells us: the inference-time model can produce the same outputs as the training-time.}
\label{fig:add}
\end{figure}
\subsubsection{Scaling Coefficient}
As mentioned above, a key requirement is that the corresponding two parallel binary filters should have the same scaling factors to guarantee the sum of them are ternary. Taking the two parallel binary filters into consideration, we obtain the appropriate $\{\alpha_1, \alpha_2\}$ by minimizing the following weight approximation problem:
\begin{equation}
\begin{aligned}
J(\alpha_1,\alpha_2,B_1,B_2) &= \Vert W_1- \alpha_1 B_1 \Vert^2_2
+ \Vert W_2- \alpha_2 B_2 \Vert^2_2 \\
&s.t. ~~\alpha_1 = \alpha_2; ~\alpha_1, \alpha_2 \ge 0
\label{eq:cal_alpha}
\end{aligned}
\end{equation}
Here $B_1$ and $B_2$ are the two parallel binary filters. And $W_1$ and $W_2$ are the corresponding float point filters. With constraint $\alpha_1 = \alpha_2$, we use $\alpha$ to denote them. Through expanding Eq.(\ref{eq:cal_alpha}), we have
\begin{equation}
\resizebox{.99\linewidth}{!}{$
\begin{aligned}
J(\alpha_1,\alpha_2,B_1,B_2) =
\alpha^2(B_1^T B_1 +B_2^T B_2)
-2\alpha(B_1^T W_1 + B_2^T W_2) + C
\label{eq:alpha_b}
\end{aligned}
$}
\end{equation}
$C=W_1^TW_1 + W_2^TW_2$ is a constant because $W_1$ and $W_2$ are known variables. In order to get $\alpha^*$, the optimal $B_1^*$ and $B_2^*$ should be determined. Since $B_1,B2\in \{+1,-1\}^{n\times chw}$, $B_1^T B_1+B_2^T B_2 = 2nchw$ is also a constant. From Eq.(\ref{eq:alpha_b}), $B_1^*$ and $B_2^*$ can be achieved by maximizing $B_1^T W_1 + B_2^T W_2$ with constraint condition that $B_1,B_2 \in \{+1,-1\}^{n\times chw}$. Obviously the optimal solution can be obtained when binary kernel has the same sign with the corresponding float point kernel at the same positions, i.e. $B_1^*=sign(W_1)$, $B_2^*=sign(W_2)$.
Based on the optimal $B_1^*$ and $B_2^*$, $\alpha^*$ can be easily calculated as:
\begin{equation}
\resizebox{.89\linewidth}{!}{$
\begin{aligned}
\alpha^* = \frac{B_1^T W_1 +B_2^T W_2}{B_1^T B_1 + B_2^T B_2}
=\frac{1}{2N}(\sum_{i=1}^N |W_{1i}| + \sum_{i=1}^N |W_{2i}|)
\label{eq:my_alpha}
\end{aligned}
$}
\end{equation}
where $N=nchw$, is the number of elements in each weight. $W_{1i}$ and $W_{2i}$ are elements of $W_{1}$ and $W_{2}$, respectively.
\subsubsection{Backward Approximation}
Since we decompose one ternary filter into two parallel binary filters at training time, binary weights approximation is needed in both forward and backward processes. During the forward propagation, the two related weights can be binarized through sign function along with the same scaling factor calculated by Eq.(\ref{eq:my_alpha}). However, during the backward propagation, the gradient of sign function is almost everywhere zero. Assume $\ell$ as the loss function and $\widetilde{W}=\alpha B=\alpha \cdot sign(W)$ as the approximated weights, XNOR-Net~\cite{rastegari2016xnor} alleviates this problem through Straight-Through Estimator (STE):
\begin{equation}
\begin{aligned}
\frac{\partial\ell}{\partial W_i} = \frac{\partial \ell}{\partial \widetilde{W_i}}\frac{\partial \widetilde{W_i}}{\partial W_i} = \frac{\partial \ell}{\partial \widetilde{W_i}}\Big(\frac{1}{N}+ \alpha \frac{\partial sign(W_i)}{\partial W_i}\Big)
\label{eq:xnor_derivate_w}
\end{aligned}
\end{equation}
Here, $sign(W_i)$ is approximated with $W_i\boldsymbol{1}_{|W_i| \le 1}$. $N$ is the number of elements in each weight.
However, note that an important requirement in our STTN is that the related two parallel binary filters should have the same scaling factors. The exact approximated weights should be:
\begin{equation}
\widetilde{W}=\alpha B= \frac{1}{2N}(\sum_{i=1}^N |W_{1i}| + \sum_{i=1}^N |W_{2i}|) \cdot sign(W)
\end{equation}
Because Eq.(\ref{eq:my_alpha}) indicates that $\alpha$ is dependent on both $W_1$ and $W_2$. When calculating the derivatives of $W_{1i}$, the effect of other kernels $W_{1j}$ and $W_{2j}$ should be considered. But Eq.(\ref{eq:xnor_derivate_w}) ignores the effect of $W_{1j}$ and $W_{2j}$, which is not suitable for our backward approximation.
Taking above analysis into consideration, we propose to calculate derivatives of $W$ in a more precise way:
\begin{equation}
\resizebox{.98\linewidth}{!}{$
\begin{aligned}
\frac{\partial \ell}{\partial W_{1i}}
&=\sum_{k=1}^{2} \sum_{j=1}^{N}
\frac{\partial \ell}{\partial \widetilde{W}_{kj}} \frac{\partial \widetilde{W}_{kj}}{\partial W_{1i}} \\
&=\sum_{k=1}^{2} \sum_{j=1}^{N} \frac{\partial \ell}{\partial \widetilde{W}_{kj}} \Big[ \frac{1}{2N} \frac{\partial \vert W_{1i}\vert}{\partial W_{1i}} sign(W_{kj}) + \alpha \frac{\partial sign(W_{1i})}{\partial W_{1i}} \Big]
\\
&=\frac{1}{2N}sign(W_{1i}) \sum_{k=1}^{2} \sum_{j=1}^{N} \Big[ \frac{\partial \ell}{\partial \widetilde{W}_{kj}} sign(W_{kj})\Big]
\\
&{}+\alpha \frac{\partial sign(W_{1i})}{\partial W_{1i}} \sum_{k=1}^{2} \sum_{j=1}^{N} \frac{\partial \ell}{\partial \widetilde{W}_{kj}}
\end{aligned}
$}
\label{eq:gradient_of_W}
\end{equation}
Here $W_k=[W_{k1}, W_{k2},...,W_{kN}]$ (~$k\in\{1,2\}$) are the two parallel kernels respectively.
$\partial \ell / \partial W_{2i}$ can be calculated in the same way as Eq.({\ref{eq:gradient_of_W}).
\subsubsection{Activation}
In this paper, we also convert activations into ternary values. We use the same ternarization function as RTN \cite{li2019rtn}. Given floating point activation $X$, the ternary activation is calculated by the following equation. The difference between ours and RTN is that we do not introduce extra parameters or calculations.
\begin{equation}
X^t_i = Ternarize(X_i)=
\begin{cases}
sign(X_i),& \text{if $|X_i|>0.5$} \\
0,& \text{otherwise} \\
\end{cases}
\end{equation}
During the backward process, as previous binary/ternary works~\cite{courbariaux2016binarized,rastegari2016xnor,liu2018bi}, gradients propagated through ternarization function are estimated by Straight-Through Estimator (STE).
\section{Experiments}
In this section, we evaluate the proposed STTN in terms of qualitative and quantitative studies. Our experiments are conducted on three popular image classification datasets: CIFAR-10, CIFAR-100 and ImageNet (ILSVRC12). We test on several representative CNNs including: AlexNet, VGG-Net, and ResNet.
\subsection{Implementation Details}
We adopt the standard data augmentation scheme. In all CIFAR experiments, we pad 2 pixels in each side of images and randomly crop 32$\times$32 size from padded images during training. As for ImageNet experiments, we first proportionally resize images to 256$\times N$ ($N \times 256$) with the short edge to 256. Then we randomly sub-crop them to 224$\times$224 patches with mean subtraction and randomly flipping. No other data augmentation tricks are used during training.
Following RTN~\cite{li2019rtn}, we modify the block structure as BatchNorm $\rightarrow$ Ternarization $\rightarrow$ TernaryConv $\rightarrow$ Activation. Following XNOR-Net~\cite{rastegari2016xnor}, we place a dropout layer with $p = 0.5$ before the last layer for AlexNet. For VGG-Net, we use the same architecture VGG-7 as TWN~\cite{li2016ternary} and TBN~\cite{wan2018tbn}. We do not quantize the first and the last layer as previous binary/ternary works. We replace all $1\times1$ downsampling layers with max-pooling in ResNet.
We use Adam with default settings in all our experiments. The batch size for ImageNet is 256. We set weight decay as $1e^{-6}$ and momentum as 0.9. All networks on ImageNet are trained for 110 epochs. The initial learning rate is 0.005, and we use cosine learning rate decay policy. All our models are trained from scratch.
\subsection{Weight Approximation Evaluation}
\begin{table}
\centering
\begin{tabular}{lllr}
\toprule
Method &TWN &TTQ &Ours \\
\midrule
$\Vert W-\alpha T\Vert_2^2$ &537.76 &439.46 &\textbf{379.25}\\
\bottomrule
\end{tabular}
\caption{$L2$ distance between approximated ternary weights and float point weights. (We add up distances of all convolution layers together.)}
\label{tab:weight_appro}
\end{table}
In this section, we explore the effect of the proposed STTN from qualitative view
Previous works quantize weights into $\{-1,0,1\}$ by setting a hard threshold $\Delta$. Different from them, the proposed STTN
generates soft threshold, quantizing weights more flexibly.
We illustrate the impact of threshold calculation on the performance of TNN based on ResNet-18. We first calculate the distance between trained floating weights $W$ and trained ternary weights $T$. $L2$ norm is used as the criterion for measurement as Eq.(\ref{eq:opt1}). We compare our method with TW
and TT
. The results are shown in Table~\ref{tab:weight_appro}.
We can see that STTN obtains the smallest gap between trained floating weights and trained ternary weights, which indicates that our methods can realize a good weight approximation. Intuitively, the smaller the approximated weight error we get, the higher precision the model can obtain. The results show the effect of STTN qualitatively and further quantitative analyses are given in section 4.3.
We also analyze the reason why our STTN can obtain a smaller approximation error than previous ternary method. An essential effect of STTN is to quantize weights into $\{+1,0,-1\}$ in a soft manner, just as shown in Figure \ref{fig:soft_threshold}(a) and Figure \ref{fig:weight_distribution}. The effect comes from that STTN throws the hard threshold away. That is, Eq.(\ref{eq:wt}) is no more a constraint to the weight approximation optimization problem. By comparing Figure\ref{fig:soft_threshold}(a) with (b): this provides more flexible ternarization intervals.
Besides, we find that STTN adjust the weight sparsity regularly. Figure~\ref{fig:sparsity} shows the weight sparsity rates (the percentage of zeros) of different layers in our STTN on ResNet-18. From the figure we can see that the sparsity rates gradually decrease from the first layer to the last layer. This probably because high-level semantics need dense kernels to encode.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in]{figs/fig4.pdf}
\caption{Weight sparsity rate of different layer in our STTN on ImageNet with ResNet-18. Here we illustrate it with 16 convolution layers in building blocks of ResNet.
\label{fig:sparsity}
\end{figure}
\subsection{Network Ternarization Results}
In this section, we evaluate the STTN from quantitative view by comparing with the state-of-the-art low-bit networks on various architectures. We ternarize both weights and activations. Experiments on only quantizing weights are also given.
\subsubsection{Results on CIFAR-10}
We first conduct experiments on CIFAR-10 dataset. We use the same network architecture as TWN, denoted as VGG-7. Compared with the architecture VGG-9 adopted in BNN and XNOR, the last two fully connection layers are removed. Table~\ref{tab:vgg7_cifar10} shows the STTN results. Note that for VGG-7, STTN with ternary weights and activations can even obtain better performance than the full-precision model.
\begin{table}
\centering
\begin{tabular}{lrr}
\toprule
Bit-width &Method &Error($\%$) \\
\midrule
32+32 &Floating point~\cite{li2016ternary} &7.12 \\
\midrule
1+32$^\dag$ &BWN~\cite{courbariaux2015binaryconnect} &8.27 \\
1+1$^\dag$ &BNN~\cite{courbariaux2016binarized} &10.15 \\
1+1$^\dag$ &XNOR~\cite{rastegari2016xnor} &9.98 \\
2+32 &TWN~\cite{li2016ternary} &7.44 \\
1+2 &TBN~\cite{wan2018tbn} &9.15 \\
\midrule
2+2 &Ours &\textbf{7.07} \\
\bottomrule
\end{tabular}
\caption{The error rates on CIFAR-10 with VGG-7. The number before and after "+" in the first column denotes the weight and activation bit-width respectively. $\dag$ denotes the architecture is VGG-9, which adds two more FC layers at last.}
\label{tab:vgg7_cifar10}
\end{table}
\subsubsection{Results on CIFAR-100}
In addition, we also evaluate STTN on CIFAR-100 dataset. We compare our STTN with a strong multi-bit baseline CBCN~\cite{gu2019bonn}. CBCN replaces each convolution layer with several parallel binarized convolution layers. For fair comparisons, we use the same architecture as CBCN (ResNet-18 with 32-64-128-256 kernel stage). Note that in CBCN, the number of channels in one layer is $4\times$. Table~\ref{tab:res18_cifar100} shows our results that although CBCN uses $4\times$ channels than ours, we obtain higher accuracy with fewer computations.
From this experiment, we argue that ternary networks can be considered before resorting to multi-bit methods.
\begin{table}
\centering
\begin{tabular}{llr}
\toprule
Model &Kernel Stage &Accuracy($\%$) \\
\midrule
Float(32+32) &32-64-128-256 &73.62 \\
\midrule
CBCN(1+1) &(32-64-128-256)$\times$4 &70.07 \\
Ours(2+2) &32-64-128-256 &\textbf{72.10} \\
\bottomrule
\end{tabular}
\caption{Accuracy on CIFAR-100 with ResNet-18 (32-64-128-256). Our ternary networks can outperform multi-bit method significantly.}
\label{tab:res18_cifar100}
\end{table}
\subsubsection{Results on ImageNet}
For the large-scale dataset, we evaluate our STTN over AlexNet and ResNet-18 on ImageNet. We compare our method with several exiting state-of-the-art low-bit quantization methods: 1) only quantizing weights: BWN~\cite{courbariaux2015binaryconnect}, TWN~\cite{li2016ternary} and TTQ~\cite{zhu2016trained}; 2) quantizing both weights and activations: XNOR~\cite{rastegari2016xnor}, Bi-Real~\cite{liu2018bi}, ABC~\cite{lin2017towards},
TBN~\cite{wan2018tbn}, HWGQ~\cite{cai2017deep}, PACT~\cite{choi2018pact}, RTN~\cite{li2019rtn} and FATNN~\cite{chen2021fatnn}.
The overall results based on AlexNet and ResNet-18 are shown in Tabel~\ref{tab:alex_imagenet} and~\ref{tab:res18_imagenet}. We highlight our accuracy improvement (up to \textbf{15$\%$} absolute improvement compared with XNOR-Net and up to \textbf{1.7$\%$} compared with state-of-the-art ternary models, \textbf{without pre-training}). These results show that the STTN outperforms the best previous ternary methods. Such improvement indicates that our soft threshold significantly benefits extreme low-bit networks.
What's more, compared with PACT and RTN, we highlight additional improvements apart from accuracy: 1) Both PACT and RTN introduce extra floating point parameters (the activation clipping level parameter in PACT and reparameterized scale/offset in RTN) into the networks, which needs extra storage space and computation. 2) The extra introduced learnable parameters in PACT and RTN need careful manual adjustments, such as learning rate, weight decay and so on. Extensive manual tuning have to be tested for different networks on different datasets. However, our method is free of extra hyper-parameters to be tuned. Further more, we argue that our method can combine with those methods for further accuracy improvement. 3) For RTN, they argue that initialization from pre-trained full-precision models is vitally important for their methods (since the small architecture modification such as changing the order of BN and Conv layer in low-bit quantization, full-precision models released by open model zoo can not be used directly).
It is widely known that minor architecture modification (i.g. changing the order of BN and Conv) is needed in low-bit quantization. Therefore float models released by open-source model zoo can not be used. Most previous works first train full-precision weights from scratch and then use it to initialize the low-bit networks. Compared to prior works, pre-trained full precision models are not needed in our method thus much training time can be saved. And we show that by eliminating the hard threshold constraint, training from scratch can still obtain state-of-the-art results.
\begin{table}
\centering
\begin{tabular}{lccc}
\toprule
Model &Bit-width &Top-1($\%$) &Top-5($\%$)\\
\midrule
XNOR &1+1 &44.2 &69.2 \\
TBN &1+2 &49.7 &74.2 \\
HWGQ &2+2 &52.7 &76.3 \\
PACT$^*$ &2+2 &55.0 &$-$ \\
RTN &2+2 &53.9 &$-$ \\
Ours &2+2 &\textbf{55.6} &\textbf{78.9} \\
\bottomrule
\end{tabular}
\caption{Comparison with the state-of-the-art methods on ImageNet with AlexNet. "$-$" means the accuracy is not reported. "$*$" indicates the networks use quaternary values instead of ternary values for 2 bits representation.}
\label{tab:alex_imagenet}
\end{table}
\begin{table}
\centering
\begin{tabular}{lcccc}
\toprule
Model &Bit-width &Top-1($\%$) &Top-5($\%$)\\
\midrule
Floating &32+32 &69.3 &89.2 \\
\midrule
TWN &2+32 &61.8 &84.2 \\
TWN$^{**}$ &2+32 &65.3 &86.2 \\
TTQ$^{**}$ &2+32 &66.6 &87.2 \\
RTN$^{***}$ &2+32 &68.5 &$-$ \\
Ours &2+32 &\textbf{68.8} &\textbf{88.3} \\
Ours (update) &2+32 &\textbf{69.2} &\textbf{89.3} \\
\midrule
XNOR &1+1 &51.2 &73.2 \\
Bi-Real$^{***}$ &1+1 &56.4 &79.5 \\
ABC-5 &(1$\times$5)+(1$\times$5) &65.0 &85.9 \\
TBN &1+2 &55.6 &79.0 \\
HWGQ &1+2 &56.1 &79.7 \\
PACT$^*$ &2+2 &64.4 &$-$ \\
RTN$^{***}$ &2+2 &64.5 &$-$ \\
FATNN &2+2 &66.0 &86.4 \\
Ours &2+2 &\textbf{66.2} &\textbf{86.4} \\
Ours (update) &2+2 &\textbf{68.2} &\textbf{87.9} \\
\bottomrule
\end{tabular}
\caption{Comparison with the state-of-the-art methods on ImageNet with ResNet-18. "$*$" indicates the model uses quaternary values instead of ternary values for 2 bits representation. "$**$" indicates the filter number of the network is $1.5\times$. "$***$" indicates the model needs full-precision pre-trained models to initialize. "$\times$" in "ABC-5" denotes multi-bit networks with multi-branch.}
\label{tab:res18_imagenet}
\end{table}
\section{Conclusion}
In this paper, we propose a simple yet effective ternarization method, Soft Threshold Ternary Networks. We divide previous ternary works into two catalogues and show that their hard threshold is suboptimal. By simply replacing the original ternary kernel with two parallel binary kernels at training, our model can automatically determine which weights to be -1/0/1 instead of depending on a hard threshold. Experiments on various datasets show that STTN dramatically outperforms current state-of-the-arts, lowering the performance gap between full-precision networks and extreme low bit networks.
\bibliographystyle{named}
|
\section{Overview}\label{sec-intro}
We consider the problem of implementing
linearizable objects that support both read
and read-modify-write (RMW) operations in message-passing systems with process crashes.
Since in many systems read operations vastly outnumber RMW operations,
we are interested in implementations that emphasize the efficiency of read operations.
We present a parametrized, leader-based algorithm for partially synchronous systems
where processes have access to clocks that are synchronized within
$\epsilon$;
such clocks can be provided by external devices such as GPS~\cite{CD+12}
which provide a very small $\epsilon$.
With this algorithm, every read operation is local (intuitively, it does not trigger messages).
If a read is not concurrent with a conflicting RMW, it is performed immediately with no waiting;
furthermore, even with a concurrent conflicting RMW,
a read experiences very little delay in the worst-case.
For example, the algorithm's parameters can be set to ensure that (after the system stabilizes) every read takes $\epsilon$ time in the worst-case.
If $\epsilon\le\delta/2$, where $\delta$ is the maximum message delay,
this nearly matches a lower bound by Chandra \emph{et al.} (Theorem~4.1 in~\cite{CHT16}).
To~the best of our knowledge this is the first algorithm
to achieve this for \emph{linearizable} object implementations in the partially synchronous systems that~we~assume~here.
The algorithm's parameters can be used to trade-off the worst-case times for read and RMW operations.
They can also be used to take advantage of the fact that in many message-passing systems
the delay of most messages is orders of magnitude smaller than the maximum message delay~$\delta$:
for example,
the parameters can be set so that,
in ``nice'' periods where message delays are $\delta^* \ll \delta$,
reads take at most $\epsilon$ time,
while the RMWs issued by the leader take
at most $3 \delta^*$.
Our parametrized algorithm generalizes the (non-parameterized) lease-based algorithm of~\cite{CHT16}
(henceforth referred to as the ``CHT algorithm'')
where the worst-case time for reads is~$3\delta$.
This generalization is achieved by adding two novel mechanisms,
each of which is controlled by a parameter.
Roughly speaking, the first mechanism
decreases the worst-case time for reads and
enables
a continuous trade-off between the worst-case times for read and RMW operations,
and the second mechanism
allows us to take advantage of ``nice'' periods when message delays are very short.
These mechanisms may be useful to achieve similar benefits in other lease-based algorithms.
We now describe our algorithm and the results in more detail.
To do so, we first explain our model, we then describe the CHT algorithm
and the two mechanisms that we added to generalize it, and finally we
compare the performance of the two algorithms for some parameter settings.
\textbf{Model sketch.} We consider message-passing systems where fewer than half of the processes may crash.\footnote{ If half of the processes or more crash,
it is impossible to implement even linearizable registers, let alone objects that support arbitrary RMW operations,
in our model of partial synchrony.
This is easy to show using a standard partitioning argument.}
Initially, processes take steps at arbitrary speeds and
messages take arbitrarily long and can even be lost.
There is, however, an \emph{unknown} time $\tau$ after which
no process crash occurs, processes take steps at some known minimum speed,
and every message that is sent is received
within some known time bound $\delta$~\cite{DLS88}.
To simplify the exposition, we assume that after time $\tau$
the time between consecutive steps of each nonfaulty process
is negligible compared to $\delta$.
We use the terms ``after the system stabilizes'' and ``stable period''
to refer to the time after~$\tau$.
When discussing the performance of an algorithm,
\emph{we focus exclusively on the period after the system stabilizes}.
The correctness of our algorithms, however, is always preserved: in particular safety is never violated and all operations issued by correct processes, even those issued before the system stabilizes, terminate.
Processes have local clocks
that are \emph{always} synchronized
within some known $\epsilon\ge0$ of each other;
such synchronized clocks can be provided
by devices such as GPS~\cite{CD+12}.
To simplify the exposition, we first assume here that $\epsilon=0$.
In Section~\ref{sec:epsilon} we explain how to deal
with an arbitrary clock skew $\epsilon > 0$, and how the
clock skew affects the performance of our algorithms.
\textbf{The CHT algorithm}.
This algorithm
has the following desirable properties.
Every read operation is ``local'';
furthermore, after the system stabilizes,
(a) every read operation is ``non-blocking''
unless it is concurrent with a RMW operation that conflicts with it, and
(b) even if a read blocks, it completes in a bounded period of time.
We say that read operations are \emph{local}
if they do not result in messages being sent;
more precisely, the number of messages sent during the execution of the algorithm
does not depend on the number of read operations performed in the execution.
A read operation issued by process $p$
is \emph{non-blocking}
if it completes
within a constant number of steps of $p$,
without waiting for a message to arrive
or for the process's clock to reach a certain value.\footnote{Because we do not assume a \emph{maximum} process speed, it is not possible to simulate waiting for a certain period of time by requiring the process to execute a minimum number of local steps.}
A read operation $r$ \emph{conflicts with} a RMW operation $w$
if there is an object state such that if we execute $r$ and $w$ starting from this state,
$r$ reads different values
depending on whether it executes before or after $w$.
Intuitively, the CHT algorithm works by combining two well-known mechanisms:
(a) a consensus algorithm to process all RMW operations, and
(b) a lease mechanism to allow local reads.
Both mechanisms rely on an eventual leader elector.
Roughly speaking, the (current) leader executes a ``two-phase commit'' algorithm
to linearize all RMW operations across the object replicas.
The leader also issues \emph{read leases}: the holder
of a read lease that expires at some time $t$ can read its local copy of the object until time $t$,
unless it is aware of a concurrent conflicting operation.
The blocking time of an operation is the time that elapses
from the moment a process issues this operation to the moment it completes it with a return value.
In the rest of this paper, we consider only the blocking time of RMW operations
\emph{issued by the leader} when it is not currently processing other RMW operations.
Note that if a RMW operation is not issued by the leader,
its blocking time may be longer by up to a round-trip delay
($2\delta$ in the worst-case, and at most $2 \delta^*$ in the ``nice'' periods):
this accounts for the time it takes for the issuer to send this operation to the leader
and to learn from the leader that this operation was committed.
We now explain why and for how long operations block in the CHT algorithm,
and we introduce the main ideas of our algorithm for decreasing
the blocking time of reads
with only a small or even no increase in the blocking time of RMW operations.
To see why operations may block with the CHT algorithm,
suppose a process $p$ has a read lease that expires far in the future,
but the leader $\ell$ wants to process a RMW operation that conflicts with the read.
To do so, $\ell$ first sends prepare messages to notify processes of the impending operation;
then, when $\ell$ receives ``enough'' acknowledgements, it commits the operation
(the state of the object is now changed);
finally $\ell$ sends commit messages to notify processes that the operation
was indeed committed.
Note that when $p$ receives the prepare message,
it does not know whether the state of the object already changed or not.
So if $p$ wants to do a read now, it cannot read its local copy of the object (because it could be stale):
it must wait until it gets the commit message from the leader.
Since
messages take at most~$\delta$,
it is clear that up to $3\delta$ time may elapse
from the moment $p$ receives the prepare message
to the moment $p$ receives the commit message;
during that period the read of $p$ is blocked.
The blocking time of a RMW operation issued by the leader $\ell$ is the time that elapses from
the moment $\ell$ starts processing the operation by sending prepare messages to
the moment $\ell$ commits it
having received enough acknowledgements.
This takes at most $2\delta$ time.
In summary,
with the CHT algorithm,
a read operation that is concurrent with a conflicting RMW operation
may block for up to $3 \delta$; and
a RMW operation issued by the leader~may block~for~up~to~$2 \delta$.
In this paper we introduce a parametrized algorithm
that can reduce the blocking time of reads without
affecting the maximum blocking time of RMW operations;
or can eliminate the blocking of reads altogether (more precisely,
reduces the blocking time to just $\epsilon$,
if clocks are not perfectly synchronized)
at the cost of slightly increasing
the maximum blocking time of RMW operations.
We do so by adding the two mechanisms described below.
\textbf{Two new mechanisms.} Our algorithm generalizes
the CHT algorithm by adding two mechanisms.
For pedagogical reasons we present our algorithm in two stages:
``Algorithm~1'' incorporates only one of the mechanisms,
and is parameterized by a quantity we denote $\alpha$.
The CHT algorithm is the special case of this algorithm
with $\alpha$ set to~0.
``Algorithm~2'' adds to Algorithm~1 the second mechanism,
and is parameterized by an additional quantity we denote $\beta$.
Algorithm~1 is the special case of Algorithm~2
with $\beta$ set to $\infty$.
\emph{Promise mechanism.} Roughly speaking,
the parameter $\alpha$ of Algorithm~1 is used as follows:
when the leader $\ell$ starts processing a RMW operation $\op$
at some time $t$,
it sends prepare messages for $op$
with the \emph{promise} not to commit $op$ before time $t+\alpha$,
the expiration time of that promise.
Now when a process $p$ (that has a valid lease) receives this message,
it knows that the state of the object
will not change before time $t+\alpha$,
so it can read its local copy up to that time.
We call this the \emph{promise mechanism}.
Process
$p$ will receive the commit message by time $t+3\delta$,
and so the reads of $p$ are blocked only during the period
$[t+\alpha, t+3\delta]$,
i.e., for up to $3\delta-\alpha$ time.
By setting $\alpha=3\delta$ we get an algorithm where
\emph{all} reads are non-blocking.
Note, however, that this setting also causes
all RMW operations issued by $\ell$ to block for $3\delta$ time.
Thus, with this setting of $\alpha$
Algorithm~1 achieves the desirable goal of non-blocking reads,
but at a considerable cost for RMW operations
in comparison to the CHT algorithm:
In the CHT algorithm a RMW operation blocks only for
the \emph{actual} delay of a round-trip message
while now \emph{all} RMW operations block for $3\delta$ time,
even if messages flow fast.
This is a problem because in many systems
the \emph{worst-case} message delay $\delta$
is orders of magnitude greater than the delay
experienced by most messages.
In particular,
there can be long periods of time after the system stabilizes
during which all messages take at most some $\delta^* \ll \delta$ time;
we call these \emph{nice periods}.
It is desirable to optimize the performance of algorithms
during such periods.
Here our goal is to decrease the maximum blocking time of reads without increasing
(or increasing only by little) the maximum blocking time of RMWs in the nice periods.
This is achieved by Algorithm~2, as we now explain.
\emph{Status mechanism.} The main idea behind Algorithm~2 is to keep the promises short,
and extend them as needed.
Instead of sending prepare messages with a long promise,
the leader $\ell$ sends ``status'' messages with a short promise~$\alpha$.
If $\ell$ does not receive enough acknowledgements
to commit an operation within a period $\beta$,
it sends another round of status messages with a new promise of~$\alpha$.\footnote{Note that it is possible for a promise to expire before the next one is received, and this may occur \emph{even in the stable period}. This is in contrast to the behaviour of read leases in the stable period.}
This is repeated until $\ell$ receives enough acknowledgements,
at which point it sends commit messages as before.
We call this the \emph{status mechanism}.
The cost of the status mechanism is the additional number of messages,
but if we set $\beta \ge 2\delta^*$
this cost is not incurred in nice periods,
because the leader receives enough acknowledgements
within $2\delta^*$ in these periods.
Thus we focus on the behaviour of Algorithm~2 only for settings of $\alpha$ and $\beta$ where $\beta \ge 2 \delta^*$.
With Algorithm~2,
we can set $\alpha$ (the length of the promise) to a small value
to reduce the blocking time of RMW operations in nice periods,
and with a suitable setting of $\beta$
(the time between successive status messages)
we can also keep the blocking time of reads short.
\textbf{Performance and comparison with CHT.} Tables~\ref{table1} and~\ref{table2} summarize
the maximum blocking times of operations
during the stable period and nice periods
under our two algorithms for certain interesting settings of their parameters $\alpha$ and $\beta$.
(The maximum blocking times of the two algorithms,
expressed as a function of $\alpha$ and $\beta$,
are given in Table~\ref{TableAlgo2Generic}.)
The column labeled ``CHT''
in both tables
shows the maximum blocking times of the CHT algorithm,
and serves as a baseline.
\input{table1.tex}
Table~\ref{table1} shows
parameter settings aimed at
\emph{improving} the blocking of reads
without increasing the blocking of RMW operations.
By setting $\alpha=2\delta$ in Algorithm~1
we reduce the blocking time of reads
to one-third of the CHT algorithm during the stable period,
and make all reads non-blocking during nice periods
(provided $\delta^*\le 2\delta/3$,
which holds because $\delta^*\ll\delta$).
This setting, however,
increases the maximum blocking of RMW operations during
nice periods from $2\delta^*$ to $2\delta$.
We can avoid this drawback by using Algorithm~2 with parameters $\alpha=\beta=2\delta^*$.
This decreases the maximum blocking time of reads
to one-third of the CHT algorithm
during both the stable period and during the nice periods,
without increasing the maximum blocking time of RMW operations
during either type of period,
and without incurring the overhead of additional status messages
during nice periods.
\input{table2.tex}
\vspace*{-6mm}
Table~\ref{table2} shows
parameter settings aimed at
\emph{eliminating} blocking of reads altogether,
even if at the cost of some increase in the blocking time of RMWs.
As we have seen in our earlier discussion,
by setting $\alpha = 3\delta$,
Algorithm~1 ensures that read operations never block;
but this setting increases the maximum
blocking time of RMW operations to $3\delta$
even during nice periods.
With a suitable choice of its two parameters,
Algorithm~2 can do better.
For example, by setting $\alpha = \beta = 3\delta^*$:
(1)~read operations block
for at most $\delta$,
and
(2)~reads never block during nice periods;
this is achieved at the cost of increasing
the maximum blocking time of RMW operations only by~$\delta^*$,
and only for nice periods.
Finally, the parameters can also be set so that
\emph{all} reads are non-blocking;
this is at the cost of an additional increase
of the maximum blocking time of RMW operations
by a single $\delta$
(see last column of Table~\ref{table2}).
\textbf{Roadmap.} In Section~\ref{sec-algo} we describe our algorithm and its performance under the simplifying assumption that $\epsilon = 0$, and we consider the case where $\epsilon \ge 0$ in Section~\ref{sec:epsilon}. In Section~\ref{sec:discussion},
we discuss our assumption of known message delays and the adaptiveness of the algorithm.
We briefly review some related work in Section~\ref{sec:related} and conclude the~paper~in~Section~\ref{sec:conclusion}.
\section{The algorithm}\label{sec-algo}
Algorithms~1 and~2 are described
in sufficient detail but informally in English
in Sections~\ref{subsec:promise} and~\ref{subsec:status}, respectively.
The pseudocode of Algorithms~1 and~2 are given in Figures~\ref{ObjectAlgo-alg1-code} and~\ref{ObjectAlgo-alg2-code}
(pages~\pageref{ObjectAlgo-alg1-code} and~\pageref{ObjectAlgo-alg2-code}), respectively.
Both algorithms use the same variables, so they are given only in Figure~\ref{ObjectAlgo-alg1-code}.
The code differences between Algorithm~1 and 2 are small and
are highlighted in {\color{blue} blue} in Figure~\ref{ObjectAlgo-alg2-code}.
Reading the detailed pseudocode may be skipped, but our English description of the algorithms has line references
to the pseudocode to help the reader who wishes
to follow it.
A complete proof of the correctness of Algorithm~1 is given in
Appendix~\ref{RO-Proof}.
\subsection{Eventual leader election}
Our algorithms use a leader election procedure $\Leader()$
with the following property:
there is a time after which every call to $\Leader()$ returns the \emph{same} correct process.
This procedure is the failure detector $\Omega$ \cite{CHT96}; it can be implemented efficiently in partially synchronous systems
(even without synchronized clocks)~\cite{ADFT08,ST08}.
Throughout the paper $\ell$ refers to this process.
Our algorithms also use the procedure $\AL(t,t')$,
which can be implemented from $\Leader()$ in our model~\cite{CHT16}.
Intuitively, $\AL(t,t')$ returns \textsc{True}
if and only if the process that invoked it has been the
leader \emph{continuously} during the entire time interval $[t, t']$;
$\AL(-,-)$ also ensures that no two distinct processes can consider
themselves to be leaders for two intersecting time intervals.
\begin{itemize}
\item
If the calls $\AL(t_1,t_2)$ and $\AL(t'_1,t'_2)$ by \emph{distinct} processes
both return \textsc{True},
then the time intervals $[t_1,t_2]$ and $[t'_1,t'_2]$ are disjoint.
\item
There is a time $t^*$ such that if
$\ell$ calls $\AL(t_1,t_2)$ at time $t \ge t_2\ge t_1\ge t^*$,
then this call returns \textsc{True},
and if a process
$q \neq \ell$ calls $\AL(t_1,t_2)$ with $t_2 \ge t^*$,
then this call returns \textsc{False}.
\end{itemize}
Our algorithms use the procedure $\ML{t_1}{t_2}$
to effectively divide time
into a sequence
of maximal non-overlapping intervals,
during each of which at most one process is continuously the leader,
and the last of which is infinite and
has a nonfaulty leader~$\ell$.
Intuitively, a leader has two functions:
(i)~it linearizes the RMW operations using a consensus mechanism, and
(ii)~it issues ``read leases'', which makes it possible
to execute read operations efficiently.
We now describe how each of these functions work in our two algorithms.
\subsection{Algorithm~1: The promise mechanism}\label{subsec:promise}
For the first function,
the leader collects into \emph{batches}
the RMW operations submitted by processes
(lines~\ref{rr1}--\ref{rr2}),\footnote{In this subsection line numbers refer to Figure~\ref{ObjectAlgo-alg1-code}.}
and it uses the two-phase commit protocol
outlined in the introduction as follows
(lines~\ref{nextops}--\ref{doops2failed} and
procedure $\DO$ in lines~\ref{acceptcheck1}--\ref{done},
called in line~\ref{second-doops}).
To commit a batch, the leader first attaches to the batch a sequence number $j$ and
a \emph{promise time} $t+\alpha$,
where $t$ is the current time and $\alpha$ is the
parameter of the algorithm
(line~\ref{set-promisetime}).
Intuitively, the leader guarantees that this batch of operations ``will not take effect'' before the promise time $t+\alpha$.
The leader then sends prepare messages to notify processes of batch~$j$
(line~\ref{prep-send}).
When a process receives this message we say that
it \emph{becomes aware} of batch $j$,
and it responds with an acknowledgment
(lines~\ref{a3b}--\ref{sendPack}).
When the leader receives enough acknowledgements,
it commits this batch $j$ and sends commit messages to all processes
(lines~\ref{prep-condition}--\ref{sendcommit}).
Note that when a batch is committed, it does \emph{not} mean that the operations in this batch have taken effect:
the algorithm ensures that these operations are not visible to users
(and in particular they do not return) before the batch's promise time.
Roughly speaking, a batch of RMW operations takes effect when it has been committed \emph{and} its promise time has been reached.
Each process applies to its local replica the committed batches in sequence, and
applies the operations of each batch in some pre-determined order,
the same for all processes
(procedure $\EB$, lines~\ref{EB-start}--\ref{EB-end}).
When a process applies one of its own RMW operations to its replica,
it determines the response of that operation,
\emph{and then it waits until the promise time
of the batch containing
that operation before
returning this response}
(lines~\ref{rmw-until}--\ref{rmw-return}).
Since all processes apply the same sequence of RMW operations in the same order
(which is consistent with the order of non-concurrent operations)
the execution of RMW operations is linearizable.
The second function of the leader is to periodically issue read leases to allow processes to read locally, as we now explain.
Recall that the leader starts processing batch $j$ at some time $t$ and commits this batch with promise time
$t+\alpha$.
After committing batch $j$,
the leader issues the read lease $(j,s)$ with $s = t+\alpha$
by sending a lease message to all processes;
this message is combined with the commit message
(line~\ref{sendcommit}).
We say that the read lease $(j,s)$ \emph{starts} at time $s$ and
\emph{expires} at time $s+\lambda$,
where $\lambda$ is the \emph{lease period};
we also say that the lease $(j,s)$ is \emph{valid at time $t'$} if $t' < s+\lambda$.
At some time $s'$ before the read lease $(j,s)$ expires,
the leader renews the lease by issuing the lease $(j,s')$.
Such lease renewals for batch $j$ occur periodically
until the leader commits batch $j+1$
(line~\ref{sendcommit2} within the main loop of the $\LW$ procedure,
lines~\ref{mainwhile}--\ref{doops2failed}).\footnote{The lease period $\lambda$
and the frequency of lease renewals are chosen so that after the system stabilizes
all the correct processes always have valid leases.}
Note that when the leader issues
the \emph{first} read lease $(j,s)$ for batch~$j$
(line~\ref{sendcommit}),
the start time $s=t+\alpha$ of this lease \emph{can be in the future},
but whenever the leader issues
a lease \emph{renewal} $(j,s')$ for batch $j$
(line~\ref{sendcommit2}),
the start time $s'$ is when this lease is issued.
We now explain the semantics of read leases,
and how they are used by processes to read from their local replicas.
If a process $p$ has a valid lease $(k^*,t^*)$ at time $t'$ then the following two lease properties hold:
\begin{enumerate}
\item
No batch $j > k^*$ takes effect before time $t^*$.
This property is ensured as follows.
If $(k^*,t^*)$ is the \emph{first} read lease
that the leader issued for batch $k^*$ (line~\ref{sendcommit}),
then the leader ``promised'' that batch $k^*$ will not take effect before time $t^*$
(and the algorithm ensures this promise is kept);
this implies that no batch $j > k^*$ takes effect before time $t^*$.
If $(k^*,t^*)$ is a read lease renewal (line~\ref{sendcommit2}),
then when the leader issues it at time $t^*$
it has not yet committed any batch $j > k^*$.
\item
No batch $j >k^*$ takes effect during the interval $[t^*,t^* +\lambda)$
before $p$ is aware of batch $j$.
Intuitively, this property is ensured as follows.
The leader keeps track of the processes that may hold a valid read lease on
the last batch it committed
(these are the $\LeaseHolders$);
before the leader commits a new batch $j$
it waits until all the $\LeaseHolders$
acknowledge
the prepare messages for
this batch (so they are now aware of batch $j$);
if some of them do not acknowledge batch $j$
then the leader waits until time $t^* +\lambda$, i.e., until all read leases expire
(lines~\ref{wait2}--\ref{lh2})
before committing the new batch $j$.
\end{enumerate}
Now suppose that a process $p$ wants to read the object at some time $t'$
(lines~\ref{read-invoke}--\ref{read-respond}).
To do so, intuitively $p$ needs to determine the maximum number $\hat{k}$
such that batch $\hat{k}$
took
effect by time $t'$:
$p$ can then read the state of the object after batch $\hat{k}$,
i.e., after applying all the operations in batches $0$ to $\hat{k}$ to its local replica.
If $p$ holds a valid lease $(k^*,t^*)$ at the time $t'$ when it wants to read, it can determine this $\hat{k}$ by using
the lease properties and the promise mechanism as follows:
\begin{case}
\item $t' < t^*$.
By the first lease property, only batches with
sequence number at most $k^*$ can take effect by time $t' < t^*$.
By the promise mechanism, only batches with a promise time at most $t'$ can take effect by time $t'$.
Process $p$ determines the maximum batch number $\hat{k}$ such that $\hat{k} \le k^*$
\emph{and} the promise time of batch $\hat{k}$ is at most~$t'$.
Note that batch $\hat{k}$ took effect by time $t'$: this is because it was committed by time $t'$\footnote{Since no leader can issue the lease $(k^*,-)$ before batches $0$, $1$, $2, \ldots , \hat{k} , \ldots , k^*$ have been committed.}
\emph{and} the promise time of batch $\hat{k}$ is at most $t'$.
Thus $\hat{k}$ is the maximum batch number such that batch $\hat{k}$ took effect by time~$t'$.
Our algorithm ensures that
because $p$ holds a lease $(k^*,t^*)$ at time $t'$,
it has already received all the batches up to and including $k^*$
by time $t'$.
After determining $\hat{k}$, process $p$ just reads the state of the object after batch $\hat{k}$
at time $t'$ without any waiting.
\item $t' \ge t^*$.
First note that batch $k^*$ took effect by time $t'$:
this is because $k^*$ was committed by time $t^* \le t'$
\emph{and} the promise time of batch $k^*$ is at most $t^* \le t'$.
Thus $\hat{k} \ge k^*$.
Since the lease $(k^*,t^*)$ is valid at time $t'$, we have $t^* \le t' < t^* +\lambda$.
By the second lease property, the only batches with sequence number $j > k^*$ that can take effect by time $t'$ are those
that $p$ is aware of at time $t'$.
By the promise mechanism, the only batches that can take effect by time~$t'$
are those with a promise time at most $t'$.
Process $p$ determines the set $B$ of batches
with sequence numbers $j > k^*$ such that:
(a) $p$ is aware of batch $j$ at time~$t'$, and
(b) the promise time of batch $j$ is at most $t'$.
From the above, $B$ consists of \emph{all} the batches with a sequence number greater than $k^*$ that could have taken effect by time~$t'$.
Thus, process $p$ can now compute
$\hat{k}$ to be the maximum batch number in $B$ if $B$ is not empty,
and $\hat{k} = k^*$ otherwise.
From the above, $\hat{k}$ is the maximum number such that
batch $\hat{k}$ could have taken effect by time~$t'$.
After computing $\hat{k}$, process $p$ first waits until it has all batches up to $\hat{k}$
and until the promise time of batch $\hat{k}$ has passed.\footnote{The promise time of batch $\hat{k}$ can change (and increase) since the time $p$ determined the set $B$ if and only if the leader trying to commit batch $\hat{k}$ changes. As an optimization, it turns out that waiting for the promise time of $\hat{k}$ to pass is not necessary!}
It then reads the state of the object after batch $\hat{k}$.\footnote{Like the CHT algorithm, our algorithm
incorporates
a further optimization that ensures no read blocks
unless it is concurrent with a conflicting RMW operation:
to determine $\hat{k}$, $p$ eliminates from the set $B$ every batch that contains only RMW operations that do \emph{not}
conflict with its read operation.
It can do so because the operations in these batches do not affect the value that it reads.}
\end{case}
Having explained how the read operations work
with the new semantics of read leases under the promise mechanism,
we now point out a subtelty with how promise times
must be handled when a new leader takes over.
Note that the leaders must ensure that,
\emph{even across leadership changes},
all nonfaulty processes agree
on the same sequence of batches,
and
that each RMW operation is included
in exactly one batch.
To do so, the first thing that a new leader does
is to wait long enough for all leases issued by previous leaders
to expire (line~\ref {wait-lease-expire}).
It then commits or recommits the last batch $j$ that the previous leader
attempted to commit but may have left half-done
(lines~\ref{est-request}--\ref{first-doops}).
The new leader should not give a future promise time to batch $j$
because doing so would allow processes
to read the state of the object
before the operations of batch~$j$ have been applied to it,
even though batch $j$ could have already taken effect
under the previous leader.
So, to be safe, the new leader uses the promise time~0 for batch $j$;
effectively giving no promise for batch~$j$
(line~\ref{first-doops}).
\textbf{Maximum blocking time analysis.}
The column of Table \ref{TableAlgo2Generic}
labeled ``Algorithm~1''
gives the maximum blocking times of RMW and read operations
during the stable period (where all messages take at most $\delta$) and during nice periods (where all messages take at most $\delta^* \ll \delta$) for arbitrary values of $\alpha \le 3 \delta$.
Setting $\alpha > 3 \delta$ only increases the blocking of RMW
operations without any benefit for the reads.
We now justify the entries of that column.
Consider the system in the stable period.
Suppose that a process $p$ wants to read at time~$t'$ and
holds a valid lease $(k^*,t^*)$ at time $t'$.
If $t' < t^*$, then by Case 1 above this read does not block.
If $t' \ge t^*$, then by Case 2 above the read may block because $p$ waits until it knows all batches up to $\hat{k}$
and until the promise time of batch $\hat{k}$ has passed.
If $\hat{k} = k^*$ then the read does not block since these two conditions are already met by time $t'$:
this is because $p$ has the read lease $(k^*,t^*)$ at time $t'$.
Now assume that $\hat{k} > k^*$, so $\hat{k} \in B$.
Let $t$ be the time when the leader sent the prepare messages for batch $\hat{k}$;
so the promise time of batch $\hat{k}$ is $t +\alpha$.
Since batch $\hat{k}$ is in the set $B$, $p$ is aware of batch $\hat{k}$ and the promise time of $\hat{k}$
is at most $t'$, i.e., $t +\alpha \le t'$.
Because the system is in the stable period, $p$ will receive all batches up to $\hat{k}$ by time $t + 3\delta$.
So $p$ blocks from time $t' \ge t +\alpha$ to at most time $t + 3\delta$, i.e., for at most $3\delta - \alpha$.
Now suppose the leader wants to issue a RMW operation at time $t$.
To process this operation, the leader
waits for acknowledgments for the batch that contains the RMW operation; this will be done by time $t+2\delta$.
It must also wait until the promise time $t +\alpha$ before it returns the response to the RMW operation.
So the RMW completes by time $\max(t+2\delta , t +\alpha)$, i.e., it blocks for $\max(2\delta , \alpha)$
The analysis for the nice periods is similar.
\subsection{Algorithm~2: The status mechanism}\label{subsec:status}
Recall that in Algorithm~1 each batch $j$ has a promise time,
which is a lower bound on the time when the batch takes effect.
In Algorithm~2, a batch does not have a fixed promise time but
a sequence of increasing promise times,
and thus a sequence of increasing lower bounds
on the time when it takes effect.
To accomplish this,
when the leader wants to commit a new batch $j$
it does not send prepare messages that notify processes of
the batch~$j$ and its associated promise time, as in Algorithm~1.
Instead, every $\beta$ time units the leader sends
a new round of so-called \emph{status} messages for batch~$j$
with promise time $t+\alpha$,
where $t$ is the time when this round of status messages is sent
(lines~\ref{algo2-dop-repeat}--\ref{algo2-dop-sendstatus-d}).\footnote{In this subsection line numbers refer to Figure~\ref{ObjectAlgo-alg2-code}.}
The leader stops sending status messages for batch~$j$
as soon as it receives enough acknowledgements
(line~\ref{algo2-dop-condition}).
It then sends commit messages for batch~$j$ to all processes,
just as in Algorithm~1.
By choosing the parameter $\beta \ge 2 \delta^*$,
in nice periods only one round of status messages is sent per batch.
This round replaces the prepare messages of Algorithm~1,
and so the algorithm does not incur extra messages during nice periods.
In fact, with such a $\beta$, Algorithm~2 behaves exactly as Algorithm~1 during nice periods.
The leader also sends read leases:
The \emph{first} lease $(j,s)$ for batch $j$
is sent alongside the commit message for that batch
with a start time
equal to
the promise time of the \emph{last} status message for batch $j$
that the leader sent --- i.e., a time that could be in the future
(line~\ref{algo2-dop-sl2}).
As in Algorithm~1, the start time of each lease \emph{renewal} for batch $j$ is the time when it is sent
(lines~\ref{algo2-taketime1}--\ref{algo2-sl1}).
Read leases have the same two properties as in Algorithm~1.
A subtlety that concerns the initialization of a new leader
is worth pointing out.
As with Algorithm~1, the new leader first commits or recommits the last batch $j$ that the previous leader
attempted to commit but may have left half-done,
and to be safe the new leader uses the promise time 0 for batch $j$.
So Algorithm~2 uses the exact same procedure as Algorithm~1
to commit batch $j$ during its initialization
(see procedure $\DO$).
To commit subsequent batches, Algorithm~2
uses the procedure described above,
which sends successive rounds of status messages
with increasing promise times
(see procedure $\DO'$ in Figure~\ref{ObjectAlgo-alg2-code}).
\textbf{Maximum blocking time analysis.}
We now analyse the maximum blocking time of reads after the system stabilizes.
This
analysis
also shows how the ``status mechanism'' unblocks certain read operations
that would remain blocked for a longer period under Algorithm~1.
Suppose that a process $p$ holding a valid lease $(k^*,t^*)$ at time $t'$
wishes to perform a read at time $t'$ and is blocked.
As with Algorithm~1, this blocking can occur only in Case~2,
i.e., when $t' \ge t^*$
and the read is blocked
because $p$ is aware of a batch $j>k^*$
that has promise time at most $t'$.
Under Algorithm~1, such a read will remain blocked until
$p$ has all batches up to $j$
which may take $3\delta - \alpha$
(see the first column of Table~\ref{TableAlgo2Generic}).
Consider now the same scenario under Algorithm~2.
Every $\beta$ units of time
the leader sends a status message (with a new promise) for batch $j$,
or it has already sent a commit message for batch $j$.
If it sends a status message after time $t'-\alpha$,
the associated promise time is greater than $t'$.
So by time $t'-\alpha+\beta$
the leader sends a status message with a promise time greater than $t'$,
or it has already sent a commit message, for batch $j$.
Process $p$ receives that message by time $t'-\alpha+\beta+\delta$,
and this unblocks the read:
if it is a status message with a promise time greater than $t'$, then $p$ can read before batch $j$;
if it is a commit message, $p$ can read after batch $j$.
Therefore, under Algorithm~2
$p$'s read operation is blocked only
during the interval $[t',t'-\alpha+\beta+\delta]$,
i.e., for at most $\delta+\beta-\alpha$ units of time.
\input{Table4}
\vspace*{-6mm}
For the analysis of the maximum blocking time of RMW operations,
it is convenient to assume that $\beta$ divides $2 \delta$.
Suppose the leader wants to issue a RMW operation at time $t$.
Before it returns the response to this RMW operation,
the leader waits for acknowledgments
for the batch that contains the RMW operation;
this will be done by time $t+2\delta$.
It must also wait until the promise time of the \emph{last} status message
that it sent for that batch;
since $\beta$ divides $2\delta$,
that sending occurs by time $t+2\delta - \beta$,
and so the promise time of that status message is
at most $t+2\delta - \beta +\alpha$.
So the RMW completes by time $\max(t+2\delta, t+2\delta - \beta +\alpha)$,
i.e., it blocks for $\max(2\delta, 2\delta - \beta +\alpha)$.
Since we assume that $\beta \ge 2\delta^*$,
and in this case
Algorithm~2 behaves exactly as Algorithm~1 during nice periods,
the blocking times during these periods are the same as in Algorithm~1.
The maximum blocking times with Algorithm~2 are shown
in the second column of Table~\ref{TableAlgo2Generic}.
\section{Approximately Synchronized Clocks}\label{sec:epsilon}
Recall that in our model all local clocks are always synchronized within $\epsilon$ with each other.
To simplify the presentation,
so far we have been assuming that $\epsilon=0$.
In this section we explain how to modify our algorithms so that they work
even when local clocks are not perfectly synchronized,
i.e., when $\epsilon > 0$, and give their performance.
We refer to the values of local (process) clocks as \emph{local time}
to distinguish it from \emph{real time}.
The main challenge when $\epsilon>0$ is that
processes may not agree whether, at some real time, a batch has taken effect yet, and
they may execute operations that violate linearizability.
For example, suppose that at every real time
the clock of process $p^-$ shows local time $\epsilon/2$ less than real time
while the clock of process $p^+$ shows local time $\epsilon/2$ more than real time.
Suppose now that batch $j$ has promise time $s$.
At real time~$s$,
when the clock of $p^+$ shows $s+\epsilon/2>s$,
$p^+$ reads the state of the object after batch $j$.
At the later real time $s+\epsilon/4$,
when the clock of $p^-$ shows $s+\epsilon/4-\epsilon/2<s$,
$p^-$ reads the state of the object before batch $j$.
This violates linearizability.
We address this problem in the same way in both Algorithms 1 and 2 as follows.
Whenever a process $p$ waits for the promise time $s$ of some batch $j$ to expire,
we require $p$ to wait for an extra $\epsilon$,
i.e., until its clock reaches $s+\epsilon$.
Thus, if a process $p$ wants
to read the state of the object after batch~$j$
(line~\ref{wait-promise-2})
or to return the response from a RMW operation contained in batch $j$
(line~\ref{rmw-wait-promise}),
$p$ now waits until its clock shows time $s+\epsilon$.
(Throughout this section,
line numbers refer to the pseudocode of Algorithm~1.)
Perhaps surprisingly, the computation of $\hat{k}$
(lines~\ref{get-k-hat-then} and~\ref{get-k-hat-else}--\ref{get-k-hat-5})
does not change when $\epsilon>0$.
To see this suppose that
process $p$ wishes to perform a read operation
at real time $\tau$ and local time $t'$, and
$p$ is aware of a batch $j$ with promise time $s > t'$.
At real time $\tau$,
the local clock of every process is at most $t' +\epsilon$.
Since $t'+\epsilon < s+\epsilon$,
and each process $q$ waits until its local clock is
at least $s+\epsilon$ before the promise of batch $j$ expires at $q$,
by real time $\tau$
no process could have read the state of the object
after the operations of batch~$j$ have been applied,
and no process could have returned the response from a RMW operation
contained in batch~$j$.
So at real time $\tau$,
$p$ can safely read the state of the object
before the operations of batch~$j$ are applied,
without violating linearizability.
This shows that process $p$ can compute $\hat k$
in the same way as with $\epsilon = 0$,
i.e., by considering only the batches~$j$ with promise $s \le t'$
(as opposed to those with $s \le t' +\epsilon$).
To retain the property that $p$'s read does not block if there are no conflicting concurrent RMW operations,
$p$ actually considers only the batches~$j$ with promise $s \le t'$ that
contain RMW operations that conflict with $p$'s read.
(This is already done when computing $\hat{k}$ in lines \ref{get-k-hat-else}--\ref{get-k-hat-5},
and the same must be done now also in line~\ref{get-k-hat-then}.)
\input{EpsilonTable}
There is a similar problem, and a similar solution,
with the lease mechanism when $\epsilon > 0$.
To see the problem
suppose all processes except $p^-$ (a process that is not the leader)
have clocks that show real time,
and process $p^-$ has a clock that shows $\epsilon$ less than real time.
Suppose that $p^-$ holds a lease $(j,t_j)$,
and the leader that issued that lease
wishes to commit a new batch $j+1$
with a promise time of $t_j+\lambda-\epsilon$.
If $p^-$ does not receive the prepare message for batch~$j+1$
(and therefore does not send an acknowledgement to the leader),
the leader waits until the lease $(j,t_j)$ expires at real time $t_j+\lambda$.
At that real time the leader commits batch~$j+1$,
issues a lease for that batch,
and reads the state of the object after batch~$j+1$.
The lease $(j,t_j)$ that $p^-$ holds
is valid at $p^-$ until local time $t_j+\lambda$,
i.e., until real time $t_j+\lambda+\epsilon$.
So, $p^-$ can read the state of the object before batch~$j+1$
during the real time interval $(t_j+\lambda,t_j+\lambda+\epsilon)$,
which follows the time when
the leader has read the state of the object after batch~$j+1$.
This violates linearizability.
The solution to this problem is similar to
the solution for the corresponding problem with promises:
Whenever
the leader waits for a lease $(j,t_j)$ to expire
(lines~\ref{wait-lease-expire} and~\ref{wait-alg1}),
we require it to wait for an extra $\epsilon$,
i.e., until its clock reaches $t_j+\lambda+\epsilon$.
This implies that when the leader stops waiting,
the lease $(j,t_j)$ has expired at all processes
and thus it cannot be used to read.
With the above modifications to handle the case that $\epsilon\ge0$,
the worst-case blocking times of our algorithms are shown in Table~\ref{TableWithEpsilon}.
As shown in this table, the maximum blocking times of RMW and read operations increase by at most $\epsilon$ compared to
the special case that $\epsilon =0$.
As with~\cite{CHT16}, however, with our algorithms every read operation that does not conflict
with a concurrent RMW operation remains non-blocking.
From Table~\ref{TableWithEpsilon} it is clear that we can set the algorithms' parameters
so that the maximum blocking time for read operations is $\epsilon$;
for example, $\epsilon$ is achieved by setting $\alpha=3\delta$ in Algorithm~1 or
$\alpha=\delta+\beta$ in Algorithm~2.
If $\epsilon\le\delta/2$, this nearly matches a lower bound by Chandra \emph{et al.} (Theorem~4.1 in ~\cite{CHT16}).
Note that $\epsilon \le \delta/2$ holds in geo-distributed systems
where, with present technology, clock skew can be under 10msec~\cite{CD+12}
and message delays (say between data centres located in different continents)
can be in the order of~100msec~or~more~\cite{KP+12}.
\section{Discussion}\label{sec:discussion}
\textbf{Knowing $\delta$ and $\delta^*$.}
Recall that our algorithms use two message delay estimates:
$\delta$ (the maximum message delay after the system stabilizes)
and $\delta^*$ (the maximum message delay during nice periods).
The reader may wonder whether it is reasonable to assume that
$\delta$ and $\delta^*$ are known,
and what happens if their assumed values are incorrect.
We first note that the assumption of a known $\delta$ is made routinely.
For example, distributed algorithms that use timeouts on remote machines
(say for detecting whether they are still alive)
include an estimate of $\delta$ to determine the timeout period.
Also, many practical \emph{lease-based} distributed algorithms (e.g.,~\cite{B06})
also use a known $\delta$ to calculate the length~of~the~lease.
What is the effect of assuming the wrong $\delta$?
In our algorithms, safety does not depend on having a correct estimate on $\delta$;
it is always preserved.
\emph{Underestimating} $\delta$ can affect liveness:
during ``bad'' periods where some messages take more than $\delta$
it is possible that no progress is made.
\emph{Overestimating} $\delta$ may increase worst-case blocking times.
What is the effect of assuming the wrong $\delta^*$?
It turns out that neither safety \emph{nor} liveness depends
on having a correct estimate on $\delta^*$.
The only consequence of \emph{underestimating} $\delta^*$ is
that nice periods would be less frequent and shorter,
so the maximum blocking times that we achieve for nice periods
would be less useful.
The consequence of \emph{overestimating} $\delta^*$ is
a possible increase in the worst-case blocking times.
But since safety and liveness do not depend on the choice of $\delta^*$,
one can easily readjust the estimate of $\delta^*$ dynamically
to match the ``current'' state of the system.
\textbf{Adaptiveness.}
Related to the question of the algorithm making use of $\delta$ and $\delta^*$
is the property of ``adaptiveness'', in the following sense:
One of the advantages of the (completely) asynchronous model is that,
because there are no known bounds on message delays,
algorithms designed to work in that model
tend to adapt to the actual operating conditions without
making worst-case assumptions:
if messages flow fast, such algorithms are correspondingly fast;
if messages slow down, so does the algorithm.
This is a desirable property because, in practice, operating conditions
are often favourable.
Unfortunately there are limits to implementing fault-tolerant objects
in completely asynchronous systems;
in particular,
it is not possible to implement objects with arbitrary RMW operations as we do here~\cite{FLP83,Herlihy91}.
Note that in our algorithm \emph{all the read operations are adaptive},
regardless of the parameter settings.
For RMW operations, our algorithm exhibits the flexibility of
trading off their adaptivity with the worst-case blocking time of reads:
if we set the parameter $\alpha$ to 0
(i.e., the special case that is the CHT algorithm),
the RMW operations are also adaptive;
but in that case the (adaptive) reads may block for up to $3\delta$ time.
If, on the other hand, we prefer to optimize reads,
we can set the parameters to reduce their worst-case blocking time
at the cost of decreasing the adaptivity of the RMWs.
The best parameter setting for this trade-off depends
on the relative frequency of read and RMW operations and
on what one wants to achieve.
An advantage of our algorithm is that it allows for parameter settings
that best fit different operating conditions and user objectives.
\section{Related work}\label{sec:related}
\textbf{Lower bounds.}
Attiya and Welch have shown some lower bounds on the time
to read and write for linearizable implementations of registers~\cite{AW94}.
These bounds apply to systems where processes
have clocks that run at the same rate as real time
and \emph{all} the message delays are in the range $[\delta-u,\delta]$ for some known $\delta$ and message uncertainty $u$,
where $0 \le u \le \delta$.
For $u = 0$, they prove that
the \emph{sum} of the times to do a read and a write operation is
at least $\delta$ (Theorem 4.1 in~\cite{AW94}).
For $u > 0$, they prove that
a read operation requires at least $u/4$ time
and a write operation requires at least $u/2$ time
(Theorems 3.1 and 3.2 in~\cite{AW94}).
These bounds do not apply to the algorithms that we presented here
because our model is incomparable
to the model in~\cite{AW94}.
On one hand, our model is weaker because the maximum message delay applies only to messages sent after (an unknown)
stabilization time.
On the other hand, it is also stronger because we assume that processes are equipped with external clocks
that are synchronized within some $\epsilon \ge 0$.
In our model, after stabilization time we have $u = \delta$.
Note that for some parameter settings,
reads in our algorithm take at most $\epsilon$ time which could be less than the $u/4$ lower bound of~\cite{AW94}
\emph{if} the clocks are highly synchronized (e.g., via special devices such as atomic clocks and GPS signals,
such as in the Spanner system~\cite{CD+12},
or via special high priority messages).
This demonstrates a benefit of adding highly synchronized external clocks to partially synchronous systems.
\textbf{Algorithms.}
Replication is used extensively in distributed systems
ranging from synchronous, tightly coupled ones,
to asynchronous, geographically dispersed ones.
Below we highlight the main points of some replication algorithms that are most closely related~to~our~work.
Megastore~\cite{BB+11} is an early Google system designed
to support distributed transaction processing with efficient reads.
Megastore implements a replicated log
that can be written (by appending entries to it) and read.
Write operations are linearized using
a version of the Paxos algorithm~\cite{LL98,LL01paxos},
and read operations are local and non-blocking
when there are no concurrent write operations.
To write the log
Megastore requires the leader to receive acknowledgements
from \emph{all} processes, or
for crashed or disconnected processes to time out.
Thus, a process that crashes or becomes disconnected
delays \emph{all} write operations issued while it is unresponsive.
In contrast, in our algorithms
the leader keeps track of the current leaseholders, i.e.,
the processes that acknowledged the last RMW operation,
and in subsequent RMW operations it waits for acknowledgements
only from them: so a process that crashes can
delay at most one write operation.
As noted in~\cite{BB+11}, an asymmetric network partition
can cause write operations to block indefinitely
because of Megastore's reliance on the Chubby lock service
(another Google system~\cite{B06}) for failure detection,
a problem that requires operator intervention to resolve.
Paxos Quorum Leases (PQL)~\cite{MAK14} is an algorithm that addresses
the above-mentioned problems with Megastore.
Similar to our algorithms,
in PQL the leader keeps track of the current leaseholders
and waits for acknowledgements to RMW operations only from them.
Lease renewals, however, are more expensive in PQL than
in our algorithms:
Leases are granted not by the leader but by a majority of processes
called ``lease grantors''.
Each lease renewal requires a quadratic number of messages
in the number of participating processes
(compared to linear, in our algorithms),
and two message delays
(compared to one, in our algorithm).
Furthermore, in PQL each change in the set of leaseholders triggers
the use of a consensus algorithm (specifically of Paxos)
among the lease grantors,
whereas in our algorithm the leader manages this set on its own
simply by noting the processes that acknowledge the last RMW operation.
Finally, in PQL a RMW operation revokes the current leases,
and so a steady stream of RMW operations can disable
local reads for arbitrarily long.
In our algorithms, all reads are local and block only for a bounded time.
Spanner~\cite{CD+12} is another Google system that,
like its predecessor Megastore,
supports distributed transactions and implements replicated objects.
Spanner is the first system we know of that uses the model
we adopted in our paper:
a partially asynchronous message-passing system
equipped with accurately synchronized clocks.
Spanner uses Google's TrueTime service, which maintains synchronized clocks,
to attach timestamps to read and write operations,
and executes these operations in timestamp order
at each of the processes that manage a replicated object.
Thus, to execute a read operation with timestamp $t$,
a process must know
the write operation with the maximum timestamp $t'$ such that $t'<t$.
A process cannot determine this locally
unless it blocks until it receives a write operation
with timestamp $t''>t$.
Thus a read operation either
must involve communication with other processes and
is therefore not local, or
it may block indefinitely to wait for a write with a higher timestamp, or
it may risk reading a stale value.
Hermes~\cite{KG+20} is a more recent system that supports replicated objects,
designed with the express purpose of reducing the latency of operations.
To achieve this, Hermes allows any process to initiate a RMW operation,
rather than channeling all such operations through the leader,
as in our algorithms.
By doing so, RMW operations that are not issued by the leader
save the round-trip delay of being sent to the leader and
receiving the commit message.
To also achieve local reads, Hermes requires all processes to acknowledge
each RMW operation, like Megastore.
If some process does not do so in a timely manner,
a relatively expensive reconfiguration operation is triggered
for a majority of processes to agree
on the new set of processes that manage the replicated object.
This is done using a variant of Paxos called Vertical Paxos~\cite{LMZ09}.
In contrast, our algorithms weather permanent or transient
disconnections of processes from the leader
using the more lightweight leaseholder mechanism.
As noted in~\cite{KG+20}, due to the lack of coordination by a leader,
concurrent RMW operations in Hermes may abort,
and thus they do not have a bounded blocking time.
Finally, as in PQL, a steady stream of write operations
can disable local reads for arbitrarily~long.
\section{Conclusion}\label{sec:conclusion}
We presented a parameterized algorithm that works
in partially synchronous systems where processes are equipped
with clocks that are synchronized within $\epsilon$.
This algorithm generalizes the (non-parameterized) CHT algorithm,
and for some settings of its parameters
it ensures that no read takes more than $\epsilon$ time
\emph{even in the presence of concurrent conflicting operations}.
A novel feature of our algorithm is that its parameters can be used for two benefits:
They enable a continuous trade-off
between the maximum blocking times of read and RMW operations,
and they can be used to reduce these blocking times during ``nice'' periods where messages delays are smaller than the maximum message delay.
This is achieved by leveraging two new ideas,
the promise mechanism and the status mechanism,
which modify the semantics of leases.
Leases are used in a variety of settings in distributed computing,
and we believe that our promise and status mechanisms can be used
to achieve similar benefits in other lease-based algorithms.
\vfill
\input{algo1-code-small}\label{algo1page}
\input{algo2-code-small}\label{algo2page}
\newpage
|
\section{Introduction}
Digital transformation trends such as Artificial Intelligence (AI), Digital Twins and internet of things plays an increasing and integral role in driving innovation and becoming ubiquitous in various domains and applications such as intelligent driving \cite{Khan19, Thevendran21}. These trends enable smart systems with novel capabilities that were never possible without AI. However, with the increasing adoption of these complex AI driven systems, new challenges emerge, especially related to the acceptance and trustworthiness of these systems by human users. Hence, there are increasing voices that demand more transparent and explainable AI models and systems.
AI engines for predicting lane changes can be implemented using white-box models that come with the advantage of being explainable by default. Alternatively, black-box classifiers currently outperform white-box models on the lane change prediction task but struggle as they are not explainable \cite{xing19}. For mission critical tasks that involve human, such as predicting lane changes, both explainability and performance are crucial and equally important.
This paper explores how the decisions making process of a complex black-box classifier can be made explicit and explained to a user. We present a state-of-the-art approach for lane change predictions that are explainable and a novel technical proof-of-concept implementation.
\section{Related Work}
Predicting lane changes for vehicles is an ongoing field of research. In their paper, Xing et al. \cite{xing19} presented a general discussion and a survey of the latest technology trends around this topic.
The literature suggests various approaches to predict lane changing behaviour that mainly varying in the data and AI architecture used. Chen et al. \cite{chen19} aimed to train an attention-based deep reinforcement agent based on visual data in a simulated environment that predicts lane changes. Another approach, by Tang et al. \cite{Tang19}, uses tabular data to train an adaptive fuzzy neural network to predict if a lane change takes place soon. Furthermore, a popular machine learning architecture for this type of problem is the recurrent neural network, as it is optimized to deal with problems related to time sequence analyses \cite{Patel18}.
With the increasing attention on trustworthiness and transparency of machine learning models and systems, we are seeing a focus in recent literature on explainable models for the lane change prediction task, to be able to explain the reasons behind a predicted vehicle lane change for a human user or road stakeholder. The main goal is to move away from predicting with black-box models and aiming to increase the performance of white-box models like expert systems and other explainable classifiers \cite{Gallitz19, Gallitz20}. One paper by Dank et al. reformulated the prediction task based on tabular data to a regression problem \cite{furnkranz17}.
While white-box models come with the advantage of being explainable \cite{schwalbe21}, they are outperformed on the lane change prediction task by black-box classifiers \cite{xing19}. Nonetheless, the latter are not explainable. For safety critical tasks, both explainability and performance are crucial and equally important.
AI systems cannot be implemented without reliable data resources. Advancement in the area of Internet of Things (IoT) and Digital Twins within the automotive area, especially around autonomous and intelligent driving, can be seen in recent literature and successfully deployed projects and systems \cite{steyn2021development, kumar2018novel, el2020roads}.
Alongside the vehicles themselves, infrastructure, such as roads and highways have also undergone modernization in places, so that these elements also can relay their "state" back to operators in real time. A real example of this is the Providentia++ Digital Twin \cite{kraemmer2019providentia}, which covers a section of Autobahn between Munich and Munich Airport. Here, the Providentia++ team decided to use cameras placed at regular points along the road, combined with visual recognition to identify vehicles. The setup is capable of relaying the position of every vehicle on the track, with a high level of accuracy and frequent update cycles.
This paper shows an approach to combining data from the Providentia++ Digital Twin with an explainable machine learning model to predict lane changes in real-time.
The following section shall introduce this approach.
\section{Approach}
In this section, we present our suggested approach towards explainable lane change prediction supported by an extensible technical implementation.
\subsection{Lane Change Predictions by a Layer Normalized LSTM}
The lane change predictions are computed by a layer normalized long short-term memory proposed for this purpose by Patel et al. in 2018 \cite{Patel18}. This section shall introduce the prediction model and its input features regarding relevant perspectives for generating explanations of its predictions. Please consult Patel et al.'s paper \cite{Patel18} for further information and evaluation of the ML model.
First, a layer normalized LSTM \cite{ba16} considers at each time step \(t\) a 1-dimensional array of vehicles \([v^i_t| \ \forall i \in [0,1,2,3,4,5,q]]\). The vehicles \(v^i\) with \(i \in [0,1,2,3,4,5]\) are the closest existing neighbours of \(v^q\).
Each vehicle \(v^i\) at the time \(k\) is represented by the following array of features:
\begin{equation}
v^i_k = [v_{x_k}^i, v_{y_k}^i, \psi_k^i, x_k^i, y_k^i, n_l^i, n_r^i],
\end{equation}
where \(x_k^i\) and \(y_k^i\) are the absolute world-fixed positions in meters, \(v_{x_k}^i\), \(v_{y_k}^i\) the respective velocities in meters per second, \(\psi_k^i\) is the heading angle of the vehicle in radiance and \(n_l^i\), \(n_r^i\) the number of lanes to the left and right.
Furthermore, at each time step, the layer normalized LSTM considers as an input the previous cell state \(c_{k-0.5s}\) and the previous recurrent state \(h_{k-0.5s}\) \cite{hochreiter97}. Formally, the layer normalized LSTM is defined in Equation \ref{eqn:lstm} \cite{Patel18}.
\begin{equation} \label{eqn:lstm}
(h_{k},c_{k}) = lnLSTM([v_{k}^{0};...;v_{k}^{5};v_{k}^{q}],h_{k-1},c_{k-1})
\end{equation}
Its output at each time step is the cell state \(c_{k}\) and the recurrent state \(h_{k}\). For each prediction \(k \in [t_{-1.5s},t_{-1.0s},t_{-0.5s},t]\) time steps are shown to the layer normalized LSTM layer. \(t\) is the time the prediction is generated. Therefore, the layer normalized LSTM observes four frames of a vehicle and its surroundings within 1.5 seconds, before it creates a prediction.
Layer normalization \cite{ba16} is applied before the non-linearities of the LSTM to increase its robustness.
Layer normalization \(\vartheta(\cdot)\) is defined as follows:
\begin{equation}
\vartheta(a) = f \left[ \frac{g}{\sigma} \odot (a - \mu) + b \right] \quad
\sigma = \sqrt{\frac{1}{H} \sum_{i = 1}^{H} ( a_{i} - \mu)^{2}} \quad
\mu = \frac{1}{H} \sum_{i = 1}^{H} a_{i},
\label{eqn:normalizationlayer}
\end{equation}
where \( \mu \) is the mean of \( a \). \( a \) is the activation vector along the feature axis before the non-linearities of the gated interactions inside an LSTM cell. \( H \)
denotes the number of hidden units in a layer, \( \sigma \) is the standard deviation of \( a \), \(g\) are the learned gain parameters, and \(b\) is a learned bias \cite{ba16}.
Ba et al. showed that layer normalization stabilizes the gradient \cite{ba16}.
This results in a more stable and faster convergence of the validation loss to an optimum at training time and increases classification performance at inference time.
The model's output represents if \(v^q\) changes to the left or right lane or stays on the same lane within the next 2.5 seconds. The labels are a one-hot encoding of the three classes.
\subsection{Explanations of the Prediction Generated by LRP}
Layer normalized LSTM's show state-of-the-art performance at the lane change prediction task \cite{Patel18, xing19}. However, their decision-making process is considered a black box, as it is too complex and complicated to be understood by a user.
We follow the increasing demand and research efforts to explain the decision-making process of a black-box classifier.
The core of our proposed prediction engine applies the Layer-wise Relevance Propagation (LRP) attribution method on the layer normalized LSTM. We aim to make the decision-making process of the lane change prediction explicit, by identifying which part of the input is relevant for the classification.
LRP assigns each input dimension of the layer normalized LSTM a relevance value. The relevance values represent how much each input dimension contributed to the prediction.
LRP starts at the output layer, where the relevance for each neuron is set to be the prediction function value of the class to be explained \(f_c(x)\).
Layer by layer, the relevance is completely redistributed, from higher layer neurons to lower layer neurons by employing layer-specific LRP rules, where neurons that contribute most to the higher layer receive the most relevance from it, as explained in \cite{Bach15}.
Arras et al. propose a chaining of (1) the LRP-\(\epsilon\) rule for the linear mappings, (2) the LRP-\(all\) rule for the gated interactions, and (3) the LRP accumulation rule to explain the interactions of a standard LSTM \cite{Arras2019}.
We extend their approach to layer normalized LSTM's by applying in addition
the LRP-\(\Omega\) rule to the model-specific interaction of a layer normalized LSTM. In particular, we propose the novel LRP-\(\Omega\) rule to explain layer normalization.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{figures/LRP_stepslnlstm_4_paper-Page-1.drawio.png}
\caption{Step by step chaining of the LPR rules for layer normalized LSTM's.} \label{fig:lrp_for_lnlstm}
\end{figure}
Figure \ref{fig:lrp_for_lnlstm} visualizes the LRP rule chaining for the layer normalized LSTM architecture.
\emph{LRP-\(all\) Rule for the Gated Interactions}
The relevance flow of the gated interactions in step (1), (5), and (6) of Figure \ref{fig:lrp_for_lnlstm} are retraced by the LRP-\(all\) rule.
With the LRP-\(all\) rule, all relevance
flows to the source units.
However, the gate units
receive no relevance, as they do not hold information themselves but control the information flow \cite{Arras2019}.
\emph{LRP Accumulation Rule}
Accumulations are interactions appearing in step (4) and (7) on Figure \ref{fig:lrp_for_lnlstm}.
At accumulations, the relevance is split proportional to the magnitude of each addend, as suggested by
\cite{ArrasMMS17}.
\emph{LRP-\(\epsilon\) Rule for Linear Mappings.}
Linear mappings are the interactions depicted in step (2)\footnote{Step (2) in Figure \ref{fig:lrp_for_lnlstm} is called the copy LRP rule. The copy LRP rule is a particular case of the LRP-\(\epsilon\) Rule, where one lower-level node and \(n\) upper-layer nodes exist, the weights are set to one, the bias is zero, and the activation function is linear.}, (9) and (11) of Figure \ref{fig:lrp_for_lnlstm}.
As suggested by \cite{Arras2019}, the LRP-\(\epsilon\) rule
is used to retrace the relevance flow of the linear mapping. The linear mapping is equivalent to a dense layer with a linear activation function and a zero bias.
\emph{LRP-\(\Omega\) Rule for Layer Normalization}
Layer normalization requires a specific LRP rule. According to our knowledge, LRP for layer normalization is not yet explored by the literature. In principle, layer normalization is similar to batch normalization \cite{IoffeS15}. However, they differ in the normalization dimension. While batch normalization normalizes over the whole batch, layer normalization normalizes over one instance \cite{IoffeS15, ba16}. We have explored and applied different LRP approaches for batch normalization to layer normalization, including the LRP identity rule\cite{Arras2019}, LRP-\(|z|\) \cite{Hui19}, LRP-\(\epsilon\) \cite{Hui19}, LRP fusion \cite{Guillemot20}, and LRP heuristic rule \cite{Alber19}.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{figures/lrp_for_bn.drawio.png}
\caption{Heuristic relevance decomposition of layer normalization in Equation \ref{eqn:normalizationlayer} by the LRP-\(\Omega\) rule. A black arrow signalizes no relevance flows to the term according to the LRP-\(\Omega\) rule. A red arrow signalizes relevance flows to the term, and the LRP-\(\Omega\) rule explicitly calculates it. Finally, a blue arrow signalizes relevance flows to the term, but it is not explicitly calculated in the LRP-\(\Omega\) rule as the term is a relevance sink.} \label{fig:lrp_for_ln}
\end{figure}
While the previous approaches made tremendous progress on explaining batch normalization, none fully consider the mean's impact on the relevance. Thus, we propose the novel LRP-\(\Omega\) rule for layer normalization. The LRP-\(\Omega\) rule decomposes layer normalization into a series of summations and scalings. This is shown in Figure \ref{fig:lrp_for_ln}. In particular, the LRP-\(\Omega\) rule acknowledges the impact of the mean shift in Equation \ref{eqn:normalizationlayer} on the relevance flow. Furthermore, it propagates the relevance assigned to the mean further down to the input of the layer normalization.
The LRP-\(\Omega\) rule is formalized in Equation \ref{eqn:lrpnorm}.
\begin{equation}
R_{i\leftarrow j} = (z_{i} - \frac{z_{i}}{H}) \cdot \frac{g_{i}}
{\sigma} \cdot \frac{ R_j }{z_{j}}
\label{eqn:lrpnorm}
\end{equation}
\(R\) is the relevance signal from the input value \(z\) of the higher layer unit \(j\) to the input value of the layer normalization unit \(i\). \(H\) is the length of the input to the layer normalization.
We outlined in this section how LRP calculates relevance values as explanations. The following section shall introduce how the relevance values are transformed into comprehensible explanations.
\subsection{Comprehensible Explanations} \label{subsec:comprehensible_explanations}
LRP assigns relevance values to the \(4\times 49\) dimensions of the layer normalized LSTM's input. This is LRP's explanation of the model's prediction.
The user of the maneuver anticipation system, i.e. the driver, will find 196 relevance values to be incomprehensible. Hence, we have adopted an aggregation approach that utilizes the adaptive nature of LRP in terms of dimensionality reduction. Due to LRPs' relevance conservation and redistribution property \cite{Bach15}, relevance values of terminal units can be added together without invalidating them. The aggregated value represents the relevance of the aggregated units. Thus, features in the input space can be aggregated to meaningful virtual super-features, and their aggregated attribution values represents their relevance for the classification.
The aggregation of relevance values makes it possible to communicate comprehensible explanations to the user. To explain the lane change prediction of the layer normalized LSTM, the time dimensions are aggregated. Therefore, the relevance values of input features representing a vehicle are added together over every time step:
\begin{equation}
\forall i : R_{v^i} = \sum_{\forall k} R_{v^i_k}
\end{equation}
Furthermore, the relevance values of the individual features of the vehicles are aggregated to the virtual and weighted super-features movement \(m\) and position \(p\):
\begin{gather}
\forall i : R_{v^i} = [R_{m^i}, R_{p^i}]; \nonumber \\
with \ R_{m^i} = \frac{R_{v_{x}^i} + R_{v_{y}^i} + R_{\psi^i}}{3}
\ and \ R_{p^i} = \frac{R_{x_k^i} + R_{y_k^i} + R_{n_l^i} + R_{n_r^i}}{4}
\end{gather}
Finally, the three most relevant super-features are communicated to the user via the demonstrator in real-time, as shown in Figure \ref{fig:driver_perspective}.
The three most relevant features are visualized via their name and logo to the driver. In addition, a color scheme describes their relative impact on the classification.
This section described how high dimensional explanations by LRP are reduced to make them comprehensible by the driver while steering a vehicle. Up next, the implementation details of the prototype and an evaluation shall be provided.
\subsection{Prototype architecture}
This prototype has been designed as a distributed set of containers, and as such can be deployed on any Kubernetes cluster with minimal configuration.
This approach was chosen in order to maximize resiliency and redundancy across the application, whilst also logically separating concerns, permitting independent horizontal scaling. Up next, the elements that describe this prototype are detailed.
\emph{Live Adaptor}
The Live Adaptor takes the protobuf stream from the digital twin, decodes and enriches it, so it can be consumed by other parts of the application. This optional step improved the workflow for the rest of the application by propagating the data as JSON. It also checks each vehicle that comes through the digital twin and assigns a UUID. This was necessary since the digital twin itself only assigns vehicle ID's in the cycle 1-10,000, meaning that we lose vehicle uniqueness if we record data that contains over 10,000 vehicles directly from the digital twin. To address this, we looked at each original vehicle ID coming through the digital twin, and checked to see when the ID was last present, if the original vehicle ID has not been present for a period of time, we assume that the vehicle is new, and it is issued with a new UUID. To store UUIDs we use redis as the in-memory cache, preserving state across application restarts and failures. Vehicle IDs contain no identifying information about the vehicle itself.
\emph{Prediction Engine}
The prediction engine is composed of the \textbf{Prediction Model} and the \textbf{Service Broker}, that together enable the user to consume live predictions on demand, in a scalable way. To realize this, we kept the prediction model in a python container that communicates with the service broker over a standard HTTP protocol. Instead of having the prediction model handle connections to the user, we created a "sessioning" platform in the service broker which listens for user requests to open a "prediction session" for a specific UUID. While the session is open, snapshots are repeatedly collected from the live adaptor and are then sent off for inference. The novel element here is that the service broker can handle many connections at once, enabling multiple users, and handles all internal state about user sessions. Because of this, the prediction model itself is stateless and can be scaled horizontally. Once a vehicle leaves the digital twin, the service broker will automatically terminate the session, running any garbage collection.
\emph{General Considerations}
Because of the nature of this domain, specifically our data source being a live digital-twin, considerations were made across every facet of this project to make sure we utilised an event driven architecture. In practice, this meant heavily utilising technologies such as websockets for two-way communication between the system and the user, as well as using websockets to manage state across the system itself. Kubernetes was chosen as our platform as it allowed us to deploy highly customised containers with relative ease.
The architecture of the prototype is fully mapped. Up next, the explanations of the lane change predictions shall be evaluated and the prototype's GUI shall be discussed.
\section{Evaluation and Discussion}
This section evaluates the explanations provided by LRP in terms of their faithfulness to the layer normalized LSTM's behavior. Furthermore, the GUI of the prototype is presented and critically discussed. Please consult \cite{Patel18} for an in-depth evaluation and comparison of the layer normalized LSTM in contrast to other machine learning models for predicting lane changes.
\subsection{Evaluation of the Explanations}
A perturbation test is deployed to evaluate the explanation.
The perturbation test is a behaviouristic approach to evaluate the faithfulness of an explanation. It asks if the explanations reflect the model's behaviour.
For the perturbation test, classifications and attributions, i.e. relevance values, of a representative amount of instances are calculated. Next, the instances are split into correct and wrong classified instances. For the correct classified instances, the most important super-feature is occluded. The occluded instances are classified, and the accuracy is measured. Again, the remaining most relevant super-feature is occluded, and the model's accuracy on the newly created instances is measured. The previous step may be repeated until there is no more feature to occlude.
A faithful explanation method produces results that decrease the accuracy significantly more than randomly occluding features.
To set the faithfulness of LRP for layer normalized LSTM's into context, LRP is compared with the attribution method Integrated Gradients \cite{Mukund2017}.
Furthermore, two versions of LRP are compared: LRP with the LRP identity rule \cite{Arras2019} applied to layer normalization and LRP with the LRP-\(\Omega\) rule for layer normalization.
The results of the perturbation test are depicted in Figure \ref{fig:perturbation_test}.
\begin{figure}[t]
\centering
\includegraphics[width=0.55\textwidth]{figures/perturbation_test.png}
\caption{Perturbation test on the layer normalized LSTM for the lane change prediction task. The perturbation test is conducted on 2315 instances. The instances are randomly drawn from a set, uniformly distributed over the labels.} \label{fig:perturbation_test}
\end{figure}
LRP with the LRP identity rule applied to layer normalization performs worst in the perturbation test. The perturbation test converges towards 50\% accuracy for this rule combination. After nine perturbation steps, it is outperformed by random occlusion. The rule captures the most relevant features accurately, but fails to distribute relevance to minor impactful parts of the input.
The LRP-\(\Omega\) rule outperforms the LRP identity rule significantly. Heuristically redistributing the relevance from the layer's output to its input, while fully considering the impact of the mean on the relevance flow, increases the performance. The LRP-\(\Omega\) rule allows capturing the impact of every part of the input accurately.
Integrated Gradients outperforms LRP for the layer normalized LSTM in terms of faithfulness. At first glance, this is surprising. Gradient-based attribution methods tend to not perform well on standard LSTM's \cite{ArrasMMS17} because the gradient of the sigmoid- and tanh- non-linearities of the LSTM cell is close to zero for activations outside the interval [-4; 4] and respectively [-2; 2]. However, through the layer normalization, the inputs to the non-linearities are brought closer to those intervals, stabilizing the gradient and leading to faithful explanations of gradient-based attribution methods.
\emph{On the Computational Expenses of LRP}
We implemented Integrated Gradients and LRP in Tensorflow 2.4, running on a workstation with two Nvidia 2080TI, CUDA 11.1, 64GB RAM, and an AMD Ryzen Threadripper 2920X. LRP computed the explanations on average 10.47 times faster than Integrated Gradients for 2335 randomly drawn instances. This is due to approximating integrals being computationally expensive.
LRP performs in terms of faithfulness comparably to Integrated Gradients while being significantly more computationally efficient. Thus, LRP is our method of choice for explaining the online lane changes predictions.
Next, the GUI, where the prediction and explanation by LRP are presented, is discussed.
\subsection{Discussion of the Prototype's GUI}
The visualization component is the user facing web application that shows the capabilities of the demonstrator (Figure \ref{fig:driver_perspective}). This allows the user to "jump in" to a vehicle and get various insights as if they were driving the vehicle themselves. The user can see real-time stats such as the nearest neighbours, number of vehicles on the road, the next prediction and the reasons associated with it. The UI also instructs the prediction engine to start or stop predictions for a specific vehicle, rendering the output. In the explainability domain, our chosen approach here is to use a heatmap, where the neighbouring vehicles change color depending upon their actions and how much they impacted the latest prediction.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\textwidth]{figures/screenshot_super_features.png}
\caption{The "driver perspective", showing our selected vehicle (highlighted with "Q"), the surrounding vehicles, and their impact on the latest prediction.} \label{fig:driver_perspective}
\end{figure}
\section{Conclusion}
This paper showcased how to predict and explain lane changes given live data of a digital twin. For that reason, the layer normalized LSTM is outlined as a state-of-art prediction model. However, its decision-making process is too complicated and complex to be understood by a user. In safety-critical applications like lane change predictions, a user must understand the reasoning of the prediction engine. Thus, we developed LRP for layer normalized LSTM's to make the decision-making process of the layer normalized LSTM explicit. LRP shows performant results in terms of faithfulness to the models' behavior while being computationally lightweight. Thus, it is the method of choice for explaining the lane change predictions in real-time.
Furthermore, this paper gave implementation insides on how to realize a scalable, high-performance prototype for making explainable lane change predictions. In addition, we presented the user interface and critically discussed it.
Future work includes implementing multimodal communication strategies of the computed relevance values beyond heatmaps. And the communication strategies of the prototype shall be evaluated in a user study.
Furthermore, it is an interesting open question on how to use the explanation of the model's prediction so that the user interactively improves the prediction model to make it more performant and trustful.
Attribution methods provide deep insights into a black box machine learning model's decision-making process. Let us use those insights to create more trustful and safe machine learning applications.
\bibliographystyle{splncs04}
|
\section{Introduction}
\label{sec:intro}
\by
\yj{Face detection research experienced significant performance improvement after advent of recent deep neural network based} general object detection approaches such as one-stage detector ($e.g.,$ SSD~\cite{ssd}, YOLO~\cite{redmon2016you(yolo)}, RetinaNet~\cite{lin2017focal(retinanet)}, EfficientDet~\cite{tan2020efficientdet(efficientdet)}) and two-stage detector ($e.g.,$ Faster R-CNN~\cite{ren2015faster}, FPN~\cite{fpn}, Mask R-CNN~\cite{he2017mask(mrcnn)}, Cascade R-CNN~\cite{cai2018cascade(cascadercnn)}).
For applicable in a real-world scenario, real-time face detection has attracted more attention, and recent face detectors commonly adopt the one-stage approach that is more simple and efficient than the two-stage approach.
}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{fig/retinaface_3head_backbone_comparison.png}
\caption{Latency-Accuracy pareto curve. We investigate the latency trend of various backbones including MobileNets family and ResNet according to adjusting width multiplier \JH{on RetinaFace\cite{deng2019retinaface} framework}. ResNet shows much lower latency compared with MobileNet family even though its mAP is higher than others.
\yj{One step forward to the observation, we propose a modified ResNet backbone for face detection task, abbreviated as EResNet, which reports superior face detection performance among other backbones.}
}
\label{fig:latency}
\end{figure}
\by{
Recent studies for real-time face detection methods frequently use the lightweight model consisting of depth-wise separable convolution, which is used in MobileNet~\cite{howard2017mobilenets(mobilenet)} and ShuffleNet~\cite{zhang2018shufflenet(sufflenet)}.
Specifically, \yj{a line of recent lightweight face detectors} including RetinaFace~\cite{deng2019retinaface}, SCRFD~\cite{vesdapunt2021crface}, CRFace~\cite{guo2021sample}) employ the MobileNet-V1 architecture~\cite{howard2017mobilenets} as the backbone network and reduce the number of channels in depthwise separable convolution layers by adjusting the width mutiplier.
Following the paradigm of residual block~\cite{resnet}, BlazeFace~\cite{bazarevsky2019blazeface} proposes BlazeBlock that consists of depthwise separable convolution layers with skip connection, achieving a stronger performance.
In practice, adopting the depthwise separable convolution is a reasonable choice to save the number of floating point operations (FLOPs), which is one of the important measurement for the real-time application.
\yj{Summing up the} following common practice, most real-time face detectors utilize the depthwise separable convolution layers in their model by default.
}
\by{
In this paper, we rethink the common belief for the depthwise separable convolution layer and found out that the standard convolution with reducing the number of channels can achieve a better trade-off between latency and detection performance than depthwise separable convolution.
Here, we use ResNet18~\cite{resnet} as our baseline backbone network for the standard convolution and compare with the depthwise separable convolution based backbone networks (MobileNetV1~\cite{howard2017mobilenets(mobilenet)}, MobileNetV2~\cite{sandler2018mobilenetv2}, and MobileNetV3~\cite{howard2019searching}).
\figurename~\ref{fig:latency} shows the latency and average of mean average precision (mAP) scores on WIDER FACE~\cite{widerface} Easy, Medium, and Hard subsets. ResNet18 demands much higher latency than MobileNet when width multiplier is not applied.
However, ResNet18 becomes much faster than MobileNet with higher mAP when reducing the number of channels using the width multiplier.
Note that ResNet18-0.5x denotes that width multiplier 0.5 is applied, and it is 2.1 times faster than MobileNetV1-1.0x and 1.85 times faster than MobieNetV2-1.0x even though its mAP is higher than others.
}
\yj{Based on the observation}, we propose EResFD, which is ResNet based real-time face detector.
\JH{We \yj{firstly propose a slimmed version of} ResNet architecture, namely EResNet, by redesigning the new stem layer, and changing the block configuration.}
Those methods can effectively reduce the inference latency and achieve higher detection accuracy compared with ResNet18.
\yj{Secondly}, we also propose the new feature map enhancement modules; Separated Feature Pyramid Network (SepFPN) and Cascade Context Prediction Module (CCPM).
SepFPN aggregates information from high level and low level feature maps separately, and CCPM further effectively captures diverse receptive fields by employing a cascade design.
\JH{Equipped with these architectural designs, our EResFD achieved 3.1\% higher mAP on WiderFace Hard subset compared to the state-of-the-arts lightweight face detectors such as FaceBoxes\cite{zhang2019faceboxes}. }
We summarize the main contributions as follows:
\begin{itemize}
\item We propose a ResNet-based extremely lightweight backbone architecture, which is much faster than previous work on the CPU and mobile device achieving the state-of-the-art detection performance.
\item We analyze the behavior of both standard convolution and depthwise separable convolution, and we found that the standard convolution is much faster than the depthwise separable convolution under extremely lightweight condition.
\item We propose a channel dimension preserving strategy to reduce the latency, fitting the number of layer in each layer group to recover the performance degeneration.
\item We propose a latency-aware enhance module, SepFPN and CCPM. These enhance modules enable to improve the detection performance on all (large, medium, small) face scales, with much faster speed compared to previous feature enhance modules.
\end{itemize}
\section{Related Works}
\paragraph{Face Detectors}
Recent face detectors~\cite{s3fd,ssh,pyramidbox,dsfd,guo2021sample,deng2019retinaface,zhu2020tinaface,cao2021emface,li2019pyramidbox++,zhang2019accurate,zhu2020progressface,najibi2019fa,liu2020bfbox,yoo2019extd,liu2020hambox,ramos2020swiftface,bazarevsky2019blazeface,hoang2020deface,zhang2020asfd,song2020kpnet} achieved impressive performance enhancement.
These face detectors apply the architectural improvement of the general object detectors such as SSD~\cite{ssd} and RetinaNet~\cite{lin2017focal(retinanet)} or two-stage detectors such as Faster R-CNN~\cite{ren2015faster} and FPN~\cite{fpn}.
The improvement of the face detection enabled to detect faces with various density and scales.
To detect dense and small-scaled faces, current state-of-the-art group of detectors\cite{pyramidbox,li2019pyramidbox++,dsfd,guo2021sample,deng2019retinaface,zhu2020tinaface,liu2020hambox,li2018dsfd} mostly employ large-scaled classification networks with custom-designed upsampling blocks.
Besides the ResNet families~\cite{resnet}, prototypical choice, various attempts including those from architecture search~\cite{zhang2020asfd} has been applied.
PyramidBox series~\cite{pyramidbox,li2019pyramidbox++} and DSFD~\cite{li2018dsfd} suggested own upsampling blocks to improve the expressiveness of the features for dealing with finer faces.
RetinaFace~\cite{deng2019retinaface}, currently dominant one, infers five-point keypoint landmarks of the faces: eyes, nose, mouth, in addition to the detection box, similar to MTCNN~\cite{zhang2016joint}.
HAMBox~\cite{liu2020hambox} focuses on unmatched box which can be used for further performance improvement.
The large memory size requirements of the face detectors critically hinder their applicability on edge-devices. Here, we target on reducing the weight parameters of the backbone network to increase the usability of the face detectors.
\begin{figure*}[t]
\centering
\includegraphics[width=0.8\textwidth]{fig/EResFD_architecture.png}
\caption{Entire architecture of EResFD.
The proposed architecture consists of EResNet with 31 weighted layers,
and SepFPN and Cascade Context Prediction Module (CCPM) are also introduced.
ResBlock denotes basic residual block, which proposed in \cite{resnet}.
The first ResBlock of each stage has stride of 2, and every ResBlock has the same number of output channels as 16 in case of EResFD-1x.
For the classification and regression head, single 1x1 convolution layer is used.}
\label{fig:eresfd_arch}
\end{figure*}
\paragraph{Lightweight Face Detectors}
To run the above mentioned face detectors on mobile or CPU devices, some of the detectors provide their lighter version, mostly substituting their backbones to lighter classification networks utilizing depthwise convolution~\cite{howard2017mobilenets}.
After advent of the pioneering works from MobileNet-V1~\cite{howard2017mobilenets} and V2~\cite{sandler2018mobilenetv2} utilizing depthwise separable convolution and inverted bottleneck block, more refinements~\cite{wu2018shift,howard2019searching,han2020ghostnet,tan2019efficientnet,tan2021efficientnetv2} on the architectures have brought the performance enhancements. These architectures show the competitive ImageNet~\cite{deng2009imagenet} classification accuracy to larger classification models and also for the transferred tasks like object detection~\cite{ren2015faster,ssd} and segmentation~\cite{chen2018encoder}.
Following the improvement of the lightweight backbone networks, RetinaFace~\cite{deng2019retinaface} and SCRFD~\cite{guo2021sample} use channel-width pruned version of MobileNet~\cite{howard2017mobilenets}. BlazeFace~\cite{bazarevsky2019blazeface} and MCUNetV2~\cite{lin2021mcunetv2} proposed new variants of MobileNet targeting on mobile GPU and CPU environment, and EXTD~\cite{yoo2019extd} recursively uses the inverted bottleneck block of the MobileNet for further slimming the network size.
Besides the overall tendency of using depthwise separable convolution based backbones, KPNet~\cite{song2020kpnet} proposes its own backbone network consisting of standard convolutional network, but its size is still large for edge devices, about 1 million parameters, and focuses on sparse and large scaled faces.
In this paper, we rediscover the efficiency of standard convolution layers, which can cover faces with various scale and density, under extremely lightweight model size and minimal inference time.
\section{EResFD}
As we mentioned in Section~\ref{sec:intro}, \yj{ResNet achieves both the faster inference time and higher detection performance.
From this observation}, we revisit the ResNet architecture.
\figurename~\ref{fig:eresfd_arch} illustrates the proposed face detection architecture, named as efficient-ResNet (EResNet) based Face Detector, EResFD.
It consists of two main parts; modified ResNet backbone architecture and \yj{ newly proposed feature enhancement modules.}
We modify several parts of ResNet to reduce the latency while preserving the detection performance \yj{based on empirical analysis on the network,} and we also propose both the new feature pyramid module and context prediction module, which are called Separated Feature Pyramid Network (SepFPN) and Cascade Context Prediction Module (CCPM), respectively.
Both modules improve the detection performance, \yj{and also} show comparable or even faster latency compared with previous state-of-the-arts CPU detectors \cite{faceboxes_improved, jin2019learning}.
\subsection{Rethinking ResNet Architecture}
\subsubsection{Convolutional Layer Analysis}
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.48\textwidth}
\includegraphics[width=0.99\columnwidth]{fig/fig3_layer_latency.png}
\caption{standard convolution vs depthwise separable convolution}
\label{fig:conv_comparison}
\end{subfigure}
\begin{subfigure}[t]{0.48\textwidth}
\includegraphics[width=0.99\columnwidth]{fig/fig4_block_latency.png}
\caption{residual block (ResNet) vs inverted bottleneck block (MobileNet-V2)}
\label{fig:block_comparison}
\end{subfigure}
\caption{Illustration of latency comparison with varying channel size: (a) standard and depthwise separable convolution, (b) residual block (ResNet) and inverted bottleneck block (MobileNet-V2).}
\label{fig:latency_comp}
\vskip -0.1in
\end{figure*}
Depthwise separable convolution is introduced to reduce the multiplication and accumulation \yj{cost of the convolution}, which occupy most of computation time during the inference.
Table~\ref{tab:dwc} shows the comparison of computational cost between the standard and depthwise separable convolution for each stage.
Depthwise separable convolution has much smaller FLOPs, \yj{
and hence} it can significantly reduce the computational cost.
However, previous work~\cite{bello2021revisiting} claims that FLOPs is not always matched with actual latency.
Latency can be bounded by memory access and hardware accelerator, $i.e.,$ CPU or GPU, so the target hardware characteristic should be considered \yj{for the architecture design.}
\input{table1_dwc}
To check the relationship between FLOPs and latency, we investigate the behavior of both standard convolution and depthwise separable convolution on CPU.
We measured the latency of both convolutional layers on CPU, and \figurename~\ref{fig:conv_comparison} shows the comparison result. \JH{Considering faster inference with small-sized input image ($e.g.,$ less than 320x), we tested with input sizes $8{\times}8$, $16{\times}16$, and $32{\times}32$.}
As input size increases, the latency of standard convolution is steeply increasing, but depthwise separable convolution shows the small amount of latency growth.
However, standard convolution achieves smaller latency than \yj{depthwise separable convolution on the extremely light weight condition.} For all the input sizes, standard convolution is faster than depthwise separable convolution when its channel dimension is equal to or smaller than 16 as shown in \figurename~\ref{fig:conv_comparison}.
\yj{Since ResNet and MobileNetV1 each cosists of standard convolution and depthwise separable convolution, we can safely conclude} that ResNet has a chance to become faster than MobilNetV1 when we extremely reduce the channel size.
Furthermore, we also analyze the block level behavior of each convolutional layer.
We use residual block and inverted residual block, which consist of standard and depthwise separable convolution, respectively.
The residual block consists of two standard $3{\times}3$ convolution.
MobileNet-V2 has inverted residual block, which includes one depthwise convolution and two pointwise convolution.
Inverted residual block commonly expand the number of channels for the depthwise convolution, which is called expansion ratio.
In MobileNet-V2, expansion ratio is set to 6 for most inverted residual blocks \cite{sandler2018mobilenetv2}, which is reported to preserve the classification ability of the block compared to standard convolution counterpart~\cite{han2022learning}.
Here, we use the equivalent expansion ratio for the latency comparison.
\figurename~\ref{fig:block_comparison} shows the block level latency, and we found that residual block is much faster than inverted residual block in most cases.
Residual block has $9.43$M FLOPs and inverted residual block has $7.18$M FLOPs when input size is $16{\times}16$ and the number of channels is 32.
Even though residual block has more multiply-add operations, its latency is faster than inverted residual block.
The latency trend of each layer and block shows that standard convolution has a chance to surpass the depthwise separable convolution according to the latency.
This trend is also the same on network level analysis as we mentioned in Section~\ref{sec:intro}.
Therefore, we propose efficient backbone originated from the ResNet
\subsubsection{Stem Layer Modification}
\input{table2_breakdown}
We analyze the ResNet architecture, and we observed that stem layer occupies large amount of entire latency.
Table \ref{tab:latency_breakdown} shows the latency breakdown of ResNet-18 with width multiplier $0.25$, and it shows that almost half of total latency is originated from stem layers.
The backbone network is highly lightened by applying the small width multiplier, so proportion of stem layer becomes larger.
Moreover, ResNet stem layer consists of $7{\times}7$ convolution with stride of 2, so it requires large amount of computation compared than others.
The number of computation (FLOPs) is proportional to square of kernel size ($K^2)$ and reciprocal-square of stride ($1/S^2)$.
If kernel size is reduced to $5$, its FLOPs becomes about $50$\% of $7{\times}7$ convolution, and FLOPs further decrease to about $13$\% when its stride of 4 is applied simultaneously.
To reduce latency of stem layers, we first change stride to 4 for the convolutional layer, which is already adopted in the previous work \cite{faceboxes_improved}
We also reduce the kernel size from $7$ to $5$, but it can hurt the detection performance because it is directly related to receptive field.
To alleviate this problem, we introduce two additional convolutional layers right after $5{\times}5$ convolution.
Thanks to those convolutional layers, receptive field size becomes larger than original stem layer, but its computation complexity is still much lower than original.
\input{table3_stemlayer}
Table \ref{table:stemlayer_modification_latency} shows comparison results on the stem layer.
By adopting smaller kernel size and bigger stride, EResNet stem layer has much smaller FLOPs, and also achieves much shorter latency compared with ResNet stem layer.
\subsubsection{Architecture Re-Configuration for Face}
\input{table4_channelpreserve}
In modern backbone architecture, The number of channels is continuously increasing from bottom to top layer \cite{szegedy2017inception, resnet, tan2019efficientnet}.
In ResNet, for example, channel dimension is doubled when its spatial dimension is decreased (stride of 2).
This designing trend is based on that high-level features are highly related to the specific classes \cite{zeiler2014visualizing}.
When the number of object \JH{classes} increases, high-level feature dimension has to be enlarged.
However, there is only one object \JH{class} in the face detection task, so we suppose that the channel dimension may be reduced compared with object detection task.
To accelerate face detection speed, EResNet backbone is designed by reducing the number of channels.
From the assumption, we propose the channel dimension preserving strategy, which means that we do not double the channel dimension for every stage.
Table~\ref{table:channel_preserve} shows the number of FLOPs and latency when our channel preserving strategy is applied, and it shows that our method significantly reduced the latency and FLOPs of each stage.
\yj{We note that the number of FLOPs is also proportional to input and output channel dimension, and hence the amount of the reduction is huge.}
Latency reduction amount is \yj{not as large as} FLOPs reduction, but it is still remarkable.
\yj{As already shown in \figurename~\ref{fig:eresfd_arch}, we set the channel dimension of our EResNet architecture to 16.}
The network capacity also decreases due to the channel dimension reduction, so we adjust stage configuration to compensate the performance degeneration.
\figurename~\ref{fig:eresfd_arch} shows the detailed stage configurations.
We insert one more blocks for the stage 2$\sim$4 to improve small face detection performance, and add two \JH{extra} stages (stage 5 and 6) for the large face.
The number of residual block increases from $8$ to $14$, so additional residual block can increase the inference time.
However, the channel preserving strategy significantly reduce FLOPs, and the computational cost of each block is also much smaller than original residual block.
For this reason, EResNet architecture is still much faster than baseline ResNet architecture \JH{as shown in Figure \ref{fig:latency}.}
\subsection{Feature Enhance Modules}
\subsubsection{SepFPN}
To improve the detection performance for small objects, feature pyramid network (FPN)~\cite{fpn} is widely used.
FPN propagates \yj{context of high-level feature into low-level feature (top-down) to strengthen the feature information for the small object.}
However, previous work claimed that aggregating high-level feature onto low-level feature can hurt detection performance for small face because receptive field of high-level feature is too large for the small face \cite{pyramidbox}. Also, the state-of-the-art object detectors use not only top-down but also bottom-up path \cite{liu2018path, tan2020efficientdet}, which can induce large latency overhead on FPN module.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{fig/SepFPN.png}
\caption{Architecture of SepFPN. High level features and low level features are aggregated separately, based on fast-normalized fusion \cite{tan2020efficientdet}. Also, final output node is fused with original input node ($i.e.,$ lateral connection), following \cite{tan2020efficientdet}.}
\label{fig:sepfpn}
\end{figure}
To resolve the problems, we propose the a FPN module, separated feature pyramid network (SepFPN).
As illustrated in \figurename~\ref{fig:sepfpn}, we remove the bottom-up path to reduce the latency overhead from the BiFPN~\cite{tan2020efficientdet}, and preserve the lateral connection for the feature aggregation. Features are aggregated with a simple element-wise weighted summation~\cite{tan2020efficientdet}, so its latency overhead is negligible.
Also, we change the top-bottom aggregation method.
From the previous observation \cite{pyramidbox}, we assume that large difference of receptive field \yj{among each aggregation group} causes performance degeneration for the feature aggregation.
Based on this intuition, \yj{we adopt the two separated top-bottom path to let each aggregation group have similar receptive field.}
While low-level feature aggregation has been demonstrated to noticeably improve detection performance for the small face detection \cite{pyramidbox}, we claim that high-level feature aggregation also makes a improvement of large face detection.
\subsubsection{CCPM}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{fig/CCPM.png}
\caption{Architecture of CCPM in case of EResFD-1x. We only visualize CCPM for feature map C1 in Figure \ref{fig:eresfd_arch} for simplicity.}
\label{fig:ccpm}
\end{figure}
\yj{To supplement the feature information,} we also provide cascade context prediction module (CCPM), a latency-aware context module design.
The context prediction module is proposed to expand the receptive field \cite{ssh}, and \yj{we also design CCPM to achieve the equivalent purpose.}
\figurename~\ref{fig:ccpm} shows our CCPM architecture.
\JH{The cascade structure can enrich larger size of receptive field than previous work \cite{ssh}, while ensuring faster speed than the previous heavy-weight enhance modules using densely connected convolution layers~\cite{li2019pyramidbox++} and convolution layers with large asymmetric kernel~\cite{refineface}.}
Thanks to this advantages, CCPM achieves lower latency with comparable detection performance.
\section{Experiment}
\JH{In this section, we evaluate our proposed EResFD by analyzing effectiveness of each component of EResFD and by comparing with the state-of-the-art (SOTA) face detectors. For quantitatively measuring accuracy of detection, we used WIDER FACE~\cite{widerface} dataset. \JH{For training on WIDER FACE, color distortion, zoom-in and out augmentation, max-out background label, multi-task loss are used, following S3FD~\cite{s3fd} \footnote{We obtained source code from \url{https://github.com/yxlijun/S3FD.pytorch}}. For evaluation, we employed flip and multi-scale testing~\cite{s3fd}, where all these predictions are merged by Box voting~\cite{gidaris2015object} with intersection-over-union (IoU) threshold at 0.3. In case of using RetinaFace framework~\cite{deng2019retinaface} \footnote{We obtained source code from \url{https://github.com/biubug6/Pytorch_Retinaface}}, we used single-scale testing where original image size is maintained.
}
For measuring latency, we used Intel Xeon CPU (E5-2660v3@2.60 GHz) with VGA input resolution (480${\times}$640).}
\subsection{Component Study}
\paragraph{Backbone Network.}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.48\linewidth}
\includegraphics[width=1\linewidth]{fig/backbone_comparison_latency.png}
\caption{latency vs accuracy}
\label{fig:latency_vs_accuracy}
\end{subfigure}
\begin{subfigure}[t]{0.48\linewidth}
\includegraphics[width=1\linewidth]{fig/backbone_comparison_params.png}
\centering
\caption{parameter vs accuracy}
\label{fig:params_vs_accuracy}
\end{subfigure}
\caption{Performance comparison of various backbone networks in terms of CPU latency and number of parameters. We applied width multipliers 1x, 1.5x, 2x, 3x, 4x, 6x, 8x for EResNet. For the other backbones, we applied width mulitplier 0.25x, 0.5x, 0.75x.}
\label{fig:backbone_comparison}
\end{figure}
\figurename~\ref{fig:backbone_comparison} shows the comparison result for widely used light-weight backbone; ResNet-18, MobileNet V1, V2, and V3.
\yj{The experimental results show that EResNet achieved superior inference latency given the same mAP condition} and also has higher mAP given the same latency condition, as shown in \figurename~\ref{fig:latency_vs_accuracy}.
Our proposed stem layer and channel dimension preserving strategy \yj{are shown to be very helpful for the latency reduction, while maintaining the powerful face detection performance.
In addition, EResNet also outperformed other comparison methods in terms of the number of parameters.
In \figurename~\ref{fig:params_vs_accuracy}, EResNet shows the highest mAP with much smaller number of parameters.
\JH{To further prove general effectiveness of EResNet backbone, we additionally compared with various backbone architectures on RetinaFace framework in Figure \ref{fig:latency}.
For all the backbones, we only employed 3 detection heads from P2 $\sim$ P4 in Figure \ref{fig:eresfd_arch}, following \cite{deng2019retinaface}.
The results further corroborate that our EResNet architecture has the best latency-accuracy trade-off among the various backbones.}
From those experiments, we found that the proposed methods effectively reduce both latency and parameters without causing mAP degeneration.
\paragraph{SepFPN.}
\input{table5_fpn}
\input{table6_sepfpn_ablation_study}
We measure the detection performance and latency of several FPN modules on the EResNet backbone, and Table~\ref{table:fpn} shows the comparison result.
Thanks to the bottom-up aggregation path, PANet \cite{liu2018path} and BiFPN \cite{tan2020efficientdet} have much higher mAP than others.
However, SepFPN achieves comparable or higher detection performance than both PANet and BiFPN although SepFPN does not have any bottom-up path.
This experiment shows that bottom-up path \yj{would not be an essential block} for the face detection performance, and about $42$\% of the FPN latency can be reduced by removing the bottom-up path.
We further empirically studied on the separation position of SepFPN, and Table~\ref{table:sepfpn_ablation} shows the result.
The latency is not significantly affected from separation position, but the accuracy is very sensitive according to the separation position.
We observed that P5 achieves best mAP for all different kinds of face size, and hence we applied P5 for all other experiments.
\paragraph{CCPM.}
\input{table_ccpm}
Table \ref{table:cpm} shows comparison result of the effect of various feature enhance modules.
The feature enhance module makes large performance gain, but several previous works~\cite{pyramidbox, dsfd, li2019pyramidbox++} show similar detection performance.
Our CCPM module mainly focuses on latency reduction, and experimental result shows that CCPM achieves the smallest latency, satisfying the purpose.
In addition, CCPM also achieves higher overall mAP than SSH, which is the fastest among all the previous methods mentioned in the Table.
\subsection{Comparison with SOTA Detectors}
\input{table7_sota_comp}
We compare our proposed method with the SOTA real-time CPU detectors on WIDER FACE validation dataset.
Table~\ref{table:sota_comp} shows the comparison result.
\textbf{EResNet-1x} indicates EResNet backbone architecture shown in \figurename~\ref{fig:eresfd_arch}, with width multiplier 1.
In case of RetinaFace \cite{deng2019retinaface} backbone, width multiplier 0.25 is applied.
MTCNN shows the smallest FLOPs and Latency, but it has large mAP degradation for medium and hard case.
EResFD has the smallest number of parameters, and also achieves the highest overall mAP.
The latency of EResFD is similar to that of FaceBoxes \cite{zhang2019faceboxes}, but its detection performance is much higher.
Moreover, proposed method achieves similar or slightly higher mAP compared with RetinaFace \cite{deng2019retinaface}, but its latency is about $64$\% of RetinaFace.
Table~\ref{table:sota_comp} (bottom part) also shows the ablation study for the proposed modules; SepFPN and CCPM.
SepFPN improves the overall mAP about $7.5$\%, but its latency only increases by $6$ms.
Moreover, we also achieve $1.9$\% overall mAP improvement when CCPM is further applied.
We observed that proposed modules can make a large performance improvement even when jointly applied.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig/comparison_cpu_gpu.png}
\vspace{-3mm}
\caption{Performance comparison of EResFD with other SOTA CPU target detectors~\cite{zhang2017detecting, mtcnn, faceboxes_improved, jin2019learning, deng2019retinaface} and GPU target detectors~\cite{dsfd, pyramidbox, s3fd, ssh, he2019lffd}. For RetinaFace~\cite{deng2019retinaface},
MobileNetV1-0.25x backbone was used.}
\vspace{-6mm}
\label{fig:comparison_with_cpu_gpu_detectors}
\end{figure}
In addition, we also compare the latency and detection performance with other CPU and GPU target face detectors in \figurename~\ref{fig:comparison_with_cpu_gpu_detectors}.
As we already mentioned above, our method achieves the highest mAP among all the CPU target face detectors.
In addition, we found that it also shows comparable detection performance compared with GPU target detectors.
EResFD shows the similar detection accuracy with S3FD and SSH, but it is about $19$x faster.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.45\linewidth}
\centering
\includegraphics[width=1\linewidth]{fig/retinaface_comparison_detection.png}
\subcaption{Detection}
\vspace{3mm}
\end{subfigure}
\begin{subfigure}[t]{0.45\linewidth}
\includegraphics[width=1\linewidth]{fig/retinaface_comparison_landmark.png}
\centering
\subcaption{Landmark detection}
\end{subfigure}
\vspace{-3mm}
\caption{Performance of face detection on WIDER FACE and landmark detection on AFLW~\cite{koestinger2011annotated} dataset. Based on RetinaFace framework where face bounding boxes with facial landmarks can be jointly detected, we only replaced backbone network from MobileNet-V1 to EResNet, while FPN and SSH are replaced by our proposed SepFPN and CCPM, respectively for EResFD.}
\vspace{-6mm}
\label{fig:retinaface_comparison}
\end{figure}
Furthermore, we compare the proposed method with RetinaFace, \yj{one standard lightweight face detector in this field.}
Since RetinaFace detects face and facial landmarks at the same time, we covered the landmark detection as well.
We measure the face detection performance on WIDER FACE and landmark detection performance on AFLW.
\figurename~\ref{fig:retinaface_comparison} shows comparison on the face and landmark detection.
The results show that our EResFD achieves higher mAP for face detection task and lower NME for the landmark detection task.
\section{Conclusion}
This paper re-highlights the efficiency of standard convolution-based architecture for the lightweight face detection. The extensive experimental results showed that the standard convolutional block achieved superior performance compared to depthwise separable convolution, contrary to the common trend in this field. Based on the observation, we propose an efficient architecture EResNet, which includes modified stem layer and channel dimension preserving strategy. Also, we propose SepFPN and CCPM for the feature enhancement, which boosts the detection performance without sacrificing latency and parameter size. Finally, summing up the observations and architecture suggestions for face detection, we establish a new state-of-the-art real-time CPU face detector, EResFD, achieving the SOTA face detection performance among the lightweight detectors.
\section{Implementaion details}
\subsection{Anchor Settings} As shown in Table \ref{tab:anchor_setting}, anchors whose sizes are ranged from 16 to 512 are assigned from detection layers from D1 to D6, following \cite{s3fd}. Anchor width ratio to the height is set to 1:1.25.
\input{table_appendix_anchor}
\subsection{Training}
During training, we adopt data pre-processing steps, augmentation strategies, and loss functions used in \cite{s3fd}. Specifically, we used color distortion by by changing hue, saturation, value (brightness) of an image. Also, for generating various scales of faces, zoom-in and out operation are applied to image along with its face bounding boxes. Consequently, the resultant image is resized to 640 $\times$ 640 after horizontal flipping. Also, max-out background label~\cite{s3fd} is applied on D1 detection head for reducing false positives with regard to small faces. We also employed multi-task loss, where both classification and regression loss are normalized by number of positive anchors. We set balancing parameter between classification loss and regression loss as 1.
With regard to optimization hyper-parameters, we used ADAM~\cite{kingma2014adam} optimizer with initial learning rate 0.001, weight decay 5e-4, and batch size 32. The maximum number of iterations is 330k and learning rate is decayed at [250k, 300k, 320k] by 0.1.
\subsection{Evaluation} For evaluation on WIDER FACE, we employed flip and multi-scale testing~\cite{s3fd}, where all these predictions are merged by Box voting~\cite{gidaris2015object} with IOU threshold at 0.3.
In case of using RetinaFace framework~\cite{deng2019retinaface}, we used single-scale testing where original image size is maintained.
\section{Stem Modification}
\input{table3_stemlayer_acc}
In Table \ref{table:stemlayer_modification_performance}, we compare detection performance depending on the design of stem layer. It is worth notable that our proposed stem layer requires much more reduced computational costs with significantly improved detection accuracy scores compared to that of ResNet.
\section{SepFPN}
In Figure \ref{fig:sepfpn_p3_p4}, we visualize the architecture of variations for our proposed SepFPN, where its separation position is varied from P3 to P4.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=1.1\linewidth]{fig/SepFPN_seppos_P3.png}
\centering
\subcaption{Separation position: P3}
\vspace{3mm}
\end{subfigure}
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=1.1\linewidth]{fig/SepFPN_seppos_P4.png}
\centering
\subcaption{Separation position: P4}
\end{subfigure}
\caption{Architectures of SepFPN with various separation positions (P3, P4).}
\label{fig:sepfpn_p3_p4}
\end{figure}
|
\section{Introduction}
\label{intro}
\add{Visual attributes describe discriminative visual properties of objects shared among different classes. Attributes have been shown to be important for zero- and few-shot learning, i.e. any-shot learning, as they allow semantic knowledge transfer from known classes with abundant training samples to novel classes with only a handful of images.
Most zero-shot learning~(ZSL) methods~\citep{romera2015embarrassingly,CCGS16,ALE,DEM} rely on image representations extracted from a deep neural network pretrained on ImageNet, and essentially learn a compatibility function between the image representations and attributes. Early solutions for few-shot learning~(FSL) are dominated by metric learning~\citep{snell2017prototypical,vinyals2016matching} which aims to find a better distance function, and meta-learning~\citep{finn2017,finn2018probabilistic,munkhdalai2017meta} which learns better model initialization or optimizer. Focusing on image representations that directly allow attribute localization is relatively unexplored. We argue that any-shot learning can significantly benefit from an image representation that allows localizing visual attributes in images, especially for fine-grained~\citep{tokmakov2019learning,tang2020revisiting} scenarios where local attributes are critical to distinguish two similar categories.
In this work, we refer to the ability of an image representation to localize and associate an image region with a visual attribute as locality. Our goal is to improve the locality of image representations for any-shot learning.}
While modern deep neural networks~\citep{he2016deep} encode local information and some CNN neurons are linked to object parts~\citep{zhou2018interpreting}, the encoded local information is not necessarily best suited for any-shot learning.
There have been attempts to improve the locality of image representations by learning visual attention~\citep{li2018discriminative,SGMA,tang2020revisiting} or attribute classifiers~\citep{sylvain2019locality}. \citet{SGMA} propose to learn channel-wise attention for bird body parts. Similarly, ~\citet{zhu2019learning} apply the channel grouping model~\citep{zheng2017learning} to learn part-based representations and part prototypes. However, the learned latent attentions/prototypes only localize a small number of object parts, and the semantic meaning of the attentions is inducted by post-hoc observation.
Although visual attention accurately focuses on some object parts, the discovered parts and attributes are often biased towards training classes due to the learned correlations.
For instance, the attributes \textit{yellow crown} and \textit{yellow belly} co-occur frequently (e.g. in Yellow Warbler).
The model may learn such correlations as a shortcut to maximize the likelihood of training data and therefore fail to deal with unseen attributes configurations in novel classes, such as \textit{black crown} and \textit{yellow belly} (e.g. in Scott Oriole), as this attribute combination has not been observed before.
To improve locality and mitigate the above weaknesses of image representations, we develop a weakly supervised representation learning framework that localizes and decorrelates visual attributes. More specifically, we learn local features by injecting losses on intermediate layers of CNNs and enforcing these features to encode visual attributes defining visual characteristics of objects. To achieve this, we learn prototypes in the feature space which define the property for each attribute,
at the same time, the local image features are encouraged to be similar to the corresponding attribute prototype. It is worth noting that we use only class-level attributes and semantic relatedness of them as supervisory signal, in other words, no human-annotated association between the local features and visual attributes is given during training.
We propose to alleviate the impact of incidentally correlated attributes by leveraging their semantic relatedness while learning these local features.
As an additional benefit, our model points to the visual evidence of the attributes in an image, confirming the improved attribute localization ability of our image representation.
\revise{The attribute attention map is obtained by measuring the similarity between local image features and attribute prototypes. We evaluate the attribute localization ability quantitatively with the ground truth part annotations, and qualitatively with visualizations. For datasets without attribute location annotations, we propose two user studies to assess the accuracy and semantic consistency of attribute attention maps, and compare our \texttt{APN} model with the baseline visualized by two model explanation methods, \texttt{Grad-CAM} and \texttt{CAM}.}
\add{This paper extends our NeurIPS 2020 conference paper~\citep{xu2020attribute} with the following additional contributions.
(1) We propose to apply the Attribute Prototype Network for the any-shot image classification task, where we improve the locality of image representations. In addition to zero-shot learning, our \texttt{APN} is extended to few-shot learning and the more realistic generalized few-shot learning setting. We evaluate our model under both N-way-K-shot and all-way scenarios and demonstrate that the local representations encoding semantic visual attributes are beneficial for the any-shot regime in discriminating categories with only a few training samples. (2) In addition to performing classification with the original image, we propose to highlight the informative image features discovered by the attribute prototypes, which helps the network to focus on informative attribute regions and discard the noisy background. (3) We verify the effectiveness of our locality-enhanced image representations on top of five
generative models and demonstrate consistent improvement over the state-of-the-art on three challenging benchmark datasets.
\revise{(4) Through qualitative analysis among three benchmark datasets and qualitative ablation study, we demonstrate that our proposed model helps to learn accurate attribute prototypes and produce compact attribute attention maps for each image. Quantitative evaluation indicates that our model outperforms other weakly supervised attribute localization methods by a large margin.}
(5) We propose two well-designed user studies to evaluate the accuracy and semantic consistency of the attribute attention maps, which is an effective evaluation protocol in the absence of ground truth annotations. }
\section{Related work}
\label{Related}
\myparagraph{Zero-shot learning.} The aim of zero-shot learning is to classify the object classes that are not observed during training~\citep{32_awa}. The key insight is to transfer knowledge learned from seen classes to unseen classes with class embeddings that capture similarities between them. Many classical approaches~\citep{romera2015embarrassingly,CCGS16,ALE,DEM,XASNHS16,wang2017zero} learn a compatibility function between image and class embedding spaces. Recent advances in zero-shot learning mainly focus on learning better visual-semantic embeddings~\citep{liu2018generalized,DEM,jiang2019transferable,cacheux2019modeling,wan2021visual,li2020transferrable} or training generative models to synthesize features~\citep{CLSWGAN,xian2019,ABP,zhu2018generative,kumar2018generalized,schonfeld2019generalized,changpinyo2020classifier}. Those approaches are limited by their image representations, which are often extracted from ImageNet-pretrained CNNs or finetuned CNNs on the target dataset with a cross-entropy loss.
Despite its importance, image representation learning is relatively under-explored in ZSL. Recently, \citet{ji2018stacked} propose to weigh different local image regions by learning attentions from class embeddings. \citet{SGMA} extend the attention idea to learn multiple channel-wise part attentions.~\citet{sylvain2019locality} show the importance of locality and compositionality of image representations for ZSL. In our work, instead of learning visual attention like \citet{SGMA} and \citet{ji2018stacked}, we propose to improve the locality of image features by learning a prototype network that is able to localize different attributes in an image.
\myparagraph{Few-shot learning.}
\add{Given a large number of training samples from the base classes, few-shot learning~(FSL) aims to recognize novel classes with a handful of (typically 1–10) labeled examples. The data efficiency problem challenges the traditional classification task, and much effort has been devoted to overcome the data sparsity issue. Metric learning based methods tackle this problem by comparing the distance between images, i.e. Euclidean distance to class prototypes~\citep{snell2017prototypical}, cosine similarity~\citep{vinyals2016matching},
etc. Meta learning based methods aim to learn a good model initialization~\citep{finn2017,finn2018probabilistic,rusu2018meta} or optimizer~\citep{ravi2016optimization,munkhdalai2017meta}, so that the classifiers for novel classes can be learned with a few labeled examples and a small number of gradient update steps.
However, meta-learning may not sufficiently handle the domain shift between base and novel classes~\citep{chen2019closer}, and has difficulty when scaling to a large number of training samples~\citep{xian2019}. This motivates data augmentation via feature synthesis which directly tackles the data deficiency problem by generating novel training examples~\citep{guan2020zero,xian2019}, e.g. ~\citet{hallucinate2017low} and ~\citet{qi2018low} propose to hallucinate additional training examples for novel classes. }
\add{Existing FSL methods usually rely on prior knowledge from only visual modality, while in zero-shot learning, multi-modality data such as word embeddings~\citep{xian2018zero} and attributes~\citep{32_awa} have been adopted and achieve promising results. There are some attempts in applying multi-modality to aid representation learning in FSL~\citep{sylvain2019locality,tokmakov2019learning,tang2020revisiting} or to train a feature generator with GANs~\citep{xian2019,guan2020zero}. A few works have focused on enhancing the compositionality~\citep{tokmakov2019learning} with the help of attributes or learning pose-normalized image representations~\citep{tang2020revisiting}, making the representation invariant to the spatial deformations and environmental changes. In this paper, we focus on learning locality-augmented image representations with the help of class attributes to aid the classification task under a low-data regime. We demonstrate that local information encoded in the image representation helps generative models to synthesize discriminative features.}
\myparagraph{Prototype learning.}
Prototype networks~\citep{yang2018robust,wang2019panet} learn a metric space where the labeling is done by calculating the distance between the test image and prototypes of each class.
Prototype learning is considered to be more robust when handling open-set recognition~\citep{yang2018robust,shu2019p}
and few-shot learning~\citep{snell2017prototypical,gao2019hybrid,oreshkin2018tadam}.
Some methods~\citep{arik2019attention,yeh2018representer,li2018deep} base the network classification decision on learned prototypes.
Instead of building sample-based prototypes, \citet{chen2019looks} dissect the image and find several prototypical parts for each object category, then classify images by combining evidences from prototypes.
Similarly,~\citet{zhu2019learning} use the channel grouping model~\citep{zheng2017learning} to learn part-based representations and part prototypes.
In contrast, we treat each channel equally and use spatial features associated with input image patches to learn attribute prototypes. \citet{chen2019looks} and \citet{SGMA,zhu2019learning} learn latent attention or prototypes during training and induct the semantic meaning of the prototypes in a post-hoc manner. Their attribute or part localization ability is limited, e.g. \citet{SGMA} can only localize two parts. To address those limitations, our method learns prototypes that represent the attributes/parts where each prototype corresponds to a specific attribute.
The attribute prototypes are shared among different classes and encourage knowledge transfer from seen classes to unseen classes, yielding better image representation for zero-shot learning and few-shot learning.
\myparagraph{Locality and representation learning.} Here we define the local feature as the image feature encoded from a local image region. Local features have been extensively investigated for representation learning~\citep{hjelm2018learning,wei2019iterative,noroozi2016unsupervised}, and are commonly used in person re-identification~\citep{sun2018beyond,wang2018learning}, image captioning~\citep{updown,li2017image} and fine-grained classification~\citep{zheng2017learning,fu2017look,zhang2016picking}.
\citet{hjelm2018learning} indicate that maximizing the mutual information between the representation and local regions of the image can significantly improve a representation’s suitability for downstream tasks.
Thanks to its locality-aware architecture, CNNs~\citep{he2016deep} exploit local information intrinsically.
Our work is related to methods that draw attention to local features~\citep{kim2018bilinear,sun2018beyond}.
\citet{zheng2017learning} generate the attention for discriminative bird parts by clustering spatially-correlated channels. Instead of operating on feature channels, we focus on the spatial configuration of image features and improve the locality of our representation. Besides, our work is supervised by the class attributes, and no part or bounding box annotation is required.
\new{Our work is also related to the methods that localize the main object in a weakly supervised way and discard the irrelevant background for image classification~\citep{zhang2021multi,wei2017selective}. \citet{wei2017selective} aggregate the activation map from the last CNN layer~(before global average pooling) and select the highest activated location to filter the final representation for the target image. \citet{zhang2021multi} improve the localization accuracy by introducing multi-layer activation maps. While in our network, we aim to search for the informative image regions that contain important attributes for zero- and few-shot learning. Unlike previous pioneers utilizing intermediate activation maps to localize the informative area, we select the image regions highlighted by the attribute prototypes.} \revise{Since most of the datasets lack ground truth attribute location annotation, e.g. AWA2 and SUN, we design two user studies to assess the accuracy and semantic consistency.}
\section{Attribute Prototype Network}
In the following, we describe our end-to-end trained attribute prototype network~(\texttt{APN}) that improves the attribute localization ability of the image representation, i.e. locality. We first define our zero-shot learning and few-shot learning problem. We then introduce in detail the three modules in our framework, the base module~(\texttt{BaseMod}), the prototype module~(\texttt{ProtoMod}), and the Zoom-In Module~(\texttt{ZoomInMod}) as shown in Figure~\ref{fig:APN}.
At the end of the section, we describe how we perform ZSL and FSL
and how the locality enables attribute localization.
\myparagraph{Problem Definition}
The training set consists of labeled images and attributes from seen classes, i.e. $S=\{x, y, \phi(y)| x \in \mathcal{X}, y \in \mathcal{Y}^s \}$. Here, $x$ denotes an image in the RGB image space $\mathcal{X}$, $y$ is its class label, and $\phi(y) \in \mathbb{R}^{K}$ is the class embedding (i.e. a class-level attribute vector
annotated with $K$ different visual attributes).
Here we use $\mathcal{Y}^n$ to denote the unseen class label set in ZSL and the novel class in FSL for convenience. The class embeddings of unseen classes, i.e. $\{\phi(y)| y \in \mathcal{Y}^n \}$, are also known.
The goal for ZSL is to predict the label of images from unseen classes, i.e. $\mathcal{X} \rightarrow \mathcal{Y}^n$, while for generalized ZSL~(GZSL)~\citep{xian2018zero} the goal is to predict images from both seen and unseen classes, i.e. $\mathcal{X} \rightarrow \mathcal{Y}^n \cup \mathcal{Y}^s$.
\add{Few-shot learning~(FSL) and generalized few-shot learning~(GFSL) are defined similarly. The main difference lies that instead of only knowing the attributes of novel classes in ZSL, FSL also gets a few training samples from each novel class. }
\begin{figure*}[tb]
\label{APN}
\centering
\includegraphics[width=\linewidth]{./figures/APN.pdf}
\caption{
\new{Our attribute prototype network~(\texttt{APN}) consists of an \texttt{Image Encoder} extracting image features $f(x)$, a \texttt{BaseMod} performing classification, a \texttt{ProtoMod} learning attribute prototypes $p_k$ and localizing them with similarity maps $M^k$, and a \texttt{ZoomInMod} cropping the informative image region $\tilde{x}$ covered by attribute similarity maps. The end-to-end training of \texttt{APN} encourages the image feature to contain both global information which is discriminative for classification and local information which is crucial to localize and predict attributes.}
}
\label{fig:APN}
\end{figure*}
\subsection{Base Module~(\texttt{BaseMod}) for global feature learning}
The base module~(\texttt{BaseMod}) learns discriminative visual features for classification. Given an input image $x$, the \texttt{Image Encoder}~(a CNN backbone) converts it into a feature representation $f(x)\in \mathbb{R}^{H\times W \times C}$ where $H$, $W$ and $C$ denote the height, width, and channel respectively.
\texttt{BaseMod} then applies global average pooling over the $H$ and $W$ to learn a global discriminative feature $g(x)\in \mathbb{R}^C$:
\begin{linenomath*}
\begin{equation}
\label{eq:GAP}
g(x) = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} {f_{i,j}(x)} \,,
\end{equation}
\end{linenomath*}
where $f_{i,j}(x) \in \mathbb{R}^C$ is extracted from the feature $f(x)$ at spatial location $(i, j)$~(blue box in Figure~\ref{fig:APN}).
\textbf{Visual-semantic embedding layer.} In contrast to standard CNNs with fully connected layers to compute class logits, we further improve the expressiveness of the image representation using attributes. In detail, a linear layer with parameter $V \in \mathbb{R}^{C\times K}$ maps the visual feature $g(x)$ into the class embedding~(e.g. attribute) space.
The dot product between the projected visual feature and every class embedding is computed to produce class logits $s = g(x)^T V \phi(y)$, followed by the cross-entropy loss encouraging the image to have a high compatibility score with its corresponding attribute vector.
Given a training image $x$ with a label $y$ and an attribute vector $\phi(y)$, the classification loss $\mathcal{L}_{\Scale[0.6]{CLS}}$ is:
\begin{linenomath*}
\begin{equation}
\mathcal{L}_{\Scale[0.6]{CLS}} = -\log \frac{\exp(s)}{\sum_{\mathcal{Y}^s}\exp(s_j)} \,,
\label{equ:CE}
\end{equation}
\end{linenomath*}
where $s_j = g(x)^T V \phi(y_j)$, $y_j \in \mathcal{Y}^s$.
The visual-semantic embedding layer and the CNN backbone are optimized jointly to finetune the image representation guided by the attribute vectors.
\subsection{Prototype Module~(\texttt{ProtoMod}) for local feature learning}
The global features learned from \texttt{BaseMod} may be biased to seen classes because they mainly capture global context, shapes and other discriminative features that may be indicative of training classes. To improve the locality of the image representation, we propose a prototype module~(\texttt{ProtoMod}) focusing on the local features that are often shared across seen and unseen classes.
\begin{comment}
\myparagraph{Attribute prototypes.} \texttt{ProtoMod} takes as input $f(x)\in \mathbb{R}^{H\times W \times C}$ coming from the \texttt{Image Encoder} composed by the local features $f_{i,j}(x)\in \mathbb{R}^{C}$ that encode the information of local image regions where local attributes and object parts appear.
We learn a set of attribute prototypes $P = \left \{ p_k \in \mathbb{R}^{C} \right \}_{k=1}^K$ to predict attributes from those local features, where $p_k$ denotes the prototype for the $k$-th attribute.
As a schematic illustration, $p_1$ and $p_2$ in Figure~\ref{fig:APN} correspond to the prototypes for \textit{black eye} and \textit{blue crown} respectively. For each attribute~(e.g. $k$-th attribute), we compute a similarity map $M^k \in \mathbb{R}^{H\times W}$ that measures the compatibility between its prototype $p_k$ and all the local features, where each element of $M^k$ is computed by $M_{i, j}^k=\langle p_k, f_{i,j}(x)\rangle$. Afterwards, we predict the $k$-th attribute $\hat{a}_k$ by taking the maximum value in the similarity map $M^k$:
\end{comment}
\myparagraph{Attribute prototypes.} \texttt{ProtoMod} takes as input the feature $f(x)\in \mathbb{R}^{H\times W \times C}$ produced by the \texttt{Image Encoder} where the local feature $f_{i,j}(x)\in \mathbb{R}^{C}$ at spatial location $(i, j)$ encodes information of local image regions. Our main idea is to improve the locality of the image representation by enforcing those local features to encode visual attributes that are critical for ZSL.
Specifically, we learn a set of attribute prototypes $P = \left \{ p_k \in \mathbb{R}^{C} \right \}_{k=1}^K$ to predict attributes from those local features, where $p_k$ denotes the prototype for the $k$-th attribute. As a schematic illustration, $p_1$ and $p_2$ in Figure~\ref{fig:APN} correspond to the prototypes for \textit{black eye} and \textit{blue crown} respectively. For each attribute~(e.g. $k$-th attribute), we produce a similarity map $M^k \in \mathbb{R}^{H\times W}$ where each element is computed by a dot product between the attribute prototype $p_k$ and each local feature, i.e. $M_{i, j}^k=\langle p_k, f_{i,j}(x)\rangle$.
Afterwards, we predict the $k$-th attribute $\hat{a}_k$ by taking the maximum value in the similarity map $M^k$:
\begin{linenomath*}
\begin{equation}
\label{eq:pred_attr}
\hat{a}_{k} = \max_{i, j} M_{i, j}^k \,.
\end{equation}
\end{linenomath*}
Other alternative operations such as average pooling and weighted average pooling are also performed for consideration, while max-pooling works the best, since it associates each visual attribute with its closest local feature and allows the network to efficiently localize attributes.
\myparagraph{Attribute regression loss.} The class-level attribute vectors supervise the learning of attribute prototypes. We consider the attribute prediction task as a regression problem and minimize the Mean Square Error~(MSE) between the ground truth attributes $\phi(y)$ and the predicted attributes $\hat{a}$:
\begin{linenomath*}
\begin{equation}
\label{eq:reg}
\mathcal{L}_{\Scale[0.6]{Reg}} = ||\hat{a} - \phi(y) ||^2_2 \,,
\end{equation}
\end{linenomath*}
where $y$ is the ground truth class. By optimizing the regression loss, we enforce the local features to encode semantic attributes, improving the locality of the image representation.
\myparagraph{Attribute decorrelation loss.}
Visual attributes are often correlated with each other as they frequently co-occur, e.g. \textit{blue crown} and \textit{blue back} for Blue Jay birds.
Consequently, the network may use those correlations as a useful signal and fails to recognize unknown combinations of attributes in novel classes.
Therefore, we propose to constrain the attribute prototypes by encouraging feature competition among unrelated attributes and feature sharing among related attributes. To represent the semantic relation of attributes, we divide all $K$ attributes into $L$ disjoint groups, encoded as $L$ sets of attribute indices $S_1,\dots, S_L$. We directly adopt the disjoint attribute groups defined by the datasets~\citep{26_wah2011caltech,32_awa,25_SUNdataset}.
Two attributes are in the same group if they have some semantic tie, e.g. \textit{blue eye} and \textit{black eye} are in the same group as they describe the same body part, while \textit{blue back} belongs to another group.
For each attribute group $S_l$, its attribute prototypes $\{p_k|k \in S_l\}$ can be concatenated into a matrix $P^{S_l} \in \mathbb{R}^{C \times |S_l|}$, and $P^{S_l}_c$ is the $c$-th row of $P^{S_l}$. We adopt the attribute decorrelation~(AD) loss inspired from~\citet{jayaraman2014decorrelating}:
\begin{linenomath*}
\begin{equation}
\label{eq:ad}
\mathcal{L}_{\Scale[0.6]{AD}} = \sum\limits_{c = 1}^{C} \sum\limits_{l = 1}^{L} {\left \| P^{S_l}_c \right \|_2} \,.
\end{equation}
\end{linenomath*}
This regularizer enforces feature competition across attribute prototypes from different groups and feature sharing across prototypes within the same groups, which helps decorrelate unrelated attributes.
\myparagraph{Similarity map compactness regularizer.}
In addition, we would like to constrain the similarity map such that it concentrates on its peak region rather than disperses on other locations. Therefore, we apply the following compactness regularizer~\citep{zheng2017learning} on each similarity map $M^k$,
\begin{linenomath*}
\begin{equation}
\mathcal{L}_{\Scale[0.6]{CPT}} = \frac{1}{KHW} \sum_{k=1}^{K} \sum_{i=1}^{H} \sum_{j=1}^{W} {M_{i,j}^k \left [ {\left ( i - \tilde{i} \right )}^2 + {\left ( j - \tilde{j} \right )}^2 \right ]} \,,
\end{equation}
\end{linenomath*}
where $(\tilde{i}, \tilde{j}) = \arg\max_{i, j} M^k_{i, j}$ denotes the coordinate for the maximum value in $M^k$. This objective enforces the attribute prototype to resemble only a small number of local features, resulting in a compact similarity map.
\subsection{Zoom-In Module (ZoomInMod) for attribute prototype-informed feature learning}
\new{Previous works have shown that the informative attributes are critical to the knowledge transfer in zero-shot learning~\citep{liu2019attribute,guo2018zero,liu2014automatic}.
We propose a Zoom-In Module~(\texttt{ZoomInMod}) to highlight the image regions covered by the informative attribute similarity maps and discard the irrelevant image regions.
Instead of performing classification in \texttt{BaseMod} with the original image $x$~(the orange pipeline in Figure~\ref{fig:APN}), the Zoom-In Module crops out the illuminating image region $\tilde{x}$ that are attended by informative attributes and feed the image into \texttt{BaseMod}~(the green pipeline). As illustrated in Figure~\ref{fig:APN}~(left), we sum up the attribute similarity maps for the most informative attribute in each attribute group to form the attention map $\tilde{M}$:
\begin{linenomath*}
\begin{equation}
\tilde{M} = \sum_{l=1}^{L} M^{n_l} \,,
\text{ where }
n_l = \mathop{\rm arg\,max}\limits_{k \in S_l} a_k \,.
\end{equation}
\end{linenomath*}
$M^{n_l}$ indicates the attribute similarity map, and $n_l$ is the index of the highest predicted~(most informative) attribute in the $l$-th attribute group (e.g. the attention maps for ``white belly'' in ``belly'' attribute group).}
\new{We follow ~\citet{zhang2021multi} to binarize the informative attention map $\tilde{M}$ with the average attention value to form a mask $A$:
\begin{linenomath*}
\begin{equation}
A_{i,j} =
\begin{cases}
1 & \text{ if } \tilde{M}_{i,j} \geq \overline{m} \\
0 & \text{ if } \tilde{M}_{i,j} < \overline{m}
\end{cases} \,,
\text{ where } \overline{m} = \frac{1}{HW} \sum_{i=1}^{H} \sum_{j=1}^{W} \tilde{M}_{i,j} \,.
\end{equation}
\end{linenomath*}
We upsample the binary mask $A$ to the size of the input image, and use the smallest bounding box covering the non-zero area to crop the original image. Then we feed the cropped image $\tilde{x}$ into the \texttt{Image Encoder}.
Note that there are no parameters in the \texttt{ZoomInMod}. When \texttt{ZoomInMod} is working, the \texttt{BaseMod} takes into two inputs, i.e. the original image $x$ and the Zoom-In image $\tilde{x}$, and maps the visual feature $g(x)$ and $g(\tilde{x})$ into the class embedding space with visual-semantic embedding layer $V$. We sum up the class logits for each image to induct the predicted class. So the overall compatibility scores are as follows:
\begin{linenomath*}
\begin{equation}
s = g(x)^T V \phi(y) + g(\tilde{x})^T V \phi(y) \,,
\end{equation}
\end{linenomath*}
used to optimize the classification loss in Equation~(\ref{equ:CE}).}
\myparagraph{Joint global and local feature learning.}
Our full model optimizes the CNN backbone, \texttt{BaseMod} and \texttt{ProtoMod} simultaneously with the following objective function,
\begin{linenomath*}
\begin{equation}
\label{equ:APN_loss}
\mathcal{L}_{\Scale[0.6]{APN}} = \mathcal{L}_{\Scale[0.6]{CLS}} + \lambda_1\mathcal{L}_{\Scale[0.6]{Reg}} + \lambda_2\mathcal{L}_{\Scale[0.6]{AD}} + \lambda_3\mathcal{L}_{\Scale[0.6]{CPT}} \,,
\end{equation}
\end{linenomath*}
where $\lambda_1, \lambda_2$, and $\lambda_3$ are hyper-parameters. The joint training improves the locality of the image representation that is critical for any-shot generalization as well as the discriminability of the features. In the following, we will explain how we perform any-shot inference and attribute localization.
\subsection{Zero- and few-shot learning}
\new{Once our full model is trained, the visual-semantic embedding layer of the \texttt{BaseMod} can be directly used for zero-shot learning inference, which is similar to ALE~\citep{ALE}.
For ZSL, given an image $x$, we generate the ZoomIn image $\tilde{x}$ through the \texttt{ProtoMod} and \texttt{ZoomInMod}, and feed them into the \texttt{BaseMod}. The classifier searches for the class embedding with the highest compatibility via
\begin{linenomath*}
\begin{equation}
\hat{y} = \mathop{\rm arg\,max}\limits_{\tilde{y} \in \mathcal{Y}^n} ~ (g(x)^\mathrm{T}V\phi(\tilde{y}) + g(\tilde{x})^\mathrm{T}V\phi(\tilde{y}))\,.
\end{equation}
\end{linenomath*}
For generalized zero-shot learning~(GZSL), we need to predict both seen and unseen classes. The extreme data imbalance issue will result in predictions to be biased towards seen classes~\citep{chao2016empirical}. To fix this issue, we apply Calibrated Stacking~(CS)~\citep{chao2016empirical} to reduce the seen class scores by a constant factor.
Specifically, the GZSL classifier is defined as,
\begin{linenomath*}
\begin{equation}
\hat{y} = \mathop{\rm arg\,max}\limits_{\tilde{y} \in \mathcal{Y}^n \cup \mathcal{Y}^s} ~ (g(x)^\mathrm{T}V\phi(\tilde{y}) + g(\tilde{x})^\mathrm{T}V\phi(\tilde{y})) - \gamma \mathbb{I} [\tilde{y} \in \mathcal{Y}^s] \,,
\end{equation}
\end{linenomath*}
where the indicator $\mathbb{I}[\cdot]=1$ if $\tilde{y}$ is a seen class and 0 otherwise, $\gamma$ is the calibration factor tuned on a held-out validation set.}
Our model aims to improve the image representation for novel class generalization and is applicable to other ZSL methods~\citep{ABP,CLSWGAN,CCGS16}, i.e. once learned, our features can be applied to any ZSL model~\citep{ABP,CLSWGAN,CCGS16}. Therefore, in addition to the above classifiers, we use image features $g(x)$ extracted from the \texttt{Image Encoder}, and train several state-of-the-art ZSL approaches on top of our features, e.g. ABP~\citep{ABP}, f-VAEGAN-D2~\citep{xian2019}, and TF-VAEGAN~\citep{tfvaegan}.
\add{Our APN network can be adapted to the task of few-shot learning~(FSL) by replacing the feature extraction network in FSL methods with our APN network. During the representation learning stage, we train the feature extractor $f(\cdot)$ using the training examples in the base classes $S=\{x, y, \phi(y)| x \in \mathcal{X}, y \in \mathcal{Y}^s \}$, and train the network with the original FSL training loss as well as our $\mathcal{L}_{\Scale[0.6]{APN}}$ loss.}
\add{With the help of the attribute prototype network, we can learn locality augmented representations that are discriminative for FSL models~\citep{yang2020dpgn} and boost their performance.}
\add{Besides, the locality augmented image representations in our model are applicable to the data synthesis based few-shot learning methods~\citep{xian2019, guan2020zero}. We use image features $g(x)$ extracted from the \texttt{Image Encoder} to train several state-of-the-art generative FSL approaches~\citep{tfvaegan,xian2019} and improve their performance.}
\begin{comment}
\add{We also focus on Few-shot learning~(FSL) problem, to identify a large number of novel classes where only a few samples are provided for each category. \citet{chen2019closer} demonstrate that a simple baseline approach, a cosine classifier on top of fixed CNN backbone becomes a strong baseline. Thus we choose to verify the effectiveness of our \textit{APN features} on top of a cosine classifier. We first fine-tune the feature extractor $f(\cdot)$ (as shown in Figure.~\ref{fig:APN}) using the base classes $S=\{x, y, \phi(y)| x \in \mathcal{X}, y \in \mathcal{Y}^s \}$ with the loss $\mathcal{L}_{\Scale[0.6]{APN}}$ specified in Equation~\ref{equ:APN_loss}. Then we extract the locality enhanced feature $g(x) \in \mathbb{R}^{C}$ following Equation~\ref{eq:GAP}. In the training stage, we optimize a cosine classifier $h\left ( g(t) |\mathrm{\Theta } \right )$ to recognize novel classes. The parameter $\mathrm{\Theta} \in \mathbb{R}^{C \times N} $ can be writen as $\left [ \mathrm{\theta}_{1},\cdots, \mathrm{\theta}_{\Scale[0.5]{N}} \right ]$, where each class has a $C$-dimensional weight vector $\theta$, and $N$ is the novel classes number. Given image feature $g(x)$, we computer its' cosine similarity with the $i$-th weight vector as follows: }
\begin{linenomath*}
\begin{equation}
s_i = \frac{g(x)^\mathrm{T}\mathrm{\theta}_{i}}{\left \| g(x) \right \|\left \| \mathrm{\theta}_{i} \right \|} \,.
\end{equation}
\end{linenomath*}
\add{The predicted probability $P_i \in \mathbb{R}^{N}$ for the $i$-th novel class is obtained by normalizing these similarity scores with a softmax function}
\begin{linenomath*}
\begin{equation}
P_i =\frac{s_i} {\textstyle \sum_{ j = 1}^{N} {s_j}} \,.
\end{equation}
\end{linenomath*}
\add{And we train the classifier with cross-entropy loss $\mathcal{L}_{\Scale[0.6]{FSL}}$}
\begin{linenomath*}
\begin{equation}
\mathcal{L}_{\Scale[0.6]{FSL}} = - \log (P_y) \,,
\end{equation}
\end{linenomath*}
\add{where $y$ is the label of image $x$.}
\add{During inference, the classifier searches for the class with highest cosine similarity via }
\begin{linenomath*}
\begin{equation}
\hat{y} = \mathop{\rm arg\,max}\limits_{\tilde{y} \in \mathcal{Y}^n} ~ P_i \,.
\end{equation}
\end{linenomath*}
\add{Intuitively, the learned weight vectors can be regarded as the class prototypes~\citep{fu2017look}), and we can explicitly reduce the inter-class variations by training the distance-based classifier. The generalized few-shot learning classifier is trained similarly, except that we adopt training samples from both novel and base classes. }
\add{Similar to Zero-shot learning, the locality augmented image representations in our model are applicable to other data synthesis based FSL methods~\citep{xian2019, guan2020zero} and improve their performance. Thus we also test the effectiveness of our \texttt{APN features} on top of the generative models. The representation learning process is the same as above.
We fix the \texttt{Image Encoder} $f$, and use the global discriminative feature $g(x)$ to train a GAN-based few-shot learning approach that aims to augment training samples for novel classes. }
\end{comment}
\subsection{Attribute localization}
\label{sec:Attri_loc}
As a benefit of the improved local features, our approach is capable of localizing different attributes in the image by inspecting the similarity maps produced by the attribute prototypes.
More specifically, we upsample the similarity map $M^k$ to the size of the input image with bilinear interpolation. The area with the maximum responses then encodes the image region that gets associated with the $k$-th attribute.
Figure~\ref{fig:APN} illustrates the attribute regions of \textit{black eye}, \textit{blue crown} and \textit{solid belly} from the learned similarity maps. It is worth noting that our model only relies on class-level attributes and semantic relatedness of them, i.e. attribute groups, as the auxiliary information and does not need any annotation of part locations.
\begin{table*}[t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.2}
\centering
\resizebox{.85\linewidth}{!}{
\begin{tabular}{l|ccc| cccccc | c}
& \multicolumn{3}{c|}{\textbf{ZSL}} & \multicolumn{7}{c}{\textbf{Part localization on CUB}} \\
Method & CUB & AWA2 & SUN & breast & belly & back & head & wing & leg & Mean \\
\hline
$\texttt{BaseMod}$ & $70.0$ & $64.9$ & $60.0$ & $40.3$ & $40.0$ & $27.2$ & $24.2$ & $36.0$ & $16.5$ & $30.7$ \\
+ $\mathcal{L}_{\Scale[0.6]{Reg}}$ & $71.5$ & $66.3$ & $60.9$ & $41.6$ & $43.6$ & $25.2$ & $38.8$ & $31.6$ & $30.2$ & $35.2$ \\
\hspace{2mm}+ $\mathcal{L}_{\Scale[0.6]{AD}}$ & $71.8$ & $67.7$ & $61.4$ & $60.4$ & $52.7$ & $25.9$ & $60.2$ & $52.1$ & $42.4$ & $49.0$ \\
\hspace{4mm}+ $\mathcal{L}_{\Scale[0.6]{CPT}}$ &$72.0$ & $68.4$ & $\textbf{61.6}$ &$63.1$ & $54.6$ & $\textbf{30.5}$ & $64.1$ & $\textbf{55.9}$ & $\textbf{50.5}$ & $52.8$ \\
\hspace{6mm}+ $\texttt{ZoomInMod}$ &$\textbf{75.0}$ & $\textbf{69.9}$ & $61.5$ &$\textbf{67.8}$ & $\textbf{55.9}$ & $29.4$ & $\textbf{68.7}$ & $49.2$ & $49.1$ & $\textbf{53.4}$
\end{tabular}}
\caption{\new{Ablation study of ZSL on CUB, AWA2, SUN~(left, top-1 accuracy) and part localization on CUB~(right, PCP). We train a single \texttt{BaseMod} with the original input $x$ as the baseline. Note that the last row represents our full model \texttt{APN}, which combines \texttt{BaseMod} and \texttt{ProtoMod}~(trained with $\mathcal{L}_{\Scale[0.6]{CLS}}$, $\mathcal{L}_{\Scale[0.6]{Reg}}$, $\mathcal{L}_{\Scale[0.6]{AD}}$, $\mathcal{L}_{\Scale[0.6]{CPT}}$), and the \texttt{ZoomInMod}. }
}
\label{tab:ablation}
\end{table*}
\section{Experiments}
\label{Experiments}
In the following, we first introduce the datasets. Then we evaluate our attribute prototype network in both zero-shot learning as well as the attribute localization performance. We finally evaluate our network on few-shot learning.
\myparagraph{Datasets.}
We conduct the experiments on three widely used benchmark datasets. CUB~\citep{26_wah2011caltech} is a fine-grained dataset containing $11,788$ images from 200 bird classes with $312$ attributes. Following \texttt{SPDA-CNN}~\citep{spdacnn}, we define $7$ body parts for all the birds in CUB dataset, i.e. \textit{belly}, \textit{breast}, \textit{back}, \textit{wing}, \textit{head}, \textit{leg}, \textit{tail}.
Part related attributes are divided into
seven part groups.
SUN~\citep{25_SUNdataset} is a fine-grained dataset consisting of $14,340$ images from $717$ scene classes, with $102$ attributes divided into $4$ groups, describing the \textit{functions}, \textit{materials}, \textit{surface properties} and \textit{spatial envelope} of scene images.
AwA2~\citep{xian2018zero} is a coarse-grained dataset containing $37,322$ images of $50$ animal classes with $85$
attributes. We follow~\citet{32_awa} to divide $85$ attributes into $9$ groups, describing various properties of animals, i.e. \textit{color}, \textit{texture}, \textit{shape}, \textit{body parts}, \textit{behaviour}, \textit{nutrition}, \textit{activity}, \textit{habitat} and \textit{character}. For zero-shot learning, we split the seen/unseen classes following~\citet{xian2018zero}
to avoid overlap between novel images and ImageNet 1K images.
\new{For few-shot learning, we follow two kinds of evaluation protocols. In the N-way-K-shot scenario, the task is to train a classifier with $K$ samples from $N$ classes, to correctly classify the query samples. Following \citet{yang2020dpgn}, we randomly split the CUB dataset into 100 base, 50 validation, and 50 novel classes. In addition to the widely used N-way-K-Shot benchmarks, we also focus on a more realistic setting, i.e. the all-way benchmark~\citep{tang2020revisiting,imaginarywang2018low,hallucinate2017low}. The classifiers are supposed to recognize all the categories simultaneously, i.e. $\mathcal{X} \rightarrow \mathcal{Y}^n$ for FSL and $\mathcal{X} \rightarrow \mathcal{Y}^n \cup \mathcal{Y}^s$ for generalized FSL. In the all-way scenario, we follow~\citet{xian2019} to split three benchmark datasets, i.e. CUB, AWA2, and SUN.}
\begin{figure*}[t]
\centering
\includegraphics[width=.9\linewidth]{./figures/AD_CPT_Ablation.pdf}
\caption{\new{The qualitative ablation study. We display the original image, the Zoom-In image from \texttt{ZoomInMod} in the first and second row. From the third to the fifth row, we show the attribute similarity maps from our \texttt{APN} model, our model trained without $\mathcal{L}_{\Scale[0.6]{CPT}}$, and our model trained without $\mathcal{L}_{\Scale[0.6]{AD}}$, respectively. The text above the attribute similarity maps indicates the attribute name. Green~(purple) box outside the image indicates a correct~(incorrect) localization by our model.}
}
\label{fig:ad_cpt_ablation}
\end{figure*}
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.2}
\begin{table*}[t]
\centering
\small
\resizebox{\linewidth}{!}
{\begin{tabular}{ll| x{1.2cm} x{1.2cm} x{1.2cm} |c c c |c c c |c c c}
& & \multicolumn{3}{c|}{\textbf{Zero-Shot Learning (ZSL)}} & \multicolumn{9}{c}{\textbf{Generalized Zero-Shot Learning (GZSL)}} \\
& & \textbf{CUB} & \textbf{AWA2} & \textbf{SUN} & \multicolumn{3}{c}{\textbf{CUB}} & \multicolumn{3}{c}{\textbf{AWA2}} & \multicolumn{3}{c}{\textbf{SUN}} \\
& Method & \textbf{T1} & \textbf{T1} & \textbf{T1} & \textbf{u} & \textbf{s} & \textbf{H} & \textbf{u} & \textbf{s} & \textbf{H} & \textbf{u} & \textbf{s} & \textbf{H}\\
\hline
\multirow{5}{*}{$\S$}
& \texttt{SGMA}~\citep{SGMA} & $71.0$ & $68.8$ & $-$ & $36.7$ & $71.3$ & $48.5$ & $37.6$ & $87.1$ & $52.5$ & $-$ & $-$ & $-$ \\
&\texttt{AREN}~\citep{AREN} & $71.8$ & $67.9$ & $60.6$ & $63.2$ & $69.0$ & $66.0$ & $54.7$ & $79.1$ & $64.7$ & $40.3$ & $32.3$ & $35.9$ \\
&\texttt{LFGAA+Hybrid}~\citep{liu2019attribute} & $67.6$ & $68.1$ & $61.5$ & $36.2$ & $80.9$ & $50.0$ & $27.0$ & $93.4$ & $41.9$ & $18.5$ & $40.4$ & $25.3$ \\
&\texttt{APN$^{\ddagger}$}~\citep{xu2020attribute} & $72.0$ & $68.4$ & $\textbf{61.6}$& $65.3$ & $69.3$ & $67.2$& $56.5$ & $78.0$ & $65.5$& $41.9$ & $34.0$ & $\textbf{37.6}$\\
&\texttt{APN}~(Ours) & $\textbf{75.0}$ & $\textbf{69.9}$ & $61.5$& $67.4$ & $71.6$ & $\textbf{69.4}$& $61.9$ & $79.4$ & $\textbf{69.6}$& $40.2$ & $35.2$ & $37.5$\\
\hline
\multirow{11}{*}{$\dagger$}
&\texttt{LisGAN}~\citep{li2019leveraging} & $58.8$ & $70.6$ & $61.7$& $46.5$ & $57.9$ & $51.6$& $52.6$ & $76.3$ & $62.3$& $42.9$ & $37.8$ & $40.2$ \\
&\texttt{CLSWGAN}~\citep{CLSWGAN} & $57.3$ & $68.2$ & $60.8$& $43.7$ & $57.7$ & $49.7$& $57.9$ & $61.4$ & $59.6$& $42.6$ & $36.6$ & $39.4$ \\
&\texttt{FREE}~\citep{free2021} & $-$ & $-$ & $-$ & $55.7$ & $59.9$ & $57.7$& $60.4$ & $75.4$ & $67.1$& $47.4$ & $37.2$ & $41.7$ \\
\cline{2-14}
& \texttt{ABP$^*$}~\citep{ABP} & $70.7$ & $68.5$ & $62.6$& $61.6$ & $73.0$ & $66.8$& $53.7$ & $72.1$ & $61.6$& $43.3$ & $39.3$ & $41.2$\\
&\texttt{APN+ABP}~(Ours) & $73.2$ & $\underline{\textbf{73.9}}$ & $63.4$& $65.5$ & $74.6$ & $ 69.8$& $57.4$ & $72.3$ & $64.0$& $46.2$ & $37.8$ & $41.6$\\
\cline{2-14}
& \texttt{CVC$^*$}~\citep{CVC} & $70.0$ & $64.6$ & $61.0$& $61.1$ & $74.2$ & $67.0$& $57.4$ & $83.1$ & $67.9$& $36.3$ & $42.8$ & $39.3$ \\
&\texttt{APN+CVC}~(Ours) & $71.0$ & $71.2$ & $60.6$& $62.0$ & $74.5$ & $67.7$& $63.2$ & $81.0$ & $\underline{\textbf{71.0}}$& $37.9$ & $45.2$ & $41.2$ \\
\cline{2-14}
& \texttt{GDAN$^*$}~\citep{GDAN} & $-$ & $-$ & $-$& $65.7$ & $66.7$ & $ 66.2$& $32.1$ & $67.5$ & $43.5$& $38.1$ & $89.9$ & $53.4$ \\
&\texttt{APN+GDAN}~(Ours) & $-$ & $-$ & $-$& $67.9$ & $66.7$ & $67.3$& $35.5$ & $67.5$ & $46.5$& $41.4$ & $89.9$ & $\underline{\textbf{56.7}}$ \\
\cline{2-14}
& \texttt{f-VAEGAN-D2$^*$}~\citep{xian2019} & $72.9$ & $70.3$ & $65.6$& $63.2$ & $75.6$ & $68.9$& $57.1$ & $76.1$ & $65.2$& $50.1$ & $37.8$ & $43.1$ \\
&\texttt{APN+f-VAEGAN-D2$^*$}~(Ours) & $73.9$ & $73.4$ & $65.9$& $65.5$ & $75.6$ &$70.2$ & $62.7$ & $68.9$ &$65.7$& $49.9$ & $39.3$ & $43.8$\\
\cline{2-14}
& \texttt{TF-VAEGAN$^*$}~\citep{tfvaegan} & $74.3$ & $73.4$ & $65.4$& $65.5$ & $75.1$ & $70.0$& $58.3$ & $81.6$ & $68.0$& $45.3$ & $40.7$ & $42.8$ \\
&\texttt{APN+TF-VAEGAN}~(Ours) & $\underline{\textbf{74.7}}$ & $73.5$ & $\underline{\textbf{66.3}}$& $65.6$ & $76.3$ & $\underline{\textbf{70.6}}$& $60.9$ & $79.1$ & $68.8$& $52.6$ & $37.3$ & $43.7$ \\
\end{tabular}
}
\caption{\new{Comparing our \texttt{APN} model with the state-of-the-art on CUB, AWA2 and SUN datasets. $\dagger$ and $\S$ indicate generative and non-generative representation learning methods respectively. \texttt{AREN}~\citep{AREN} and \texttt{APN}~(Ours) uses Calibrated Stacking~\citep{chao2016empirical} for GZSL. We denote the attribute prototype network trained without \texttt{ZoomInMod} in~\citet{xu2020attribute} as \texttt{APN}$^\ddagger$, and represent our full model with \texttt{ZoomInMod} as \texttt{APN} in this paper. For fair comparison, the generative models marked with $^*$~(e.g. \texttt{ABP$^*$}) use \textit{finetuned features} extracted from ResNet101, while the models below, e.g. \texttt{APN+ABP}(Ours), use our \textit{APN features}. We measure top-1 accuracy~(\textbf{T1}) in ZSL, top-1 accuracy on seen/unseen~(\textbf{s/u}) classes and their harmonic mean~(\textbf{H}) in GZSL.}
}
\label{tab:ZSL_acc}
\end{table*}
\myparagraph{Implementation.}
To train the attribute prototype network, we adopt ResNet101~\citep{he2016deep} pretrained on ImageNet~\citep{39_Imagenet} as the backbone, and jointly finetune the entire model in an end-to-end fashion to improve the image representation.
We use the Adam optimizer~\citep{kingma2014adam} with $\beta_1=0.5$ and $\beta_2$=0.999. The learning rate is initialized as $10^{-6}$ and decreased every ten epochs by a factor of $0.9$. Hyperparameters in our model are obtained by grid search on the validation set~\citep{xian2018zero}. We set $\lambda_1$ as ranges from $0.01$ to $0.1$ for three datasets, $\lambda_2$ as $0.01$, and $\lambda_3$ as $0.2$.
The factor $\gamma$ for Calibrated Stacking is set to $0.7$ for CUB, $0.85$ for AWA2, and $0.4$ for SUN.
\myparagraph{Evaluation metrics}
We follow the same evaluation protocol as demonstrated in~\citet{xian2018zero} and~\citet{xian2019}: for ZSL/FSL we report average top-1 accuracy for unseen~(novel) classes; for GZSL we report average top-1 accuracy for both seen~(s) and unseen~(u) classes, as well as their harmonic mean~(H); for GFSL we report average top-1 accuracy for both seen and novel classes.
\subsection{Zero-shot learning}
In this section, we
present an ablation study of our framework in the ZSL setting, and then we present a comparison with the state-of-the-art in ZSL and GZSL settings.
\myparagraph{Ablation study.} To measure the influence of each model component on the extracted image representation, we design an ablation study where we train a single \texttt{BaseMod} with cross-entropy loss as the baseline, and four variants of \texttt{APN} by adding the \texttt{ProtoMod} and the three loss functions, attribute regression loss $\mathcal{L}_{\Scale[0.6]{Reg}}$, attribute decorrelation loss $\mathcal{L}_{\Scale[0.6]{AD}}$, and compactness regularizer $\mathcal{L}_{\Scale[0.6]{CPT}}$ gradually, and finaly we add the \texttt{ZoomInMod} to generate a ZoomIn image as the input for the \texttt{BaseMod}.
\new{Our results on CUB, AWA2 and SUN presented in Table~\ref{tab:ablation}~(left) demonstrate that the full \texttt{APN} model improve ZSL accuracy over \texttt{BaseMod} consistently, by $5.0\%$~(CUB), $5.0\%$(AWA2), and $1.5\%$~(SUN). }
The attribute regression loss supervises the learning of each attribute prototype and enforces the local image features to contain attribute information, which boosts the performance by $1.5\%$~(CUB), $1.4\%$~(AWA2), and $0.9\%$~(SUN). This indicates that adding locality to the image representation can help the network to learn discriminative features and significantly improve the performance of unseen classes. The attribute decorrelation loss, which suppresses the unwanted attribute co-occurrence and helps to recognize unknown combinations of attributes in novel classes, provides accuracy gains on AWA2~($1.4\%$), CUB~($0.3\%$), and SUN~($0.5\%$). The compactness loss, which constrains the attribute attention maps, does not influence ZSL accuracy much. \new{The \texttt{ZoomInMod} highlights the informative image region and provides significant performance gain for CUB~($3.0\%$) and AWA2~($1.5\%$).}
\revise{The accuracy improvement on SUN is not as great as that on CUB and AWA2 for the following reasons. First, APN aims to learn local image features by regressing and decorrelating attributes, which works well for local and visually-grounded attributes. Most attributes in CUB and AWA2 are related to the local visual properties of the birds and animals. However, the attributes of SUN are designed to be global and abstract~\citep{25_SUNdataset}, thus sometimes can not help with learning locality image features. Second, the data distribution of SUN is quite different from CUB and AWA2. The datasets for bird~(CUB) and animal~(AWA2) classification contain main object and discriminative regions. However, dataset SUN is for scene classification, where each element in a scene is crucial in discriminating it from other categories. For instance, the ``glacier'' in Figure 4~(right, column 2) consists of mountains, snow, and sky, and the global feature of the whole scene can lead to good predictions. Emphasizing local features as APN and the ZoomInMod can not help much.}
\new{The qualitative ablation of our \texttt{APN} network is shown in Figure~\ref{fig:ad_cpt_ablation}. We display the Zoom-In image $\tilde{x}$ generated by the \texttt{ZoomInMod} in the second row. The Zoom-In image $\tilde{x}$ accurately crops out the objects, e.g. the birds in the corner~(row 2, column 3), and discards the noisy background. We qualitatively ablate the attribute decorrelation loss $\mathcal{L}_{\Scale[0.6]{AD}}$ and the compactness regularizer $\mathcal{L}_{\Scale[0.6]{CPT}}$ by visualizing the attribute similarity maps generated by different models. }
\revise{Adding the \texttt{ZoomInMod} increases the model complexity by around two times, e.g. increasing FLOPs from 7.9 GMac to 15.7 GMac, increasing memory from 5,216 MiB to 9,052 MiB. However, the training process is still efficient, which will finish in 48 minutes on CUB dataset with a single NVIDIA V100 GPU.}
\new{From Figure~\ref{fig:ad_cpt_ablation}~(row 3-5), we have the following observations. First, the compactness regularizer $\mathcal{L}_{\Scale[0.6]{CPT}}$ helps the model to generate compact attention maps that focus on the most salient attribute region. For instance, our \texttt{APN} model attends to the correct image region when predicting attributes, and the model works well even when localizing small attribute regions, e.g. the ``yellow nape'', ``yellow forehead'' and ``multi-colored wing'' (row 3, column 3,4,6). While the $\texttt{APN}-\mathcal{L}_{\Scale[0.6]{CPT}}$ model produces disperse attention maps and results in inaccurate attribute similarity maps. When predicting attributes ``yellow nape'' and ``yellow forehead''~(row 4, column 3,4), the model spreads attention over the whole bird body. This observation agrees with the quantitative results in Table~\ref{tab:ablation} where $\mathcal{L}_{\Scale[0.6]{CPT}}$ loss improves the part localization score.}
\new{Second, the decorrelation loss $\mathcal{L}_{\Scale[0.6]{AD}}$ helps the model to avoid attribute correlations and results in precise attention maps. For example, the model $\texttt{APN} - \mathcal{L}_{\Scale[0.6]{AD}}$ misunderstands ``white throat'' with ``white belly''~(row 5, column 2), and ``green throat'' with ``green belly''(row 5, column 5). The reason might be the model $\texttt{APN} - \mathcal{L}_{\Scale[0.6]{AD}}$ learns similar prototypes for two attributes since they share often co-occurring color properties. While the model $\texttt{APN}$ trained with attribute decorrelation loss forces attribute prototypes from different body parts to be different and correctly localizes the corresponding area. }
\myparagraph{Comparing with the SOTA.} We compare our attribute prototype network~(\texttt{APN}) with two groups of state-of-the-art models: non-generative models, i.e. \texttt{SGMA}~\citep{SGMA}, \texttt{AREN}~\citep{AREN}, \texttt{LFGAA+Hybrid}~\citep{liu2019attribute}; and generative models, i.e. \texttt{LisGAN}~\citep{li2019leveraging}, \texttt{CLSWGAN}~\citep{CLSWGAN}, \new{\texttt{FREE}~\citep{free2021}}, \texttt{ABP}~\citep{ABP}, \new{\texttt{CVC}~\citep{CVC}}, \new{\texttt{GDAN}~\citep{GDAN}}, \texttt{f-VAEGAN-D2}~\citep{xian2019}, and \new{\texttt{TF-VAEGAN}~\citep{tfvaegan}} on ZSL and GZSL settings.
As shown in Table~\ref{tab:ZSL_acc}, our \texttt{APN} is
comparable to or better than SOTA non-generative methods in terms of ZSL accuracy. It indicates that our model learns an image representation that generalizes better to unseen classes.
In the more challenging generalized ZSL setting, our \texttt{APN}
achieves impressive gains over state-of-the-art non-generative models for the harmonic mean~(H): we achieve $69.4\%$ on CUB and $37.5\%$ on SUN. On AWA2, it obtains $69.6\%$, which is even better than other generative models, e.g. \texttt{LisGAN} with $62.3\%$, \texttt{CLSWGAN} with $59.6\%$, and \texttt{FREE} with $67.1\%$.
This shows that our network is able to balance the performance of seen and unseen classes well, since our attribute prototypes enforce local features to encode visual attributes facilitating more effective knowledge transfer.
\add{Image features extracted from our model also boosts the performance of generative models that synthesize CNN image features for unseen classes. We choose five SOTA methods \texttt{ABP}~\citep{ABP}, \texttt{GDAN}~\citep{GDAN}, \texttt{CVC}~\citep{CVC}, \texttt{f-VAEGAN-D2}~\citep{xian2019}, and \texttt{TF-VAEGAN}~\citep{tfvaegan} as generative models, and follow the same training and evaluation protocol as stated in these approaches. For fair comparison, we train these models with \textit{finetuned features}~\citep{xian2019} extracted from ResNet101 (denoted with $^*$). We also report the setting where the feature generating models are trained with our \textit{APN feature} $g(x)$~(e.g.~\texttt{APN + TF-VAEGAN}).}
\add{In the ZSL setting, we observe the following. First, for five generative models, our \texttt{APN} consistently boosts their performance on three datasets. For instance, for AWA2, we improve the accuracy of \texttt{ABP} from $68.5\%$ to $73.8\%$, the performance of \texttt{CVC} from $64.6\%$ to $71.2\%$. On the fine-grained datasets CUB and SUN, we also gain performance, e.g. the accuracy of \texttt{ABP} is improved from $70.7\%$ to $73.3\%$ on CUB, from $62.6\%$ to $63.1\%$ on SUN.}
\revise{This indicates that compared to standard finetuning with only cross-entropy loss, our \textit{APN features} provide more local information helping the generative models to synthesize discriminative image features.}
\add{In the GZSL setting, the model predicts both seen and unseen images.
We observe that applying our feature to the generative models consistently boosts the harmonic mean, e.g. we improve \texttt{ABP} by $2.7\%$~(CUB) and $2.3\%$~(AWA2).
Training with our \texttt{APN} feature achieves a more balanced accuracy with much better performance on unseen classes. Compared to \texttt{CVC$^*$} on AWA2, \texttt{APN+CVC} gains $5.8\%$ on unseen while sacrificing only $2.1\%$ on seen. These results demonstrate that our learned locality-enforced image representation lends itself better for knowledge transfer from seen to unseen classes, as the attribute decorrelation loss achieves de-biasing the label prediction. }
\begin{table*}[t]
\centering
\resizebox{0.9\linewidth}{!}{%
\begin{tabular}{l|c|c|cccccc|c}
Method & Parts Annotation & BB size & Breast & Belly & Back & Head & Wing & Leg & Mean \\
\hline
\texttt{SPDA-CNN}~\citep{spdacnn}& \multirow{1}{*}{\cmark} & \multirow{1}{*}{$1/4$} & $67.5$ & $63.2$ & $75.9$ & $90.9$ & $64.8$ & $79.7$ & $73.6$\\
\hline
\texttt{BaseMod}~(uses $\mathcal{L}_{\Scale[0.6]{CLS}}$) & \multirow{2}{*}{\xmark} & \multirow{2}{*}{$1/4$} & $40.3$ & $40.0$ & $27.2$ & $24.2$ & $36.0$ & $16.5$ & $30.7$ \\
\texttt{APN}~(Ours) & & & $67.8$ & $55.9$ & $29.4$ & $68.7$ & $49.2$ & $49.1$ & $53.4$ \\
\hline
\texttt{SGMA}~\citep{SGMA} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{$1/\sqrt{2}$} & $-$ & $-$ & $-$ & $74.9$ & $-$ & $48.1$ & $61.5$ \\
\texttt{APN}~(Ours) & & & $88.1$ & $81.3$ & $71.6$ & $91.4$ & $76.2$ & $70.8$ & $79.9$\\
\end{tabular}
}
\caption{\add{Body part localization in CUB dataset. We comparing our \texttt{APN} with detection model \texttt{SPDA-CNN} trained with part annotations~(row 2), and a ZSL model \texttt{SGMA}. The baseline \texttt{BaseMod} takes the original image feature $g(x)$ as input and is trained with $\mathcal{L}_{\Scale[0.6]{CLS}}$. For BB~(bounding box) size, $1/4$ means each part bounding box has the size $\frac{1}{4}W_b \times \frac{1}{4}H_b$
, where $W_b$ and $H_b$ are the width and height of the bird.
We use gradient-based visual explanation method \texttt{CAM} to visualize the attribute attention map for the baseline \texttt{BaseMod}.
For a fair comparison, we use the same evaluation protocol as \texttt{SGMA} in the last two rows.}
}
\label{tab:atten_loc}
\end{table*}
\subsection{Attribute and part localization in ZSL setting}
We first evaluate the part localization capability of our method quantitatively. We provide an ablation study and comparison with other methods in CUB~\citep{26_wah2011caltech}. In addition, we provide qualitative results of attribute localization on three benchmark datasets. Two user studies are performed to evaluate the accuracy and semantic consistency of attribute similarity maps.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{./figures/CUB_quali.pdf}
\caption{ Part and attribute localization on CUB.
Left:
Attention maps for each body part of Mallard generated by our \texttt{APN}~(first row) and the baseline model \texttt{BaseMod} visualized by \texttt{CAM}~(\texttt{BaseMod(C)}, second row). Boxes mark out the area with the highest attention.
Attention maps are min-max normalized for visualization. \add{Right: Various attribute similarity maps generated by our model. Red, blue, yellow bounding boxes in the image represent the ground truth part bounding box, our results, and the results of \texttt{BaseMod(C)} respectively.} Green~(purple) box outside the image indicates a correct~(incorrect) localization.
}
\label{fig:all_part}
\end{figure*}
\subsubsection{Body part localization}
Given the attribute similarity maps related to six body parts of birds, we report the part localization accuracy by calculating the Percentage of Correctly Localized Parts~(PCP) following SGMA~\citep{SGMA}.
As shown in Figure~\ref{fig:all_part}, the bounding box marks
the image region with the highest attention in each attribute similarity map, then compared with the ground truth part annotation provided by the CUB~\citep{26_wah2011caltech} dataset.
\myparagraph{Ablation study.}
We evaluates the effectiveness of our \texttt{APN} framework in terms of the influence of the attribute regression loss $\mathcal{L}_{\Scale[0.6]{Reg}}$, attribute decorrelation loss $\mathcal{L}_{\Scale[0.6]{AD}}$, the similarity compactness loss $\mathcal{L}_{\Scale[0.6]{CPT}}$, and the \texttt{ZoomInMod}.
As shown in Table~\ref{tab:ablation}~(right), when trained with the joint losses, \texttt{APN} significantly improves the accuracy of \textit{breast}, \textit{head}, \textit{wing} and \textit{leg} by $27.5\%$, $44.5\%$, $13.2\%$, and $32.6\%$ respectively, while the accuracy of \textit{belly} and \textit{back} are improved less. This observation agrees with the qualitative results in Figure~\ref{fig:all_part} that \texttt{BaseMod} tends to focus on the center body of the bird, while \texttt{APN} results in more accurate and concentrated attention maps.
Moreover, $\mathcal{L}_{\Scale[0.6]{AD}}$ boosts the localization accuracy, which highlights the importance of encouraging in-group similarity and between-group diversity when learning attribute prototypes.
\myparagraph{Comparing with SOTA.}
We report PCP in Table~\ref{tab:atten_loc}.
As the baseline, we train a single \texttt{BaseMod} with cross-entropy loss $\mathcal{L}_{\Scale[0.6]{CLS}}$,
and use gradient-based visual explanation method \texttt{CAM}~\citep{CAM} to investigate the image area \texttt{BaseMod} used to predict each attribute. As state of the art, we report the part localization accuracy of a fine-grained classification model \texttt{SPDA-CNN}~\citep{spdacnn} which contains a bird part detection branch supervised by parts annotations. In the last two rows, we compare with the weakly supervised model~(without part annotation) \texttt{SGMA}~\citep{SGMA}, which learns part attention for \textit{head} and \textit{leg} by clustering feature channels
On average, our \texttt{APN} improves PCP over \texttt{BaseMod} by $22.7\%$~($53.4\%$ vs $30.7\%$). The majority of the improvements come from better leg and head localization, e.g. from $24.2\%$ to $68.7\%$~(head), and from $16.5\%$ to $49.1\%$~(leg). Compared to the supervised method \texttt{SPDA-CNN}, our method achieves comparable accuracy on \textit{breast} and \textit{wing}. Although there is still a gap to the supervised method \texttt{SPDA-CNN}, especially on the \textit{back} and \textit{leg} which are hard to localize, the results are encouraging since we do not need require part annotation during training.
Since the feature channels encode more pattern information rather than local information~\citep{geirhos2018imagenet,zhou2018interpreting}, enforcing locality over spatial dimension is more accurate than over channel. Under the same evaluation protocol, we significantly improve the localization accuracy over SGMA~($79.9\%$ vs $61.5\%$ on average). Besides, our model is able to localize attributes related to six body parts, while SGMA can only localize two main parts.
\subsubsection{Qualitative results}
Our model localizes attributes by inspecting the attention maps produced by the attribute prototypes. In this section, we qualitatively investigate the part localization ability on the CUB dataset, as well as the attribute localization results on three benchmark datasets.
\myparagraph{Part localization in CUB.}
We first investigate the difference between our \texttt{APN} and the baseline \texttt{BaseMod} for localizing different body parts in CUB dataset. In Figure~\ref{fig:all_part} (left), for each part of the bird Mallard, we display one attribute similarity map generated by our model \texttt{APN}, and the baseline model \texttt{BaseMod} visualized by \texttt{CAM}~(\texttt{BaseMod(C)}). The baseline model tends to generate disperse attention maps covering the whole bird, as it utilizes more global information, e.g. correlated bird parts and context, to predict attributes. For instance, when predicting attributes of \textit{belly} and \textit{back}, the baseline model utilizes pixels scattered on the bird body.
On the other hand, the similarity maps of our \texttt{APN} are more concentrated and diverse and therefore they localize different bird body parts more accurately. The improvement is lead by the attribute prototypes and compactness loss, which helps the model to focus on local image features when learning attributes.
\myparagraph{Attribute localization in CUB.}
\add{In addition, unlike other models~\citep{spdacnn,uijlings2013selective,SGMA} that can only localize body parts, our \texttt{APN} model can provide attribute-level localization for CUB dataset, as shown in Figure~\ref{fig:all_part} (right). Our model can localize back with various shapes ~(row 1, column 1 and 2) and wings with different postures ~(row 2, column 1 and 2). We can also locate the \textit{pointed belly} of the occluded bird~(row 2, column 4). Compared with \texttt{BaseMod(C)}, our approach produces more accurate bounding boxes that localize the predicted attributes.
For example, while \texttt{BaseMod(C)} wrongly learns the \textit{red leg} from the image region of the belly~(row 1, column 3), our model precisely localizes the \textit{red leg} at the correct region. Specifically, when predicting the attributes for wings and legs, the model tends to focus on both two wings and legs even when they are physically separated~(row 1, column 3 for \textit{red leg} and row 2, column 2 for \textit{gray wing}). These results are interesting because our model is trained on only class-level attributes without accessing any bounding box annotation. }
\add{As a side benefit, the attribute localization ability introduces a certain level of interpretability that supports the zero-shot inference with attribute-level visual evidence.
The last two columns in Figure~\ref{fig:all_part} (right) show some failure examples where our model makes wrong predictions. For example, when the yellow breast is wrongly predicted as black~(row 1, column 4), the attention map tends to spread over the tail and background; when the black breast is recognized as multi-colored~(row 1, column 5), the attention map points to the region of the black and white wing.
Besides, although our attribute decorrelation loss in Equation~\ref{eq:ad} alleviates the correlation issue to some extent~(as shown in the previous results in Table 1 and Figure 2), we observe that our \texttt{APN} seems to still conflate the \textit{yellow belly} and \textit{yellow breast}~(row 2, column 5) in some cases, indicating the attribute correlation issue as a challenging problem for future research. }
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth]{./figures/SUN_AWA_Grad_CAM+CAM_arrow.pdf}
\caption{\add{Attribute attention maps for AWA2~(left) and SUN~(right) from our \texttt{APN} model~(row 2), \texttt{BaseMod} visualized with \texttt{CAM}~(row 3), and \texttt{BaseMod} with \texttt{Grad-CAM}~(row 4). The attribute similarity maps are min-max normalized for visualization. We cover the upsampled attention map on the original image, to show the corresponding location of the highlighted area.
Green~(purple) box outside the image indicates a correct~(incorrect) localization.}
}
\label{fig:SUN_AWA_quali}
\end{figure*}
\myparagraph{Attribute localization in AWA2 and SUN.}
\add{To show how these observations generalize in two other datasets, in Figure~\ref{fig:SUN_AWA_quali}, we compare our \texttt{APN} model with two baseline models on AWA2 and SUN. The attribute attention maps of \texttt{BaseMod} is generated with two gradient-based visual explanation method \texttt{CAM}~\citep{CAM} and \texttt{Grad-CAM}~\citep{grad_cam}.}
\add{In AWA2 dataset~(Figure~\ref{fig:SUN_AWA_quali}, left), our network produces precise similarity maps for visual attributes that describe \textit{texture} and \textit{body parts}, etc.
We can localize visual attributes with diverse appearances, e.g. the white and black \textit{stripe} of zebra, and the yellow and black \textit{stripe} of tiger~(row 2, column 3,4), while
\texttt{CAM} and \texttt{Grad-CAM} fails in localizing the stripe on tigers.
Our similarity maps for \textit{furry} and \textit{longleg} can precisely mask out the image regions of the ox and horse (row 2, column 3,4), while \texttt{BaseMod} only localizes part of the image~(row 1, column 3,4).
On AWA2 dataset we are interested in the visual attributes of animals, while our model in some cases highlights the attributes of the background, e.g. identifying the grid on the rat cage as stripes~(row 2, column 5). This can be explained by the fact that our model only relies on weak supervision, i.e. class-level attributes and their semantic relatedness.}
\add{The attribute similarity maps on SUN dataset are shown in figure~\ref{fig:SUN_AWA_quali}~(right). Our model can discriminate between different attributes with similar color or texture, e.g. correctly locating \textit{snow} and \textit{cloud} in one image~(row 2, column 2,3). The baseline models, on the other hand, cannot distinguish between cloud and snow. Although the appearance of one visual attribute may vary significantly, we can still locate them correctly, e.g. the \textit{fencing} with different colors, location, and shape~(row 2, column 4,5).
Overall, those results indicate that we can perform attribute localization in a weakly supervised manner and provide visual evidence for the inference process of ZSL.}
\begin{figure*}[t]
\centering
\includegraphics[width=0.4\linewidth]{./figures/Userstudy_1.pdf}
\includegraphics[width=0.4\linewidth]{./figures/Userstudy_2.pdf}
\caption{\add{User Study. Left: User study interface to evaluate the accuracy. Participants are required to choose the attention map that more accurately covers the attribute related region from two attention maps generated by our \texttt{APN} model and the \texttt{BaseMod}~(in random position).
Right: User study interface for the semantic consistency. Given three attention maps generated for an attribute, the user selects from two candidate attributes which one is the target of the attention maps.}
}
\label{fig:Userstudy}
\end{figure*}
{
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.2}
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{%
\centering
\begin{tabular}{l l | c }
& \textbf{Methods} & \textbf{Results} \\ \hline
Accuracy & \texttt{APN} vs \texttt{CAM} & $\textbf{76.0}\%$ vs $24.0\%$~($\pm 5.1\%$) \\
& \texttt{APN} vs \texttt{Grad-CAM} & $\textbf{74.4}\%$ vs $25.6\%$~($\pm 2.9\%$) \\ \hline
Semantic & \texttt{CAM} & $55.0\%$~($\pm 3.2\%$)\\
consistency & \texttt{Grad-CAM} & $60.0\%$~($\pm 5.5\%$) \\
& \texttt{APN} & $\textbf{89.0}\%$~($\pm 3.7\%$) \\
\end{tabular}
}
\caption{\add{User study results. Top: the percentage of time that one method is marked as more accurate than the other one by users. Bottom: the percentage of attention maps that can be correctly associated with the target attribute by users. \texttt{APN}, \texttt{CAM}, and \texttt{Grad-CAM} denote our model, baseline model visualized by \texttt{CAM}~\citep{CAM}, and \texttt{Grad-CAM}~\citep{grad_cam} respectively.}}
\label{tab:Userstudy}
\end{table}
}
\subsubsection{User study}
\add{Since the CUB dataset is the only one among the datasets considered here that contains ground truth parts, we design two user studies to assess the accuracy and semantic consistency of attribute attention maps from SUN and AWA2 datasets. We compare the performance of our \texttt{APN} model and the baseline \texttt{BaseMod} visualized by two model explanation methods, \texttt{Grad-CAM} and \texttt{CAM}, respectively.}
\myparagraph{Accuracy of attribute localization.}
\add{The goal is to evaluate whether the attribute attention maps precisely attend to the related image area. As shown in Figure~\ref{fig:Userstudy} (left), each test is a tuple $\left ( M^{i}_{\Scale[0.6]{APN}}, M^{i}_{\Scale[0.6]{BaseMod}}, a_i\right )$ for attribute $a_i$, where $M^{i}_{\Scale[0.6]{APN}}$ is the attribute attention maps generated by our \texttt{APN} model, and $M^{i}_{\Scale[0.6]{BaseMod}}$ by the \texttt{BaseMod}. The human annotators are presented with the tuple
and they are asked to choose the attention map that more accurately covers the attribute region. We randomly sample 50 attention maps from our \texttt{APN} model for 20 visual attributes, then generate the corresponding \texttt{Grad-CAM} and \texttt{CAM} attention maps for \texttt{BaseMod}, and create $100$ tuples in total. Two separate experiments are performed to compare our \texttt{APN} model with the \texttt{BaseMod} visualized by \texttt{Grad-CAM} and \texttt{CAM}.
For each experiment, we employed 5 annotators, i.e. in total 10 students (4 female) aged between 20 and 30 and majoring in computer science participated in the experiment. }
\add{We average the responses from each participant and report the standard deviation between each participant as well as the overall accuracy in Table~\ref{tab:Userstudy}~(top). As for the result, our \texttt{APN} attribute attention maps outperform the \texttt{BaseMod} by a large margin. When comparing the attribute attention maps generated by our \texttt{APN} model and the \texttt{BaseMod} visualized by \texttt{CAM}, in $76.0\%$ cases, \texttt{APN} are marked as more accurately covers the attribute-related area than \texttt{CAM}. And in $74.4\%$ cases, $\texttt{APN}$ are more accurate than \texttt{Grad-CAM}. The user study results agree with the qualitative results in Figure~\ref{fig:SUN_AWA_quali} that \texttt{APN} demonstrates more accurate attribute attention maps than the baseline model. }
\myparagraph{Semantic consistency.}
\add{Here, our aim is to measure whether the attention maps on different images for one attribute is semantically consistent and can be understood by human.
Each test is a tuple $\left ( \mathcal{M}^{i}, a_i, a_j \right )$, where $a_i$ is the target attribute, and $a_j$ is a distractor attribute that is semantically similar to $a_i$. $\mathcal{M}^{i} = \left \{ M^{i}_1, M^{i}_2, M^{i}_3\right \}$ is a random subset of attribute attention maps for the target attribute $a_i$. As shown in Figure~\ref{fig:Userstudy} (right), the human annotators are presented with the tuple via a user interface,
and their task is to identify which of the attributes does the attention map refer to. The performance is defined as the average accuracy of solving such tasks correctly.
We sample $20$ tuples for three methods~(\texttt{APN}, \texttt{BaseMod+CAM} and \texttt{BaseMod+Grad-CAM}) with two criterion: 1) the evaluated original image and attributes stay the same for the three methods, 2) the target $a_0$ and distractor $a_1$ attributes all appear in the original images. Three experiments are performed to evaluate the three methods. There are five annotators for each test, and in total 10 students (4 female) aged between 20 and 30 and majoring in computer science participated in the experiment.}
\add{We average the responses from each user and report their overall accuracy and standard deviation between each user in Table~\ref{tab:Userstudy}~(bottom).
The user study results indicate high semantic consistency for our model. \texttt{APN} attention maps can be associated with the correct attribute in $89.0\%$ cases, while attention maps from \texttt{BaseMod+CAM} only achieve $55.0\%$ accuracy. \texttt{BaseMod+Grad-CAM} gains a higher result, at $60.0\%$, but is still $29.0\%$ lower than ours.
The results indicate that our model is able to generate attribute attention maps that are semantic coherent, and can be understood by human. Although the distractor attribute $a_0$ shares a very similar semantic meaning with the correct attribute~(as shown in Figure~\ref{fig:Userstudy}~(right)), \texttt{APN} attention maps can still help users to find the correct answer. }
\subsection{Few-shot learning}
\new{In the few-shot learning~(FSL) scenario, the images are divided into base classes where plenty of training samples can be obtained and novel classes with only a handful of training samples. The goal of FSL is to learn a classifier to recognize novel classes with limited labeled examples. In the generalized few-shot learning~(GFSL) setting, the classifier is trained to recognize images from both base and novel classes.}
\begin{table*}[t]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{c | c c c c | c c c c | c c c c}
\multirow{2}{*}{\textbf{Method}} & \multicolumn{4}{c|}{\textbf{AWA2}} & \multicolumn{4}{c}{\textbf{CUB}} & \multicolumn{4}{c}{\textbf{SUN}}
\\
& \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} & \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} & \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} \\ \hline
\texttt{Analogy}~\citep{hallucinate2017low} & 62.5 & 81.3 & 82.4 & 87.8 & 56.5 & 69.5 & 78.0 & 81.1 & 40.0 & 49.0 & 67.6 & 75.5 \\
\texttt{Imprinted}~\citep{qi2018low} & 66.9 & 82.6 & 89.0 & 93.5 & 48.5 & 65.0 & 80.0 & 85.3 & 37.9 & 47.8 & 62.1 & 70.2 \\ \hline
\texttt{f-VAEGAN-D2}~\citep{xian2019} & 75.0 & 87.9 & 90.5 & 93.1 & 76.1 & 79.6 & 83.4 & 85.9 & \textbf{68.8} & 69.5 & 70.0 & 70.9 \\
\texttt{APN+f-VAEGAN-D2}~(Ours) & 82.1 & 92.4 & 94.2 & 95.6 & \textbf{77.8} & 81.1 & 84.8 & \textbf{87.1} & 68.2 & 69.4 & 72.4 & 75.1 \\ \hline
\texttt{TF-VAEGAN}~\citep{tfvaegan} & 77.3 & 84.4 & 87.7 & 89.8 & 75.6 & 81.1 & 83.5 & 85.6 & 68.1 & 68.5 & 68.9 & 74.0 \\
\texttt{APN+TF-VAEGAN}~(Ours) & \textbf{87.2} & \textbf{92.8} & \textbf{94.7} & \textbf{96.0} & 77.1 & \textbf{82.6} & \textbf{85.2} & 86.6 & \textbf{68.8} & \textbf{69.9} & \textbf{74.1} & \textbf{76.0} \\
\end{tabular}}
\caption{\new{Few-shot learning results. We apply our \textit{APN features} to feature generation model~(i.e. \texttt{APN+f-VAEGAN-D2} and \texttt{APN+TF-VAEGAN}) and compare with other data synthesis based models. }
\revise{Note that \texttt{f-VAEGAN-D2}~\citep{xian2019} and \texttt{TF-VAEGAN}~\citep{tfvaegan} generate image features given the class attributes to augment the training of novel classes. Our \texttt{APN} model uses the same attributes to learn locality enhanced image features.}}
\label{tab:fsl}
\end{table*}
\begin{table*}[t]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{c | c c c c | c c c c | c c c c}
\multirow{2}{*}{\textbf{Method}} & \multicolumn{4}{c|}{\textbf{AWA2}} & \multicolumn{4}{c}{\textbf{CUB}} & \multicolumn{4}{c}{\textbf{SUN}}
\\
& \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} & \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} & \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} \\ \hline
\texttt{Analogy}~\citep{hallucinate2017low} & 55.0 & 64.7 & 70.7 & 74.5 & 54.1 & 67.5 & 75.5 & 79.5 & 37.5 & 39.6 & 42.4 & 44.6 \\
\texttt{Imprinted}~\citep{qi2018low} & 44.7 & 50.5 & 70.0 & 88.1 & 57.0 & 67.5 & 75.5 & 79.5 & 37.7 & 38.9 & 42.4 & 42.5 \\ \hline
\texttt{f-VAEGAN-D2}~\citep{xian2019} & 72.7 & 80.7 & 85.6 & 88.9 & 71.1 & 74.5 & 77.6 & 79.7 & 43.5 & 43.5 & 43.2 & 44.9 \\
\texttt{APN+f-VAEGAN-D2}~(Ours) & 74.6 & 83.3 & 86.9 & 89.7 & \textbf{72.1} & 76.1 & \textbf{79.3} & \textbf{80.9} & 43.1 & \textbf{44.0} & 45.5 & 45.9 \\ \hline
\texttt{TF-VAEGAN}~\citep{tfvaegan} & 74.7 & 81.1 & 86.0 & 87.1 & 70.6 & 74.7 & 77.5 & 79.3 & 43.3 & 42.3 & 42.9 & 45.4 \\
\texttt{APN+TF-VAEGAN}~(Ours) & \textbf{80.1} & \textbf{85.9} & \textbf{88.6} & \textbf{90.3} & 71.7 & \textbf{76.4} & 78.9 & 80.4 & \textbf{44.3} & 42.9 & \textbf{46.1} & \textbf{46.7} \\
\end{tabular}}
\caption{\new{Generalized few-shot learning results. We apply our \textit{APN features} to feature generation model~(i.e. \texttt{APN+f-VAEGAN-D2} and \texttt{APN+TF-VAEGAN}) and compare with other data synthesis based models. }}
\label{tab:gfsl}
\end{table*}
\myparagraph{Compared methods.}
\new{In this section, we evaluate our attribute prototype network under two evaluation protocols, i.e. the all-way evaluation and N-way-K-shot evaluation.
In the all-way evaluation,
our model is compared with several state-of-the-art generative FSL methods. \texttt{Analogy}~\citep{hallucinate2017low}, \texttt{f-VAEGAN-D2}~\citep{xian2019} and \texttt{TF-VAEGAN}~\citep{tfvaegan} are data synthesis based methods that augment image features for novel classes. \texttt{Imprinted}~\citep{qi2018low} directly uses the normalized activation of novel images as the classifier weight. For a fair comparison, these models are trained with \textit{finetuned feature} extracted from ResNet101, and under the same dataset split~\citep{xian2019}. In the FSL setting, we report the averaged top-1 accuracy for novel classes with the model that only predicts novel classes. In the GFSL setting, we report the averaged top-1 accuracy of test samples of all classes, where the model predicts both base and novel classes labels.}
\new{In the N-way-K-shot evaluation, we build our attribute prototype network over the ResNet12~\citep{he2016deep} backbone in \texttt{DPGN}~\citep{yang2020dpgn}, and train the network with both the FSL training losses~\citep{yang2020dpgn} and our $\mathcal{L}_{\Scale[0.6]{APN}}$ loss. Our model is compared with state-of-the-art FSL methods under the N-way-K-shot evaluation. \texttt{MatchingNet}~\citep{vinyals2016matching}, \texttt{ProtoNet}~\citep{snell2017prototypical} and \texttt{CloserLook}~\citep{fu2017look} propose to optimize the representation learning model with metric learning methods. \texttt{MAML}~\citep{finn2017} learns to initialize the model weight so that it can adapt to novel classes efficiently. Two latest methods further enhance the meta-learning approach with graph network~\citep{yang2020dpgn} and attention agent trained with reinforcement learning~\citep{hong2021reinforced}. We report the top-1 accuracy in the 5way-1shot and 5way-5shot setting following~\citep{yang2020dpgn}.}
\myparagraph{Comparing with the SOTA.}
\new{We display the few-shot learning accuracy in Table~\ref{tab:fsl}. Under the all-way evaluation setting, our model yields consistent improvement on three datasets, i.e. CUB, AWA2, and SUN. On AWA2 dataset, compared to \texttt{TF-VAEGAN} trained with \textit{finetuned feature}, our model \texttt{APN+TF-VAEGAN} improve the FSL accuracy of by a large margin, especially in the low-shot scenario where only a small number of samples from novel classes are available, i.e. we gain $9.9\%$~(1-shot), $8.4\%$~(2-shot), $7.0\%$~(5-shot), and $6.2\%$~(10-shot). }
\add{On fine-grained dataset CUB, our \textit{APN feature} yields consistent improvement over the \textit{finetuned feature}. Compared to \texttt{f-VAEGAN-D2} trained with \textit{finetuned feature}, our model \texttt{APN+f-VAEGAN-D2} gain $1.7\%$~(1-shot), $1.5\%$~(2-shot), $1.4\%$~(5-shot), and $1.2\%$~(10-shot) on FSL. The same trend is observed on SUN dataset. Though the accuracy for the 1-shot and 2-shot regime are comparable to other methods, we manage to improve a lot when training samples increase, e.g. our model \texttt{APN+TF-VAEGAN} achieves $74.1\%$~(5-shot) and $76.0\%$~(10-shot), compared to \texttt{TF-VAEGAN} with $68.9\%$~(5-shot) and $74.0\%$~(10-shot).}
\new{Compared to other FSL models, we gain significant improvements. For instance, we achieve $77.8\%$~(1-shot) on FSL for CUB dataset, compared to $48.5\%$~(\texttt{Analogy}) and $56.5\%$~(\texttt{Imprinted}). For AWA2 dataset, we gain $87.2\%$ on FSL, compared to $66.9\%$~(\texttt{Analogy}) and $62.5\%$~(\texttt{Imprinted}). Specifically, our model using one labeled image per class approaches the accuracy of \texttt{Analogy} and \texttt{Imprinted} trained with five samples. It indicates that improving the locality of image features can better help the feature generators to mimic the real data distribution. When increasing the number of training samples~(i.e. in the 10-shot scenario), the distance between the other three methods shrinks as we are going towards the fully supervised setting. However, our model still manages to improve the performance, which denotes that even with abundant training samples, locality enhanced image features will train a more discriminative classifier than ordinary features.}
\new{The locality augmented model generates discriminative features for novel classes, especially when applied to the generalized setting where the model should predict both base and novel classes. As shown in Table~\ref{tab:gfsl}, on AWA2 dataset, our model $\texttt{APN+TF-VAEGAN}$ gains accuracy improvement for $5.4\%$~(1-shot), $4.8\%$~(2-shot), $2.6\%$~(5-shot), and $3.2\%$~(10-shot). The results demonstrate that our model generates highly discriminative image features by leveraging attribute information. In CUB dataset, we gain $1.0\%$~(1-shot), $1.6\%$~(2-shot), $1.7\%$~(5-shot), and $1.1\%$~(10-shot). In SUN dataset, we also manage to improve the accuracy consistently, e.g. we gain $1.0\%$~(1-shot), $0.7\%$~(2-shot) and $3.2\%$~(5-shot), and $1.3\%$~(10-shot) on GFSL.}
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{c |c |c c}
\textbf{Method} & \textbf{Backbone} & \textbf{5way-1shot} & \textbf{5way-5shot} \\
\hline
\texttt{MatchingNet*} & ResNet18 & $72.4\pm0.90$ & $83.6\pm0.60$ \\
\texttt{ProtoNet*} & ResNet18 & $73.0\pm0.88$ & $86.6\pm0.51$ \\
\texttt{MAML*} & ResNet18 & $68.4\pm1.07$ & $83.5\pm0.62$ \\
\texttt{CloserLook} & ConvNet & $60.5\pm0.83$ & $79.3\pm0.61$ \\
\texttt{ArL} & ConvNet & $50.6$ & $65.9$ \\
\texttt{RAP+ProtoNet} & ResNet18 & $74.1\pm0.60$ & $89.2\pm0.31$ \\
\texttt{RAP+Neg-Margin} & ResNet18 & $75.4\pm0.81$ & $90.6\pm0.39$ \\ \hline
\texttt{DPGN} & ResNet12 & $75.7\pm0.47$ & $91.5\pm0.33$ \\
\texttt{APN+DPGN} (Ours) & ResNet12 & $\textbf{77.4}\pm0.44$ & $\textbf{92.2}\pm0.24$ \\
\end{tabular}}
\caption{\new{Few-shot learning results on CUB dataset (* results are from~\citet{hong2021reinforced}). Following~\citet{yang2020dpgn}, we report top-1 accuracy in 5way-1shot/-5shot settings and compare ours with \texttt{MatchingNet}~\citep{vinyals2016matching}, \texttt{ProtoNet}~\citep{snell2017prototypical}, \texttt{MAML}~\citep{finn2017}, \texttt{CloserLook}~\citep{fu2017look}, \texttt{ArL}~\citep{zhang2021rethinking}, \texttt{RAP+ProtoNet}~\citep{hong2021reinforced}, \texttt{RAP+Neg-Margin}~\citep{hong2021reinforced}, \texttt{DPGN}~\citep{yang2020dpgn}. }}
\label{tab:meta_cub}
\end{table}
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.2}
\begin{table*}[h]
\centering
\small
\resizebox{\linewidth}{!}
{\begin{tabular}{l| x{1.2cm} x{1.2cm} x{1.2cm} |c c c |c c c |c c c}
& \multicolumn{3}{c|}{\textbf{Zero-Shot Learning (ZSL)}} & \multicolumn{9}{c}{\textbf{Generalized Zero-Shot Learning (GZSL)}} \\
& \textbf{AWA2} & \textbf{CUB} & \textbf{SUN} & \multicolumn{3}{c}{\textbf{AWA2}} & \multicolumn{3}{c}{\textbf{CUB}} & \multicolumn{3}{c}{\textbf{SUN}} \\
Method & \textbf{T1} & \textbf{T1} & \textbf{T1} & \textbf{u} & \textbf{s} & \textbf{H} & \textbf{u} & \textbf{s} & \textbf{H} & \textbf{u} & \textbf{s} & \textbf{H}\\
\hline
\texttt{CADA-VAE}~\citep{schonfeld2019} & $49.0$ & $22.5$ & $\textbf{37.8}$ & $38.6$ & $60.1$ & $47.0$ & $16.3$ & $39.7$ & $23.1$ & $26.0$ & $28.2$ & $\textbf{27.0}$ \\
\texttt{SJE}~\citep{akata2015evaluation} & $53.7$ & $14.4$ & $26.3$ & $39.7$ & $65.3$ & $48.8$ & $13.2$ & $28.6$ & $18.0$ & $19.8$ & $18.6$ & $19.2$ \\
\texttt{GEM-ZSL}~\citep{liu2021goal} & $50.2$ & $25.7$ & $-$ & $40.1$ & $80.0$ & $53.4$ & $11.2$ & $48.8$ & $ 18.2$ & $-$ & $-$ & $-$ \\
\texttt{APN(w2v)}~(Ours) & $\textbf{59.6}$ & $\textbf{27.7}$ & $32.1$ & $41.8$ & $75.0$ & $\textbf{53.7}$ & $20.6$ & $26.4$ & $\textbf{23.4}$ & $20.3$ & $21.3$ & $20.8$ \\
\end{tabular}
}
\caption{\revise{Zero-Shot Learning results from our \texttt{APN} model and other state-of-the-art on CUB, AWA2, and SUN datasets. All the models are trained with unsupervised class embedding, i.e. w2v.
\texttt{SJE}~\citep{akata2015evaluation}, \texttt{GEM-ZSL}~\citep{liu2021goal}, and our model \texttt{APN} are non-generative models, while \texttt{CADA-VAE}~\citep{schonfeld2019} is feature generation model. We measure top-1 accuracy~(\textbf{T1}) in ZSL, top-1 accuracy on seen/unseen~(\textbf{s/u}) classes and their harmonic mean~(\textbf{H}) in GZSL. }
}
\label{tab:ZSL}
\end{table*}
\begin{table*}[h]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{c | x{0.8cm} x{0.8cm} x{0.8cm} x{0.8cm} x{0.8cm} | x{0.8cm} x{0.8cm} x{0.8cm} x{0.8cm} x{0.8cm}}
\multirow{2}{*}{\textbf{Method}} & \multicolumn{5}{c|}{\textbf{Few-Shot Learning}} & \multicolumn{5}{c}{\textbf{Generalized Few-Shot Learning}}
\\
& \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} &\textbf{20} & \textbf{1} & \textbf{2} & \textbf{5} & \textbf{10} & \textbf{20} \\ \hline
\texttt{softmax} & 49.3 & 64.5 & 76.7 & 81.0 & 84.2 & 50.0 & 60.8 & 73.5 & 79.0 & 80.7 \\
\texttt{Analogy}~\citep{hallucinate2017low} & 40.5 & 50.7 & 61.6 & 69.5 & 76.0 & 51.5 & 59.8 & 67.4 & 72.0 & 77.3 \\
\texttt{f-VAEGAN-D2-ind}~\citep{xian2019} & 54.4 & 64.4 & 74.6 & 79.9 & 84.0 & 60.3 & 65.7 & 73.5 & 78.8 & 79.5 \\
\texttt{f-VAEGAN-D2}~\citep{xian2019} & 60.1 & 70.0 & 79.0 & 81.5 & 84.5 & 66.3 & 72.6 & 78.6 & 81.5 & 83.2 \\
\texttt{APN(w2v)+f-VAEGAN-D2} & \textbf{61.7} & \textbf{70.9} & \textbf{79.5} & \textbf{83.4} & \textbf{85.5} & \textbf{67.5} & \textbf{73.8} & \textbf{79.6} & \textbf{82.5} & \textbf{84.3} \\
\end{tabular}}
\caption{\revise{Few-Shot Learning results on ImageNet with increasing number of training samples per novel class (top-5 accuracy). Our \texttt{APN(w2v)} model is trained with w2v class embeddings. We apply our APN features to feature generation model~(i.e. \texttt{APN(w2v)+} \texttt{f-VAEGAN-D2}).}}
\label{tab:imagenet}
\end{table*}
\new{As shown in Table~\ref{tab:meta_cub}, in the N-way-K-shot scenario where we train \texttt{DPGN}~\citep{yang2020dpgn} with our attribute prototype network, we yield improvement over the baseline \texttt{DPGN} on CUB and achieve new state-of-the-art accuracy. We achieve $77.4\%$~(5way-1shot), compared to $68.4\%$~(\texttt{CloserLook}), $74.1\%$~(\texttt{RAP+ProtoNet}) and $75.7\%$~(\texttt{DPGN}). In the 5way-5shot setting, we also improve over \texttt{CloserLook}, \texttt{RAP+ProtoNet} and \texttt{DPGN} by $12.9\%$, $3.0\%$ and $0.7\%$ respectively. The results indicate that integrating attribute prototype network into the representation learning process helps the network to learn locality enhanced features and benefit the FSL performance.}
\subsection{Discussion}
\revise{We study the flexibility of our \texttt{APN} network by using only unsupervised embeddings, e.g. word2vec~(w2v)~\citep{mikolov2013distributed}. Similar to learning attributes, for w2v with 300 dim, we design 300 prototypes in the \texttt{ProtoMod}. Since there is no grouping for w2v, we discard the attribute decorrelation loss $\mathcal{L}_{\Scale[0.6]{AD}}$ when training the model.}
\revise{We compare the zero-shot learning results of our APN model with other state-of-the-art models in Table~\ref{tab:ZSL}. The performance of using w2v as the class embeddings could drop compared to attributes. But this issue exists in most zero-shot learning methods, as shown in Table~\ref{tab:ZSL}. In general, our method can be generalized to unsupervised class embeddings and still outperforms the baselines using w2v.
For instance, we outperform other non-generative models such as \texttt{SJE} and \texttt{GEM-ZSL} on three datasets, e.g. we achieve 59.6\% on AWA2, comparing with \texttt{SJE} with 53.7\% and \texttt{GEM-ZSL} with 50.2\%. Our model even outperform generative model \texttt{CADA-VAE} that synthesize image features on AWA2 and CUB dataset.}
\revise{We perform few-shot learning on ImageNet to study the generalization ability of our model on the large-scale dataset. We follow the data split in~\citet{xian2019} where the 1K ImageNet categories are randomly divided into 389 base classes and 611 novel classes. We use the 300-dim word2vec~\citep{mikolov2013distributed} embeddings as the class embedding for ImageNet since there is no attribute annotation. Our \texttt{APN} model is trained with train samples from the base classes.
The results shown in Table~\ref{tab:imagenet} demonstrate that our \texttt{APN} model trained with only w2v learns better representations and further boosts the few-shot learning performance on the large-scale ImageNet dataset. Notably, we gain 1.6\%~(1-shot) on FSL and 1.2\%~(1-shot) on GFSL.}
\section{Conclusion}
\add{In this work, we develop a representation learning framework for zero-shot learning and few-shot learning, i.e. attribute prototype network (\texttt{APN}), to jointly learn global and local features.
Our model improves the locality of image representations by regressing attributes with local features and decorrelating
prototypes with regularisation. }\new{We explicitly encourage the network to learn from informative attribute related image regions and discard noisy backgrounds by cropping the original image with attribute similarity maps.} \add{
We demonstrate consistent improvement over the state-of-the-art on three benchmarks. And our representations improve over finetuned ResNet representations when used in conjunction with feature generating models. We qualitatively verify that our network is able to localize attributes in images accurately. Two well-designed user studies indicate that our network can generate semantically consistent and accurate attribute attention maps. The part localization accuracy significantly outperforms a weakly supervised localization model designed for zero-shot learning. We further show that our model can be extended to the FSL scenario, and we consistently improve the classification accuracy in any-shot regimes on three datasets.}
\begin{acknowledgements}
This work has been partially funded by the ERC 853489 - DEXIM and by the DFG – EXC number 2064/1 – Project number 390727645.
\end{acknowledgements}
\bibliographystyle{spbasic}
|
\section{Emergent Gauge Fields from Lattice Consideration}
\label{sec:emergent-gauge-derivation}
In this section, we elucidate the procedure by which the emergent gauge fields are obtained from the lattice consideration. Initially the discretized phase integrals were written as areal products $W_{e_x,m}$ and $W_{e_y,m}$:
\begin{align}
W_{e_x , m }=& \prod_{i \in \cal{A}} \left( b_i\right)^{y_{x0} -y_i }, &
W_{e_y , m } =& \prod_{i \in \cal{A}} \left( b_i\right)^{x_i - x_{y0} }. \label{eq:weg-loop-e}
\end{align}
They are illustrated in Fig. \ref{fig:tx} (a) and (c) in the case of the rectangular boundary ${\cal A} = [x_0, x_1] \times [y_0 , y_1 ]$. By substituting the definition of $b_i$ in terms of spin operators, we obtain equivalent, Wegner-Wilson (WW) loop expressions as a product of various $X$ operators along the boundary as illustrated in Fig. \ref{fig:tx} (b) and (d). We can further decompose them into products of four line operators as
\begin{align}
W_{e_x , m} & = T_x (x_0\rightarrow x_1|y_0) T_x (y_0\rightarrow y_1|x_1 ) \left[T_x (x_0\rightarrow x_1| y_1)\right]^{-1} \left[T_x (y_0\rightarrow y_1|x_0)\right]^{-1} , \nonumber \\
W_{e_y , m} & = T_y (x_0\rightarrow x_1|y_0) T_y (y_0\rightarrow y_1|x_1 ) \left[T_y (x_0\rightarrow x_1| y_1)\right]^{-1} \left[T_y (y_0\rightarrow y_1|x_0)\right]^{-1} .
\label{eq:weg-t-decompose}
\end{align}
For instance, $T_x (x_0 \rightarrow x_1 |y_0)$ involves the product of operators at $x=x_0$ through $x=x_1$ for a fixed $y=y_0$. Specifically,
\begin{align}
& T_x (x_1\rightarrow x_2|y) \!=\! \prod_{x=x_1}^{x_2-1} \left( X_{1,i} \right)^{-1} \left( X_{1,i}X_{1,i \!-\! \hat{y} }^{-1}\right) ^{y-y_{x0}}, &
& T_x (y_1\rightarrow y_2|x) \!=\! \prod_{y=y_1}^{y_2-1} \left( X_{0,i} X_{0,i \!-\!\hat{y}}^{-1} X_{2,i}^{-1} X_{2,i\!-\!\hat{x}} \right)^{y\!-\!y_{x0}}, \nonumber \\
& T_y (x_1\rightarrow x_2|y) \!=\! \prod_{x=x_1}^{x_2-1} \left( X_{0,i} X_{0,i \!-\!\hat{x}}^{-1} X_{1,i}^{-1} X_{1,i\!-\!\hat{y}} \right)^{x\!-\!x_{y0}}, &
& T_y (y_1\rightarrow y_2|x) \!=\! \prod_{y=y_1}^{y_2-1} \left( X_{2,i} \right)^{-1} \left( X_{2,i}X_{2,i \!-\! \hat{x} }^{-1}\right) ^{x-x_{y0}}.
\label{eq:e-translate}
\end{align}
The shape of the boundary can be relaxed. For any type of boundary made by putting together several rectangles of arbitrary sizes, one still finds complete cancellation of operators except those at the boundary, which can be decomposed into the products of $T$ operators given in Eq. (\ref{eq:e-translate}).
\begin{figure*}[tb]
\includegraphics[width=0.99\textwidth]{./figures5-bdf/prl-tx.eps}
\caption{Graphical illustration for (a) $T_x(x_0 \rightarrow x_1|y_0)$ and (b) $T_x(y_0 \rightarrow y_1|x_1)$. The navy line represents the domain of $T_x$. The cyan square represents $e_x$ monopole with charge $\pm 1$ together with other charges comprising the auxiliary dipoles. The orange square represents the other pair that make up the auxiliary dipole. The auxiliary dipoles are connected by the gray line.
}
\label{fig:tx}
\end{figure*}
The first operator in $W_{e_x,m}$ is $T_x (x_0 \rightarrow x_1 | y_0 )$, which is illustrated in Fig. \ref{fig:tx}(a). It reduces to $\prod_{x=x_0}^{x_1-1} \left( X_{1,i} \right)^{-1}$ when $y_0 -y_{x0} ~{\rm mod}~ N = 0$ and creates $e_x$ monopoles with charge $-1$ at $(x_0,y_0)$ and charge $+1$ at $(x_1,y_0)$. In other words, the operator translates the $e_x$ monopole of charge +1 from $( x_0 , y_0) $ to $(x_1 , y_0)$. For $y_0-y_{x0} ~{\rm mod}~ N \neq 0$, we need to take account of the action by the auxiliary factor $\left( X_{1,i}X_{1,i \!-\! \hat{y} }^{-1}\right) ^{y_0-y_{x0}}$ as well, which is the creation of a pair of $e_x$ dipoles. One of the dipoles created consists of two $e_x$ monopoles with charges $(y_0-y_{x0})$ at $(x_0,y_0)$ and $-(y_0-y_{x0})$ at $(x_0, y_0-1)$. The other dipole consists of two $e_x$ monopoles with charges $-(y_0-y_{x0})$ at $(x_1,y_0)$ and $(y_0-y_{x0})$ at $(x_1,y_0-1)$.
We next look into the action of $T_x (y_0\rightarrow y_1| x_1)$ operator, which is illustrated in Fig. \ref{fig:tx}(b). First of all one can write $T_x (y_0\rightarrow y_1| x_1) = T_x (y_0 \rightarrow y_{x0} | x_1 ) T_x (y_{x0} \rightarrow y_1 | x_1 )$. When both $y_0-y_{x0} ~{\rm mod}~ N = 0$ and $y_1-y_{x0} ~{\rm mod}~ N =0$, one can think of it as a product of two translation operators of $e_x$, namely $T_x (y_0 \rightarrow y_{x0} | x_1 )$ and $T_x (y_{x0} \rightarrow y_1 | x_1 )$, moving $e_x$ from $(x_1 , y_0)$ to $(x_1 , y_{x0})$ and then subsequently from $(x_1 , y_{x0})$ to $(x_1, y_1)$. When $y_0-y_{x0} ~{\rm mod}~ N \neq 0$ or $y_1-y_{x0} ~{\rm mod}~ N \neq 0$, either the action of $T_x (y_0 \rightarrow y_{x0} | x_1 )$ or $T_x (y_{x0} \rightarrow y_1 | x_1 )$ is accompanied by the creation of an auxiliary dipole near $(x_1,y_0)$ or $(x_1,y_1)$, respectively. We can understand the action of $T_x$ and its inverse $T_x^{-1}$ as performing the translation of $e_x$ monopole while either creating or annihilating an auxiliary dipole. At the end of the $W_{e_x,m}$ operation, all the auxiliary dipoles disappear and the $e_x$ monopole has completed a loop.
By following the same procedure, one can interpret $T_y$ as the translation operator moving the $e_y$ monopole from one point to another and with or without the accompanying dipole depending on the modality of the coordinates with respect to $N$. The auxiliary dipoles are aligned in the $x$-direction and disappear at the completion of the loop.
\begin{figure*}[tb]
\includegraphics[width=0.99\textwidth]{./figures5-bdf/prl-e-detect1-2.eps}
\caption{
Graphical illustrations for the operators (a) $W_{m,e_x}$ and (b) $W_{m,e_y}$ in Eq. (\ref{eq:weg-loop-m}). The boundary of Wegner-Wilson operators are expressed by the navy lines.
}
\label{fig:weg-loop-e}
\end{figure*}
The operators $W_{m,e_x}$ and $W_{m,e_y}$ are defined as
\begin{align}
W_{m,e_x} = & \prod_{i \in {\cal A} } \left(a_{i}^x \right)^{y_i - y_{m0}}, &
W_{m,e_y} = & \prod_{i \in {\cal A} } \left(a_{i}^y \right)^{x_{m0} -x_i}.
\label{eq:weg-loop-m}
\end{align}
In the case of the rectangular boundary ${\cal A} = [x_0, x_1] \times [y_0 , y_1 ]$, $W_{m,e_x}$ and $W_{m,e_y}$ are expressed in Fig. \ref{fig:weg-loop-e}. The presence of $Z_0$ operators in the interior of the boundary makes it impossible to decompose $W_{m,e_x}$ and $W_{m,e_y}$ into the product of line operators as was the case in $W_{e_x,m}$ and $W_{e_y,m}$. On the other hand, on inspecting Fig. \ref{fig:weg-loop-e}, one realizes that the $Z_0$ operators in the interior cancel out by multiplying $W_{m,e_x}$ and $W_{m,e_y}$. Hence, we can decompose $W_{m,e} \equiv W_{m,e_x}W_{m,e_y}$ as the product of line operators:
\begin{align}
W_{m,e} & = T_m (x_0 \rightarrow x_1 | y_0) T_m (y_0 \rightarrow y_1 | x_1 ) \left[T_m (x_0 \rightarrow x_1 | y_1) \right]^{-1} \left[ T_m (y_0 \rightarrow y_1 | x_0 ) \right]^{-1} .
\label{eq:weg-loop-m2}
\end{align}
Here, the $T_m$ operator is defined by
\begin{align}
&T_m (x_1 \rightarrow x_2 | y) = \prod_{x=x_1\!+\!1}^{x_2} \left(Z_{2,i} \right)^{x - x_{m0}} \left( Z_{0,i-\hat{x}}\right)^{y_{m0}-y } , &
& T_m(y_1 \rightarrow y_2 | x) = \prod_{y=y_1\!+\!1}^{y_2}\left( Z_{1,i} \right)^{y-y_{m0} } \left(Z_{0,i-\hat{y}} \right)^{x_{m0}-x } .
\label{eq:m-translate}
\end{align}
The complete cancellation of $Z_0$ takes place for other shapes of the boundary as well, and $W_{m,e}$ can always be decomposed into the product of $T_m$'s.
\begin{figure*}[tb]
\includegraphics[width=0.98\textwidth]{./figures5-bdf/prl-tm.eps}
\caption{Graphical illustration for (a) $T_m(x_0 \rightarrow x_1|y_0)$ and (b) $T_m(y_0 \rightarrow y_1|x_1)$. The navy line represents the domain of $T_x$. The cyan square represents $e_x$ monopole with charge $\pm 1$ together with other charges composing the auxiliary dipole. The orange square represents the other pair that make up the auxiliary dipole. The auxiliary dipoles are connected by the gray line.
}
\label{fig:tm}
\end{figure*}
The first operator in $W_{m,e}$ is $T_m (x_0 \rightarrow x_1 | y_0)$, which is illustrated in Fig. \ref{fig:tm}(a). Its role is to translate the $m$ monopoles from $(x_0,y_0)$ to $(x_1,y_0)$ and creating the auxiliary dipoles depending on the modality of the coordinates with respect to $N$. When $y_0-y_{m0} ~{\rm mod}~ N \neq 0$, two auxiliary $m$ dipoles aligned in the $y$-direction are created near $(x_0,y_0)$ and $(x_1,y_0)$, respectively. The dipole near the point $(x_0,y_0)$ has dipole moment $y_0-y_{m0}$, and the other dipole near the point $(x_1,y_0)$ has dipole moment $-(y_0-y_{m0})$. When $x_0-x_{m0} ~{\rm mod}~ N \neq 0$, $T_m (x_0 \rightarrow x_1 | y_0)$ creates an $x$-directed auxiliary dipole near $(x_0,y_0)$ with the dipole moment $x_0-x_{m0}$. When $x_1-x_{m0} ~{\rm mod}~ N \neq 0$, $T_m (x_0 \rightarrow x_1 | y_0)$ creates an $x$-directed auxiliary dipole near $(x_1,y_0)$ with the dipole moment is $-(x_1-x_{m0})$. In contrast to $T_x$ or $T_y$ creating the auxiliary dipole only with the $x$- or $y$-component dipole moment, the auxiliary dipole created by $T_m$ has both $x$- and $y$-components and it consists of three monopoles. For example, the auxiliary dipole created by $T_m (x_0 \rightarrow x_1 | y_0)$ near $(x_1,y_0)$ consists of three monopoles with charge $-(y_0-y_{m0})$ at $(x_1,y_0+1)$, charge $-(x_1-x_{m0})$ at $(x_1+1,y_0)$, and charge $(y_0-y_{m0})+(x_1-x_{m0})$ at $(x_1,y_0)$.
The graphical illustration of $T_m(y_0\rightarrow y_1|x_1)$ is given in Fig. \ref{fig:tm}(b). After the application of $T_m (x_0 \rightarrow x_1 | y_0)$ to the vacuum, applying $T_m(y_0\rightarrow y_1|x_1)$ in succession translates the $m$ monopole with charge $+1$ and the auxiliary dipole from $(x_1,y_0)$ to $(x_1,y_1)$. Subsequently, applying $\left[T_m (x_0 \rightarrow x_1 | y_1) \right]^{-1}$ and $\left[T_m (y_0 \rightarrow y_1 | x_0 ) \right]^{-1}$ moves the $m$ monopoles and the auxiliary dipole from $(x_1,y_1)$ to $(x_0,y_1)$ to $(x_0,y_0)$, which completes the entire loop. In summary, during the operation of $W_{m,e}$, the consecutive $T_m$'s operation braids the $m$ monopole and the accompanied auxiliary dipole around the boundary of the $W_{m,e}$ operator.
\section{Gauge transformations of emergent gauge fields}
\label{sec:gauge-transf}
Under the gauge transformation
\begin{align}
A^{xx}\rightarrow & A^{xx} + \partial_x f_x, &
A^{xy}\rightarrow & A^{xy} + \partial_x f_y \!+\! \partial_y f_x, &
A^{yy}\rightarrow & A^{yy} + \partial_y f_y, \nonumber \\
E^{xx}\rightarrow & E^{xx} + \partial_y^2 g, &
E^{xy}\rightarrow & E^{xy} - \partial_x \partial_y g, &
E^{yy}\rightarrow & E^{yy} + \partial_x^2 g,
\end{align}
the emergent vector potentials transform as ${\bm a}_x \rightarrow {\bm a}_x + {\bm \nabla} F_x$ and ${\bm a}_y \rightarrow {\bm a}_y + {\bm \nabla} F_y$ where
\begin{align}
F_x &= -f_x + (y-y_{x0} )\partial_{y}f_x ,
&
F_y &= f^y - (x-x_{y0} ) \partial_{x}f_y,
\end{align}
and ${\bm a}_m \rightarrow {\bm a}_m + {\bm \nabla} G $ where
\begin{align}
G = (x - x_{m0}) \partial_x g + (y-y_{m0}) \partial_y g - g.
\end{align}
\section{Conservation Laws for $m$, $e_x$, and $e_y$ quasiparticles}
\label{sec:conservation}
The continuity equation for $m$ particles is given by
\begin{align}
\partial_t {\rho}^m + \partial_x^2 J_m^{xx} +\partial_x \partial_y J_m^{xy} + \partial_{y}^2 J_m^{yy} & = 0.
\label{eq:continuity-m}
\end{align}
Assuming vanishing currents at infinity, one can derive the charge conservation law:
\begin{align}
\frac{d Q^m }{dt} = \int d^2 {\bm r} \partial_t \rho^m = - \int d^2 {\bm r} \left[
\partial_x^2 J_m^{xx} +\partial_x \partial_y J_m^{xy} + \partial_{y}^2 J_m^{yy} \right] = 0.
\end{align}
Since second-order derivatives of the currents are involved, we find additional conservation law for the dipole moment ${\bm \mu}^m$:
\begin{align}
\frac{d {\bm \mu}^m }{dt} = \int d^2 {\bm r} ~ {\bm r} \partial_t \rho^m
= - \int d^2 {\bm r} ~ {\bm r} \left[\partial_x^2 J_m^{xx} +\partial_x \partial_y J_m^{xy} + \partial_{y}^2 J_m^{yy} \right] = 0.
\end{align}
The $e_x$ and $e_y$ quasiparticles obey different kinds of continuity equations:
\begin{align}
\partial_t {\rho}^{x}+ \partial_x J_e^{xx} + \partial_y J_e^{xy} & = 0, &
\partial_t {\rho}^{y}+ \partial_x J_e^{xy} + \partial_y J_e^{yy} & = 0.
\label{eq:continuity-e}
\end{align}
The total charge of $e_x$ and $e_y$ quasiparticles are conserved individually. The dipole moment of the individual quasiparticle is not conserved, but the particular combination $\mu^e = \int d^2 {\bm r} \left[x \rho^y - y \rho^x \right]$ is:
\begin{align}
\frac{d \mu^e }{dt} = - \int d^2 {\bm r} \left[ x (\partial_x J^{xy}_e +\partial_y J^{yy}_e ) - y ( \partial_x J^{xx}_e + \partial_y J^{xy}) \right] = - \int d^2 {\bm r} \left[ x \partial_x J^{xy}_e - y \partial_y J^{xy}_e \right] = 0.
\end{align}
\section{Auxiliary dipole and continuity equations}
\label{sec:aux-dipole}
In this section, we show how to write the charge densities of $e_x$, $e_y$, and $m$ quasiparticles in the presence of their auxiliary dipoles in the continuum theory, and solve the appropriate continuity equations to find the corresponding current densities that match the charge densities. In Sec. \ref{sec:emergent-gauge-derivation} we have shown that in the spin system, the translation of $e_x$, $e_y$, and $m$ monopoles are accompanied by the auxiliary dipoles. The auxiliary dipole moment for $e_x$ ($e_y$) varies in a way that conserves the $y$-component ($x$-component) of the total dipole moment as the $e_x$ ($e_y$) monopole is translated. On the other hand, the auxiliary dipole for $m$ varies in a way to conserve both the $x$- and $y$-component for the $m$ monopole and the accompanying dipole.
First, consider the $m$ quasiparticle with the unit charge moving around by the applications of $T_m$ operators. When the $m$ particle is translated from the initial point ${\bm r}_{m0} = (x_{m0},y_{m0})$ to some point ${\bm r}_m = (x_m,y_m)$, it is accompanied by the auxiliary dipole. The auxiliary dipole, according to the lattice calculation, consists of three $m$ monopoles with charge $-(x_m-x_{m0})$ at $(x_m+1,y_m)$, charge $(x_m-x_{m0})+(y_m-y_{m0})$ at $(x_m,y_m)$, and charge $-(y_m-y_{m0})$ at $(x_m,y_m+1)$. The combined dipole moment of the auxiliary dipole is
\begin{align}
{\bm \mu}^m = - ( x_m-x_{m0} , y_m -y_{m0} ) ,
\label{eq:aux_dm_em}
\end{align}
which precisely compensates for the dipole moment incurred by the motion of the $m$ particle from ${\bm r}_{m0}$ to ${\bm r}_m$.
To take the continuum limit we switch the lattice constant from 1 to $\Delta$ and take the limit $\Delta \rightarrow 0$. The positions of three monopolse composing the auxiliary dipole become $(x_m+\Delta,y_m)$, $(x_m,y_m)$, and $(x_m,y_m+\Delta)$. The monopole charges are modified accordingly to $-\frac{1}{\Delta}(x_m-x_{m0})$, $\frac{1}{\Delta}(x_m-x_{m0})+\frac{1}{\Delta}(y_m-y_{m0})$, and $-\frac{1}{\Delta}(y_m-y_{m0})$. We can write the charge density for the auxiliary dipole as
\begin{align}
\rho_{{\bm \mu}^m}({\bm r}) =& \frac{\left(x_m\!-\!x_{m0}\right)}{\Delta} \Big[ -\! \delta\left(x\!-\!x_m\!-\!\Delta\right)\delta\left( y\!-\!y_m\right) \!+\! \delta\left(x\!-\!x_m \right)\delta\left( y\!-\!y_m\right) \Big] \nonumber \\
&+\!
\frac{\left(y_m\!-\!y_{m0}\right)}{\Delta}\Big[ -\!\delta\left(x\!-\!x_m\right)\delta\left( y\!-\!y_m\!-\!\Delta\right)
\!+\!\delta\left(x\!-\!x_m\right)\delta\left( y\!-\!y_m\right) \Big],
\label{eq:rho-m-1}
\end{align}
which becomes, in the $\Delta\rightarrow0$ limit,
\begin{align}
\rho_{{\bm \mu}^m}({\bm r}) =& (x_m -x_{m0}) \partial_x \delta(x-x_m ) \delta(y-y_m ) + (y_m -y_{m0}) \delta(x-x_m ) \partial_y \delta(y-y_m ) \nonumber \\
=& ( {\bm r}_m - {\bm r}_{m0} ) \cdot {\bm \nabla} \delta^2({\bm r}-{\bm r}_m ),
\end{align}
where $\delta^2({\bm r}-{\bm r}_m ) = \delta(x-x_m ) \delta(y-y_m )$.
Together with the original $m$ monopole density $\rho_{m}^0 ({\bm r}) =\delta^2 ({\bm r} - {\bm r}_m )$, the net density becomes
\begin{align}
\rho_{m}({\bm r}) = \rho_{m}^0({\bm r}) + \rho_{{\bm \mu}^m}({\bm r}) = \delta^2({\bm r}-{\bm r}_m )+ ( {\bm r}_m - {\bm r}_{m0} ) \cdot {\bm \nabla} \delta^2({\bm r}-{\bm r}_m ).
\end{align}
When the position ${\bm r}_m$ varies over time, the time derivative of the total density $\rho_{m}$ becomes
\begin{align}
\partial_t \rho_{m}({\bm r}) =& - \dot{x}_m \left[(x_m -x_{m0}) \partial_x^2 \delta^2({\bm r}-{\bm r}_m) + (y_m - y_{m0}) \partial_x \partial_y \delta^2({\bm r}-{\bm r}_m ) \right] \nonumber \\
&- \dot{y}_m \left[(x_m -x_{m0}) \partial_x \partial_y \delta^2({\bm r}-{\bm r}_m ) + (y_m -y_{m0}) \partial_y^2 \delta^2({\bm r}-{\bm r}_m ) \right].
\label{eq:rho_m_t_dev}
\end{align}
By entering this expression to the continuity equation in Eq. (\ref{eq:continuity-m}), one can deduce the accompanying current density as
\begin{align}
J_m^{xx} & = \dot{x}_m (x_m-x_{m0}) \delta^2({\bm r} - {\bm r}_m), \nonumber \\
J_m^{xy} & = \left(\dot{x}_m (x_m-x_{m0}) + \dot{y}_m (y_m-y_{m0}) \right)\delta^2({\bm r} - {\bm r}_m), \nonumber \\
J_m^{yy} &= \dot{y}_m (y_m-y_{m0}) \delta^2({\bm r} - {\bm r}_m).
\label{eq:current-m}
\end{align}
When the $e_x$ monopole is translated from ${\bm r}_{x0} = (x_{x0},y_{x0})$ to ${\bm r}_x = (x_x, y_x)$, it is accompanied by an auxiliary dipole consisting of $e_x$ monopole with charge $(x_x-x_{x0})$ at $(x_x,y_x-1)$ and charge $-(x_x-x_{x0})$ at $(x_x,y_x)$. By following the same procedure for taking the continuum limit as before, we obtain the charge density of $e_x$ particle as
\begin{align}
\rho_x = \rho_x^0 + \rho_{\mu^x} = \delta^2({\bm r}-{\bm r}_x) + (y_x-y_{x0}) \partial_y \delta^2({\bm r}-{\bm r}_x) , ~~ \rho_y = 0,
\end{align}
where $\rho_x^0 ({\bm r}) = \delta^2 ({\bm r}-{\bm r}_x)$ is the $e_x$ monopole density, and $\rho_{\mu^x}$ is the density of the auxiliary dipole. Note that we need to consider the density of $e_y$ monopole simultaneously, $\rho_y=0$, since it is coupled to $\rho_x$ by the continuity equations in Eq. (\ref{eq:continuity-e}).
The accompanying current densities are deduced from the continuity equations as
\begin{align}
J^{xx}_e =& \dot{x}_x \left[ 1+(y_x-y_{x0}) \partial_y \right] \delta^2({\bm r}-{\bm r}_x) , \nonumber \\
J^{xy}_e =& \dot{y}_x (y_x -y_{x0}) \partial_y \delta^2({\bm r}-{\bm r}_x) , \nonumber \\
J^{yy}_e =& -\dot{y}_x (y_x -y_{x0}) \partial_x \delta^2({\bm r}-{\bm r}_x).
\label{eq:current-ex}
\end{align}
Similarly, when the monopole $e_y$ is translated from ${\bm r}_{y0} = (x_{y0},y_{y0})$ to ${\bm r}_y = (x_y,y_y)$, it is accompanied by the auxiliary dipole. The charge density of $e_y$ in the continuum limit is given by
\begin{align}
\rho_y = \rho_y^0 + \rho_{\mu^y} = \delta^2({\bm r}-{\bm r}_x) + (x_y -x_{y0}) \partial_x \delta^2({\bm r}-{\bm r}_x) , ~~ \rho_x = 0.
\end{align}
Again, we consider $\rho_x$ simultaneously since it is coupled with $\rho_y$ by the continuity equations in (\ref{eq:continuity-e}). The proper current densities that one can deduce from entering the above expression to the continuous equations in Eq. (\ref{eq:continuity-e}) are given by
\begin{align}
J^{xx}_e =& -\dot{x}_y (x_y -x_{y0}) \partial_y \delta^2({\bm r}-{\bm r}_y), \nonumber \\
J^{xy}_e =& \dot{x}_y (x_y -x_{y0}) \partial_x \delta^2({\bm r}-{\bm r}_y) , \nonumber \\
J^{yy}_e =& \dot{y}_y \left[ 1+(x_y-x_{y0}) \partial_x \right] \delta^2({\bm r}-{\bm r}_y) .
\label{eq:current-ey}
\end{align}
\section{Geometrical phases}
\label{sec:geo-phase}
Inserting the current densities obtained in the previous section to the dipolar-BF (dBF) action, we recover the phase of the dipolar braiding statistics. First, we consider a $m$ monopole and an auxiliary dipole making a round. Inserting the current density obtained in Eq. (\ref{eq:current-m}) to the dBF action, we get
\begin{align}
\phi &= \frac{2\pi}{N}\int dt \int d^2 {\bm r} \left[E^{xx}J_m^{yy} - E^{xy} J_m^{xy}+E^{yy}J_m^{xx} \right] \nonumber \\
&= \frac{2\pi}{N}\int dt \Big[\dot{x}_m \left[ (x_m - x_{m0} ) E^{yy} - (y_m -y_{m0} ) E^{xy} \right] + \dot{y}_m \left[ (y_m - y_{m0} ) E^{xx} - (x_m- x_{m0} ) E^{xy}\right] \Big] \nonumber \\
&= \frac{2\pi}{N} \oint dx_m \left[ (x_m - x_{m0} ) E^{yy} - (y_m -y_{m0} ) E^{xy} \right] + \frac{2\pi}{N} \oint dy_m \left[ (y_m - y_{m0} ) E^{xx} - (x_m- x_{m0} ) E^{xy}\right] \nonumber \\
& = \oint d{\bm r}_m \cdot {\bm a}_m ({\bm r}_m ) = \phi_{m, e} .
\end{align}
Similarly, when the $e_x$ monopole and the accompanying auxiliary dipole moves around, the relevant phase factor is obtained by entering the current density in Eq. (\ref{eq:current-ex}) to the dBF action:
\begin{align}
\phi = & -\int dt \int dxdy \left[ A^{xx} J_e^{xx} + A^{xy} J_e^{xy} + A^{yy} J_e^{yy} \right] \nonumber \\
=& \int dt \Big[ \dot{x}_x\left[-A^{xx} + (y_x - y_{x0}) \partial_{y_x} A^{yy} \right] + \dot{y}_x (y_x-y_{x0}) \left(\partial_{y_x} A^{xy} - \partial_{x_x} A^{yy} \right)\Big] \nonumber \\
= & \oint dx_y \left[-A^{xx} + (y_x - y_{x0}) \partial_{y_x} A^{yy} \right]+ \oint dy_x (y_x-y_{x0}) \left[\partial_{y_x} A^{xy} - \partial_{x_x} A^{yy} \right] \nonumber \\
=& \oint d{\bm r}_x \cdot {\bm a}_x({\bm r}_x) = \phi_{e_x,m}.
\end{align}
Finially, when the $e_y$ monopole and the accompanying auxiliary dipole makes a round, the phase factor obtained by entering the current density in Eq. (\ref{eq:current-ey}) to the dBF action is
\begin{align}
\phi = & -\int dt \int dxdy \left[ A^{xx} J_e^{xx} + A^{xy} J_e^{xy} + A^{yy} J_e^{yy} \right] \nonumber \\
=& \int dt \Big[ \dot{x}_y (x_y - x_{y0} ) \left[ - \partial_{y_y} A^{xx} + \partial_{x_y} A^{xy} \right] + \dot{y}_y \left(-A^{yy} + \left(x_y-x_{y0} \right) \partial_{x_y} A^{yy} \right)\Big] \nonumber \\
= & \oint dx_y (x_y - x_{y0} ) \left[ - \partial_{y_y} A^{xx} + \partial_{x_y} A^{xy} \right] + \oint dy_y \left(-A^{yy} + \left(x_y-x_{y0} \right) \partial_{x_y} A^{yy} \right) \nonumber \\
=& \oint d{\bm r}_y \cdot {\bm a}_y({\bm r}_y) = \phi_{e_y,m}.
\end{align}
In all three cases, we nicely recover the dipolar braiding phases that were obtained by other considerations, verifying the validity of the dBF action as the correct description of the dipolar braiding.
|
\section{Introduction}
\label{sec:intro}
\begin{figure*}[t]
\scriptsize
\centering
\includegraphics[width=1\linewidth]{figures/fig_teaser.pdf}\\
\caption{Top: illustration of top-down methods~\cite{neverova2020continuous} early commitment and overlapping ambiguity issues (red boxes). Bottom: data sparsity, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot the ratio of human pixels over the whole image pixels.
}
\label{fig:teaser}
\end{figure*}
Human dense pose estimation aims to learn the correspondence between RGB images and 3D human model surfaces.
It is a fundamental and essential problem in human-centric analysis and synthesis applications,
such as 3D body shape estimation~\cite{xu2019denserac,zhang2020learning}, human pose transfer~\cite{grigorev2019coordinate,sarkar2020neural}, unselfie~\cite{ma2020unselfie}, and character animation~\cite{gafni2019vid2game,wang2018vid2vid}.
G{\"u}ler~\emph{et al}\onedot ~\cite{alp2018densepose}
densely map each person pixel of an `in the wild'' RGB image to a location of a 3D human model surface using a two-stage top-down framework based on Mask R-CNN~\cite{he2017mask}.
They sparsely annotate a human subset of the COCO dataset~\cite{lin2014microsoft} for training. Follow-up works also adopt the same top-down principle to further improve the performance via uncertainty modeling~\cite{neverova2019correlated}, multi-scale strategy~\cite{guo2019adaptive}, knowledge-transfer~\cite{wang2020ktn}, simulated data~\cite{zhu2020simpose}, or continuous embedding~\cite{neverova2020continuous}.
All these methods are built upon Mask R-CNN~\cite{he2017mask} framework and follow the two-stage top-down detect-then-segment pipeline.
In particular, these top-down methods first employ an object detector based on Faster R-CNN~\cite{RenHG017} to predict a bounding box for each person, and then crop and resize regions-of-interest (ROIs) feature maps into a fixed size for each person.
Such a top-down pipeline has several limitations.
(1) The bounding box causes early commitment, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot if the bounding box fails to detect the whole body, there is no path to recovery, as shown in Fig.~\ref{fig:teaser}. Furthermore, several people may exist in the same bounding box which may not well align with a human due to overlap ambiguity (also mentioned in~\cite{wang2020ktn}),
(2) The detected person patches are cropped and resized, thus the resolution per person is usually heavily compressed.
(3) Top-down methods cannot fully leverage the sharing computation mechanism of convolutional networks, and thus, their inference time
scales unfavorably with the number of instances
(see Fig.~\ref{fig:infer_time}).
In contrast, we propose an end-to-end direct method inspired by the success of direct instance segmentation methods~\cite{tian2020conditional,wang2020solo}, which is faster and more accurate under multi-person occlusion than previous methods.
In particular, we formulate the dense pose estimation task into two inter-related sub-tasks: global dense pose IUV representation
\footnote{The IUV representation, introduced by~\cite{alp2018densepose}, is an image-based UV map with multiple channels.}
estimation and instance segmentation.
Inferring IUVs in a direct approach avoids early commitment, overlapping ambiguity, and heavy resolution compression, as well as makes computational complexity less dependent on the instance number.
Moreover, a simple fully convolutional implementation (as experimented in~\cite{alp2018densepose}) produces inferior results due to the interference from multi-person feature normalization
and wastes much computation and GPU memory in the empty background area. To address this, we propose an instance-aware normalization (IAN) technique to improve the results and sparse residual FCN to save computation.
Furthermore, previous methods are designed for per-frame prediction and produce flicker on video sequences. To alleviate flickering, we propose a simple and effective 2D temporal smoothing scheme which naturally fits our direct method.
We make three contributions:
1) We propose a direct framework DDP for human dense pose estimation, which runs faster and detects humans better in multiple person overlap situations than top-down methods.
2) We propose an instance-aware normalization (IAN) technique to remove the interference from multi-person feature normalization and utilize sparse convolution to skip the background computation.
3) We introduce an effective 2D temporal-smoothing scheme tailored for dense 2D inference problems, which preserves the coherence of projected 2D shapes.
\section{Related Work}
\label{sec:related}
\myparagraph{Instance segmentation.}
Instance segmentation methods combine instance-level object detection and pixel-level semantic segmentation.
Most existing approaches can be categorized into top-down methods and bottom-up methods. Top-down methods~\cite{li2017fully,he2017mask,liu2018path,bolya2019yolact,huang2019mask,chen2019tensormask} operate in a detect-then-segment way, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot detect the bounding box and then segment the object within the box. Among these methods, Mask R-CNN~\cite{he2017mask} is the most widely known top-down framework and has seen improved variants~\cite{liu2018path,huang2019mask}.
Due to the high performance of Mask R-CNN, all prior dense pose methods build upon it.
Bottom-up methods~\cite{newell2017associative,de2017semantic,liu2017sgn} operate in a label-then-cluster way, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot learn pixel-level embeddings and then cluster them into groups.
Recently, some effective direct methods~\cite{tian2020conditional,wang2020solo,wang2020solov2}
solved instance segmentation in one stage, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot without a detect-then-segment or label-then-cluster strategy.
In their SOLO approach, Wang~\emph{et al}\onedot~\cite{wang2020solo} segment objects by locations, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot using instance location categorization to predict object center locations.
They further improved accuracy and speed in their improved SOLOv2 method~\cite{wang2020solov2}.
Tian~\emph{et al}\onedot propose CondInst~\cite{tian2020conditional}, which employs dynamic instance-aware networks to infer an instance mask from global feature maps directly.
In this work, we take inspiration from CondInst~\cite{tian2020conditional} and propose a new direct framework for dense pose estimation.
\myparagraph{Multi-person 2D pose estimation.}
Multi-person 2D pose estimation methods estimate the number of people, their positions, and their body keypoints (joints) from an image. Similar to instance segmentation methods, multi-person 2D pose estimation algorithms can be classified into top-down and bottom-up methods.
Top-down methods~\cite{papandreou2017towards,xiao2018simple,huang2017coarse,sun2019deep,wang2020graph,huang2020devil,zhang2020distribution} employ off-the-shelf person detectors to obtain a bounding box for each person and then apply a single-person pose estimator within each box.
Bottom-up methods~\cite{cao2017realtime,insafutdinov2016deepercut,insafutdinov2017arttrack,newell2017associative,kreiss2019pifpaf,jin2020differentiable,cheng2020higherhrnet} first locate all the body joints in one image, and group them into individual person instances during post-processing.
Both top-down and bottom-up methods require multiple steps to obtain the final keypoint detection results.
Recently, Tian~\emph{et al}\onedot propose DirectPose~\cite{tian2019directpose} to handle the multi-person pose estimation. They extend the anchor-free single-stage object detector FCOS~\cite{tian2019fcos}, with one new output branch for keypoint detection.
Our method DDP follows a similar design principle to avoid the limitations of top-down methods.
Note that, unlike sparse keypoints pose estimation, it is very challenging to formulate the dense correspondence prediction problem as a keypoint classification problem like~\cite{tian2019directpose}.
Thus, we predict the instance mask and global IUV separately.
\myparagraph{Dense human pose estimation.}
Some methods~\cite{bogo2016keep,kanazawa2018end} estimate pixel-wise correspondence to a 3D surface model by fitting a prior deformable surface model to the image via indirect supervision, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot through body silhouette and key-joints.
In~\cite{alp2018densepose}, G{\"u}ler~\emph{et al}\onedot propose to directly map each pixel of a person region in the RGB image to the 3D human surface.
The dense correspondence is represented in a chart-based format called IUV, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot 24 body surfaces plus one background category (25 classes in total). With each surface patch, the local correspondence is further represented by a local UV coordinate system. Therefore, their dense pose IUV representation has 25$\times$3=75 dimensions, which are then summarized into three dimensions according to the 25 classes.
\cite{alp2018densepose} introduces a two-stage DensePose-RCNN built upon Mask R-CNN~\cite{he2017mask} and collects a large-scale image-to-surface sparse mapping dataset for direct supervision. DensePose-RCNN shows promising results on in-the-wild data.
Following this pipeline, Guo~\emph{et al}\onedot~\cite{guo2019adaptive} propose a multi-scale method called AMANet with improved performance under scale variations.
Wang~\emph{et al}\onedot~\cite{wang2020ktn} improve dense pose estimation by utilizing external commonsense knowledge.
Neverova~\emph{et al}\onedot~\cite{neverova2019slim} propose to augment the data annotations with motion cues for performance improvement.
Zhu~\emph{et al}\onedot~\cite{zhu2020simpose} introduce a new synthetic dense human pose dataset, together with a new estimator using a domain adaptation strategy to achieve good performance on real-world data.
Recently, Neverova~\emph{et al}\onedot~\cite{neverova2020continuous} propose a more straightforward and universal representation, Continuous Surface Embeddings (CSE), to better represent dense correspondences and improve the performance.
These works all follow the two-stage top-down strategy that suffers from early commitment, overlap ambiguities, heavy resolution compression, and unfavorable runtime scaling with the number of instances.
In contrast, we propose a direct dense pose estimation framework with faster runtime and better resilience to the aforementioned challenges.
\section{Framework}
\label{sec:framework}
\begin{figure*}[t]
\scriptsize
\centering
\includegraphics[width=1\linewidth]{figures/fig_framework.pdf}\\
\caption{
Framework overview.
The ResNet-FPN based backbone is first used to extract a feature pyramid.
Then, the feature aggregation module is applied to aggregate the feature pyramid into a global feature representation. Such a global feature is then fed into the instance branch and global IUV branch, respectively, to estimate the instance-level masks and the global IUV representation.
Note that, for each instance, the Mask FCN weights are generated dynamically via a weight generator.
}
\label{fig:framework}
\end{figure*}
\begin{figure}[t]
\scriptsize
\centering
\includegraphics[width=1\linewidth]{figures/fig_sparse_block.pdf}\\
\caption{The sparse residual block contains three sparse sub-blocks. Each sparse sub-block consists of a sparse convolution layer~\cite{graham20183d}, an instance-aware normalization (IAN) layer, and a ReLU layer.}
\label{fig:sparse_block}
\end{figure}
\subsection{Overview}
\label{sec:overview}
We propose a direct framework to divide the multi-person dense pose estimation problem into two subtasks: instance prediction and global IUV dense pose representation prediction.
As illustrated in Fig.~\ref{fig:framework}, our framework starts with a feature extractor backbone - a ResNet network (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot ResNet-50) followed by a feature pyramid network (FPN)~\cite{lin2017feature}. Then, the extracted feature pyramid (1/4 to 1/32 scales) is aggregated via a feature aggregation (Sec.~\ref{sec:feat_agg}) module, and then fed into the instance branch (Sec.~\ref{sec:instance_branch}) and the global IUV branch (Sec.~\ref{sec:global_iuv_branch}), respectively. The instance branch aims to predict the instance-level information, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot instance mask $\boldsymbol{M}_{ins}$ and dense pose mask $\boldsymbol{M}_{dp}$, for each person. In
contrast, the global IUV branch aims to predict the dense pose IUV representation for the whole image. The final per-person dense pose estimation can be obtained simply by multiplying the dense pose mask $\boldsymbol{M}_{dp}$ with the global IUV representation in an element-wise way.
In addition, we also propose a 2D temporal-smoothing scheme (Sec.~\ref{sec:smoothing}) to alleviate the temporal jitters when dealing with video data.
\subsection{Feature aggregation module}
\label{sec:feat_agg}
To alleviate instance scale variation issue in our direct method, we adopt a feature pyramid aggregation module~\cite{kirillov2019panoptic} to aggregate features of different scales. Specifically, each FPN level is upsampled by convolutions and bilinear upsampling until it reaches 1/4 scale. These upscaled features are summed together as $\boldsymbol{X}_{agg}$ which are fed into our instance branch and global IUV branch, respectively.
\subsection{Instance branch}
\label{sec:instance_branch}
The instance branch is built on top of a direct instance segmentation method CondInst~\cite{tian2020conditional} whose core idea is to dynamically generate the specific mask FCN head parameters for each instance\footnote{Similar to CondInst~\cite{tian2020conditional}, there are centerness and box heads in our instance branch, but omitted in the figure for simplicity.}.
For an image with $K$ instances, $K$ different mask FCN heads will be dynamically generated, each containing the characteristics of its target instance in the filters.
In particular, as shown in Fig.~\ref{fig:framework} top,
the instance branch first processes the received aggregated features $\boldsymbol{X}_{agg}$ with a Down Sample Conv module to extract global instance features $\boldsymbol{X}_D$ and reduce the feature resolution to 1/8 scale for saving computation.
This Down Sample Conv module contains three convolution layers of stride 1 and one convolution layer with stride 2.
Moreover, $\boldsymbol{X}_D$ is combined with a map of the coordinates, which are relative coordinates from all the locations on $\boldsymbol{X}_D$ to the location (h, w) (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot where the filters of the mask head are generated). Then, the combination, termed as $\Tilde{\boldsymbol{X}}_D$, is sent to the mask FCN head, whose parameters $\theta_{h,w}$ is dynamically generated by the weight generator module, to predict the instance mask.
In other words, $\Tilde{\boldsymbol{X}}_D$ contains the global information for the whole image, while the parameters $\theta_{h,w}$ encode instance-aware characteristics (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot relative position, shape, and appearance).
Different from CondInst, our mask FCN head will produce the instance mask $\boldsymbol{M}_{ins}$ and dense pose mask $\boldsymbol{M}_{dp}$ jointly for each instance.
This joint learning design takes the advantages of the existing rich instance annotations for stabilizing training (explained in Sec.~\ref{sec:learning}), as well as provides instance masks for the Global IUV branch (explained in Sec.~\ref{sec:global_iuv_branch}).
Note that one can also adopt other advanced direct instance segmentation networks for our instance branch, like SOLO~\cite{wang2020solo} and SOLOv2~\cite{wang2020solov2}.
\subsection{Global IUV branch}
\label{sec:global_iuv_branch}
We aim to predict the dense pose IUV information globally where all IUVs of different people are represented in the same image plane.
However, merely performing convolution over the large background region is computationally quite expensive and wasteful. Therefore, we propose a background suppression operation followed by a sparse residual FCN module to only deal with the person region of interest. We also propose an instance-aware normalization (IAN) technique that performs better than normalizing all the instances together as done in~\cite{alp2018densepose}.
\myparagraph{Background suppression.}
As for the human dense pose estimation task, we are only interested in the people region, which usually occupies a small part of the whole image. For example, the sparsity (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot ratio of the people region over the whole image) on the DensePose-COCO dataset~\cite{alp2018densepose} is on average 23.7\% and ranges from 1.7\% to 90.9\%\footnote{Calculated on the instance annotations of DensePose-COCO minival split.}. Some sparsity examples are illustrated in Fig.~\ref{fig:teaser}.
These large background regions will waste a large amount of computation and interfere with the learning procedure as well.
Previous top-down multi-person dense pose estimation methods~\cite{alp2018densepose,guo2019adaptive,neverova2019correlated,wang2020ktn,zhu2020simpose,wu2019detectron2,neverova2020continuous} alleviate this issue by cropping the feature maps according to the detected bounding boxes and process them separately in a single person style.
Such a cropping-based approach is sub-optimal as it still introduces some background or content from other instances. Besides, this suffers from the early commitment issue.
In contrast, we propose a background suppression strategy to suppress the background interference explicitly.
Specifically, we first resize the estimated instance masks $\{\boldsymbol{M}_{ins}^i\}_{i=1,\dots,N}$ to the resolution of features $\boldsymbol{X}_{agg}$. The resized masks $\{\Tilde{\boldsymbol{M}}_{ins}^i\}_{i=1,\dots,N}$ are then combined into a foreground mask $\boldsymbol{M}_{fg}$ which is applied to mask the features in a point-wise manner as follows,
\begin{align}
\boldsymbol{X}_{fg}=\boldsymbol{M}_{fg} \odot \boldsymbol{X}_{agg},
~\text{where}~ \boldsymbol{M}_{fg}=\bigcup\limits_{i=1}^{N} \Tilde{\boldsymbol{M}}_{ins}^i,
\end{align}
where $\boldsymbol{X}_{fg}$ denote the masked foreground features.
\myparagraph{Sparse residual FCN.}
As mentioned above, computing the features on background regions is wasteful, especially considering that we need to maintain high-resolution features to achieve dense global IUV prediction. Therefore, we perform convolution in a sparse manner to reduce the computation and memory cost. Specifically, the dense features are first transformed into a sparse format and then processed by our sparse residual FCN module and finally transformed back to a dense format.
The sparse residual FCN module consists of three sparse residual blocks. As illustrated in Fig.~\ref{fig:sparse_block}, each sparse residual block contains two sparse sub-blocks with a residual skip connection followed by a third sparse sub-block (except for the last sparse residual block). Each sparse sub-block includes a sparse convolution layer, an instance-aware normalization (IAN) layer, and a ReLU layer.
Regarding the sparse convolution layer, we adopt the submanifold sparse convolution (SSC) layer~\cite{graham20183d}, which can fix the location of active sites and thus maintain the same level of sparsity throughout the network.
The instance-aware normalization (IAN) layer is described in the next paragraph.
\myparagraph{Instance-aware normalization.}
We propose the instance-aware normalization to perform feature normalization for each instance separately.
In particular, we integrate the resized instance masks $\{\Tilde{\boldsymbol{M}}_{ins}\}_{i=1,\dots,N}$ into our normalization operation as follows,
\begin{align}
\mu_{nci} &= \frac{1}{\text{Numel}(\Tilde{\boldsymbol{M}}_{ins}^i)} \sum\limits_{h,w \in \Tilde{\boldsymbol{M}}_{ins}^i} x_{nchw},
\label{eq:IAN_mu} \\
\sigma^2_{nci} &= \frac{1}{\text{Numel}(\Tilde{\boldsymbol{M}}_{ins}^i)} \sum\limits_{h,w \in \Tilde{\boldsymbol{M}}_{ins}^i} (x_{nchw}-\mu_{nci})^2,
\label{eq:IAN_sigma} \\
\hat{x} &= \bigcup\limits_{i=1}^{N} \bigcup\limits_{h,w \in \Tilde{\boldsymbol{M}}_{ins}^i} \Big( \frac{x_{nchw}-\mu_{nci}}{\sqrt{\sigma^2_{nci}+\epsilon}} \gamma + \beta \Big),
\label{eq:IAN}
\end{align}
where $x_{nchw}$ denotes the feature point at location (h,w) exists in $\Tilde{\boldsymbol{M}}_{ins}^i$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot the region of the $i$-th instance. Numel($\cdot$) means the point number in the region of the $i$-th instance.
We calculate the mean $\mu_{nci}$ and variance $\sigma^2_{nci}$ for each instance mask $\Tilde{\boldsymbol{M}}_{ins}^i$ individually, and then apply instance normalization~\cite{ulyanov2016instance} due to its better performance over other normalization techniques in our pilot experiment. Note that, all the instances share the same learnable affine parameters $\gamma$ and $\beta$ as they all belong to the person category.
\section{2D temporal-smoothing scheme}
\label{sec:smoothing}
Previous dense pose estimation methods are designed for image-based prediction because it is expensive to collect accurate dense pose annotations for video data. However, directly applying image-based methods to video data leads to the undesirable flickering issue, since the temporal information is not considered. To address this, we introduce a simple and effective 2D temporal-smoothing scheme, which fits well our global IUV representation.
The main idea is to use the temporal constraint from the original RGB video.
Specifically, given the present RGB frame $I_t$ and its temporally adjacent frames $\{{I_{t+j}}\}_{j=-r,\cdots,-1,1,\cdots,r}$,
we adopt an optical flow estimation model (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot RAFT~\cite{teed2020raft}) to predict the optical flows $\{f_{t\rightarrow t+j}\}_{j=-r,\cdots,-1,1,\cdots,r}$, which are used to warp the global IUV dense pose representation $\boldsymbol{C}$ of adjacent frames to one $\boldsymbol{C}_{temp}$ for the present frame using the following weighted sum:
\begin{align}
\boldsymbol{C}_{t+j \rightarrow t} &= \text{Warp}(\boldsymbol{C_{t+j}}, f_{t\rightarrow t+j}),
\label{eq:warp} \\
\boldsymbol{C}_{temp} &= \sum\limits_{j=-r}^r \alpha_j \boldsymbol{C}_{t+j \rightarrow t},
\label{eq:IUV_temp}
\end{align}
where $r$ is the temporal window interval and $\alpha_j$ is the sum weight. We set $r=2$ and $\{\alpha_j\} = [0.2,0.2,0.2,0.2,0.2]$ by default.
Here we perform the warping and weighted sum operations on the continuous logit-level, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\boldsymbol{C}$ is a 75-dim logit representation.
\section{Learning}
\label{sec:learning}
As mentioned before, we predict instance mask $\boldsymbol{M}_{ins}$ and dense pose mask $\boldsymbol{M}_{dp}$ jointly in our instance branch (Sec.~\ref{sec:instance_branch}).
We take this joint approach because of the limited dense pose annotation and the limited body coverage.
Regarding the annotation, only part of human instances was selected for dense pose annotation on the DensePose-COCO dataset. Training with annotated dense pose masks only will lead to inferior instance prediction performance.
Furthermore, the dense pose masks do not cover the whole person, leading to important information missing in our background suppression. (Sec.~\ref{sec:global_iuv_branch}).
Therefore, we jointly regress both instance masks and dense pose masks in our instance branch to improve the learning with rich instance annotations and suppress the background with instance masks.
Besides, we apply the ground truth instance masks for background suppression in our global IUV branch during training, considering that the estimated instance masks contain many errors in the early training stage. Such errors will break the learning of the global IUV branch.
\begin{table*}[t]
\centering
\setlength{\tabcolsep}{4.3pt}
\begin{tabular
{@{\extracolsep{\fill}} l| c c c| c c| c c c| c c|c}
\toprule
Method & AP & AP$_{50}$ & AP$_{75}$ & AP$_{M}$ & AP$_{L}$ & AR & AR$_{50}$ & AR$_{75}$ & AR$_{M}$ & AR$_{L}$ & time (s) \\
\midrule[0.6pt]
\multicolumn{11}{c}{Top-down methods} \\
\midrule[0.6pt]
DP-cascade (Res50)~\cite{alp2018densepose} & 51.6 & 83.9 & 55.2 & 41.9 & 53.4 & 60.4 & 88.9 & 65.3 & 43.3 & 61.6 & 0.583 \\
DP-cascade (Res50)+masks~\cite{alp2018densepose} & 52.8 & 85.5 & 56.1 & 40.3 & 54.6 & 62.0 & 89.7 & 67.0 & 42.4 & 63.3 & - \\
DP-cascade (Res50)+keypoints~\cite{alp2018densepose} & 55.8 & 87.5 & 61.2 & 48.4 & 57.1 & 63.9 & 91.0 & 69.7 & 50.3 & 64.8 & - \\
Parsing (Res50)~\cite{yang2019parsing} & 55.0 & 87.6 & 59.8 & 50.6 & 56.6 & - & - & - & - & - & 0.392 \\
Parsing (Res50)+keypoints~\cite{yang2019parsing} & 58.3 & 90.1 & 66.9 & 51.8 & 61.9 & - & - & - & - & - & - \\
Parsing (ResNeXt101)~\cite{yang2019parsing} & 59.1 & 91.0 & 66.9 & 51.8 & 61.9 & - & - & - & - & - & - \\
Parsing (ResNeXt101)+keypoints~\cite{yang2019parsing} & 61.6 & 91.6 & 72.3 & 54.8 & 64.8 & - & - & - & - & - & - \\
SimPose (Res50)*~\cite{zhu2020simpose} & 57.3 & 88.4 & 67.3 & 60.1 & 59.3 & 66.4 & 95.1 & 77.8 & 62.4 & 66.7 & - \\
AMA-Net (Res50)~\cite{guo2019adaptive} & 64.1 & 91.4 & 72.9 & 59.3 & 65.3 & 71.6 & 94.7 & 79.8 & 61.3 & 72.3 & - \\
KTN (Res50)~\cite{wang2020ktn} & 66.5 & 91.5 & 75.5 & 61.9 & 68.0 & 74.2 & 95.2 & 82.3 & \textbf{64.2} & \textbf{74.9} & 0.518 \\
DP R-CNN DeepLab (Res50)~\cite{neverova2020continuous} & 66.8 & 92.8 & \textbf{79.7} & 60.7 & 68.0 & 72.1 & 95.8 & 82.9 & 62.2 & 72.4 & 0.242 \\
DP R-CNN DeepLab (Res101)~\cite{neverova2020continuous} & \textbf{67.7} & \textbf{93.5} & \textbf{79.7} & \textbf{62.6} & \textbf{69.1} & \textbf{73.6} & \textbf{96.5} & \textbf{84.7} & \textbf{64.2} & 74.2 & 0.382 \\
\midrule[0.6pt]
\multicolumn{11}{c}{Direct methods} \\
\midrule[0.6pt]
Ours (Res50) & 64.0 & 92.4 & 76.0 & 57.2 & 65.7 & 70.9 & 96.4 & 82.4 & 59.9 & 71.6 & \textbf{0.209} \\
\bottomrule[1pt]
\end{tabular}
\caption{The quantitative results on DensePose-COCO minival split. * Models are trained with a simulated dataset.
}
\label{tab:comparisons}
\end{table*}
\myparagraph{Losses.}
The overall loss function of DDP is formulated as,
\begin{align}
L_{all} &= L_{mask} + L_{IUV},
\label{eq:loss_all} \\
L_{mask} &= L_{fcos} + \lambda_1 (L_{M_{ins}} + L_{M_{dp}})
\label{eq:loss_ins} \\
L_{IUV} &= L_I + \lambda_2 L_{UV} + \lambda_3 L_s
\label{eq:loss_IUV}
\end{align}
where $L_{mask}$ and $L_{IUV}$ denote the loss for instance branch and global IUV branch, respectively.
$\lambda_1=5$, $\lambda_2=10$, $\lambda_3=1$ are used to balance the losses.
Similar to CondInst,
our instance prediction loss $L_{mask}$ includes $L_{fcos}$ and two mask losses $L_{M_{ins}}$, $L_{M_{dp}}$.
We refer the reader to FCOS~\cite{tian2019fcos} for the details of $L_{fcos}$.
$L_{M_{ins}}$ and $L_{M_{dp}}$ are defined as,
\begin{align}
\resizebox{1\hsize}{!}{$
L_{M_{ins}} = \frac{1}{N_{pos}} \sum\limits_{h,w} \mathbbm{1}_{\{c^*_{h,w}>0\}} L_{dice} (\boldsymbol{M}^{h,w}_{ins},\boldsymbol{M}^{h,w*}_{ins}),
$}
\label{eq:mask_ins} \\
\resizebox{1\hsize}{!}{$
L_{M_{dp}} = \frac{1}{N_{pos}} \sum\limits_{h,w} \mathbbm{1}_{\{c^*_{h,w}>0\}} L_{dice} (\boldsymbol{M}^{h,w}_{dp},\boldsymbol{M}^{h,w*}_{dp}),
$}
\label{eq:mask_dp}
\end{align}
where $[\boldsymbol{M}^{h,w}_{ins},\boldsymbol{M}^{h,w}_{dp}]=\text{MaskHead}(\Tilde{\boldsymbol{X}}_D;\theta_{h,w})$ are the estimated instance and dense pose masks, and $\boldsymbol{M}^{h,w*}_{ins},\boldsymbol{M}^{h,w*}_{dp}$ are the corresponding ground truth.
The $c^*_{h,w}$ is the classification label of location (h, w), which is the class of the instance associated with the location or 0 (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot background) if the location is not
associated with any instance. $N_{pos}$ is the number of locations where $c^*_{h,w}>0$.
Our global dense pose IUV prediction loss $L_{IUV}$ contains one cross-entropy loss $L_I$ for body parts classification, one smooth L1-loss $L_{UV}$ for local UV coordinates regression, and one smoothing loss $L_s$.
The smoothing loss $L_s$ is used to encourage the model to produce less noisy IUV representation, since the dense pose annotation is a set of sparse points. In particular, we adopt an edge-aware smoothness regularization~\cite{godard2017unsupervised} as follows,
\begin{align}
L_s = \frac{1}{N} \sum\limits_{i=1}^N &|\nabla_h \boldsymbol{C}|e^{-|\nabla_h \boldsymbol{M}_{ins}^i|} \nonumber \\
+ &|\nabla_w \boldsymbol{C}|e^{-|\nabla_w \boldsymbol{M}_{ins}^i|},
\label{eq:loss_smooth}
\end{align}
where $\boldsymbol{C}$ is the predicted global IUV representation.
\section{Experiments}
\label{sec:exp}
For evaluation, we present qualitative and quantitative results, a runtime analysis, and user study results.
\begin{table*}[t]
\centering
\setlength{\tabcolsep}{3.7pt}
\begin{tabular
{@{\extracolsep{\fill}} l| c c c| c c| c c c| c c | c}
\toprule
Method & AP & AP$_{50}$ & AP$_{75}$ & AP$_{M}$ & AP$_{L}$ & AR & AR$_{50}$ & AR$_{75}$ & AR$_{M}$ & AR$_{L}$ & time (s) \\
\midrule[0.6pt]
Ours (Res50) w/o $L_s$, w/o IAN, w/o Sparse & 62.2 & 91.5 & 73.4 & 55.9 & 64.0 & 69.3 & 95.9 & 80.3 & 58.2 & 70.1 & 0.380 \\
Ours (Res50) w/o $L_s$, w/o IAN & 62.2 & 92.3 & 73.2 & 55.0 & 63.9 & 69.3 & 96.2 & 80.4 & 58.9 & 70.0 & 0.234 \\
Ours (Res50) w/o $L_s$ & 63.8 & 91.8 & 75.9 & 57.1 & \textbf{65.8} & \textbf{71.2} & 96.2 & \textbf{83.4} & 59.0 & \textbf{71.9} & 0.211 \\
\midrule[0.6pt]
Ours (Res50) & \textbf{64.0} & \textbf{92.4} & \textbf{76.0} & \textbf{57.2} & 65.7 & 70.9 & \textbf{96.4} & 82.4 & \textbf{59.9} & 71.6 & \textbf{0.209} \\
\bottomrule[1pt]
\end{tabular}
\caption{Ablation study and average inference time on DensePose-COCO minival split.
}
\label{tab:ablation}
\end{table*}
\begin{table}[t]
\centering
\setlength{\tabcolsep}{5.5pt}
\begin{tabular
{@{\extracolsep{\fill}} l c}
\toprule
Method & Human preference \\
\midrule[0.6pt]
DP R-CNN DeepLab (Res50)~\cite{neverova2020continuous} & 0.076 \\
Ours (Res50) w/ temporal smoothing & 0.924 \\
\midrule[0.6pt]
Ours (Res50) w/o temporal smoothing & 0.016 \\
Ours (Res50) w/ temporal smoothing & 0.984 \\
\bottomrule[1pt]
\end{tabular}
\caption{User study for temporal smoothing.
}
\label{tab:user_study}
\end{table}
\myparagraph{Implementation details.}
Unless specified otherwise, we use the following implementation details.
We use a ResNet-50~\cite{he2016deep} architecture as the backbone, followed by a 4-level FPN (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot 1/4, 1/8, 1/16, 1/32 levels). The ResNet weights are initialized by the pre-trained keypoints estimation models from COCO~\cite{lin2014microsoft}. Our models are trained with stochastic gradient descent
(SGD) for 130K iterations with an initial learning rate of 0.01 and a mini-batch of 8 images. The learning rate is reduced by a factor of 10 at iteration 100K and 120K, respectively. Weight decay and momentum are set as 0.0001 and 0.9, respectively.
Following~\cite{alp2018densepose}, two quantitative metrics are used for evaluation, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot
Average Precision (AP) and Average Recall (AR). Both metrics are calculated at a number of geodesic point similarity (GPS) ranging from 0.5 to 0.95. In addition, other evaluation metrics, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot AP$_M$ and AR$_M$ for medium people and AP$_L$ and AR$_L$ for large people, are also reported.
Our method is implemented based on the Detectron2 framework~\cite{wu2019detectron2}.
No data augmentation is used during training or testing.
Inference times are measurd on a single V100 GPU with one image per batch.
\myparagraph{Dataset.}
Our method is evaluated on DensePose-COCO dataset~\cite{alp2018densepose},
which has manually annotated correspondences on a subset of the COCO dataset~\cite{lin2014microsoft}.
There are about 50K labeled human instances
each of which is annotated with 100 points on average. In total, there are about 5 million manually annotated correspondences. The dataset is split into a training set and a validation set with 32K images and 1.5k images, respectively.
\subsection{Comparisons}
We evaluate the proposed end-to-end DDP framework on DensePose-COCO minival split and compare it with the SOTA top-down methods in Table~\ref{tab:comparisons}. Our method achieves 64.0\% AP and 70.9\% AR with ResNet-50.
The performance of our model is better than the strong baselines DP-cascade~\cite{alp2018densepose} (64.0\% vs. 55.8\% AP) and Parsing~\cite{yang2019parsing} (64.0\% vs. 61.6\% AP), and is comparable to AMA-Net~\cite{guo2019adaptive} (64.0\% vs. 64.1\% AP).
Our model is still behind the top-performing top-down methods KTN~\cite{wang2020ktn} and DP R-CNN DeepLab~\cite{neverova2020continuous}.
Nevertheless, it is worth noting that the top-down strategy suffers from the early commitment issue, overlap ambiguities as shown in Fig.~\ref{fig:teaser}.
Besides, top-down methods run slower in larger multi-person scenes and compress image resolution heavily.
\begin{figure}[t]
\scriptsize
\centering
\includegraphics[width=0.8\linewidth]{figures/fig_infer_time_two.pdf}
\caption{Inference time. X-axis: number of person instances on each image. Y-axis: average inference time in seconds for each image.
}
\label{fig:infer_time}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{figures/fig_smooth_ablation.pdf}\\
Ours (Res50) w/o $L_s$ ~~~~~~~~~~~~~~~~~~ Ours (Res50)~~~~~
\caption{Effectiveness of smoothing loss $L_s$.
}
\label{fig:smooth}
\end{figure}
\subsection{Ablation study and inference time}
We perform an ablation study for each component of DDP presented in Table~\ref{tab:ablation}. The ablative settings are as follows. \textbf{w/o $L_{smooth}$}: Removing the smoothing loss. \textbf{w/o IAN}: Removing the instance-aware normalization. \textbf{w/o Sparse}: We use full dense feature maps instead of sparse feature maps, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\boldsymbol{M_{fg}}$ is a tensor with all elements set to one.
We observe that the proposed IAN technique improves the results, and the sparse technique aids to significantly reduce the computation time.
Regarding the smoothing loss $L_s$, the quantitative results are almost the same, while the qualitative results are spatially smoothed. As shown in the isocontour visualization in Fig.~\ref{fig:smooth}, the isocontour becomes less noisy when applying the smoothing loss $L_s$.
Such smoothness is desirable because of the continuity nature of human body surface.
Furthermore, we also compare the inference time on images of different instance numbers with the state-of-the-art top-down method DP R-CNN DeepLab~\cite{neverova2020continuous} as shown in Fig.~\ref{fig:infer_time}.
We observe that our method's inference time increases more slowly with the number of people in the scene,
because our approach's inference time mainly depends on the image's sparsity.
\begin{figure*}[t]
\scriptsize
\centering
\includegraphics[width=1\linewidth]{figures/fig_qualitative_small.pdf}\\
\caption{Qualitative results on DensePose-COCO minival split. The red and yellow circles spot the failure cases (see Sec.~\ref{sec:limitation}).}
\label{fig:qualitative}
\end{figure*}
\subsection{Temporal smoothing evaluation}
\myparagraph{User study.} To demonstrate the effectiveness of our 2D temporal-smoothing scheme, we perform a user study on 11 YouTube video sequences of 20 seconds each.
As shown in Table~\ref{tab:user_study}, we ask 30 users to assess the temporal smoothing of the results of two comparisons: (1) our method versus the top-performing top-down method - DP R-CNN DeepLab (Res50)~\cite{neverova2020continuous} (row 1-2), and (2) our method versus our non-smoothing setting (row 3-4).
For each video sequence, given two randomly ordered video results of dense pose isocontour visualization from two methods, users are asked to pick one that looks temporally smoother than the other. The results in Table~\ref{tab:user_study} illustrate that our method is obviously preferred over the alternative.
\myparagraph{Quantitative evaluation.}
Furthermore, we also perform a quantitative evaluation measured with the most widely used metrics in the video stabilization field: Interframe Transformation Fidelity (ITF) index~\cite{marcenaro2001image} and Interframe Similarity Index (ISI)~\cite{guilluy2018performance}, which are based on the video inter-frame PSNR and SSIM scores, respectively.
We report the average scores (see Tab.~\ref{tab:temporal_eval}) of 11 isocontour visualization videos used in the user study. We can see that our temporal smoothing strategy improves the temporal smoothness on both evaluation metrics.
\myparagraph{Image-to-image translation.} We also apply our method on the human pose transfer task,
\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot translate the dense pose IUV representation input to an RGB image.
We adopt the popular pix2pixHD~\cite{wang2018high} as our translation model and generate the image frame-by-frame.
We observe that our temporally smooth dense pose can help alleviate video flickering issue and generate stable video results.
Please refer to our supplementary video for the visualization results.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{5.5pt}
\begin{tabular
{@{\extracolsep{\fill}} l c c}
\toprule
Model & ITF & ISI \\
\midrule[0.6pt]
DP R-CNN DeepLab (Res50)~\cite{neverova2020continuous} & 35.04 & 0.9202 \\
Ours (Res50) w/o temporal smoothing & 35.24 & 0.9253 \\
Ours (Res50) w/ temporal smoothing & \textbf{35.68} & \textbf{0.9294} \\
\bottomrule[1pt]
\end{tabular}
\caption{Quantitative evaluation for temporal smoothing.
}
\label{tab:temporal_eval}
\end{table}
\begin{figure}[t]
\centering
\scriptsize
\includegraphics[width=1\linewidth]{figures/rebuttal_compare3.pdf}
\caption{Results of~\cite{neverova2020continuous} (top) and Ours (bottom) with Res50. The red and blue circles spot our advantages.
}
\label{fig:compare}
\end{figure}
\section{Qualitative results and discussion}
\label{sec:limitation}
In Fig.~\ref{fig:compare}, we show the advantages of our method compared to the top-down methods~\cite{neverova2020continuous} of avoiding early commitment (red circles) and better boundaries (blue circles).
In Fig.~\ref{fig:qualitative}, we illustrate more isocontour visualization results. We observe that our DDP can predict smooth and accurate 3D body correspondences in diverse real-world scenarios exhibiting different challenges, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot illumination variations (indoor and outdoor), occlusions (self occlusion, inter occlusion, and background occlusion), diverse body poses and views.
Although our method achieves comparable performance to SOTA methods, it may produce noisy results for large-scale or occluded people (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot the red circle in Fig.~\ref{fig:qualitative}) and fail to detect some occluded small people (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot the yellow circle in Fig.~\ref{fig:qualitative}).
As for our 2D temporal-smoothing scheme, it can alleviate the temporal flickering for most cases, but may fail in textureless region or varying illumination where the optical flow estimation is not robust.
\section{Conclusion}
We present a direct human dense pose estimation method DDP that deals with the multi-person dense pose estimation via two inter-related sub-tasks: global IUV estimation and instance segmentation. The proposed method achieves comparable results to strong
top-down benchmark methods, and avoids issues like early commitment and overlap ambiguity. Our method also runs in weak instance number dependent time.
Furthermore, we introduce a temporal smoothing pose-processing which naturally fits the global IUV representation and enables temporal smoothing dense pose estimation. We believe such a temporal smoothing ability can benefit human analysis and synthesis tasks.
\section{Introduction}
Please follow the steps outlined below when submitting your manuscript to
the IEEE Computer Society Press. This style guide now has several
important modifications (for example, you are no longer warned against the
use of sticky tape to attach your artwork to the paper), so all authors
should read this new version.
\subsection{Language}
All manuscripts must be in English.
\subsection{Dual submission}
By submitting a manuscript to 3DV, the authors assert that it has not been
previously published in substantially similar form. Furthermore, no paper
which contains significant overlap with the contributions of this paper
either has been or will be submitted during the 3DV 2021 review period to
{\bf either a journal} or any conference or any
workshop.
{\bf Papers violating this condition will be rejected.}
If there are papers that may appear to the reviewers
to violate this condition, then it is your responsibility to: (1)~cite
these papers (preserving anonymity as described in Section 1.6 below),
(2)~argue in the body of your paper why your 3DV paper is non-trivially
different from these concurrent submissions, and (3)~include anonymized
versions of those papers in the supplemental material.
\subsection{Paper length}
3DV papers should be no longer than 8 pages, excluding references.
The references section will not be included in the page count, and
there is no limit on the length of the references section. Overlength
papers will simply not be reviewed. This includes papers where the
margins and formatting are deemed to have been significantly altered
from those laid down by this style guide. Note that this \LaTeX\
guide already sets figure captions and references in a smaller font.
The reason such papers will not be reviewed is that there is no
provision for supervised revisions of manuscripts. The reviewing
process cannot determine the suitability of the paper for presentation
in eight pages if it is reviewed in eleven.
\subsection{The ruler}
The \LaTeX\ style defines a printed ruler which should be present in the
version submitted for review. The ruler is provided in order that
reviewers may comment on particular lines in the paper without
circumlocution. If you are preparing a document using a non-\LaTeX\
document preparation system, please arrange for an equivalent ruler to
appear on the final output pages. The presence or absence of the ruler
should not change the appearance of any other content on the page. The
camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment
the \verb'\threedvfinalcopy' command in the document preamble.) Reviewers:
note that the ruler measurements do not align well with lines in the paper
--- this turns out to be very difficult to do well when the paper contains
many figures and equations, and, when done, looks ugly. Just use fractional
references (e.g.\ this line is $097.5$), although in most cases one would
expect that the approximate location will be adequate.
\subsection{Mathematics}
Please number all of your sections and displayed equations. It is
important for readers to be able to refer to any particular equation. Just
because you didn't refer to it in the text doesn't mean some future reader
might not need to refer to it. It is cumbersome to have to use
circumlocutions like ``the equation second from the top of page 3 column
1''. (Note that the ruler will not be present in the final copy, so is not
an alternative to equation numbers). All authors will benefit from reading
Mermin's description of how to write mathematics
\subsection{Blind review}
Many authors misunderstand the concept of anonymizing for blind
review. Blind review does not mean that one must remove
citations to one's own work---in fact it is often impossible to
review a paper unless the previous citations are known and
available.
Blind review means that you do not use the words ``my'' or ``our''
when citing previous work. That is all. (But see below for
techreports)
Saying ``this builds on the work of Lucy Smith [1]'' does not say
that you are Lucy Smith, it says that you are building on her
work. If you are Smith and Jones, do not say ``as we show in
[7]'', say ``as Smith and Jones show in [7]'' and at the end of the
paper, include reference 7 as you would any other cited work.
An example of a bad paper just asking to be rejected:
\begin{quote}
\begin{center}
An analysis of the frobnicatable foo filter.
\end{center}
In this paper we present a performance analysis of our
previous paper [1], and show it to be inferior to all
previously known methods. Why the previous paper was
accepted without this analysis is beyond me.
[1] Removed for blind review
\end{quote}
An example of an acceptable paper:
\begin{quote}
\begin{center}
An analysis of the frobnicatable foo filter.
\end{center}
In this paper we present a performance analysis of the
paper of Smith \emph{et al}\onedot [1], and show it to be inferior to
all previously known methods. Why the previous paper
was accepted without this analysis is beyond me.
[1] Smith, L and Jones, C. ``The frobnicatable foo
filter, a fundamental contribution to human knowledge''.
Nature 381(12), 1-213.
\end{quote}
If you are making a submission to another conference at the same time,
which covers similar or overlapping material, you may need to refer to that
submission in order to explain the differences, just as you would if you
had previously published related work. In such cases, include the
anonymized parallel submission~\cite{Authors12final} as additional material and
cite it as
\begin{quote}
[1] Authors. ``The frobnicatable foo filter'', F\&G 2021 Submission ID 324,
Supplied as additional material {\tt fg324.pdf}.
\end{quote}
Finally, you may feel you need to tell the reader that more details can be
found elsewhere, and refer them to a technical report. For conference
submissions, the paper must stand on its own, and not {\em require} the
reviewer to go to a techreport for further details. Thus, you may say in
the body of the paper ``further details may be found
in~\cite{Authors12bfinal}''. Then submit the techreport as additional material.
Again, you may not assume the reviewers will read this material.
Sometimes your paper is about a problem which you tested using a tool which
is widely known to be restricted to a single institution. For example,
let's say it's 1969, you have solved a key problem on the Apollo lander,
and you believe that the 3DV70 audience would like to hear about your
solution. The work is a development of your celebrated 1968 paper entitled
``Zero-g frobnication: How being the only people in the world with access to
the Apollo lander source code makes us a wow at parties'', by Zeus \emph{et al}\onedot.
You can handle this paper like any other. Don't write ``We show how to
improve our previous work [Anonymous, 1968]. This time we tested the
algorithm on a lunar lander [name of lander removed for blind review]''.
That would be silly, and would immediately identify the authors. Instead
write the following:
\begin{quotation}
\noindent
We describe a system for zero-g frobnication. This
system is new because it handles the following cases:
A, B. Previous systems [Zeus et al. 1968] didn't
handle case B properly. Ours handles it by including
a foo term in the bar integral.
...
The proposed system was integrated with the Apollo
lunar lander, and went all the way to the moon, don't
you know. It displayed the following behaviours
which show how well we solved cases A and B: ...
\end{quotation}
As you can see, the above text follows standard scientific convention,
reads better than the first version, and does not explicitly name you as
the authors. A reviewer might think it likely that the new paper was
written by Zeus \emph{et al}\onedot, but cannot make any decision based on that guess.
He or she would have to be sure that no other authors could have been
contracted to solve problem B.
FAQ: Are acknowledgments OK? No. Leave them for the final copy.
\begin{figure}[t]
\begin{center}
\fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
\end{center}
\caption{Example of caption. It is set in Roman so that mathematics
(always set in Roman: $B \sin A = A \sin B$) may be included without an
ugly clash.}
\label{fig:long}
\label{fig:onecol}
\end{figure}
\subsection{Miscellaneous}
\noindent
Compare the following:\\
\begin{tabular}{ll}
\verb'$conf_a$' & $conf_a$ \\
\verb'$\mathit{conf}_a$' & $\mathit{conf}_a$
\end{tabular}\\
See The \TeX book, p165.
The space after \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, meaning ``for example'', should not be a
sentence-ending space. So \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot is correct, {\em e.g.} is not. The provided
\verb'\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot' macro takes care of this.
When citing a multi-author paper, you may save space by using ``et alia'',
shortened to ``\emph{et al}\onedot'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.)
However, use it only when there are three or more authors. Thus, the
following is correct: ``
Frobnication has been trendy lately.
It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by
Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \emph{et al}\onedot~\cite{Alpher04}.''
This is incorrect: ``... subsequently developed by Alpher \emph{et al}\onedot~\cite{Alpher03} ...''
because reference~\cite{Alpher03} has just two authors. If you use the
\verb'\emph{et al}\onedot' macro provided, then you need not worry about double periods
when used at the end of a sentence as in Alpher \emph{et al}\onedot.
\begin{figure*}
\begin{center}
\fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}}
\end{center}
\caption{Example of a short caption, which should be centered.}
\label{fig:short}
\end{figure*}
\section{Formatting your paper}
All text must be in a two-column format. The total allowable width of the
text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54
cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a
$\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the
first page) should begin 1.0 inch (2.54 cm) from the top edge of the
page. The second and following pages should begin 1.0 inch (2.54 cm) from
the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86
cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4
paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the
page.
\subsection{Margins and page numbering}
All printed material, including text, illustrations, and charts, must be kept
within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm)
high. Page numbers should be in footer with page numbers, centered and .75
inches from the bottom of the page and make it start at the correct page
number rather than the 4321 in the example. To do this find the line (around
line 23)
\begin{verbatim}
\setcounter{page}{4321}
\end{verbatim}
where the number 4321 is your assigned starting page.
Make sure the first page is numbered by commenting out the first page being
empty on line 47
\begin{verbatim}
\end{verbatim}
\subsection{Type-style and fonts}
Wherever Times is specified, Times Roman may also be used. If neither is
available on your word processor, please use the font closest in
appearance to Times to which you have access.
MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of
the first page. The title should be in Times 14-point, boldface type.
Capitalize the first letter of nouns, pronouns, verbs, adjectives, and
adverbs; do not capitalize articles, coordinate conjunctions, or
prepositions (unless the title begins with such a word). Leave two blank
lines after the title.
AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title
and printed in Times 12-point, non-boldface type. This information is to
be followed by two blank lines.
The ABSTRACT and MAIN TEXT are to be in a two-column format.
MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use
double-spacing. All paragraphs should be indented 1 pica (approx. 1/6
inch or 0.422 cm). Make sure your text is fully justified---that is,
flush left and flush right. Please do not place any additional blank
lines between paragraphs.
Figure and table captions should be 9-point Roman type as in
Figures~\ref{fig:onecol} and~\ref{fig:short}. Short captions should be centred.
\noindent Callouts should be 9-point Helvetica, non-boldface type.
Initially capitalize only the first word of section titles and first-,
second-, and third-order headings.
FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction})
should be Times 12-point boldface, initially capitalized, flush left,
with one blank line before, and one blank line after.
SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements})
should be Times 11-point boldface, initially capitalized, flush left,
with one blank line before, and one after. If you require a third-order
heading (we discourage it), use 10-point Times, boldface, initially
capitalized, flush left, preceded by one blank line, followed by a period
and your text on the same line.
\subsection{Footnotes}
Please use footnotes\footnote {This is what a footnote looks like. It
often distracts the reader from the main flow of the argument.} sparingly.
Indeed, try to avoid footnotes altogether and include necessary peripheral
observations in
the text (within parentheses, if you prefer, as in this sentence). If you
wish to use a footnote, place it at the bottom of the column on the page on
which it is referenced. Use Times 8-point type, single-spaced.
\subsection{References}
List and number all bibliographical references in 9-point Times,
single-spaced, at the end of your paper. When referenced in the text,
enclose the citation number in square brackets, for
example~\cite{Authors12final}. Where appropriate, include the name(s) of
editors of referenced books.
\begin{table}
\begin{center}
\begin{tabular}{|l|c|}
\hline
Method & Frobnability \\
\hline\hline
Theirs & Frumpy \\
Yours & Frobbly \\
Ours & Makes one's heart Frob\\
\hline
\end{tabular}
\end{center}
\caption{Results. Ours is better.}
\end{table}
\subsection{Illustrations, graphs, and photographs}
All graphics should be centered. Please ensure that any point you wish to
make is resolvable in a printed copy of the paper. Resize fonts in figures
to match the font in the body text, and choose line widths which render
effectively in print. Many readers (and reviewers), even of an electronic
copy, will choose to print your paper in order to read it. You cannot
insist that they do otherwise, and therefore must not assume that they can
zoom in to see tiny details on a graphic.
When placing figures in \LaTeX, it's almost always best to use
\verb+\includegraphics+, and to specify the figure width as a multiple of
the line width as in the example below
{\small\begin{verbatim}
\usepackage[dvips]{graphicx} ...
\includegraphics[width=0.8\linewidth]
{myfile.eps}
\end{verbatim}
}
\subsection{Color}
Color is valuable, and will be visible to readers of the electronic copy.
However ensure that, when printed on a monochrome printer, no important
information is lost by the conversion to grayscale.
\section{Final copy}
You must include your signed IEEE copyright release form when you submit
your finished paper. We MUST have this form before your paper can be
published in the proceedings.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\label{sec:introduction}
It has transpired recently that a number of operator-theoretical techniques which have been under active development for the past 60 years or so are extremely useful in the asymptotic analysis of highly inhomogeneous media. Apart from yielding sharp asymptotics of the corresponding Hamiltonians in the norm-resolvent topology, this research has resulted in a number of new important, yet mostly unexplored, connections between certain areas of the modern operator and spectral theory. These include the theory of dilations and functional models of dissipative and non-selfadjoint operators in Hilbert spaces, the boundary triples theory in the analysis of symmetric operators, zero-range models with an internal structure and, finally, the theory of generalised resolvents and their out-of-space ``dilations''.
The present survey, based on our results published in \cite{KCher,CEK,CherErKis,CEKN,CENS,CherednichenkoKiselevSilva,CherKisSilva1,CherKisSilva2,CherKisSilva3,CENS,MR2330831,Ryzh_spec,Ryzhov_closed}, attempts to shed some light on these connections and to thus present the subject area of strongly inhomogeneous media under the spotlight of modern spectral theory. We aim to show that in many ways this novel outlook allows one to gain a better understanding of the mentioned area by providing a universal abstraction layer for all the main objects to be found in the asymptotic analysis. Moreover, in most cases one can then proceed in the analysis on a purely abstract level surprisingly far, essentially postponing the use of the specific features of the problem at hand till the very last stages.
For readers' convenience, we have included a rather detailed exposition of the relevant areas of operator and spectral theory, keeping in mind that some papers laying the foundations of these areas have been poorly accessible to date.
We start with Section 2, devoted to the now-classical theory of dilations of dissipative operators. The role of dissipative operators as opposed to self-adjoint ones is that whereas the latter represent physical systems with the energy conservation law (``closed'', or conservative, systems), the former allow for the consideration of a more realistic setup, where the loss of the total energy is factored in. The importance of dissipative systems has been a common place since at least the works of I.~Prigogine; it is well-known that such systems may possess certain rather unexpected properties. The main difference between the self-adjoint and dissipative theories can be clarified,
following M.~G.~Kre\u\i n, as follows: the major instruments of
self-adjoint spectral analysis arise from the Hilbert space geometry,
whereas this geometry doesn't work very well in the
non-selfadjoint situation, with modern complex analysis taking
the role of the main tool in the investigation.
Since the seminal contribution of B.~Sz.-Nagy and C.~Foia\c{s}, the main object of dissipative spectral analysis has been the so-called dilation, representing an out-of-space self-adjoint extension, in the sense of \eqref{DilationCondition} below, of the original dissipative operator $L$. Our argument actually goes as far as to suggest that this concept underpins the whole set of ideas and notions presented in the paper. B.~S.~Pavlov's explicit construction of dilation relies upon the second major ingredient, which is the characteristic function $S(z),$ see \eqref{DissCharF}, which is an analytic operator-valued contraction in $\mathbb{C}_+$. The analysis of the dissipative operator $L$ is reduced to the study of the function $S(z),$
and hence from this point onward it belongs to the domain of complex analysis. It turns out that the sole knowledge of $S(z)$ yields an explicit spectral representation of the dilation. Moreover, Naboko has shown that, in the same representation, a whole family of operators ``close'' to $L$, self-adjoint and non-selfadjoint alike, are modelled in an effective way. This idea in particular led to the description of absolutely continuous subspaces of all the operators considered as the closure of the so-called smooth vectors set. This latter is characterised as the collection of vectors such that the resolvent of the operator in question in the spectral representation maps them as the multiplication operator. An explicit construction of wave operators and scattering matrices then follows almost immediately.
In Section \ref{sec:ryzh-funct-model}, we give a systematic exposition of this approach applied to the family of extensions of a symmetric densely defined operator on a Hilbert space $H$ possessing equal deficiency indices. In so doing, we follow closely the strategy suggested by Sergey Naboko which he had applied in the analysis of additive relatively bounded perturbations of self-adjoint operators. We thus hope to provide a coherent presentation of the major contribution by Naboko to the spectral analysis of non-selfadjoint operators.
Our analysis is facilitated by the boundary triples theory, being an abstract framework, from which the extensions theory of symmetric operators, especially differential operators, greatly benefits. That's why we start our exposition by introducing the main concepts of this theory. The formula obtained for the scattering operator in the functional model representation allows us to derive an explicit formula for the scattering matrix, formulated in terms of Weyl-Titchmarsh $M$-matrices, i.e., in the natural terms associated with the problem. In Section \ref{sec:theor-appl-funct}, we consider an application of this technique to an inverse scattering problem on a quantum graph, where we are able to give an explicit solution to the problem of reconstructing matching conditions at graph vertices.
In Section \ref{sec:funct-model-family}, we consider the possible generalisation of the approach described above to the case of partial differential operators (PDO), associated with boundary value problems (BVP). Although the theory of boundary triples has been successfully applied to the spectral analysis of BVP for ordinary differential operators and related setups, in its original form this theory is not suited for dealing with BVP for partial differential equations (PDE), see \cite[Section
7]{MR2418300} for a relevant discussion.
Recently, when the works \cite{Grubb_Robin, Grubb_mixed, BehrndtLanger2007, Gesztesy_Mitrea, Ryzh_spec, MR2418300} started to appear, it has transpired that, suitably modified, the boundary triples approach nevertheless admits a natural generalisation to the BVP setup, see also the seminal contributions by
J. W. Calkin \cite{Calkin}, M.\,S.\,Birman \cite{Birman}, L.\,Boutet de Monvel \cite{BdeM}, M.\,S.\,Birman and M.\,Z.\,Solomyak \cite{Birman_Solomyak}, G.\,Grubb \cite{Grubb_classic}, and M.\,Agranovich \cite{Agranovich}, which provide the analytic backbone for the related operator-theoretic constructions.
In all cases mentioned above, one can see the fundamental r\^{o}le of
a certain Herglotz operator-valued analytic function, which in
problems where a boundary is present (and sometimes even without an
explicit boundary \cite{Amrein}) turns out to be a natural
generalisation of the classical notion of a Dirichlet-to-Neumann
map. Moreover, it is precisely this object that permits to define the characteristic function which in turn facilitates the functional model construction.
In Section \ref{sec:models-potent-zero}, we pass over to the discussion of zero-range models with an internal structure. The idea of replacing a model of short-range interactions by an explicitly solvable one with a zero-radius potential (possibly with an internal structure)
\cite{Berezin_Faddeev,Pavlov_internal_structure,Pavlov_Helmholtz_resonator, MR0080271,MR0024574,MR0024575, MR0051404} has paved the way for an influx of methods of the theory of extensions (both
self-adjoint and non-selfadjoint) of symmetric operators to problems
of mathematical physics. In particular, we view zero-range potentials with an internal structure as a particular case of out-of-space self-adjoint extensions of symmetric operators, the theory of which is intrinsically related to the analysis of generalised resolvents. The latter area is introduced in Section \ref{gen_res1}. We argue that out-of-space self-adjoint extensions corresponding to generalised resolvents naturally supersedes the dilation theory as presented in Section \ref{sec:pavl-funct-model}.
On yet another level, we claim that zero-range perturbations (and more precisely, zero-range perturbations with an internal structure) appear naturally as the norm-resolvent limits of Hamiltonians in the asymptotic analysis of inhomogeneous media. This relationship is established using the apparatus of generalised resolvents, as explained in Section \ref{gen_res2}.
Finally, we mention here that the theory of functional models as presented in Section \ref{sec:functional-model} is directly applicable to the treatment of models of zero-range potentials with an internal structure. Its development yields a complete spectral analysis and an explicit construction of the scattering theory for the latter.
Two different models are considered in
Section \ref{sec:models-potent-zero}, one of these being a one-dimensional periodic model with critical contrast, unitary equivalent to the double porosity one. The PDE counterpart of the latter is discussed in Section \ref{sec:appl-cont-mech}. The second mentioned model pertains to the problem with a low-index inclusion in a homogeneous material. Our argument shows that the leading order term in the asymptotic expansion of its resolvent admits the same form as expected of a zero-range model; the difference is that here the effective model of the media is no longer ``zero-range'' per se; rather it pertains to a singular perturbation supported by a manifold. Therefore, this allows us to extend the notion of internal structure to the case of distributional perturbations supported by manifolds.
The discussion started in Section \ref{sec:models-potent-zero} is then continued in Section \ref{sec:appl-cont-mech}. We note that in every model considered so far, the internal structure of the limiting zero-range model is necessarily the simplest possible, i.e., pertains to the out-of-space extensions defined on $H\oplus \mathbb{C}^1$, where $H$ is the original Hilbert space. It turns out that this is due to the fact that we only consider norm-resolvent convergence when the spectral parameter $z$ is restricted to a compact set in $\mathbb{C}.$
Passing over to a generic setup with $z$ not necessarily in a compact, we are able to claim that in some sense the internal structure can be arbitrarily complex, provided that the spectral parameter $z$ is allowed to grow with the large parameter $a$, which describes the inhomogeneity, increasing to $+\infty.$ This allows us to present an explicit example of a non-trivial internal structure in the leading order term of the norm-resolvent asymptotics in Section 4.4, which is supplemented with the discussion of the so-called scaling regimes which we introduce in Section 5.1.
The remainder of Section 5 is devoted to the analysis of a double porosity model of high-contrast homogenisation, where the leading order term of the asymptotic expansion is obtained by an application of the operator-theoretical technique based on the generalised resolvents, as in Section 4.
\section{Functional models for dissipative and nonselfadjoint operators}
\label{sec:functional-model}
Functional model construction for a contractive linear operator~$T$
acting on a Hilbert space~$H$ is a well developed domain of the
operator theory.
Since the pioneering works by B.~Sz.-Nagy, C.~Foia\c{s}~\cite{MR2760647},
P.~D.~Lax, R.~S.~Phillips~\cite{MR0217440}, L.~de~Branges,
J.~Rovnyak~\cite{deBranges, deBrangesRovnyak}, and
M.~S.~Liv\v{s}ic~\cite{Livsic1954}, this field of research has attracted many specialists in operator theory, complex analysis, system control,
Gaussian processes and other disciplines.
Multiple studies culminated in the development of a comprehensive theory
complemented by various applications,
see~\cite{DyM,Fu,Nikolski,NikolskiiKhrushchev, NikolskiiVasyunin1998}
and references therein.
The underlying idea of a functional model is the fundamental theorem
of B.~Sz.-Nagy and C.~Foias establishing the existence of a unitary dilation for
any contractive (linear) Hilbert space operator~$T$, $\|T\| \leq 1$.
The unitary dilation~$U$ of $T$ is a unitary operator on a
Hilbert space~$\mathcal H \supset H$
such that $P_H U^n|_H = T^n$ for all
$n = 1,2,\dots$.
Here $P_H : {\mathcal H} \rightarrow H$ is an orthoprojection
from~$\mathcal H$ to its subspace~$H$.
The dilation~$U$ is called minimal if the linear set~$\vee_{n>0} U^n H$
is dense in~$\mathcal H$.
The minimal dilation~$U$ of a contraction is unique up to unitary
equivalence.
The spectrum of~$U$ is absolutely continuous and covers the
unit circle~$\mathbb T = \{z \in \mathbb C : |z| = 1\}$.
If one denotes by~$\mu$ the spectral measure of~$U$,
the spectral theorem yields that operator~$T$ is unitarily equivalent
to its model~$T = P_H z |_H$, where ~$f \mapsto z f$ is the
operator of multiplication on the spectral representation space $L^2(\mathbb T, \mu)$ of~$U$.
Due to its abstract nature, a significant part of the functional model
research for contractions took place among specialists
in complex analysis and operator theory.
The parallel theory for unbounded
operators is based on the Cayley
transform~$T \mapsto -{\rm i}(T+I)(T-I)^{-1}$ applied to a
contraction~$\|T \| \leq 1$.
Assuming that~$\clos\ran (T -I) = H$, the Cayley transform of~$T$,
$\|T\| < 1$ is a dissipative densely defined
operator~$L = -{\rm i}(T+I)(T-I)^{-1}$, not necessarily bounded in~$H$.
It is easily seen that the spectrum of~$L$ is situated in the closed upper half
plane~$\overline{\mathbb C}_+$ of the complex plane.
The imaginary part of~$L$ (defined in the sense of forms if needed)
is a non-negative operator.
Alongside the developments in operator theory, the second half
of the 20th century witnessed a huge progress in the spectral analysis
of linear operators pertaining to physical disciplines.
The principal tool of this was the method of Riesz projections, i.e., the contour
integration of the operator's resolvent in the complex plane of
spectral parameter.
The spectral analysis of self-adjoint operators of quantum mechanics
can be viewed as the prime example of highly successful application of contour
integration in the study of conservative systems, i.e., closed
systems with the energy preserved in the course of evolution.
Topical questions concerning the behavior of non-conservative systems, where
the total energy is not preserved, and of resonant systems motivated in-depth studies of (unbounded)
non-selfadjoint operators.
The analysis of non-conservative dynamical systems and of
non-selfadjoint operators especially relevant to the functional
model theory was pioneered
in the works
of M.~S.~Brodski\v{i},
M.~S.~Liv\v{s}ic and their colleagues,
see~\cite{Brodski, BrodskiLivsic, LivsicBook}.
Starting with a (bounded) self-adjoint
operator~$A = A^*$
as the main operator of a closed conservative system,
these authors considered the
coupling of this system to the outside
world by means of externally attached channels.
This construction represents a model of the so-called
``open system'', that is, of a physical system
connected to its external environment.
The energy of such modified system
can dissipate through the external channels,
while at the same time the energy can be fed into
the system from the outside in the course
of its evolution.
In works of M.~S.~Brodski\v{\i} and
M.~S.~Liv\v{s}ic, the external channels are modelled as an
additive perturbation of the main self-adjoint
operator~$A$ by a (bounded) non-selfadjoint
perturbation: $A \to L = A +{\rm i}V$, $V = V^*$.
The ``channel vectors'' form the Hilbert
space~$E = \clos \ran |V|$.
If~$V \geq 0$, the operator~$L$ is dissipative
(i.e., $\im (Lu,u) > 0$, $u \in H$);
it describes a non-conservative
system losing the total energy.
In turn, and quite analogously to the
case of contractions,
under the assumption~$ \mathbb C_- \subset \rho(L)$
(recall that dissipative operators satisfying this condition are
called {\em maximal}),
the self-adjoint dilation of~$L$ is a self-adjoint
operator~$\mathscr L$ on a wider
space~$\mathcal H \supset H$
such that
\begin{equation}\label{DilationCondition}
(L-zI)^{-1} = P_H (\mathscr L - zI)^{-1}|_H, \quad
z \in \mathbb C_-,
\end{equation}
where~$P_H$ is an orthogonal projection
from $\mathcal H$ onto $H$.
The operator~$\mathscr L$ describes a (larger) system with
the state space~$\mathcal H$, in which
the energy is conserved, whereas~$L$ describes
its subsystem losing its total energy.
In the general case, a non-dissipative~$L$
corresponds to an
open system where both the
energy loss and the energy supply
coexist.
The analysis of a non-selfadjoint operator~$L$ relies on the
notion of its characteristic function~\cite{MR0020719,Strauss1960}
discovered by M.~S.~Liv\v{s}ic in 1943--1944.
It is a bounded analytic
operator-function~$\Theta(z)$, $z \in \rho(L^*)$
defined on the resolvent set of~$L^*$ and acting
on the ``channel vectors'' from the space~$E$.
For dissipative $L$ the function~$\Theta$ coincides with the characteristic
function of a contraction~$T = (L-{\rm i}I)(L+{\rm i}I)^{-1}$
(the inverse Cayley transform of $L$), featured prominently in the works by B.~Sz.-Nagy and C.~Foias.
The characteristic function of a non-selfadjoint
operator~$L$ (or, alternatively, of its Cayley transform) determines the original operator $L$ uniquely up to a unitary
equivalence (see~\cite{Livsic1954, MR2760647}),
provided~$L$ has no non-trivial self-adjoint ``parts''.
Therefore, the study of non-selfadjoint operators
is reduced to the study of
operator-valued analytic functions.
In and of itself, this does not mean much as these functions might
be as complicated as the operators themselves.
A simplification is achieved when the values of
these functions are either matrices or belong to Schatten-von Neumann classes of
compact operators, which is often the case in physical applications.
Closely related to the Sz.-Nagy-Foias model for
contractions and to the open
systems framework are the Lax-Phillips
scattering theory~\cite{MR0217440}
and the ``canonical model'' due to L.~de Branges
and J.~Rovnyak~\cite{deBrangesRovnyak}.
The latter is developed
for completely non-isometric
contractions and their adjoints
with quantum-mechanical
applications in mind.
The Lax-Phillips theory was
originally developed to facilitate the analysis of
scattering problems for hyperbolic wave equations in exterior domains to compact scatterers.
It provides useful
intuition into the underpinnings of
the functional model construction and
this connection will be exploited
in the next section.
It was realized very early~\cite{MR0206711} that the three
theories, i.e., the open systems theory, the Sz.-Nagy-Foia\c{s} model, and
the Lax-Phillips scattering, all deal with
essentially the same objects.
In particular, the characteristic function of a contraction (or of a
dissipative operator) emerges, albeit under disguise, in all three theories.
Being a purely theoretical abstract object in the Sz.-Nagy-Foia\c{s} theory, the characteristic function emerges as a
transfer function of a linear system according to M.~S.~Brodski\v{\i} and
M.~S.~Liv\v{s}ic, and as the scattering matrix in the Lax-Phillips theory.
The characteristic function of a contraction is
also the central component in
the L.~de Branges and J.~Rovnyak
model theory~\cite{deBrangesRovnyak}.
Deep connections between the Sz-Nagy-Foias and the
de Branges-Rovnyak
models are clarified in a series of
papers by N.~Nikolskii and V.~Vasyunin~\cite{NikolskiiVasyunin1986,
NikolskiiVasyunin1989, NikolskiiVasyunin1998}.
\subsection{Lax-Phillips theory}
The Lax-Phillips scattering theory~\cite{MR0217440}
for the acoustic waves by
a smooth compact obstacle in~$\mathbb R^n$ with $n \geq 3$ odd
provides an excellent
illustration of the intrinsic links
between the operator theory and
mathematical physics.
A number of concepts found in the theory of functional
models of dissipative operators
find their direct counterparts here, expressed in the
language of realistic physical processes.
For instance, the characteristic function
of the operator governing the
scattering process
is realized as the scattering matrix,
the self-adjoint dilation corresponds to
the operator of ``free'' dynamics, i.e., the
wave propagation process observed
in absence of the obstacle, and the scattering
channels are a direct analogue of the channels found in
the Brodski\v{\i}-Liv\v{s}ic constructions.
In this section we briefly recall
the main concepts of Lax-Phillips scattering theory.
Let $\cal H$ be a Hilbert space
with two mutually
orthogonal subspaces $\cal
D_\pm \subset H$, $\cal D_- \oplus \cal D_+ \neq \mathcal H$.
Denote by $\mathcal K$ the orthogonal
complement of~$\cal D_- \oplus \cal D_+$
in $\mathcal H$.
Assume the existence of a single parameter evolution group of
unitary operators~$\{U(t)\}_{t\in \mathbb R}$ with the following properties
\begin{equation}\label{conditionsUt}
\begin{aligned}
U(t)\mathcal{D}_- & \subseteq \mathcal{D}_-, \quad t \leq 0, \\[0.3em]
U(t) \mathcal{D}_+ & \subseteq \mathcal{D}_+, \quad t \geq 0, \\[0.3em]
\cap_{t \in \mathbb R}U(t) \mathcal{D}_\pm &= \{0\},\\[0.3em]
\clos \cup_{t \in \mathbb R}U(t) \mathcal{D}_\pm & = \mathcal{H}.
\end{aligned}
\end{equation}
In the acoustic scattering,
the space~$\mathcal H$ consists of
solutions to the wave equation
(i.e., acoustic waves) and is
endowed with the energy norm.
The group~$U(t)$ describes the
evolution of ``free'' waves in~$\mathcal H$,
that is,
the group~$U(t)$ maps
the Cauchy data of solutions at the time~$t = 0$
to their Cauchy data at the time $t$.
Since~$U(t)$ is unitary for all $t\in\mathbb R$,
the energy of solutions is preserved
under the time
evolution~$f = U(0) f \mapsto U(t)f$, for any
Cauchy data~$f \in \mathcal H$.
Correspondingly, the infinitesimal
generator~$\mathscr B$ of~$U(t)$ is a
self-adjoint
operator in $\mathcal H$ with purely absolutely continuous spectrum
covering the whole real line.
The subspaces~$\cal D_\pm$ are called incoming and
outgoing subspaces of~$U(t)$.
These names are well justified. Indeed,
the subspace~$\mathcal D_-$ consists of solutions
that do not interact with the
obstacle prior to the moment~$t = 0$,
whereas $\mathcal D_+$ consists of scattered waves
which do not interact with the obstacle after $t = 0$.
There exist two representations for the generator~$\mathscr B$
associated with $\mathcal D_\pm$ (the so called incoming and
outgoing translation representations), in which
the group~$U(t)$ acts as the right shift
operator~$U(t): u(x)\mapsto u(x-t)$ on $L^2(\mathbb R, E)$
with some auxiliary Hilbert space~$E$.
In these representations the subspaces~$\mathcal D_\pm$
are mapped to $L^2(\mathbb R_\pm, E)$.
It is not difficult to see that this
construction satisfies the assumptions~(\ref{conditionsUt}).
Denote by
$P_\pm : \mathcal{H} \to [\mathcal{D}_\pm]^\perp$
the orthogonal projections to the complements of
$\mathcal{D}_\pm$ in $\cal H$.
The elements of~$\mathcal K$ are the scattering waves that are neither
incoming in the past, nor outgoing in the future,
i.e., the waves localized in the vicinity of the obstacle.
The interaction of incoming waves with the obstacle, i.e.,
the scattering
process, is then described by the compression of the group~$U(t)$ to the
neighborhood of the obstacle:
$$
Z(t) = P_+ U(t) P_- = P_{\mathcal K} U(t) P_{\mathcal K}, \quad t \geq 0.
$$
Here~$ P_{\mathcal K} = P_- P_+$ is an orthoprojection on~$\mathcal H$.
The operator family~$\{Z(t)\}_{t \geq 0}$ forms a strongly
continuous semigroup acting on~$\mathcal{K}$.
Since $Z(t)$ is a compression of the unitary group, it is clear that
$\|Z(t)\| \leq 1$ for all $t \geq 0$.
The infinitesimal generator~$B$ of the semigroup~$\{Z(t)\}_{t \geq 0}$
turns out to be a
linear operator with purely discrete spectrum~$\{\lambda_k\}$
with~$\re(\lambda_k) < 0$,
$k = 1,2,\dots$.
The poles~$\{\lambda_k\}$ of the resolvent of~$B$
and the corresponding
eigenvectors are interpreted
as the scattering resonances.
%
These resonances correspond
to the poles of the scattering matrix
defined as an operator-valued function acting in the
space $L^2(\mathbb R_-, E)$, i.e., the space of vector
functions taking values in~$E$.
The scattering matrix is mapped by the Fourier transform to the
analytic in the lower half-plane operator function~$S(z)$, $z\in\mathbb C_-$
with zeroes at~$z_k = -{\rm i}\lambda_k$.
The boundary values $S(k - {\rm i}0)$ on the real
axis exist almost
everywhere in the strong operator topology
and are
unitary for almost all $k \in \mathbb R$.
The function~$S(z)$
permits an analytic continuation~$S(z) = [S^*(\bar z)]^{-1}$
to the upper half-plane, where it is meromorphic.
The results of~\cite{MR0206711} show
that $\theta(z): = S^*(\bar z)$ coincides with
the Liv\v{s}ic's characteristic
function of a dissipative operator being
unitary equivalent to the infinitesimal
generator of~$Z(t)$.
Consequently,~$\theta\bigl((z+{\rm i})/(z-{\rm i})\bigr)$
is the characteristic function
of its Cayley transform as
defined by B.~Sz.-Nagy
and C.~Foia\c{s}~\cite{MR2760647}.
Finally,
the resolvents of
operators~$\mathscr L = -{\rm i}\mathscr B$ and $L = -{\rm i}B$
satisfy the dilation equation~(\ref{DilationCondition}).
In other words, the operator corresponding to the free dynamics
is the self-adjoint dilation of the dissipative operator
that governs the scattering process.
\subsubsection{Minimality, non-selfadjointness, resolvent}
The beautiful geometric interpretation of scattering processes provided by the Lax-Phillips
theory is not entirely transferable to
the modelling of a general
dissipative (or contractive) operator.
For instance, given an arbitrary
dissipative operator~$L$ on a Hilbert
space~$K$, its selfadjoint
dilation $\mathscr L$ does
not exists {\it a priori} and must be
explicitly constructed first.
In addition, such a
dilation~$\mathscr L$ should
be {\it minimal}, that is,
it must contain no
reducing self-adjoint parts
unrelated to the operator~$L$.
Mathematically, the minimality condition is
expressed by the equality
\[
\clos\bigvee_{z \notin R} {(\mathscr L - zI)^{-1}
\mid_K} =
\mathcal H
\]
where $\mathcal H$ is the dilation space~$\mathcal H
\supset K$.
The construction of a dilation satisfying this
condition is a highly non-trivial task which was successfully addressed for contractions
by B.~Sz.-Nagy and C.~Foia\c{s}~\cite{MR2760647}, aided by a theorem of M.~A.~Na\u{\i}mark \cite{Naimark1943},
and then by B.~Pavlov~\cite{MR0510053, Drogobych}
in two
important cases of dissipative
operators arising in
mathematical physics. Later, this construction was generalized to a generic setting. We dwell on this further in the following sections.
For obvious reasons, the functional model theory of non-selfadjoint operators deals with operators possessing no non-trivial reducing
self-adjoint parts.
Such operators are called {\em completely non-selfadjoint} or, using
a somewhat less accurate term, {\em simple}.
The rationale behind this condition is easy to
illustrate within the Lax-Phillips framework.
Let the dissipative operator~$L = -{\rm i}B$ governing
the wave dynamics in a vicinity of an obstacle possess
a non-trivial self-adjoint part.
This part is then a self-adjoint operator
acting on the subspace spanned
by the eigenvectors of~$L$
corresponding to its real eigenvalues.
The restriction of $Z(t)$ to this subspace
is an isometry for all~$t\in \mathbb R$, and
the energy of these states remains constant
(recall that the space~$K$ is equipped with
the energy norm).
Therefore these waves stay in
a bounded region adjacent
to the obstacle at all times.
These bound states do
not participate in scattering, as they
are invisible to the scattering process
describing the asymptotic
behaviour as $t \to\pm\infty$, so that
the matrix~$S$ (or the characteristic
function of~$L$) contains no
information pertaining to them.
In operator theory, this is known as the
claim that the characteristic function of a dissipative
operator is oblivious to its self-adjoint part.
It is also well-known, that the characteristic
function uniquely determines the
completely non-selfadjoint part of a
dissipative operator.
The interaction dynamics of the Lax-Phillips scattering
supposes neither minimality nor complete
non-selfadjointness.
One can envision incoming waves that do not interact with the
obstacle during their evolution and eventually become outgoing as~$t\to+\infty$.
It is also easy to conceive of trapping
obstacles preventing waves
from leaving the neighbourhood of the
obstacle at~$t\to+\infty$.
The geometry of such
obstacles cannot be fully recovered from
the scattering data because, as explained above,
these standing waves
do not participate in the
scattering process.
In the applications discussed below,
unless explicitly stated otherwise,
all non-selfadjoint operators are
assumed closed, densely defined
with regular points in both lower
and upper half planes.
The latter condition can be relaxed but is adopted in what follows for the sake of convenience.
\subsection{Pavlov's functional model and its spectral form}
\label{sec:pavl-funct-model}
Functional models for prototypical dissipative operators
of mathematical physics (as
opposed to the model for contractions), alongside explicit constructions of self-adjoint
dilations, were investigated by
B.~Pavlov in his
works~\cite{MR0365199, MR0510053, Drogobych}.
%
Two classes of dissipative operators were considered:
the Schr\"odigner operator in $L^2(\mathbb R^3)$ with
a complex-valued potential, and the operator
generated by the differential
expression~$-y'' + q(x)y $ on the interval~$[0, \infty)$
with a dissipative boundary condition at $x = 0$.
In both cases the self-adjoint dilations are constructed explicitly in terms of the problem at hand, and
supplemented by the model
representations known today
as ``symmetric''
and commonly referred to
as the Pavlov's model.
The results of~\cite{MR0365199, MR0510053, Drogobych}
were extensively employed in
various applications and provided a foundation for the subsequent constructions
of self-adjoint dilations and functional models for
general non-selfadjoint
operators.
\subsubsection{Additive perturbations~\cite{MR0365199, MR0510053}}\label{sec:add_perturbations}
Let $A = A^*$ be a selfadjoint unbounded operator on a Hilbert space~$K$ and
$V$ a bounded non-negative operator~$V = V^* = \alpha^2/2\geq 0$, where
$\alpha := (2V)^{1/2}$.
The paper~\cite{MR0510053} studies the dissipative
Schr\"odinger operator~$L = A +({\rm i}/2)\alpha^2$ in $\mathbb R^3$
defined by the differential expression~ $-\Delta + q(x)+({\rm i}/2)\alpha^2(x)$
with real continuous functions~$q$ and $\alpha$
such that $0 \leq \alpha\leq C < \infty$.
The operators~$A = -\Delta + q$ and~$V = \alpha^2/2$ are the
real and imaginary parts
of~$L$ defined on $\dom(L) = \dom(A)$.
Assuming the operator~$L$ has no non-trivial self-adjoint components
and the resolvent set of~$L$ contains points in both upper and
lower half planes,
the operator $L$ is a maximal completely non-selfadjoint, densely defined
dissipative operator on~$K$.
According to the general theory, there exists a minimal dilation of
$L$, which is a self-adjoint operator~$\mathscr L$ on a Hilbert space $\mathcal
H \supset K$ such that
\begin{equation}\label{eqn:DilationEq}
(L - zI)^{-1} = P_K ( {\mathscr L} - zI)^{-1}|_K, \quad z \in\mathbb C_-,
\end{equation}
where $P_K$ is the orthogonal projection
from $\mathcal H$ onto its subspace~$K$.
The dilation constructed in~\cite{MR0510053} closely
resembles the generator $\mathscr B$ of the unitary group $U(t)$ in the Lax-Phillips theory: Pavlov realised that a natural way to construct a self-adjoint dilation would be to add the missing ``incoming'' and ``outgoing'' energy channels to the original non-conservative dynamics, thus mimicking the starting point of Lax and Phillips.
The challenge here is to determine the operator that describes the ``free''
evolution of the dynamical system given only its ``internal'' part, thus in some sense ``reversing" the Lax-Phillips approach.
Denote $E:= \clos \ran \alpha$ and define the dilation
space as the direct sum of
$K$ and the equivalents of incoming and
outgoing channels~ $\mathcal D_\pm = L^2(\mathbb
R_\pm, E)$,
\begin{equation*}
\mathcal H = \mathcal D_- \oplus K \oplus \mathcal D_+
\end{equation*}
Elements of~$\mathcal H$ are represented as
three-component vectors~$(v_-, u,
v_+)$ with
$v_\pm \in \mathcal D_\pm$ and $u \in K$.
The Lax-Phillips theory suggests that the
dilation~$\mathscr L$
restricted
to~$\mathcal D_- \oplus \{0\} \oplus\mathcal D_+$
should be the self-adoint generator~$A$ of
the continuous unitary
group of right shifts~$\exp({\rm i}At) =U(t) : v(x) \mapsto v(x -t)$
in $L^2(\mathbb R, E)$.
By Stone's theorem, one has
\[
{\rm i}A v = \lim\limits_{t\downarrow 0}t^{-1}[U(t) v - v]
= \lim\limits_{t\downarrow 0}t^{-1}\bigl(v(x-t) - v(x)\bigr) = - v^\prime (x),
\]
so that
the generator of~$U(t)$ is the operator~$A : v \mapsto {\rm i}dv/dx$.
Hence, the action of~$\mathscr L$ on the channels~$\mathcal D_\pm$
is defined by
$\mathscr L : (v_-,0, v_+) \mapsto ( {\rm i}v_-^\prime,0,{\rm i}v_+^\prime)$.
The self-adjointness of~$\mathscr L = \mathscr L^*$ and
the requirement~(\ref{eqn:DilationEq}) yield
the form of dilation~$\mathscr L$
as found in~\cite{MR0510053},
\begin{equation}\label{DilationForm}
\mathscr L \left(
\begin{array}{l} v_-
\\ u
\\ v_+
\end{array}\!\!\right)
=
\begin{pmatrix} {\rm i}\dfrac{dv_-}{dx}
\\[0.7em]A u + \dfrac{\alpha}{2} \left[v_+(0) + v_-(0) \right]
\\[0.5em] {\rm i}\dfrac{dv_+}{dx}
\end{pmatrix},
\end{equation}
defined on the domain
\[
\dom(\mathscr L) = \left \{(v_-, u, v_+ ) \in \mathcal H \mid
v_\pm \in W_2^1(\mathbb R_\pm, E), u \in \dom(A),
v_+(0) - v_-(0) = i\alpha u
\right\}
\]
Embedding theorems for the Sobolev space~$W_2^1$
guarantee the existence of boundary values $v_\pm(0)$.
The ``boundary condition''~$v_+(0) - v_-(0) = {\rm i}\alpha u$
can be interpreted
as a concrete form of coupling between the incoming and outgoing channels
$\mathcal D_\pm$ realized by the
imaginary part of~$L$ acting
on $E.$
When $\alpha = 0$, the right hand side of~(\ref{DilationForm}) is the
orthogonal sum of two self-adjoint operators, that is,
the operator~$A$ on $K$
and the operator~${\rm i}d/dx$ acting
in the orthogonal sum of
channels~$L^2(\mathbb R, E )
= \mathcal D_-\oplus\mathcal D_+$.
The characteristic function of $L$ is the contractive
operator-valued
function defined by the formula
\begin{equation}
\label{DissCharF}
S(z) = I_E + {\rm i}\alpha(L^* - zI)^{-1}\alpha : E \to E,
\quad z \in \mathbb C_+
\end{equation}
According to the fundamental result of Adamyan and Arov~\cite{MR0206711}, the
function~$S$ coincides with
the scattering operator of the pair $(\mathscr L,
\mathscr
L_0)$ where $\mathscr L_0$ is defined by (\ref{DilationForm})
with $u = 0$ and $\alpha = 0$.
\subsubsection{Extensions of symmetric operators~\cite{Drogobych}}
Consider the differential expression
\[
\ell y = -y'' +q(x)y
\]
in $K= L^2(\mathbb R_+)$ with a real function~$q$
such that the Weyl limit point case takes place.
Denote by $\varphi$ and $\psi$ the standard solutions to
the equation $\ell y = z y$ with $z
\in \mathbb C_+$, satisfying the boundary conditions
\[
\varphi(0,z) = 0, \quad \varphi^\prime (0,z) = -1,
\quad \psi(0,z) = 1, \quad \psi^\prime(0,z) = 0
\]
Then the Weyl solution~$\chi = \varphi + m_\infty(z)\psi\in L^2(\mathbb R_+)$, where $m_\infty(z)$ is the Weyl
function pertaining to~$\ell$ and
corresponding to the boundary condition~$y(0) = 0$, is defined
uniquely.
The function $m_\infty(z)$ is analytic with positive
imaginary part for $z
\in \mathbb C_+$.
Define the operator $L$ in $K =L^2(\mathbb R_+)$
by the expression~$\ell$
supplied with the non-selfadjoint
boundary condition at $x = 0$
\[
( y^\prime - \left . h y) \right |_{x =0} = 0 , \quad \text{where} \quad \im h
= \frac{\alpha^2}{2}, \quad \alpha > 0
\]
A short calculation ascertains that $L$ is dissipative indeed.
The Pavlov's dilation of $L$ is the operator~$\mathscr L$ in the
space~$\mathcal H = \mathcal D_- \oplus K \oplus \mathcal D_+$,
where $\mathcal D_\pm = L^2(\mathbb R_\pm)$, defined
on elements
$( v_- , u, v_+) \in \mathcal H$
which satisfy
\begin{equation}\label{eqn:ExtensionsDomain}
\begin{aligned}
& v_\pm \in W_2^1(\mathbb R_\pm), \quad u,\ell u \in L^2({\mathbb R}_+),
\\[0.3em]
& \left . u^\prime - h u \right|_0 = \alpha v_-(0), \quad
\left . u^\prime - \bar h u \right|_0 = \alpha v_+(0)
\end{aligned}
\end{equation}
The action of the operator $\mathscr L$ on this domain is set by the formula
\begin{equation}\label{eqn:ExtensionsAction}
\mathscr L \left(
\begin{array}{l} v_-
\\ u
\\ v_+
\end{array}\!\!\right)
=
\begin{pmatrix} {\rm i}\dfrac{dv_-}{dx}
\\[0.8em] \ell u
\\[0.4em] {\rm i}\dfrac{dv_+}{dx}
\end{pmatrix}.
\end{equation}
The characteristic function of~$L$ is the scalar analytic in the
upper half-plane function~$S(z)$
given by
\[
S(z) = \frac{ m_\infty(z) - h }{m_\infty(z) - \bar h}, \quad \im(z) > 0.
\]
Note that since $\im h =\alpha^2/2$, the function $S(z)$ can be
rewritten in a form similar to that of the characteristic
function~(\ref{DissCharF}), i.~e.,
\begin{equation}\label{CharF_h}
S(z) = 1 + {\rm i}\alpha\bigl(\bar h -
m_\infty(z)\bigr)^{-1}\alpha.
\end{equation}
\subsubsection{Pavlov's symmetric form of the dilation}\label{sec:pavlov_model}
According to the general theory~\cite{MR2760647},
once the characteristic function~$S$ is known, the
analysis of the completely non-selfadjoint part of
the operator~$L$ is reduced to the
analysis of~$S$.
Hence, the typical questions of the operator theory
(the spectral analysis, description of invariant
subspaces) are reformulated as problems pertaining
to analytic (operator-valued) functions.
Assume that $L$ is completely non-selfadjoint
and $\rho(L) \cap \mathbb C_\pm \neq \varnothing$.
Let $\mathscr L $ be its minimal self-adjoint dilation,
$S$ being the characteristic function of~$L$.
Owing to the general theory~\cite{MR2760647},
the operator~$L$ is unitary equivalent
to its model acting
in the spectral representation of~$\mathscr L$
in accordance with~(\ref{eqn:DilationEq}).
Recall that the characteristic
function~$S(z)$, $z \in \mathbb C_+$
is analytic in the upper half-plane taking values in
the set of contractions of~$E$,
\[
S(z) : E \to E, \quad \|S(z) \| \leq 1, \quad z \in \mathbb C_+
\]
Due to the operator version of Fatou's
theorem~\cite{MR2760647},
the nontangential boundary values of the
function~$S$
exist in the strong operator topology
almost everywhere on the real line.
Put~$S = S(k): = \slim_{\varepsilon\downarrow 0}
S(k + {\rm i}\varepsilon)$
and~$S^* = S^*(k)
:= \slim_{\varepsilon\downarrow 0}[S(k + {\rm i}\varepsilon)]^*$,
both limits existing for almost all~$k \in \mathbb R$.
The Fatou theorem guarantees that
the operators~$S(k)$ and $S^*(k)$ are
contractions on~$E$
for almost all~$k\in\mathbb R$.
The symmetric form of the dilation
is obtained
by completion
of the dense linear set in~$L^2(E) \oplus L^2(E)$
with respect to the norm
\begin{equation}\label{ModelSpace}
\left\|\binom{\tilde g}{g}\right\|_{\mathscr H}^2 :=
\int\limits_{\mathbb R}
\left\langle
\left(
\begin{array}{cc} I & S^* \\ S & I \end{array}
\right)
\binom{\tilde g}{g},
\binom{\tilde g}{g}
\right\rangle_{ E\oplus E}
dk,
\end{equation}
followed by factorisation by the elements of zero norm.
In the symmetric representation, the incoming and outgoing
subspaces~$\mathcal D_\pm$ admit their simplest possible
form.
On the other hand,
calculations related to
the space~$K$ can meet
certain difficulties, since
the ``weight'' in~(\ref{ModelSpace})
can be singular.
Also note that the
elements of~$\mathscr H$ are not individual functions from
$L^2(E) \oplus L^2(E)$ but rather equivalence classes~\cite{NikolskiiKhrushchev, NikolskiiVasyunin1989}.
%
Despite these complications,
the Pavlov's symmetric model has been widely accepted in
the analysis of non-selfadjoint operators, and in particular of the operators
of mathematical physics.
%
Two alternative and equivalent forms
of the norm $\|\cdot\|_{\mathscr H}$ that
are easy to derive,
\begin{equation*}
\left\|\binom{\tilde g}{g}\right\|_{\mathscr H}^2 =
\left\| S\tilde g + g \right\|_{L^2( E)}^2 +
\left\|\Delta_* g \right\|_{L^2( E)}^2 =
\left\|\tilde g + S^* g \right\|_{L^2( E)}^2 +
\left\|\Delta \tilde g \right\|_{L^2( E)}^2,
\end{equation*}
where $\Delta := \sqrt{I- S^* S}$ and
$\Delta_* := \sqrt{I- S S^*}$, show
that for each $\binom{\tilde g}{g} \in \mathscr H$
expressions
$ S\tilde g + g$, $\tilde g + S^* g$, $\Delta
\tilde g$,
and $\Delta_* g$ are in fact usual square summable
vector-functions
from~$L^2( E)$.
Moreover, due to these equalities the form~(\ref{ModelSpace}) is
positive-definite indeed and thus represents a norm.
The space
\[
\mathscr H = L^2\left(\begin{matrix}
I & S^* \\ S & I
\end{matrix}\right)
\]
with the norm defined
by~(\ref{ModelSpace}) is the space of spectral
representation for the self-adjoint
dilation~$\mathscr L$ of the operator~$L$. Henceforth we will denote the corresponding unitary mapping of $\mathcal{H}$ onto $\mathscr{H}$ by $\Phi$.
It means that
the operator of multiplication by the independent variable
acting on~$\mathscr H$, i.e., the
operator~$f(k) \mapsto k f(k) $, is unitary equivalent
to the dilation~$\mathscr L$.
Hence, for~$z \in\mathbb
C \setminus \mathbb R$, the
mapping~$\binom{\tilde g}{g} \mapsto (k - z)^{-1}
\binom{\tilde g}{g}$
is unitary equivalent to the resolvent~$(\mathscr L-z)^{-1}$
and therefore
\begin{equation*}
(L - zI)^{-1} \simeq\left . P_{\mathscr K} (k - z)^{-1} \right
|_{\mathscr K}, \quad z \in \mathbb C_-,
\end{equation*}
where the $\simeq$ sign is utilised to denote unitary equivalence.
The incoming and outgoing subspaces of the dilation
space~$\mathscr H$ admit the form
\begin{equation*}
{\mathscr D}_+ := \binom{H^2_+(E)}{0}, \quad
{\mathscr D}_- := \binom{0}{H^2_-( E)}, \quad
\mathscr K := \mathscr H \ominus
\left[
{\mathscr D}_+ \oplus {\mathscr D}_-
\right]
\end{equation*}
where $H_2^\pm(E)$ are the Hardy classes of $
E$-valued vector functions analytic in $\mathbb C_\pm$.
As usual~\cite{MR822228}, the functions from vector-valued Hardy
classes~$H_2^\pm(E)$ are identified with their
boundary values existing almost everywhere
on the real line.
They form two complementary mutually orthogonal subspaces
so that
$L^2(E) = H^2_+(E)\oplus H^2_-(E)$.
The image~$\mathscr K$ of $K$ under the spectral
mapping $\Phi$ of the dilation space~$\mathcal H$
to~$\mathscr H$ is the subspace
\begin{equation*
\mathscr K =
\left\{
\binom{\tilde g}{g} \in \mathscr H \,\, :\,\,
\tilde g + S^* g \in H^2_-( E), \,
S \tilde g + g \in H^2_+ ( E)
\right\}
\end{equation*}
The orthogonal projection~$P_\mathscr K$ from
$\mathscr H$ onto
$\mathscr K$ is defined by formula~(\ref{2.1.3.2})
on a dense set of functions from
$L^2(E) \oplus L^2(E)$ in $\mathscr H$
\begin{equation}\label{2.1.3.2}
P_{\mathscr K} \binom{\tilde g}{g} =
\binom{\tilde g - P_+ ( \tilde g + S^* g)}
{ g - P_- ( S \tilde g + g)},
\quad \tilde{g} \in L^2( E), \; g \in L^2( E).
\end{equation}
Here $P_\pm$ are the orthogonal projections of $L^2$
onto the Hardy
classes~$H_2^\pm$.
Further information on model representations can be found in the
series of papers~\cite{NikolskiiKhrushchev,NikolskiiVasyunin1986,
NikolskiiVasyunin1989, NikolskiiVasyunin1998}
and the treatise~\cite{Nikolski}.
\subsubsection{Naboko's functional model of non-selfadjoint operators}
The development of the functional
model approach for
contractions inspired the
search for such models of
non-dissipative operators.
The attempts to follow the blueprints of Sz.-Nagy-Foias
and Lax-Philitps meet serious challenges
rooted in the absence of a proper self-adjoint dilation for
non-dissipative operators: the dilatation in this case
is a self-adjoint operator acting on a space with an
indefinite metric~\cite{Davis}.
Consequently, the characteristic function of a non-dissipative
operator is an analytic operator-function, contractive with
respect to an indefinite metric~\cite{DavisFoias}, which
considerably hinders any further progress in this direction.
We mention the works~\cite{Ball, McEnnis}, the
monograph~\cite{Kuzhel} and references therein
for more details and examples.
An alternative approach was suggested in the late seventies with the
publication of papers~\cite{MR0500225,nabokozapiski2} and
especially~\cite{MR573902} by S.~Naboko who
found a way to represent a non-dissipative
operator in a model space of a suitably
chosen dissipative one.
We refer the reader to the relevant section of the paper on Sergey Naboko's mathematical heritage in the present volume for the details of
the mentioned approach and the relevant references. In the next section of the present paper we outline the main ingredients of an adaptation of the latter to the setting of extensions of symmetric operators, which was developed by Sergey's students.
We mention that this set of techniques allows one to significantly advance the spectral analysis of non-selfadjoint operators, including the definition of the absolutely continuous and singular subspaces and the study of spectral resolutions of identity. In particular, of major importance is the possibility to construct the wave and scattering operators in a natural representation. It should be noted that the self-adjoint scattering theory (and all the major versions of the latter) turns out to be included as a particular case of a much more general non-selfadjoint one.
\subsection{Functional model for a family of
extensions of a symmetric operator}
\label{sec:ryzh-funct-model}
The generic model constructed in~\cite{Mak_Vas, Ryzhov_closed}
lends itself as a powerful and universal tool for the analysis of
(completely) non-selfadjoint and (completely) non-unitary operators.
Since characteristic functions of such operators are
essentially unique and define the operators up to unitary
equivalence, all model considerations are immediately
available once this
function is known.
In many applications, however, the results sought need to be
formulated in terms of the problem itself (i.e., in the natural terms), rather than in the abstract language
of characteristic functions (and their transforms).
One prominent example when the general theory is not
sufficient is the setting of extensions of symmetric operators and the associated setting of operators pertaining to boundary-value problems.
Within this setup, the results are expected to be formulated
as statements concerning
the symmetric operator itself and the
relevant properties of the
extension parameters.
Some results in this direction were obtained by B.~Solomyak
in \cite{Solomyak2}, but the related calculations tend to be rather tedious due to the reduction to the case of contractions which is required.
The extension theory of symmetric operators, especially
differential operators, greatly benefits from the abstract
framework known as the boundary triples theory.
The basic concepts of this operator-theoretic approach
can be found in the textbook~\cite{MR2953553}
by K.~Schm{\"u}dgen.
The recent monograph~\cite{BHS} contains a detailed
treatment of this area.
It is therefore quite natural to utilise this approach in conjunction with the functional models techniques outlined above in the analysis of non-selfadjoint extensions of symmetric operators.
This section briefly outlines the results pertaining to the functional model construction for dissipative and non-dissipative extensions of symmetric operators and the related developments, including an explicit construction of the wave and scattering operators and of the scattering matrices. Since all the considerations in this area are essentially parallel to the ones of Naboko in his development of spectral theory for additive perturbations of self-adjoint operators, one can consider this narrative as a rather detailed exposition of Naboko's ideas and results in a particular case, important for applications.
\subsubsection{Boundary triples}
The fundamentals of the boundary triples theory
have been introduced
in~\cite{BHS, MR2953553}, see also
references therein.
Denote by~$A$ a closed and densely defined symmetric
operator on the separable Hilbert space $H$ with the
domain~$\dom A$, having equal deficiency
indices~$0 < n_{+}(A)= n_{-}(A) \leq\infty$.
\begin{definition}[\cite{MR0365218}]\label{def:bonudaryTriple}
A triple $\{\mathcal{K}, \Gamma_0, \Gamma_1 \}$ consisting of an auxiliary
Hilbert space $\mathcal{K}$ and linear mappings $\Gamma_0, \Gamma_1$ defined
everywhere on $\dom A^*$ is called a \emph{boundary triple} for $A^*$ if the following
conditions are satisfied:
\begin{enumerate}
\item
The abstract Green's formula is valid
\begin{equation}\label{GreenFormula}
(A^*f,g)_H - (f, A^*g)_H = (\Gamma_1 f, \Gamma_0 g)_{\mathcal{K}} -
(\Gamma_0 f, \Gamma_1 g)_{\mathcal{K}},\quad f,g \in \dom A^*
\end{equation}
\item
For any $Y_0, Y_1 \in{\mathcal{K}}$ there exist $f \in
\dom A^*$, such that $\Gamma_0 f = Y_0$, $\Gamma_1 f = Y_1$.
In other words,
the mapping~$f \mapsto \Gamma_0 f \oplus \Gamma_1 f $, $f \in \dom A^*$
to ${\mathcal{K}}\oplus {\mathcal{K}}$ is surjective.
\end{enumerate}
\end{definition}
\noindent
It can be shown~(see \cite{MR0365218}) that a boundary
triple for~$A^*$
exists assuming only~$ n_{+}(A)=n_{-}(A) $.
Note also that a boundary triple
is not unique.
Given any bounded self-adjoint operator~$\Lambda = \Lambda^*$
on $\mathcal K$, the
collection~$\{\mathcal K, \Gamma_0 , \Gamma_1 + \Lambda \Gamma_0\}$
is a
boundary triple for~$A^*$ as well,
provided that~$\Gamma_1 + \Lambda \Gamma_0$
is surjective.
\begin{definition}\label{def:WeylFunction}
Let $\mathscr T = \{\mathcal{K}, \Gamma_0, \Gamma_1 \}$
be a boundary triple
of~$A^*$.
The Weyl function of $A^*$ corresponding to~$\mathscr T$ and
denoted $M(z)$, $z \in\mathbb C\setminus \mathbb R$ is an analytic
operator-function with a positive imaginary part for~$z \in \mathbb C_+$
(i.e., an operator $R$\nobreakdash-function) with values in
the algebra of bounded operators on $\mathcal K$ such that
\[
M(z)\Gamma_0 f_z = \Gamma_1 f_z, \quad f_z \in {\rm ker}(A^* -zI),
\quad z \notin \mathbb R
\]
For $z\in\mathbb C \setminus\mathbb R$ we
have~$(M(z))^* = (M(\bar z))$ and
$\im (z) \cdot \im(M(z)) > 0$.
\end{definition}
\begin{definition}
An extension~$\mathcal A$ of a closed densely defined
symmetric operator~$A$ is
called \emph{almost solvable (a.s.)} and denoted $\mathcal A = A_B$ if
there exist a boundary
triple~$\{\mathcal{K}, \Gamma_0, \Gamma_1 \}$ for~$A^*$
and a bounded operator~$B : \mathcal{K} \to \mathcal K$ defined
everywhere in $\mathcal K$ such that
\begin{equation*}
f \in \dom A_B \iff \Gamma_1 f = B \Gamma_0 f
\end{equation*}
\end{definition}
\noindent
This definition implies the
inclusion~$\dom A_B \subset \dom A^*$
and that~$A_B$ is a restriction of~$A^*$ to
the linear set $\dom A_B := \{f \in \dom A^* : \Gamma_1 f =
B \Gamma_0 f\}$.
In this context, the operator~$B$ plays the role of a
parameter for the
family of extensions~$\{A_B \mid B : \mathcal K \to \mathcal K\}$.
It can be shown (see~\cite{CherKisSilva1} for references)
that if the
deficiency indices $n_\pm(A)$ are equal
and $A_B$ is an almost
solvable extension of~$A$, then
the resolvent set of~$A_B$ is
not empty (i.e. $A_B$ is maximal),
both $A_B$ and $(A_B)^* = A_{B^*}$
are restrictions of $A^*$ to their
domains, and
$A_B$ and $B$
are selfadjont (dissipative) simultaneously.
The spectrum of~$A_B$ coincides with the
set of points~$z_0 \in \mathbb C$ such that $(M(z_0) -B)^{-1}$
does not admit analytic continuation into it.
\subsubsection{Characteristic functions}\label{sec:char_functions_ext}
Assume that the parameter~$B$ of
an almost solvable extension $A_B$
is completely non-selfadjoint.
It can be represented as the sum of its real
and imaginary part
\[
B = B_R + {\rm i}B_I, \quad B_R =B_R^*, \quad B_I = B_I^*
\]
These parts are well defined since $B$ is bounded.
The Green's formula implies that
for~$B_I \neq 0$ the imaginary part
of~$A_B$ (in the sense of its form) is non-trivial,
i.~e.,~$\im(A_B u, u) \neq 0$
at least for some $u \in\dom(A_B)$.
Hence~$A_B$ in this case is not a
self-adjoint operator.
It appears highly plausible that complete non-selfadjointness of~$A_B$
can be derived solely from complete non-selfadjointness of~$B$, assuming that~$A$
has no reducing self-adjoint parts.
However, no direct proof of this assertion
seems to be available in the existing
literature.
According to~(\ref{DissCharF}),
the characteristic function of~$B$
has the form
\[
\Theta_B(z) = I_E + {\rm i}J\alpha(B^* - zI)^{-1}\alpha : E \to E,
\quad z \in \rho(B^*),
\]
where $\alpha:=\sqrt{2|B_I|}$, $J:={\rm sign} B_I,$
and~$E:=\clos\ran(\alpha)$.
On the other hand, direct calculations according
to~\cite{Strauss1960} lead to the following
representation for the characteristic
function~$\Theta_{A_B}: E \to E$ of the
non-selfadjoint part of the extension~$A_B$
\[
\Theta_{A_B} = I_E + {\rm i}J \alpha (B^* - M(z))^{-1}\alpha,
\quad z \in \rho(A_B^*).
\]
These two formulae confirm an earlier observation
that goes back to B.~S.~Pavlov's work~\cite{Drogobych},
see~(\ref{CharF_h}) above.
The function~$\Theta_{A_B}$ is obtained
from $\Theta_B$ by the
substitution of~$M(z)$ for $zI_E$,
$z\in \mathbb C_+$
\[
\Theta_{A_B}(z) = \Theta_B(M(z)), \quad z \in \rho(A_B^*).
\]
Alongside $A_B$
introduce the dissipative almost solvable extension~$A_+$
parameterized by~$B_+:= B_R + {\rm i}|B_I|$. Note that the characteristic function~$S$
of~$A_+$
is given by ({\it cf.} (\ref{DissCharF}))
\begin{equation}\label{eq:charf_ext}
S(z) = I_E + {\rm i}\alpha(B_+^* - zI)^{-1}\alpha : E \to E,
\quad z \in \mathbb{C}_+.
\end{equation}
Calculations of~\cite{MR2330831} (cf. \cite{MR0500225}) show that the
characteristic functions of $A_B$ and
$A_+$ are
related via an
operator linear-fractional transform known as
Potapov-Ginzburg transformation,
or PG-trans\-form~\cite{AzizovIokhvidov}.
This fact is essentially geometric.
It connects contractions on Kre\u{\i}n spaces
(i.e., the spaces with an indefinite metric defined
by the involution~$J = J^* = J^{-1}$) with contractions
on Hilbert spaces endowed with the regular metric.
The PG-trans\-form is invertible and
the following assertion pointed out in~\cite{MR0500225}
holds.
\begin{proposition}
\!\!The characteristic function~$\Theta_{A_B}$ is $J$\nobreakdash-contractive on its domain
and the PG\nobreakdash-trans\-form
maps it to the contractive characteristic
function $S$ of~$A_+$
as follows:
\begin{equation}\label{PGtransform}
\Theta_{A_B} \mapsto S = - (\chi^+ - \Theta_{A_B} \chi^-)^{-1}(\chi^- - \Theta_{A_B} \chi^+),
\qquad
S \mapsto \Theta_{A_B} = (\chi^- + \chi^+ S)(\chi^+ + \chi^- S)^{-1}
\end{equation}
where $\chi^\pm =(I_E \pm J)/2$ are orthogonal projections onto
subspaces of $\chi^+ E$ and $\chi^- E$, respectively.
\end{proposition}
\noindent
It appears somewhat unexpected that two operator-valued functions
connected by formulae~(\ref{PGtransform}) can be explicitly
written down in terms of their ``main
operators''~$A_{B}$ and $A_+$.
This relationship between the characteristic
functions of~$A_B$ and
$A_+$ goes in fact
much deeper, see~\cite{Arov,AzizovIokhvidov}.
In particular, the self-adjoint dilation
of~$A_+$
and the $J$-self-adjoint dilation of $A_B$ are also related via a suitably
adjusted version of the PG\nobreakdash-transform.
Similar statements hold for the corresponding
linear systems or ``generating operators'' of the
functions~$\Theta_{A_B}$ and $S$, cf.~\cite{Arov,AzizovIokhvidov}.
This fact is crucial for the construction of a model of a
general closed and densely defined
non-selfadjoint operator,
see~\cite{Ryzhov_closed}.
%
\subsubsection{Functional model for a family of
extensions}\label{sect:BoundaryTriplesModel}
Formulae of the previous secion are essentially the same as the
formulae of \cite{MR0500225}
connecting characteristic functions of
non-dissipative and dissipative operators.
Reasoning by analogy, this suggests an existence
of certain identities that would connect the resolvents of $A_B$ and $A_+$
corresponding to
parameters~$B = B_R + iB_I$ and $B_+ = B_R + i|B_I|$.
Such identities indeed exist; they are the
celebrated Kre\u{\i}n formulae for
resolvents of two extensions of a symmetric operator.
Their variant is readily derived within the framework of boundary
triplets, see~\cite{MR2330831} for calculations,
where all details of the following results can also be found.
\smallskip
The functional model of the dissipative extension~$A_+$ begins
with the derivation of its minimal
selfadjoint dilation~$\mathscr A$.
It is constructed following the recipe of
B. Pavlov~\cite{MR0365199,Drogobych,MR0510053}
and takes a form quite
similar to (\ref{eqn:ExtensionsDomain}), (\ref{eqn:ExtensionsAction})
\[
\mathscr A \begin{pmatrix}
v_- \\ u \\ v_+
\end{pmatrix} =
\begin{pmatrix}
{\rm i}v^\prime_- \\[0.3em] A_+^* u \\[0.3em] {\rm i}v^\prime_+
\end{pmatrix}, \quad
\begin{pmatrix}
v_- \\ u \\ v_+
\end{pmatrix} \in \dom(\mathscr A)
\]
where $\dom(\mathscr A)$ consists of vectors~$(v_-,u,v_+) \in \mathcal H =
\mathcal D_- \oplus H\oplus \mathcal D_+$, with $v_\pm \in W_2^1(\mathbb R_\pm,
E)$, $u \in \dom(A_+^*)$ under two ``boundary
conditions'' imposed on $v_\pm$ and $u$:
\[
\Gamma_1 u - B_+ \Gamma_0 u = \alpha v_-(0), \quad
\Gamma_1 u - B_+^* \Gamma_0 u = \alpha v_+(0)
\]
The functional model construction for $A_+$ follows the recipe
by S.~Naboko~\cite{MR573902}. The following theorem holds.
\begin{theorem}
\label{ModelTheoremDiss}
There exists a mapping $\Phi$ from the
dilation space~$\mathcal H$ onto Pavlov's model space~$\mathscr H$
defined by~(\ref{ModelSpace})
with the following properties
\begin{enumerate}
\item
$\Phi$ is isometric.
\item
$\tilde g + S^* g = \mathscr F_+ h$,
$ S \tilde g + g = \mathscr F_- h$, where
$\binom{\tilde g}{g} = \Phi h$, $ h \in \mathcal H$
\ite
$
\Phi \circ (\mathscr L - zI )^{-1} = (k - z)^{-1} \circ \Phi,
\quad z \in \mathbb C \setminus\mathbb R
$
\item
$ \Phi\mathcal H = \mathscr H$, \quad $\Phi \mathcal D_\pm = {\mathscr
D}_\pm$, \quad $\Phi \mathcal K = \mathscr K$
\item
$\mathscr F_\pm \circ (\mathscr L - zI)^{-1} = (k - z)^{-1} \circ
\mathscr F_\pm,
\quad z\in \mathbb C \setminus \mathbb R$.
\end{enumerate}
where bounded maps~$\mathscr F_\pm : \mathcal H \to L^2(\mathbb R, E)$
are defined by the formulae
\begin{align*}
\mathscr F_+ : h & \mapsto - \frac{1}{\sqrt{2\pi}}\,
\alpha \Gamma_0( A_+ - k + {\rm i}0)^{-1}
u +
S_+^*(k) {\hat{v}}_-(k) + \hat{v}_+(k),
\\
\mathscr F_- : h & \mapsto - \frac{1}{\sqrt{2\pi}}\,
\alpha \Gamma_0(A_+^*-k-{\rm i}0)^{-1}
u + {\hat{v}}_-(k) + S_+(k) \hat{v}_+(k),
\end{align*}
where $h = (v_-, u, v_+) \in \mathcal H$ and $\hat v_\pm$ are the Fourier
transforms of $v_\pm \in L^2(\mathbb R_\pm, E)$.
\end{theorem}
Using the
Kre\u{\i}n formulae, which play the same role as
Hilbert resolvent identities in the case of additive perturbations,
one can obtain results similar to those of \cite{MR573902} and obtain an explicit description of $(A_+-z)^{-1}$ in the functional model representation.
An analogue of this result
for more general extensions of~$A$
corresponding to a choice of parameter~$B$ in the
form~$B_R+\alpha\varkappa\alpha/2$
with bounded~$\varkappa:E\to E$ and~$B_R = B_R^*$
is proven along the same lines.
This program was realized in~\cite{Ryzhov_closed} for a particular
case~$\varkappa = {\rm i}J$ and in~\cite{CherKisSilva1}
for the family of
extensions~$A_\varkappa$ parameterized
by~$B_\varkappa =\alpha \varkappa\alpha/2.$
The latter form of extension parameter
utilizes the possibility of ``absorbing''
the part~$B_R= B_R^*$ into
the map~$\Gamma_1$, i.~e., passing from the
boundary triple~$\{\mathcal K, \Gamma_0, \Gamma_1\}$
to the triple~$\{\mathcal K,
\Gamma_0, \Gamma_1 + B_R \Gamma_0\}$.
\subsubsection{Smooth vectors and the absolutely continuous subspace}
\label{sec:smooth-vectors-scatt}
Here we characterise the absolutely continuous spectral subspace for an almost solvable extension of a densely defined symmetric operator with equal (possibly infinite) deficiency indices. The procedure we follow is heavily influenced by the ideas of Sergey Naboko, see \cite{MR573902, MR1252228} and is carried out essentially in parallel to the exposition of \cite{MR573902}. In contrast to the mentioned works, dealing with additive perturbations of self-adjoint operators, we are dealing with the case of extensions, self-adjoint and non-self-adjoint alike. The narrative below follows the argument presented in our papers \cite{CherednichenkoKiselevSilva,CherKisSilva1}.
Since we are not limiting the consideration to the case of self-adjoint operators, we first require the notion of the absolutely continuous spectral subspace applicable in the non-self-adjoint setup. In the functional model space $\mathscr{H}$ introduced in Section \ref{sec:pavlov_model} constructed based on the characteristic function $S(z)$ introduced in Section \ref{sec:char_functions_ext}
consider two subspaces $\mathscr{N}^\varkappa_\pm$ defined as follows:
\begin{equation*}
\mathscr{N}^\varkappa_\pm:=\left\{\binom{\widetilde{g}}{g}\in\mathscr{H}:
P_\pm\left(\chi_\varkappa^+(\widetilde{g}+S^*g)+\chi_\varkappa^-(S\widetilde{g}
+g)\right)=0\right\},
\end{equation*}
where \begin{equation*}
\chi_\varkappa^\pm:=\frac{I\pm{\rm i}\varkappa}{2}.
\end{equation*}
and $P_\pm$ are orthogonal projections onto their respective Hardy classes, as above.
These subspaces have a characterisation in terms of the resolvent of the operator $A_\varkappa.$ This, again, can be seen as a consequence of a much more general argument (see {\it e.g.} \cite{Ryzhov_closed, Ryzh_ac_sing}).
\begin{theorem}
\label{lem:similar-to-naboko-thm-4}
Suppose that $\ker \alpha =0.$ The following characterisation holds:
\begin{equation*}
\mathscr{N}^\varkappa_\pm=\left\{\binom{\widetilde{g}}{g}\in\mathscr{H}:
\Phi(A_{\varkappa}-z I)^{-1}\Phi^*P_K\binom{\widetilde{g}}{g}=
P_K\frac{1}{k-z}\binom{\widetilde{g}}{g}
\text{ for all } z\in\mathbb{C}_\pm\right\}\,.
\end{equation*}
Here $\Phi$ denotes the unitary mapping of the dilation space $\mathcal{H}$ onto $\mathscr{H}$, as above.
\end{theorem}
Consider the counterparts of $\mathscr{N}^\varkappa_\pm$ in the original
Hilbert space $H:$
\begin{equation*}
\widetilde{N}_\pm^\varkappa:=\Phi^*P_K\mathscr{N}^\varkappa_\pm\,,
\end{equation*}
which are linear sets albeit not necessarily subspaces.
In a way similar to \cite{MR573902}, one introduces the set
\begin{equation*}
\widetilde{N}_{\rm e}^\varkappa:=\widetilde{N}_+^\varkappa\cap
\widetilde{N}_-^\varkappa
\end{equation*}
of so-called \emph{smooth vectors} and its closure $N_{\rm e}^\varkappa:=\clos(\widetilde{N}_{\rm e}^\varkappa).$
The next assertion ({\it cf. e.g.} \cite{Ryzhov_closed, Ryzh_ac_sing}, for the case of general non-selfadjoint operators), is an alternative non-model characterisation of the linear sets
$\widetilde{N}_\pm^\varkappa$.
\begin{theorem}
\label{lem:on-smooth-vectors-other-form}
The sets $\widetilde{N}_\pm^\varkappa$ are described as follows:
\begin{equation*}
\widetilde{N}_\pm^\varkappa=\{u\in{\cal H}: \alpha\Gamma_0(A_{\varkappa}-z
I)^{-1}u\in H^2_\pm(E)\}.
\end{equation*}
\end{theorem}
Moreover, one shows that for the functional model image of $\tilde{N}^\varkappa_{\rm e}$ the following representation holds:
\begin{align}
&\Phi\widetilde{N}^\varkappa_{\rm e}=\biggl\{P_K\binom{\widetilde{g}}{g}\in\mathscr{H}:\nonumber
\\
&\binom{\widetilde{g}}{g}\in\mathscr{H}\ {\rm satisfies}\
\Phi(A_{\varkappa}-z I)^{-1}\Phi^*P_K\binom{\widetilde{g}}{g}=
P_K\frac{1}{k-z}\binom{\widetilde{g}}{g}\ \ \ \forall\,z\in{\mathbb C}_-\cup{\mathbb C}_+\biggr\},
\label{New_Representation}
\end{align}
which motivates the term ``the set of smooth vectors'' used for $\widetilde{N}^\varkappa_{\rm e}$.
(Note that the inclusion of the right-hand side of (\ref{New_Representation}) into $\Phi\tilde{N}^\varkappa_{\rm e}$ follows immediately from Theorem \ref{lem:similar-to-naboko-thm-4}.)
The above Theorem together with Theorem \ref{thm:on-smooth-vectors-a.c.equality} below motivates generalising the notion of the absolutely continuous subspace ${\cal H}_{\rm ac}(A_{\varkappa})$ to the case of non-selfadjoint extensions $A_\varkappa$ of a symmetric operator $A,$ by identifying it with the set $N^\varkappa_{\rm e}.$ This generalisation follows in the footsteps of the corresponding definition by Naboko \cite{MR573902} in the case of additive perturbations (see also \cite{Ryzhov_closed, Ryzh_ac_sing} for the general case).
\begin{definition}
\label{abs_cont_subspace}
For a symmetric operator $A,$ in the case of a non-selfadjoint extension $A_\varkappa$ the absolutely continuous subspace ${\cal H}_{\rm ac}(A_{\varkappa})$ is defined by the formula ${\cal H}_{\rm ac}(A_{\varkappa}):=N^\varkappa_{\rm e}.$
In the case of a self-adjoint extension $A_\varkappa$, we understand ${\cal H}_{\rm ac}(A_{\varkappa})$ in the sense of the classical definition of the absolutely continuous subspace of a self-adjoint operator.
\end{definition}
It turns out that in the case of self-adjoint extensions a rather mild additional condition guarantees that the non-self-adjoint definition above is equivalent to the classical self-adjoint one. Namely, we have the following
\begin{theorem}
\label{thm:on-smooth-vectors-a.c.equality}
Assume that $\varkappa=\varkappa^*,$ ${\rm ker}(\alpha)=\{0\}$
and let $\alpha\Gamma_0(A_{\varkappa}-z I)^{-1}$ be a Hilbert-Schmidt
operator for at least one point $z\in\rho(A_\varkappa)$. If
$A$ is completely non-selfadjoint, then the definition ${\cal H}_{\rm ac}(A_{\varkappa})=N^\varkappa_{\rm e}$ is
equivalent to the classical definition of the absolutely continuous subspace of a self-adjoint operator, {\it i.e.}
\begin{equation*}
N_{\rm e}^\varkappa = {\cal H}_{\rm ac}(A_{\varkappa})\,.
\end{equation*}
\end{theorem}
\begin{remark}
Alternative conditions, which are even less restrictive in general, that guarantee the validity of the assertion of Theorem
\ref{thm:on-smooth-vectors-a.c.equality} can be obtained along the lines of \cite{MR1252228}.
\end{remark}
\subsubsection{Wave and scattering operators}
\label{sec:wave-operators}
The results of the preceding section allow us, see \cite{CherednichenkoKiselevSilva,CherKisSilva1}, to calculate the wave
operators for any pair $A_{\varkappa_1},A_{\varkappa_2}$, where
$A_{\varkappa_1}$ and $A_{\varkappa_2}$ are two different extensions of a symmetric operator $A$, under the additional
assumption that the operator $\alpha$ has a trivial kernel.
For simplicity,
in what follows we set $\varkappa_2=0$ and write
$\varkappa$ instead of $\varkappa_1$. Note that $A_0$ is a self-adjoint
operator, which is convenient for presentation purposes.
In order to compute the wave operators of this pair, one first establishes the model representation for the function
$\exp(iA_\varkappa t)$, $t\in\mathbb{R}$, of the
operator $A_\varkappa,$ evaluated on the set of smooth vectors $\widetilde{N}_{\rm e}^\varkappa.$ Due to \eqref{New_Representation}, it is easily shown that on this set $\exp(iA_\varkappa t)$ acts as an operator of multiplication by $\exp(ikt)$. We then utilise the following result.
\begin{proposition}(\cite[Section 4]{MR573902})
\label{prop:pre-wave-op}
If $\Phi^*P_K\binom{\widetilde{g}}{g}\in\widetilde{N}_{\rm e}^\varkappa$
and $\Phi^*P_K\binom{\widehat{g}}{g}\in\widetilde{N}_{\rm e}^0$
(with the same element\footnote{Despite the fact that
$\binom{\widetilde{g}}{g}\in\mathscr{H}$ is nothing but a symbol,
still $\widetilde{g}$ and $g$ can be identified with vectors in
certain $L^2(E)$ spaces with operators ``weights'', see
details below in Section~\ref{sec:spectral-repr-ac}. Further, we recall that even then
for $\binom{\widetilde{g}}{g}\in\mathscr{H}$, the components $\widetilde{g}$ and $g$ are not, in general, \emph{independent} of each other.} $g$), then
\begin{equation*}
\norm{\exp(-{\rm i}A_\varkappa t)\Phi^*P_K\binom{\widetilde{g}}{g}-\exp(-{\rm i}A_0
t)\Phi^* P_K\binom{\widehat{g}}{g}}_{\mathscr H}\convergesto{t}{-\infty}0.
\end{equation*}
\end{proposition}
It follows from Proposition \ref{prop:pre-wave-op} that whenever $\Phi^*P_K\binom{\widetilde{g}}{g}\in\widetilde{N}_{\rm e}^\varkappa$
and $\Phi^*P_K\binom{\widehat{g}}{g}\in\widetilde{N}_{\rm e}^0$ (with the same second component $g$), formally one has
\begin{align*}
\lim_{t\to-\infty}{\rm e}^{{\rm i}A_0t}{\rm e}^{-{\rm i}A_\varkappa t}\Phi^*P_K\binom{\widetilde{g}}{g}&=\Phi^*P_K\binom{\widehat{g}}{g}\\
&=\Phi^*P_K\binom{-(I+S)^{-1}(I+S^*)g}{g}\,.
\end{align*}
In view of the classical definition of the wave operator of a pair of self-adjoint operators, see {\it e.g.} \cite{MR0407617},
\begin{equation*}
W_\pm(A_0,A_\varkappa):=\slim_{t\to\pm\infty}{\rm e}^{{\rm i}A_0t}{\rm e}^{-{\rm i}A_\varkappa t}P_{\rm ac}^\varkappa,
\end{equation*}
where $P_{\rm ac}^\varkappa$ is the projection onto the absolutely continuous subspace of $A^\varkappa,$ we obtain that, at least formally, for $\Phi^*P_K\binom{\widetilde{g}}{g}\in\widetilde{N}_{\rm e}^\varkappa$ one has
\begin{equation}
\label{eq:formula-0-kappaw-}
W_-(A_0,A_\varkappa)\Phi^*P_K\binom{\widetilde{g}}{g}=\Phi^*P_K\binom{-(I+S)^{-1}(I+S^*)g}{g}\,.
\end{equation}
By considering the case $t\to+\infty$, one also obtains
\begin{equation*}
W_+(A_0,A_\varkappa)\Phi^*P_K\binom{\widetilde{g}}{g}=
\lim_{t\to+\infty}{\rm e}^{{\rm i}A_0t}{\rm e}^{-{\rm i}A_\varkappa t}\Phi^*P_K\binom{\widetilde{g}}{g}
=\Phi^*P_K\binom{\widetilde{g}}{-(I+S^*)^{-1}(I+S)\widetilde{g}}
\end{equation*}
again for $\Phi^*P_K\binom{\widetilde{g}}{g}\in\widetilde{N}_{\rm e}^\varkappa$.
Further, the definition of the wave operators $W_\pm(A_\varkappa,A_0)$
\begin{equation*}
\norm{{\rm e}^{-{\rm i}A_\varkappa
t}W_\pm(A_\varkappa,A_0)\Phi^*P_K\binom{\widetilde{g}}{g}-{\rm e}^{-{\rm i}A_0t}\Phi^*P_K\binom{\widetilde{g}}{g}
}_{\mathscr H}\convergesto{t}{\pm\infty}0
\end{equation*}
yields, for all $\Phi^*P_K\binom{\widetilde{g}}{g}\in\widetilde{N}_{\rm e}^0,$
\begin{equation*}
W_-(A_\varkappa,A_0)\Phi^*P_K\binom{\widetilde{g}}{g}=\Phi^*P_K\binom{-(I+\chi_\varkappa^-(S-I))^{-1}(I+\chi_\varkappa^+(S^*-I))g}{g}
\end{equation*}
and
\begin{equation}
\label{eq:formula-kappa-0w+}
W_+(A_\varkappa,A_0)\Phi^*P_K\binom{\widetilde{g}}{g}=\Phi^*P_K
\binom{\widetilde{g}}{-(I+\chi_\varkappa^+(S^*-I))^{-1}
(I+\chi_\varkappa^-(S-I))\widetilde{g}}.
\end{equation}
In order to rigorously justify the above formal argument, {\it i.e.} in order to prove the existence and completeness of the wave operators, one needs to first show that
the right-hand sides of the formulae \eqref{eq:formula-0-kappaw-}--\eqref{eq:formula-kappa-0w+} make sense on dense subsets of the corresponding absolutely continuous subspaces, which is done in a similar way to \cite {MR1252228}. Below, we show how this argument works in relation to the wave operator \eqref{eq:formula-0-kappaw-} only, skipping the technical details in view of making the exposition more transparent.
Let $S(z)-I$ be of the class $\mathfrak {S}_\infty(\overline{ \mathbb {C}}_+),$ {\it i.e.} a compact analytic operator function in the upper half-plane up to the real line. Then so is $(S(z)-I)/2$, which is also uniformly bounded in the upper half-plane along with $S(z)$. We next use the result of \cite[Theorem 3]{MR1252228} about the non-tangential boundedness of operators of the form $(I+T(z))^{-1}$ for $T(z)$ compact up to the real line. We infer that, provided $(I+(S(z_0)-I)/2)^{-1}$ exists for some $z_0\in \mathbb C_+$ (and hence, see \cite{Brodski}, everywhere in $\mathbb C_+$ except for a countable set of points accumulating only to the real line), one has non-tangential boundedness of $(I+(S(z)-I)/2)^{-1}$, and therefore also of $(I+S(z))^{-1}$, for almost all points of the real line.
On the other hand, the latter inverse can be computed in $\mathbb C_+$:
\begin{equation}\label{id1}
\bigl(I+S(z)\bigr)^{-1}=\frac{1}{2}\bigl(I+ {\rm i}\alpha M(z)^{-1}\alpha/2\bigr).
\end{equation}
It follows from (\ref{id1}) and the analytic properties of $M(z)$ that the inverse $(I+S(z))^{-1}$ exists everywhere in the upper half-plane. Thus, Theorem 3 of \cite{MR1252228} is indeed applicable, which yields that
$(I+S(z))^{-1}$ is $\mathbb R$-a.e. nontangentially bounded and, by the operator generalisation of the Calderon theorem (see \cite{Calderon}), which was extended to the operator context in \cite[Theorem 1]{MR1252228}, it admits measurable non-tangential limits in the strong operator topology almost everywhere on $\mathbb R$. As it is easily seen, these limits must then coincide with $(I+ S(k))^{-1}$ for almost all $k\in \mathbb R$.
Then the correctness of the formula (\ref{eq:formula-0-kappaw-}) for the wave operators follows: indeed, consider $\mathbbm{1}_n(k),$ the indicator of the set $\{k\in \mathbb R: \|(I+S(k))^{-1}\|\leq n\}.$
Clearly, $\mathbbm{1}_n(k)\to 1$ as $n\to \infty$ for almost all $k\in \mathbb R$. Next, suppose that $P_K(\tilde g, g)\in \tilde N_{\rm e}^\varkappa$. Then $P_K\mathbbm{1}_n(\tilde g, g)$ is shown to be a smooth vector as well as
$$
\binom{-(I+S)^{-1}\mathbbm{1}_n (I+S^*) g}{\mathbbm{1}_n g}\in \mathscr H.
$$
It follows, by the Lebesgue dominated convergence theorem, that the set of vectors $P_K\mathbbm{1}_n(\tilde g, g)$ is dense in $N_{\rm e}^\varkappa$.
Thus the following theorem holds.
\begin{theorem}
\label{thm:existence-completeness-wave-operators}
Let $A$ be a closed, symmetric, completely nonselfadjoint operator
with equal deficiency indices and consider its extension $A_\varkappa$ under the assumptions that ${\rm ker}(\alpha)=\{0\}$ and that $A_\varkappa$ has at least one regular point in ${\mathbb C}_+$ and in ${\mathbb C}_-.$ If $S-I\in\mathfrak {S}_\infty(\overline{ \mathbb {C}}_+),$
then the wave operators
$W_\pm(A_0,A_\varkappa)$ and $W_\pm(A_\varkappa,A_0)$ exist on dense sets in $N_{\rm e}^\varkappa$ and
$\mathcal{H}_{\rm ac}(A_0)$, respectively, and are given by the formulae
(\ref{eq:formula-0-kappaw-})--(\ref{eq:formula-kappa-0w+}). The
ranges of $W_\pm(A_0,A_\varkappa)$ and $W_\pm(A_\varkappa,A_0)$ are dense in $\mathcal{H}_{\rm ac}(A_0)$ and $N_{\rm e}^\varkappa,$ respectively.\footnote{In the case when $A_\varkappa$ is self-adjoint, or, in general, the named wave operators are bounded, the claims of the theorem are equivalent (by the classical Banach-Steinhaus theorem) to the statement of the existence and completeness of the wave operators for the pair $A_0, A_\varkappa.$ Sufficient conditions of boundedness of these wave operators are contained in {\it e.g.} \cite[Section 4]{MR573902}, \cite {MR1252228} and references therein.}
\end{theorem}
\begin{remark}
\label{long_remark}
1. The condition $S(z)-I\in \mathfrak S_\infty(\overline{ \mathbb {C}}_+)$ can be replaced by the following equivalent condition: $\alpha M(z)^{-1}\alpha$ is nontangentially bounded almost everywhere on the real line, and $\alpha M(z)^{-1}\alpha\in \mathfrak S_\infty(\overline{ \mathbb {C}}_+)$ for $\Im z\geq 0$.
2. The latter condition is satisfied \cite{Krein}, as long as the scalar function $\|\alpha M(z)^{-1}\alpha\|_{\mathfrak S_p}$ is nontangentially bounded almost everywhere on the real line for some $p<\infty,$ where ${\mathfrak S_p},$
$p\in(0, \infty],$ are the standard Schatten -- von Neumann classes of compact operators.
3. An alternative sufficient condition is the condition $\alpha\in\mathfrak S_2$ (and therefore $B_\varkappa \in \mathfrak S_1$), or, more generally, $\alpha M(z)^{-1}\alpha\in\mathfrak S_1,$ see \cite{MR1036844} for details.
\end{remark}
Finally, the scattering operator $\Sigma$ for the pair $A_\varkappa,$
$A_0$
is defined by
$$
\Sigma=W_{+}^{-1}(A_\varkappa,A_0) W_{-}(A_\varkappa,A_0).
$$
The above formulae for the wave operators lead ({\it cf.} \cite{MR573902}) to the following formula for the action of $\Sigma$ in the model representation:
\begin{equation}
\Phi\Sigma\Phi^*P_K \binom{\tilde g}{g}=
P_K \binom
{-(I+\chi_\varkappa^-(S-I))^{-1}(I+\chi_\varkappa^+(S^*-I))g}
{(I+S^*)^{-1}(I+S)(I+\chi_\varkappa^-(S-I))^{-1}(I+\chi_\varkappa^+(S^*-I))g},
\label{last_formula}
\end{equation}
whenever $\Phi^*P_K \binom{\tilde g}{g}\in\widetilde N_{\rm e}^0$. In fact, as explained
above, this representation holds on a dense linear set in
$\widetilde N_{\rm e}^0$ within the conditions of Theorem \ref{thm:existence-completeness-wave-operators}, which guarantees that
all the objects on the right-hand side of the formula (\ref{last_formula}) are correctly defined.
\subsubsection{Spectral representation for the absolutely continuous part of the operator $A_0$ and the scattering matrix}
\label{sec:spectral-repr-ac}
The identity
$$
\biggl\|P_K\binom{\tilde g}{g}\biggr\|^2_{\mathscr H}=\bigl\langle(I-S^*S)\tilde g, \tilde g \bigr\rangle
$$
which is derived in \cite[Section 7]{MR573902} for all
$P_K \binom{\tilde g}{g}\in\widetilde N_{\rm e}^0$ allows us to consider the
isometry
$F: \Phi\widetilde N_{\rm e}^0\mapsto L^2(E; I-S^*S)$
defined by the formula
\begin{equation*}
FP_K \binom{\tilde g}{g}=\tilde g.
\end{equation*}
Here $L^2(E; I-S^*S)$ is the Hilbert space of $E$-valued functions on
$\mathbb R$ square summable with the matrix ``weight'' $I-S^*S$.
Under the assumptions of Theorem \ref{thm:existence-completeness-wave-operators} one can show that the range of the operator $F$ is dense in the space $L^2(E; I-S^*S)$.
Thus, the operator $F$ admits an extension to the
unitary mapping between $\Phi N_{\rm e}^0$ and
$L^2(E; I-S^*S)$.
It follows that the self-adjoint operator $(A_0-z)^{-1}$
considered on $\widetilde N_{\rm e}^0$ acts as the multiplication by
$(k-z)^{-1},$ $k\in{\mathbb R},$
in $L^2(E; I-S^*S)$. In particular,
if one considers the absolutely continuous ``part'' of the
operator $A_0$, namely the operator $A_0^{({\rm e})}:=A_0|_{N_{\rm e}^0},$ then $F\Phi A_0^{({\rm e})}\Phi^*F^*$
is the operator of multiplication by the independent variable in the
space $L^2(E; I-S^*S)$.
In order to obtain a spectral representation from the above result, it
is necessary to diagonalise the ``weight'' in the definition of the
above $L^2$-space. The corresponding transformation is straightforward when, e.g.,
$\alpha=\sqrt{2}I.$ (This choice of $\alpha$ satisfies the conditions of Theorem \ref{thm:existence-completeness-wave-operators} {\it e.g.} when the boundary space $\mathcal K$ is finite-dimensional). In this particular case one has
\begin{equation*}
S=(M-{\rm i}I)(M+{\rm i}I)^{-1},
\end{equation*}
and
consequently
\begin{equation*}
I-S^*S=-2i (M^*-{\rm i}I)^{-1}(M-M^*)(M+{\rm i}I)^{-1}.
\end{equation*}
Introducing the unitary transformation
\begin{equation*}
G: L^2(E; I-S^*S)\mapsto L^2(E; -2{\rm i}(M-M^*)),
\end{equation*}
by the
formula $g\mapsto (M+{\rm i}I)^{-1}g$, one arrives at the fact that
$
GF\Phi A_0^{({\rm e})}\Phi^* F^*G^*$
is the operator of multiplication by the independent
variable in the space $L^2(E;-2{\rm i}(M-M^*))$.
\begin{remark}
The weight $M^*-M$ can be assumed to be naturally diagonal in many physically relevant settings, including the setting of quantum graphs considered in Section \ref{sec:theor-appl-funct}.
\end{remark}
The above result
only pertains to the absolutely
continuous part of the self-adjoint operator $A_0$, unlike {\it e.g.}
the passage to the classical von Neumann direct integral, under which
the whole of the self-adjoint operator gets mapped to the
multiplication operator in a weighted $L^2$-space (see {\it e.g.}
\cite[Chapter 7]{MR1192782}). Nevertheless, it proves useful in scattering theory, since it
yields an explicit expression for the scattering matrix $\widehat{\Sigma}$ for the pair
$A_\varkappa,$ $A_0,$ which is the image of the scattering operator
$\Sigma$ in the spectral representation of the operator $A_0.$ Namely, one arrives at:
\begin{theorem}
The following formula holds:
\begin{equation}
\label{scat2}
\widehat{\Sigma}=GF\Sigma(GF)^*=
(M-\varkappa)^{-1}(M^*-\varkappa)(M^*)^{-1}M,
\end{equation}
where the right-hand side
represents the operator of multiplication by the corresponding function in the space $L^2(E;-2{\rm i}(M-M^*))$.
\end{theorem}
\subsection{Functional models for operators of boundary value problems}
\label{sec:funct-model-family}
The surjectivity condition in Definition~\ref{def:bonudaryTriple}
is a strong limitation that excludes many important problems
for extensions of symmetric operators with
infinite deficiency indices.
The standard textbook version of a boundary value
problem for the Laplace operator in a
bounded domain~$\Omega \subset \mathbb R^3$
with smooth boundary~$\partial \Omega$
is a typical example.
The ``natural'' boundary maps $\Gamma_0$ and $\Gamma_1$ are
two trace
operators~$\Gamma_0 : u \mapsto u|_{\partial\Omega}$,
$\Gamma_1 : u \mapsto -\partial u/\partial n|_{\partial\Omega}$,
where $\partial/\partial n$ denotes the derivative along the
exterior normal to the boundary~$\partial\Omega$.
The ranges of these operators do not coincide
with~$\mathcal H = L^2(\Omega)$
(the simplest possible Hilbert space
of functions defined on the boundary)
so the assumption of surjectivity does not
hold.
A simple argument reveals the source of
this problem: it appears due to
the limited
compatibility of the Green's formula
required to hold
on all of~$\dom(A^*)$ and the required
surjectivity of both
boundary maps~$\Gamma_0$,
$\Gamma_1$ also defined on
the same domain~$\dom(A^*)$.
This limitation of the boundary triples
formalism can be relaxed and the framework
extended to cover more general cases,
albeit at the cost of increased complexity,
see~\cite{BehrndtLanger2007, BehrndtLanger2012},
the book~\cite{BHS} and the
references therein for a detailed account.
Formally a more restrictive approach
applicable to semibounded
symmetric operators~$A$ and
not based on
the description of~$\dom A^*$
was developed
by M.~Birman, M.~Kre\u{\i}n and
M.~Vishik.
Despite its limited scope, this
theory proves to be indispensable
in applications to various
problems of ordinary and
partial differential operators.
The publication~\cite{AlonsoSimon}
contains a concise
exposition of these results.
It was realized later that the
Birman-Kre\u{\i}n-Vishik
method is closely related to the
theory of linear systems with
boundary control and to the
original ideas of M.~Liv\v{s}ic
from the open systems theory,
see e.~g.~\cite{Ryzhov_systems, Ryzh_spec}
in this connection.
Let us give a brief account of
relevant results
derived from the works cited above
and tailored to the purposes
of current presentation.
\subsubsection{Boundary value problem}
Let~$H$, $E$ be two separable
Hilbert spaces, $A_0$ an unbounded closed
linear operator on $H$ with the dense domain~$\dom A_0$
and $\Pi : E \to H$ a bounded linear
operator defined everywhere in~$E$.
\begin{theorem}\label{thm:BVP1}
Assume the following:
\begin{itemize}
\item $A_0$ is self-adjoint and boundedly invertible;
\item There exists the left inverse~$\tilde
\Gamma_0$ of $\Pi$ so that $\tilde
\Gamma_0 \Pi \varphi = \varphi$ for all $\varphi \in E;$
\item The intersection of $\dom A_0$ and $\ran \Pi$
is trivial: $\dom A_0\cap \ran\Pi = \{0\}$.
\end{itemize}
Since~$\dom A_0$ and $\ran\Pi$ have trivial intersection,
the direct sum~$\dom A_0 \dotplus \ran\Pi$
form a dense linear set in~$H$ that
can be described as
$\{A_0^{-1} f + \Pi \varphi \mid f\in H, \varphi \in E\}$.
Define two linear operators~$A$ and $\Gamma_0$ with
the common
domain~$\dom A_0 \dotplus \ran\Pi$
as ``null extensions'' of $A_0$
and $\tilde \Gamma_0$ to the complementary
component of~$\dom A_0\dotplus \ran\Pi$
\begin{equation*}
A: A_0^{-1} f + \Pi \varphi \mapsto f, \quad
\Gamma_0 : A_0^{-1} f + \Pi\varphi \mapsto \varphi ,
\qquad f \in H, \varphi \in E
\end{equation*}
The spectral ``boundary value problem'' associated
with the pair~$\{A_0,\Pi\}$ satisfying these conditions
is the system of two linear
equations for the unknown
vector~$u \in \dom A := \dom A_0\dotplus \ran\Pi:$
\begin{equation}\label{eqn:BVP_system}
\left\{\,\,
\begin{aligned}
(&A - zI) u = f \\
&\Gamma_0 u = \varphi
\end{aligned}
\right . \qquad f \in H, \quad \varphi \in E,
\end{equation}
where $z \in \mathbb C$ is the spectral parameter.
Let $z\in \rho(A_0)$, $f\in H$, $\varphi \in E$.
Then the system~(\ref{eqn:BVP_system}) admits
the unique
solution~$u_z^{f,\varphi}$
given by the formula
\begin{equation*}
u_z^{f,\varphi} = (A_0 - zI)^{-1} f + (I - z A_0^{-1})^{-1} \Pi \varphi
\end{equation*}
If the expression on the right hand side
is null for some~$f\in H$, $\varphi \in E$,
then $f = 0$ and $\varphi = 0$.
\end{theorem}
Let~$\Lambda$ be a linear operator on $E$ with the
domain~$\dom \Lambda \subset E$ not necessarily dense in $E$.
Define the linear operator~$\Gamma_1$ on~$\dom\Gamma_1
:= \{A_0^{-1}f +\Pi\varphi \mid f \in H, \varphi \in \dom\Lambda \}$ as
the mapping
\[
\Gamma_1 : A_0^{-1}f +\Pi\varphi \mapsto \Pi^* f + \Lambda \varphi, \qquad
f\in H, \quad \varphi \in \dom\Lambda
\]
This definition implies~$\Lambda = \Gamma_1\Pi|_{\dom\Lambda}$.
Denote~$\mathscr D :=\dom\Gamma_1 =
\{A_0^{-1}f +\Pi\varphi \mid f \in H, \varphi \in \dom\Lambda \}$.
%
Obviously $\mathscr D \subset \dom A$.
%
The next theorem is a form of the Green's formula for the operator~$A$.
\begin{theorem}\label{thm:GreenBVP}
Assume that $\Lambda$ is
selfadjoint (and therefore densely defined) in $E$. Then
\[
(A u, v ) - (u, Av) = (\Gamma_1 u, \Gamma_0 v)_E -
(\Gamma_0 u, \Gamma_1 v)_E, \qquad u, v \in \mathscr D
\]
\end{theorem}
Notice the difference with the boundary
triples version~(\ref{GreenFormula}),
where the operator on the left hand side is the
adjoint of a symmetric operator.
In contrast,
Theorem~\ref{thm:GreenBVP} has no relation to
symmetric operators. The Green's formula is valid on a
set defined by the selfadjoint~$A_0$
and an arbitrarily chosen selfadjoint
operator~$\Lambda$.
\smallskip
Under the assumptions of Theorems~\ref{thm:BVP1}
and~\ref{thm:GreenBVP}
the operator-valued analytic function
\[
M(z) = \Lambda + z\Pi^*(I -zA_0^{-1})^{-1}\Pi, \quad z \in \rho(A_0)
\]
defined on~$\dom M(z) = \dom\Lambda$ is the
Weyl function (cf.~\cite{BHS}) of the boundary value
problem~(\ref{eqn:BVP_system})
in the sense of equality (cf. Definition~\ref{def:WeylFunction})
\[
M(z)\Gamma_0 u_z = \Gamma_1 u_z, \quad z \in \rho(A_0)
\]
where~$u_z= u_z^{f,\varphi}$ is the solution
to~(\ref{eqn:BVP_system}) with~$f = 0$
and~$\varphi \in \dom\Lambda$.
\smallskip
For the boundary value problem pertaining to the Laplace operator in a
bounded domain~$\Omega \subset \mathbb R^3$ with a
smooth boundary~$\partial \Omega$ the boundary maps~$\Gamma_0$,
$\Gamma_1$ are defined as
$\Gamma_0 : u \mapsto u|_{\partial\Omega}$,
$\Gamma_1 : u \mapsto -\partial u/\partial n|_{\partial\Omega}$.
Then~$A_0$ is the Dirichlet Laplacian in~$L^2(\Omega)$
and~$\Pi$ is the operator of harmonic continuation from
the boundary space~$E = L^2(\partial \Omega)$ to $\Omega$.
The conditions~$\dom(A_0) \cap \ran(\Pi) = \{0\}$ and
$\Gamma_0\Pi = I_E$ are satisfied by virtue of
the embedding theorems for Sobolev
classes.
In this setting~$M(\cdot)$ is known as the Dirichlet-to-Neumann
map, which is a pseudodifferential operator
defined on~$H^1(\partial \Omega)$.
A special role of the operator~$\Lambda = M(0)$ for the study
of boundary value
problems was pointed out by M.~Vishik in his work~\cite{MR0051404} and
sometimes~$\Lambda$ in the settings of elliptic partial differential
operators is referred to as the {\it Vishik operator}.
\subsubsection{Family of boundary value problems}
General boundary value problems for the operator~$A$
have the form (cf.~\cite{MR0051404})
\begin{equation}\label{eqn:BVP_system_family}
\left\{\,\,
\begin{aligned}
(&A - zI) u = f \\
&(\alpha\Gamma_0 + \beta\Gamma_1)u = \varphi
\end{aligned}
\right . \qquad f \in H, \quad \varphi \in E.
\end{equation}
Here $\alpha$, $\beta$ are linear operators on~$E$ such that
$\beta$ is bounded (and defined everywhere in $E$)
and $\alpha$ can be unbounded in which case $\dom\alpha \supset
\dom\Lambda = \mathscr D$.
Under certain verifiable conditions the
solutions to~(\ref{eqn:BVP_system_family}) exist and
are described by the following theorem.
\begin{theorem}[see~\cite{Ryzh_spec}]\label{thm:BVP_def}
Assume that the conditions of Theorems~\ref{thm:BVP1} and
\ref{thm:GreenBVP} are satisfied
and that the operator
sum~$\alpha + \beta \Lambda$ is correctly defined
on~$\dom\Lambda$ and closable in $E$.
Then~$\alpha + \beta M(z)$, $z \in \rho(A_0)$ is also
closable as an
additive perturbation of $\alpha + \beta \Lambda$
by the bounded operator
$M(z) - \Lambda$.
Denote by~$\mathscr B(z)$ the closure
of $\alpha + \beta M(z)$, $z\in
\rho(A_0)$ and
let $\mathscr B = \mathscr B(0)$.
\begin{itemize}
\item
Consider the Hilbert space~$\mathscr H_{\mathscr B}$
formed by the vectors~$\left\{u = A_0^{-1} f + \Pi\varphi \mid
f \in H, \varphi \in \dom\mathscr B\right\}$ and endowed
with the norm
\[
\|u\|_{\mathscr B} = \left( \|f\|^2 + \|\varphi\|^2 + \|\mathscr B \varphi
\|^2\right)^{1/2}
\]
The formal sum~$\alpha \Gamma_0 + \beta\Gamma_1$
is a bounded map from the Hilbert space~$\mathscr H_{\mathscr B}$ to $E$.
Note that
the summands in~$(\alpha \Gamma_0 + \beta\Gamma_1)u$,
$u\in \mathscr H_{\mathscr B}$ need not be defined individually.
\item
Assume that for some~$z \in \rho(A_0)$ the operator~$\mathscr B(z)$ has a bounded
inverse $[\mathscr B(z)]^{-1}$.
Then the problem~(\ref{eqn:BVP_system_family}) is uniquely solvable.
Under this condition there exists a closed operator~$A_{\alpha, \beta}$
with dense domain
\[
\dom A_{\alpha, \beta} = \left\{
u \in {\mathscr H}_{\mathscr B} \mid
(\alpha \Gamma_0 + \beta\Gamma_1)u = 0
\right\} = \Ker(\alpha \Gamma_0 + \beta\Gamma_1)
\]
and the resolvent (Kre\u{\i}n formula) holds:
\[
(A_{\alpha, \beta} - zI)^{-1} = (A_0 -zI)^{-1} -
(I -zA_0^{-1})^{-1}\Pi [\mathscr B(z)]^{-1} \beta \Pi^* (I -zA_0^{-1})^{-1}
\]
\item
Denote by~$A_{00}$ the restriction of~$A_0$ to the
set~$\Ker{\Gamma_1}$,
that is,
$A_{00} = \left. A\right|_{\Ker\Gamma_0\cap\Ker\Gamma_1}$.
Then $A_{00}$ is a symmetric operator with its domain not necessarily
dense in $H$ and
\[
A_{00} \subset A_{\alpha, \beta} \subset A
\]
\end{itemize}
\end{theorem}
\noindent
Notice that~$A_0$ is a self-adjoint extension of~$A_{00}$ contained in~$A$.
It is not difficult to recognize
the parallel with the von Neumann
theory of self-adjoint extensions of
symmetric operators.
The operator~$A_{00}$ is the ``minimal'' operator with
the ``maximal'' equal
to~$A_{00}^{*}$ (whenever the latter exists)
and all self-adjoint extensions~$A_{s.a.}$ of~$A_{00}$
satisfy~$A_{00} \subset A_{s.a.} \subset A_{00}^*$.
Within the framework of Theorem~\ref{thm:BVP_def}
the equivalent of~$A_{00}^*$ is the operator~$A$
of the boundary value problem~(\ref{eqn:BVP_system})
defined on the domain~$\dom(A)$.
The semiboundness condition for~$A_{00}$ is relaxed and
replaced by the bounded invertibility of~$A_0$, i.~e.,
the existence of a regular point of~$A_0$ on the real line.
\subsubsection{Functional model}
The results of previous sections hint at the possibility
of a functional model construction for the
family of operators~$A_{\alpha, \beta}$
with a suitably chosen pair~$(\alpha, \beta)$.
Having in mind the model
space~(\ref{ModelSpace}),
the selection of a ``close'' dissipative
operator is typically guided
by the properties of the problem at hand.
In the most general case when
parameters~$(\alpha, \beta)$ are
unspecified,
a reasonable approach seems to be
to construct a model suitable
for the widest possible range
of~$(\alpha, \beta)$.
In accordance with the work by
S.~Naboko~\cite{MR573902},
the action of the operator~$A_{\alpha, \beta}$
will then be
explicitly described in the functional model representation.
This program is realized in the recent paper~\cite{CherKisSilva2}.
The ``model'' dissipative operator~$L = A_{-iI, I}$
corresponds
to the boundary
condition~$(\Gamma_1 - i\Gamma_0)u = 0$
for $u \in \dom(L)$ in the
notation~(\ref{eqn:BVP_system_family}).
The characteristic function of~$L$ then coincides
with the Cayley transform of the Weyl
function,
\[
S(z) = (M(z) - {\rm i}I)(M(z) +{\rm i}I)^{-1} : E \to E,
\quad z \in \mathbb C_+.
\]
Under the mapping of the upper half plane to the unit disk,
the function~$S((z-{\rm i})/(z+{\rm i}))$ is the Sz.Nagy-Foia\c{s}
characteristic function of a contraction,
namely, the Cayley transform~$V_0$ of~$A_{00}$
extended to~$H \ominus \dom(V_0)$ by the null
operator, hence resulting in a partial isometry.
If~$A_{00}$ has no non-trivial
self-adjoint parts,
the dissipative operator~$L = A_{-{\rm i}I, I}$ is
also completely non-selfadjoint.
The standard assumptions of complete
non-selfadjoint\-ness and maximality of
$L$ are thus met.
The minimal self-adjoint dilation~$\mathscr A$ of~$L$ formally
coincides with the dilation obtained in
Section~\ref{sect:BoundaryTriplesModel}
for the case of boundary triples with
$A_+^* = A$, $B_+ = {\rm i}I_E$, $B_+^* = - {\rm i}I_E$ and
$\alpha = \sqrt{2} I_E$.
The description of operators~$A_{\alpha, \beta}$ in the
spectral representation of dilation~$\mathscr A$,
i.e. in the model space~(\ref{ModelSpace}),
cannot be easily obtained for arbitrary~$(\alpha,\beta)$.
However, under certain conditions imposed on the parameters~$(\alpha, \beta)$
the model construction becomes tractable. Namely, one assumes that the operator $\beta$ is boundedly invertible,
the operator~$B = \beta^{-1}\alpha$ is
bounded in~$E$ and that the operator-valued function~$M(z)$ is invertible
and $B M(z)^{-1}$ is compact at least for
some~$z$ in the upper and lower
half-plane of
$\mathbb C$ (and therefore for all $z \in \mathbb C \setminus \mathbb R$).
It follows, that the operator~$A_{\alpha, \beta}$ has at most discrete spectrum
in~$\mathbb C \setminus \mathbb R$ with possible
accumulation to the real line only.
Moreover, the resolvent set of~$A_{\alpha, \beta}$ coincides with the open set of
complex numbers~$z\in \mathbb C$ such that the closed operator~$\overline{B +
M(z)}$ has a bounded inverse, i.~e. $\rho(A_{\alpha, \beta}) = \{z \in \mathbb
C \mid 0 \in \rho(\overline{B + M(z)})\}$.
Finally, the
model representation of the resolvent~$(A_{\alpha,\beta} - zI)^{-1}$, $z \in
\rho(A_{\alpha,\beta})$ is explicitly computed in the model space~(\ref{ModelSpace}).
Once the latter are established, it is natural to expect that the absolutely continuous subspaces can be characterised for the operators of boundary value problems in the case of exterior domains and the scattering theory can then be constructed following the recipe of Naboko, as presented in Section 2.2. If this programme is pursued, this would yield a natural representation of the corresponding scattering matrix purely in terms of the $M-$operator defined above. A paper devoted to this subject is presently being prepared for publication.
\subsection{Generalised resolvents}
\label{gen_res1}
In the present Section, we briefly recall the notion of generalised resolvents (see \cite{AG} for details) of symmetric operators, which play a major role in the asymptotic analysis of highly inhomogeneous media as presented in the present paper. It turns out that generalised resolvents and their underlying self-adjoint operators in larger (dilated) spaces feature prominently in our approach; moreover, their setup turns out to be natural in the theory of time-dispersive and frequency-dispersive media. On the mathematical level, this area is closely interrelated with the theory of dilations and functional models of dissipative operators, the latter (at least, in the case of dissipative extensions of symmetric operators) being an important particular case of the former.
We start with an operator-function $R(z)$ in the Hilbert space $H$, analytic in $z\in \mathbb C_+$. Assuming that $\im R(z)\geq 0$ for $z\in \mathbb{C}_+$, and under the well-known asymptotic condition $$\limsup_{\tau\to+\infty} \tau \|R({\rm i}\tau)\|<+\infty,$$ one has due to the operator generalisation of Herglotz theorem by Neumark \cite{Naimark1943}:
$$
R(z)=\int_{-\infty}^{\infty} \frac{1}{t-z} dB(t),
$$
where $B(t)$ is a uniquely defined left-continuous operator-function such that $B(-\infty)=0$, $B(t_2)-B(t_1)\geq 0$ for $t_2>t_1$ and $B(+\infty)$ bounded. By the argument of \cite{Figotin_Schenker_2005}, it follows from the Neumark theorem \cite{Naimark1940,Naimark1943} (cf., e.g., \cite{MR1036844}) that there exists a bounded operator $X: \mathcal H \mapsto H$ with an auxiliary Hilbert space $\mathcal H$ and a self-adjoint operator $\mathcal A$ in $\mathcal H$ such that
$$
R(z)=X(\mathcal A -z)^{-1} X^*
$$
with
$$
X X^* =\text{\rm s-}\!\lim_{t\to +\infty} B(t)=\text{\rm s-}\limsup_{\tau\to +\infty} \tau \im R(i\tau).
$$
A particular case of this result, see \cite{Strauss}, holds when (a) for some $z_0\in \mathbb{C}_+$ there exists a subspace $\mathfrak L\subset H$ such that (i) for all non-real $z$ and all $f\in \mathfrak{L}$ one has
$$
R(z)f -R(z_0)f= (z-z_0)R(z)R(z_0)f,
$$
(ii) for any $z\in \mathbb{C}_+$ and any $g\in \mathfrak{L}^\bot$ one has
$$
\|R(z)g\|^2\leq \frac{1}{\im z} \im \langle R(z)g,g \rangle,
$$
(iii) for all $g\in \mathfrak{L}^\bot$ the function $R(z)g$ is regular in $\mathbb{C}_+$; (b) $\overline{R(z_0) \mathfrak{L}}= H$.
Under these assumptions, the function $R(z)$ is ascertained to be a generalised resolvent of a densely defined symmetric operator $A$ in $H$. Moreover, the deficiency index of $A$ in $\mathbb{C}_+$ is equal to $\dim \mathfrak{L}^\bot$. Precisely, this means that $\mathcal H \supset H$ and $X=P$, where $P$ is the orthogonal projection of $\mathcal H$ to $H$, i.e.,
\begin{equation}\label{eq:genres}
R(z)=P(\mathcal A -z)^{-1}|_H,\quad \im z\not=0,
\end{equation}
where $\mathcal A$ is a self-adjoint out-of-space extension of the symmetric operator $A$ (or, alternatively, a zero-range model with an internal structure, see Section 4 below). Moreover, under the minimality condition $\bigvee_{\im z\not=0} (\mathcal A-z)^{-1}H=\mathcal H$, it is defined uniquely up to a unitary transform which acts as unity on $H$, see \cite{Naimark1940}.
The latter representation takes precisely the same form as the dilation condition \eqref{eqn:DilationEq} in the case of maximal dissipative extensions of symmetric operators, with the generalised resolvent $R(z)$ replacing the resolvent of a dissipative operator. It is in fact shown that the property \eqref{eq:genres} generalises \eqref{eqn:DilationEq}.
Namely, it turns out \cite{Strauss,Strauss_ext} that
$$
R(z)=(A_{B(z)}-z)^{-1} \text{ for } z\in \mathbb{C}_+\cup \mathbb{C}_-.
$$
Here in the particular case of equal deficiency indices, which is of interest to us from the point of view of zero-range models with an internal structure, $A_{B(z)}$ is a $z-$dependant extension of $A$ such that there exists a boundary triple $(\mathscr K, \Gamma_0, \Gamma_1)$ defining this extension as follows:
$$
\dom A_{B(z)}=\{u\in \dom A^* | \Gamma_1 u= B(z)\Gamma_0 u\}
$$
with $B(z)$ being a $-R$ function (i.e., an analytic operator-function with a non-positive imaginary part in $\mathbb{C}_+$).
Because of $B^*(\bar z)=B(z)$, which is the standard extension of an $R-$function into $\mathbb{C}_-$ implied here, the extension $A_{B(z)}$ turns out to be dissipative for $z\in \mathbb{C}_-$ and anti-dissipative for $z\in \mathbb{C}_+$. We henceforth refer to $\mathcal A$ as the Neumark-Strauss dilation of the generalised resolvent $R(z)=(A_{B(z)}-z)^{-1}$. In a particular case of constant $B(z)=B$ such that $\im B\leq 0$, we have
$$
(A_{B}-z)^{-1}=P(\mathcal A -z)^{-1}|_H \text{ for } z\in \mathbb{C}_-
\text{ and }
(A_{B^*}-z)^{-1}=P(\mathcal A -z)^{-1}|_H \text{ for } z\in \mathbb{C}_+,
$$
which are precisely \eqref{eqn:DilationEq} for both $A_B$ and $A_{B^*}$ at the same time.
From what has been said above it follows that generalised resolvents appear when one conceals certain degrees of freedom in a conservative physical system, either for the sake of convenience or because these are not known. In particular, we refer the reader to the papers \cite{Figotin_Schenker_2005,Figotin_Schenker_2007b}, where systems with time dispersion are analysed, with prescribed ``memory'' term. It turns out that passing over to the frequency domain one ends up with a generalised resolvent. It then proves possible to explicitly restore a conservative Hamiltonian (the operator $\mathcal A$ in our notation) which yields precisely the postulated time dispersion. In a nutshell, the idea here is to work with an explicit and simple enough model of the part of the space pertaining to the ``hidden degrees of freedom'' instead of the unnecessarily complicated physical equations which govern them. Similar ideas have been utilised in \cite{Tip_1998,Tip_2006}. The same technique has found its applications in numerics, and in particular in the so-called theory of absorbing boundary conditions, see, e.g., \cite{Absorbing1,Absorbing2}.
The problem of constructing a spectral representation for a Neumark-Strauss dilation of a given generalised resolvent thus naturally arises. In a number of special cases, where $\mathcal H$ and $\mathcal A$ admit an explicit construction (and in particular one has $\mathcal H = H\oplus \mathbb{C}^k$ for $k\geq 1$), this can be done following essentially the same path as outlined in Section \ref{sec:ryzh-funct-model} above. This is due to the fact that in this case the operator $\mathcal A$ can be realised as a von Neumann extension of a symmetric operator in $\mathcal H$ with equal and finite deficiency indices. The corresponding construction in the case when $k=1$ is presented in \cite{CENS}. Surprisingly, this rather simple model already has a number of topical applications, see Sections 4 and 5 of the present paper, and also the papers \cite{Exner,KuchmentZeng,KuchmentZeng2004}, where a generalised resolvent of precisely the same class appears in the setting of thin networks converging to quantum graphs.
The generic case has been studied by Strauss in \cite{Strauss_survey}, where three spectral representations of the dilation are constructed, analogous to the ones of L.~de~Branges and J.~Rovnyak, B.~S.~Pavlov, and B.~Sz.-Nagy and C.~Foia\c{s}. These results however present but theoretical interest, as they are formulated in terms which apparently cannot be related to the original problem setup and are therefore not usable in applications.
A different approach was suggested by M.~D.~Faddeev and B.~S.~Pavlov in \cite{PavlovFaddeev}, where a problem originally studied by P.~D.~Lax and
R.~S.~Phillips in \cite{Lax1973} was considered. In \cite{PavlovFaddeev}, a five-component representation of the dilation was constructed, which further allowed to obtain the scattering matrix in an explicit form. It therefore comes as no surprise that, precisely as in the Lax-Phillips approach, the resonances are revealed to play a fundamental role in this analysis (cf. the analysis of the so-called Regge poles in the physics literature).
Later on, and again motivated in particular by applications to scattering, Neumark-Strauss dilations were constructed in some special cases by J.~Behrndt et al., see \cite{Behrndt2007,OQS_Malamud}.
We remark that all the above results, except \cite{CENS} and \cite{Strauss_survey}, have stopped short of attempting to construct a spectral form of the Neumark-Strauss dilation. Any generic construction leading to the latter and formulated in ``natural'' terms is presently unknown, to the best of our knowledge.
\subsection{Universality of the model construction}
The general form of the
functional model of an unbounded
closed
operator~\cite{Ryzhov_closed}
is a generalization
of the special cases, as developed
in the papers by B.~Pavlov
and S.~Naboko cited above.
This section aims to clarify the relationship
between the models pertaining to
different representations
of the characteristic function of a
non-selfadjoint operator.
As an illustration, we consider
two special cases of
operators of mathematical physics
described above and link them to
the general model
construction of~\cite{Ryzhov_closed}.
\subsubsection{Characteristic function of a linear operator~\cite{Strauss1960}}
Let~$L$ be a closed linear operator on a (separable)
Hilbert space~$H$ with the
domain~$\dom(L)$.
Consider the form~$\Psi_L(\cdot, \cdot)$ defined
on~$\dom(L)\times \dom(L)$:
\begin{equation}\label{Psi_L}
\Psi_L(f,g) = \frac{1}{\rm i}\left[ (Lf,g)_H - (f, Lg)_H\right],
\quad f,g \in \dom(L)
\end{equation}
\begin{definition}\label{defBoundarySpaceAndOperator}
The {\it boundary space} of~$L$ is a
linear
space~$E$ with a possibly indefinite
scalar product~$(\cdot, \cdot)_E$ such that
there exists a closed
linear operator~$\Gamma$ defined
on~$\dom(\Gamma) = \dom(L)$ and the
following identity holds
\begin{equation}\label{BoundaryOperator}
\Psi_L(f,g) = (\Gamma f, \Gamma g)_E,
\quad f,g \in\dom(L)
\end{equation}
The operator~$\Gamma$ is called the
{\it boundary operator} of~$L$.
\end{definition}
This definition is meaningful for any
linear operator on~$H$.
For the purposes of model construction,
it is sufficient to focus only on the case
when $L$ is densely defined and dissipative.
The model representation of a non-dissipative
operator is given in the model space of an
auxiliary dissipative one, as explained above.
When~$L$ is dissipative, one has
$\Psi_L(f,f) \geq 0$, $f\in\dom(L)$ and
therefore the space~$E$ can be chosen
as the Hilbert space obtained by
factorization and completion
of~$\{\Gamma f \mid f\in \dom(L)\}$
with respect to the norm~$\|\Gamma f\|_E$,
$f\in \dom(L)$.
Note that the boundary
operator~$\Gamma$ defined
in~(\ref{BoundaryOperator})
is not uniquely defined.
Due to the Hilbert structure of~$E$,
for any isometry~$\pi$ on $E$
the operator~$\pi\Gamma$
also satisfies the
condition~(\ref{BoundaryOperator}).
Moreover, if (\ref{BoundaryOperator})
holds for some
operator~$\Gamma^\prime$
and space~$E^\prime$,
then there exists an
isometry~$\pi : E^\prime \to E$
such that
$\Gamma = \pi \Gamma^\prime$.
Denote by~$E_*$ and $\Gamma_*$
the boundary space
and the boundary operator for
the dissipative operator~$-L^*$ endowed
with the Hilbert metric.
Assume that~$L$ is maximal, i.~e.,
$\mathbb C_- \subset \rho(L)$.
Then $L^*$ is also maximal
and~$\mathbb C_+ \subset \rho(L^*)$.
The following definition is valid for general
non-selfadjiont operators.
\begin{definition}
Let~$E$ and $\Gamma$
be the boundary space and the boundary
operator for a closed
densely defined operator~$L$. Let~$E_*$
and $\Gamma_*$ be the boundary space
and the boundary operator
for the operator~$-L^*$.
The characteristic function of the operator~$L$
is the analytic operator-valued
function~$S(z): E \to E_*$
defined by
\[
S(z) \Gamma f =
\Gamma_* (L^* - zI)^{-1}(L - zI)f,
\quad f \in \dom(L),\quad z \in \rho(L^*)
\]
If $L$ is dissipative, then the spaces~$E$ and $E_*$
are Hilbert spaces, and
the operator~$S(z)$ is a contraction on $E$ for each $z\in\mathbb C_+$.
\end{definition}
Note that the actual form of~$S(z)$
depends on the choice
of boundary spaces and boundary
operators.
If~$\Gamma^\prime : \dom(L) \to E^\prime$
and $\Gamma_*^\prime : \dom(L^*) \to E_*^\prime$
satisfy the condition~(\ref{BoundaryOperator})
and $S^\prime(z)$ is the
corresponding characteristic function, then
there exist two isometries~$\pi_* : E_* \to E_*^\prime$
and $\pi : E \to E^\prime$ such that
$\pi_* S(z) = S^\prime(z)\pi$, $z\in \rho(L^*)$.
Such characteristic functions of the operator~$L$
are often called {\it equivalent}.
All steps involved in the model
construction outlined above do not depend
on the concrete form
of the characteristic function~\cite{BMNW2018}.
In particular cases when the characteristic
function can be expressed in terms of
the original problem, the model
admits a ``natural'' form in relation to the
problem setup.
Examples of such calculations are
provided towards the end of this section.
In order to compute a
characteristic function of~$L$ one has to
come up with a suitable definition
of boundary spaces and operators.
Consider first the general case where
no specific assumptions on the
operator~$L$ are made, and
introduce the Cayley
transform of~$L$, i.e.,
$T = (L -{\rm i}I)(L+{\rm i}I)^{-1}$.
The operator~$T$ is clearly contractive.
The operators
\[
Q :=\frac{1}{\sqrt{2}}(I-T^*T)^{1/2},\qquad
Q_* :=\frac{1}{\sqrt{2}}(I-T T^*)^{1/2}
\]
are thus non-negative.
A straightforward calculation~\cite{Ryzhov_closed}
shows that
the boundary spaces~$E$, $E_*$ and
the boundary operators~$\Gamma$,
$\Gamma_*$ can be defined as follows:
\[
E = \clos\ran(Q), \quad E_* = \clos\ran(Q_*),
\quad \Gamma = \clos Q(L + {\rm i}I),
\quad \Gamma_* = \clos Q_*(L^* - {\rm i}I).
\]
Here the operators $\Gamma$ and $\Gamma_*$
are the closures of the respective mappings
initially defined on~$\dom(L)$ and $\dom(L^*)$.
This choice leads to the following expression
for the characteristic function of
the operator~$L:$
\begin{equation}\label{abstractCharFunc}
S(z) = \left.\bigl(T - (z-{\rm i})\Gamma_*(L^* - zI)^{-1}Q\bigr)\right|_E, \quad
z \in \mathbb C_+.
\end{equation}
An explicit calculation
reveals that~$S(z)$ is closely related to
the characteristic function
of~$T$,
\[
S(z) = - \vartheta_T\left(\frac{z -{\rm i}}{z + {\rm i}}\right), \quad z \in \mathbb C_+,
\]
where
\[
\vartheta_T(\lambda) =\left.
\bigl(- T + 2 \lambda Q_*(I - \lambda T^*)^{-1}Q\bigr)\right|_E,
\qquad |\lambda| < 1
\]
is the Sz.-Nagy-Foia\c{s}
characteristic function of the contractive operator~$T$.
Therefore, the formula~(\ref{abstractCharFunc})
is the abstract form of the characteristic function
of~$L$ regardless the ``concrete'' realization of
the operator~$L$.
\subsubsection{Examples}
The actual choice of boundary spaces
and operators is guided by the specifics
of the problem at hand.
Let us demonstrate the ``natural'' selection for these objects for some
of the models introduced in Section 2.
\paragraph{Additive perturbations}
This is the simplest (and canonical)
case of the characteristic
function calculations included here solely
for the completeness of exposition.
Let~$L$ be a dissipative operator of Section \ref{sec:add_perturbations},
defined as an additive perturbation
of a self-adjoint operator.
Then for~$f, g\in\dom(A) = \dom(L)$
one has
\[
\Psi_L(f,g) = \frac{1}{\rm i}\left[(Lf ,g) - (f, Lg)\right] =
\frac{1}{\rm i}\left[ \left({\rm i}\frac{\alpha^2}{2}f,g\right) -
\left(f,{\rm i}\frac{\alpha^2}{2}g\right)\right] =
(\alpha f, \alpha g)
\]
and therefore, the boundary
space can be chosen as~$E = \clos\ran(\alpha)$
with
the boundary operator~$\Gamma$
defined as a
mapping~$\Gamma f \mapsto \alpha f$. In a similar way, $E_*=E$ and $\Gamma_*=\Gamma.$
The characteristic function of~$L$ corresponding
to this selection of boundary spaces
and operators is then computed as \eqref{DissCharF}.
As explained above, this characteristic function is equivalent
to the function~(\ref{abstractCharFunc}).
\paragraph{Almost solvable extensions}
In the notation of Section~\ref{sec:ryzh-funct-model},
let~$L = A_B$ be a dissipative almost solvable
extension of a symmetric operator~$A$
corresponding to the bounded
operator~$B = B_R + i B_I$
with $B_R = B_R^*$ and $B_I = B_I^* \geq 0 $
defined on the space~$\mathcal K$.
Denote~$\alpha := \sqrt{2}(B_I)^{1/2}$.
From the Green's formula~(\ref{GreenFormula})
and the condition~$\Gamma_1 f= B \Gamma_0 f$, $f\in\dom(L)$
we obtain for~$f,g \in \dom(L)$:
\[
\Psi_L(f,g) = \frac{1}{\rm i}
\left[ (Lf,g) - (f,Lg) \right]= \frac{1}{\rm i}
\bigl((B - B^*)\Gamma_0 f, \Gamma_0 g\bigr)_\mathcal K =
(\alpha\Gamma_0 f, \alpha\Gamma_0 g)_{\mathcal K}.
\]
Next we demonstrate two alternative approaches to
the derivation of the characteristic function of~$L$.
\paragraph{Approach 1.}
Define the boundary space~$\mathcal E$ of
the operator~$L = A_B$ as
the factorization and completion of
the linear set~$\mathcal L = \{ \Gamma_0 f, \mid f \in \dom (L)\}$
endowed with the norm~$\|u\|_{\mathcal E} =
\|\alpha^2 u\|_\mathcal K$,
$u \in \mathcal L$.
The norm~$\|\cdot\|_{\mathcal E}$ is degenerate
if~$\ker(\alpha)$ is non-trivial, thus the factorization
becomes necessary.
The corresponding boundary operator~$\Gamma$ is the
map~$\Gamma : f \mapsto \Gamma_0 f$
on the domain~$\dom(\Gamma) = \dom(L)$.
In a similar way, $\mathscr E_*$ is defined
as the factorization and completion of the linear
set ~$\mathcal L_*= \{ \Gamma_0 g, \mid g \in \dom (L^*)\}$
with respect to the norm~$\|u\|_{\mathcal E_*} =
\|\alpha^2 u\|_\mathcal K$, $u \in \mathcal L_*$.
The boundary operator~$\Gamma_*$ is the
mapping~$\Gamma_* : g \mapsto \Gamma_0 g$
defined on~$\dom(\Gamma_*) = \dom(L^*)$.
Thus, both boundary spaces~$\mathcal E$
and $\mathcal E_*$ are Hilbert spaces
with the norm associated with the ``weight''
equal to~$\alpha^2$.
An explicit computation then yields the following expression for the characteristic function $\mathscr S$:
\[
\mathscr S(z) =\bigl(B^* - M(z)\bigr)^{-1}\bigl(B - M(z)\bigr),
\quad z \in\rho(L^*),
\]
where $M(z)$ is the Weyl-Titchmarsh $M-$function of Section~\ref{sec:ryzh-funct-model}.
\paragraph{Approach 2.}
An alternative form of the characteristic function
is obtained based on the
boundary operators~$\Gamma$ and $\Gamma_*$ introduced
as the closures of
the mappings~$f \mapsto \alpha\Gamma_0 f$
and $g\mapsto \alpha\Gamma_0 g$
defined on the linear sets~$ \dom(L)$ and
$\dom(L^*)$,
respectively.
The boundary spaces~$E$ and $E_*$
in this case are chosen
as
\[
E = \clos\ran\bigl(\alpha\Gamma_0|_{\dom(L)}\bigr),\qquad
E_* = \clos\ran\bigl(\alpha\Gamma_0|_{\dom(L^*)}\bigr).
\]
In all applications considered in this paper, these
spaces coincide:~$E= E_*$.
Similarly to the situation of additive perturbations,
it is often convenient (and common) to
extend these spaces to~$\clos\ran(\alpha)$.
The
corresponding characteristic function is then represented
by the formula~(\ref{eq:charf_ext})
repeated here for the sake of readers' convenience:
\[
S(z) = I_E + {\rm i}\alpha\bigl(B^* - M(z)\bigr)^{-1}\alpha : E \to E,
\quad z \in \rho(L^*).
\]
In contrast to Approach 1, this form captures the specifics
of the extension parameter~$B$.
In particular, the dimension of the space~$E$ equals
the dimension of the range of~$\alpha$.
If the operator~$B$ is a compact perturbation
of a self-adjoint, i.~e., $B_I = \alpha^2/2 \in \mathfrak S_\infty$, then
the characteristic function~$S(z)$ is an operator-valued
function of the form $I + \mathfrak{S_\infty}$ defined
on the
(unweighted) Hilbert space~$\clos\ran(\alpha)$.
\paragraph{Equivalence of $\mathscr S$ and $S$.}
The mapping~$\widehat\alpha: f \mapsto \alpha f$,
$f \in \mathcal E$
is an isometry from the ``weighted'' space~$\mathcal E$
to the space~$\mathcal K$.
The equality~$\widehat\alpha \mathscr S(z) = S(z)\widehat\alpha$,
$z\in\rho(L^*)$
expresses equivalence of the characteristic
functions~$\mathscr S$ and $S$ corresponding to
different choices of boundary spaces and operators.
Both $\mathscr S$ and $S$ functions are equivalent to the characteristic
function of~$L$ written in its abstract form~(\ref{abstractCharFunc}).
It is easy to see that the
boundary operators of Approach 1 and Approach 2
are also related by means of the isometric
mapping~$\widehat\alpha$.
In conclusion, we point out the recent
paper~\cite{BMNW2018}
where the
construction of the selfadjoint dilation and the
functional model
of a dissipative operator
is based entirely on the concept of
Strauss boundary
spaces and operators~(\ref{Psi_L}),
(\ref{BoundaryOperator}) with no reference to their
``concrete'' realizations.
\section{An application: inverse scattering problem for quantum graphs}
\label{sec:theor-appl-funct}
In the present section, we present an application of the theory introduced in Section \ref{sec:ryzh-funct-model}, and in particular of the explicit construction of wave operators and scattering matrices facilitated by the approach based on the functional model due to Sergey Naboko. We first obtain an explicit expression for the scattering matrix of a quantum graph which we take to be the Laplacian on a finite non-compact metric graph, subject to $\delta-$type coupling at graph vertices. Then we present an explicit constructive solution to the inverse scattering problem for this graph, i.e., explicit formulae for the coupling constants at graph vertices. The narrative of this Section is based upon the papers \cite{CherednichenkoKiselevSilva,CherKisSilva1}.
For simplicity of presentation we will only consider the case of a finite non-compact quantum graph, when the deficiency indices are finite. However, the same approach allows us to consider the general setting of infinite deficiency indices, which in the quantum graph setting leads to an infinite graph. In particular, one could consider the case of an infinite compact part of the graph.
In what follows, we denote by ${\mathbb G}={\mathbb G}({\mathcal E},\sigma)$ a
finite metric graph, {\it i.e.} a collection of a finite non-empty set
${\mathcal E}$ of compact or semi-infinite intervals
$e_j=[x_{2j-1},x_{2j}]$ (for semi-infinite intervals we set
$x_{2j}=+\infty$), $j=1,2,\ldots, n,$ which we refer to as
\emph{edges}, and of a partition $\sigma$ of the set of endpoints
${\mathcal V}:=\{x_k: 1\le k\le 2n,\ x_k<+\infty\}$ into $N$ equivalence classes $V_m,$
$m=1,2,\ldots,N,$ which we call \emph{vertices}:
${\mathcal V}=\bigcup^N_{m=1} V_m.$ The degree, or valence, ${\rm deg}(V_m)$ of the vertex $V_m$ is defined as the number of elements in $V_m,$ {\it i.e.} ${\rm card}(V_m).$
Further, we partition the set ${\mathcal V}$ into the two non-overlapping
sets of \emph{internal} ${\mathcal V}^{({\rm i})}$ and \emph{external}
${\mathcal V}^{({\rm e})}$ vertices, where a vertex $V$ is classed as internal
if it is incident to no non-compact edge and external otherwise. Similarly, we partition the set of edges
${\mathcal E}={\mathcal E}^{({\rm i})}\cup {\mathcal E}^{({\rm e})}$, into the collection of
compact (${\mathcal E}^{({\rm i})}$) and non-compact (${\mathcal E}^{({\rm e})}$)
edges. We assume for simplicity that the number of non-compact edges
incident to any graph vertex is not greater than one.
For a finite metric graph ${\mathbb G},$ we consider the Hilbert spaces
\[
L^2({\mathbb G}):=\bigoplus_{j=1}^n L^2(e_j),\qquad W^{2,2}({\mathbb G}):=\bigoplus_{j=1}^n W^{2,2}(e_j).
\]
Further, for a function $f\in W^{2,2}({\mathbb G})$, we
define
the normal derivative at each vertex along each of the adjacent edges,
as follows:
\begin{equation*}
\partial_n f(x_j):=\left\{ \begin{array}{ll} f'(x_j),&\mbox{ if } x_j \mbox{ is the left endpoint of the edge},\\[0.35em]
-f'(x_j),&\mbox{ if } x_j \mbox{ is the right endpoint of the
edge.}
\end{array}\right.
\end{equation*}
In the case of semi-infinite edges we only apply this
definition at the left endpoint of the edge.
\begin{definition}
For $f\in W^{2,2}({\mathbb G})$
and ${a_m}\in{\mathbb C}$ (below referred to as the ``coupling
constant"), the condition of continuity of the function $f$ through
the vertex $V_m$ ({\it i.e.} $f(x_j)=f(x_k)$ if $x_j,x_k\in V_m$)
together with the condition
\[
\sum_{x_j \in V_m} \partial _n f(x_j)={a_m} f(V_m)
\]
is called the $\delta$-type matching at the vertex $V_m$.
\end{definition}
\begin{remark}
Note that the $\delta$-type matching condition in a particular case
when ${a_m}=0$ reduces to the standard Kirchhoff matching
condition at the vertex $V_m$, see {\it e.g.} \cite{MR3013208}.
\end{remark}
\begin{definition}
\label{def6}
The quantum graph Laplacian $A_a,$ $a:=(a_1,...,a_N),$ on a graph ${\mathbb G}$ with
$\delta$-type
matching conditions is the operator of minus second derivative $-d^2/dx^2$
in the Hilbert space $L^2({\mathbb G})$ on the domain of functions that
belong to the Sobolev space $W^{2,2}({\mathbb G})$ and satisfy the
$\delta$-type matching conditions at every vertex $V_m$,
$m=1,2,\dots,N.$ The Schr\"odinger operator on the same graph is
defined likewise on the same domain in the case of summable edge
potentials ({\it cf.} \cite{MR3484377}).
\end{definition}
If all coupling constants ${a_m}$, $m=1,\dots, N$, are real, it is
shown that the operator $A_a$ is a proper self-adjoint
extension of a closed
symmetric operator $A$ in $L^2({\mathbb G})$
\cite{MR1459512}. Note that, without loss of generality,
each edge $e_j$ of the graph ${\mathbb G}$ can be considered to be an
interval $[0,l_j]$, where $l_j:=x_{2j}-x_{2j-1}$, $j=1,\dots, n$ is
the length of the corresponding edge. Throughout the present Section we
will therefore only consider this situation.
In \cite{MR3484377}, the following result is obtained for the case of
finite \emph{compact} metric graphs.
\begin{proposition}[\cite{MR3484377}]\label{Prop_M}
Let ${\mathbb G}$ be a finite compact metric graph with $\delta$-type
coupling at all vertices. There exists a closed densely defined
symmetric operator $A$ and a boundary triple such that the
operator $A_a$ is an almost solvable extension of $A$,
for which the parametrising matrix $\varkappa$
is given by
$\varkappa=\mathrm{diag}\{a_1,\dots,a_N\}$, whereas the Weyl
function is an $N\times N$ matrix with elements
\begin{equation}\label{Eq_Weyl_Func_Delta}
m_{jk}(z)=
\begin{cases}
-\sqrt{z}\biggl(\sum\limits_{e_p\in E_k}\cot\sqrt{z} l_p- 2\sum\limits_{e_p\in L_k}\tan\dfrac{\sqrt{z}
l_p}{2}\biggr),
& j=k,\\[0.4em]
\sqrt{z}\sum\limits_{e_p\in C_{jk}}\dfrac{1}{\sin\sqrt{z} l_p},
& j\not=k;\ V_j, V_k\ \mbox{adjacent},\\[0.5em]
0, & j\not=k;\ V_j, V_k\ \mbox{non-adjacent}.\\
\end{cases}
\end{equation}
Here the branch of the square root is chosen so that $\Im\sqrt{z}\geq 0,$
$l_p$ is the length of the edge $e_p$, $E_k$ is the set of non-loop graph edges
incident to the vertex $V_k$, $L_k$ is the set of loops at the
vertex $V_k,$ and $C_{jk}$ is the set of graph edges
connecting vertices $V_j$ and $V_k.$
\end{proposition}
In \cite{CherednichenkoKiselevSilva} this is extended to non-compact metric graphs as follows.
Denote by ${\mathbb G}^{({\rm i})}$ the compact part of the graph
${\mathbb G}$, {\it i.e.} the graph ${\mathbb G}$ with all the non-compact edges
removed. Proposition \ref{Prop_M} yields an expression for the Weyl
function $M^{({\rm i})}$ pertaining to the graph ${\mathbb G}^{({\rm i})}$.
\begin{lemma}
The matrix functions $M,$ $M^{({\rm i})}$ described above are related by the formula
\begin{equation}
M(z)=M^{({\rm i})}(z) + {\rm i}\sqrt{z} P_{\rm e},\quad\quad z\in{\mathbb C}_+,
\label{M_Mi}
\end{equation}
where $P_{\rm e}$ is the orthogonal projection in the boundary space
$\mathcal K$ onto the set of external vertices $V_{\mathbb G}^{({\rm e})}$, {\it
i.e.} the matrix $P_{\rm e}$ such that $(P_{\rm e})_{ij}=1$ if $i=j,$ $V_i\in V_{\mathbb G}^{({\rm e})},$ and
$(P_{\rm e})_{ij}=0$ otherwise.
\end{lemma}
The formula (\ref{M_Mi}) leads to
$M(s)-M^*(s)=2{\rm i}\sqrt{s}P_{\rm e}$ a.e. $s\in{\mathbb R}$, and the
expression (\ref{scat2}) for $\widehat{\Sigma}$ leads to the classical
scattering matrix $\widehat{\Sigma}_{\rm e}(k)$ of the pair of operators $A_0$ (which is the Laplacian on the graph ${\mathbb G}$ with standard Kirchhoff matching at all the vertices) and $A_\varkappa,$ where $\varkappa=\varkappa=\mathrm{diag}\{a_1,\dots,a_N\}:$
\begin{equation}
\widehat{\Sigma}_{\rm e}(s)=
P_{\rm e} (M(s)-\varkappa)^{-1}(M(s)^*-
\varkappa)(M(s)^*)^{-1}M(s) P_{\rm e},\ \ \ \ s\in{\mathbb R},
\label{sigma_hat}
\end{equation}
which acts as the operator of multiplication in the space
$L^2(P_{\rm e}{\mathcal K}; 4\sqrt{s}ds)$.
We remark that in the more common approach to the construction of scattering matrices, based on comparing the asymptotic expansions of solutions to spectral equations, see {\it e.g.} \cite{{Faddeev_additional}}, one obtains
$\widehat{\Sigma}_{\rm e}$ as the scattering matrix. Our approach yields an explicit factorisation of
$\widehat{\Sigma}_{\rm e}$
into expressions involving the matrices $M$ and $\varkappa$ only, sandwiched between two projections. (Recall that $M$ and $\varkappa$ contain the information about the geometry of the graph and the coupling constants, respectively.) From the same formula (\ref{sigma_hat}), it is obvious that without the factorisation the pieces of information pertaining to the geometry of the graph and the coupling constants at the vertices are present in the final answer in an entangled form.
We reiterate that the analysis above pertains not only to the cases when the coupling constants are real, leading to self-adjoint operators $A_a,$ but also to the case of non-selfadjoint extensions, {\it cf.} Theorem \ref{thm:existence-completeness-wave-operators}.
In what follows we often drop the argument $s\in{\mathbb R}$ of the
Weyl function $M$ and the scattering matrices $\widehat{\Sigma},$
$\widehat{\Sigma}_{\rm e}.$
It is easily seen that
a factorisation of $\widehat{\Sigma}_{\rm e}$ into a product of
$\varkappa-$dependent and $\varkappa-$independent factors ({\it cf.}
(\ref{scat2})) still holds in this case in $P_{\rm e} {\mathcal K},$ namely
\begin{equation}
\widehat{\Sigma}_{\rm e}=
\bigl[P_{\rm e} (M-\varkappa)^{-1}(M^*-\varkappa)P_{\rm e}\bigr]\bigl[P_{\rm e}(M^*)^{-1}M P_{\rm e}\bigr].
\label{bigstar}
\end{equation}
We will now exploit the above approach in the analysis of the inverse
scattering problem for Laplace operators on finite metric graphs,
whereby the scattering matrix $\widehat{\Sigma}_{\rm e}(s),$ defined by (\ref{bigstar}), is assumed
to be known for almost all positive ``energies'' $s\in{\mathbb R},$
along with the graph ${\mathbb G}$ itself. The data to be determined is the
set of coupling constants $\{{a_j}\}_{j=1}^N$. For simplicity, in what follows we treat the inverse problem for graphs with
real coupling constants, which corresponds to self-adjoint operators.
First, for given $M,$ $\widehat{\Sigma}_{\rm e}$ we reconstruct the expression $P_{\rm e}(M^{({\rm i})}-\varkappa)^{-1}P_{\rm e}$ for almost all $s>0:$
\begin{equation}
P_{\rm e}(M^{({\rm i})}-\varkappa)^{-1}P_{\rm e}
=\frac{1}{{\rm i}\sqrt{s}}\biggl(2\bigl(P_{\rm e}+\widehat{\Sigma}_{\rm e}[P_{\rm e}(M^*)^{-1}MP_{\rm e}]^{-1}\bigr)^{-1}-I\biggr)P_{\rm e}.
\label{DtD}
\end{equation}
In particular, due to the property of analytic continuation, the expression $P_{\rm e}(M^{({\rm i})}-\varkappa)^{-1}P_{\rm e}$ is determined uniquely in the whole of $\mathbb C$ with the
exception of a countable set of poles, which coincides with the
set of eigenvalues of the self-adjoint Laplacian $A_\varkappa^{({\rm i})}$ on the
compact part ${\mathbb G}^{({\rm i})}$ of the graph ${\mathbb G}$ with matching
conditions at the graph vertices given by the matrix $\varkappa,$ {\it cf.} Proposition \ref{Prop_M}.
\begin{definition}
\label{definition7}
Given a partition ${\mathcal V}_1\cup{\mathcal V}_2$ of the set of graph vertices, for $z\in{\mathbb C}$ consider the linear set $U(z)$ of functions that satisfy the differential equation $-u_z''=zu_z$ on each edge, subject to the conditions of continuity at all vertices of the graph and the $\delta$-type matching conditions at the vertices in the set ${\mathcal V}_2.$ For each function $f\in U(z),$ consider the vectors
\[
\Gamma_1^{{\mathcal V}_1}u_z:=\Bigl\{\sum_{x_j \in V_m} \partial _n f(x_j)\Bigr\}_{V_m\in{\mathcal V}_1},\quad\quad\Gamma_0^{{\mathcal V}_1}u_z:=\bigl\{f(V_m)\bigr\}_{V_m\in{\mathcal V}_1}.
\]
The {\it Robin-to-Dirichlet map} of the set ${\mathcal V}_1$ maps the vector $(\Gamma_1^{{\mathcal V}_1}-\varkappa^{{\mathcal V}_1}\Gamma_0^{{\mathcal V}_1})u_z$ to $\Gamma_0^{{\mathcal V}_1}u_z,$ where $\varkappa^{{\mathcal V}_1}:=\diag \{a_m:\ V_m\in{\mathcal V}_1\}$. (Note that the function $u_z\in U(z)$ is determined uniquely by $(\Gamma_1^{{\mathcal V}_1}-\varkappa^{{\mathcal V}_1}\Gamma_0^{{\mathcal V}_1})u_z$ for all $z\in\mathbb C$ except a countable set of real points accumulating to infinity).
\end{definition}
The above definition is a natural generalisation of the corresponding definitions of Dirichlet-to-Neumann and Neumann-to-Dirichlet maps pertaining to the graph boundary, considered
in {\it e.g.} \cite{MR3013208}, \cite{MR2600145}.
We argue that the matrix $P_{\rm e}(M^{({\rm i})}-\varkappa)^{-1}P_{\rm e}$ is the Robin-to-Dirichlet map
for the set ${\mathcal V}^{({\rm e})}.$ Indeed, assuming $\phi:=\Gamma_1 u_z-\varkappa \Gamma_0 u_z$ and $\phi=P_{\rm e} \phi,$ where the latter condition ensures the correct $\delta$-type matching on the set $\mathcal{V}^{({\rm i})},$ one has
$P_{\rm e}\phi=(M^{({\rm i})}-\varkappa)\Gamma_0 u_z$ and hence $\Gamma_0 u_z=(M^{({\rm i})}-\varkappa)^{-1}P_{\rm e} \phi$. Applying $P_{\rm e}$ to the last identity yields the claim, in accordance with Definition \ref{definition7}.
We thus have the following theorem.
\begin{theorem}
\label{thm9.1}
The Robin-to-Dirichlet map for the vertices ${\mathcal V}^{({\rm e})}$
is determined uniquely by the scattering matrix
$\widehat{\Sigma}_{\rm e}(s),$ $s\in{\mathbb R},$ via the formula (\ref{DtD}).
\end{theorem}
The following definition, required for the formulation of the next theorem, is a generalisation of the procedure of graph contraction, well studied in the algebraic graph theory, see {\it e.g.} \cite{Tutte}.
\begin{definition}[Contraction procedure for graphs and associated quantum graph Laplacians]
\label{def8}
For a given graph ${\mathbb G}$ vertices $V$ and $W$ connected by an edge $e$ are ``glued'' together to form a
new vertex $(VW)$
of the contracted graph $\widetilde{\mathbb G}$ while simultaneously the edge $e$ is removed, whereas the rest of the graph remains
unchanged. We do allow the situation of multiple edges, when $V$ and
$W$ are connected in ${\mathbb G}$ by more than one edge, in which case all such edges but the edge $e$ become loops of their
respective lengths attached to the vertex $(VW)$. The corresponding quantum graph Laplacian $A_a$ defined on ${\mathbb G}$ is contracted to the
quantum graph Laplacian $\widetilde{A}_{\widetilde{a}}$ by the application of the following rule pertaining to the coupling constants:
a coupling constant
at any unaffected vertex remains the same, whereas the coupling
constant at the new vertex $(VW)$ is set to be the sum of the coupling constants at $V$ and $W.$
Here it is always assumed that all quantum graph Laplacians are described by Definition \ref{def6}.
\end{definition}
\begin{theorem}
\label{NtD_NtD}
Suppose that the edge lengths of the graph ${\mathbb G}^{({\rm i})}$ are rationally independent.
The element\footnote{By renumbering if necessary, this does not lead to loss of generality.} $(1,1)$ of the Robin-to-Dirichlet map described above yields the element $(1,1)$ of the ``contracted'' graph
$\widetilde{\mathbb G}^{({\rm i})}$ obtained from the graph ${\mathbb G}^{({\rm i})}$ by removing a non-loop edge $e$
emanating from $V_1.$ The procedure of passing from the graph ${\mathbb G}^{({\rm i})}$ to the contracted graph
$\widetilde{\mathbb G}^{({\rm i})}$ is given in Definition \ref{def8}.
\end{theorem}
\begin{proof}
Due to the assumption
that the edge lengths of the graph ${\mathbb G}^{({\rm i})}$ are rationally
independent, the element (1,1), which we denote by $f_1,$ is expressed explicitly as a function of
$\sqrt{z}$ \emph{and} all the edge lengths $l_j,$ $j=1,2,\dots, n,$ in particular, of the length of the edge $e,$ which we assume to be $l_1$ without loss of generality. This is
an immediate consequence of the explicit form of the matrix
$M^{({\rm i})},$ see (\ref{Eq_Weyl_Func_Delta}). Again without loss of generality, we also assume that the edge $e$ connects the vertices $V_1$ and $V_2.$
Further, consider the expression $\lim_{l_1\to 0}
f_1(\sqrt{z}; l_1,\dots,l_n; a)$. On the one hand, this limit is known from the explicit expression for $f_1$ mentioned above. On the
other hand, $f_1$ is the ratio of the determinant ${\mathcal D}^{(1)}(\sqrt{z}; l_1, \dots, l_n; a)$ of the principal minor of the matrix $M^{({\rm i})}(z)-\varkappa$ obtained by removing its first row and and first column and the
determinant of $M^{({\rm i})}(z)-\varkappa$ itself:
\[
f_1(\sqrt{z}; l_1,\dots,l_n; a)=\frac{{\mathcal D}^{(1)}(\sqrt{z}; l_1, \dots, l_n; a)}{{\rm det}\bigl(M^{({\rm i})}(z)-\varkappa\bigr)}
\]
Next, we multiply by $-l_1$ both the numerator and denominator of this ratio, and pass to the limit in each of them separately:
\begin{equation}
\lim_{l_1\to0}f_1(\sqrt{z}; l_1,\dots,l_n; a)=\frac{\lim\limits_{l_1\to0}(-l_1){\mathcal D}^{(1)}(\sqrt{z}; l_1, \dots, l_n; a)}{\lim\limits_{l_1\to0}(-l_1){\rm det}\bigl(M^{({\rm i})}(z)-\varkappa\bigr)}
\label{bulky_ratio}
\end{equation}
The numerator of (\ref{bulky_ratio}) is easily computed as the
determinant ${\mathcal D}^{(2)}(z; l_1, \dots, l_n; a)$ of the minor of $M^{({\rm i})}(z)-\varkappa$ obtained by removing its first two rows and first two columns.
As for the denominator of (\ref{bulky_ratio}), we add to the second row of the matrix
$M^{({\rm i})}(z)-\varkappa$ its first row multiplied by $\cos(\sqrt{z}l_1),$ which leaves the determinant unchanged. This operation, due to the identity
\[
-\cot(\sqrt{z}l_1)\cos(\sqrt{z}l_1)+\frac{1}{\sin(\sqrt{z}l_1)}=\sin(\sqrt{z}l_1),
\]
cancels out the singularity of all matrix elements of the second row at the point $l_1=0.$ We introduce the factor $-l_1$ ({\it cf.} \ref{bulky_ratio}) into the first row and pass to the limit as $l_1\to 0.$ Clearly, all rows but the first are regular at $l_1=0$ and hence converge to their limits as $l_1\to0.$ Finally, we add to the second column of the limit its first column, which again does not affect the determinant, and note that the first row of the resulting matrix has
one non-zero element, namely the $(1,1)$ entry. This procedure reduces the denominator in (\ref{bulky_ratio}) to the determinant of a matrix of the size reduced by one. As in \cite{MR3430381}, it is checked that this determinant is nothing but ${\rm det}(\widetilde{M}^{({\rm i})}-\widetilde\varkappa)$, where $\widetilde M^{({\rm i})}$ and $\widetilde\varkappa$ are the Weyl matrix and the (diagonal) matrix of coupling constants pertaining to the contracted graph $\widetilde{\mathbb G}^{({\rm i})}.$ This immediately implies that the ratio obtained as a result of the above procedure coincides with the entry (1,1) of the matrix $(\widetilde M^{({\rm i})}-\widetilde\varkappa)^{-1},$ {\it i.e.}
\begin{equation*}
\lim_{l_1\to0}f_1(\sqrt{z}; l_1,\dots,l_n; a)=f_1^{(1)}(\sqrt{z}; l_2,\dots,l_n; \widetilde{a}),
\end{equation*}
where $f_1^{(1)}$ is the element (1,1) of the Robin-to-Dirichlet map of the contracted graph $\widetilde{\mathbb G}^{({\rm i})},$ and $\widetilde{a}$ is given by Definition \ref{def8}.
\end{proof}
The main result of this section is the theorem below, which is obtained as a corollary of Theorems \ref{thm9.1} and \ref{NtD_NtD}. We assume without loss of generality that $V_1\in{\mathcal V}^{{\rm (e)}}$ and denote by $f_1(\sqrt{z})$ the (1,1)-entry of the Robin-to-Dirichlet map for the set ${\mathcal V}^{{\rm (e)}}.$ We set the following notation. Fix a spanning tree ${\mathbb T}$ (see {\it e.g.} \cite{Tutte}) of the graph ${\mathbb G}^{({\rm i})}.$ We let the vertex $V_1$ to be the root of $\mathbb T$ and assume, again without loss of generality, that the number of edges in the path $\gamma_m$ connecting $V_m$ and the root is a non-decreasing function of $m.$
Denote by $N^{(m)}$ the number of vertices in the path $\gamma_m,$
and by $\bigl\{l^{(m)}_k\bigr\},$ $k=1,\dots, N^{(m)}-1,$ the associated sequence of lengths of the edges in $\gamma_m,$ ordered along the path from the root $V_1$ to
$V_m.$ Note that each of the lengths $l^{(m)}_k$ is clearly one of the edge lengths $l_j$ of the compact part of the original graph ${\mathbb G}.$
\begin{theorem}
\label{last_theorem}
Assume that the graph ${\mathbb G}$ is connected and the lengths of its compact edges are rationally independent. Given the scattering matrix $\widehat{\Sigma}_{\rm e}(s),$ $s\in{\mathbb R},$ the Robin-to-Dirichlet map for the set ${\mathcal V}^{{\rm (e)}}$ and the matrix of coupling constants $\varkappa$ are determined constructively in a unique way. Namely,
the following formulae hold for $l=1, 2,\dots, N$ and determine $a_m,$ $m=1,\dots, N:$
\[
\sum_{m: V_m\in\gamma_l}a_m
=\lim_{\tau\to+\infty}\Biggl\{-\tau\Bigl(\sum_{V_m\in\gamma_l}{\rm deg} (V_m)-2(N^{(l)}-1)\Bigr)-\frac{1}{f_1^{(l)}(i\tau)}\Biggr\},
\]
where
\begin{equation}
f_1^{(l)}(\sqrt{z}):=\lim_{l^{(l)}_{N^{(l)}-1}\to0}\dots\lim_{l^{(l)}_2\to0}\lim_{l^{(l)}_1\to0}f_1(\sqrt{z}),
\label{last}
\end{equation}
where in the case $l=1$ no limits are taken in (\ref{last}).
\end{theorem}
\section{Zero-range potentials with an internal structure}
\label{sec:models-potent-zero}
\subsection{Zero-range models}
In many models of mathematical physics, most notably in the analysis of Schrodinger operators, an explicit solution can be obtained in a very limited number of special cases (essentially, those that admit separation of variables and thus yield solutions in terms of special functions). This deficit of explicitly solvable models has led physicists, starting with E.~Fermi in 1934, to the idea to replace potentials with some boundary condition at a point of three-dimensional space, i.e., a zero-range potential.
The rigorous mathematical treatment of this idea was initiated in \cite{Berezin_Faddeev}. It was shown that the corresponding model Hamiltonians are in fact self-adjoint extensions of a Laplacian which has been restricted to the set of $W^{2,2}$ functions vanishing in a vicinity of a fixed point $x_0$ in $\mathbb{R}^3$. These ideas were further developed in a vast series of papers and books, culminating in the monograph \cite{kura}, which also contains an comprehensive list of references.
Physical applications of zero-range models have been treated in, e.g., \cite{demkov}. It has been conjectured that zero-range models provide a good approximation of realistic physical systems in at least a far-away zone, where the concrete shape of the potential might be discarded, making them especially useful in the analysis of scattering problems. Here we also mention the celebrated Kronig-Penney model where a periodic array of zero-range potentials is used to model the atoms in a crystal lattice.
Despite the obvious success of the idea explained above, it still carries a number of serious drawbacks. In particular, it can be successfully applied to model spherically symmetric scatterers only. If one attempts to model a scatterer of a more involved structure, i.e., possessing a richer spectrum, by a finite set of zero-range potentials, the complexity of the model grows rapidly, essentially eliminating the main selling point of the model, i.e., its explicit solvability.
In 1980s, based in part on earlier physics papers by Shirokov et. al. where the idea was presented in an implicit form,
B.~S.~Pavlov \cite{Pavlov_internal_structure} rigorously introduced a model of zero-range potential with an internal structure. This idea was further developed by Pavlov and his students, see e.g. \cite{Pavlov1988,AdamyanPavlov} and references therein. In the mentioned works of Pavlov, one starts by considering the operator $A_0$ being a Laplacian restricted to the set of $W^{2,2}$ functions vanishing in a vicinity of a fixed point in $\mathbb{R}^3$, precisely as in \cite{Berezin_Faddeev}. Then, instead of considering von Neumann self-adjoint extensions of the latter, one passes over to the consideration of the so-called \emph{out of space} extensions, i.e., extensions to self-adjoint operators in a larger Hilbert space. The theory of out-of-space extensions generalising that of J.~von~Neumann was constructed by M.~A.~Neumark in \cite{Neumark1, Naimark1940} in the case of densely defined symmetric operators and by M.~A.~Krasnoselskii \cite{Krasnoselskii} and A.~V.~Strauss \cite{Strauss_ext} in the case opposite, see also \cite{Strauss_survey} for the connections with the theory of functional models.
In fact, Pavlov, being quite possibly unaware of these theoretical developments, has reinvented this technique in the following way. Alongside the original Hilbert space $H=L^2(\mathbb{R}^3)$, consider an auxiliary \emph{internal} Hilbert space $E$ (which can be in many important cases considered to be finite-dimensional) and a self-adjoint operator $A$ with simple spectrum in it. Let $\phi$ be its generating vector and consider the restriction $A_\phi$ of $A$ (non-densely defined) to the space
$$
\dom A_\phi := \{(A-{\rm i})^{-1}\psi: \psi\in E, \langle \phi,\psi\rangle =0\}.
$$
This leads to the symmetric operator $\mathcal{A}_0$ on the Hilbert space $H\oplus E$, defined as $A_0\oplus A_\phi$ on the domain
$$
\dom \mathcal{A}_0:= \left\{ \binom{f}{v}: f\in \dom A_0, v\in \dom A_\phi \right\},
$$
where $A_0$ is the restricted Laplacian on $H$ introduced above. The operator $\mathcal{A}_0$ is then a symmetric non-densely defined operator with equal deficiency indices, and one can consider its self-adjoint extensions $\mathcal{A}$. Among them, we will single out those which non-trivially couple the spaces $H$ and $E$ by feeding the boundary data at $x_0$ of a function $f\in W^{2,2}(\mathbb{R}^3)$ to the ``part'' of operator acting in $E$. The latter then serves as the operator of the ``internal structure'', which can be chosen arbitrarily complex. We elect not to dwell on the precise way in which the extensions $\mathcal{A}$ are constructed as an explicit examples of operators of this class will be presented below in the present section.
\subsection{Connections with inhomogeneous media}
\label{connections_inhom}
Leaving the subject of zero-range models with an internal structure aside for a moment, let us briefly consider a number of physically motivated models giving rise to zero-range potentials in general. In particular, we will be interested in those models which lead to a distribution ``potential'' $\delta'$, where $\delta$ is the Dirac delta function. It is well-known, see, e.g., \cite{kura}, that the question of relating an operator of the form $-\Delta + \alpha\delta'$ to one of self-adjoint von Neumann extensions of a properly selected symmetric restriction $A_0$ of $-\Delta$ is far from being trivial, as $\delta'$, unlike $\delta$, is form-unbounded.
For the same reason, it is non-trivial to construct an explicit ``regularisation'' of a $\delta'-$perturbed Laplacian, i.e., a sequence of operators $A_\varepsilon$ being either potential perturbations of the Laplacian or, in general, any perturbations of the latter which converge in some sense (say, in the sense of resolvent convergence) to the Laplacian with a $\delta'$ perturbation. In particular, we point out among many others the paper \cite{Carreau} where $A_\varepsilon$ are chosen as first-order differential non-self-adjoint perturbations of the Laplacian of a special form and the paper \cite{Golovaty} where the perturbation is assumed to admit the form $\varepsilon^{-2} v(x/\varepsilon)$.
It turns out that additive $\varepsilon-$dependant perturbations are not the most straightforward choice for the task described, as zero-range perturbations (and more precisely, zero-range perturbations with an internal structure) appear naturally in the asymptotic analysis of inhomogeneous media. In particular, in the paper \cite{KCher} we studied the norm-resolvent asymptotics of differential operators $A_\varepsilon$ with periodic coefficients with high contrast, defined by their resolvents $(A_\varepsilon-z)^{-1}:f\mapsto u$ as follows:
\begin{equation}
-\bigl(a^\varepsilon\bigl(x\bigr)u'\bigr)'-zu=f,\ \ \ \ f\in L^2({\mathbb R}),\ \ \ \varepsilon>0,\ \ \ z\in{\mathbb C},
\label{orig_problem}
\end{equation}
where, for all $\varepsilon>0,$ the coefficient $a^\varepsilon$ is $1$-periodic and
\begin{equation*}
a^\varepsilon(y):=\left\{\begin{array}{lll}a \varepsilon^{-2},\ \ \ y\in[0, l_1),\\[0.3em] 1,\ \ \ y\in[l_1, l_1+l_2),\\[0.3em] a \varepsilon^{-2},\ \ \ y\in[l_1+l_2, 1),\end{array}\right.
\end{equation*}
with $a>0,$ and $0<l_1<l_1+l_2<1.$ Here in \eqref{orig_problem} the ``natural'' matching conditions are imposed at the points of discontinuity of the symbol $a^\varepsilon(x)$, i.e., the continuity of both the function itself and of its conormal derivative, so that the operators $A_\varepsilon$ can be thought as being defined by the form $\int a^\varepsilon(x)|u'(x)|^2dx$. We remark that the operators $A_\varepsilon$ are unitary equivalent to the operators of the double-porosity model of homogenisation theory in dimension one, see, e.g., \cite{HempelLienau_2000}.
The main result of the named paper can be reformulated as follows.
\begin{theorem}\label{kc}
The norm-resolvent limit of the sequence $A_\varepsilon$ is unitarily equivalent to the operator $A_{\rm hom}$
in $L^2({\mathbb R})$ given by the differential expression
$-l_2^{-2}d^2/{dx^2}$ on
\begin{multline*}
{\rm dom}({ A}_{\rm hom})\\[0.25em]
=\bigl\{U: \forall n\in\mathbb Z\ \ U\in W^{2,2}(n, n+1),\ \ U'\in C(\mathbb R),\ \forall n\in\mathbb Z\ \ U(n+0)-U(n-0)=
l_2^{-1}(l_1+l_3)U'(n)\bigr\},
\end{multline*}
where $l_3:=1-(l_1+l_2)$.
Moreover, for all $z$ in a compact set $K_\sigma$ such that the distance of the latter from the positive real line is not less than a fixed $\sigma>0$, this norm resolvent convergence is uniform, with the (uniform) error bound $O(\varepsilon^2)$.
\end{theorem}
By inspection, the operator $A_{\rm hom}$ defined above corresponds to the formal differential expression
$$
-l_2^{-2}\frac{d^2}{dx^2} + \frac{(l_1+l_3)} {l_2}\sum_{n\in \mathbb{Z}} \delta'(x-n),
$$
i.e., it is the operator of the Kronig-Penney dipole-type model on the real line. It is also quite clear that the periodicity of the model considered has nothing to do with the fact that the effective operator acquires the $\delta'$-type potential perturbation. Thus it leads to the understanding that strong inhomogeneities in the media in generic (i.e., not necessarily periodic) case naturally give rise to zero-range potentials of $\delta'$-type.
In order to relate this exposition to our subject of zero-range potentials with an internal structure, let us describe the main ingredient leading to the result formulated above. As usual in dealing with periodic problems, we apply the Gelfand transform
\begin{equation}
\hat{U}(y,\tau)=(2\pi)^{-d/2}\sum_{n\in{\mathbb Z}^d}U(y+n)\exp\bigl(-{\rm i}\tau\cdot(y+n)\big),\ \ \ y\in
[0,1],
\ \tau\in [-\pi,\pi),
\label{Gelfand_formula}
\end{equation}
to the operator family $A_\varepsilon$, which yields the operator family $A_\varepsilon^{(\tau)}$ corresponding to the differential expression
\[
-\biggl(\frac d{dx}+{\rm i}\tau\biggr){a^\varepsilon(x)}\biggl(\frac d{dx}+{\rm i}\tau\biggr)
\]
on the interval $[0,1]$ with periodic boundary conditions at the endpoints. Here $\tau\in[-\pi, \pi)$ is the quasimomentum. As above, the matching conditions at the points of discontinuity of the symbol $a^\varepsilon(x)$ are assumed to be natural.
The asymptotic analysis of the operator family $A_\varepsilon^{(\tau)}$, as shown in \cite{KCher,CEKN}, yields the following operator as its norm-resolvent asymptotics.
Let $H_\text{\rm hom}=H_{\rm soft}\oplus \mathbb C^1$. For all values $\tau\in[-\pi, \pi),$ consider a self-adjoint operator $\mathcal A_{\rm hom}^{(\tau)}$ on the space $H_\text{\rm hom},$ defined as follows. Let the domain $\dom \mathcal A_{\text{\rm hom}}^{(\tau)}$ be defined as
\begin{equation*}
\dom \mathcal A_{\text{\rm hom}}^{(\tau)}=\Bigl\{(u,\beta)^\top\in H_\text{\rm hom}:\ u\in W^{2,2}(0,l_2), u(0)=\overline{\xi^{(\tau)}} u(l_2)=\beta/\sqrt{l_1+l_3}\Bigr\}.
\end{equation*}
On $\dom \mathcal A_{\text{\rm hom}}^{(\tau)}$ the action of the operator is set by
$$
\mathcal A_{\text{\rm hom}}^{(\tau)}\binom{u}{\beta}=
\left(\begin{array}{c}\biggl(\dfrac{1}{\rm i}\dfrac{d}{dx}+\tau\biggr)^2\\[0.7em]
-\dfrac{1}{\sqrt{l_1+l_3}}
\bigl(\partial^{(\tau)} u\bigr|_0 - \overline{\xi^{(\tau)}}\partial^{(\tau)} u\bigr|_{l_2}\bigr)
\end{array}\right).
$$
Here
\[
\xi^{(\tau)}:=\exp\bigl({\rm i}(l_1+l_3) \tau\bigr),\qquad
\partial^{(\tau)} u:=\biggl(\frac d {dx}+{\rm i}\tau\biggr) u.
\]
\begin{theorem}\label{thm:ex0}
The resolvent $(A_\varepsilon^{(\tau)}-z)^{-1}$ admits the following estimate in the uniform operator norm topology:
$$
(A_\varepsilon^{(\tau)}-z)^{-1}-\Psi^* (\mathcal A_{\text{\rm hom}}^{(\tau)}-z)^{-1}\Psi=O(\varepsilon^2),
$$
where $\Psi$ is a partial isometry from $H=L^2(0,1)$ to $H_\text{\rm hom}$.
This estimate is uniform in $\tau\in[-\pi,\pi)$ and $z\in K_\sigma$.
\end{theorem}
It is clear now that the operator $\mathcal A_{\text{\rm hom}}^{(\tau)}$ is nothing but the simplest possible example of Pavlov's zero-range perturbations with an internal structure, corresponding to the case where the dimension of the internal space $E$ is equal to one. The definition of $\mathcal A_{\text{\rm hom}}^{(\tau)}$ implies that the support of the zero-range potential here is located at the point $x_0=0$, which is identified due to quasi-periodic (of Datta--Das Sarma type) boundary conditions with the point $x=l_2$.
Next it is shown (see \cite{KCher} for details) that under an explicit unitary transform the operator family $\mathcal A_{\text{\rm hom}}^{(\tau)}$ is unitary equivalent to the family $\mathcal A'_{\rm hom}(\tau')$ at the quasimomentum point $\tau'=\tau+\pi ({\rm mod\ } 2\pi)$. Here $\mathcal A'_{\rm hom}(\tau')$ acts in the space $L^2[0,l_2]$
and is defined by the same differential expression as $\mathcal A_{\rm hom}^{(\tau)},$ with the parameter $\tau$ replaced by
$\tau':$
\[
\biggl(\frac{1}{\rm i}\frac {d}{dx}+\tau'\biggr)^2,
\]
on the domain described by the conditions
\begin{eqnarray*}
u(0)+{\rm e}^{-{\rm i}(l_1+l_3)\tau'} u(l_2)=(l_1+l_3) {\partial^{(\tau')}} u\bigr|_0,
\\[0.4em]
{\partial^{(\tau')}} u\bigl|_0=-{\rm e}^{-{\rm i}(l_1+l_3)\tau'}{\partial^{(\tau')}} u\bigr|_{l_2}.
\end{eqnarray*}
An application of the inverse Gelfand transform then yields Theorem \ref{kc}. This shows that the operator $\mathcal A_{\text{\rm hom}}^{(\tau)}$ which is an operator of a zero-range model with the internal space $E$ of dimension one is in fact a differential operator with a $\delta'-$potential, up to a unitary transformation. In view of \cite{Shkalikov_1983,Kurasov_supersingular} it is plausible that by a similar argument it could be shown, that an operator with a $\delta^{(n)}-$potential can be realized as a zero-range model with $\dim E= n$, for any natural $n$.
It is interesting to note that an operator admitting the same form as $\mathcal A_{\text{\rm hom}}^{(\tau)}$ (with $\tau=0$) appears naturally
in the setting of
\cite{Exner, KuchmentZeng, KuchmentZeng2004},
who discuss the behaviour of the spectra of operator sequences associated with domains
``shrinking'' as $\varepsilon\to 0$ to a metric graph embedded into $\mathbb{R}^d$.
Here the rate of shrinking of the ``edge'' parts
is assumed to be related to the rate of shrinking of the ``vertex'' parts of the domain via
\begin{equation}
\frac{\text{vol}(V^\varepsilon_{\rm vertex})}{\text{vol}(V^\varepsilon_{\rm edge})}\to\alpha>0, \ \ \ \varepsilon\to0.
\label{shrinking_rate}
\end{equation}
It is shown in the above works
that the spectra of the corresponding Laplacian operators with Neumann boundary conditions
converge to the spectrum of a
quantum graph associated with a Laplacian on the metric graph obtained as the limit of the domain as $\varepsilon\to0.$
The ``weight'' $l_1+l_3$ in $\mathcal A_{\text{\rm hom}}^{(\tau)}$
plays the r\^{o}le of the constant $\alpha$ in (\ref{shrinking_rate}).
By a similar argument to the one presented above one can show, that in the case of domains shrinking to a graph under the ``resonant'' condition (\ref{shrinking_rate}) one obtains, under a suitable unitary transform, the matching condition of $\delta'$ type at the internal graph vertices, with the corresponding coupling constant equal to $\alpha$.
\subsection{A PDE model: BVPs with a large coupling}
\label{PDE_model_bounded}
\subsubsection{Problem setup}
In \cite{CherKisSilva3}, we studied a prototype large-coupling transmission problem, posed on a bounded domain $\Omega\subset{\mathbb R}^d,$ $d=2,3,$ see Fig.\,\ref{fig:kaplya}, containing a ``low-index" (equivalently, ``high propagation speed") inclusion $\Omega_-,$ located at a positive distance to the boundary $\partial\Omega.$ Mathematically, this is modelled by a ``weighted" Laplacian $-a_{\pm}\Delta$, where $a_+=1$ (the weight on the domain
$\Omega_+:=\Omega\setminus\overline{\Omega}_-$), and $a_-\equiv a$ (the weight on the domain $\Omega_-$) is assumed to be large, $a_-\gg1.$ This is supplemented by the Neumann boundary condition
$\partial u/\partial n=0$ on the outer boundary $\partial\Omega,$
where $n$ is the exterior normal to $\partial\Omega,$ and ``natural" continuity conditions on the ``interface" $\Gamma:=\partial\Omega_-.$ For each $a,$ we consider time-harmonic vibrations
of the physical domain
represented by $\Omega,$ described by the eigenvalue problem for an appropriate operator in $L^2(\Omega).$
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.45]{kaplya_smooth.pdf}
\end{center}
\caption{Domain with a ``stiff" inclusion.\label{fig:kaplya}}
\end{figure}
It is easily seen that eigenfunction sequences for these eigenvalue problems converge, as $a\to\infty,$ to either a constant or a function of the form
\[
v-\frac{1}{|\Omega|}\int_{\Omega_+}v,
\]
where
$v$ satisfies the spectral boundary-value problem (BVP)
\begin{equation}
\label{electrostatic}
-\Delta v=z\biggl(v-\frac{1}{|\Omega|}\int_{\Omega_+}v\biggr)\ \ \ {\rm in}\ \Omega_+,\ \quad v\vert_\Gamma=0,
\qquad \dfrac{\partial v}{\partial n}\biggr\vert_{\partial\Omega}=0.
\end{equation}
Here the spectral parameter $z$ represents
the ratio of the size of the original physical domain to the wavelength in its part
represented by $\Omega_+.$
The problem (\ref{electrostatic})
is isospectral to the so-called ``electrostatic problem'' discussed in \cite[Lemma 3.4]{Zhikov_2004}, see also \cite{AKKL} and references therein, namely the eigenvalue problem for the self-adjoint operator $Q$ defined by the quadratic form
\begin{equation*}
q(u, u)=\int_{\Omega_+}\nabla v\cdot\overline{\nabla v},\qquad u=v+c,\quad v\in H^1_{0,\Gamma}:=\bigl\{v\in H^1(\Omega_+),\ v\vert_\Gamma=0\bigr\},\quad c\in{\mathbb C}
\end{equation*}
on the Hilbert space $L^2(\Omega_+)\dotplus \mathbb{C}$, treated as a subspace of $L^2(\Omega)$.
Denote by $A_0^+$ the Laplacian $-\Delta$ on $\Omega_+,$ subject to the Dirichlet condition on $\Gamma$ and the Neumann boundary condition
on $\partial\Omega$ and let
$\lambda^+_j,$ $\phi^+_j,$ $j=1,2,\dots,$ be the eigenvalues and the corresponding orthonormal eigenfunctions, respectively, of $A_0^+.$
It is then easily shown, that the spectrum of the electrostatic problem is the union of two sets: a) the set of $z$ solving the equation
\begin{equation*}
z\Biggl[|\Omega|+z\sum_{j=1}^\infty(\lambda_j^+-z)^{-1}\biggl|\int_{\Omega_+}\phi^+_j\biggr|^2\Biggr]=0.
\end{equation*}
and b) the set of those eigenvalues $\lambda^+_j$
for which the corresponding eigenfunction $\phi^+_j$ has zero mean over $\Omega_+.$
\subsubsection{Norm-resolvent convergence to a zero-range model with an internal structure}
\label{large_coupling_lim}
Suppose that $\Omega$ is a bounded $C^{1,1}$ domain, and $\Gamma\subset\Omega$ is a closed $C^{1,1}$ curve, so that $\Gamma=\partial\Omega_-$ is the common boundary of domains $\Omega_+$ and $\Omega_-,$ where $\Omega_-$ is strictly contained in $\Omega,$ such that $\overline{\Omega}_+\cup\overline{\Omega}_-=\overline{\Omega},$ see Fig.\,\ref{fig:kaplya}.
For $a>0,$ $z\in{\mathbb C}$
we consider the ``transmission" eigenvalue problem ({\it cf.} \cite{Schechter})
\begin{equation}
\label{eq:transmissionBVP}
\begin{cases}
&-\Delta u_+=zu_+\ \ {\rm in\ } \Omega_+,\\[0.4em]
&-a\Delta u_-=zu_-\ \ {\rm in\ } \Omega_-,\\[0.4em]
&u_+=u_-,\quad \dfrac{\partial u_+}{\partial n_+}+a\dfrac{\partial u_-}{\partial n_-}=0\ \ {\rm on\ } \Gamma,\\[0.7em]
&
\dfrac{\partial u_+}{\partial n_+}=0\ \
{\rm \ on\ } \partial\Omega,
\end{cases}
\end{equation}
where $n_\pm$ denotes the exterior normal (defined a.e.) to the corresponding part of the boundary. The above problem is understood in the strong sense,
{i.e.} $u_\pm\in H^2(\Omega_\pm),$ the Laplacian differential expression $\Delta$ is the corresponding combination of second-order weak derivatives, and the boundary values of $u_\pm$ and their normal derivatives are understood in the sense of traces according to the embeddings of $H^2(\Omega_\pm)$ into $H^s(\Gamma),$ $H^s(\partial\Omega),$ where $s=3/2$ or $s=1/2.$
Denote by $A_a$ the operator of the above boundary value problem. Its precise definition is given on the basis of the boundary triples theory in the form of \cite{Ryzh_spec}.
Consider the space $H_{\rm eff}=L^2(\Omega_+)\oplus \mathbb C$ and the following linear subset of $L^2(\Omega):$
\begin{equation*}
\dom \mathcal A_{\rm eff}=\biggl\{
\binom{u_+}{\eta}\in H_{\rm eff}:\ u_+\in H^2(\Omega_+),\ \ u_+|_\Gamma=\frac{\eta}{\sqrt{|\Omega_-|}}{\mathbbm 1}_\Gamma, \ \ \dfrac{\partial u_+}{\partial n_+}\biggr\vert_{\partial\Omega}=0\biggr\},
\end{equation*}
where $u|_\Gamma$
is the trace of the function $u$ and ${\mathbbm 1}_\Gamma$ is the unity function on $\Gamma.$
On $\dom \mathcal A_{\rm eff}$ we set the action of the operator ${\mathcal A}_{\rm eff}$ by the formula
\vskip -0.3cm
\begin{equation}
\mathcal A_{\rm eff}\binom{u_+}{\eta}=
\left(\begin{array}{c}-\Delta u_+\\[0.6em]
\dfrac{1}{\sqrt{|\Omega_-|}}\mathop{{{\int}}}_{\Gamma}\dfrac{\partial u_+}{\partial n_+}
\end{array}\right).
\label{A_eff_form}
\end{equation}
\begin{theorem}
\label{mmm}
The operator $\mathcal A_{\rm eff}$ is the norm-resolvent limit of the operator family $A_a$. This convergence is uniform for $z\in K_\sigma$, with an error estimate by $O(a^{-1})$.
\end{theorem}
This theorem yields in particular the convergence (in the sense of Hausdorff) of the spectra of $A_a$ to that of $\mathcal A_{\rm eff}$. This convergence is uniform in $K_\sigma$, and its rate is estimated as $O(a^{-1})$. Moreover, it is shown that the spectrum of $\mathcal A_{\rm eff}$ coincides with the spectrum of the electrostatic problem (\ref{electrostatic}).
Note that the form of $\mathcal A_{\rm eff}$ is once again identical to that of a zero-range model with an internal structure in the case when the internal space $E$ is one-dimensional. The obvious difference is that here the effective model of the medium is no longer ``zero-range'' per se; rather it pertains to a singular perturbation supported by the boundary $\Gamma$. Therefore, the result described above allows one to extend the notion of internal structure to the case of distributional perturbations supported by a curve, see also \cite{Kurasov_surfaces} where this idea was first suggested, although unlike above no asymptotic regularisation procedure was considered. Moreover, well in line with the narrative of preceding sections, the internal structure appears owing exclusively to the strong inhomogeneity of the medium considered.
We remark that a ``classical'' zero-range perturbation with an internal structure can still be obtained by a rather simple modification of the problem considered. Namely, let $a$ be fixed, and let the \emph{volume} of the inclusion $\Omega_-$ now wane to zero as the new parameter $\varepsilon\to 0$. This represents a model that has been studied in detail, see, e.g., \cite{Cap} and references therein. In this modified setup, a virtually unchanged argument leads to the inclusion being asymptotically modelled by a zero-range potential with an internal structure. Moreover, the dimension of the internal space $E$ is again equal to one, provided that a uniform norm-resolvent convergence is sought for the spectral parameter belonging to the compact $K_\sigma$.
\subsubsection{Internal structure with higher dimensions of internal space $E$}
\label{internal_sec}
A natural question must therefore be posed: can strongly inhomogeneous media only give rise to simplest possible zero-range models with an internal structure, pertaining to the case of $\dim E=1$, or is it possible to obtain effective models with more involved internal structures? It turns out that the second mentioned possibility is realised, which we will demonstrate briefly using the material of the preceding section.
Recall that in all the results formulated above the uniform convergence was claimed under the additional assumption that the spectral parameter belongs to a fixed compact. If one drops this assumption, within the setup of the previous section one has the following statement.
\begin{theorem}
Up to a unitary equivalence, for and $k\in{\mathbb N}$ there exists a self-adjoint operator $\mathcal A_{\rm eff}$ of a zero-range model with an internal structure on the space $H_{\rm eff}:=L^2(\Omega_+)\oplus \mathbb{C}^k$ such that
\begin{equation}
(A_a-z)^{-1}\simeq \mathfrak{P}(\mathcal A_{\rm eff}-z)^{-1}\mathfrak{P} + O\bigl(\max\{a^{-1},|z|^{k+1}a^{-k}\}\bigr)
\label{double_star}
\end{equation}
in the uniform operator norm topology. Here $\mathfrak{P}$ is the orthogonal projection of $H_{\rm eff}$ onto $L^2(\Omega_+)\oplus \mathbb C$ (i.e., the space $H_{\rm eff}$ of the previous section).
\end{theorem}
Note that unlike the results pertaining to the situation of the spectral parameter contained in a compact, here the leading-order term of the asymptotic expansion of the resolvent of the original operator $A_a$ is not the resolvent of some self-adjoint operator (unless $k=1$), but rather a generalised resolvent. It is also obvious that the concrete choice of $k$ to be used in the last theorem depends on the concrete relationship between $z$ and $a$ and on the error estimate sought: the error estimate of the theorem becomes tighter at $k$ increases. In essence, this brings about the understanding that despite the fact that on the face of it the problem at hand is one-parametric, it must be treated as having two parameters, $z$ and $a$.
The operator $\mathcal A_{\rm eff}$ of the last Theorem admits an explicit description for any $k\in\mathbb{N}$, but this description is rather involved. In view of better readability of the paper, we only present its explicit form in the case $k=2$:
\begin{equation*}
\mathcal A_{\rm eff}\left(\begin{array}{c}{u_+}\\{\eta_1}\\\eta_2\end{array}\right):=
\left(\begin{array}{c}-\Delta u_+\\[0.2em]
\dfrac{1}{\kappa}\mathop{{{\int}}}_{\Gamma}\dfrac{\partial u_+}{\partial n_+ }+ a(B^2 D^{-1}\eta_1 + B\eta_2)\\[0.8em]
a({B}\eta_1+D \eta_2)
\end{array}\right).
\end{equation*}
Here $B,D$ and $\kappa$ are real parameters, which are explicitly computed.
It should be noted that similar results can be obtained in the homogenisation-related setup of the previous section, see also Section \ref{sec:appl-cont-mech}.
\medskip
We can therefore conclude that zero-range models with an internal structure appear naturally in the asymptotic analysis of highly inhomogeneous media. Moreover, in the generic case they appear as Neumark-Strauss dilations (see \cite{Naimark1940, Naimark1943, Strauss, Strauss_survey}) of main order terms in the asymptotic expansions of the resolvents of problems considered. The complexity of the internal structure can be arbitrarily high (i.e., the dimension of the internal space $E$ can be made as high as required), provided that the spectral parameter $z$ is allowed to grow with the parameter $a\to\infty$ (or $\varepsilon\to 0$). Further, owing to the remark made above that an operator with a $\delta^{(n)}$-potential could be realized as a zero-range model with $\dim E= n$ for any natural $n$, we expect models with strong inhomogeneities to admit the role of the tool of choice in the regularisation of singular and super-singular perturbations, beyond the form-bounded case and including the case of singular perturbations supported by a curve or a surface.
\subsection{The r\^{o}le of generalised resolvents}
\label{gen_res2}
We close this section with a brief exposition of how precisely the asymptotic results formulated above are obtained. The analysis starts with the family of resolvents, say (in the case of Section \ref{connections_inhom}) $(A_\varepsilon -z)^{-1}$, describing the inhomogeneous medium at hand. One then passes over to the generalised resolvent $R_\varepsilon(z):=P(A_\varepsilon-z)^{-1}P^*,$ where $P$ denotes the orthogonal projection onto the ``part'' of the medium which is obtained by removing the inhomogeneities. Note that the generalised resolvent thus defined is a solution operator of a BVP pertaining to homogeneous medium, albeit subject to non-local $z-$dependant boundary conditions. The problem considered therefore reduces to the asymptotic analysis of the operator $B_\varepsilon(z)$, parameterising these conditions. As such, it becomes a classical problem of perturbation theory.
Assuming now, for the sake of argument, that $R_\varepsilon(z)$ has a limit, as $\varepsilon\to0,$ in the uniform operator topology for $z$ in a domain $D\subset\mathbb C$, and, further,
that the resolvent $({A}_\varepsilon-z)^{-1}$ also admits such limit, one clearly has
\begin{equation}
P\bigl(A_{\text{eff}}-z\bigr)^{-1}P^*=R_0(z),\quad z\in D\subset \mathbb C,
\label{R_def}
\end{equation}
where $R_0$ and $A_{\text{eff}}$ are the limits introduced above. The idea of simplifying the required analysis by passing to the resolvent ``sandwiched'' by orthogonal projections onto a carefully chosen subspace is in fact the same as in \cite{MR0217440}, where the resulting sandwiched operator is shown to be the resolvent of a dissipative operator.
The function $R_0$ defined by (\ref{R_def}) is a generalised resolvent, whereas $A_{\text{eff}}$ is its out-of-space self-adjoint extension (or {\it Neumark-Strauss dilation} \cite{Strauss}). By a theorem of Neumark \cite{Naimark1940} (see Section \ref{gen_res1} of the present paper) this dilation is defined uniquely up to a unitary transformation of a special form, provided that the minimality condition
holds. The latter can be reformulated along the following lines: one has minimality, provided that there are no eigenmodes in the effective media modelled by the operator $A_\varepsilon,$ and therefore in the medium modelled by the operator $A_{\text{eff}}$ as well, such that they ``never enter'' the part of the medium without inhomogeneities. A quick glance at the setup of our models helps one immediately convince oneself that this must be true.
It then follows that the effective medium is completely determined, up to a unitary transformation, by $R_0(z).$ Once this is established, it is tempting to construct its Neumark-Strauss dilation and conjecture, that it is precisely this dilation that the original operator family converges to in the norm-resolvent sense (of course, up to a unitary transformation).
This conjecture in fact holds true, although it is impossible to prove it on the abstract level: taking into account no specifics of problems at hand, one can claim weak convergence at best. Still, the approach suggested seems to be very transparent in allowing to grasp the substance of the problem and to almost immediately ``guess'' correctly the operator modelling the effective medium.
\section{Applications to continuum mechanics and wave propagation}
\label{sec:appl-cont-mech}
Parameter-dependent problems for differential equations have traditionally attracted much interest within applied mathematics, by virtue of their potential for
replacing complicated formulations with more straightforward, and often explicitly solvable, ones. This drive has led to a plethora of asymptotic techniques, from perturbation theory to multi-scale analysis, covering a variety of applications to physics, engineering, and materials science. While this subject area can be viewed as ``classical", problems that require new ideas continue emerging, often motivated by novel wave phenomena. One of the recent application areas of this kind is provided by composites and structures involving components with highly contrasting material properties (stiffness, density, refractive index). Mathematically, such problems lead to boundary-value formulations for differential operators with parameter-dependent coefficients. For example, problems of this kind have arisen in the study of periodic composite media with ``high contrast" (or ``large coupling")
between the material properties of the components, see \cite{HempelLienau_2000, Zhikov2000, CherErKis}.
In what follows, we outline how the contrast parameter emerges as a result of dimensional analysis, using a scalar elliptic equation of second order with periodic coefficients as a prototype example.
\subsection{Scaling regimes for high-contrast setups}
We will consider the physical context of elastic waves propagating through a medium with whose elastic moduli vary periodically in a chosen plane (say $(x_1,x_2)$-plane) and are constant in the third, orthogonal, direction (say, the $x_3$ direction). For example, one could think of a periodic arrangement of parallel fibres of a homogeneous elastic material within a ``matrix" of another homogeneous elastic material. We will look at the ``polarised" anti-plane shear waves, which can be described completely by a scalar function representing the displacement of the medium in the $x_3$ direction. In the case of the fibre geometry mentioned above, the relevant elastic moduli ${G}$ then have the form
\[
{G}(y)=\left\{\begin{array}{c}{G}_0,\quad y\in Q_0,\\[0.2em] {G}_1,\quad y\in Q_1,\end{array}\right.\ \ =:\left\{\begin{array}{c}{G}_0\\[0.1em] {G}_1\end{array}\right\}(y).
\]
where $Q_0,$ $Q_1$ are the mutually complementary cross-sections of the fibre and matrix components, respectively, so that $\overline{Q}_0\cup\overline{Q}_1=[0,1]^2.$ the mass density of the described composite medium is assumed to be constant. (The constants $G_0,$ $G_1$ are the so-called shear moduli of the materials occupying $Q_0,$ $Q_1.$) This physical setup was considered in \cite{MMP, poulton}.
Denote by $d$ the period of the original ``physical" medium and consider time-harmonic wave motions, i.e. solutions of the wave equation that have the form
\begin{equation}
U(x, t)={\rm e}^{{\rm i}\omega t}u(x),\qquad x\in{\mathbb R}^2,\ \ t\ge0,
\label{harmonic}
\end{equation}
where $\omega$ is a fixed frequency. In the setting of time-harmonic waves, see (\ref{harmonic}), the function $u=u(x)$ satisfies the following equation, written in terms of the original physical units:
\begin{equation}
-\nabla_x\cdot\left\{\begin{array}{c}{G}_0\\[0.1em] {G}_1\end{array}\right\}(x/d)\nabla_xu=\rho\omega^2u,
\label{sigma_eq}
\end{equation}
Multiply both sides by ${G}_1^{-1}$ and denote $\delta:={G}_0/{G}_1.$ the parameter $\delta$ represents the ``inverse contrast", which will be assumed ``small" later, and corresponds to the value $a^{-1}$ of the ``large" parameter of Sections \ref{large_coupling_lim}, \ref{internal_sec}. The equation (\ref{sigma_eq}) takes the form
\[
-\nabla_x\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(x/d)\nabla_xu=\frac{\rho}{{G}_1}\omega^2u
\]
Note that
\begin{equation}
\omega=\frac{2\pi c_1}{\lambda_1}=\frac{2\pi c_0}{\lambda_0},
\label{omega_form}
\end{equation}
where $c_j,$ $\lambda_j$ are the wave speed and wavelength in the relevant media ($j=0,1$).
Introduce a non-dimensional spatial variable $\tilde{x}=2\pi x/\lambda_1:$
\[
-\frac{4\pi^2}{\lambda_1^2}\nabla_{\tilde{x}}\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}
\biggl(\frac{\tilde{x}}{2\pi d/\lambda_1}\biggr)\nabla_{\tilde{x}}u=\frac{\rho}{{G}_1}\omega^2u,
\]
equivalently, with $\varepsilon:=2\pi d/\lambda_1:$
\begin{equation*}
-\nabla_{\tilde{x}}\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(\tilde{x}/\varepsilon)\nabla_{\tilde{x}}u=\frac{\rho}{{G}_1}c_1^2u,
\end{equation*}
where we have used (\ref{omega_form}). Note that $c_1\sqrt{\rho/{G}_1}=1$
and relabel $\tilde{x}$ by $x:$
\[
-\nabla_x\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(x/\varepsilon)\nabla_xu=u,
\]
Let us ``scale to the period one" i.e. consider the change of variable $y=\tilde{x}/\varepsilon=x/d:$
\begin{equation*}
-\varepsilon^{-2}\nabla_y\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(y)\nabla_yu=u,
\end{equation*}
or
\begin{equation*}
-\nabla_y\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(y)\nabla_yu=\biggl(\frac{2\pi d}{\lambda_1}\biggr)^2u.
\end{equation*}
The different scaling regimes, ranging from what we know as ``finite frequency, high contrast" to ``high frequency, high contrast", are described by setting
\begin{equation}
\varepsilon^2=\delta^\nu\widetilde{z},
\label{epsdelz}
\end{equation}
where $\widetilde{z}$ is obviously dimensionless is assumed to vary over the compact $K_\sigma,$ and $0\le\nu\le 1.$ Note that $\widetilde{z}$ can be alternatively expressed as
\begin{equation}
\widetilde{z}=\delta^{-\nu}\rho{G}_1^{-1}(d\omega)^2.
\label{zform}
\end{equation}
In particular, the setup analysed in the paper \cite{CherErKis} corresponds to the case $\nu=1:$
\begin{equation}
-\delta^{-1}\nabla_y\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(y)\nabla_yu=zu.
\label{ff_PDE}
\end{equation}
In terms of the original spatial variable $x$ the equation (\ref{ff_PDE}) takes the form
\[
-d^2\delta^{-1}\nabla_{x}\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(x)\nabla_xu=zu,
\]
or
\[
-\nabla_{x}\cdot\left\{\begin{array}{c}\delta\\ 1\end{array}\right\}(x)\nabla_xu=k^2u,
\]
where the wavenumber (i.e. ``spatial frequency") is given by
$k:=d^{-1}\sqrt{\delta z},$
so that
\begin{equation*}
(kd)^2=\frac{z}{n^2},
\end{equation*}
where $n^2=\delta^{-1}=G_0/G_1$ is the shear modulus of the material occupying $Q_0$ relative to the material occupying $Q_1.$
The setup (\ref{ff_PDE}) is the ``periodic" version of the formulation discussed in Section \ref{PDE_model_bounded}, see also Section \ref{connections_inhom} for the one-dimensional version of a high-contrast homogenisation problem that gives rise to the same formulation. Similarly, choosing the values $\nu=2/(k+1),$ $k=2,3,\dots,$ in (\ref{epsdelz}) gives rise to ``high-frequency large-coupling" formulations, which in turn lead, in the limit as $\delta\to0,$ to effective operators with an ``internal space" of dimension $k,$ see Section \ref{internal_sec}. The parameter $\widetilde{z}$ is the related to the spectral parameter $z$ in (\ref{double_star}) via
\[
z=\widetilde{z}\delta^{1-\nu}=\widetilde{z}\delta^{\frac{k-1}{k+1}}=\widetilde{z}a^{\frac{k+1}{k-1}},\qquad k=2,3,\dots,
\]
so that the error estimate in (\ref{double_star}) is optimal for $\widetilde{z}\in K_\sigma,$ in the sense that for such $\widetilde{z}$ it yields an error of order $a^{-1}\sim |z|^{k+1}a^{-k}$ for large $a.$
\subsection{Homogenisation of composite media with resonant components}
\subsubsection{Physical motivation}
The mathematical theory of homogenisation (see
{\it e.g.} \cite{Lions, Bakhvalov_Panasenko, Jikov_book}) aims at characterising limiting, or ``effective'', properties of small-period composites. Following an appropriate non-dimensionalisation procedure, a typical problem here is to study the asymptotic behaviour of solutions to equations of the type
\begin{equation}
-{\rm div}\bigl(A^\varepsilon(\cdot/\varepsilon)\nabla u_\varepsilon\bigr)-\widetilde{\omega}^2u_\varepsilon=f,\ \ \ \ f\in L^2({\mathbb R}^d),\quad d\ge2,\qquad \widetilde{\omega}^2\notin{\mathbb R}_+,
\label{eq:generic_hom}
\end{equation}
where for all $\varepsilon>0$ the matrix $A^\varepsilon$ is $Q$-periodic, $Q:=[0,1)^d,$ non-negative, bounded, and symmetric. The parameter $\widetilde{\omega}$ here represents a ``non-dimensional frequency": $\widetilde{\omega}^2=\widetilde{z},$ where $\widetilde{z}$ is the spectral parameter introduced in (\ref{epsdelz}), so for example for $\nu=1$ one can set $\widetilde{\omega}=d\sqrt{\rho/G_0}\,\omega,$ see (\ref{zform}).
One proves (see \cite{Zhikov_1989, BirmanSuslina} and references therein) that when $A$ is uniformly elliptic, there exists a constant matrix $A^\text{\rm hom}$ such that solutions $u_\varepsilon$ to \eqref{eq:generic_hom} converge to $u_\text{\rm hom}$ satisfying
\begin{equation}
-{\rm div}\bigl(A^\text{\rm hom} \nabla u_\text{\rm hom}\bigr)-\widetilde{\omega}^2u_\text{\rm hom}=f.
\label{eq:generic_Ahom}
\end{equation}
In what follows we write $\omega,$ $z$ in place of $\widetilde{\omega},$
$\widetilde{z},$ implying that either the dimensional or non-dimensional version of the equation is chosen.
In recent years, the subject of modelling and engineering a class of composite media with ``unusual" wave properties (such as negative refraction) has been brought to the forefront of materials science. Such media are generically referred to as metamaterials, see {\it e.g.} \cite{Phys_book}. In the context of
homogenisation, the result sought ({\it i.e.}, the ``metamaterial" behaviour in the limit of vanishing $\varepsilon$) belongs to the domain of the so-called time-dispersive media (see, {\it e.g.}, \cite{Tip_1998,Figotin_Schenker_2005,Tip_2006,Figotin_Schenker_2007b}). For such media, in the frequency domain one faces equations of the form
\begin{equation}
-{\rm div}\bigl(A\nabla u\bigr)+\mathfrak B(\omega^2)u=f,\ \ \ \ f\in L^2({\mathbb R}^d),\
\label{eq:generic_td}
\end{equation}
where $A$ is a constant matrix and $\mathfrak B(\omega^2)$ is a frequency-dependent operator in $L^2({\mathbb R}^d)$ taking the place of $-\omega^2$ in (\ref{eq:generic_hom}), if, for the sake of argument, in the time domain we started with an equation of second order in time.
If, in addition, the matrix function $\mathfrak B$ is scalar, {\it i.e.}, $\mathfrak B(\omega^2)=\beta(\omega^2)I$ with a scalar function $\beta$, the problem of the type
\begin{equation}
-{\rm div}\bigl(A^{\rm hom}(\omega^2)\nabla u\bigr)=\omega^2 u
\label{eq:generic_spectral_td}
\end{equation}
appears in place of the spectral problem after a formal division by $-\beta(\omega^2)/\omega^2$, with frequency-dependent (but independent of the spatial variable) matrix $A^{\rm hom}(\omega^2).$
In the equation (\ref{eq:generic_spectral_td}), in contrast to (\ref{eq:generic_Ahom}), the matrix elements of $A^{\rm hom}$, interpreted as material parameters of the medium, acquire a non-trivial dependence on the frequency, which may lead to their taking negative values in some frequency intervals. The possibility of electromagnetic media exhibiting negative refraction was envisaged in an early work \cite{Veselago}, who showed theoretically that the material properties of such media must be frequency-dependent, and the last two decades have seen a steady advance towards realising such media experimentally.
One may hope that upon relaxing the condition of uniform ellipticity on $A^\varepsilon$ one may be able to achieve a metamaterial-type response to wave propagation for sufficiently small values of $\varepsilon.$ It is therefore important to understand how inhomogeneity in the spatial variable in \eqref{eq:generic_hom} can lead, in the limit $\varepsilon\to0,$ to frequency dispersion as in (\ref{eq:generic_td}).
\subsubsection{Operator-theoretic motivation}
Already in the setting of finite-dimensional matrix algebra equations of the
form (see (\ref{eq:generic_td}))
\begin{equation}
Au+{\mathfrak B}(z)u=f,\qquad u\in{\mathbb C}^d,
\label{zdep_form}
\end{equation}
where $A=A^*\in{\mathbb C}^{d\times d},$
$f\in{\mathbb C}^d,$ ${\mathfrak B}$ is a Herglotz function with values in ${\mathbb C}^{d\times d},$ emerge when one seeks solutions to the standard resolvent equation for a block matrix:
\begin{equation}
\left(\begin{array}{cc}A&B\\[0.2em]B^*&C\end{array}\right)\left(\begin{array}{c}u\\[0.2em]v\end{array}\right)-z\left(\begin{array}{c}u\\[0.2em]v\end{array}\right)=\left(\begin{array}{c}f\\[0.2em]0\end{array}\right),\qquad\left(\begin{array}{c}u\\[0.2em]v\end{array}\right)\in{\mathbb C}^{d+k}.
\label{block_matrix}
\end{equation}
where $B\in{\mathbb C}^{k\times d},$ $C=C^*\in{\mathbb C}^{k\times k}.$
Indeed, it is the result of a straightforward calculation that (\ref{block_matrix}) implies
\[
Au-(B(C-z)^{-1}B^*+zI)u=f,
\]
whenever $-z$ is not an eigenvalues of $C,$ so (\ref{block_matrix}) implies (\ref{zdep_form}) with ${\mathfrak B}(z)=-B(C-z)^{-1}B^*-zI.$ Another consequence of the above calculation is that for any vector $(f, g)^\top\in {\mathbb C}^{d+k}$ one has
\[
u=P\biggl\{\left(\begin{array}{cc}A&B\\[0.2em]B^*&C\end{array}\right)+zI\biggr\}^{-1}P*\left(\begin{array}{c}f\\[0.2em]g\end{array}\right),
\]
where $P$ is the orthogonal projection of ${\mathbb C}^{d+k}$ onto ${\mathbb C}^k$ and $P^*$ is interpreted as a restriction to the $k$-dimensional subspace of vectors of the form $(f,0)^\top,$ $f\in{\mathbb C}^k.$
The above argument, in the more general setting of block operator matrices in a Hilbert space, likely appeared for the first time in \cite{Figotin_Schenker_2005}. ``Generalised resolvents", i.e. objects of the form
\begin{equation}
P({\mathcal A}-z)^{-1}P^*,
\label{AP_form}
\end{equation}
where ${\mathcal A}$ is an operator in a Hilbert space ${\mathcal H}$ and $P$ is an orthogonal projection of $\mathcal H$ onto its subspace $H$ have already been discussed in the present survey, see Sections \ref{gen_res1}, \ref{gen_res2}. As discussed in Section \ref{gen_res1}, abstract results of Neumark and Strauss \cite{Naimark1940, Strauss} establish that solution operators for formulations (\ref{zdep_form}), where $A$ is a self-adjoint operator in a Hilbert space $H$ can be written in the form (\ref{AP_form}) for a suitable ``out-of-space" extension ${\mathcal A}.$ Therefore, a natural question is whether formulations (\ref{eq:generic_td}) can be viewed as generalised resolvents obtained by an asymptotic analysis of some parameter-dependent operator family describing a heterogeneous medium. One piece of evidence pointing at the validity of such a conjecture is the result of Section \ref{large_coupling_lim}, where the role of the operator ${\mathcal A}$ in (\ref{AP_form}) is played by ${\mathcal A}_{\rm eff},$ see (\ref{A_eff_form}).
In \cite{KCher, CEK, CEKN}
a model of a high-contrast graph periodic along a single direction was considered. A unified treatment of critical-contrast homogenisation was proposed and carried out in three distinct cases: (i) where neither the soft nor the stiff component of the medium is connected; (ii) where the stiff component of the medium is connected; (iii) where the soft component of the medium is connected. The analytical toolbox presented in these works was then amplified to the PDE setting in \cite{CherErKis}. In the wider context of operator theory and its applications, this provides a route towards:
constructing explicit spectral representations and functional models for both homogenisation limits of critical-contrast composites and the related time-dispersive models, as well as solving the related direct and inverse scattering problems.
\subsubsection{Prototype problem setups in the PDE context}
\label{setup_section}
Consider the problem \eqref{eq:generic_hom} under the following assumptions:
\begin{equation*}
A^\varepsilon(y)=
\begin{cases}
aI,& y\in Q_{\text{stiff}},\\[0.25em]
\varepsilon^2I, & y\in Q_{\text{soft}},
\end{cases}
\end{equation*}
where $Q_{\text{soft}}$ ($Q_{\text{stiff}}$) is the soft (respectively, stiff) component of the unit cube $Q=[0,1)^d\subset\mathbb R^d$, so that $\overline{Q}=\overline{Q}_{\text{soft}}\cup \overline{Q}_{\text{stiff}},$ and $a>0.$
Two distinct setups were studied in \cite{CherErKis}.
For one of them (``Model I"), which is unitary equivalent to the model of \cite{HempelLienau_2000, Friedlander}, the component $Q_{\text{soft}}\subset Q$ is simply connected and
and its distance to $\partial Q$ is positive, {\it cf.} \cite{Zhikov2000, CherCoop}. For the other one (``Model II") the component $Q_{\text{stiff}}$ has the described properties.
It is assumed that the Dirichlet-to-Neumann maps for $Q_{\rm soft}$ and $Q_{\rm stiff},$ which map the boundary traces of harmonic functions in $Q_{\rm soft}$ and $Q_{\rm stiff}$ to their boundary normal derivatives, are well-defined as pseudo-differential operators of order one in the $L^2$ space on the boundary \cite{Hoermander, Friedlander_old, Arendt, AGW}.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.3]{media.pdf}
\end{center}
\caption{{\scshape Model setups.} {\small Model I: soft component $Q_{\rm soft}$ in blue, stiff component $Q_{\rm stiff}$ in green. Model II: soft component $Q_{\rm soft}$ in green, stiff component $Q_{\rm stiff}$ in blue.}\label{fig:media}}
\end{figure}
For both above setups, \cite{CherErKis} deals with the resolvent $({A}_\varepsilon-z)^{-1}$ of a self-adjoint operator in $L^2(\mathbb R^d)$ corresponding to the problem \eqref{eq:generic_hom}, so that its solutions are expressed as $u_\varepsilon=({A}_\varepsilon-z)^{-1}f$ with $z=\omega^2.$ For each $\varepsilon>0,$ the operator ${A}_\varepsilon$ is defined by the forms
$$
\int_{{\mathbb R}^d}A^\varepsilon(\cdot/\varepsilon)\nabla u\cdot \overline{\nabla u},\qquad u, v\in H^1({\mathbb R}^d).
$$
It is assumed that $z\in \mathbb C$ is separated from the spectrum of the original operator family, in particular $z\in K_\sigma,$ where $K_\sigma$ is defined in Theorem \ref{kc}.
In order to deal with operators having compact resolvents, it is customary to apply
Gelfand transform \cite{Gelfand}, which we review next.
\subsubsection{Gelfand transform and direct integral}
\label{Gelfand_section}
The version of the Gelfand transform convenient for the analysis of the operators $A_\varepsilon$ is defined on functions $u\in L^2(\mathbb R^d)$ by the formula\footnote{The formula (\ref{Gelfand_formula}) is first applied to continuous functions $U$ with compact support, and then extended to the whole of $L^2({\mathbb R}^d)$ by continuity.} ({\it cf.} (\ref{Gelfand_formula}))
\begin{equation*}
G_\varepsilon u(y,t):=
\biggl(\frac{\varepsilon^2}{2\pi}\biggr)^{d/2}\sum_{n\in{\mathbb Z}^d}u\bigl(\varepsilon(y+n)\bigr)\exp\bigl(-{\rm i}\tau\cdot(y+n)\big),\ \ \ y\in Q,\ \ \tau\in Q':=[-\pi,\pi)^d,
\end{equation*}
This yields a unitary operator $G_\varepsilon: L^2({\mathbb R}^d)\longrightarrow L^2(
Q\times Q'),$ and the inverse
with the inverse mapping given by
\begin{equation*}
u(x)=(2\pi)^{-d/2}\int_{Q'}G_\varepsilon u\biggl(\frac{x}{\varepsilon},\tau\biggr)\exp\biggl({\rm i}\tau\cdot \frac{x}{\varepsilon}\biggr)d\tau,\ \ \ \ x\in{\mathbb R}^d,
\end{equation*}
where $G_\varepsilon u$ is extended to ${\mathbb R}^d\times Q'$ by $Q$-periodicity in the spatial variable.
As in \cite{BirmanSuslina}, an application of the Gelfand transform $G$ to the operator family ${A}_\varepsilon$ corresponding to the problem \eqref{eq:generic_hom} yields the two-parametric family $A_\varepsilon^{(\tau)}$ of operators in $L^2(Q)$ given by the differential expression
$$
-(\nabla + {\rm i}\tau)A^\varepsilon(x/\varepsilon)(\nabla +{\rm i}\tau), \qquad \varepsilon>0,\quad \tau\in Q',
$$
subject to periodic boundary conditions $\partial(\varepsilon Q)$ and defined by the corresponding closed coercive sesquilinear form.
For each $\varepsilon>0,$ the operator ${A}_\varepsilon$ is then unitary equivalent to the von Neumann integral (see {\it e.g.} \cite[Chapter 7]{MR1192782}) of $A_\varepsilon^{(\tau)}:$
$$
{A}_\varepsilon=G_\varepsilon^*\biggl(\oplus \int_{Q'} {A}_\varepsilon^{(\tau)}d\tau\biggr)G_\varepsilon.
$$
Similar to \cite{Friedlander} and facilitated by the abstract framework of \cite{Ryzh_spec},
the operator ${A}_\varepsilon^{(\tau)}$ can be associated to transmission problems \cite{Schechter}, akin to those considered in Section \ref{large_coupling_lim}. To this end, consider $Q$ as a torus with the opposite parts of $\partial Q$ identified, and view $Q_{\rm soft}$ and $Q_{\rm stiff}$ as subsets of this torus.
Furthermore, in line with the notation of Section \ref{large_coupling_lim}, denote by $\Gamma$ the interface between $Q_{\rm soft}$ and $Q_{\rm stiff}.$ For each $\varepsilon,$ $\tau,$ $f\in L^2(Q),$ the transmission problem is formulated as finding a function $u\in L^2(Q)$ such that $u\vert_{Q_{\rm soft}}\in H^1(Q_{\rm soft}),$ $u\vert_{Q_{\rm stiff}}\in H^1(Q_{\rm stiff}),$
that solves, in the weak sense, the boundary-value problem ({\it cf.} (\ref{eq:transmissionBVP}))
\begin{equation*}
\begin{cases}
&-\varepsilon^{-2}(\nabla +{\rm i}\tau)^2u_+-zu_+=f\quad {\rm in}\ \ Q_{\rm stiff},\\[0.45em]
&-(\nabla +{\rm i}\tau)^2u_--zu_-=f\quad {\rm in}\ \ Q_{\rm soft},\\[0.45em]
&u_+=u_-,\qquad
\biggl(\dfrac{\partial}{\partial n_+}+{\rm i}\tau\cdot n_+\biggr)u_++\varepsilon^{-2}\biggl(\dfrac{\partial}{\partial n_-}+{\rm i}\tau\cdot n_-\biggr)u_-=0,\quad {\rm on}\ \ \Gamma.
\end{cases}
\end{equation*}
where
$n_+$ and $n_-=-n_+$ are the outward normals to $\Gamma$ with respect to $Q_{\rm soft}$ and $Q_{\rm stiff}.$
By a classical argument the weak solution of the above problem is shown to coincide with $({A}_\varepsilon^{(\tau)}-z)^{-1}f.$
\subsubsection{Homogenised operators and convergence estimates}
Throughout this section, $H_{\rm hom}:=L^2(Q_{\rm soft})\oplus \mathbb C^1,$ ${\mathcal H}:=L^2(\Gamma),$ and $\partial_n^\tau u:=-(\partial u/\partial n+{\rm i}\tau\cdot n u)|_\Gamma$ is the co-normal boundary derivative for $Q_{\rm soft}.$
{\sc Model I.}
Set
\begin{equation*}
\dom \mathcal A_{\text{\rm hom}}^{(\tau)}=\Bigl\{(u,\beta)^\top\in H_\text{\rm hom}:\ u\in H^2(Q_{\rm soft}), u|_\Gamma=\bigl\langle u|_\Gamma,\psi_0\bigr\rangle_{{\mathcal H}}\psi_0 \text{ and } \beta = \kappa\bigl\langle u|_\Gamma, \psi_0\bigr\rangle_{{\mathcal H}}\Bigr\},
\end{equation*}
where
$\kappa:=|Q_{\rm stiff}|^{1/2}/|\Gamma|^{1/2},$ $\psi_0(x)=|\Gamma|^{-1/2},$ $x\in \Gamma,$ and define
\begin{equation*}
\mathcal A_{\text{\rm hom}}^{(\tau)}\binom{u}{\beta}=
\left(\begin{array}{c}-(\nabla+{\rm i}\tau)^2 u\\[0.5em]
- \kappa^{-1}\bigl\langle\partial_n^\tau u|_\Gamma, \psi_0\bigr\rangle_{{\mathcal H}}
- \kappa^{-2}\varepsilon^{-2}(\mu_*\tau\cdot\tau)\beta
\end{array}\right),\qquad \binom{u}{\beta}\in \dom \mathcal A_{\text{\rm hom}}^{(\tau)},
\end{equation*}
where $\mu_*\tau\cdot\tau$ is the leading-order term (for small $\tau$) of the first Steklov eigenvalue for $-(\nabla+{\rm i}\tau)^2$ on $Q_{\rm soft}.$
{\rm Model II.}
Set
\begin{equation*}
\dom \mathcal A_{\text{\rm hom}}^{(\tau)}=\Bigl\{(u,\beta)^\top\in H_\text{\rm hom}:\ u\in H^2(Q_{\rm soft}), u|_\Gamma=\bigl\langle u|_\Gamma,\psi_\tau\bigr\rangle_{{\mathcal H}}\psi_\tau \text{ and } \beta = \kappa\bigl\langle u|_\Gamma, \psi_\tau\bigr\rangle_{{\mathcal H}}\Bigr\},
\end{equation*}
where $\kappa$ is as above
and $\psi_\tau(x)= |\Gamma|^{-1/2}\exp(-{\rm i}\tau\cdot x)|_\Gamma,$ $x\in\Gamma.$
The action of the operator is set by
\begin{equation*}
\mathcal A_{\text{\rm hom}}^{(\tau)}\binom{u}{\beta}=
\left(\begin{array}{c}-(\nabla+{\rm i}\tau)^2 u\\[0.5em]
-\kappa^{-1}\bigl\langle \partial_n^\tau u|_\Gamma, \psi_\tau\bigr\rangle_{{\mathcal H}}
\end{array}\right),\qquad \binom{u}{\beta}\in \dom \mathcal A_{\text{\rm hom}}^{(\tau)}.
\end{equation*}
{\sc Convergence estimate.} Set $\gamma=2/3$ for the case of Model I and $\gamma=2$ for the case of Model II. The resolvent $({A}_\varepsilon^{(\tau)}-z)^{-1}$ admits the following estimate in the uniform operator-norm topology:
\begin{equation}
\bigl({A}_\varepsilon^{(\tau)}-z\bigr)^{-1}-\Theta^*\bigl(\mathcal A_{\text{\rm hom}}^{(\tau)}-z\bigr)^{-1}\Theta=O(\varepsilon^\gamma),
\label{main_est}
\end{equation}
where $\Theta$ is a partial isometry from $L^2(Q)$ onto $H_\text{\rm hom}:$ on the subspace $L^2(Q_{\rm soft})$ it coincides with the identity, and each function from $L^2(Q_{\rm stiff})$ represented as an orthogonal sum
\[
c_\tau\Vert \Pi_{\rm stiff} \psi_\tau\Vert^{-1}\Pi_{\rm stiff} \psi_\tau\oplus \xi_\tau,\qquad c_\tau\in{\mathbb C}^1,
\]
is mapped to $c_\tau$ unitarily. Here $\Pi_{\rm stiff}$ maps $\varphi$ on $\Gamma$ to the solution $u_\varphi$ of $-(\nabla+{\rm i}\tau)^2u_\varphi=0$ in $Q_{\rm soft},$ $u_\varphi\vert_{\Gamma}=\varphi.$
The estimate (\ref{main_est}) is uniform in $\tau\in Q'$ and $z\in K_\sigma$.
\section*{Acknowledgements}
KDC is grateful for the financial support of EPSRC Grants EP/L018802/2,
EP/V013025/1. The work of all authors has been
supported by CONACyT CF-2019 No.\,304005.
|
\section{Introduction}
\label{introduction}
Recent observations with several different probes have shown accumulating evidence of cosmological-scale anisotropy, and the presence of a Hubble-scale axis. Perhaps the most notable and thoroughly studied probe showing evidence of cosmological-scale anisotropy and a cosmological-scale axis is the cosmic microwave background \citep{eriksen2004asymmetries,cline2003does,gordon2004low,campanelli2007cosmic,zhe2015quadrupole,abramo2006anomalies,mariano2013cmb,land2005examination,ade2014planck,santos2015influence,dong2015inflation,gruppuso2018evens,yeung2022directional}. Other messengers that show cosmological anisotropy and possible axes in the large-scale structure include radio sources \citep{ghosh2016probing,tiwari2015dipole,tiwari2016revisiting}, LX-T scaling \citep{migkas2020probing}, short gamma ray bursts \citep{meszaros2019oppositeness}, cosmological acceleration rates \citep{perivolaropoulos2014large,migkas2021cosmological,krishnan2021hints}, galaxy morphology types \citep{javanmardi2017anisotropy}, Ia supernova \citep{javanmardi2015probing,lin2016significance}, dark energy \citep{adhav2011kantowski,adhav2011lrs,perivolaropoulos2014large,colin2019evidence}, fine structure constant \citep{webb2011indications}, galaxy motion \citep{skeivalas2021predictive}, $H_o$ \citep{luongo2021larger}, polarization of quasars \citep{hutsemekers2005mapping,secrest2021test,zhao2021tomographic,semenaite2021cosmological}, and high-energy cosmic rays \citep{aab2017observation}.
Other studies showed that the large-scale distribution of galaxies in the Universe is not random, as discussed by \cite{de1986slice,hawkins20032df,colless20032df,jones2005scaling,deng2006super,adelman2008ma} and others. Data-driven observations based on large datasets revealed the existence of very large structures \citep{gott2005map,lietzen2016discovery,horvath2015new} that could be beyond astrophysical scale, and therefore challenge the cosmological principle.
These observations can be viewed as a certain tension with the standard cosmological models \citep{pecker1997some,perivolaropoulos2014large,bull2016beyond,velten2020hubble,krishnan2021hints,luongo2021larger}, and triggered several expansions to the standard models, as well as other cosmological theories that shift from the standard models. Possible explanations and theories include double inflation \citep{feng2003double}, primordial anisotropic vacuum pressure \citep{rodrigues2008anisotropic}, contraction prior to inflation \citep{piao2004suppressing}, flat space cosmology \citep{tatum2018flat,tatum2018clues,azarnia2021islands}, multiple vacua \citep{piao2005possible}, spinor-driven inflation \citep{bohmer2008cmb}, and moving dark energy \citep{jimenez2007cosmology}.
Other proposed theories can be related to the geometry of the Universe such as ellipsoidal universe \citep{campanelli2006ellipsoidal,campanelli2007cosmic,campanelli2011cosmic,gruppuso2007complete,cea2014ellipsoidal}, geometric inflation \citep{arciniega2020geometric,edelstein2020aspects,arciniega2020towards,jaime2021viability}, supersymmetric flows \citep{rajpoot2017supersymmetric}, and rotating universe \citep{godel1949example}. Early rotating universe theories were based on a non-expanding universe \citep{godel1949example}, and therefore conflict with the observation that the Universe is expanding. More recent models of rotating Universe were modified to support cosmological expansion \citep{ozsvath1962finite,ozsvath2001approaches,sivaram2012primordial,chechin2016rotation,seshavatharam2020integrated,camp2021}.
The existence of a cosmological-scale axis can also be associated with the theory of black hole cosmology, and can explain cosmic accelerated inflation without the assumption of dark energy \citep{pathria1972universe,stuckey1994observable,easson2001universe,chakrabarty2020toy}. Black holes spin \citep{gammie2004black,takahashi2004shapes,volonteri2005distribution,mcclintock2006spin,mudambi2020estimation,reynolds2021observational}, and their spin is inherited from the spin of the star from which the black hole was created \citep{mcclintock2006spin}. Due to the spin of the black hole, it has been proposed that a universe hosted in a black hole should have an axis and a preferred direction \citep{poplawski2010cosmology,seshavatharam2010physics,seshavatharam2014understanding,christillin2014machian,seshavatharam2020light,seshavatharam2020integrated}. Black hole cosmology is also associated with the theory of holographic universe \citep{susskind1995world,bak2000holographic,bousso2002holographic,myung2005holographic,hu2006interacting,rinaldi2022matrix}, which can also represent the large-scale structure of the Universe in a hierarchical manner \citep{sivaram2013holography,shor2021representation}.
In addition to the messengers discussed above, multiple previous studies showed substantial evidence that the distribution of the spin directions of spiral galaxies is anisotropic, and forms a cosmological-scale axis \citep{macgillivray1985anisotropy,longo2011detection,shamir2012handedness,shamir2013color,shamir2016asymmetry,shamir2017photometric,shamir2017large,shamir2017colour,shamir2019large,shamir2020pasa,shamir2020patterns,shamir2020large,shamir2020asymmetry,lee2019galaxy,lee2019mysterious,shamir2021particles,shamir2021large,shamir2022new}. These observations include different telescopes such as SDSS \citep{shamir2012handedness,shamir2020patterns,shamir2021particles,shamir2022new}, Pan-STARRS \citep{shamir2020patterns}, HST \citep{shamir2020pasa}, and DECam \citep{shamir2021large,shamir2022new}. These telescopes show consistent patterns of the asymmetry, regardless of the telescope being used or the method of annotation of the galaxies \citep{shamir2021large,shamir2022new}.
The alignment of the spin directions of galaxies was observed within cosmic web filaments, as discussed in \citep{tempel2013evidence,tempel2013galaxy,tempel2014detecting,dubois2014dancing,kraljic2021sdss} among other studies, but also between galaxies too far from each other to interact gravitationally \citep{lee2019mysterious}. That alignment is difficult to explain with the standard gravity models, and was defined as ``mysterious'' \citep{lee2019mysterious}. It has also been proposed that the galaxy spin direction is a probe for studying the early Universe \citep{motloch2021observed}.
As a relative measurement, the probe of the large-scale distribution of galaxy spin directions has the advantage of being less sensitive to background contamination such as Milky Way obstruction. The reason is that the asymmetry in galaxy spin directions in a certain field is determined by the difference between the number of galaxies spinning clockwise and the number of galaxies spinning counterclockwise in the same field. Since all galaxies are observed in the exact same field, any background contamination or other effect that affects galaxies spinning clockwise is naturally expected to affect galaxies spinning counterclockwise in the same manner. Because the background contamination affects all galaxies in the field, its existence is not expected to affect the asymmetry. That might be different from other probes such as CMB, where the measurement is absolute, and background contamination that affects a certain field can lead to the observation of anisotropy.
Another advantage of using spin direction asymmetry is that spiral galaxies are very common in the Universe, and are present in a broad redshift range. That important advantage allows to identify not merely the existence of a cosmological-scale axis, but also to determine its approximated location relative to Earth. This paper uses a dataset of $\sim$100K spiral galaxies with spectra to identify the location of the axis formed by the distribution of these galaxies.
\section{Data}
\label{data}
The dataset used in this study is made of galaxies imaged by three different telescopes: the Sloan Digital Sky Survey (SDSS), the Dark Energy Survey (DES), and the Dark Energy Spectroscopic Instrument (DESI) Legacy Survey. That datasets is compared to the distribution of the spin directions of galaxies in datasets used in previous studies.
The directions of the curves of the spiral arm of a galaxy is a reliable indication on the spin direction of the galaxy. For instance, \cite{de1958tilt} used dust silhouette and Doppler shift to determine that in all tested cases the spiral arms were trailing, and therefore allow to determine the spin direction of the galaxy. While in some rare cases galaxies can have leading arms, such as NGC 4622 \citep{freeman1991simulating}, the vast majority of spiral galaxies have trailing arms, and therefore the curve of the galaxy arms can in most cases determine the direction towards which it spins.
The primary task related to the data used in this study is the annotation of galaxies by their spin direction. Although one of the datasets used here was annotated manually \citep{shamir2020pasa}, the scale of databases acquired by modern digital sky surveys is far too large for manual annotation. The practical approach to the annotation of very large datasets of galaxy images is by automatic annotation. It should be mentioned that while pattern recognition, and specifically deep neural networks, have become the common solution to automatic annotation of galaxy images, these approaches might not be suitable for studying subtle cosmological-scale anisotropies \citep{dhar2022systematic}. Pattern recognition, and specifically deep neural networks, are based on complex data-driven rules, and are therefore subjected to subtle biases that are very difficult to identify \citep{carter2020overinterpretation,dhar2021evaluation}, and have also been detected in galaxy images \citep{dhar2022systematic}. Such biases might skew the results. A more thorough discussion about the possible impact of bias in the annotation algorithm is provided in Section~\ref{error}.
\subsection{Automatic annotation of clockwise and counterclockwise galaxies}
\label{ganalyzer}
The galaxies were annotated automatically from the 2D galaxy images by using the {\it Ganalyzer} algorithm \citep{shamir2011ganalyzer}. Ganalyzer is a model-driven algorithm that uses fully symmetric clear mathematical rules to determine the spin direction of a spiral galaxy. The algorithm is described in detail in \citep{shamir2011ganalyzer}, and brief description is also available in \citep{dojcsak2014quantitative,hoehn2014characteristics,shamir2017photometric,shamir2017large,shamir2020patterns,shamir2021particles,shamir2021large,shamir2022new}.
In summary, Ganalyzer first transforms each galaxy image into its radial intensity plot. The radial intensity plot transformation of a galaxy image is a 35$\times$360 image, such that the pixel $(x,y)$ in the radial intensity plot is the median value of the 5$\times$5 pixels around coordinates $(O_x+\sin(\theta) \cdot r,O_y-\cos(\theta)\cdot r)$ in the original galaxy image, where {\it r} is the radial distance in percentage of the galaxy radius, $\theta$ is the polar angle measured in degrees, and $(O_x,O_y)$ are the pixel coordinates of the galaxy center.
Pixels on the galaxy arms are expected to be brighter than pixels that are not on the galaxy arm at the same radial distance from the center. Therefore, peaks in the radial intensity plot are expected to correspond to pixels on the arms of the galaxy at different radial distances from the center. To identify the arms, a peak detection algorithm \citep{morhavc2000identification} is applied to the lines in the radial intensity plot.
Figure~\ref{radial_intensity_plots} shows examples of the radial intensity plots and the peaks detected in them in four DES galaxies. The figure also shows the radial intensity plot of each galaxy, and the lines formed by the detected peaks. Each line in the radial intensity plot shows the brightness of the pixels around the center of the galaxy. That is, the first value in the line is the brightness of the pixel at 0$^o$ compared to the galaxy center, and the last value in the line is the brightness of the pixel at 359$^o$. Since each line is at a different radius from the center, each radial intensity plot has multiple lines. Below each radial intensity plot shown in Figure~\ref{radial_intensity_plots}, the figure displays the peaks identified in the lines. That is, if a pixel in the radial intensity plot is identified as a peak in its line, the corresponding pixel in the image below the radial intensity plot is white. Otherwise, the pixel is black. More information about Ganalyzer can be found in \citep{shamir2011ganalyzer,dojcsak2014quantitative,shamir2017photometric,shamir2017colour,shamir2017large,shamir2019large,shamir2020patterns,shamir2021large,shamir2022new}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.44]{radial_intensity_plots.pdf}
\caption{Examples of galaxy images and their corresponding radial intensity plots. The peaks detected in the radial intensity plots are displayed below the radial intensity plots. The directions of the lines formed by the peaks reflect the curves of the arms, and consequently the spin direction of the galaxy. The algorithm is symmetric, and it is model-driven with intuitive rules. }
\label{radial_intensity_plots}
\end{figure}
As Figure~\ref{radial_intensity_plots} shows, the lines formed by the peaks identified in the radial intensity plots form lines in different directions. A linear regression is applied to the peaks in adjunct lines formed by the peaks, and the slope of the linear regression reflects the curve of the arm. As Figure~\ref{radial_intensity_plots} shows, if the galaxy spin clockwise the slope of the regression is positive, while if the galaxy spins counterclockwise the slope is negative. Therefore, the slope of the regression can be used to determine the spin direction of the galaxy.
Naturally, many galaxies are elliptical galaxies, irregular galaxies, or spiral galaxies that do not have an identifiable spin direction. To avoid galaxies that do not have an identifiable spin direction, only galaxies with 30 or more peaks that form curved lines in the radial intensity plots are used. If that criteria is not met, the galaxy is determined to have an unidentifiable spin direction. As mentioned above, the main advantage of the algorithm is that it follows defined and fully symmetric rules. Analysis of different situations when applying the algorithm to populations of galaxies are described in \cite{shamir2021particles}, and also more briefly in Section~\ref{error}.
\subsection{Galaxy images and digital sky surveys}
The dataset of galaxies with spectra used in this study is based on data collected by three different digital sky surveys: SDSS, DES, and the DESI Legacy Survey. The SDSS galaxies are 63,693 galaxies with spectra used in \citep{shamir2019large,shamir2020patterns,shamir2020pasa}. The preparation of that dataset is described in detail in \citep{shamir2020patterns}. The galaxies from the DESI Legacy Survey are the subset of galaxies used by \cite{shamir2021large} that had spectroscopic redshift through the catalog of \cite{zhou2021clustering}. The entire DESI Legacy Survey dataset contained 807,898 galaxies \citep{shamir2021large}, but only 23,715 of these galaxies had spectroscopic redshift through the catalog of \cite{zhou2021clustering}. The \citep{zhou2021clustering} catalog contains mostly the photometric redshift of the DESI Legacy Survey galaxies. As also discussed in Section~\ref{error}, since the inaccuracy of photometric redshift is greater than the expected signal, the photometric redshift cannot be used for this study, and therefore only galaxies that had spectroscopic redshift are used.
Similarly to the DESI dataset, galaxies from DES data release (DR) 1 were also used. The initial list of DES objects included all objects identified as exponential disks, de Vaucouleurs ${r}^{1/4}$ profiles, or round exponential galaxies, and were brighter than 20.5 magnitude in one or more of the g, r or z bands. That list contained an initial set of 18,869,713 objects. The galaxy images were downloaded using the {\it cutout} API of the DESI Legacy Survey server, which also provide access to DES data. The size of each image was 256$\times$256, and retrieved in the JPEG format. Each image was scaled using the Petrosian radius to ensure that the galaxy fits in the image. The process of downloading the images started on April 25th 2021, and ended about six months later on November 1st 2021.
Once the image files were downloaded, they were annotated by their spin direction using the Ganalyzer method described above and in \citep{shamir2011ganalyzer,dojcsak2014quantitative,shamir2017photometric,shamir2017colour,shamir2017large,shamir2019large,shamir2020patterns,shamir2021large,shamir2022new}. The annotation of the galaxies lasted 73 days of operation using a single Intel Xeon processor. Then, the images were mirrored using {\it ImageMagick} and annotated again to allow repeating the experiments with mirrored images. That provided a dataset of 773,068 galaxies annotated by their spin directions. To remove satelite galaxies or stars positioned inside a galaxy, objects that had another object in the dataset within 0.01$^o$ or less were removed from the dataset. That provided a dataset of 739,286 galaxies imaged by DES, and 14,365 of these galaxies had redshift through the 2dF redshift survey \citep{colless20032df,cole20052df}.
Combining all three datasets and removing objects that appeared in more than one dataset provided a dataset of 90,023 galaxies with spectra. Figure~\ref{z_distribution} shows the distribution of the redshift in the dataset, and Figure~\ref{ra_distribution} shows the distribution of the galaxies by their RA. Figure~\ref{population} shows the galaxy population density in each $5^o\times5^o$ field of the sky. The density is determined by the number of galaxies in each $5^o\times5^o$ field divided by the total number of galaxies.
\begin{figure}[h]
\centering
\includegraphics[scale=0.7]{ra_distribution.pdf}
\caption{The distribution of the galaxies in the dataset in different 30$^o$ RA ranges.}
\label{ra_distribution}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=0.65]{z_distribution.pdf}
\caption{The redshift distribution of the galaxies.}
\label{z_distribution}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=0.15]{population.pdf}
\caption{The density of the galaxy distribution in the dataset. The density in each $5^o\times5^o$ field of the sky is determined by the number of galaxies in the $5^o\times5^o$ field, divided by the total number of galaxies in the dataset.}
\label{population}
\end{figure}
An important property of the dataset used in this study is that the data are retrieved and analyzed without any prior assumptions, and are not based on any existing catalog of galaxy morphology. Because the dataset is made of data from several telescopes covering both the Northern and Southern hemispheres, there is no existing catalog of galaxy morphology that includes all galaxies used in this study. More importantly, using an existing catalog of galaxy morphology might expose the analysis to bias of a catalog that was not prepared with the normalization of spin direction in mind. For instance, morphology catalogs that were prepared manually can be biased by the human perception \citep{land2008galaxy,hayes2017nature}. Due to the complex nature of the human perception, these subtle but consistent biases are very difficult to quantify and correct.
In the past two decades, catalogs of galaxy morphology were prepared by using machine learning algorithms, and specifically deep neural networks such as \citep{gravet2015catalog,perez2015morphologies,goddard2020catalog,cheng2021galaxy}. While these catalogs are prepared by using a computer software, the rules used to make the annotation are determined automatically by using training sets of galaxies that were annotated manually. Therefore, the machine learning systems can still capture the perceptional biases of the humans who annotated the data that was used to train the machine learning system. Also, the rules used by these machine learning systems are complex non-intuitive rules generated automatically from the data by which the machine learning system is trained. Due to their complexity, it is very difficult to verify theoretically or empirically that these rules are fully symmetric, and do not lead to certain biases.
It might be reasonable to assume that if such bias existed, it would have been expected to be consistent throughout the sky, and would not exhibit inverse asymmetry in opposite hemispheres. However, due to the complex nature of these algorithms it is very difficult to prove such claim. For instance, it has been shown that by selecting a different training set, a deep neural network produces a slightly but consistently different catalog \citep{dhar2022systematic}. Therefore, using catalogs that were prepared for other purposes, and do not necessarily ensure that the algorithm is fully symmetric in terms of the spin directions of the galaxies, can introduce an additional source of bias that depends on the way the catalog was prepared, and might be carried on to the rest of the analysis.
As discussed in Section~\ref{ganalyzer}, the algorithm used to determine the spin direction of the galaxies can also perform the broad morphological classification of elliptical and spiral galaxies \citep{shamir2011ganalyzer}, and identify just galaxies with identifiable spin direction. Elliptical galaxies or other galaxies that their spin direction cannot be determined are not included in the analysis, and the removal of the galaxies is done by the same model-driven symmetric algorithm described in Section~\ref{ganalyzer}. Further discussion on the use of catalogs of galaxy morphology can be found in Section~\ref{morphology_catalogs}.
\section{Results}
\label{results}
A very simple way of observing the distribution of galaxy spin directions in a certain field in the sky is by comparing the number of galaxies spinning clockwise to the number of galaxies spinning counterclockwise. The statistical significance of the difference can be determined by using binomial distribution, such that the probability of a galaxy to spin clockwise or counterclockwise is assumed to be 0.5. The asymmetry {\it A} can be defined by $A=\frac{cw-ccw}{cw+ccw}$, where {\it cw} is the number of galaxies spinning clockwise, and {\it ccw} is the number of galaxies spinning counterclockwise.
A galaxy that seems to spin clockwise to an observer on Earth would seem to spin counterclockwise if the observer was in the opposite side of the galaxy. Therefore, the asymmetry of the distribution of galaxy spin directions in one hemisphere is expected to be inverse in the opposite hemisphere. Perhaps the most simple way of separating the sky into two hemispheres is such that one hemisphere is $(0^o<\alpha<180^o)$, and the other hemisphere is $(180^o<\alpha<360^o)$. That separation is clearly very simple, and used in this case for the sake of simplicty. Table~\ref{hemispheres} shows the number of clockwise and counterclockwise galaxies in each hemisphere, and the binomial probability to have such difference by chance.
\begin{table}
\centering
\begin{tabular}{lcccc}
\hline
Hemisphere & \# cw & \# ccw & $\frac{cw-ccw}{cw+ccw}$ & P \\
& galaxies & galaxies & & \\
\hline
$(0^o<\alpha<180^o)$ & 23,070 & 23,606 & -0.0115 & 0.007 \\
$(180^o<\alpha<360^o)$ & 21,808 & 21,539 & 0.0062 & 0.09 \\
\hline
\end{tabular}
\caption{Distribution of clockwise and counterclockwise galaxies in opposite hemispheres. The P values are the binomial distribution probability to have such difference or stronger by chance when assuming mere chance 0.5 probability for a galaxy to spin clockwise or counterclockwise.}
\label{hemispheres}
\end{table}
As the table shows, the hemisphere $(0^o<\alpha<180^o)$ has a higher number of galaxies spinning clockwise, while the opposite hemisphere has a higher number of galaxies spinning counterclockwise. While the sign of the asymmetry is inverse in the opposite hemisphere, the asymmetry in the hemisphere $(180^o<\alpha<360^o)$ is not necessarily statistically significant. However, even if assuming no asymmetry in that hemisphere, after applying a Bonferroni correction the the Bonferroni-corrected P value of the asymmetry in $(0^o<\alpha<180^o)$ is $\sim$0.014, which is statistically significant. When repeating the analysis by using the mirrored images the results inverse, as expected due to the symmetric nature of the Ganalyzer algorithm that was used to annotate the images.
The inverse asymmetry of galaxy spin direction between opposite hemispheres has been done in the past with several telescopes such as SDSS \citep{shamir2020patterns,shamir2020pasa}, DECam \citep{shamir2021large,shamir2022new}, and Pan-STARRS \citep{shamir2020patterns,shamir2021large,shamir2022new}. For instance, as shown in \citep{shamir2022new}, separating the entire dataset of galaxies imaged by DECam, mostly in the Southern hemisphere, into two hemispheres provided significant difference between the number of galaxies spinning in opposite directions in the two opposite hemispheres.
Clearly, the simple separation to two opposite hemispheres is a very simple analysis, and a more thorough analysis will be provided later in this section. One of the aspects that can be tested is the relationship between the magnitude of the difference and the redshift. Table~\ref{hemispheres_z} shows the same analysis shown in Table~\ref{hemispheres} for the hemisphere $(0^o<\alpha<180^o)$, but after separating the galaxies into galaxies with redshift of $z<0.15$ and galaxies with redshift of $z>0.15$.
\begin{table}
\begin{tabular}{lcccc}
\hline
Redshift & \# cw & \# ccw & $\frac{cw-ccw}{cw+ccw}$ & P \\
range & galaxies & galaxies & & \\
\hline
$z<0.15$ & 15,767 & 15,877 & -0.0034 & 0.26 \\
$z>0.15$ & 7,303 & 7,729 & -0.0283 & 0.0002 \\
\hline
\end{tabular}
\caption{Distribution of clockwise and counterclockwise galaxies in the hemisphere $(0^o<\alpha<180^o)$ in different redshift ranges.}
\label{hemispheres_z}
\end{table}
As the table shows, in the higher redshift ranges the asymmetry becomes significantly stronger than when limiting the redshift to lower redshift ranges. The observation is in agreement with previous results using smaller datasets of galaxies with spectra \citep{shamir2020patterns}. A more detailed analysis that shows the consistent increase of the difference and the statistical signal with the redshift is described in \citep{shamir2020patterns}.
\subsection{Analysis of a dipole axis in different redshift ranges}
\label{dipole}
The analysis by separating the sky into two simple hemispheres has the advantage of simplicity, but might not provide a full accurate analysis of the location of the most probable axis around which the galaxies are aligned. The separation into two simple hemispheres is arbitrary, and while it shows evidence of asymmetry by suing very simple statistics, it does not allow to profile or identify the location of the most likely axis. A more comprehensive analysis of the presence of a dipole axis can be done by fitting the cosine of the angular distances of the galaxies from each possible $(\alpha,\delta)$ combination to their spin directions.
That analysis can be done by assigning the galaxies with their spin direction $d$, which is 1 for galaxies spinning clockwise, and -1 for galaxies spinning counterclockwise. The cosines of the angular distances $\phi$ is then $\chi^2$ fitted into $d\cdot|\cos(\phi)|$. From each possible $(\alpha,\delta)$ integer combination, the angular distance $\phi_i$ between $(\alpha,\delta)$ and each galaxy {\it i} in the dataset is computed. The $\chi^2$ from each $(\alpha,\delta)$ is determined by Equation~\ref{chi2}
\begin{equation}
\chi^2_{\alpha,\delta}=\Sigma_i \frac{(d_i \cdot | \cos(\phi_i)| - \cos(\phi_i))^2}{\cos(\phi_i)} ,
\label{chi2}
\end{equation}
where $d_i$ is the spin direction of galaxy {\it i}, and $\phi_i$ is the angular distance between galaxy {\it i} and $(\alpha,\delta)$.
For each $(\alpha,\delta)$ combination, the $\chi^2$ is computed with the real spin directions of the galaxies, and then computed 1000 times when $d_i$ is assigned with a random spin direction. Using the $\chi^2$ from 1000 runs, the mean and standard deviation of the $\chi^2$ when the spin directions are random is computed. Then, the $\sigma$ difference between the $\chi^2$ computed with the real spin directions and the mean $\chi^2$ computed with the random spin directions is used to determine the $\sigma$ of the $\chi^2$ fitness to occur by chance. That is done for each $(\alpha,\delta)$ integer combination in the sky to determine the probability of each $(\alpha,\delta)$ to be the center of the dipole axis \citep{shamir2012handedness,shamir2019large,shamir2020pasa,shamir2020patterns,shamir2021particles,shamir2021large,shamir2022new}.
Figure~\ref{dipole_all} shows the computed probabilities of a dipole axis in different $(\alpha,\delta)$ combinations. The most probable axis is at $(\alpha=65^o,\delta=52^o)$, with probability of 4.7$\sigma$ to occur by chance. The 1$\sigma$ error of the location of that axis is $(0^o,122^o)$ for the RA, and $(42^o,-77^o)$ for the declination.
\begin{figure}
\centering
\includegraphics[scale=0.15]{dipole_all.pdf}
\caption{The probability of a dipole axis in different $(\alpha,\delta)$ combinations.}
\label{dipole_all}
\end{figure}
When assigning the galaxies with random spin directions, the asymmetry becomes insignificant \citep{shamir2012handedness,shamir2019large,shamir2020pasa,shamir2020patterns,shamir2020large,shamir2021particles,shamir2021large,shamir2022new}. Figure~\ref{z_all_random} shows the probabilities of a dipole axis in different $(\alpha,\delta)$ combinations when using the same galaxies used in Figure~\ref{dipole_all}, but when assigning these galaxies with random spin directions. The most likely axis has statistical strength of 0.91$\sigma$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.15]{z_all_random.pdf}
\caption{The probability of a dipole axis at different $(\alpha,\delta)$ combinations when the galaxies are assigned with random spin directions.}
\label{z_all_random}
\end{figure}
Figure~\ref{dipole_z} shows the profile when fitting the galaxy spin directions into a dipole axis alignment using galaxies at different redshift ranges. The figure shows that the location of the most likely axis changes with the redshift, until around the redshift range of $0.12<z<0.22$, after which it stays constant. One immediate explanation to the change in the position of the most likely location of the axis when the redshift increases is that the axis does not necessarily go through Earth.
\begin{figure*}[h]
\centering
\includegraphics[scale=0.16]{dipole_z.pdf}
\caption{The dipole axes in galaxy spin direction from different $(\alpha,\delta)$ combinations in different redshift ranges.}
\label{dipole_z}
\end{figure*}
The profiles of asymmetry observed with the low redshift galaxies can be compared to previous analyses of galaxy spin directions using Pan-STARRS \citep{shamir2020patterns} and DECam \citep{shamir2021large}. The Pan-STARRS dataset used in \cite{shamir2020patterns} includes 33,028 galaxies, and the DECam dataset contains 807,898 galaxies \citep{shamir2021large}. The vast majority of the galaxies in these datasets do not have redshift, but because the magnitude of all galaxies is limited to 19.5, the redshift is also expected to be relatively low. Figure~\ref{panstarrs_decam} shows the profile of asymmetry observed with Pan-STARRS and DECam as reported in \citep{shamir2021large,shamir2022new}. These results are compared to the results with the dataset described in Section~\ref{data} when 43,566 galaxies are selected such that the redshift distribution of these galaxies is similar to the redshift distribution of the DECam galaxies as determined by the redshift distribution of the few DECam galaxies with spectra. The agreement between different telescopes indicates that the distribution of galaxy spin direction reflects the actual Universe rather than a certain unknown anomaly in a telescope.
\begin{figure}[h]
\centering
\includegraphics[scale=0.25]{panstarrs_decam.pdf}
\caption{The profile of galaxy spin direction distribution in the dataset described in Section~\ref{data} (top), compared to the profiles observed in Pan-STARRS (bottom) and DECam (middle) as reported in \citep{shamir2021large}. The graphs shows similar profiles of galaxy spin direction distribution observed in the different telescopes.}
\label{panstarrs_decam}
\end{figure}
The asymmetry in the distribution of galaxy spin directions observed when the redshift of the galaxies is relatively high can be compared to the asymmetry profile observed by galaxies imaged by the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey \citep{grogin2011candels,koekemoer2011candels} of the Hubble Space Telescope (HST). The HST dataset contains 8,690 galaxies annotated manually by their spin direction, as explained in \citep{shamir2020pasa}. During that process of annotation, a random half of the galaxies were mirrored for the first cycle of annotation, and then all galaxies were mirrored for a second cycle of annotation as described in \citep{shamir2020pasa} to offset possible effect of human perceptional bias. That provided a complete dataset that is also not subjected to atmospheric effects. A full description of the dataset and the analysis of the distribution of galaxy spin direction is described in \citep{shamir2020pasa}. These galaxies do not have spectra, but due to the nature of the Hubble Space Telescope it is clear that the galaxies imaged by HST are of much higher redshift than the galaxies imaged by the Earth-based sky surveys. Figure~\ref{sdss_hst} shows the profile observed using the HST galaxies as described in \citep{shamir2020pasa}, and the profile of asymmetry when using the dataset described in Section~\ref{data} when the redshift of the galaxies is limited to $0.16<z<0.26$. The most likely axis observed with HST galaxies is at $(\alpha=78^o,\delta=47^o)$, with probability of 2.8$\sigma$ to have such distribution by chance. As the figure shows, although the two datasets do not share any galaxies, they show similar positions of the most likely axes observed in the distribution of galaxy spin directions. The most likely axis observed with the dataset described in Section~\ref{data} peaks at $(\alpha=48^o,\delta=67^o)$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.15]{sdss_hst.pdf}
\caption{The probability of a dipole axis in galaxy spin directions from different integer $(\alpha,\delta)$ combinations in HST (bottom) and the galaxies in the dataset described in Section~\ref{data} (top) when the redshift of the galaxies is limited to $0.16<z<0.26$.}
\label{sdss_hst}
\end{figure}
Table~\ref{dipole_redshift_ranges} shows the location of the most likely axis when the dataset described in Section~\ref{data} is separated into different redshift ranges. As the table shows, the declination of the galaxies changes consistently as the redshift gets higher, while the change in the right ascension is milder.
\begin{table}
\scriptsize
\begin{tabular}{|c|c|c|c|c|}
\hline
z & \# galaxies & RA & Dec & Statistical \\
& & (degrees) & (degrees) & significance \\
\hline
0.0-0.1 & 41,218 & 61 & 10 & 1.3$\sigma$ \\
0.01-0.11 & 45,618 & 56 & 22 & 1.2$\sigma$ \\
0.02-0.12 & 49,693 & 53 & 30 & 1.6$\sigma$ \\
0.03-0.13 & 51,027 & 55 & 32 & 1.7$\sigma$ \\
0.04-0.14 & 51,243 & 62 & 36 & 1.9$\sigma$ \\
0.05-0.15 & 50,446 & 48 & 45 & 2.1$\sigma$ \\
0.06-0.16 & 48,362 & 54 & 46 & 2.4$\sigma$ \\
0.07-0.17 & 44,739 & 60 & 49 & 3.2$\sigma$ \\
0.08-0.18 & 40,101 & 53 & 52 & 3.4$\sigma$ \\
0.09-0.19 & 35,565 & 58 & 55 & 4.0$\sigma$ \\
0.1-0.2 & 32,206 & 61 & 53 & 3.6$\sigma$ \\
0.11-0.21 & 28,486 & 61 & 50 & 2.9$\sigma$ \\
0.12-0.22 & 24,462 & 54 & 60 & 2.7$\sigma$ \\
0.13-0.23 & 21,001 & 58 & 58 & 2.8$\sigma$ \\
0.14-0.24 & 17,497 & 55 & 65 & 3.2$\sigma$ \\
0.15-0.25 & 14,785 & 57 & 64 & 3.3$\sigma$ \\
0.16-0.26 & 12,664 & 48 & 67 & 3.1$\sigma$ \\
\hline
\end{tabular}
\caption{The most likely axis when the galaxies are limited to different redshift ranges.}
\label{dipole_redshift_ranges}
\end{table}
The change in the location of the peak of the most likely axis when the redshift of the galaxies changes can be viewed as an indication that the axis does not necessarily go directly through Earth. In such case, the location of the most likely axis is expected to change at low redshifts, and then to remain nearly constant at the higher redshifts. Figure~\ref{restframe_axis} displays a simple two-dimensional illustration of a possible axis compared to Earth. The angle $\alpha$ is measured between two points determined using two different redshifts, but the two redshifts are relatively low. That angle is much larger than $\beta$, which is the difference in the position of the dipole axis as seen from Earth when using two higher redshifts. Therefore, an axis that does not go directly through Earth is expected to change its location as seen from Earth in lower redshifts, but remain nearly at the same location when observed in higher redshift ranges.
\begin{figure}[h]
\centering
\includegraphics[scale=0.25]{restframe_axis3.pdf}
\caption{Changes in the position of a non-Earth centered axis as seen from Earth when its peaks are identified at different redshifts. The change in the location of the axis in lower redshift range $\alpha$ is far larger than the change in the location in higher redshift range $\beta$.}
\label{restframe_axis}
\end{figure}
Figure~\ref{dipole_axis_points} visualizes the most likely axis points in a 3D space such that the distance {\it d} is determined by converting the mean redshift of the galaxies in each redshift range to the distance, measured in Mpc. The 3D transformation is then performed by Equation~\ref{transformation}. As expected, the points are aligned in a manner that forms a three-dimensional axis.
\begin{equation}
\begin{split}
x= & \cos(\alpha)\cdot d \cdot \cos(\delta) \\
y= & \sin(\alpha)\cdot d \cdot \cos(\delta) \\
z= & d \cdot sin(\delta) \\
\label{transformation}
\end{split}
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[scale=0.35]{dipole_axis_points.pdf}
\caption{Three-dimensional visualization of the points in Table~\ref{dipole_redshift_ranges}. The points form an axis that might not go directly through Earth.}
\label{dipole_axis_points}
\end{figure}
To estimate the three dimensional direction of the axis, a simple analysis between two points $\eta$ and $\xi$ in a three-dimensional space can be used as shown by Equation~\ref{az_alt_from_3D}. Since the points in Table~\ref{dipole_redshift_ranges} as visualized in Figure~\ref{restframe_axis} are aligned in an axis, the direction of the axis can be determined by the two most distant points from each other. These two points are also the two points with the lowest and highest redshift as shown in Table~\ref{dipole_redshift_ranges}.
The closet point to Earth shown in Table~\ref{dipole_redshift_ranges} is at $(\alpha=61^o,\delta=10^o)$, and $\sim$283 Mpc away, based on the average 0.064 mean redshift of the galaxies in that redshift range. An observer in that point would see an axis in $(\alpha=262^o,\delta=85^o)$, and the other end of that axis in $(\alpha=82^o,\delta=-85^o)$.
\begin{equation}
\begin{split}
\alpha = & atan2(y_{\eta}-y_{\xi}, \sqrt{(x_{\eta}-x_{\xi})^2 + (z_{\eta}-z_{\xi})^2}) \\
\delta = & atan2(-(x_{\eta}-x_{\xi}),-(z_{\eta}-z_{\xi})) \\
\label{az_alt_from_3D}
\end{split}
\end{equation}
\subsection{Quadrupole axis analysis}
\label{quadrupole}
Analysis of the distribution of the CMB provided consistent evidence of a quadrupole axis alignment of the CMB anisotropy \citep{efstathiou2003low,feng2003double,cline2003does,weeks2004well,gordon2004low,piao2004suppressing,piao2005possible,campanelli2006ellipsoidal,campanelli2007cosmic,gruppuso2007complete,jimenez2007cosmology,rodrigues2008anisotropic,zhe2015quadrupole,santos2015influence}. Fitting the galaxy spin directions into quandrupole axis alignment can be done as described in Section~\ref{dipole}, but instead of $\chi^2$ fitting $\cos(\phi)$ into $d\cdot|\cos(\phi)|$ as was done for identifying a dipole axis, the quadrupole alignment is identified by $\chi^2$ fitting $\cos(2\phi)$ into $d\cdot|\cos(2\phi)|$, as described in \citep{shamir2019large,shamir2020patterns,shamir2021large}.
Figure~\ref{quad_all} shows the likelihood of a quadrupole axis in different $(\alpha,\delta)$ combinations. One axis peaks at $(\alpha=52^o,\delta-8^o)$, with statistical signal of 4.3$\sigma$. The 1$\sigma$ error range for that axis is $(34^o,87^o)$ for the RA, and $(-42^o,36^o)$ for the declination. The other axis peaks at $(\alpha=151^o,\delta=31^o)$ with statistical signal of 3.1$\sigma$. The 1$\sigma$ error of that axis is $(118^o,191^o)$ for the RA, and $(-5^o,66^o)$ for the declination. Figure~\ref{quad_random} shows the same analysis when the galaxies are assigned random spin directions, showing that the axes disappear when the galaxy spin directions are random.
\begin{figure}
\centering
\includegraphics[scale=0.15]{quad_all.pdf}
\caption{The probability of a quadrupole axis in galaxy spin directions in different $(\alpha,\delta)$ combinations.}
\label{quad_all}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.15]{quad_random.pdf}
\caption{The probability of a quadrupole axis in galaxy spin directions in different $(\alpha,\delta)$ combinations when the galaxies are assigned with random spin directions.}
\label{quad_random}
\end{figure}
As with the analysis of the dipole alignment, the analysis shown in Figure~\ref{quad_all} can be compared to previous analyses \citep{shamir2020patterns,shamir2021large}. Figure~\ref{panstarrs_decam_quad} displays the same analysis with Pan-STARRS and DECam as shown in \citep{shamir2021large}, compared to the analysis of the dataset described in Section~\ref{data}, using the 43,566 that have the same redshift distribution as the subset of DECam galaxies with redshift. The figure shows that the profile of galaxy spin direction distribution observed with the dataset described in Section~\ref{data} is similar to the previous results using Pan-STARRS and DECam shown in \citep{shamir2021large}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.25]{panstarrs_decam_quad.pdf}
\caption{The quadrupole profile of galaxy spin direction distribution in the dataset described in Section~\ref{data} (top), compared to the quadrupole profiles observed in Pan-STARRS (bottom) and DECam (middle) as reported in \citep{shamir2021large}.}
\label{panstarrs_decam_quad}
\end{figure}
Figure~\ref{quad_z} shows the analysis when fitting the distribution of galaxy spin direction to quadrupole axis alignment as observed when separating the redshift of the galaxies to certain different redshift ranges. As with the dipole alignment, the statistical signal of the quadrupole alignment becomes stronger as the redshift gets higher. The locations of the most likely axes change at the relatively lower redshifts, indicating that the axis is not necessarily Earth-centered. Table~\ref{quad_redshift_ranges} shows the most likely position of the two axes in each redshift range.
\begin{figure*}
\centering
\includegraphics[scale=0.15]{quad_z.pdf}
\caption{The probability of a quadrupole axes at different $(\alpha,\delta)$ combinations in different redshift ranges.}
\label{quad_z}
\end{figure*}
\begin{table*}
\scriptsize
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
z & \# galaxies & RA & Dec & Statistical & RA & Dec & Statistical \\
& & (Axis 1) & (Axis 1) & significance & (Axis 2) & (Axis 2) & significance \\
\hline
0.0-0.1 & 41,218 & 347 & 26 & 1.2$\sigma$ & 63 & -12 & 0.8$\sigma$ \\
0.01-0.11 & 45,618 & 344 & 23 & 1.2$\sigma$ & 61 & -12 & 0.8$\sigma$ \\
0.02-0.12 & 49,693 & 342 & 21 & 1.2$\sigma$ & 61 & -14 & 0.9$\sigma$ \\
0.03-0.13 & 51,027 & 345 & 20 & 1.3$\sigma$ & 65 & -15 & 1.1$\sigma$ \\
0.04-0.14 & 51,243 & 341 & 18 & 1.6$\sigma$ & 63 & -17 & 1.2$\sigma$ \\
0.05-0.15 & 50,446 & 348 & 15 & 1.4$\sigma$ & 68 & -17 & 1.2$\sigma$ \\
0.06-0.16 & 48,362 & 348 & -27 & 1.5$\sigma$ & 71 & -20 & 1.5$\sigma$ \\
0.07-0.17 & 44,739 & 350 & -36 & 1.6$\sigma$ & 66 & -25 & 1.5$\sigma$ \\
0.08-0.18 & 40,101 & 353 & -40 & 1.5$\sigma$ & 67 & -33 & 1.7$\sigma$ \\
0.09-0.19 & 35,565 & 344 & -38 & 1.8$\sigma$ & 70 & -34 & 2.1$\sigma$ \\
0.1-0.2 & 32,206 & 349 & -35 & 2.2$\sigma$ & 64 & -37 & 2.6$\sigma$ \\
0.11-0.21 & 28,486 & 351 & -37 & 2.7$\sigma$ & 64 & -36 & 2.9$\sigma$ \\
0.12-0.22 & 24,462 & 347 & -34 & 2.9$\sigma$ & 62 & -34 & 3.4$\sigma$ \\
0.13-0.23 & 21,001 & 350 & -36 & 3.2$\sigma$ & 60 & -37 & 3.7$\sigma$ \\
0.14-0.24 & 17,497 & 345 & -35 & 3.5$\sigma$ & 61 & -35 & 3.9$\sigma$ \\
0.15-0.25 & 14,785 & 341 & -32 & 3.9$\sigma$ & 57 & -37 & 4.1$\sigma$ \\
0.16-0.26 & 12,664 & 338 & -36 & 3.1$\sigma$ & 65 & -36 & 3.9$\sigma$ \\
\hline
\end{tabular}
\caption{The most likely quadrupole axis when the galaxies are limited to different redshift ranges.}
\label{quad_redshift_ranges}
\end{table*}
Figure~\ref{quad_axis_points} displays the points in Table~\ref{quad_redshift_ranges} in a three dimensional space. The figure shows that the two axes form two lines that can be considered straight lines. Figure~\ref{x_y_z_points} shows the two-dimensional projections of the x,y plane and the x,z plane. The directions of the two axes can be determined by applying a linear regression to the points of each axis. According to the alignment of the points in Table~\ref{quad_redshift_ranges}, the two axes meet at RA 313$^o$ and declination of 65$^o$ from Earth, and at distance of 1736 Mpc.
\begin{figure}[h]
\centering
\includegraphics[scale=0.45]{quad_axis_points.pdf}
\caption{Three dimensional visualization of the points in Table~\ref{quad_redshift_ranges}.}
\label{quad_axis_points}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=0.45]{x_y_z_points.pdf}
\caption{The X,Y two-dimensional projection of the points in Table~\ref{quad_redshift_ranges} (top), and the X,Z two-dimensional projections.}
\label{x_y_z_points}
\end{figure}
\subsection{Analysis of differences in galaxy morphology}
\label{morphology}
Another analysis was focused on differentiating the morphological differences between galaxies. For that purpose, the galaxies were separated by their $\frac{r}{S}$, where $r$ is the radius (in pixels), and $S$ is the galaxy surface size, also measured in pixels as described in \citep{shamir2011ganalyzer}. The surface size of the galaxy is measured by the number of foreground pixels, which can be counted after separating the foreground and background pixels of the galaxy as described in \citep{shamir2011ganalyzer}. That allowed to separate the galaxies by their mass distribution. A smaller $\frac{r}{S}$ value indicates that the surface size of the galaxy is large compared to its radius, and therefore the galaxy is more dense. Figure~\ref{morphology_type} shows examples of different galaxies and their $\frac{r}{S}$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{morphology_type.pdf}
\caption{$\frac{r}{S}$ of different galaxies. Higher $\frac{r}{S}$ indicates that the galaxy is less dense. The galaxies on the left have a relatively low $\frac{r}{S}$ of 0.0068 and 0.0065, indicating that the galaxies are more dense. The more sparse galaxies on the right have a higher $\frac{r}{S}$ of 0.0185 and 0.0176.}
\label{morphology_type}
\end{figure}
To test for the possible effect of galaxy morphology on the analysis, the data was separated into two subsets of similar size, such that in one subset all galaxies had $\frac{r}{S}\leq0.01$, and in the other subset all galaxies satisfied $\frac{r}{S}>0.01$. The subsets contained 45,526 and 44,497 galaxies, respectively. That led to two subsets that differ from each other by their mass distribution.
Figure~\ref{dipole_morphology} shows the results of the analysis described in Section~\ref{dipole}, such that the galaxies in the dataset are separated to galaxies with $\frac{r}{S}\leq0.01$ (top) and galaxies with $\frac{r}{S}>0.01$ (bottom). The figure shows that the two subsets of galaxies show a dipole axis at roughly the same location. When using galaxies that satisfy $\frac{r}{S}>0.01$, the statistical strength of the axis is 4.6$\sigma$. When the galaxies are limited to $\frac{r}{S}\leq0.01$, the statistical significance is 3.5$\sigma$. The 1.1$\sigma$ difference might provide certain indication of a link between the observed asymmetry and morphology of the galaxies.
\begin{figure}
\centering
\includegraphics[scale=0.26]{dipole_morphology.pdf}
\caption{The probability of a dipole axis in different $(\alpha,\delta)$ combinations when the galaxies have $\frac{r}{S}\leq0.01$ (top), and $\frac{r}{S}>0.01$ (bottom).}
\label{dipole_morphology}
\end{figure}
\section{Analysis of reasons that can lead to asymmetry not originated from the real sky}
\label{error}
While several messengers have shown evidence of violation of the cosmological-scale isotropy assumption as discussed in Section~\ref{introduction}, the null hypothesis would be that the spin directions of spiral galaxies are distributed randomly. This section discusses several possible reasons that could have led to the observation of asymmetry that does not reflect an asymmetry in the local Universe.
\subsection{Error in the galaxy annotation algorithm}
An error in the annotation algorithm can lead to any form of distribution, depends on the nature of the error. However, multiple indications show that the asymmetry cannot be the result of an error in the classification algorithm. The algorithm is a model-driven symmetric method that is based on clear rules. It is not based on complex data-driven rules used by pattern recognition systems. Such systems are complex and non-intuitive, rely on the data they are trained by and even by the order of the samples in the training set, making it virtually impossible to verify their symmetricity. The algorithm used here is fully symmetric, and follows clear defined rules as discussed in Section~\ref{data}.
Another evidence that the asymmetry is not driven by an error in the annotation algorithm is that the asymmetry changes between different parts of the sky, and inverse between opposite hemispheres. Since each galaxy is analyzed independently, a bias in the annotation algorithm is expected to be consistent throughout the sky, and it is not expected to flip in opposite hemispheres. The downloading of the images and the automatic analysis of the images were all done by the same computer, to avoid unknown differences between computers that can lead to bias or unknown differences in the way galaxy images are analyzed.
Due to the theoretical and empirical evidence that the algorithm is symmetric, an error in the galaxy annotation is expected to impact clockwise and counterclockwise galaxies in a similar manner. If the galaxy annotation algorithm had a certain error in the annotation of the galaxies, the asymmetry {\it A} can be defined by Equation~\ref{asymmetry}.
\begin{equation}
A=\frac{(N_{cw}+E_{cw})-(N_{ccw}+E_{ccw})}{N_{cw}+E_{cw}+N_{ccw}+E_{ccw}},
\label{asymmetry}
\end{equation}
where $E_{cw}$ is the number of galaxies spinning clockwise incorrectly annotated as counterclockwise, and $E_{ccw}$ is the number of galaxies spinning counterclockwise incorrectly annotated as spinning clockwise. Because the algorithm is symmetric, the number of counterclockwise galaxies incorrectly annotated as clockwise is expected to be roughly the same as the number of clockwise galaxies missclassified as counterclockwise, and therefore $E_{cw} \simeq E_{ccw}$ \citep{shamir2021particles}. Therefore, the asymmetry {\it A} can be defined by Equation~\ref{asymmetry2}.
\begin{equation}
A=\frac{N_{cw}-N_{ccw}}{N_{cw}+E_{cw}+N_{ccw}+E_{ccw}}
\label{asymmetry2}
\end{equation}
Since $E_{cw}$ and $E_{ccw}$ cannot be negative, a higher rate of incorrectly annotated galaxies is expected to make {\it A} lower. Therefore, incorrect annotation of galaxies is not expected to lead to asymmetry, and can only make the asymmetry lower rather than higher.
An experiment \citep{shamir2021particles} of intentionally annotating some of the galaxies incorrectly showed that even when an error is added intentionally, the results do not change significantly even when as many as 25\% of the galaxies are assigned with incorrect spin directions, as long as the error is added to both clockwise and counterclockwise galaxies \citep{shamir2021particles}. But if the error is added in an asymmetric manner, even a small asymmetry of 2\% leads to a very strong asymmetry, and a dipole axis that peaks exactly at the celestial pole \citep{shamir2021particles}.
It should be mentioned that in one of the datasets shown here, which is the dataset acquired by HST, the annotation was done manually, and without using any automatic classification. The galaxies imaged by HST were annotated manually, and the results are in agreement with the automatic annotation of galaxies imaged by the other telescopes and annotated automatically.
\subsection{Bias in the sky survey hardware or photometric pipelines}
Autonomous digital sky surveys are some of the more complex research instruments, and involve sophisticated hardware and software to enable the collection, storage, analysis, and accessibility of the data. It is difficult to think of an error in the hardware or software that could lead to asymmetry between the number of galaxies spinning in opposite directions, but due to the complexity of these systems it is also difficult to prove that such error does not exist. That possible error is addressed here by comparing the results using data from several different telescopes. The instruments used in this study are independent from each other, and have different hardware and different photometric pipelines. As it is unlikely to have such bias in one instrument, it is very difficult to assume that all of these different instruments have such bias, and the profile created by that bias is consistent across all of them.
\subsection{Cosmic variance}
The distribution of galaxies in the universe is not completely uniform. These subtle fluctuations in the density of galaxy population can lead to ``cosmic variance'' \citep{driver2010quantifying,moster2011cosmic}, which can impact measurements at a cosmological scale \citep{kamionkowski1997getting,amarena2018impact,keenan2020biases}.
The probe of asymmetry between galaxies spinning in opposite directions is a relative measurement rather than an absolute measurement. That is, the asymmetry is determined by the difference between two measurements made in the same field, and therefore should not be affected by cosmic variance. Any cosmic variance or other effects that impacts the number of clockwise galaxies observed from Earth is expected to have a similar effect on the number of counterclockwise galaxies.
\subsection{Multiple objects at the same galaxy}
In some cases, digital sky surveys can identify several photometric objects as independent galaxies, even in cases they are part of one larger galaxy. In the datasets used here all photometric objects that are part of the same galaxy were removed by removing all objects that had another object within 0.01$^o$.
Even if such objects existed in the datasets, they are expected to be distributed evenly between galaxies that spin clockwise and galaxies that spin counterclockwise, and therefore should not introduce an asymmetry. Experiments by using datasets of galaxies assigned with random spin directions and adding artificial objects to the galaxies showed that adding objects at exactly the same position of the original galaxies does not lead to signal of asymmetry \citep{shamir2021particles}.
The experiments were made by using $\sim7.7\cdot10^4$ SDSS galaxies, and assigning the galaxies with random spin directions. Then, gradually adding more objects with the same location and spin directions as the galaxies in the original dataset, and the new artificial galaxies were assigned with the same spin direction as the galaxies in the original dataset \citep{shamir2021particles}. Adding such artificial galaxies did not lead to statistically significant signal.
\subsection{Differences in inclination}
The axis shown in Section~\ref{dipole} is profiled by computing the most likely axis when limiting the galaxies to different redshift ranges. In each redshift range, the axis is computed by the cosine dependence of all galaxies in the dataset that fit the redshift range, regardless of their location in the sky. By applying simple geometry, the change in the location of the most likely axis in different redshift ranges can be used to deduce the location of an axis that does not necessarily goes directly through Earth.
That analysis can also be affected by differences in the inclination of galaxies. The inclination of the galaxies can impact the ability to identify its spin direction, as a sharper inclination makes it more difficult to identify the spin patterns. As mentioned above, the axis is computed for each redshift range by fitting all galaxies in that redshift range, regardless of their position in the sky, and therefore it is expected that the impact of the inclination would be statistically the same among clockwise and counterclockwise galaxies.
To test that assumption, the inclination of galaxies that spin clockwise was compared to the inclination of galaxies that spin counterclockwise. The inclination of each galaxy was computed by $\cos^{-1}(\frac{short \: axis}{long \: axis})$, and the short and long axis of all galaxies were determined by {\it Ganalyzer} as described in \citep{shamir2011ganalyzer}. Table~\ref{inclination} shows the average inclination of the galaxies in different redshift ranges and different RA ranges.
\begin{table}
\centering
\scriptsize
\begin{tabular}{llll}
\hline
Z & RA & Average & Average \\
range & range & inclination & inclination \\
& & cw (rad) & ccw (rad) \\
\hline
$<0.15$ & all & 1.14372$\pm$0.001 & 1.14426$\pm$0.001 \\
$>0.15$ & all & 1.18472$\pm$0.002 & 1.18529$\pm$0.002 \\
$<0.15$ & $(0^o<\alpha<180^o)$ & 1.14387$\pm$0.002 & 1.14459$\pm$0.002 \\
$<0.15$ & $(180^o<\alpha<360^o)$ & 1.14357$\pm$0.002 & 1.14371$\pm$0.002 \\
$>0.15$ & $(0^o<\alpha<180^o)$ & 1.18488$\pm$0.003 & 1.18564$\pm$0.003 \\
$>0.15$ & $(180^o<\alpha<360^o)$ & 1.18451$\pm$0.002 & 1.18464$\pm$0.002 \\
\hline
\end{tabular}
\caption{Average inclination (in radians) of clockwise galaxies and counterclockwise galaxies in different redshift and RA ranges. The errors are the standard errors of the means.}
\label{inclination}
\end{table}
As expected, the table shows that the difference in inclination between clockwise and counterclockwise galaxies is not statistically significant, and well below the standard error. The inclination in higher redshifts is somewhat higher than the lower redshift, and that difference can be attributed to the smaller size of the galaxies. But also in the higher redshifts, there are no statistically significant differences between the inclination of clockwise galaxies and the inclination of counterclockwise galaxies. As also expected, no differences in the inclination were observed in opposite hemispheres.
\subsection{Photometric redshift}
Some of the analysis shown here is based on the redshift of the galaxies. Obtaining the spectra of a galaxy is a relatively long process, and therefore the vast majority of the galaxies do not have spectra. To estimate the redshift of galaxies that do not have spectra, the redshift can be estimated computationally from the photometric information in an approach called ``photometric redshift". While the photometric redshift is very quick to compute, it is also highly inaccurate, ambigous (meaning that one galaxy can have several different photometric redshifts), and systematically biased.
In this study the asymmetry is of magnitude smaller than 1\%. The error of the state-of-the-art photometric redshift methods is 10\% to 20\%. Since it is normally determined by complex data-driven rules of machine learning systems, it is also systematically bias in a manner that is difficult to quantify and profile. Therefore, the photometric redshift is not a suitable probe that can be used in this study. All redshifts used in this paper are the spectroscopic redshifts, and the photometric redshifts are not used in any part of this study. Some of the analyses are done with no redshift information at all, showing that the signal is not originated from an error or systematic bias in the redshift.
\subsection{Bias carried over from previous catalogs}
\label{morphology_catalogs}
Catalogs of galaxy morphology can be prepared by either manual annotation of the galaxies \citep{land2008galaxy,nair2010catalog,baillard2011efigi}, or by automatic annotation \citep{gravet2015catalog,perez2015morphologies,goddard2020catalog,cheng2021galaxy}. When not prepared specifically for the purpose of analysis of galaxy spin directions, such catalogs can be biased in some way by the process through which the galaxies were annotated. For instance, \cite{land2008galaxy} found very substantial bias in the annotation of galaxies by their spin direction when the annotation was done by anonymous volunteers. Such bias driven by the human perception is difficult to quantify and correct, and even a small but consistent bias can lead to strong signal in the analysis \citep{shamir2021particles}. Such bias can also affect the separation of galaxies into elliptical and spiral galaxies, and therefore using a catalog that was prepared manually could lead to unexpected patterns that might be driven by the human perception rather than the real sky.
Some catalogs of galaxy morphology were prepared by using certain algorithms. Examples of such catalogs include \citep{gravet2015catalog,perez2015morphologies,goddard2020catalog,cheng2021galaxy}. However, these catalogs rely in most cases on machine learning algorithms, which work by complex data-driven rules. Due to their complexity and non-intuitive nature, it is very difficult to verify that these algorithms are fully symmetric. More importantly, these machine learning systems are based on manual annotation of the galaxies, and therefore any bias in the manual annotation would be carried on to the catalog. For instance, \citep{goddard2020catalog,cheng2021galaxy} are two catalogs that made use of crowdsourcing annotations that are known to be biased \citep{land2008galaxy,hayes2017nature}, and therefore are not safe for the task shown in this study, any other task related to an analysis of anisotropy in the large-scale structure. Moreover, algorithms based on deep neural networks use any discriminative information they can find in the classes of images, and therefore also learn the background of the images, leading to unusual and unexpected biases that are difficult to profile, as explained in detail in \cite{dhar2022systematic}.
An important aspect of the experimental design of this study is that no catalog of galaxy morphology was used. The entire process of annotation was done by the model-driven symmetric algorithm explained in Section~\ref{ganalyzer}, and no assumptions are made about the distribution of the galaxies in any existing catalog of galaxy morphology. The only selection of the galaxies is by the object detection algorithms and the magnitude of the objects in HST, DES, DESI Legacy Survey, SDSS, and Pan-STARRS photometric pipelines. It is difficult to think of a bias in these algorithms that would prefer one spin direction over another, and would also inverse that preference in the opposite hemisphere. While it is difficult to think of such bias in one sky survey, it is highly unlikely that such bias would appear in several sky surveys, with consistent profiles such that the profile of the bias in one sky survey matches the profile of the bias in the other surveys.
\subsection{Atmospheric effect}
There is no known atmospheric effect that can make a galaxy that spin clockwise appear as if it spins counterclockwise. Also, because the asymmetry is always measured with galaxies imaged in the same field, any kind of atmospheric effect that affects galaxies the spin clockwise will also affect galaxies that spin counterclockwise. Therefore, it is unlikely that a certain atmospheric effect would impact the number of clockwise galaxies at a certain field, but would have different impact on galaxies spinning counterclockwise. In any case, one of the datasets used here is made of galaxies imaged by the space-based HST, and are therefore not subjected to any kind of atmospheric effect.
\subsection{Backward spiral galaxies}
In rare cases, the shape of the arms of a spiral galaxy is not an indication of the spin direction of the galaxy. An example is NGC 4622 \citep{freeman1991simulating,buta2003ringed}. A prevalent and systematically uneven distribution of backward spiral galaxies might indeed lead to asymmetry between the number of galaxies spinning clockwise and the number of galaxies spinning counterclockwise. For instance, if a relatively high percentage of galaxies that actually spin clockwise are backward spiral galaxies, it would have led to an excessive number of galaxies that seem to be spinning counterclockwise.
However, backward spiral galaxies are relatively rare. Also, these galaxies are expected to be distributed equally between galaxies that spin clockwise and galaxies that spin counterclockwise, and there is no indication of asymmetry between backwards spiral galaxies. Therefore, according to the known evidence, there is no reason to assume that the observations shown here are driven by backward spiral galaxies.
\section{Conclusion}
\label{conclusion}
Autonomous digital sky surveys powered by robotic telescopes have allowed the collection of unprecedented amounts of astronomical data, enabling to address research questions that their studying was not feasible in the pre-information era. For instance, very large structures such as the Great Wall of Sloan were not discovered until robotic telescopes that can collect very large databases were introduced.
The question addressed here is the large-scale distribution of the spin directions of spiral galaxies. Multiple previous experiments have shown that the distribution of spin directions of spiral galaxies as observed from Earth might not be random \citep{longo2011detection,shamir2012handedness,shamir2013color,shamir2016asymmetry,shamir2017large,shamir2017photometric,shamir2017colour,shamir2019large,shamir2020pasa,shamir2020patterns,lee2019galaxy,lee2019mysterious,shamir2021particles,shamir2021large}.
This study uses a large dataset of galaxies with spectra from several different telescopes. The analysis shows that the asymmetry in the spin directions of spiral galaxies increases with the redshift. The peak of the most likely axis changes consistently with the redshift, which can imply on an axis that does not go directly through Earth. The analysis uses several telescopes, covering the both the Northern and Southern hemispheres. The findings are in agreement with previous results, including space-based data acquired by the Hubble Space Telescope. Another noted observation is that the asymmetry becomes stronger as the redshift gets higher. Although the redshift range of the galaxies used in this study is naturally limited by the imaging capabilities of the telescopes, that correlation can be interpreted as higher asymmetry in the earlier Universe. If that trend is consistent also in higher redshifts not observed in this study, it can be viewed as an indication that the asymmetry patterns are primordial, and were stronger in the young Universe, but gradually becomes weaker as the Universe gets older. That can be explained by gravitational interaction between galaxies and galaxy mergers that can change the spin direction of the galaxies. Future sky surveys such as the Vera Rubin Observatory will have far greater depth, and will allow to test whether the trend continues also in higher redshifts.
Studies with smaller datasets of galaxies showed non-random spin directions of galaxies in filaments of the cosmic web, as described in \citep{tempel2013evidence,tempel2013galaxy,tempel2014detecting,dubois2014dancing,kraljic2021sdss} and others. Other studies showed alignment in the spin directions even when the galaxies are too far from each other to interact gravitationally \citep{lee2019galaxy,lee2019mysterious}, unless assuming modified Newtonian dynamics (MOND) gravity models that explain longer gravitational span \citep{sanders2003missing,darabi2014liquid,amendola2020measuring}.
Other observations of large-scale alignment in spin directions focused on quasars such as \citep{hutsemekers1998evidence,hutsemekers2005mapping,agarwal2011alignments,hutsemekers2014alignment}. Position angle of radio galaxies also showed large-scale consistency of angular momentum \citep{taylor2016alignments}. These observations agree with observations made with datasets such as the Faint Images of the Radio Sky at Twenty-centimetres (FIRST) and the TIFR GMRT Sky Survey (TGSS), showing large-scale alignment of radio galaxies \citep{contigiani2017radio,panwar2020alignment}.
These observational studies are also supported by simulations of dark matter \citep{aragon2007spin,zhang2009spin,codis2012connecting,libeskind2013velocity,libeskind2014universal,ganeshaiah2018cosmic,kraljic2020and} and galaxies \citep{dubois2014dancing,codis2018galaxy,ganeshaiah2019cosmic,kraljic2020and}, showing links between spin directions and the large-scale structure. That correlation was associated with the stellar mass and color of the galaxies \citep{wang2018spin}, and it has been proposed that the association was also linked to halo formation \citep{wang2017general}. That led to the contention that the spin direction in the halo progenitors is linked to the large-scale structure of the Universe \citep{wang2018build}. It should be mentioned that the spin direction of a galaxy might not be necessarily the same as the spin direction of the dark matter halo, as it has been proposed that in some cases a galaxy might spin in a different direction than its host dark matter halo \citep{wang2018spin}.
The analysis of spin directions done in this study provides evidence of large-scale dipole and quadrupole alignment. The observation of a large-scale axis has been proposed in the past by analyzing the cosmic microwave background (CMB), with consistent data from the Cosmic Background Explorer (COBE), Wilkinson Microwave Anisotropy Probe (WMAP) and Planck, as described in \citep{abramo2006anomalies,mariano2013cmb,land2005examination,ade2014planck,santos2015influence,gruppuso2018evens} and other studies. Observations also showed that the axis formed by the CMB temperature is aligned with other cosmic asymmetry axes such as dark energy and dark flow \citep{mariano2013cmb}. Other notable statistical anomalies in the CMB are the quandrupole-octopole alignment \citep{schwarz2004low,ralston2004virgo,copi2007uncorrelated,copi2010large,copi2015large}, the asymmetry between hemispheres \citep{eriksen2004asymmetries,land2005examination,akrami2014power}, point-parity asymmetry \citep{kim2010anomalous,kim2010anomalous2}, and the CMB Cold Spot. If these anomalies are not statistical fluctuations \citep{bennett2011seven}, they can be viewed as observations that disagree with $\Lambda$CDM, as proposed by \cite{bull2016beyond,yeung2022directional} and others.
As described in Section~\ref{introduction}, the possible axis observed in the CMB is aligned with theories related to the geometry of the Universe such as ellipsoidal universe \citep{campanelli2006ellipsoidal,campanelli2007cosmic,gruppuso2007complete,rodrigues2008anisotropic,campanelli2011cosmic,cea2014ellipsoidal}, rotating universe \citep{gamow1946rotating,godel1949example,ozsvath1962finite,ozsvath2001approaches,su2009universe,sivaram2012primordial,chechin2016rotation,chechin2017does,camp2021}, holographic big bang \citep{pourhasan2014out,altamirano2017cosmological}, negative gravitational mass \citep{le2021recent}, and Black hole cosmology \citep{pathria1972universe,stuckey1994observable,easson2001universe,seshavatharam2010physics,poplawski2010radial,poplawski2010cosmology,chakrabarty2020toy,seshavatharam2020light,rinaldi2022matrix}
The availability of robotic telescopes provides the ability to analyze a possible non-random distribution of the spin directions of spiral galaxies, and that research question was not approachable in the pre-information era. As evidence for such non-random distribution are accumulating, additional research will be needed to fully understand its nature, and match it with other messengers in addition to CMB.
\section*{Acknowledgments}
I would like to thank the anonymous reviewer for the insightful comments. This study was supported in part by NSF grants AST-1903823 and IIS-1546079. I would like to thank Ethan Nguyen for retrieving and organizing the image data from the DESI Legacy Survey.
This project used public archival data from the Dark Energy Survey (DES). Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology FacilitiesCouncil of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, the Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A\&M University, Financiadora de Estudos e Projetos, Funda{\c c}{\~a}o Carlos Chagas Filho de Amparo {\`a} Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cient{\'i}fico e Tecnol{\'o}gico and the Minist{\'e}rio da Ci{\^e}ncia, Tecnologia e Inova{\c c}{\~a}o, the Deutsche Forschungsgemeinschaft, and the Collaborating Institutions in the Dark Energy Survey.
The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energ{\'e}ticas, Medioambientales y Tecnol{\'o}gicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgen{\"o}ssische Technische Hochschule (ETH) Z{\"u}rich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ci{\`e}ncies de l'Espai (IEEC/CSIC), the Institut de F{\'i}sica d'Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig-Maximilians Universit{\"a}t M{\"u}nchen and the associated Excellence Cluster Universe, the University of Michigan, the National Optical Astronomy Observatory, the University of Nottingham, The Ohio State University, the OzDES Membership Consortium, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A\&M University.
Based in part on observations at Cerro Tololo Inter-American Observatory, National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
The research is based on observations made with the NASA/ESA Hubble Space Telescope, and obtained from the Hubble Legacy Archive, which is a collaboration between the Space Telescope Science Institute (STScI/NASA), the Space Telescope European Coordinating Facility (ST-ECF/ESA) and the Canadian Astronomy Data Centre (CADC/NRC/CSA).
The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; Proposal ID \#2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Prop. ID \#2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the Mayall z-band Legacy Survey (MzLS; Prop. ID \#2016A-0453; PI: Arjun Dey). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF’s NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. The Legacy Surveys project is honored to be permitted to conduct astronomical research on Iolkam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation.
NOIRLab is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A\&M University, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo a Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cientifico e Tecnologico and the Ministerio da Ciencia, Tecnologia e Inovacao, the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energeticas, Medioambientales y Tecnologicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgenossische Technische Hochschule (ETH) Zurich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ciencies de l’Espai (IEEC/CSIC), the Institut de Fisica d’Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig Maximilians Universitat Munchen and the associated Excellence Cluster Universe, the University of Michigan, NSF’s NOIRLab, the University of Nottingham, the Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A\&M University.
BASS is a key project of the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories of China, the Chinese Academy of Sciences (the Strategic Priority Research Program “The Emergence of Cosmological Structures” Grant \# XDB09000000), and the Special Fund for Astronomy from the Ministry of Finance. The BASS is also supported by the External Cooperation Program of Chinese Academy of Sciences (Grant \# 114A11KYSB20160057), and Chinese National Natural Science Foundation (Grant \# 11433005).
The Legacy Survey team makes use of data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), which is a project of the Jet Propulsion Laboratory/California Institute of Technology. NEOWISE is funded by the National Aeronautics and Space Administration.
The Legacy Surveys imaging of the DESI footprint is supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH1123, by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract; and by the U.S. National Science Foundation, Division of Astronomical Sciences under Contract No. AST-0950945 to NOAO.
SDSS-IV is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science, Carnegie Mellon University, the Chilean Participation Group, the French Participation Group, Harvard-Smithsonian Center for Astrophysics, Instituto de Astrofisica de Canarias, The Johns Hopkins University, Kavli Institute for the Physics and Mathematics of the Universe (IPMU) / University of Tokyo, the Korean Participation Group, Lawrence Berkeley National Laboratory, Leibniz Institut fur Astrophysik Potsdam (AIP), Max-Planck-Institut fur Astronomie (MPIA Heidelberg), Max-Planck-Institut fur Astrophysik (MPA Garching), Max-Planck-Institut fur Extraterrestrische Physik (MPE), National Astronomical Observatories of China, New Mexico State University, New York University, University of Notre Dame, Observatario Nacional / MCTI, The Ohio State University, Pennsylvania State University, Shanghai Astronomical Observatory, United Kingdom Participation Group, Universidad Nacional Autonoma de Mexico, University of Arizona, University of Colorado Boulder, University of Oxford, University of Portsmouth, University of Utah, University of Virginia, University of Washington, University of Wisconsin, Vanderbilt University, and Yale University.
The Pan-STARRS1 Surveys (PS1) and the PS1 public science archive have been made possible through contributions by the Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project Office, the Max-Planck Society and its participating institutes, the Max Planck Institute for Astronomy, Heidelberg and the Max Planck Institute for Extraterrestrial Physics, Garching, The Johns Hopkins University, Durham University, the University of Edinburgh, the Queen's University Belfast, the Harvard-Smithsonian Center for Astrophysics, the Las Cumbres Observatory Global Telescope Network Incorporated, the National Central University of Taiwan, the Space Telescope Science Institute, the National Aeronautics and Space Administration under Grant No. NNX08AR22G issued through the Planetary Science Division of the NASA Science Mission Directorate, the National Science Foundation Grant No. AST-1238877, the University of Maryland, Eotvos Lorand University (ELTE), the Los Alamos National Laboratory, and the Gordon and Betty Moore Foundation.
\bibliographystyle{apalike}
|
\section{Introduction}
Federated Learning(FL) has been widely studied recently to train machine learning models without directly accessing user's data. Despite being successful in achieving high utility performance compared to centralized training, huge communication costs induced by current FL algorithms like FedAvg\citep{mcmahan2017communication} prevents using federated data to train large scale models. Specifically, communicating the entire model between each client and the server could drastically slower the training process, and imposes communication costs on the users. Many prior efforts have focused on sparsifying the model to reduce communication cost, including masking the model updates, low precision training with quantization, distillation, etc. However, sparsification based methods usually suffers from communication cost-utility trade-off: high compression rate could hurt model quality.
To overcome such limitation, we propose a different way to think about model compression in FL. Instead of a model update, which is the same size as the original model, we propose sending a batch of carefully optimized synthetic data, which is significantly smaller in size. Each client crafts a set of synthetic data such that the model updated by the synthetic data performs well on the client's original training data. In this way, we could use a dataset that is significantly smaller than the client training set to obtain a similar model update as if we use the original data. Each client then send the synthetic data to the server. Upon receiving the synthetic data from each client, the server will use it to recover the model updated by synthetic data.
Having this intuition in mind, we formally propose a new objective in federated learning for local clients at each communication round. Using this formulation, we propose an effective solver that could adapt to a wide family of existing federated learning algorithm. Specifically, we develop an algorithm that learns synthetic data for local clients under the FedAvg\citep{mcmahan2017communication} framework. We empirically evaluate and compare our method with prior works, demonstrating advantage of transmitting synthetic data as an effective compression technique.
\section{Background and Related Work}
\paragraph{Compression in FL} Model compression techniques has been widely studied in machine learning community with a centralized dataset. Some widely used methods include gradient quantization\citep{alistarh2017qsgd}, gradient ternarization\citep{wen2017terngrad}, using the sign of gradient\citep{bernstein2018signsgd}, pruning based methods like masking\citep{beznosikov2020biased}, etc. In federated learning, compression could happen at two places: transmitting model updates from local client to the server (upstream); transmitting updated global model from the server to local client (downstream). Recent works have proposed using sparsification techniques for both upstream and downstream communication to reduce the cost of large scale federated learning \citep{reisizadeh2020fedpaq, haddadpour2021federated, sattler2020STC}. Different from our work, these works focused on communicating a sparse model between server and client, where the model performance could significantly degrade given the same number of training steps. \citet{goetz2020federated} proposed a similar scheme to transmit synthetic data via upstream communication. However, they propose optimizing synthetic data to minimize the distance between model updated by synthetic data and model updated by training data. Our proposed objective directly optimizes the synthetic data so that the resulting model achieves good performance on the true training data.
\paragraph{Dataset Distillation} A motivation of our work is to learn a small set of synthetic data that could perform equally well on a given model compared to real data used to train the model. \cite{wang2018dataset} proposed a dataset distillation algorithm that optimizes synthetic data $w_{syn}$ such that the model learned using $w_{syn}$ as the training data approximates the model learned using the true training data. Although similar to our approach, they only considered distillation from a centralized dataset at one time while in our case we learn the objective locally for each client at \textit{every} communication round. Our proposed method is also more general in the method to update the model using synthetic data (See Section \ref{sec:method:alg}) rather than restricted to SGD.
\vspace{-0.1in}
\section{Communication via Synthetic Data}
\subsection{Formulation}
Traditional Federated Learning(FL) aims at solving the following objective:
\begin{equation}
\min_w \sum_{k=1}^K p_k F_k(w)
\end{equation}
where $F_k(w)$ is the local objective for client $k$, $p_k$ are pre-defined weights such that $\sum_kp_k=1$. At each communication round, the central server selects a subset of clients and send the current model to the them. Each client then separately optimizes its local objective iteratively using stochastic gradients. Then the server collects and aggregates the model updates from every client to obtain the new global model. Note that model updates have the same size as the actual global model, which means if a large-scale model is used as the model, the client would need to send a model as large as the global model. Under our proposed method, instead of sending the model updates, each client now sends batches of synthetic data generated locally to the server. The server will then utilize the synthetic data to recover the local model updates. We formalize the optimization process of synthetic data as the following.
For client $k$, let $D_k^{tr}=(X_k, Y_k)$ be the training data and $w_k^t$ the local copy of global model at the $t$-th communication round. In traditional FL, at every communication round, client $k$ tries to solve
\begin{equation}
\label{eq:fl_traditional_local}
\min_w F_k(D^{tr}_k;w)
\end{equation}
using $w_k^t$ is as the initialization of $w$. Note that a lot of existing federated learning methods rely on using iterative gradient methods to solve for Equation \ref{eq:fl_traditional_local}. Let's define the update process for any client $k$ at communication round $t$ to be $\texttt{ClientUpdate}^k(\cdot;w_k^t)$. Thus, in the traditional FL setting, local optimization process could be written as $\texttt{ClientUpdate}^k(D_k^{tr};w_k^t)$. The goal is to find a set of synthetic data $D_k^{syn}=\{x_k^i,y_k^i\}_{i=1,\cdots,m}$ that is significantly smaller in memory size than $w$, such that $\texttt{ClientUpdate}^k(D_k^{syn};w_k^t)$ is similar to optimizing $\texttt{ClientUpdate}^k(D_k^{tr};w_k^t)$. At the end of that communication round, client $k$ will send $D^{syn}_k$ to the server and the server could recover $w^{syn}_k = \texttt{ClientUpdate}^k(D_k^{syn};w_k^t)$ and utilizes $w^{syn}_k$ as client $k$'s updated model for aggregation.
Now the problem becomes how can we find $D^{syn}_k$ that distills the knowledge from $D^{tr}_k$. The most direct way to do so is to minimize a distance metric between the model generated from synthetic data and the model generated from true train data. However, note that our purpose is that using $D^{syn}_k$, we could obtain an updated model $w^{syn}_k$ such that $F_k(D^{tr}_k;w^{syn}_k)$ is as good as $F_k(D^{tr}_k;w^{tr}_k)$. With certain purpose in mind, we propose the following objective:
\begin{equation}
\label{eq:fl_synthetic_local_exact}
\min_{D^{syn}_k} F_k\left(D^{tr}_k;\argmin_w F_k\left(D^{syn}_k;w\right)\right)
\end{equation}
Note that when Equation \ref{eq:fl_traditional_local} is convex in $w$ and Equation \ref{eq:fl_synthetic_local_exact} is convex in $D^{syn}_k$, given the same $D^{tr}_k$, both equations are essentially finding the same optimal local model. However, when there doesn't exist a closed form solution for $\argmin_w F_k\left(D^{syn}_k;w\right)$, the inner optimization problem for Equation \ref{eq:fl_synthetic_local_exact} could not be solved exactly with finite number of steps at every communication round. To find an approximate solution, most existing FL methods utilize gradient based methods like SGD. Therefore, we propose to optimize the following objective in practice instead of Equation \ref{eq:fl_synthetic_local_exact}:
\begin{equation}
\label{eq:fl_synthetic_local}
\min_{D^{syn}_k} F_k\left(D^{tr}_k;\texttt{ClientUpdate}_k\left(D_k^{syn};w_k^t\right)\right)
\end{equation}
Without loss of generality, $\texttt{ClientUpdate}$ could be any local optimization methods including GD, SGD, etc.
\subsection{Algorithm}
\label{sec:method:alg}
We summarize our algorithm for federated learning via synthetic data in Algorithm \ref{alg:1}. Our algorithm is based off of FedAvg\citep{mcmahan2017communication}, a communication efficient method widely used in federated learning. At each communication round, instead of performing SGD on the local training data, each selected client $k$ first initializes a synthetic dataset $D^{syn}_k$ (line 5). To find the best $w^{syn}_k$, synthetic updated model generated by $D^{syn}_k$ (line 7) that minimizes the its loss on the original training data $D^{tr}_k$, we propose to apply gradient descent on $D^{syn}_k$ for multiple iterations (line 8). After that, client $k$ would send $D^{syn}_k$, an entity that requires significantly less storage compared to the model weight, back to the server. To recover client $k$'s learned model, the server updates $w_k^t$ with $D^{syn}_k$ using the same process client $k$ generated $w^{syn}_k$ (line 12). We also provide an example of the \texttt{ClientUpdate} method: running SGD on $w_k^t$ using the $D^{syn}_k$ (line 16-19). This is consistent to the local update rule in FedAvg, where client applies SGD to update $w_k^t$ using its local training data. In order to fully utilize the advantage of using synthetic data to distill the information from the original training data, we also propose the following techniques while learning $D^{syn}_k$.
\textbf{Multiple batches of synthetic data} At every communication round, FedAvg allows a selected client $k$ to split its local training data into multiple batches and perform minibatch-SGD for multiple epochs. Motivated by this, we allow client $k$ to create multiple batches of synthetic data. Instead of running one step gradient descent on the entire synthetic data, client $k$ updates $w_k^t$ sequentially using different batches of synthetic data, as specified in Line 18 of Algorithm \ref{alg:1}.
\textbf{Trainable label} In a traditional supervised classification task, the data usually has fixed label $y$. When using cross entropy as the loss function, fixed $y$ is encoded as an one-hot vector in $\mathbb{R}^{|C|}$ where $C$ is the set of all labels. However, this is not necessary for synthetic data. The purpose of using synthetic data is only to generate a model that performs well on the real training data. Restricting any synthetic $x_i$ to have a fixed label $y_i$ is too stringent and limit the search space for pairs of $(x_i,y_i)$ to learn the information of the original training data. Hence, we propose randomly initializing $y_i\sim Uniform(0,1)^{|C|}$. While updating synthetic data $(x_i,y_i)$, we calculate
\begin{align}
x_i &\leftarrow x_i - \eta_x \nabla_{x_i}F_k(D^{tr}_k; w^{syn}_k)\\
y_i &\leftarrow y_i - \eta_y \nabla_{y_i}F_k(D^{tr}_k; w^{syn}_k)
\end{align}
It is worth noting that under certain scenario, we do not limit $y_i$ to be a vector representing the probability that $x_i$ belongs to a certain class. Each entry for $y_i$ could be arbitrary real numbers so that we could search in the entire $\mathbb{R}^{|C|}$ to find a good local minima for $y_i$.
\setlength{\textfloatsep}{20pt}
\begin{algorithm}[t]
\setlength{\abovedisplayskip}{0pt}
\setlength{\belowdisplayskip}{0pt}
\setlength{\abovedisplayshortskip}{0pt}
\setlength{\belowdisplayshortskip}{0pt}
\caption{FedSynth}
\label{alg:1}
\begin{algorithmic}[1]
\STATE {\bf Input:} $T$, $E$, $\eta$, $\eta_w$, $w^0$, $\{D^{tr}_k\}_{k=1,\cdots,K}$
\FOR {$t=0, \cdots, T-1$}
\STATE Server selects a subset of clients $S_t$ and broadcasts $w^t$ to $S_t$.
\FOR {all $k\in S_t$ in parallel}
\STATE Client $k$ initializes $w_k^t=w^t$ and $m$ batches of synthetic data $D^{syn}_k = \{x_i,y_i\}_{i=1,\cdots,m}$.
\FOR {$j=0,1,\cdots,E$}
\STATE Client $k$ obtains the model updated by $D^{syn}_k$
\[w^{syn}_k = \texttt{ClientUpdate}(D^{syn}_k; w_k^t)\]
\STATE Client $k$ updates $D^{syn}_k$ by
\[D^{syn}_k \leftarrow D^{syn}_k - \eta\nabla_{D^{syn}_k}F_k(D^{tr}_k; w^{syn}_k)\]
\ENDFOR
\STATE Client $k$ sends $D^{syn}_k$ back to the server.
\ENDFOR
\STATE Server recovers $\widehat{w}^{syn}_k = \texttt{ClientUpdate}(D^{syn}_k, w_k^t)$ for every $k$.
\STATE Server aggregates the weight
\[
w^{t+1} = w^{t} + \frac{1}{|S_t|}\sum_{k\in S_t} (\widehat{w}^{syn}_k-w^{t})
\]
\ENDFOR
\RETURN $w^T$
\vspace{-.05in}
\\\hrulefill
\vspace{.1em}
\STATE \texttt{ClientUpdate}($\{x_i,y_i\}_{i=1,2,\cdots,m};w$)
\FOR {$j = 1, \cdots, m$}
\STATE Client performs minibatch-SGD locally
\[ w \leftarrow w-\eta_w\nabla_{w}F_k((x_i,y_i);w)\]
\ENDFOR
\end{algorithmic}
\end{algorithm}
\section{Experiments}
In this section we empirically evaluate our Algorithm \ref{alg:1} on common large scale federated learning benchmarks. We first demonstrate that given the same compression rate our method could achieve higher test accuracy then random masking, a popular compression method used in federated learning. We also show how number of batches of synthetic data($m$) and the size of each batch could affect the resulting model's performance.
\subsection{Experimental setup details}
For fair comparison, all methods are trained for the same amount of communication rounds for each dataset. For baseline method(Random Masking), we only apply the compression technique during the upstream communication (i.e. only compress the model sent from client to the server) in order to be consistent to our method. For a fixed compression rate, we apply grid search to tune the hyperparameters($E$, $\eta$, $\eta_w$) for our method on the validation data and report the test accuracy corresponding to the best validation accuracy. We similarly finetune the hyperparameters for random masking and baseline FedAvg as well. For all our experiments, we evaluate the test accuracy and compression rate for a fixed number of communication rounds $T$. All experiments are performed on common federated learning benchmark datasets. Data for FEMNIST and Reddit are naturally partitioned among all the users.
\subsection{Comparison between FedSynth and Random Masking}
We evaluated our method on three commonly used federated benchmark datasets: FEMNIST, MNIST, and Reddit \citep{caldas2018leaf}. The results are shown in Table \ref{table:synthetic_vs_baselines}. Under all three datasets, our method achieves comparable/better performance then baseline random masking methods. Specifically, under a low compression rate, there is an advantage of using our method over random masking in all three datasets. Under Reddit next word prediction task where the utility performance is extremely sensitive to masking, our method that utilizes synthetic data without trainable $y$ achieves higher test accuracy than prior works under all compression rates we experimented. It is also worth noting that our method with trainable $y$ does not always outperform synthetic data with a fixed label, as shown in the FEMNIST experiment.
\setlength{\tabcolsep}{5pt}
\begin{table*}[t!]
\caption{\small Comparing our method with previous compression baselines. The best performance under each given compression rate is \textbf{highlighted}.}
\vspace{1em}
\centering
\label{table:synthetic_vs_baselines}
\scalebox{0.9}{
\begin{tabular}{l ccccc}
\toprule[\heavyrulewidth]
\textbf{FEMNIST} & & FedAvg & Random Masking & FedSynth(Ours) & FedSynth w/ Trainable $y$ (Ours)\\
\midrule
1x & & 69.29 & 69.29 & 69.29 & 69.29\\
5.8x & & - & 68.21 & \textbf{68.63} & 46.67\\
11.6x & & - & \textbf{67.34} & 63.27& 39.98\\
\hline
\hline
\textbf{MNIST} & & FedAvg & Random Masking & FedSynth(Ours) & FedSynth w/ Trainable $y$ (Ours)\\
\midrule
1x & & 97.74 & 97.74 & 97.74 & 97.74\\
7.8x & & - & 97.08 & 95.28 & \textbf{97.25}\\
15.6x & & - & \textbf{96.94} & 93.68 & 96.62\\
\hline
\hline
\textbf{Reddit} & & FedAvg & Random Masking & FedSynth(Ours) & FedSynth w/ Trainable $y$ (Ours)\\
\midrule
1x & & 14.19 & 14.19 & 14.19 & 14.19\\
1.3x & & - & 8.20 & \textbf{8.86}& -\\
2.6x & & - & 4.87 & \textbf{4.89} & -\\
\bottomrule[\heavyrulewidth]
\end{tabular}}
\vspace{-0.1in}
\end{table*}
\subsection{Number of synthetic batches vs. batch size}
As we mentioned in Section 3, each client could split their synthetic data into multiple batches. In Table \ref{table:num_batches_vs_batch_size}, we demonstrate how different number of synthetic batches and batch size could influence the model performance. Given the same number of data points, having more small batches significantly outperforms having few large batches. On one extreme where we treat the entire synthetic dataset as a large batch, i.e. $w_k^t$ is only updated once to get $w^{syn}_k$, model trained using our methods is barely useful. However, on the other extreme where every single piece of data is treated as a separate batch, our method is able to achieve significantly better performance. We would also like to highlight that the more synthetic data we use, the better model performance we could obtain, given the same number of synthetic batches used for getting $w^{syn}_k$.
\begin{table*}[t!]
\caption{\small The effect of number of synthetic batches and batch size on the test accuracy of FedSynth.}
\vspace{1em}
\centering
\label{table:num_batches_vs_batch_size}
\scalebox{0.9}{
\begin{tabular}{l cccc}
\toprule[\heavyrulewidth]
\multicolumn{2}{c}{\textbf{FEMNIST}} & & FedSynth & FedSynth w/ Trainable $y$\\
Synthetic Batches & Batch size & & & \\
\midrule
1 & 10 & & 11.87 & 10.93\\
5 & 5 & & 64.21 & 29.43\\
10 & 5 & & 67.92 & 46.67\\
10 & 2 & & 60.19 & 39.72\\
20 & 2 & & 68.63 & 42.32\\
10 & 1 & & 56.47 & 26.29\\
20 & 1 & & 62.37 & 39.98\\
\bottomrule[\heavyrulewidth]
\end{tabular}}
\vspace{-0.1in}
\end{table*}
\section{Conclusion and Future works}
In this work, we propose a new objective for communication efficient federated learning along with a practical algorithm to solve it. We showed empirically that our methods outperforms baseline methods at low compression level in all three datasets we evaluated. In future works, we aim at making the algorithm more scalable so that learning of synthetic data would require less iterations. We also want to look at sending differentially private synthetic data to protect the local data from potential privacy leakage.
\newpage
{\small
\bibliographystyle{abbrvnat}
|
\section{\label{Intro}Introduction}
Magnetic materials are highly crucial to the civilization with a long history. Entering the quantum era, people are keen on understanding its micro-mechanism while many are still left unsolved and the physics of cubic anisotropy~\cite{PhysRev.52.1178} is among them. The simplest model for it includes a quartic term in the Hamiltonian whose easy axes are determined by the sign of its strength~\cite{IntroMag}. With a small magnetic field, it has been studied by Landau paradigm~\cite{Sov.Phys.Usp.26} and renormalization group (RG) theory~\cite{PhysRevLett.37.565} for its nature of phase transition. However, as the field strength goes larger, the above-mentioned calculations fail to obtain the correct physics.
On the other hand, spins, as a quantum object, are often influenced by each other on nearby sites within materials through the Heisenberg superexchange interaction~\cite{heisenberg}, leading to possible explanations for peculiar phenomena such as the high-$T_c$ superconductivity and its related properties~\cite{PhysRevLett.92.227002, PhysRevB.78.134530, PhysRevB.84.041108, PhysRevLett.113.046402, scirep.6.18675, NewJPhys.19.013028, scirep.9.1719}. Thus, interplay between the superexchange and the cubic anisotropy terms may present an interesting phenomena while the brute-force quantum Monte Carlo (QMC) was not permissible because of the negative-sign problem or the local trapping due to the anisotropic term. This problem was studied many years ago but only within the mean-field analysis~\cite{J.Magn.Magn.Mater.42, Phys.StatusSolidiB129, Phys.StatusSolidiB135}. With $N$-component $\phi^4$ theory and using RG, previous studies revealed a critical $N_c$ where the cubic fixed point is stable when $N>N_c$ while the $O(N)$ symmetric fixed point is driven in the reverse scenario~\cite{PhysRevB.56.14428, PhysRevB.61.15136}, and the nature of its phase transition was also discussed~\cite{PhysRevB.59.4176}. For the real materials, recently, two-dimensional layered magnetic materials have drawn researchers' attention because of its broad range of potential practical applications~\cite{nrevphys1}.
Among them, the ferromagnetic semiconductor thin film is of special interest because of its promising features from industrial point of view \cite{nat.mat.9.965}.
Various magnetic anisotropies can be also found in these materials~\cite{RevModPhys.86.187}. It is important to notice that in many related works upon such magnetic materials, the determination of the easy-axis alignment plays an important role in their studies, in order to explain the experimental observation. However, the introduction of quantum entanglement and its possible influence are seldom discussed because most studies follow the mean-field paradigm. Moreover, despite a great amount of effort for its theoretical understanding through Zener's or other phenomenological models \cite{RevModPhys.86.187}, the microscopic
picture is often overlooked. This provides us a strong motivation to look into this problem.
More than 30 years ago, the quantum ferromagnet with cubic anisotropy under an external field was investigated by Domański and Sznajd within the mean-field scope and a peculiar phase transition between polarized ($P$) and canted ($C$) magnetic phases was proposed~\cite{DOMANSKI1988306}. The difference between these two phases lies in the fact that $C$ phase possesses magnetic orders perpendicular to the direction of field while $P$ does not. Again, its full understanding beyond the mean-field scope has not yet been attained, despite some previous works using perturbation theory and density-matrix renormalization group (DMRG) showing the existence of quadrupolar phase with easy axes aligning in $\langle 100 \rangle$~\cite{Eur.Phys.J.B.5, PhysRevB.59.13764, Eur.Phys.J.B.17}. Especially, whether the easy-axis alignment can be altered by quantum effect also stays unknown. In most of the cases, mean-field analysis can already provide a quite accurate understanding, which is qualitatively correct. However, when different states possess competing energies under certain scenarios, which is the case in our study, introduction of quantum entanglement might lead to different results. One good example comes in the studies upon $t-J$ model, which is a strong candidate in explaining the high-$T_c$ cuprate superconductivity~\cite{Phys.Scr.2002}. Although mean-field analysis always favors the $d$-wave superconducting state as the ground state while other states with instabilities sharing competing energies~\cite{scirep.6.18675}, more elaborated numerical tools demonstrate different outcomes~\cite{doi:10.1126/science.aak9546}. Therefore, in our study after the mean-field analysis, we also adopt the two-dimensional tensor network ansatz aiming to provide a more thorough understanding.
In the following content, we will first introduce our model Hamiltonian in Section~\ref{model} with some preliminary analysis. We start from studying the one-site Hamiltonian when $J=0$ in Section \ref{onesite} and conduct the mean-field analysis in Section~\ref{MF}. We then adopt iPEPS in Section~\ref{ipeps} for a better numerical calculation and discuss the results there. In Section~\ref{discussion} we tackle our Hamiltonian through the perturbation theory to the leading order and realize that it is equivalent to a HCB model, leading to the competing scenario among different $C$ states. Finally, we will argue the possibly practical usage of such field-induced softening of magnetic easy axes and summarize this work in Section~\ref{summary}.
\section{\label{model}Model Hamiltonian}
In this work, we study the quantum ferromagnet in the presence of the related lowest order single-ion term of the cubic anisotropy~\cite{PhysRevB.8.4270} and magnetic field in the square lattice, of which the Hamiltonian reads
\begin{equation}
\begin{aligned}
H=&-J\sum_{\langle i,j \rangle}\vec{S}_i \cdot \vec{S}_j+K\sum_{i,\alpha}(S^\alpha_i)^4 -h\sum_{i}S^z_i,
\end{aligned}
\label{Hamil}
\end{equation}
with $S=2$ and $\alpha=x,y,z$. Spins on nearest-neighbor sites ($\langle i,j\rangle$) are connected through a ferromagnetic Heisenberg term ($J>0$). The second term reflects the cubic anisotropy and with $K>0$, the easy axes go along $\langle 111 \rangle$
Finally, a Zeeman term along $z$-axis is also considered when a magnetic field is present. Equation~(\ref{Hamil})
does not possess the $U(1)$ symmetry and its symmetric group is reduced to a 8-order group. This can be seen by replacing $(\pm S^\beta,\pm S^{\bar{\beta}})$ with $\beta,\bar{\beta} =x,y$ and $\beta\neq\bar{\beta}$ into Eq.~(\ref{Hamil}) while leaving itself
unchanged. In abstract algebra, this group is referred to as the $D_8$ group~\cite{doi:10.1063/1.4887660}. Such symmetry is isomorphic to $Z_4\rtimes Z_2$, where $\rtimes$ refers to the semidirect product because $Z_4$ is a normal subgroup to $D_8$ while $Z_2$ is not.
As $C$ phase appears, $D_8$ symmetry is spontaneously broken due to the formation of planar magnetization in the $x$-$y$ plane. Among the possible orders,
we can have two magnetic orientations that still preserve one of the reflection symmetries of $D_8$. One is when the magnetization aligns parallel to one of the 2$\sigma_d$ (the mirror planes parallel to main diagonal axes, i.e., $\theta=\pi/4,3\pi/4,5\pi/4$, and $7\pi/4$ in Fig.~\ref{fig1}(a)). Another one is when the magnetization lies in one of the 2$\sigma_v$ (the mirror planes parallel to cubic axes, i.e., $\theta=0,\pi/2,\pi$, and $3\pi/2$ in Fig.~\ref{fig1}(a)). These two scenarios, although breaking the $D_8$ symmetry, still preserve $\textbf{C}_s$ reflection symmetry which is a subgroup of $D_8$ and thus have a 4-fold degeneracy. Besides the above two cases, the general canted phase (i.e., $0<\theta<\pi/4$ (mod $\pi/4$)) possesses a 8-fold degeneracy.
We will start our analysis by solving the one-particle Hamiltonian of Eq.~(\ref{Hamil}) with $J=0$.
Such one-particle Hamiltonian shows a three(two)-fold degeneracy at $h=0$ ($h=2K$). Due to such degeneracy, by manipulating the phase factor while we superpose among the degenerate states we are able to construct any product state whose planar order orients in any desired direction. By including the $J$ term, however, a specific orientation might be stabilized under the so-called quantum order by disorder (OBD) effect~\cite{J.Phys.41.1263, doi:10.1063/1.338570, PhysRevLett.62.2056}.
To study such effect we first conduct the mean-field calculation based on the 5 local orthonormal states on each site with non-zero $J$ and observe two $P$ phases with one $C$ phase in the phase diagram. Especially, the easy axes of $C$ phase's planar magnetization always align along $\langle110\rangle$ in the mean-field calculation. However, near the $h=2K$ region all possible canted states with different $\theta$ share very competing energies, due to the above-mentioned two-fold degeneracy at $h=2K$ and a large energy gap.
In what follows, we restrict $\theta$ to be in the range $\theta\in[0,\frac{\pi}{4}]$ with no loss of generality.
For the purpose of a more detailed calculation, we adopt the 2$D$ tensor network ansatz in the thermodynamic limit, the infinite projected entangled-pair states (iPEPS)~\cite{PhysRevLett.101.250602, NatRevPhys.1.538, RevModPhys.93.045003}, in order to take into account the quantum fluctuation beyond the mean-field treatment. According to the iPEPS estimation, the dome of the $P$ phase is larger compared to the mean-field result. In addition, our iPEPS studies reveal a very competing scenario for different $C$ states around $h=2K$, which we interpret as the field-induced easy-axis softening. This is due the huge energy gap among eigenstates that degenerates the Heisenberg term into an effective hard-core bosonic (HCB) model as the leading-order contribution.
\section{\label{onesite}The one-site ($J=0$) limit}
\begin{figure}[pbt]
\centering
\includegraphics[width=1.0\columnwidth]{Fig1.png}
\caption{(a) The spontaneous magnetization (the red arrow) and the angle $\theta$. When a canted phase appears, planar magnetic order points in a certain direction with relative angle $\theta$ to the $x$-axis ($[100]$). (b) Ground state (GS) phase diagram for the single-site Hamiltonian ($J=0$). Definition for each state is recorded in Eq.~(\ref{basis}). (c) The eigenenergy for each local state along with $h$. Energies are calculated according to Eq.~(\ref{prefactor}) and (\ref{eigenvalues}). Dashed line at $h=2K$ indicates the degenerate point.}
\label{fig1}
\end{figure}
\begin{figure*}[pbt]
\centering
\includegraphics[width=2.0\columnwidth]{Fig2.png}
\caption{(a) The phase diagram from MFA and iPEPS. For MFA, phase boundaries are obtained by solving Eq.~(\ref{HamilMF}). Two polarized ($P$) phases are connected through $h=2K$ while canted ($C$) phase forms when $J$ comes into play. Moreover, $C$ phase by iPEPS does not require $\theta=\pi/4$ (see main text). The plots of $S_z$ and $M_p$ order parameters along with $h$ by iPEPS ($D=5$ and $\chi=35$) are shown for (b) $J=0.04K$ and (c) $J=0.08K$. A $C\to P \to C\to P$ phase transition can be clearly seen in (b), while only two phases exist in (c).}
\label{fig2}
\end{figure*}
According to the previous results~\cite{DOMANSKI1988306}, Eq.~(\ref{Hamil}) undergoes a peculiar phase transition when $K$ is dominant relative to $J$. Moreover, if $J=0$, Eq.~(\ref{Hamil}) degenerates to a one-site Hamiltonian. Therefore, it is natural to start with solving Eq.~(\ref{Hamil}) with $J=0$. Because $S=2$, the local Hilbert space is five-dimensional and $S_z=1$, 0, and -1 are still good states. On the other hand, due to the quartic term $S_z=2$ and -2 are connected and we need to re-construct new eigenstates. The results are
\begin{equation}
\begin{aligned}
\begin{bmatrix}
|\psi_1 \rangle \\
|\psi_2 \rangle \\
|\psi_3 \rangle \\
|\psi_4 \rangle \\
|\psi_5 \rangle
\end{bmatrix}
=
\begin{bmatrix}
N(|2\rangle-\alpha |-2\rangle)\\
|1\rangle\\
|0\rangle\\
|-1\rangle\\
N(\alpha |2\rangle+|-2\rangle)
\end{bmatrix},
\end{aligned}
\label{basis}
\end{equation}
where $|2\rangle$, $|1\rangle$, $|0\rangle$, $|-1\rangle$, and $|-2\rangle$ represent the eigenstates of $S^z$ with
\begin{equation}
\begin{aligned}
&\alpha=\frac{\sqrt{16h^2+36K^2}-4h}{6K}\ (<1),\\
&N=\frac{1}{\sqrt{\alpha^2+1}}.
\end{aligned}
\label{prefactor}
\end{equation}
The corresponding eigenvalues are
\begin{equation}
\begin{aligned}
\begin{bmatrix}
E_{1} \\
E_{2} \\
E_{3} \\
E_{4} \\
E_{5}
\end{bmatrix}
=
\begin{bmatrix}
-3\alpha^{-1} K+21K+2h\\
18K-h\\
24K\\
18K+h\\
3\alpha K+21K+2h
\end{bmatrix},
\end{aligned}
\label{eigenvalues}
\end{equation}
with $E_n$ being the eigenenergy of $|\psi_n\rangle$. By varying $h$, we plot ground state (GS) phase diagram in Fig.~\ref{fig1}(b). There are two special points where degeneracy takes place. One is at $h=0$ with three-fold degeneracy among $|\psi_1 \rangle$, $|\psi_2 \rangle$, and $|\psi_4 \rangle$. The other lies at $h=2K$ with two-fold degeneracy between $|\psi_1 \rangle$ and $|\psi_2 \rangle$. Thanks to the degeneracy, canted phase can only be formed on those two points when $J=0$. Such phase diagram also explains the reason why one can observe a peculiar phase transition of $C\to P\to C\to P$ in the previous study~\cite{DOMANSKI1988306}. Moreover, we plot the eigenenergy along $h$ for each state in Fig.~\ref{fig1}(c). One obvious feature lies in the fact that as $h$ is large enough ($\approx2K$), $|\psi_3 \rangle$, $|\psi_4 \rangle$, and $|\psi_5 \rangle$ are largely gapped from $|\psi_1 \rangle$ and $|\psi_2\rangle$.
Moreover, $|\psi_1 \rangle$ and $|\psi_2\rangle$ share very close energies around $h=2K$. All the above facts suggest that near the exact degeneracy, other high-energy states can only serve as the ``tie-breaker'' through quantum fluctuation after we turn on $J$.
\section{Inclusion of Heisenberg coupling}
We will now study the effect while including the $J$ term. Owing to the extra phase degree of freedom while we superpose those degenerate eigenstates, we are able to compose every state standing for a specific magnetic polarization and they all share the same energy $E_{GS}$, which is equal to $18K$ at $h=0$ and $16K$ at $h=2K$. After introducing quantum effect by turning on the $J$ term, we anticipate that a certain direction will be selected due to the OBD effect, thus lifting the degeneracy.
\subsection{\label{MF}Mean-field analysis}
We first study this effect with mean-field approximation (MFA). The mean-field Hamiltonian reads
\begin{equation}
\begin{aligned}
H_{\text{MF}}=&-J\sum_{\langle i,j \rangle}\vec{S}_i \cdot \langle \vec{S}_j\rangle+K\sum_{i,\alpha}(S^\alpha_i)^4 -h\sum_{i}S^z_i.
\end{aligned}
\label{HamilMF}
\end{equation}
By introducing the mean-field of $\langle \vec{S}_j\rangle$ Eq.~(\ref{HamilMF}) again reduces to a one-site Hamiltonian and can be solved iteratively. By self-consistently solving Eq.~(\ref{HamilMF}), we plot the mean-field phase diagram in Fig.~\ref{fig2}(a) as dashed lines, along with the phase boundaries by iPEPS which will be discussed in the next section. To determine $P$ and $C$ phases, we rely on the following orders
\begin{equation}
\begin{aligned}
&M_p=\sqrt{\langle S^x\rangle^2+\langle S^y\rangle^2},\\
&S_z=\langle S^z\rangle,
\end{aligned}
\label{order}
\end{equation}
where $M_p$ is nonzero in $C$ phase. For the $P$ phase inside the dome, $S_z\simeq 1$ while $S_z> 1$ for the $P$ phase beyond $h=2K$. Notice that $S_z$ does not have to be strictly equal to 1 because $U(1)$ symmetry is not present in our Hamiltonian. Moreover, in $C$ phase MFA always favors the direction of $\theta=\pi/4$, which is consistent with previous results~\cite{DOMANSKI1988306}.
However, other $C$ states with different $\theta$ also possess competing energies around $h=2K$. For example, when $J=0.1K$ and $h=2K$, their energy difference is smaller than $10^{-5}$. This is because such energy difference comes from including $|\psi_3 \rangle$ to $|\psi_5 \rangle$, whose weights are very small due to the energy gap (Fig.~\ref{fig1}(c)). In sum, quantum OBD by MFA predicts $\theta=\pi/4$ in $C$ phase, despite a very small energy difference.
Our MFA study suggests a phase coherence among our five local eigenstates, which minimizes the energy of the Heisenberg term, $E_J$. As we turn on $J$, our wavefunction becomes a product state of the locally superposed state,
\begin{equation}
\begin{aligned}
|\psi_{\rm local}\rangle=c_0|\psi_1\rangle+\sum_{i=1}^{4}c_ie^{i\phi_i}|\psi_{i+1}\rangle,
\end{aligned}
\label{localstate}
\end{equation}
in the mean-field manner. All the pre-factors, $c_0$ and $c_i$, are related to their separate eigenenergies and always stay positive or zero. For example, near $h=0$, $c_0$, $c_1$, and $c_3$ possess dominant weights. Besides, the phase factors, $\phi_1$ to $\phi_4$, provide extra degrees of freedom to stabilize the energy (here we assign $|\psi_1\rangle$ with the trivial phase). According to our MFA, in $C$ phase $(\phi_1,\phi_2,\phi_3,\phi_4)=(5\pi/4,3\pi/2,7\pi/4,0)$, suggesting that a coherence is attained and it leads to the $\theta=\pi/4$ magnetization. It is important to note that such coherence is a consequence of our reshuffled local Hilbert space because in the sole Heisenberg model there is no easy-axis appearing after the $SU(2)$ symmetry breaking.
\subsection{\label{ipeps}iPEPS study}
In the previous section, our MFA already tells us that all the five local states become coherent in their phase factors. However, such scenario becomes exact only when each site is isolated from other sites, which is indeed the case for MFA since we consider a one-site Hamiltonian in Eq.~(\ref{HamilMF}). When we allow the inter-site correlation, the information on each site will ``leak'' into surrounding sites, causing the decoherence. Of course, our wavefunction as a whole can still reach coherence in its ground state, but such coherence does not necessarily hold for each site. Moreover, when single-site decoherence takes place, its easy axes can be altered. Thus, in order to better study the quantum effect upon these competing states, we will apply iPEPS for such purpose. The iPEPS tensor network ansatz considers one or many rank-5 tensors in the repeating unit cell and scales to the infinite size with corner-transfer-matrix renormalization group (CTMRG) technique~\cite{doi:10.1143/JPSJ.65.891, PhysRevB.80.094403, PhysRevLett.113.046402}. For bulk tensors they have four auxiliary bonds with dimension $D$ and one physical bond with dimension $d$ ($d=5$ in our case). This ansatz becomes exact only when $D\to \infty$. For environment tensors from CTMRG, their dimension is denoted as $\chi$ and generally $\chi \ge D^2$. Due to the translational invariance, we adopt the $1\times 1$ unit cell. In this work, we are going to variationally optimize the iPEPS tensor using gradients calculated by automatic differentiation (AD)~\cite{PhysRevX.9.031041, Hasik}. Details of such iPEPS calculation are included in Appendix~\ref{ipepsbasics}.
The resulting phase diagram is also shown in Fig.~\ref{fig2}(a) with dots connected by solid lines. Phase transition points are determined by scaling the planar order parameter $M_p$ with inverse bond dimension $1/D$. Details will be explained in Appendix~\ref{scaling}. As one can clearly see, close to the one-site limit ($K\gg J$) both phase boundaries by MFA and iPEPS seem to agree well. While when the quantum fluctuation intervenes, they deviate from each other. It is especially important to see that the dome of $P$ phase expands after introducing quantum effect, which cannot be detected by sole MFA. On the other hand, the other phase boundary between $C$ and $P$ phases is quite consistent with MFA, suggesting that it serves as the saturation line for this quantum spin model. Thus, the second $P$ phase can be well described by $|\psi_1\rangle$. Notice that after the saturation $S_z$ is not equal to 2 because of the $K$ term. In Fig.~\ref{fig2}(b) and (c), we demonstrate two cuts of order parameters in Eq.~(\ref{order}) for $J=0.04$ and $J=0.08$. In Fig.~\ref{fig2}(b), $M_p$ is nonzero in the beginning for small $h$ and then becomes zero entering $P$ phase dome. As the magnetic field is enhanced strongly enough, $M_p$ shows up again in $C$ phase. Finally, the state gets saturated and $M_p$ disappears again. Such phase transition is the peculiar $C\to P \to C\to P$ transition revealed in Ref.~\onlinecite{DOMANSKI1988306}. When $J$ is large enough outside the $P$ phase dome, we end up with only one $C$ and one $P$ phases, as demonstrated in Fig.~\ref{fig2}(c) for $J=0.08$.
\begin{figure*}[pbt]
\centering
\includegraphics[width=2.0\columnwidth]{Fig3.png}
\caption{
Plots of ground state energy and azimuthal angle ($\theta$) defined in Fig.~\ref{fig1}(a) during the optimization process ($D=5$ and $\chi=30$) for (a) $J=0.1K$, $h=0$, (b) $J=0.1K$, $h=2K$, (c) $J=0.2K$, $h=2K$, and (d) $J=0.5K$, $h=2K$. Dashed line represent the location of $\pi/4$. On the right-hand side of each plot, we demonstrate the energy profile of our GS ansatz during the optimization if it fulfills (a) $17.7<E_{GS}<17.72$, (b) $15.375<E_{GS}<15.395$, (c) $14.71<E_{GS}<14.73$, and (d) $12.525<E_{GS}<12.545$, where the interval we choose is around 0.1$\%$ of $E_{GS}$.}
\label{fig3}
\end{figure*}
A more interesting scenario lies in the $C$ phase area. Unlike MFA which always favors $\theta=\pi/4$, after introducing the entanglement we discover that the easy axes do not need to align along $|x|=|y|$ all the time, although it is still favored when the field stays weak.
In Fig.~\ref{fig3}, we demonstrate the energy of our GS ansatz ($E_{GS}$) and its corresponding $\theta$ during the optimization process. When $h=0$ (Fig.~\ref{fig3}(a)), one can see that as $E_{GS}$ converges, $\theta$ is relatively stably equal to $\pi/4$, suggesting that easy-axis aligning along $\langle110\rangle$ is favorable. On the other hand, when $h=2K$ (Fig.~\ref{fig3}(b)) even as $E_{GS}$ converges $\theta$ fluctuates greatly, meaning that various ansatz with different $\theta$ competes with each other energetically. For each plot we also demonstrate its energy profile along $\theta$, as long as $E_{GS}$ sits in between a chosen interval which is around 0.1$\%$ of its own value. In contract to the fact that sampled points all gather around $\theta=45^\circ$ in Fig.~\ref{fig3}(a), we have a very long profile for those points fulfilling the criteria in Fig.~\ref{fig3}(b).
In order to show the effect of $J$, we demonstrate the same plots for $J=0.2K$ (Fig.~\ref{fig3}(c)) and $J=0.5K$ (Fig.~\ref{fig3}(d)) with $h=2K$. From the energy profile, one can see that the easy axes gradually resume as $J$ increases. This is because a strong enough $J$ can lead to a tunneling which overcomes the energy gap; thus the phase coherence once again emerges. In Appendix~\ref{scaling}, We will provide the detailed values for $E_{GS}$ and $\theta$ with some sampled $J$s. In our interpretation, the above scenario represents the fact that easy axes are ``softened'' under certain conditions and can be hardly determined. We will provide a proper reasoning for such softening of easy axes in the following section.
\section{\label{discussion}Effective model around $h=2K$}
Our iPEPS calculation in the previous section suggests a scenario of softening the easy axes in the $C$ phase around $h=2K$, which can already be sensed in MFA due to the small energy difference among states with different easy axes. Its cause comes from the fact that there is a huge energy gap that separates $|\psi_1 \rangle$ and $|\psi_2\rangle$ from the rest. Since $J$ is small, we can treat the Heisenberg $J$ term as the perturbative term and we are going to consider its leading-order contribution. From our MFA, we notice that the possibility to find $|\psi_3 \rangle$ to $|\psi_5\rangle$ ($P(\psi_3)+P(\psi_4)+P(\psi_5)$) is smaller than 0.01 ($\approx0.004$ for $J=0.1K$, $h=2K$). Thus, the local Hilbert space boils down to be two-dimensional in the leading order, composed of $|\psi_1 \rangle$ and $|\psi_2 \rangle$. Therefore, we can reconstruct the perturbative Hamiltonian by considering the algebra using $|\psi_1 \rangle$ and $|\psi_2 \rangle$. For the original $S=2$ system, we have
\begin{equation}
\begin{aligned}
&\langle \psi_1 |S^z|\psi_1 \rangle =(N_1)^2(2a^--2),\\
&\langle \psi_2 |S^z|\psi_2 \rangle =1,\\
&\langle \psi_1 |S^+|\psi_2 \rangle =\langle \psi_2 |S^-|\psi_1 \rangle =2N_1a^-,
\end{aligned}
\label{algebra}
\end{equation}
while all the other terms are zero. As a result, the perturbative term can be re-written as
\begin{equation}
\begin{aligned}
H^{(1)}=&-J\sum_{\langle i,j \rangle}\vec{\tau}_i \cdot \vec{\tau}_j,
\end{aligned}
\label{HamilFO}
\end{equation}
with
\begin{gather}
\tau^+=(\tau^-)^\dagger
=
\begin{bmatrix}
0 &
2N_1a^- \\
0 &
0
\end{bmatrix},
\tau^z
=
\begin{bmatrix}
(N_1)^2(2a^--2) &
0 \\
0 &
1
\end{bmatrix}.
\label{tau}
\end{gather}
With further elaboration, Eq.~(\ref{HamilFO}) can be expressed as (see Appendix~\ref{HCB})
\begin{equation}
\begin{aligned}
H^{(1)}_{\text{HCB}}=&-t\sum_{\langle i,j \rangle}(\sigma^+_i \sigma^-_j + \sigma^-_i \sigma^+_j) +V\sum_{\langle i,j \rangle}\sigma^z_i \sigma^z_j -B\sum_i\sigma^z_i,
\end{aligned}
\label{HamilFO2}
\end{equation}
where $t=2J(N_1a^-)^2$, $V=-J\gamma_1^2$, and $B=4J\gamma_1\gamma_2$ with $\gamma_1=\frac{1}{2}[(N_1)^2(2a^--2)-1]$ and $\gamma_2=\frac{1}{2}[(N_1)^2(2a^--2)+1]$.
$\sigma$ stands for the Pauli matrix. This effective Hamiltonian suggests an emergence of $U(1)$ symmetry, due to the accidental two-fold degeneracy at $h=2K$ and a huge gap that separates the competing eigenenergies of $|\psi_1 \rangle$ and $|\psi_2\rangle$ from other local states around that point.
At $h=2K$, we have $(t,V,B)=(1.8J,-0.81J,-0.36J)$ and thus Eq.~(\ref{HamilFO2}) represents a HCB Hamiltonian with an attractive potential under an external magnetic field. Since $|t|>|V|$, this model is expected to realize a superfluid phase with emergent $U(1)$ symmetry breaking, which does not favor any specific easy axes for the planar magnetization. Thus, the energy difference among $C$ states cannot be obtained with only the leading-order correction, while higher-order effect stays negligible due to the energy gap. Such fact also suggests that different $C$ states could be very competitive in their energies and therefore cannot be easily determined with iPEPS. It also explains the reason why our preliminary trials using QMC suffer from a very slow and difficult convergence, which is also seen in the iPEPS studies. Thus, to pin down the true GS, one might need a more powerful numerical tool or a proper higher-order correction. However, it is not easy or even practical due to the nature of a large-energy-gap existence, leading to a competing scenario among different $C$ states.
Recall that our previous MFA demonstrates a phase coherence among local states and, on the other hand, a decoherence can be triggered while entanglement can be properly taken care of. This happens at around $h=2K$ and can be understood through an emergent $U(1)$ symmetry breaking. If $J$ becomes large enough to introduce more weights for the other three eigenstates, the coherence will be restored to minimize $E_J$. At $h=0$, however, such decoherence is not found because there are three states, $|\psi_1\rangle$, $|\psi_2\rangle$, and $|\psi_4\rangle$, possessing dominant weight and thus the coherence stays robust to weak entanglement. In conclusion, our analysis and numerical calculation suggest a scenario of local eigenstate decoherence caused by the magnetic field and a ferromagnetic Heisenberg coupling, representing the softening of magnetic easy axes.
A simple way for further confirming our effective theory is by adding a small in-plane magnetic field and see if the magnetic moment can be guided. We will demonstrate this in the next section.
\section{\label{summary}Conclusion}
\begin{figure}[pbt]
\centering
\includegraphics[width=1.0\columnwidth]{Fig4.png}
\caption{Plot of the azimuthal angle ($\theta$) from Fig.~\ref{fig1}(a) along $\phi$ in Eq.~(\ref{Hamilin}) after solving Eq.~(\ref{Hamil}) with (\ref{Hamilin}) by using iPEPS ($D=5$ and $\chi=30$). Dashed line stands for the trending line for $h=0$ case and diagonal solid line represents $\theta=\phi$. Here, we have chosen $J=0.1$ with two $h$s, 0 and $2K$.}
\label{fig4}
\end{figure}
In this article, we studied the simplest non-trivial quantum magnetic phenomena caused by the cubic spin anisotropy.
Instead of $U(1)$ symmetry, our mother Hamiltonian enjoys the $D_8$ symmetry and planar magnetization appears after symmetry breaking. Within the classical limit, mean-field calculation reveals two $P$ and one $C$ phases. The first $P$ phase exists within a dome area while the other one appears after the saturation. Within $C$ area, MFA always favors $\theta=\pi/4$ easy axes under a phase coherence of local eigenstates, while other possibilities possess competing energies near $h=2K$, due to the ground state degeneracy as $J=0$ and a large energy gap. Once we try to push further using iPEPS, we discover that the first $P$ dome grows while the second phase boundary agrees well with the MFA result. Moreover, the easy axes for $C$ phase's in-plane magnetization does no longer align along $\langle 110 \rangle$. Instead, $C$ states with various easy axes share competing energies which are not easy to distinguish. Our calculation suggests a softening of easy axes once we turn on the magnetic field, caused by an effective HCB model as the leading-order correction.
Our discovery provides a new possibility in manipulating the magnetism of diluted magnetic semiconductor (DMS)~\cite{RevModPhys.86.187}, where such manipulation plays an important role in designing new devices for future spintronics~\cite{nat.mat.9.952}. Previously, a well studied way for such purpose comes from the electrical manipulation for Mn-doped III-V DMS, such as (Ga,Mn)As and (In,Mn)As, which possesses ferromagnetism due to the Mn dopant.
By applying an external electric field, the magnetic easy axes can be shifted due to the altering of carrier concentration led by the field~\cite{nature.408.944, doi:10.1126/science.1086608, nature.455.515, nat.nano.10.209}. Particularly, the Mn-doped gallium phosphide ceramic, $\text{Ga}_\text{1-x}\text{Mn}_\text{x}\text{P}$, provides a good platform in realization of our proposed scenario. For GaP it has the neutral Mn acceptor ($\text{Mn}^\text{3+}$) with spin $S=2$~\cite{PhysRevB.54.10508} as well as the ferromagnetic exchange~\cite{PhysRevLett.95.207204}, whose Curie temperature (and thus the strength of superexchange coupling) can be adjusted by the Mn concentration~\cite{6971732}. Under a low temperature, moreover, it has a $\langle011\rangle$ cubic anisotropy along with a negligible in-plane uniaxial anisotropy, leading to an easy axis close to $[01\bar{1}]$~\cite{PhysRevB.75.214419, PhysRevB.81.205210}. This easy axis can also be rotated from in-plane $[01\bar{1}]$ to the film normal $[100]$ through alloying nitrogen which changes the strain state of the film~\cite{PhysRevB.81.205210}. We believe that our proposal here can be used to describe the physical scenario of a well-doped $\text{Ga}_\text{1-x}\text{Mn}_\text{x}\text{P}_\text{1-y}\text{N}_\text{y}$ whose easy axis goes along $[11\bar{1}]$. After casting a controlled external magnetic field $\textbf{h}||[100]$, the softening of in-plane easy axis takes place, leading to an easy manipulation of magnetism by applying a small in-plane field.
As a demonstration, we introduce an extra in-plane field to Eq.~(\ref{Hamil}),
\begin{equation}
\begin{aligned}
H_{\text{in-plane}}=-\sum_{i}\textbf{B}\cdot \textbf{S}_i,
\end{aligned}
\label{Hamilin}
\end{equation}
where $\textbf{B}=B(\hat{e}_x\cos{\phi}+\hat{e}_y\sin{\phi})$ with $\hat{e}_x$ and $\hat{e}_y$ being the unit vectors along $x$ and $y$ directions. Here we set $B=0.05K$ and present the $\theta$ of resulting GS, obtained by iPEPS, along with $\phi$ for $h=0$ and $2K$ in Fig.~\ref{fig4}. As one can see, with $h=2K$ the resulting $\theta$ (red dots) aligns well along the $\theta=\phi$ line, suggesting the guiding of magnetic moment by the in-plane field. On the other hand, when $h=0$ since the easy axes remain ``hard'', tilting of magnetization is less observant. Moreover, among red dots their energy difference is negligible ($\approx 10^{-4}$, approximately $0.001 \permil$ of the energy). The tiny energy difference for $h=2K$ case supports our emergent $U(1)$ symmetry breaking scenario proposed in the previous section.
From Ref.~\onlinecite{PhysRevB.75.214419} we notice that the contribution of cubic anisotropy decreases rapidly after 5 Kelvin, suggesting that this easy-axis softening exists in a very low temperature.
Moreover, our scenario is suitable for any easy axis from $[01\bar{1}]$ to $[100]$ because by including a XXZ anisotropy to the Heisenberg term of our model, an easy-axis rotation can be realized while varying the anisotropy strength. An easy way to investigate this phenomenon is to detect the angular dependence of ferromagnetic resonance field under a perpendicular external magnetic field~\cite{PhysRevB.75.214419}, which has, to the best of our knowledge, not been done by any group yet.
In sum, the easy-axis softening proposed in this work unveils a new possibility in manipulating the magnetism for spinful semiconductors, leading to a potential usage for future semiconductor devices.
\section{\label{Acknowledgement}Acknowledgement}
W.-L.T. and H.-Y.L. are supported by National Research Foundation of Korea under the grant numbers NRF-2020R1I1A3074769. H.-Y.L. was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(2014R1A6A1030732).
S.R.G. and N.K. are supported by JSPS KAKENHI Grant No. JP19H01809.
H.-K.W. is supported by JQI-NSF-PFC (NSF Grant No. PHY-1607611).
|
\section{Background and attack model \label{sec:background}}
\begin{figure*}[t]
\centering
\includegraphics[width=6in]{figs/background_CAN_dataframe.png}
\caption{CAN2.0B data frame format.}
\label{fig:can_data_frame}
\end{figure*}
\begin{figure}[t]
\centering
\subfloat[DoS attack \label{fig:dos}]{{\includegraphics[width=0.25\textwidth]{figs/background_dos.png}}}\hfill
\subfloat[Spoofing attack \label{fig:spoof}]{{\includegraphics[width=0.25\textwidth]{figs/background_impersonate.png}}}
\subfloat[Fuzzy attack \label{fig:fuzzy}]{{\includegraphics[width=0.25\textwidth]{figs/background_fuzzy.png}}}
\caption{Depiction of three attack models.}
\label{fig:attack_models}
\end{figure}
\subsection{CAN bus system}
The CAN bus system \cite{1991BOSCH2.0}, which was first introduced by Robert Bosh GmbH in 1985, is a channel for communication between ECUs. Because the CAN bus reduces the complexity and cost of in-vehicle wiring, it has become a de-facto standard for in-vehicle communication systems. A modern vehicle can comprise multiple CAN buses. Each bus supports a specific function (e.g., engine control, powertrain control, brake control, and others \cite{Yoshiyasu2018VehicleSecurity}). These buses are connected through a gateway, as illustrated in \cref{fig:can_bus_sys}.
ECUs exchange information with others through the CAN bus system by broadcasting messages following a predefined data frame format. Each type of message is identified by a CAN ID, which describes the meaning of the data. Therefore, the CAN ID helps receiver ECUs choose appropriate messages for specific functions. In addition, ECUs can be added to the vehicle without any modification to the CAN messages, rendering the CAN bus system more flexible. According to the length of the CAN ID, CAN messages are classified into two types: CAN 2.0A and CAN 2.0B. The ID of CAN 2.0A is only 11 bits (including the base identifier), whereas the ID of CAN 2.0B is 29 bits (including both base and extended identifiers) (see \cref{fig:can_data_frame}). In addition, the CAN ID is used to determine the priority of multiple messages. For example, a message with more leading zero bits in its ID wins the right to be transmitted. In addition, there is a 64-bit data field that contains the information that the sender ECUs want to transmit. The other fields are not involved in IDS research and thus are not explained here. We refer to the CAN specification \cite{1991BOSCH2.0} for additional information on how the CAN bus system works.
\subsection{Attack model}
Because of the working principle of the CAN bus system: broadcast-based system and message priority scheme without encryption and authentication mechanisms, the system is vulnerable to cyber-attacks. Before launching an actual attack, an attacker must access the CAN-bus system. This assumption is practical as many studies launched attacks on the in-vehicle network successfully both directly (via an OBD-II port) and indirectly (via wireless technologies such as WiFi, cellular network, and Bluetooth) \cite{Koscher2010ExperimentalAutomobile, Hoppe2011SecurityCountermeasures, Jo2021ACountermeasures}. After the network is accessed, the attacker can inject malicious messages, resulting in various kinds of attacks such as DoS, spoofing, and fuzzy, which are described in detail below and illustrated in \cref{fig:attack_models}:
\begin{itemize}
\item DoS: The DoS attack aims to disable the availability of the network. By utilizing the message priority mechanism, the attacker can inject high-priority messages to win the bus and prevent transmission of other legitimate messages. For example, the attacker (node B) in \cref{fig:dos} injects the highest priority message whose ID is all zero bits (0x000). As a result, legitimate messages from other nodes (A and C) are delayed.
\item Spoofing: The spoofing attack aims to impersonate a CAN ID to control a specific function of the vehicle as the attacker desires. To accomplish that, the attacker must first determine the target CAN ID and then inject messages with that ID. For example, the attacker in \cref{fig:spoof} masquerades CAN ID of (0x2C0) and sends manipulated messages because there is no authentication.
\item Fuzzy: The fuzzy attack aims to make the vehicle malfunction as well as delay other messages. The attacker can inject multiple messages with a random CAN ID and data. For example, the attacker in \cref{fig:fuzzy} sends malicious messages with random CAN IDs, causing the vehicle to malfunction.
\end{itemize}
\section{Conclusion \label{sec:conclusion}}
The aim of this research is to develop a lightweight and efficient deep learning model for the in-vehicle IDS, using less labeled data as much as possible. We proposed a CAAE model, which is a semi-supervised learning-based in-vehicle IDS to detect CAN bus message injection. The training process consists of three phases, including reconstruction, regularization, and supervised learning phases. With the main idea of utilizing a large amount of unlabeled data, the model learns manifolds of both normal and attack patterns. Therefore, the model provides good results with only a small number of labeled data.
We conducted various experiments to illustrate the performance of the proposed model. A real-world car dataset with four types of message injection attacks was utilized for all the experiments. The results show that feeding only 40\% labeled CAN ID frames to the model achieves the lowest ER of 0.1\% with an F1 score of 0.9984, compared to other supervised methods. In addition, the model can detect unknown attacks, which are not labeled during training, with an F1 score of approximately 0.98 on average. Moreover, the proposed model consists of approximately 2 million trainable parameters, which is very small compared to other deep learning models. Therefore, the proposed model can detect a new attack sample within a millisecond.
We believe that our model can be helpful in reducing the effort required for labeling and collecting attack data, which is time-consuming in real life. Our model is suitable when there is a large number of unlabeled data but only a small number of those are labeled. In addition, we can collect real car data from users. With the proposed scheme, the model can detect stealthy attacks if they exist in the data collection without requiring a labeling process.
However, the research only focused on message injection attacks, there are also other kinds of attacks, such as replay or drop attacks. Our future work will include investigating other kinds of attacks in the CAN bus system and applying the proposed model to these data.
\section{Introduction}
This bundle provides two classfiles, namely \verb+cas-sc.cls+ and
\verb+cas-dc.cls+ and corresponding template files for typesetting
journal articles supposed to go through Elsevier's updated workflow.
\verb+cas-sc.cls+ is meant for one-column, the other
\verb+cas-dc.cls+ for two-column layout. These are now accepted for
submitting articles both in Elsevier's electronic submission system and
elsewhere.
\subsection{Usage}
\begin{enumerate}
\item \verb+cas-sc.cls+ for single column journals.
\begin{vquote}
\documentclass[<options>]{cas-sc}
\end{vquote}
\item \verb+cas-dc.cls+ for single column journals.
\begin{vquote}
\documentclass[<options>]{cas-dc}
\end{vquote}
\end{enumerate}
and have an option \verb+longmktitle+ to handle long front matter.
\section{Front matter}
\begin{vquote}
\title [mode = title]{This is a specimen $a_b$ title}
\tnotemark[1,2]
\tnotetext[1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[2]{The second title footnote which is a longer text
matter to fill through the whole text width and overflow
into another line in the footnotes area of the first page.}
\end{vquote}
\begin{vquote}
\author[1,3]{J.K. Krishnan}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-0000-0000]
\cormark[1]
\fnmark[1]
\ead{<EMAIL>}
\ead[url]{www.jkkrishnan.in}
\credit{Conceptualization of this study,
Methodology, Software}
\affiliation[1]{organization={Department of Physics,
J.K. Institute of Science},
addressline={Jawahar Nagar},
city={Trivandrum},
postcode={695013},
state={Kerala},
country={India}}
\author[2,4]{Han Thane}[style=chinese]
\author[2,3]{William {J. Hansen}}[%
role=Co-ordinator,
suffix=Jr,
]
\fnmark[2]
\ead{<EMAIL>}
\ead[URL]{https://www.university.org}
\credit{Data curation, Writing - Original draft preparation}
\end{vquote}
\begin{vquote}
\affiliation[2]{organization={World Scientific University},
addressline={Street 29},
postcode={1011 NX},
postcodesep={},
city={Amsterdam},
country={The Netherlands}}
\author[1,3]{T. Rafeeq}
\cormark[2]
\fnmark[1,3]
\ead{<EMAIL>}
\ead[URL]{www.campus.in}
\affiliation[3]{organization={University of Intelligent
Studies},
addressline={Street 15},
city={Jabaldesh},
postcode={825001},
state={Orissa},
country={India}}
\cortext[cor1]{Corresponding author}
\cortext[cor2]{Principal corresponding author}
\fntext[fn1]{This is the first author footnote, but is common
to third author as well.}
\fntext[fn2]{Another author footnote, this is a very long
footnote and it should be a really long footnote. But this
footnote is not yet sufficiently long enough to make two
lines of footnote text.}
\nonumnote{This note has no numbers. In this work we
demonstrate $a_b$ the formation Y\_1 of a new type of
polariton on the interface between a cuprous oxide slab
and a polystyrene micro-sphere placed on the slab.
}
\end{vquote}
\begin{vquote}
\begin{abstract}[S U M M A R Y]
This template helps you to create a properly formatted
\LaTeX\ manuscript.
\begin{abstract} ... \end{abstract} and \begin{keyword}
... \end{keyword} which contain the abstract and keywords
respectively. Each keyword shall be separated by
a \sep command.
\end{abstract}
\begin{keywords}
quadrupole exciton \sep polariton \sep WGM \sep BEC
\end{keywords}
\maketitle
\end{vquote}
\subsection{Title}
\verb+\title+ command have the below options:
\begin{enumerate}
\item \verb+title:+ Document title
\item \verb+alt:+ Alternate title
\item \verb+sub:+ Sub title
\item \verb+trans:+ Translated title
\item \verb+transsub:+ Translated sub title
\end{enumerate}
\begin{vquote}
\title[mode=title]{This is a title}
\title[mode=alt]{This is a alternate title}
\title[mode=sub]{This is a sub title}
\title[mode=trans]{This is a translated title}
\title[mode=transsub]{This is a translated sub title}
\end{vquote}
\subsection{Author}
\verb+\author+ command have the below options:
\begin{enumerate}
\item \verb+auid:+ Author id
\item \verb+bioid:+ Biography id
\item \verb+alt:+ Alternate author
\item \verb+style:+ Style of author name, eg.\ chinese
\item \verb+prefix:+ Prefix, eg.\ Sir
\item \verb+suffix:+ Suffix
\item \verb+degree:+ Degree
\item \verb+role:+ Role
\item \verb+orcid:+ ORCID
\item \verb+collab:+ Collaboration
\item \verb+anon:+ Anonymous author
\item \verb+deceased:+ Deceased author
\item \verb+twitter:+ Twitter account
\item \verb+facebook:+ Facebook account
\item \verb+linkedin:+ LinkedIn account
\item \verb+plus:+ Google plus account
\item \verb+gplus:+ Google plus account
\end{enumerate}
\begin{vquote}
\author[1,3]{Author Name}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-0000-0000,
facebook=<facebook id>,
twitter=<twitter id>,
linkedin=<linkedin id>,
gplus=<gplus id>]
\end{vquote}
\begin{figure}
\includegraphics[width=\textwidth]{sc-sample.pdf}
\caption{Single column output (classfile: cas-sc.cls).}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{dc-sample.pdf}
\caption{Double column output (classfile: cas-dc.cls).}
\end{figure}
\subsection{Various Marks in the Front Matter}
The front matter becomes complicated due to various kinds
of notes and marks to the title and author names. Marks in
the title will be denoted by a star ($\star$) mark;
footnotes are denoted by super scripted Arabic numerals,
corresponding author by an Conformal asterisk (*) mark.
\subsubsection{Title marks}
Title mark can be entered by the command, \verb+\tnotemark[<num>]+
and the corresponding text can be entered with the command
\verb+\tnotetext[<num>]+ \verb+{<text>}+. An example will be:
\begin{vquote}
\title[mode=title]{Leveraging social media news to predict
stock index movement using RNN-boost}
\tnotemark[1,2]
\tnotetext[1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[2]{The second title footnote which is a longer
text matter to fill through the whole text width and
overflow into another line in the footnotes area of
the first page.}
\end{vquote}
\verb+\tnotemark+ and \verb+\tnotetext+ can be anywhere in
the front matter, but should be before \verb+\maketitle+ command.
\subsubsection{Author marks}
Author names can have many kinds of marks and notes:
\begin{vquote}
footnote mark : \fnmark[<num>]
footnote text : \fntext[<num>]{<text>}
affiliation mark : \author[<num>]
email : \ead{<emailid>}
url : \ead[url]{<url>}
corresponding author mark : \cormark[<num>]
corresponding author text : \cortext[<num>]{<text>}
\end{vquote}
\subsubsection{Other marks}
At times, authors want footnotes which leave no marks in
the author names. The note text shall be listed as part of
the front matter notes. Class files provides
\verb+\nonumnote+ for this purpose. The usage
\begin{vquote}
\nonumnote{<text>}
\end{vquote}
\noindent and should be entered anywhere before the \verb+\maketitle+
command for this to take effect.
\subsection{Abstract and Keywords}
Abstract shall be entered in an environment that starts
with\break \verb+\begin{abstract}+ and ends with
\verb+\end{abstract}+. Longer abstracts spanning more than
one page is also possible in slass file even in double
column mode. We need to invoke \verb+longmktitle+ option in the
class loading line for this to happen smoothly.
The key words are enclosed in a \verb+{keyword}+
environment.
\begin{vquote}
\begin{abstract}
This is an abstract. \lipsum[3]
\end{abstract}
\begin{keywords}
First keyword \sep Second keyword \sep Third
keyword \sep Fourth keyword
\end{keywords}
\end{vquote}
\section{Main Matter}
Main matter contains sections, paragraphs, equations and floats like
tables, figures, textboxes etc.
\subsection{Tables}
\subsubsection{Normal tables}
\begin{vquote}
\begin{table}
\caption{This is a test caption.}
\begin{tabular*}{\tblwidth}{@{} LLLL@{} }
\toprule
Col 1 & Col 2\\
\midrule
12345 & 12345\\
12345 & 12345\\
12345 & 12345\\
12345 & 12345\\
12345 & 12345\\
12345 & 12345\\
\bottomrule
\end{tabular*}
\end{table}
\end{vquote}
\subsubsection{Span tables}
\begin{vquote}
\begin{table*}[width=.9\textwidth,cols=4,pos=h]
\caption{This is a test caption.}
\begin{tabular*}{\tblwidth}{@{} LLLLLL@{} }
\toprule
Col 1 & Col 2 & Col 3 & Col4 & Col5 & Col6 & Col7\\
\midrule
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
\bottomrule
\end{tabular*}
\end{table*}
\end{vquote}
\subsection{Figures}
\subsubsection{Normal figures}
\begin{vquote}
\begin{figure}
\centering
\includegraphics[scale=.75]{Fig1.pdf}
\caption{The evanescent light - $1S$ quadrupole coupling.
See also Fig. \protect\ref{FIG:2}).}
\label{FIG:1}
\end{figure}
\end{vquote}
\subsubsection{Span figures}
\begin{vquote}
\begin{figure*}
\centering
\includegraphics[width=\textwidth,height=2in]{Fig2.pdf}
\caption{Schematic of formation of the evanescent polariton on
linear chain of \PMS.}
\label{FIG:2}
\end{figure*}\end{vquote}
\subsection{Theorem and theorem like environments}
CAS class file provides a few hooks to format theorems and
theorem like environments with ease. All commands the
options that are used with \verb+\newtheorem+ command will work
exactly in the same manner. Class file provides three
commands to format theorem or theorem like environments:
\begin{enumerate}
\item \verb+\newtheorem+ command formats a theorem in
\LaTeX's default style with italicized font for theorem
statement, bold weight for theorem heading and theorem
number typeset at the right of theorem heading. It also
optionally accepts an argument which will be printed as an
extra heading in parentheses. Here is an example coding and
output:
\begin{vquote}
\newtheorem{theorem}{Theorem}
\begin{theorem}\label{thm}
The \WGM evanescent field penetration depth into the
cuprous oxide adjacent crystal is much larger than the
\QE radius:
\begin{equation*}
\lambda_{1S}/2 \pi \left({\epsilon_{Cu2O}-1}
\right)^{1/2} = 414 \mbox{ \AA} \gg a_B = 4.6
\mbox{ \AA}
\end{equation*}
\end{theorem}
\end{vquote}
\item \verb+\newdefinition+ command does exactly the same
thing as with except that the body font is up-shape instead
of italic. See the example below:
\begin{vquote}
\newdefinition{definition}{Definition}
\begin{definition}
The bulk and evanescent polaritons in cuprous oxide
are formed through the quadrupole part of the light-matter
interaction:
\begin{equation*}
H_{int} = \frac{i e }{m \omega_{1S}} {\bf E}_{i,s}
\cdot {\bf p}
\end{equation*}
\end{definition}
\end{vquote}
\item \verb+\newproof+ command helps to define proof and
custom proof environments without counters as provided in
the example code. Given below is an example of proof of
theorem kind.
\begin{vquote}
\newproof{pot}{Proof of Theorem \ref{thm}}
\begin{pot}
The photon part of the polariton trapped inside the \PMS
moves as it would move in a micro-cavity of the effective
modal volume $V \ll 4 \pi r_{0}^{3} /3$. Consequently, it
can escape through the evanescent field. This evanescent
field essentially has a quantum origin and is due to
tunneling through the potential caused by dielectric
mismatch on the \PMS surface. Therefore, we define the
\emph{evanescent} polariton (\EP) as an evanescent light -
\QE coherent superposition.
\end{pot}
\end{vquote}
\end{enumerate}
\subsection{Enumerated and Itemized Lists}
CAS class files provides an extended list processing macros
which makes the usage a bit more user friendly than the
default LaTeX list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list
counter type and its attributes. You can see the coding and
typeset copy.
\begin{vquote}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.' so that the item counter will be suffixed
by a period as in the optional argument.
\item If you provide a closing parenthesis to the number in the
optional argument, the output will have closing
parenthesis for all the item counters.
\item You can use `(a)' for alphabetical counter and `(i)' for
roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\begin{enumerate}[(i)]
\item This item has roman numeral counter.
\end{vquote}
\begin{vquote}
\item Another one before we close the third level.
\end{enumerate}
\item Third item in second level.
\end{enumerate}
\item All list items conclude with this step.
\end{enumerate}
\section{Biography}
\verb+\bio+ command have the below options:
\begin{enumerate}
\item \verb+width:+ Width of the author photo (default is 1in).
\item \verb+pos:+ Position of author photo.
\end{enumerate}
\begin{vquote}
\bio[width=10mm,pos=l]{tuglogo.jpg}
\textbf{Another Biography:}
Recent experimental \cite{HARA:2005} and theoretical
\cite{DEYCH:2006} studies have shown that the \WGM can travel
along the chain as "heavy photons". Therefore the \WGM
acquires the spatial dispersion, and the evanescent
quadrupole polariton has the form (See Fig.\ref{FIG:3}):
\endbio
\end{vquote}
\section[CRediT...]{CRediT authorship contribution statement}
Give the authorship contribution after each author as
\begin{vquote}
\credit{Conceptualization of this study, Methodology,
Software}
\end{vquote}
To print the details use \verb+\printcredits+
\begin{vquote}
\author[1,3]{J.K. Krishnan}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-0000-0000]
\end{vquote}
\begin{vquote}
\cormark[1]
\fnmark[1]
\ead{<EMAIL>}
\ead[url]{www.jkkrishnan.in}
\credit{Conceptualization of this study, Methodology, Software}
\affiliation[1]{organization={Department of Physics,
J.K. Institute of Science},
addressline={Jawahar Nagar},
city={Trivandrum},
postcode={695013},
state={Kerala},
country={India}}
\author[2,4]{Han Thane}[style=chinese]
\author[2,3]{William {J. Hansen}}[%
role=Co-ordinator,
suffix=Jr,
]
\fnmark[2]
\ead{<EMAIL>}
\ead[URL]{https://www.university.org}
\credit{Data curation, Writing - Original draft preparation}
. . .
. . .
. . .
\printcredits
\end{vquote}
\section{Bibliography}
For CAS categories, two reference models are recommended.
They are \file{model1-num-names.bst} and \file{cas-model2-names.bst}.
Former will format the reference list and their citations according to
numbered scheme whereas the latter will format according name-date or
author-year style. Authors are requested to choose any one of these
according to the journal style. You may download these from
The above bsts are available in the following location for you to
download:
\url{https://support.stmdocs.in/wiki/index.php?title=Model-wise_bibliographic_style_files}
\hfill $\Box$
\end{document}
\section{Introduction}
\begin{figure*}[ht!]
\centering
\includegraphics[width=5in]{figs/background_CAN_bus_system.png}
\caption{Overview of an in-vehicle network using a CAN bus. Many buses installed in the vehicle include multiple ECUs to serve a specific function and are connected through a gateway. The network can be accessed through an OBD-II port and wireless technologies such as WiFi, Bluetooth, and cellular network.}
\label{fig:can_bus_sys}
\end{figure*}
To enable autonomous vehicles, many sensors or electrical control units (ECUs) have been installed on vehicles \cite{Tuohy2015Intra-VehicleReview}. The ECUs provide useful information about the real-life environment to the vehicles, which facilitates the learning process of the vehicles through deep learning \cite{Deo2018Multi-ModalLSTMs, Lee2021FastTransformation} or reinforcement learning methods \cite{Isele2017NavigatingLearning, Wang2018AManeuvers}. In addition, these connected ECUs communicate with each other through a controller area network (CAN) bus system \cite{1991BOSCH2.0}. Since 1986, the CAN bus system has been used widely because of its high speed and efficiency. Furthermore, the CAN bus operates based on broadcast and message priority mechanisms. However, there is no authentication and encryption in the CAN bus. Therefore, the system is vulnerable to various cyberattacks, particularly message injection attacks. It can be risky when an attacker accesses the CAN bus system and sends anomalous messages. For example, the attacker can control the brake, lock the doors, or even steer the vehicle. Many studies have shown that we can inject messages to the CAN bus system directly through an on-board diagnostics II (OBD-II) port or indirectly through WiFi or bluetooth channels \cite{Koscher2010ExperimentalAutomobile, Hoppe2011SecurityCountermeasures, Jo2021ACountermeasures}.
An intrusion detection system (IDS) has been introduced to monitor and detect attacks in CAN buses \cite{Hoppe2011SecurityCountermeasures}. IDS, which is usually placed in an ECU, receives and analyses incoming messages. It will raise an alert to other ECUs if any anomalous message is detected. The machine learning-based IDS has piqued the interest of many researchers because it can adapt to the complexity and instability of cybersecurity in the in-vehicle network. In terms of detecting manner, the machine learning-based IDS can be divided into two groups: unsupervised and supervised models. On the one hand, unsupervised models learn the representation of normal patterns and then detect an abnormal state based on its deviation from the normal. The problem with this scheme is that the model produces a high false-positive rate. In addition, we need to find an optimal threshold to achieve the best result. On the other hand, supervised models learn to classify a predefined number of classes provided by a labeled dataset. Therefore, we usually must provide a large amount of labeled data to supervised models to achieve a good outcome.
In this study, we propose a novel semi-supervised deep learning-based IDS, in which the model learns to extract appropriate features from unlabeled data by itself before being trained in a supervised manner. Therefore, the proposed model can handle not only limited data environments but also unknown attacks.
Specifically, our main contributions can be summarized as follows:
\begin{itemize}
\item We propose a convolutional adversarial autoencoder (CAAE) based IDS by combining convolutional autoencoder and generative adversarial networks (GAN) to counter both known and unknown attacks. Because the proposed model is trained in semi-supervised learning, only a small number of labeled data is required during training. We believe that the proposed model can reduce the time to collect and annotate data significantly. To the best of our knowledge, this is the first time that CAAE is applied to the in-vehicle IDS.
\item To demonstrate the performance of our model, we conducted a comprehensive experiment using a real-car dataset
with both known and unknown attacks. In addition, we provide the source code\footnote{Source code is available at https://github.com/htn274/CanBus-IDS} to facilitate future studies on this topic.
\item Using approximately 60k labeled samples, which accounts for only 40\% of the total training data, the proposed method achieved a high F1 score of 0.99 and a low error rate of 0.1\%, compared to other supervised and unsupervised models. Moreover, the proposed model successfully reduced the number of model parameters by 5 times and 8 times for the inference time. Therefore, the proposed model is efficient for real-time detection.
\end{itemize}
The remaining part of the paper proceeds as follows: Section \ref{sec:background} introduces the background of the CAN bus system and attack models related to the study. Related works are presented in Section \ref{sec:related_works}. Furthermore, Section \ref{sec:method} describes our proposed method in detail. The experimental results and conclusion are described in Section \ref{sec:results} and Section \ref{sec:conclusion}, respectively.
\section{Methodology \label{sec:method}}
We propose our methodology to address the limitations of related works, such as low detection rate for unknown attacks in supervised models and high false positive rate in unsupervised models. The proposed deep learning-based IDS is developed from an adversarial autoencoder (AAE) architecture, which is a combination of autoencoder (AE) and generative adversarial networks (GAN). The AAE scheme is suitable for in-vehicle IDS for two reasons. First, the AAE-based model can handle the data scarcity problem. This is because it does not require a large amount of data, which consumes a long time to collect and label. Particularly, safety has the highest priority in the vehicle domain. Therefore, the data source for attack samples is limited. Second, the AAE-based model can detect unknown attacks. Because the security of the CAN-bus system is extremely weak, the system is vulnerable to various types of attacks, which are updated frequently by intelligent adversaries. In some cases, security researchers are unaware of a new type of attack, therefore it is not labeled.
In this section, we first explain the fundamental knowledge about AE, GAN, and AAE. Then, the details of our proposed system are presented.
\subsection{Autoencoder}
\begin{figure}[h!]
\centering
\includegraphics[width=2.5in]{figs/method_AE.png}
\caption{Autoencoder architecture.}
\label{fig:AE}
\end{figure}
An autoencoder (AE) \cite{Ian2016DeepLearning} is an unsupervised neural network that is trained to reconstruct its input. The network (\cref{fig:AE}) consists of two parts: an encoder maps the input to latent features and a decoder attempts to reconstruct the input with the encoder output. With the condition that the dimensionality of the latent space is assumed to be much lower than the dimensionality of the input, the autoencoder can learn useful properties of the data. Therefore, it is usually used for dimensionality reduction. The goal of the autoencoder is to minimize the reconstruction loss $L_R$, which can be defined as the squared error between the input $X$ and the reconstructed output $X'$ with $N$, the number of samples, as follows:
\begin{equation}\label{eq:reconstruction_loss}
L_R(\textbf{X}, \textbf{X}^\prime) = \frac{1}{N}||\textbf{X} - \textbf{X}^\prime||^2.
\end{equation}
\subsection{Generative Adversarial Networks}
\begin{figure}[h!]
\centering
\includegraphics[width=3.5in]{figs/method_GAN.png}
\caption{GAN architecture.}
\label{fig:GAN}
\end{figure}
Generative Adversarial Networks (GAN) \cite{GoodfellowGenerativeNets} is a deep learning-based model that uses a training set assumed to follow a distribution $p_{data}$. GAN learns to represent an estimate of that distribution and returns a probability distribution $p_{model}$ as the result. As illustrated in \cref{fig:GAN}, GAN is composed of two deep neural networks: generator ($G$) and discriminator ($D$). Although $G$ attempts to generate new data plausibly, $D$ is trained to distinguish real data derived from the training dataset and generated data from $G$.
To train $G$, we obtain the noise input {$\boldsymbol{z}$} following a predefined distribution. $G(\boldsymbol{z}; \boldsymbol{\theta_g})$ is new data generated from $G$ with parameter $\boldsymbol{\theta_g}$. Furthermore, $D(\boldsymbol{x}; \boldsymbol{\theta_d})$ represents the probability that $\boldsymbol{x}$ is real or fake, and $D(G(\textbf{z}))$ evaluates the authenticity of data generated from $G$.
In addition, $D$ is trained to maximize the probability of assigning the correct label to both training examples and samples from $G$. The loss function of $D$ can be formulated as follows:
\begin{equation}
\begin{split}
\mathop{\text{max}}_{D} V(D) &= \mathbb{E}_{\boldsymbol{x} \sim p_{data}(\boldsymbol{x})}[\text{log}(D(\boldsymbol{x}))] \\
& + \mathbb{E}_{\boldsymbol{z} \sim p_{\boldsymbol{z}}(\boldsymbol{z})} [ \text{log}(1 - D(G(\boldsymbol{z}))].
\end{split}
\end{equation}
By contrast, $G$ wants to create new data that appear similar to the training data to decieve $D$. As a result, $G$ minimizes $log(1 - D(z))$. The loss function of $G$ is
\begin{equation}
\mathop{\text{min}}_{G} V(G) = \mathbb{E}_{\boldsymbol{z} \sim p_{\boldsymbol{z}}(\boldsymbol{z})} [1 - \text{log}(D(G(\boldsymbol{z}))].
\end{equation}
To summarize, GAN loss can be written as
\begin{equation} \label{eq:gan_loss}
\begin{split}
\mathop{\text{min}}_{G} \mathop{\text{max}}_{D} (V(D, G))
& = \mathbb{E}_{\boldsymbol{x} \sim p_{data}(x)}[\text{log}(D(\boldsymbol{x}))] \\
& + \mathbb{E}_{\boldsymbol{z} \sim p_z(\boldsymbol{z})} [1 - \text{log}(D(G(\boldsymbol{z}))].
\end{split}
\end{equation}
\subsection{Adversarial Autoencoder}
\begin{figure}[h!]
\centering
\includegraphics[width=3in]{figs/method_AAE.png}
\caption{AAE architecture.}
\label{fig:AAE}
\end{figure}
AAE \cite{Makhzani2015AdversarialAutoencoders} is a technique to regularize a vanilla autoencoder by applying adversarial training. The goal is to force the latent space to follow an arbitrary distribution. The overview of AAE architecture is shown in \cref{fig:AAE}.
For example, we assume $\boldsymbol{x}$ is the input and $\boldsymbol{z}$ is the latent variable of the AE. Furthermore, assuming $p(\boldsymbol{z})$ is the prior distribution that we want to impose on the latent variable, $q(\boldsymbol{z}|\boldsymbol{x})$ is an encoding distribution, and $p(\boldsymbol{x}|\boldsymbol{z})$ is a decoding distribution. The encoding function of the AE, $q(\boldsymbol{z}|\boldsymbol{x})$, defines an aggregated posterior distribution of $q(\boldsymbol{z})$ on the latent variable of the AE as follows:
\begin{equation} \label{eq:aae_loss}
\begin{split}
q(\boldsymbol{z}) = \int_x q(\boldsymbol{z}|\boldsymbol{x})p(\boldsymbol{x}|\boldsymbol{z})dx.
\end{split}
\end{equation}
Thus, the AAE can match the prior distribution $p(\boldsymbol{z})$ to the aggregated posterior $q(\boldsymbol{z})$ of latent variable $\boldsymbol{z}$. In other words, the latent variable $\boldsymbol{z}$ follows the prior distribution, and the encoder of AE plays as the generator of GAN, which generates the latent variable $\boldsymbol{z}$. The discriminator scores the similarity between $\boldsymbol{z}$ and samples $\boldsymbol{\boldsymbol{z^\prime}}$ drawn from prior distribution of $p(\boldsymbol{z})$. AAE is trained with stochastic gradient descent (SGD) in two phases as follows:
\begin{enumerate}
\item Reconstruction phase: only the autoencoder part is trained with reconstruction loss in this phase. First, the input is fed into the encoder to obtain the latent features. The decoder decodes those latent features to retrieve the input. This phase will optimize the reconstruction loss, $L_R$ defined in \eqref{eq:reconstruction_loss}.
\item Regularization phase: the generator (encoder) and discriminator are trained together. First, the discriminator learns how to classify the encoder output and random input from the prior distribution. The random input is labeled as 1, and the encoder output is labeled as 0. Then, we fix the discriminator and train the generator to produce output following the prior distribution. To achieve that, the generator is trained to minimize the cross-entropy between target values, which are all set to be 1, and the scores that the discriminator provides to the encoder outputs. This phase attempts to minimize the GAN loss defined in \eqref{eq:gan_loss} by replacing $\boldsymbol{x}$ to $\boldsymbol{\boldsymbol{z^\prime}}$.
\end{enumerate}
\subsection{Proposed model}
\begin{figure}[h!]
\centering
\includegraphics[width=3in]{figs/method_Workflow.png}
\caption{Workflow overview.}
\label{fig:workflow}
\end{figure}
\subsubsection{Workflow overview}
The workflow (\cref{fig:workflow}) consists of 2 phases: training and testing. We propose a deep learning model named the convolutional adversarial autoencoder (CAAE). As suggested by \cite{Song2020In-vehicleNetwork}, we utilize only the CAN ID in messages to capture its sequential pattern. After data is collected, we construct a frame with a size of $29 \times 29$ by stacking 29 consecutive CAN IDs in a 29-bit representation. Then, the frame is fed into the CAAE model.
In the training phase, we label the frame as abnormal if there is at least one injected message. However, we do not need to label all of them because our model only needs a small number of labeled data, which helps us save time on labeling frames as well as messages. The training process will be completed offline, whereas the detection occurs online to serve in real-time.
\begin{figure*}[H!]
\centering
\includegraphics[width=5in]{figs/method_frameCANID.png}
\caption{CAN ID frame - input to CAAE model}
\label{fig:frame}
\end{figure*}
\subsubsection{Data preprocessing}
Based on the working principles of the CAN bus system, the proposed method utilizes the CAN IDs as features because there is a pattern in the CAN IDs sequence. Whenever an injected message exists, the pattern will be broken. Therefore, using the CAN IDs sequence, the model can capture the normal and abnormal patterns and classify them correctly. To enable the model to adapt to any version of CAN messages and increase efficiency, the CAN IDs are represented in 29 bits, which means that each CAN ID is illustrated as follows:
\begin{equation}
\text{ID} = b_i (\text{for } i = 0, ..., 28),
\end{equation}
where $b_i$ is the bit at $i^{th}$ position. The \cref{fig:frame} illustrates a frame that is constructed by stacking 29 continuous CAN IDs together. We chose 29 because a square matrix is easier for CNN to process the input data. The frames are the inputs fed into the CAAE model, which will be described in detail in the next section.
\subsubsection{Convolutional Adversarial Autoencoder (CAAE)}
\begin{figure}[h!]
\centering
\includegraphics[width=3.3in]{figs/method_CAAE.png}
\caption{CAAE architecture.}
\label{fig:CAAE}
\end{figure}
To reduce the amount of labeled data, we train the CAAE in semi-supervised learning, which combines a small amount of labeled data with a large amount of unlabeled data during training. The architecture of our model is presented in \cref{fig:CAAE}. The input consists of $n$ labeled samples denoted as $\{\boldsymbol{X_l}, \boldsymbol{Y_l}\}$ and $m$ unlabeled samples denoted as $\{\boldsymbol{X_{ul}}\}$, where $n \ll m$.
The encoder generates two latent variables: $\boldsymbol{\widehat{Y}} \in \mathbb{R}^2$ is for class information (normal and abnormal) and $\boldsymbol{\widehat{Z}} \in \mathbb{R}^{10}$ is for other features. Therefore, the proposed model needs two discriminators: $D_{cat}$ forces $\boldsymbol{\widehat{Y}}$ to follow the categorical distribution $Cat(2)$, whereas $D_{gaus}$ forces $\boldsymbol{\widehat{Z}}$ to follow the Gaussian distribution with the mean of zero and identity covariance. Because the representations of normal and abnormal patterns can be very complicated and cannot be described by only the class information $\boldsymbol{\widehat{Y}}$, we use $\boldsymbol{\widehat{Z}}$ to preserve other specific features, and then samples are mapped accurately in the latent space.
The training process for CAAE is similar to that for AAE, except
that a supervised phase is added. First, we train the autoencoder part by feeding a batch of $\boldsymbol{\{X_{ul}\}}$. The next step is training two discriminators corresponding to each type of latent feature and the encoder in an adversarial way. Each discriminator comprises two layers, with 1000 neurons for each layer. The discriminator output is the probability generated by the sigmoid activation function. To make the model more stable and converge faster, we use the Wasserstein GAN (WGAN) loss with gradient penalty (GP) \cite{GulrajaniImprovedAlgorithms}.
Rather than using Jensen–Shannon divergence, WGAN uses Wasserstein distance to calculate the difference between real and fake distributions. We consider the categorical discriminator as an example. The categorical discriminator $D_{cat}$ attempts to minimize the loss defined as follows:
\begin{equation} \label{eq:wgan}
L_{WGAN} = \mathbb{E}[D_{cat}(\boldsymbol{\widehat{Y}})] - \mathbb{E}[D_{cat}(\boldsymbol{Y^\prime})],
\end{equation}
where $\boldsymbol{Y^\prime}$ is the samples drawn from the categorical distribution $Cat(2)$.
A gradient penalty, which is added to the loss to ensure the 1-Lipschitz constraint in GAN, is defined as follows:
\begin{equation} \label{eq:gp}
\text{GP}_{cat} = \mathbb{E}[(\| \nabla_{\boldsymbol{\tilde{Y}}} D(\tilde{Y}) \| - 1)^2],
\end{equation}
where $\tilde{Y} = \epsilon \widehat{Y} + (1 - \epsilon) \boldsymbol{Y^\prime}$, $\epsilon$ is a random variable, and $\epsilon \sim \mathcal{U}[0,1]$. The final loss is the sum of (\ref{eq:wgan}) and (\ref{eq:gp})
\begin{equation}
L_{cat} = L_{WGAN} + \lambda \text{GP}_{cat},
\end{equation}
where $\lambda$ is a penalty factor and is usually set to 10 \cite{GulrajaniImprovedAlgorithms}. These formulas are similar when applied for Gaussian discriminator $D_{gaus}$.
Following this step, the model learns to extract useful features from unlabeled data. Finally, we train the encoder with labeled samples $\{\boldsymbol{X_l}, \boldsymbol{Y_l}\}$ by minimizing the cross-entropy loss. The aforementioned steps are summarized in \cref{alg:caae}, where $\boldsymbol{\theta}$ indicates the parameters of the model. Although the training process appears complicated, only the encoder's weights are saved and utilized for online detection. Consequently, our model is lightweight and efficient for the in-vehicle IDS.
\begin{figure*}[h!]
\centering
\includegraphics{figs/method_CAE.png}
\caption{Detail of convolutional layers in Autoencoder.}
\label{fig:CAE}
\end{figure*}
\begin{algorithm}
\caption{CAAE training process}\label{alg:caae}
\textbf{Input:}
\begin{itemize}
\item $\boldsymbol{\{X_{ul}\}}$: Unlabaled dataset.
\item ${\{\boldsymbol{X_l, Y_l}\}}$: Labeled dataset.
\item $n\_epochs$: Number of epochs.
\item $batch\_size$: Batch size.
\end{itemize}
\textbf{Output:}
\begin{itemize}
\item $\boldsymbol{\theta}_{En}$: Encoder's weights.
\end{itemize}
\begin{algorithmic}[1]
\FOR{$epoch$:= 1 \TO $n\_epochs$}
\STATE Sample minibatch for $\{\boldsymbol{X_l}, \boldsymbol{Y_l}\}$, $\boldsymbol{\{X_{ul}\}}$.
\STATE Draw samples for $\boldsymbol{Y^\prime} \sim Cat(2)$, $\boldsymbol{Z^\prime} \sim N(0, I_{10})$.
\STATE \textbf{Reconstruction phase}
\STATE Update $\boldsymbol{\theta}_{En}$ and $\boldsymbol{\theta}_{De}$ by minimizing
\(L_R = \frac{1}{batch\_size}\|\boldsymbol{X}_{ul} - \boldsymbol{\widehat{X}}\|^2\).
\STATE \textbf{Regularization phase}
\STATE Update $\boldsymbol{\theta}_{D_{cat}}$ by minimizing
\(L_{cat} = \mathbb{E}[D_{cat}(\boldsymbol{\widehat{Y}})] - \mathbb{E}[D_{cat}(\boldsymbol{Y^\prime})] + \lambda \text{GP}_{cat}\).
\STATE Update $\boldsymbol{\theta_{D_{gaus}}}$ by minimizing
\(L_{gaus} = \mathbb{E}[D_{gaus}(\boldsymbol{\widehat{Z}})] - \mathbb{E}[D_{gaus}(\boldsymbol{Z^\prime})] + \lambda \text{GP}_{gaus}\).
\STATE Update $\boldsymbol{\theta}_{En}$ by minimizing
\(L_{gen} = -\mathbb{E}[D_{cat}(\boldsymbol{\widehat{Y}})] - \mathbb{E}[D_{gaus}(\boldsymbol{\widehat{Z}})]\).
\STATE \textbf{Supervised learning phase}
\STATE Using $\{\boldsymbol{X_l}, \boldsymbol{Y_l}\}$ to update $\boldsymbol{\theta}_{En}$ by minimizing
\(L_{sup} = -\mathbb{E}[\text{log}(\boldsymbol{\widehat{Y}}) * \boldsymbol{Y_l}]\).
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{table}[t!]
\centering
\caption{Hyperparameter value of the CAAE model}
\begin{tabular}{c|c}
\hline
Hyperparameter & Value \\
\hline Batch size & 64 \\
Supervised learning rate & $10^{-4}$\\
Reconstruction learning rate & $10^{-4}$ \\
Regularization learning rate & $10^{-4}$ \\
Decay for learning rate update & 0.1 \\
Latent space dimension & 2 + 10 \\
\hline
\end{tabular}
\label{tab:hyperparameter}
\end{table}
Convolution, which is well-established for spatial and sequential patterns, is extremely suitable for the CAN IDs data. Therefore, we added 2D convolutional layers to our autoencoder structure. Because we trained the model in a semi-supervised manner, our convolution neural network is very simple. Thus, the proposed model still adapts to the real-time requirement for in-vehicle IDS. The $29 \times 29$ frame input is transformed into a $32 \times 32$ input by padding. This is because an even-numbered size is more suitable for the convolutional autoencoder. For the encoder, we applied convolution with a kernel size of $3 \times 3$ and max-pooling layers. Then the result is flattened and applied to two fully connected networks to generate two types of latent features. Then, the encoder outputs are concatenated before being fed into the decoder. By contrast, the decoder includes deconvolution layers with the same kernel size as the encoder and upsampling layers. The final result is cropped to regain the $29 \times 29$ frame. To prevent overfitting during training, we added a dropout layer \cite{Srivastava2014Dropout:Overfitting} with a rate of 0.15 before the fully connected layers. In addition, we used ReLU \cite{Nair2010RectifiedMachines} for activation functions and the ADAM optimizer \cite{Kingma2014Adam:Optimization} for backpropagation. We also used the learning rate decay technique, which decreases the learning rate by 10 after the $50^{th}$ epoch. The details of the architecture of the convolutional autoencoder and hyperparameters for training the CAAE model are shown in \cref{fig:CAE} and \cref{tab:hyperparameter} respectively.
\section{Related works \label{sec:related_works}}
An IDS monitors and detects malicious behaviors in a network. For the in-vehicle network, an IDS can be installed in an ECU to serve as an additional node in the CAN bus. Because each ECU broadcasts messages, the IDS analyzes incoming messages and alerts if any abnormality exists in the network. In terms of design, the in-vehicle IDS can be grouped into four categories: fingerprint-based methods (bus level), parameter monitoring-based methods (message level), information theoretic-based methods (data-flow level), and machine learning-based methods (functional level) \cite{Wu2020ANetworks}. Previous studies showed that machine learning methods are efficient for detecting an attack from the application layer. Our study also falls into this category. In this section, we will review state-of-the-art machine learning studies related to in-vehicle IDS, which are summarized in \cref{tab:litrev}.
For intruder detection problems, machine learning or deep learning models can be trained in supervised or unsupervised manners. Regarding supervised approaches, a large and accurate labeled dataset is required. The IDS problem can be formulated as a binary classification with two classes: normal and abnormal. In \cite{Kang2016ASecurity}, the authors proposed a deep neural network consisting of 64 neurons at the input layer, which represents the data field in the CAN message. They tested the proposed model with a simulation dataset and achieved a high detection rate for both the classes. By contrast, the authors from \cite{Song2020In-vehicleNetwork} published a real dataset - the car hacking dataset, which has been widely used to test IDS models. They also proposed a deep convolutional neural network (DCNN) that accepts a 29-CAN-ID frame as input. Although the DCNN model has a low false alarm rate, it requires high computation costs. Furthermore, the authors in \cite{Hossain2020LSTM-basedCommunications} leverage the time-series information for IDS, using long short-term memory (LSTM) neural networks. In addition, a combination of CNN and attention-based gated recurrent unit (GRU) was proposed in \cite{Javed2021CANintelliIDS:GRU}. Similar to DCNN, the proposed models are extremely complicated to deploy in real life. Conversely, in \cite{Moulahi2021ComparativeBus}, simple machine learning models are used for faster training and inference. However, the models achieve low accuracy, particularly for DoS and fuzzy attacks.
Although the supervised models provide good results, it is difficult to collect sufficient labeled data for learning. In addition, some models cannot detect zero-day attacks because the supervised models can only learn existing patterns in the training dataset. Consequently, unsupervised models have been proposed, in which only normal samples are used in the training phase. In the testing phase, any sample deviating very far from the learned patterns is classified as abnormal. Based on this concept, some in-vehicle IDS studies proposed traditional machine learning techniques, such as K-means and K-nearest neighbors \cite{DAngelo2020AVehicles}, one-class SVM (OSVM) \cite{Avatefipour2019AnLearning}, and Kohonen SOM network (KSOM) \cite{Barletta2020IntrusionApproach}, whereas others proposed deep learning models, such as hierarchical temporal memory (HTM) learning algorithm \cite{Wang2018AHTM} and LSTM based autoencoder \cite{Ashraf2020NovelSystems} to improve the detection performance. However, the unsupervised models perform worse than the supervised models because of the high false-positive rate.
To fill the gap between supervised and unsupervised IDS, the authors in \cite{Seo2018GIDS:Network} proposed a two-stage deep neural network: the first classifier is trained in a supervised manner, whereas the second one is a discriminator in a GAN network and is used for detecting unknown attacks. They evaluated the two classifiers separately, and the combined result was not reported. A new idea presented in \cite{Song2021Self-SupervisedData} is to generate attack samples by an LSTM-based GAN model, and then the generated samples and available normal samples are fed into a DCNN model. The study is promising but achieved low accuracy and needs to be further developed. The authors from \cite{Yang2021MTH-IDS:Vehicles} used tree-based machine learning algorithms and focused on developing a complicated data preprocessing framework to improve the accuracy.
Compared to existing studies related to in-vehicle IDS, our proposed model has some advantages as follows: 1) It is trained end-to-end using a small number of labeled data without any complicated data preprocessing; 2) It can detect both known and unknown attacks with high precision and recall, compared to other models; 3) It processes a new sample within a millisecond, which meets the real-time requirement for the in-vehicle IDS.
\begin{table*}[t]
\centering
\setlength{\leftmargini}{0.4cm}
\caption{Literature reviews of machine-learning-based IDS for the in-vehicle network.}
\begin{tabular}{|M{1.7cm}|M{1.5cm}|M{2cm}|M{2.5cm}|M{3cm}|M{3cm}|}
\hline \textbf{Categories} & \textbf{Research Work} &\textbf{ ML algorithm} & \textbf{Features} & \textbf{Contributions} & \textbf{Limitations} \\
\hline
\multirow{4}{*}{Supervised} & \cite{Kang2016ASecurity} &
\begin{minipage}[t][1.3cm]{\linewidth}
\begin{itemize}
\item DNN
\end{itemize}
\end{minipage}
&
\begin{minipage}[t]{\linewidth}
\begin{itemize}
\item Data payload
\end{itemize}
\end{minipage}
&
\begin{minipage}[t]{\linewidth}
\begin{itemize}
\item Lightweight and \newline fast model
\end{itemize}
\end{minipage} &
\begin{minipage}[t]{\linewidth}
\begin{itemize}
\item Train and test on a \newline simuluation dataset
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Song2020In-vehicleNetwork} &
\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item DCNN
\end{itemize}
\end{minipage}
& \begin{minipage}[t]{\linewidth}
\begin{itemize}\item CAN IDs \end{itemize}
\end{minipage} & \begin{minipage}[t][1.3cm]{\linewidth}
\begin{itemize}
\item Novel data \newline processing technique
\end{itemize}
\end{minipage} & \begin{minipage}[t]{\linewidth}
\begin{itemize}
\item Complex model
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Hossain2020LSTM-basedCommunications} & \begin{minipage}[t][1.3cm]{\linewidth}
\begin{itemize}
\item LSTM
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Best tuned \newline parameters for \newline LSTM-based IDS
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Need a large \newline labeled training \newline dataset
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Javed2021CANintelliIDS:GRU} &
\begin{minipage}[t][1.5cm]{\linewidth}
\begin{itemize}
\item 1-D CNN
\item Attention-based GRU
\end{itemize}
\end{minipage}
& \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Time stamp
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Novel way for \newline features extraction
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Complex model
\item No real-time \newline evaluation
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Moulahi2021ComparativeBus} &\begin{minipage}[t][2cm]{\linewidth}
\begin{itemize}
\item SVM
\item DT
\item RF
\item MLP
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Time stamp
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Short training time
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Inefficient for DoS and fuzzy attacks
\end{itemize}
\end{minipage} \\
\hline
\multirow{4}{*}{Unsupervised} & \cite{DAngelo2020AVehicles} & \begin{minipage}[t][1.8cm]{\linewidth}
\begin{itemize}
\item K-means
\item KNN
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.8cm]{\linewidth}
\begin{itemize}
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Efficient for \newline message level
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Sensitive to noise
\item High computational cost
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Avatefipour2019AnLearning} &\begin{minipage}[t][1.85cm]{\linewidth}
\begin{itemize}
\item OSVM
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Novel \newline meta-heuristic optimization algorithm
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Train and test on a simulation dataset
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Wang2018AHTM} & \begin{minipage}[t][1.3cm]{\linewidth}
\begin{itemize}
\item HTM
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.2cm]{\linewidth}
\begin{itemize}
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Novel distributed anomaly detection system
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item High time \newline complexity
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Ashraf2020NovelSystems} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item LSTM \newline autoencoder
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.5cm]{\linewidth}
\begin{itemize}
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Solve both \newline internal and \newline external attacks
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item No time complexity evaluation
\end{itemize}
\end{minipage} \\
\hline
\multirow{4}{*}{Hybrid} & \cite{Seo2018GIDS:Network} &\begin{minipage}[t][0.8cm]{\linewidth}
\begin{itemize}
\item GAN
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}\item CAN IDs\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Detect unknown \newline attacks
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Low accuracy
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Song2021Self-SupervisedData} &\begin{minipage}[t][1.5cm]{\linewidth}
\begin{itemize}
\item LSTM-based GAN
\item DCNN
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}\item CAN IDs\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Novel approach for \newline data generation \newline and sampling
\end{itemize}
\end{minipage} & \begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Low accuracy
\end{itemize}
\end{minipage} \\
\cline{2-6} & \cite{Yang2021MTH-IDS:Vehicles} &\begin{minipage}[t][1.8cm]{\linewidth}
\begin{itemize}
\item Supervised tree-based models
\item K-means
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item CAN IDs
\item Data payload
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Detect unknown \newline attacks
\end{itemize}
\end{minipage} &\begin{minipage}[t][1.1cm]{\linewidth}
\begin{itemize}
\item Complicated data \newline preprocessing
\end{itemize}
\end{minipage} \\
\hline
\end{tabular}
\label{tab:litrev}
\end{table*}
\section{Experimental results \label{sec:results}}
\subsection{Datasets}
\begin{table}[t!]
\centering
\caption{Car hacking dataset overview}
\begin{tabular}{c|c|c}
\hline Attack type & Normal messages & Injected messages \\
\hline DoS Attack & 3,078,250 (84\%) & 587,521 (16\%) \\
Fuzzy Attack & 3,347,013 (87\%) & 491,847 (13\%) \\
Gear Spoofing & 2,766,522 (82\%) & 597,252 (18\%) \\
RPM Spoofing & 2,290,185 (78\%) & 654,897 (22\%) \\
\hline
\end{tabular}
\label{tab:datasets}
\end{table}
\begin{table}[t!]
\centering
\caption{Preprocessed datasets overview.}
\begin{tabular}{c|c}
\hline Attack type & Frames \\
\hline Normal & 352,767 (62\%) \\
DoS Attack & 37,451 (7\%) \\
Fuzzy Attack & 44,486 (8\%) \\
Gear Spoofing & 65,283 (11\%) \\
RPM Spoofing & 71,372 (12\%) \\
\hline
\end{tabular}
\label{tab:dataframes}
\end{table}
We used the car hacking datasets \cite{Song2020In-vehicleNetwork} produced by the Hacking and Countermeasure Research Lab (HCRL) of Korea University. The dataset was constructed by logging CAN traffic via the OBD-II port of a real vehicle while malfunctioning messages are injected. There are four types of attacks: DoS, fuzzy, spoofing RPM, and spoofing gear information, which are saved in the different comma-separated value files. Table \ref{tab:datasets} shows the details of this dataset.
Each message includes timestamp, CAN ID in HEX, the number of data bytes, 8-byte data, and a flag with two values, which are T for an injected message and R for a normal message. We extracted CAN IDs and transformed them from hexadecimal to a 29-bit representation. Then, the data frame was constructed by stacking 29 sequential samples together as shown in \cref{fig:frame}. In our model, there are 2 classes: normal (0) and abnormal (1). The frame was labeled as abnormal if there was at least one injected message. Table \ref{tab:dataframes} shows the information about data frames after preprocessing.
\subsection{Experiment setup}
With normal frames, we divided the data frames into the training set, validation set, and test set with the percentage of 70\%, 15\%, and 15\%, respectively. The validation set was used for checking the overfitting and tuning hyperparameters during training. The test set was kept hidden until the training progress is complete. To demonstrate that our model can detect intrusion activities with a small number of attack samples, we adjusted the number of attack samples. We used 10\%, 30\%, 50\%, and 70\% of total attack data for training. Moreover, the labeled data comprised only 10\% of the total training data. Table \ref{tab:dataset_structure} displays the details of the training dataset structure.
The computer configuration for the experiment included a 64-bit Intel (R) Core(TM) i7-7700 CPU @ 3.6 GHz and a Geforce GTX 1060 6GB GPU. GPU was used only for accelerating the training process. We trained and tested the proposed model with Tensorflow version 1.15 and Python 3.6.
\subsection{Evaluation metrics}
Because the test set can be imbalanced between normal and abnormal data, we used precision, recall, and F1 score to evaluate the performance of our model. In addition, the error rate (ER) is very important in the in-vehicle IDS. For example, if the IDS results in a high false negative rate, it can be dangerous for the driver; or if the IDS produces a high false positive rate, it can affect the user experience. Our goal is to achieve a high F1 score and a low ER. These metrics can be calculated by using true positive (TP), true negative (TN), false positive (FP), and false negative (FN):
\begin{equation}
\text{Error rate (ER)} = \frac{FP + FN}{TP + TN + FP + FN}
\end{equation}
\begin{equation}
\text{Recall (Rec)} = \frac{TP}{TP + FN}
\end{equation}
\begin{equation}
\text{Precision (Pre)} = \frac{TP}{TP + FP}
\end{equation}
\begin{equation}
\text{F1-score (F1)} = 2\times \frac{Prec \times Rec}{Prec + Recall}.
\end{equation}
\begin{table}[t]
\caption{Training settings}
\begin{threeparttable}
\begin{tabular}{c|c|c|c|c|c}
\hline
\multirow{2}{*}{Train ratio\tnote{*}} & \multicolumn{5}{c}{Labeled} \\
\cline{2-6} & DoS & Fuzzy & Gear & RPM & Total \\
\hline 0.1 & 400 & 450 & 650 & 700 & 2.20k \\
0.3 & 1.20k & 1.35k & 1.95k & 2.10k & 6.60k \\
0.5 & 2.00k & 2.25k & 3.25k & 3.50k & 11k \\
0.7 & 2.80k & 3.15k & 4.55k & 4.90k & 15.4k \\
\hline
\end{tabular}
\begin{tablenotes}
\item[*] The train ratio indicates the number of training samples over the total samples in the dataset.
\end{tablenotes}
\end{threeparttable}
\label{tab:dataset_structure}
\end{table}
\begin{table}[t]
\centering
\caption{Detection results for different amounts of attack data for known attacks}
\begin{threeparttable}
\begin{tabular}{c|c|c|c|c}
\hline Train ratio\tnote{*} & ER & Rec & Prec & F1 \\
\hline 0.1 & 3.2\% & 0.9620 & 0.9999 & 0.9806 \\
0.3 & 0.9\% & 0.9893 & 0.9997 & 0.9945 \\
0.5 & 1.0\% & 0.9821 & 0.9985 & 0.9902 \\
0.7 & 0.4\% & 0.9899 & 0.9996 & 0.9947 \\
\hline
\end{tabular}
\begin{tablenotes}
\item[*] The train ratio indicates the number of training samples over the total samples in the dataset.
\end{tablenotes}
\end{threeparttable}
\label{tab:detection_result}
\end{table}
\begin{table}[t]
\centering
\caption{Detection results for different amounts of labeled data for known attacks}
\begin{threeparttable}
\begin{tabular}{c|c|c|c|c}
\hline Label ratio\tnote{*} & ER & Rec & Prec & F1 \\
\hline 0.1 & 0.4\% & 0.9899 & 0.9996 & 0.9947 \\
0.2 & 0.2\% & 0.9942 & 0.9998 & 0.9970 \\
0.3 & 0.2\% & 0.9958 & 0.9998 & 0.9978 \\
0.4 & 0.1\% & 0.9972 & 0.9997 & 0.9984 \\
\hline
\end{tabular}
\begin{tablenotes}
\item[*] The label ratio indicates the number of labeled samples over the number of training samples.
\end{tablenotes}
\end{threeparttable}
\label{tab:labeled_detection_results}
\end{table}
\subsection{Detection for known attacks}
In this section, the results of detection for known attacks are investigated. First, we attempt with a different amount of attack data where only $10\%$ of the data is labeled. Next, we also test the ability to detect the model when the number of labeled training data is increased. Finally, we compare our scheme with other supervised models.
\subsubsection{Results for known attacks}
Table \ref{tab:detection_result} shows the result of detection phase with different settings. The training ratio is defined as the ratio of the number of training samples to the total number of samples in the dataset. Therefore, the greater the number of samples, the better the detection results. For example, the ER decreases by 2.8\% if we change the training ratio from 10\% to 70\%. This is because 10\% of the total data is a very small number, which is not sufficient for the model to learn. The model achieves the best result of 0.4\% of ER when we use 70\% of the total amount of attack data for training. This result is very impressive because we only used 10\% labeled data. Moreover, the results can be improved if we increase the labeled data, which will be shown in the following experiment.
To evaluate the impact of the amount of labeled data, we used $70\%$ of total attack data for training and adjusted the labeled ratio with $10\%$, $20\%$, $30\%$, and $40\%$. We defined the labeled ratio as the fraction between the number of labeled training samples over the total training samples. As shown in Table \ref{tab:labeled_detection_results}, the more the number of labeled data, the lower ER and the higher F1 score. Moreover, the recall increased significantly from 0.9899 to 0.9972 if we increased the labeled ratio from 0.1 to 0.4, whereas the precision was considerably stable. The results also indicated that the false negative was reduced when more labeled attack data were fed. The model achieved $0.1\%$ and 0.9984 in terms of ER and F1 score with $70\%$ of total data in which $40\%$ of them were labeled. The result is very competitive to other supervised methods, which will be discussed in the next section.
\subsubsection{Comparison with other supervised methods}
Table \ref{tab:cmp_ml} presents the comparison between our model and other machine learning algorithms. The proposed model is trained with 70\% of total attack data, of which 40\% of the data were labeled. In addition, the other supervised models were trained with 100\% labeled data. The data processing for all these models is the same. We chose these models for diversification purposes: SVM for a kernel-based model, DT for a tree-based model, ANN for a neural network model, and DCNN for a deep learning model.
First, compared to the simple models, such as SVM, DT, and ANN, the results show that our model achieved the lowest ER and the highest F1 score. Most traditional machine learning models have low recall because they usually suffer from imbalanced classes. DT model performs the worst, with an ER of 1.77\% and an F1 score of approximately 0.98.
In addition, the ERs of the SVM and ANN models are 0.21\% and 0.15\%, respectively. Although the results of the two models are slightly worse than ours, it is noticeable that they are trained with 100\% labeled data, whereas our model used only 40\% of them. The next step is the comparison between the proposed model and the DCNN model, which is the state-of-the-art for the in-vehicle IDS.
Interestingly, there is no significant difference between the CAAE model and the DCNN model, with 0.03\% of ER and 0.0007 of F1 score. It should be noted that our model used only 40\% labeled data, whereas DCNN used 100\% labeled data. The result suggests that the CAAE model can save time and effort for collecting and labeling data considerably, particularly when a new attack occurs. Moreover, the proposed model can detect unknown attacks, whereas DCNN and other traditional machine learning models cannot, because these models are trained in supervised learning, which can only classify patterns existing in training samples.
\begin{table}[t]
\centering
\caption{Comparison to other supervised methods for known attacks}
\begin{threeparttable}
\begin{tabular}{M{2cm}|c|c|c|c}
\hline Model & ER & Rec & Prec & F1 \\
\hline SVM & 0.21\% & 0.9947 & 0.9998 & 0.9972 \\
Decision Tree & 1.77\% & 0.9799 & 0.9740 & 0.9770 \\
ANN & 0.15\% & 0.9962 & 0.9999 & 0.9980 \\
DCNN \cite{Song2020In-vehicleNetwork} & 0.07\% & 0.9984 & 0.9998 & 0.9991 \\
Ours \tnote{*} & 0.1\% & 0.9972 & 0.9997 & 0.9984 \\
\hline Compared to the best & 0.03\% & -0.0012 & -0.0010 & -0.0007 \\
\hline
\end{tabular}
\begin{tablenotes}
\item[*] Note: All the supervised models were trained with 100\% labeled samples, whereas our model used only 40\% of them.
\end{tablenotes}
\end{threeparttable}
\label{tab:cmp_ml}
\end{table}
\begin{table*}[ht]
\centering
\caption{Detection results for unknown attacks}
\begin{tabular}{c|c|c|c|c|c|c}
\hline
\multirow{2}{*}{Unknown attack} & \multicolumn{3}{c|}{Unknown Results} & \multicolumn{3}{c}{Known Results}\\
\cline{2-7} & Rec & Prec & F1 & Rec & Prec & F1 \\
\hline \centering DoS & 0.9823 & 0.9992 & 0.9907 & 0.9865 & 0.9999 & 0.9931 \\
\centering Fuzzy & 0.8426 & 0.9999 & 0.9145 & 0.9953 & 1.0 & 0.9976 \\
\centering Gear Spoofing & 0.9978 & 0.9977 & 0.9977 & 0.9674 & 0.9992 & 0.9831 \\
\centering RPM Spoofing & 0.9955 & 0.9984 & 0.9970 & 0.9789 & 0.9994 & 0.9890 \\
\hline
\end{tabular}
\label{tab:unknown_result}
\end{table*}
\begin{table*}[ht]
\centering
\caption{Comparison for unknown attacks}
\begin{tabular}{c|c|c|c|c}
\hline Unknown attack & Model & Rec & Prec & F1 \\
\hline
\multirow{3}{*}{DoS} & Deep Autoencoder \cite{Ca2010StackedCriterion}& \textbf{0.9988} & 0.9127 & 0.9538 \\
& Self-supervised learning \cite{Song2021Self-SupervisedData}& 0.9916 & 0.9751 & 0.9833 \\
& Ours & 0.9823 & \textbf{0.9992} & \textbf{0.9907} \\
\hline
\multirow{3}{*}{Fuzzy} & Deep Autoencoder \cite{Ca2010StackedCriterion}& \textbf{0.9626} & 0.9005 & \textbf{0.9305} \\
& Self-supervised learning \cite{Song2021Self-SupervisedData}& 0.8345 & 0.9445 & 0.8861 \\
& Ours & 0.8426 & \textbf{0.9999} & 0.9145 \\
\hline
\multirow{3}{*}{Gear Spoofing} & Deep Autoencoder \cite{Ca2010StackedCriterion}& 0.8180 & 0.9463 & 0.8775 \\
& Self-supervised learning \cite{Song2021Self-SupervisedData}& 0.8803 & 0.9768 & 0.9261 \\
& Ours & \textbf{0.9978} & \textbf{0.9977} & \textbf{0.9977} \\
\hline
\multirow{3}{*}{RPM Spoofing} & Deep Autoencoder \cite{Ca2010StackedCriterion}& 0.9573 & 0.9573 & 0.9573 \\
& Self-supervised learning \cite{Song2021Self-SupervisedData}& \textbf{0.9997} & 0.9720 & 0.9850 \\
& Ours & 0.9955 & \textbf{0.9984} & \textbf{0.9970} \\
\hline
\end{tabular}
\label{tab:cmp_unknown}
\end{table*}
\begin{table}[ht]
\centering
\caption{Model complexity comparison}
\begin{tabular}{M{1cm}|M{2cm}|M{2cm}|M{2cm}}
\hline \multirow{2}{*}{Model} & \#Parameters & \multicolumn{2}{c}{Inference time (ms)} \\
\cline{3-4}& (million) & GPU & CPU \\
\hline DCNN & 9.80 & 5.00 & 6.70 \\
Ours & 2.15 & 0.63 & 0.69 \\
\hline
\end{tabular}
\label{tab:cmp_dcnn_time}
\end{table}
\subsection{Detection for unknown attacks}
We define an unknown attack as an attack that can stealthily occur in training data but is not labeled. Our model can detect this kind of attack. To test it, we considered each kind of attack as an unknown attack by eliminating its labeled data from the training data. For example, if the unknown attack is DoS, we trained the model with labeled data including normal, fuzzy, gear, and RPM attacks. In this case, we produced unknown results by using the test sets of normal and DoS attack; and known results by using the test sets of normal and the other types. It is similar to three other kinds of attacks. In this experiment, we used only 30\% of total data, of which 10\% of data were labeled.
Table \ref{tab:unknown_result} presents the results of unknown attack testing. Gear and RPM achieved the highest F1 scores of more than 0.99. Consequently, there is a possibility that gear and RPM all belong to the spoofing attacks. When one of them is removed, the model still can detect the other. By contrast, the recalls of DoS and fuzzy are 0.98 and 0.84, respectively. The results can be acceptable because the labeled data were not used. Furthermore, the results are evidence that the proposed model can detect unknown attacks. However, it can be observed that there is a trade-off between known and unknown attacks. There is only an F1 score of DoS attack achieve higher than 0.99 for both unknown and known tests.
We compared our results with those of two other methods which can detect unknown attacks. The first model is a deep autoencoder (DAE) \cite{Ca2010StackedCriterion} trained with only normal data. The model detects attacks by checking whether the reconstruction loss of the new sample is higher than a predefined threshold. The second model is the self-supervised learning method presented in \cite{Song2021Self-SupervisedData}. The model is trained with normal and generated data, which were labeled as attack samples. An additional RPM data is used as hint data to improve the final results.
As shown in \cref{tab:cmp_unknown}, the proposed method achieves the highest F1 score in different kinds of attacks, except the fuzzy attack. In addition, the DAE model achieved an impressive F1 score of 0.9305 for fuzzy attack. However, the other results of this model are not good, with the worst F1 score at 0.8775 for gear spoofing attack. In addition, the DAE model achieved the recall higher than the precision, which means the model usually causes miss alarm alerts. Regarding the self-supervised learning method, the F1 score of the model is low, particularly only 0.8861 for the Fuzzy attack. Although the labeled data of RPM is included in the training set, the model achieved the F1 score of only 0.9850, lower by approximately 1\% compared to our method. Thus, we can conclude that our model is more stable compared to the other models because it has F1 scores higher than 0.99 for all kinds of attacks.
\subsection{Model complexity analysis}
This section presents model complexity analysis in terms of the number of parameters and inference time. We also compared our model with the DCNN model \cite{Song2020In-vehicleNetwork} to demonstrate that our model is lightweight and fast. The number of parameters directly affects the training and testing time of a model. In theory, the lower the number of parameters, the faster the training and testing model. For the empirical aspect, we also consider the inference time, which is related to the detection latency. The study from \cite{Song2020In-vehicleNetwork} indicated that a small batch size and optimizing inference time reduce the overall detection latency. We set the batch size to one frame and measure the inference time of our model. It is noticeable that the number of parameters of our model includes those of encoder, decoder, and two discriminators, whereas only encoder is used for measuring the inference time.
As illustrated in Table \ref{tab:cmp_dcnn_time}, the total number of parameters of the proposed model is 2.15 million, which is nearly one-fifth of those of the DCNN. Moreover,
we decrease the inference time by approximately eight times with GPU processing and ten times with CPU processing.
|
\section{Introduction}
Ferroelectricity was first discovered in Rochelle salt (KNaC$_4$H$_4$O$_6\cdot$4H$_2$O),
a chiral metal-organic compound \cite{PhysRev.17.475,doi:10.1098/rspa.1941.0010,PhysRevB.53.5217,Mo2015}.
Indeed, chirality and ferroelectricity are closely related in symmetry; mirror symmetry breaks in the former, while inversion symmetry breaks in the latter.
Chiral molecules are easier to crystallize in ferroelectric structures and can serve as an attractive system for the search of functional materials \cite{Li5878}.
In this study, we focus on the physical properties of a chiral molecule 1,1’-bi-2-naphthol (BINOL), where two enantiomers are distinguished by the axial chirality \cite{https://doi.org/10.1002/cphc.201800950, molecules21111541}.
By using enantiomerically-pure BINOL, the resultant crystal structure also becomes right- or left-handed depending on the handedness of BINOL.
Some organic crystals composed of BINOL molecules are reported by Lee and Peng \cite{doi:10.1021/cg1004648}.
Among them, we pick up an enantiomorphic system (R)-(+)-BINOL$\cdot$2DMSO [(S)-(-)-BINOL$\cdot$2DMSO], where the crystal contains two dimethylsulfoxide (DMSO) molecules as guests [Fig. \ref{fig:crystal}(a)].
Here, the alignments of the guest DMSO molecules are partially disordered in the chiral framework of BINOL [Fig. \ref{fig:crystal}(b, c)].
Since a DMSO has a large dipole moment of 3.96D \cite{NSRDS}, the dipole-dipole interactions (and/or hydrogen bonding with BINOL) can result in an order-disorder transition of DMSO at low temperatures.
Besides, the ordering of the polar molecules DMSO indicates ferroelectricity or antiferroelectricity in the ordered phase.
Recently, organic ferroelectrics catch more attentions because of their environment-friendly features \cite{doi:10.1126/science.1229675,doi:10.1126/science.aas9330,Horiuchi2008}.
In this paper, we report a ferroelectric transition of BINOL$\cdot$2DMSO studied by the dielectric, thermodynamical, acoustic, and optical techniques.
Besides, the crystal structure at low temperature and the related ordering of DMSO molecules are discussed based on the low-temperature powder x-ray diffraction (XRD).
\section{Experiment}
We purchased enantiomerically-pure BINOL powder from Fuji Molecular Planning Co., Ltd.
Single crystals of (R)-(+)-BINOL$\cdot$2DMSO [(S)-(-)-BINOL$\cdot$2DMSO] were grown by slow evaporation of the saturated DMSO solutions at room temperature.
Transparent single crystals with the typical size of $1\times1\times1$ mm$^3$ were obtained in 2--3 days.
We confirmed that the results of the dielectric measurements were the same for the right- and left-handed crystals.
In the following, we do not distinguish the results for (R)-(+)-BINOL$\cdot$2DMSO or (S)-(-)-BINOL$\cdot$2DMSO.
In this paper, we do not discuss the properties of the racemic compound, which also shows dielectric anomalies at low temperatures.
We measured the dielectric properties on single crystals with applied electric fields along $E||c$ and $E||a$.
We attached two gold wires to the as-grown surfaces by using Ag paint as electrodes with the typical area of $\sim0.5$~mm$^2$.
The dielectric constant $\varepsilon^{\prime}$ and dissipation factor $D$ were measured by the LCR meter Keysight E4980A.
The pyrocurrent ($I$) measurements were performed by using the electrometer Keithley 6517A/B.
We measured a displacement current $I(t)$ induced by sweeping an applied voltage or temperature.
Polarization $P$ was obtained by numerically integrating $I(t)$.
We performed specific-heat measurements by using the relaxation technique of the Physical Property Measurement System (Quantum Design Ltd.).
We used 3 mg of single crystals and attached them to the sample stage by using Apiezon N grease.
We performed sound-velocity measurements by using an ultrasound pulse-echo technique.
We measured the longitudinal ($c_{33}$, $k||u||[001]$) and transverse ($c_{44}$, $k||[001]$, $u||[100]$) mode by using Y-36$^{\circ}$- and X-41$^{\circ}$-cut LiNbO$_3$ transducers, respectively.
We measured at the fundamental frequencies of the transducers, 38 MHz ($c_{33}$ mode) and 18 MHz ($c_{44}$ mode).
We attached two transducers to the $c$ surfaces of the single crystal by using Thiokol.
Temperature dependence of the echo pattern was recorded by a digital storage oscilloscope (LeCroy HDO4104A).
The phase of the 0th echo was numerically analyzed and the relative change of the sound velocity $\Delta v/v_0$ was obtained.
The estimated error for the sound velocity was around 5\%, while the error for $\Delta v/v_0$ was around 1\%.
We performed optical absorption spectroscopy with incident light along the $c$ axis by using JASCO V-570.
The sample thickness was 2.6 and 1.1 mm for the measurement run \#1 and \#2, respectively.
The sample was cooled in a He-gas-flow optical cryostat.
In addition to the spectroscopy, we also observed by eyes how the outlook of the crystal changes at low temperatures.
We performed low-temperature powder XRD measurements by using an x-ray diffractometer (SmartLab Rigaku) with Cu $\rm{K\alpha}_{1}$ radiation monochromated by a Ge(111)-Johansson-type monochromator.
We ground single crystals of BINOL$\cdot$2DMSO to paste in a mortar.
Here, we added a small amount of DMSO to avoid deterioration because the BINOL$\cdot$2DMSO crystals can lose the guest DMSO molecules in a vacuum.
We analyzed the XRD powder patterns by using PDXL2 software (Rigaku).
\section{Results}
\subsection{Dielectric properties}
Figure \ref{fig:eps}(a) shows the temperature dependence of the dielectric constant $\varepsilon^{\prime}$ at various frequencies.
A clear anomaly is observed at $T_\mathrm{c1}=190$ K, indicating a ferroelectric or antiferroelectric phase transition.
Slightly below $T_\mathrm{c1}$, a broad peak in the dissipation factor $D$ is observed at a certain temperature $T_\mathrm{peak}$ [Fig. \ref{fig:eps}(b)] probably due to the domain dynamics \cite{PhysRevB.55.16159}.
The frequency dependencies of $T_\mathrm{c1}$ and $T_\mathrm{peak}$ with the error bar of FWHM are plotted in Fig.~\ref{fig:eps}(c).
$T_\mathrm{peak}$ shifts from 190 K to 120~K towards a low-frequency limit, while the transition temperature $T_\mathrm{c1}$ is frequency independent.
This frequency dependence indicates that the domain motion is frozen at around 120~K.
Noteworthy, a small kink is observed in the dielectric constant at $T_\mathrm{c2} \sim 120$~K measured at 1~kHz, which indicates the freezing temperature of domains or another phase transition.
The inset of Fig. \ref{fig:cw} shows the inverse dielectric constant 1/$\varepsilon^{\prime}$ measured at 1~MHz.
Blue (red) line shows the fitting result below (above) $T_\mathrm{c1}$ based on the Curie-Weiss law
\begin{equation}
1/{\varepsilon}^{\prime} = |T-T_{\mathrm{c1}}|/C,
\label{eq:CW}
\end{equation}
where $C$ is the Curie constant.
Both curves below and above $T_\mathrm{c1}$ are well fitted by Eq.~(\ref{eq:CW}).
However, ${\varepsilon}^{\prime}(T)$ below $T_\mathrm{c1}$ deviates from the Curie-Weiss law at frequencies lower than 10~kHz [the hump around 160 K in Fig. \ref{fig:eps}(a)].
The frequency dependence of the Curie constant is plotted in Fig. \ref{fig:cw}.
$C(T>T_\mathrm{c1})$ is almost frequency independent, while $C(T<T_\mathrm{c1})$ decreases at higher frequencies, reflecting the domain-wall dynamics.
At a high-frequency limit, $C(T>T_\mathrm{c1})/C(T<T_\mathrm{c1}) \sim 3.8$, which is close to the Landau-theory value of 2.
The estimated Curie constant around 1000 K is in a typical energy scale for ferroelectrics driven by the order-disorder mechanism, NaNO$_2$\cite{doi:10.1143/JPSJ.16.2207} and tri-glycine sulfate (TGS) \cite{PhysRev.107.1255}.
Figure \ref{fig:cole} shows the Cole-Cole diagram of the complex dielectric constant $\varepsilon^\ast =
\varepsilon^\prime -i\varepsilon^{\prime \prime}$ at 170 K.
The frequency dependence of the real and imaginary parts of the dielectric constant shows a single-dome behavior, indicating only one relaxation mechanism is relevant in this frequency range.
The blue curve shows the Cole-Cole law \cite{doi:10.1143/JPSJ.16.2207}
\begin{equation}
\frac{{\varepsilon}^{\ast}-{\varepsilon}^0}{{\varepsilon}^0-{\varepsilon}^{\infty}}
=\frac{1}{1+(2{\pi}f{\tau}i)^{1-\alpha}},
\label{eq:cole}
\end{equation}
where $\varepsilon^0$ and $\varepsilon^\infty$ are the dielectric constants at low and high frequency limits, and $\tau$ is the relaxation time.
$\alpha$ is a deviation parameter from the Debye model.
With the parameters of $\alpha=0.3$ and $\tau=5$ ms, the experimental plots are reasonably fitted.
The relaxation time exponentially increases towards 120 K as seen from the temperature dependence of $T_\mathrm{peak}$ [Fig.~\ref{fig:eps}(c)].
Next, we show the polarization versus electric field ($P-E$) curves measured at the frequency of 5~mHz [Fig. \ref{fig:pe}(a)].
Single hysteresis loops are observed, meaning that the phase below $T_\mathrm{c1}$ is ferroelectric, not antiferroelectric.
The temperature dependence of the coercive field $E_\mathrm{c}$ is plotted in Fig. \ref{fig:pe}(b) right axis.
$E_\mathrm{c}$ nonlinearly increases towards low temperatures and goes over 400~kV/m already at 131~K.
The uncommon shape of the hysteresis loop for 180 K [orange curve in Fig. \ref{fig:pe}(a)] might be related to the domain-wall motion observed in the dielectric measurements (Fig. \ref{fig:eps}).
Figure \ref{fig:pe}(c) shows the result of pyrocurrent measurement with the temperature sweep rate of $\sim+0.1$ K/s.
In addition to the clear anomaly at $T_\mathrm{c1}=190$ K, a sharp peak is observed at $T_\mathrm{c2}=125$ K, indicating another phase transition.
By integrating $I(t)$ and assuming no spontaneous polarization at room temperature, the polarization as a function of temperature is obtained as Fig. \ref{fig:pe}(b).
The polarization starts to increase at $T_\mathrm{c1}$ and almost saturates below $T_\mathrm{c2}$ at the value of $\sim 20$ mC/m$^2$.
This polarization value is again comparable to other ferroelectric materials driven by the order-disorder transition of polar molecules \cite{doi:10.1143/JPSJ.16.2207,PhysRev.107.1255}.
\subsection{Specific heat}
Figure \ref{fig:cp}(a) shows the specific heat as a function temperature.
Two sets of results coincide well, including the anomalies at $T_\mathrm{c1}=190$ K and $T_\mathrm{c2}= 125$ K.
The $\lambda$-shaped anomalies suggest that these transitions are of second-order.
Indeed, all other measurements do not show sizable hysteresis.
The entropy change related to these transitions $\Delta S$ are estimated in Fig.~\ref{fig:cp}(b).
Here, the phonon contribution is subtracted as the gray curve in Fig. \ref{fig:cp}(a), and the entropy is normalized for one DMSO molecule.
The relatively large $\Delta S$ at $T_\mathrm{c1}$ indicates that the ferroelectric transition is driven by the order-disorder mechanism \cite{doi:10.1143/JPSJ.20.2180}.
Here, we discuss the theoretical entropy change for an order-disorder transition of DMSO molecules in this system.
Previous structural analysis suggests that there are two molecular alignments for each DMSO molecule at room temperature, with the configurational probabilities of major:minor $\sim$ 2/3:1/3 [Fig. \ref{fig:crystal}(b, c)] \cite{doi:10.1021/cg1004648}.
The minor configuration is preferred for the hydrogen bonding between DMSO and BINOL.
However, in this packing geometry, the DMSO molecule is subjected to larger stress from the BINOL framework, observed as the stretching of the S-C bonding.
By assuming this configurational degree of freedom freezes at $T_\mathrm{c1}$, the expected entropy change per one DMSO is calculated as $\Delta S = (1/3\mathrm{ln}(1/3) + 2/3\mathrm{ln}(2/3))k_\mathrm{B} = 0.64k_\mathrm{B}$.
This value is comparable to the entropy change from 120 K to 190 K, indicating that the fluctuation of the DMSO molecules gradually freezes towards $T_\mathrm{c2}$.
This picture is in line with the temperature dependence of $T_\mathrm{peak}$, which also suggests that the dynamics freezes towards $T_\mathrm{c2}$.
Here, we note on the possible error range of our results.
We notice that the sample becomes opaque and the mass of the sample slightly decreases after measuring the specific heat in a vacuum environment.
This indicates that the guest DMSO molecules can escape from the crystal with the help of vacuum.
We estimate that possible errors of the heat capacity and entropy change are up to $\pm$~5\% and $\pm$~10\%, respectively.
The estimated error does not affect the above discussions.
\subsection{Ultrasound velocity}
Figure \ref{fig:sound} shows the temperature dependence of the sound velocity of the $c_{33}$ and $c_{44}$ acoustic modes.
A clear anomaly is observed at $T_\mathrm{c1} = 190$ K, which is consistent with other measurements.
The phase transition at $T_\mathrm{c2}$ is not detected with these acoustic modes, indicating that the corresponding strains $\epsilon_{zz}$ and $\epsilon_{zx}$ ($\epsilon_{yz}$) are not involved at this transition.
The slope change at $T_\mathrm{c1}$ is reminiscent of the change of the polarization in Fig. \ref{fig:pe}, indicating that the developing dipole-dipole interaction contributes to the stiffness.
We note that the ultrasound frequency of 10--100~MHz is much higher than the relaxation time estimated by the Cole-Cole plot ($\sim5$~ms), and the effect of domain dynamics is not observed in the sound velocity.
We should note that the echo signal becomes discontinuously worse below 150 K.
This is probably related to the domain formation below the ferroelectric transition and sound scattering at the domain boundaries.
The curves in Fig. \ref{fig:sound} show the results obtained by two or three runs of measurements, which are well reproduced.
The effect of the domain boundaries is observed in the optical spectroscopy as well.
\subsection{Optical spectroscopy}
Figure \ref{fig:abs}(a) shows the absorption spectra of the measurement run \#1 (2.6 mm thickness) at temperatures from 300 K to 5 K.
The absorbance at 300 K in the visible range is assumed to be zero \cite{https://doi.org/10.1002/cphc.201800950} and the background due to light scattering is subtracted.
The absorption peaks below 1 eV come from the intramolecular vibrational modes and their higher harmonics \cite{molecules21111541,Wallace2015}, which are almost temperature independent.
The absorption edge at 3.3 eV corresponds to the bandgap, consisting mainly of the $\pi-\pi^*$ transitions of BINOL \cite{doi:10.1021/cg1004648}.
The temperature dependence of the bandgap is plotted in Fig.~\ref{fig:abs}(b).
The bandgap increases towards lower temperatures and saturates at $T_\mathrm{c2}$.
This probably reflects the change of the packing structure which affects the intermolecular hybridization \cite{PhysRevB.101.165102,singleton}.
The absorbance change at 2.5 eV, where no optical absorption exists, is plotted in Fig.~\ref{fig:abs}(c).
The drastic change occurs at around 120~K, where the absorbance increases in all energy ranges.
By eye observations, we find that the crystal becomes opaque and the light scattering increases at this temperature.
The absorbance tends to be larger in the higher photon energy (shorter wavelength), indicating Rayleigh scattering \cite{doi:10.1080/00150193.2011.577673}.
Based on the other measurements above, we speculate that this light scattering is due to crystallographic domains, probably related to the structural transition from tetragonal to lower-symmetry phase (orthorhombic or monoclinic) at $T_\mathrm{c2}$.
The mismatch of the refractive index at the domain boundaries causes strong light scattering \cite{PhysRevLett.112.247201}.
In contrast, the light scattering does not change at $T_\mathrm{c1}$.
This indicates that the ferroelectric transition at $T_\mathrm{c1}$ does not form domain boundaries by lowering the crystallographic symmetry from tetragonal.
\subsection{Powder XRD}
We have performed the low-temperature powder XRD measurements at 220, 150, 100, and 50 K.
The full diffraction patterns are presented in Supplemental Material (SM) \cite{suppl}.
The diffraction pattern at 220 K can be reasonably fitted by the reported crystal structure at room temperature \cite{doi:10.1021/cg1004648}.
Subsequent symmetry lowering is observed in the low-temperature results.
Figures \ref{fig:101}(a-d) show the diffraction patterns around the $101$ and $114$ peaks.
The $101$ and $114$ peaks do not split at 150 K [Figs. \ref{fig:101}(a, b)], however, an additional peak appears at 10.2 deg. at 150~K.
This indicates that the transition at $T_\mathrm{c1}$ changes the structure from higher-symmetry tetragonal to lower-symmetry tetragonal ones.
In the low temperature phase below $T_\mathrm{c2}$ [Figs. \ref{fig:101}(c, d)], clear splittings of $101$ and $114$ peaks are observed, indicating that the symmetry changes from tetragonal to monoclinic (or triclinic).
Indeed, the overall diffraction patterns of the low-$T$ phase can be fitted by a monoclinic structure (unique axis $c$).
Figures~\ref{fig:lattice}(a-d) summarize the lattice constants, monoclinic angle, and unit-cell volume.
For convenience, we call the tetragonal paraelectric phase as PE, the tetragonal ferroelectric phase as FE1 ($T_\mathrm{c2}<T<T_\mathrm{c1}$), and the monoclinic ferroelectric phase as FE2 ($T<T_\mathrm{c2}$).
Next, we discuss the space groups of FE1 and FE2.
The candidate space groups can be significantly reduced to the product set of the chiral and polar space groups, lacking the mirror and inversion symmetries, respectively.
$P4$, $P4_1$, $P_2$, $P4_3$, $I4$, and $I4_1$ are the candidate space groups for the ferroelectric tetragonal structure (FE1), while $P112$, $P112_1$, and $A112$ are the candidates for the ferroelectric monoclinic structure (FE2), satisfying these conditions.
It is notable that the orthogonal structures do not satisfy both chiral and polar conditions simultaneously.
From these candidates, we narrow down the space group based on the extinction rule.
Figure \ref{fig:101}(a) indicates that FE1 shows a $100$ peak which is canceled in PE.
The appearance of this index narrows the candidates down to $P4$, $P4_1$, $P_2$, $P4_3$.
Among the $00l$ peaks, only the $004$ and $008$ peaks are observed for PE1, indicating $4_1$ ($4_3$) screw axis along the $c$ axis (see SM \cite{suppl}).
Thus, we conclude that the space group of FE1 is $P4_1$ ($P4_3$).
These are the maximal subgroups from $P4_12_12$ ($P4_32_12$) of PE.
The active irreducible representation (IR) describing the crystal-symmetry breaking is $A_2$.
Figure \ref{fig:101}(c) indicates that PE2 shows the $010$ peak, but not the $100$ peak expected at 10.06 deg.
Among the candidate space groups, only $P112_1$ satisfies this condition.
Thus, we conclude the space group of FE2 as $P112_1$.
Again, $P112_1$ is the maximal subgroup from $P4_1$ ($P4_3$) with the active IR of $B$.
The group-subgroup relationship is consistent with the observed second-order transitions at $T_\mathrm{c1}$ and $T_\mathrm{c2}$.
\section{Discussion}
Here, we discuss the consistency of the experimental results.
The ultrasound velocity clearly detects the transition at $T_\mathrm{c1}$, while does no the one at $T_\mathrm{c2}$.
This reflects the soft mode at $T_\mathrm{c2}$.
From FE1 to FE2, $\epsilon_{xy}$ and $\epsilon_\mathrm{T}=(\epsilon_{xx}-\epsilon_{yy})/2$ belonging to the IR $B$ are the active strains.
The expected acoustic soft modes are the $c_{66}$ and $c_\mathrm{T}=(c_{11}-c_{12})/2$.
In this study, we have only measured the $c_{33}$ and $c_{44}$ modes because of the crystal habit.
Therefore, the insensitivity at $T_\mathrm{c2}$ is consistent with the symmetry arguments.
Next, we discuss the results of the optical spectroscopy.
The drastic increase of light scattering is observed at $T_\mathrm{c2}$, while it does not change at $T_\mathrm{c1}$.
The difference indicates that the crystallographic domain is the origin of the light scattering.
At $T_\mathrm{c2}$, the unit cell deforms from tetragonal to monoclinic, leading to two domain structures with the shear strains.
Such domain boundaries cause the refractive-index mismatch and lead to the opaque appearance of the crystal.
A similar effect, the decrease of the acoustic signal, is observed in the ultrasound measurement as well.
On the other hand, the transition at $T_\mathrm{c1}$ does not change the tetragonal unit cell, although the ferroelectric domain structures are formed as indicated by the hysteresis loop [Fig.~\ref{fig:pe}(a)].
Such ferroelectric domains would also be observed by using a polarizing microscope.
Here, we comment on the order-disorder transition of DMSO molecules at $T_\mathrm{c1}$.
With the symmetry lowering from $P4_12_12$ to $P4_1$, $2_1$ and $2$ axes in the $a$-$b$ plane are lost.
Because of this symmetry lowering, the nearest neighbor DMSO sites [Fig. \ref{fig:crystal}(b, c)] become nonequivalent.
Most probably, at $T_\mathrm{c1}$, one DMSO site settles in the major configuration and another in the minor one.
Detailed studies on the crystal structures at low temperatures are left for future works.
Last, we comment on the possible formation of electric Bloch skyrmions in this compound.
The point group changes from $422$ to $4$ at $T_\mathrm{c1}$ satisfies the necessary condition to realize electric Bloch skyrmions in a bulk system \cite{PhysRevB.102.024110}.
The strange hysteresis curve in the $P-E$ curve [Fig.~\ref{fig:pe}(a), 180 K] and the broad peak of $D$ [Fig.~\ref{fig:eps}] might be related to this kind of ferroelectric domains.
\section{Conclusion}
We have characterized the physical properties of a chiral organic compound BINOL$\cdot$2DMSO.
We have found two second-order phase transitions at $T_\mathrm{c1}=190$ K and $T_\mathrm{c2}=125$ K, where the space group changes as $P4_12_12$ ($P4_32_12$) $\rightarrow$ $P4_1$ ($P4_3$) $\rightarrow$ $P112_1$.
The transition at $T_\mathrm{c1}$ is characterized by the spontaneous polarization driven by the order-disorder transition of the DMSO molecules.
At $T_\mathrm{c2}$, the crystal symmetry is further lowered by the monoclinic distortion.
Because of the domain boundaries, the transmission of light and acoustic waves is strongly disturbed in the monoclinic phase.
In this paper, we report a new member of the chiral organic ferroelectrics.
Although the first ferroelectricity had been reported on a chiral system (Rochelle salt), chiral ferroelectrics in purely organic systems are still seldom \cite{Horiuchi2008,li2022highest,Li5878,acsnano7b07090}.
Besides, most of the reported chiral organic ferroelectrics show hysteretic behavior in temperature dependence.
BINOL$\cdot$2DMSO, which shows a second-order ferroelectric transition, could be an attractive platform to deepen the understanding of organic ferroelectrics.
\begin{acknowledgment}
We thank M. Akatsu for sharing the ultrasound measurement software.
This work was partly supported by JSPS KAKENHI, Grant-in-Aid for Scientific Research (No. 20K14403).
\end{acknowledgment}
\bibliographystyle{jpsj}
|
\section{Continuous Generative Cellular Automata}
\begin{figure}[t]
\includegraphics[width=\textwidth]{./figures/method_overview}
\centering
\vspace{-2em}
\caption[] {
Overview of our method.
The implicit function of continuous shape can be encoded as sparse voxel embedding $s$ and decoded back (left).
%
The colors in the sparse voxel embedding represent the clustered labels of latent code $z_c$ for each cell $c$. %
The sampling procedure of cGCA (right) involves $T$ steps of sampling the stochastic transition kernel $p_\theta$, followed by $T'$ mode seeking steps which remove cells with low probability.
%
%
From the final sparse voxel embedding $s^{T + T'}$, the decoder can recover the implicit representation for the complete continuous shape.
%
}
\label{fig:method_overview}
\end{figure}
\vspace{-1em}
In Sec.~\ref{sec:embedding}, we formally introduce an extension of sparse occupancy voxels to represent continuous geometry named sparse voxel embedding, where each occupied voxel contains latent code representing local implicit fields.
We train an autoencoder that can compress the implicit fields into the embeddings and vice versa.
Then we present the sampling procedure of cGCA that generates 3D shape in Sec.~\ref{sec:sampling}, which is the inference step for shape completion.
Sec.~\ref{sec:training} shows the training objective of cGCA, which approximately maximizes the variational lower bound for the distribution of the complete continuous geometry.
\vspace{1.0em}
\subsection{Sparse Voxel Embedding}
\label{sec:embedding}
In addition to the sparse occupied voxels of GCA, the state of cGCA, named sparse voxel embedding, contains the associated latent code, which can be decoded into continuous surface.
Formally, the state $s$ of cGCA is defined as a set of pair of binary occupancy $o_c$ and the latent code $z_c$, for the cells $c$ in a three-dimensional grid $\mathbb{Z}^3$
\begin{equation}
s = \{(c, o_c, z_c) | c \in \mathbb{Z}^3, o_c \in \{0, 1\}, z_c \in \mathbb{R}^K \}.
\label{eq:state_cgca}
\end{equation}
Similar to GCA, cGCA maintains the representation sparse by storing only the set of occupied voxels and their latent codes, and sets $z_c=0$ if $o_c=0$.
The sparse voxel embedding $s$ can be converted to and from the implicit representation of local geometry with neural networks, inspired by the work of \cite{peng2020conv_onet}, \cite{chabra2020deep_local_shapes}, and \cite{chibane20ifnet}.
We utilize the (signed) distance to the surface as the implicit representation, and use autoencoder for the conversion.
The encoder $g_\phi$ produces the sparse voxel embedding $s$ from the coordinate-distance pairs $P = \{(p, d_p) | p \in \mathbb{R}^3, d_p \in \mathbb{R} \}$, where $p$ is a 3D coordinate and $d_p$ is the distance to the surface, $s = g_\phi(P)$.
The decoder $f_\omega$, on the other hand, regresses the local implicit field value $d_q$ at the 3D position $q \in \mathbb{R}^3$ given the sparse voxel embedding $s$, $\hat{d}_q=f_\omega(s, q)$ for given coordinate-distance pairs $Q = \{(q, d_q)| q\in \mathbb{R}^3, d_q \in \mathbb{R}\}$.
The detailed architecture of the autoencoder is described in Appendix~\ref{app:sec:implementation_details}, where the decoder $f_\omega$ generates continuous geometry by interpolating hierarchical features extracted from sparse voxel embedding.
An example of the conversion is presented on the left side of Fig.~\ref{fig:method_overview}, where the color of the sparse voxel embedding represents clustered labels of the latent codes with k-means clustering~(\cite{hartigan1979kmeans}).
Note that the embedding of a similar local geometry, such as the seat of the chair, exhibits similar values of latent codes.
The parameters $\phi, \omega$ of the autoencoder are jointly optimized by minimizing the following loss function:
\begin{equation}
\label{eq:autoencoder_loss}
\mathcal{L}(\phi, \omega) = \frac{1}{|Q|} \sum_{(q, d_q) \in Q}{|f_\omega(s, q) - \max\left(\min\left(\frac{d_q}{\epsilon}, 1\right), -1\right)|} + \beta \frac{1}{|s|}\sum_{c \in s}{\|z_c\|},
\end{equation}
where $s=g_\phi(P)$ and $\epsilon$ is the size of a single voxel.
The first term in Eq.~(\ref{eq:autoencoder_loss}) corresponds to minimizing the normalized distance and the second is the regularization term for the latent code weighted by hyperparameter $\beta$.
Clamping the maximum distance makes the network focus on predicting accurate values at the vicinity of the surface (\cite{park2019deepsdf, chibane2020ndf}).
\subsection{Sampling from continuous Generative Cellular Automata}
\label{sec:sampling}
The generation process of cGCA echos the formulation of GCA (\cite{zhang2021gca}), and repeats $T$ steps of sampling from the transition kernel %
to progressively grow the shape.
Each transition kernel $p(s^{t + 1} | s^{t})$ is factorized into cells within the local neighborhood of the occupied cells of the current state, $\mathcal{N}(s^t)=\{ c' \in \mathbb{Z}^3 \mid d(c, c') \leq r, c\in s^t \}$ \footnote{We use the notation $c \in s$ if $o_c = 1$ for $c\in \mathbb{Z}^3$ to denote occupied cells.} given a distance metric $d$ and the radius $r$:
\begin{equation}
p(s^{t + 1} | s^{t})
= \prod_{c \in \mathcal{N}(s^t)} p_\theta(o_c, z_c| s^t) \label{eq:transition_kernel} \\
= \prod_{c \in \mathcal{N}(s^t)} p_\theta(o_c|s^t) p_\theta(z_c|s^t, o_c).
\end{equation}
Note that the distribution is further decomposed into the occupancy $o_c$ and the latent code $z_c$, %
where we denote $o_c$ and $z_c$ as the random variable of occupancy and latent code for cell $c$ in state $s^{t + 1}$.
Therefore the shape is generated by progressively sampling the occupancy and the latent codes for the occupied voxels which are decoded and fused into a continuous geometry.
The binary occupancy is represented with the Bernoulli distribution
\begin{equation}
p_\theta(o_c|s^t) = Ber(\lambda_{\theta, c}),
\end{equation}
where $\lambda_{\theta, c} \in [0, 1]$ is the estimated occupancy probability at the corresponding cell $c$.
With our sparse representation, the distribution of the latent codes is
\begin{equation}
p_\theta(z_c|s^t, o_c) =
\begin{cases}
\delta_0 &\text{if } o_c = 0\\
N(\mu_{\theta, c}, \sigma^t {\bm{I}}) &\text{if } o_c = 1.
\end{cases}
\end{equation}
$\delta_0$ is a Dirac delta distribution at $0$ indicating that $z_c = 0$ when $o_c=0$.
For the occupied voxels ($o_c=1$), $z_c$ follows the normal distribution with the estimated mean of the latent code $\mu_{\theta, c} \in \mathbb{R}^K$ and the predefined standard deviation $\sigma^t{\bm{I}}$, where $\sigma^t$ decreases with respect to $t$.
\paragraph{Initial State.}
Given an incomplete point cloud, we set the initial state $s^0$ of the sampling chain by setting the occupancy $o_c$ to be 1 for the cells that contain point cloud and associating the occupied cells with a latent code sampled from the isotropic normal distribution.
However, the input can better describe the provided partial geometry if we encode the latent code $z_c$ of the occupied cells with the encoder $g_\phi$.
The final completion is more precise when all the transitions $p_\theta$ are conditioned with the initial state containing the encoded latent code.
Further details are described in Appendix~\ref{app:sec:implementation_details}.
\paragraph{Mode Seeking.}
While we effectively model the probabilistic distribution of multi-modal shapes, the final reconstruction needs to converge to a single coherent shape.
Na\"ive sampling of the stochastic transition kernel in Eq.~(\ref{eq:transition_kernel}) can include noisy voxels with low-occupancy probability.
As a simple trick, we augment \textit{mode seeking steps} that determine the most probable mode of the current result instead of probabilistic sampling.
Specifically, we run additional $T'$ steps of the transition kernel but we select the cells with probability higher than 0.5 and set the latent code as the mean of the distribution $\mu_{\theta, c}$.
The mode seeking steps ensure that the final shape discovers the dominant mode that is closest to $s^T$ as depicted in Fig.~\ref{fig:method_overview}, where it can be transformed into implicit function with the pretrained decoder $f_w$.
\subsection{Training Continuous Generative Cellular Automata}\label{sec:training}
We train a homogeneous transition kernel $p_\theta(s^{t + 1} | s^t)$, whose repetitive applications eventually yield the samples that follow the learned distribution.
However, the data contains only the initial $s^0$ and the ground truth state $x$, and we need to emulate the sequence for training.
We adapt infusion training (\cite{bordes2017infusion}), which induces the intermediate transitions to converge to the desired complete state.
To this end, we define a function $G_x(s)$ that finds the valid cells that are closest to the complete shape $x$ within the neighborhood of the current state $\mathcal{N}(s)$:
\begin{equation}
G_{x}(s) = \{\mathrm{argmin}_{c \in \mathcal{N}(s)} d(c, c') \mid c' \in x\}.
\end{equation}
Then, we define the infusion kernel $q^t$ factorized similarly as the sampling kernel in Eq.~(\ref{eq:transition_kernel}):
\begin{equation}
q_\theta^t(s^{t + 1} | s^{t}, x)
= \prod_{c \in \mathcal{N}(s^t)} q^t_\theta(o_c, z_c| s^t, x)
= \prod_{c \in \mathcal{N}(s^t)} q^t_\theta(o_c|s^t, x) q^t_\theta(z_c|s^t, o_c, x). \label{eq:infusion_kernel}
\end{equation}
The distributions for both $o_c$ and $z_c$ are gradually biased towards the ground truth final shape $x$ with the infusion rate $\alpha^t$, which increases linearly with respect to time step, i.e., $\alpha^t = \max(\alpha_1t + \alpha_0, 1)$, with $\alpha_1 > 0$:
\begin{equation}
q^t_\theta(o_c|s^t, x) = Ber( (1 - \alpha^t) \lambda_{\theta, c} + \alpha^t \mathds{1}[c \in G_{x}(s^t) ]),
\end{equation}
\begin{equation}
q^t_\theta(z_c|s^t, o_c, x) =
\begin{cases}
\delta_{0} &\text{if } o_c = 0 \\
{N}((1 - \alpha^t) \mu_{\theta, c} + \alpha^t z^x_c, ~\sigma^t {\bm{I}}) &\text{if } o_c = 1.
\end{cases}
\end{equation}
Here $\mathds{1}$ is an indicator function, and we will denote $o^x_c, z^x_c$ as the occupancy and latent code of the ground truth complete shape $x$ at coordinate $c$.
cGCA aims to optimize the log-likelihood of the ground truth sparse voxel embedding $\log p_\theta (x)$.
However, since the direct optimization of the exact log-likelihood is intractable, we modify the variational lower bound using the derivation of diffusion-based models (\cite{sohl2015nonequil}):
\begin{align}
\label{eq:elbo}
\log p_\theta(x)
&\ge \sum_{s^{0: T-1}} q_\theta(s^{0:T-1}|x) \log \frac{p_\theta(s^{0:T-1}, x)}{q_\theta(s^{0:T-1}|x)} \\
&=
\underbrace{\log \frac{p(s^0)}{q(s^0)}}_{\mathcal{L}_{\mathrm{init}}}
+ \sum_{0 \le t < T - 1} \underbrace{-D_{KL}(q_\theta(s^{t + 1} | s^t, x) \| p_\theta(s^{t + 1} | s^t))}_{\mathcal{L}_t} + \underbrace{\mathbb{E}_{q_\theta}[\log p_\theta(x | s^{T - 1})]}_{\mathcal{L}_{\mathrm{final}}} , \notag
\end{align}
where the full derivation is in Appendix~\ref{app:elbo_derivation}.
We now analyze $\mathcal{L}_{\mathrm{init}}, \mathcal{L}_t, \mathcal{L}_{\mathrm{final}}$ separately.
We ignore the term $\mathcal{L}_{\mathrm{init}}$ during optimization since it contains no trainable parameters.
$\mathcal{L}_t$ for $0 \le t < T - 1$ can be decomposed as the following :
\begin{equation}
\label{eq:transition_loss}
\begin{split}
\mathcal{L}_t =
-\sum_{c \in \mathcal{N} (s^t)} &
\underbrace{D_{KL} (q_\theta (o_c | s^t, x) \| p_\theta (o_c | s^t) )}_{\mathcal{L}_{o}} \\
+ & q_\theta (o_c = 1 | s^t, x) \underbrace{D_{KL} (q_\theta (z_c| s^t, x, o_c=1) \| p_\theta (z_c| s^t, o_c=1))}_{\mathcal{L}_{z}},
\end{split}
\end{equation}
where the full derivation is in Appendix~\ref{app:kl_decomposition}.
Since $\mathcal{L}_{o}$ and $\mathcal{L}_{z}$ are the KL divergence between Bernoulli and normal distributions, respectively, $L_t$ can be written in a closed-form.
In practice, the scale of $\mathcal{L}_{z}$ can be much larger than that of $\mathcal{L}_{o}$.
This results in local minima in the gradient-based optimization and reduces the occupancy probability $q_\theta (o_c= 1 | s^t, x) $ for every cell.
So we balance the two losses by multiplying a hyperparameter $\gamma$ at $\mathcal{L}_{z}$, which is fixed as $\gamma= 0.01$ for all experiments.
The last term $\mathcal{L}_{\mathrm{final}}$ can be written as following:
\begin{equation}
\mathcal{L}_{\mathrm{final}}= \sum_{c \in \mathcal{N} (s^{T-1})} \log \{
(1 - \lambda_{\theta, c} ) \mathds{1}[o_c^x=0] \delta_{0}(z_c^x) + \lambda_{\theta, c} \mathds{1}[o_c^x=1] {N}(z^x_c; \; \mu_{\theta, c}, \sigma^{T-1} I)
\}.
\end{equation}
A problem rises when computing $\mathcal{L}_{\mathrm{final}}$, since the usage of Dirac distribution makes $\mathcal{L}_{\mathrm{final}} \rightarrow \infty$ if $o_c^x = 0$, which does not produce a valid gradient for optimization.
However, we can replace the loss $\mathcal{L}_{\mathrm{final}}$ with a well-behaved loss $\mathcal{L}_t$ for $t = T - 1$ , by using the following proposition:
\begin{restatable}{prop}{auxloss}
By replacing $\delta_{0}(z_c^x)$ with the indicator function $\mathds{1}[z_c^x=0]$ when computing $\mathcal{L}_\mathrm{final}$, $\nabla \mathcal{L}_{T - 1} = \nabla \mathcal{L}_{\mathrm{final}}$, for $T \gg 1$
\end{restatable}
\begin{proof}
\vspace{-0.7em}
The proof is found in Appendix~\ref{app:final_state_loss}.
\vspace{-0.7em}
\end{proof}
The proposition above serves as a justification for approximating $\nabla \mathcal{L}_{\mathrm{final}}$ as $ \nabla \mathcal{L}_{T - 1}$, with the benefits of having a simpler training procedure and easier implementation.
Replacing $\delta_{0}(z_c^T)$ with $\mathds{1}[z_c^x=0]$ can be regarded as a reweighting technique that naturally avoids divergence in the lower bound, since both functions output a non-zero value only at $0$.
Further discussions about the replacement are in Appendix~\ref{app:final_state_loss}.
In conclusion, the training procedure is outlined as follows:
\begin{enumerate}
\item Sample $s^{0: T}$ by $s^0 \sim q^0, \; s^{t + 1} \sim q_\theta^t(\cdot | s^t, x)$.
\item For $t < T$, update $\theta$ with $\theta \leftarrow \theta + \eta \nabla_\theta \mathcal{L}_t$, where $\eta$ is the learning rate.
\end{enumerate}
Note that the original infusion training (\cite{bordes2017infusion}) also attempts to minimize the variational lower bound, employing the Monte Carlo approximation with reparameterization trick (\cite{kingma2014vae}) to compute the gradients.
However, our objective avoids the approximations and can compute the exact lower bound for a single training step.
The proposed simplification can be applied to infusion training with any data structure including images.
We also summarize the difference of our formulation compared to GCA in Appendix~\ref{app:gca_difference}.
\subsection{Reconstructing Implicit Surfaces from Sparse Voxels}
We formulate the reconstruction problem as a two-step procedure: (i) stochastic voxel generation and (ii) implicit function reconstruction.
The overall procedure is depicted in Figure~\ref{fig:method_overview}.
The stochastic generation of the sparse voxels directly follows the formulation of GCA~\cite{zhang2021gca}.
Basically, we repeat the $T$ steps of sampling from the aggregated transition kernel in Equation~(\ref{eq:cell_sampling}) to progressively grow the shape.
Then the state $s^T$ is transformed into an implicit surface by evaluating the distance value for each query point $q \in \mathbb{R}^3$.
Note that an efficient sampling of query points near the surface is possible by sampling near the sparse voxel, which is already an approximation of the surface.
The final reconstruction is the implicit representation created from $s^T$.
The implicit function reconstruction is composed of the encoder $g_\phi$ and the decoder $f_\omega$.
The encoder receives $s^T$ sampled from the voxel generation step using the formulation of GCA.
In addition, the transition kernel $p_\theta$ outputs the $K$ dimensional feature $l_c \in \mathbb{R}^K$ for each occupied cell $c \in s^T$, which is consumed as the auxiliary feature $e^T=\{l_c\in \mathbb{R}^K\mid c \in s^{T}\}$ for the encoder.
The encoder $g_\phi$ transforms the sparse set of the coordinate-feature pairs $(s^T, e^T)$ into the hierarchical feature representation that summarizes the local context of each grid cell.
Then the decoder $f_\omega$ regresses the distance values~\cite{chibane20ifnet,peng2020conv_onet} using the information from the encoded feature $g_\phi(s^T, e^T)$.
Note that we intentionally re-use the deep sparse CNN for the transition kernel $p_\theta$ to train the entire network in an end-to-end fashion.
The right side of Figure~\ref{fig:method_overview} illustrates the implicit surface reconstruction module.
The encoder $g_\phi$ aggregates the information into multi-level grids in $n$ different resolutions
\begin{equation}
g_\phi(s^{T}, e^{T}) = F_1, F_2, ..., F_{n},
\end{equation}
similar to IF-Net~\cite{chibane20ifnet}.
Starting from the original resolution $F_1$, $F_k$ contains a grid that is downsampled $k$ times.
However, in contrast to the original IF-Net, our grid features are sparse, since they only store the downsampled occupied cells of state $s^T$.
The sparse representation $F_k = (s_k, e_k)$ is composed of coordinates $s_k \subset \mathbb{Z}^3$ and $L$-dimensional features $e_k = \{l_{c} \in \mathbb{R}^L \mid c \in s_k \}$.
The grid points that are not in $s_k$ are considered as having zero features.
Throughout the paper, we use $n=3$ levels.
The feature for a query point $q$ is obtained via adding all the multi-scale features: $\Psi (g_{\phi}(s^T, e^T), q) = \sum_k \psi_k(F_k, q)$. %
At the $k$-th scale, the feature at a particular position $q$ is evaluated by trilinear interpolation $\psi_k(F_k, q) \in \mathbb{R}^L$ of the feature $F_k$ at $k$-th level grid as in IF-Net~\cite{chibane20ifnet}.
The decoder $f_\omega: \mathbb{R}^L \rightarrow \mathbb{R}$ is an MLP that maps the feature at the query point
$\Psi (g_{\phi}(s^T, e^T), q)$
to the implicit function value $\hat{d}_q$.
We use truncated signed distance within the range of $[-1,1]$.
Note that we can also use unsigned distance function as in NDF~\cite{chibane2020ndf} by changing the codomain of the decoder to $[0, 1]$.
While the sampling procedure in Equation~(\ref{eq:transition_kernel}) enables efficient multi-modal shape generation, the output needs to converge to the single coherent shape at the final step.
However, na\"i{\bm{e}} sampling can produce blurry shapes when few voxels with low-occupancy probability are included.
As a simple trick, we augment \textit{mode seeking steps}.
During the mode seeking steps, we run a few additional $T'$ steps of the transition kernel but we select the cells with probability higher than 0.5 instead of implementing stochastic sampling.
The mode seeking steps prevent the cells with low probability from being sampled and the final continuous surface of the shape converges to a single mode that is closest to $s^T$.
The effects of mode seeking steps are visualized in Appendix~\ref{appendix:mode_seeking_steps}.
\subsection{Training Continuous Generative Cellular Automata}
\label{sec:training_cgca}
Since cGCA is composed of two modules, stochastic voxel generation and implicit function reconstruction, we introduce the training procedure of each module separately.
\paragraph{Stochastic Voxel Generation.}
At training time, we aim to learn the parameters $\theta$ of the transition kernel $p_\theta(s^{t + 1} \mid s^t)$, whose finite number of applications yields the samples from the learned distribution.
The training data contains an initial incomplete state $s^0$ and complete state $x$, which represents the surface as sparse voxels for each data.
Since we are given only the initial and final state, we need to emulate the sequence by using infusion training~\cite{bordes2017infusion}, which induces the chain to converge to the complete state.
To do so, we define a function $G_x(s)$ that finds the state that is closest to the complete state $x$ from the current state $s$ within the neighborhood $\mathcal{N}(s)$:
\begin{equation}
G_x(s) = \{\mathrm{argmin}_{c \in \mathcal{N}(s)} d(c, c') \mid c' \in x\}.
\end{equation}
Then, we define the infusion kernel $q^t$:
\begin{equation}
\quad q^t_\theta(s^{t+1} \mid s^t, x) = \prod_{c \in \mathcal{N}(s^t)} (1 - \alpha^t) p_\theta(c \mid s^t) + \alpha^t \delta_{G_x(s^t)}(c),
\end{equation}
where $\delta_{G_x(s^t)}(c)$ is formulated as Bernoulli distribution with probability 1, if $c \in G_x(s^t)$, else 0.
$\alpha^t$ is the infusion rate set to increase linearly with respect to time step, i.e., $\alpha^t = \max(wt + \alpha_0, 1)$, with $w > 0$.
Thus the infusion kernel for each cell is defined to be a mixture of $p_\theta$ and $\delta_{G_x(s^t)}$, where $\delta_{G_x(s^t)}$ is crucial to guarantee that the sequence ultimately reaches the ground truth shape.
The following proposition proves that the infusion chain converges to the complete data in a finite number of steps.
\begin{restatable}{prop}{infusionconvergence} \label{prop:infusion_convergence}
If the infusion rate $\alpha^t$ increases until it reaches a maximum value of 1 and every state $s^t$ is non-empty, then there exists $T$ such that $s^T = x$.
\label{prop1}
\end{restatable}
\begin{proof}
\vspace{-0.7em}
The proof is found in Appendix~\ref{appendix:prop1}.
\vspace{-0.7em}
\end{proof}
Thus, the training procedure for each step $s^t$ is to sample $s^{t + 1} \sim q_\theta(\cdot \mid s^t)$ and minimize the negative log-likelihood of its next state $s^{t + 1}$
\begin{equation}
\mathcal{L_{\textit{voxel}}} = - \log p_\theta(s^{t + 1} \mid s^t).
\label{eq:loss_voxel}
\end{equation}
While the training procedure of the voxel generation is similar to that of GCA, our voxel generation module is more flexible in the aspect that we do not assume any connectivity of the shape for convergence.
Instead, the infusion using the auxiliary function $G_x(s)$ ensures the convergence which is further verified in Appendix~\ref{appendix:gca_comparison}.
Also, while the likelihood objective of GCA is based on heuristics, our objective can be shown to maximize the lower bound of the log-likelihood of the complete state shown in Section~\ref{sec:lower_bound_derivation}.
\paragraph{Implicit Function Reconstruction.}
To train the network that regresses continuous surface, we train the neural network with a set $Q = \{(q, d_q)| q\in\mathbb{R}^3\}$, where $q$ is the points near the surface and $d_q$ is the signed distance from the surface to query point.
Since the voxel generation progressively completes the sparse voxels as if it were expanding the surface, we train the network to reconstruct the voxels for every state $s^t$ during the training step.
The loss function is defined as
\begin{equation}
\mathcal{L_{\textit{implicit}}} = \sum_{q \in Q} |f_w(\Psi (g_{\phi}(s^T, e^T), q)) - \max(\min(\frac{d_q}{\epsilon}, 1), -1)|,
\label{eq:loss_implicit}
\end{equation}
where $s^{t + 1}$ and $e^{t + 1}$ are sampled from the transition kernel $p_\theta$ and $\epsilon$ is the size of a single voxel.
Clamping the maximum distance lets the network focus on accurately predicting in the vicinity of the surface~\cite{park2019deepsdf, chibane2020ndf}.
To summarize, the total loss $\mathcal{L_{\textit{total}}}$ at state $s^t$ is the weighted sum of $\mathcal{L_{\textit{voxel}}}$ and $\mathcal{L_{\textit{implicit}}}$:
\begin{equation}
\mathcal{L_{\textit{total}}} = \mathcal{L_{\textit{voxel}}} + \beta \mathcal{L_{\textit{implicit}}},
\label{eq:loss_total}
\end{equation}
where $\beta$ is a hyperparameter.
To stabilize the training regardless of the varying level of incompleteness, the full training algorithm normalizes the loss function with the number of occupied voxels and utilizes a data buffer to control the length of the transition steps~\cite{zhang2021gca}.
The details of the end-to-end training can be found in Appendix~\ref{appendix:implementation_details}.
\subsection {Derivation of Maximizing Lower Bound of Log-likelihood}
\label{sec:lower_bound_derivation}
In this section, we verify that the training objective leads to a valid generative model.
The derivation focuses only on the final generated shape instead of the intermediate steps in the Markov chain.
We show that the process indeed maximizes the lower bound for the marginal log-likelihood of the complete state that we want to achieve at the last step.
We first use Proposition~\ref{prop:infusion_convergence} to find $T$ such that $x = s^T$, inducing $\log p_\theta(x) = \log p_\theta(s^T)$.
However, since the direct optimization of the exact log-likelihood is intractable, we show that our training objective in Equation~(\ref{eq:loss_voxel}) maximizes its variational lower bound, using the derivation from Luo \& Hu~\cite{luo2021diffusion}:
\begin{align}
\log p_\theta(s^T)
&\ge \sum_{s^{0: T-1}} q_\eta(s^{0:T-1}|s^T) \log \frac{p_\theta(s^{0:T})}{q_\eta(s^{0:T-1}|s^T)} \\
&= \log \frac{p(s^0)}{q(s^0)} + \mathbb{E}_{q_\eta}[\log p_\theta(s^T | s^{T - 1})] - \sum_{0 \le t < T - 1} D_{KL}(q_\eta(s^{t + 1} | s^t, s^T) \| p_\theta(s^{t + 1} | s^t)))
\label{eq:elbo}
\end{align}
where the full derivation is in Appendix~\ref{section:detailed_derivation} for the reader's convenience.
The parameter $\eta$ of $q$ is substituted with $\theta$ every time $\theta$ is updated.
Note that the first term of Equation~(\ref{eq:elbo}) has no trainable parameters and the second term is maximized directly from the training objective.
The following proposition shows that the training objective minimizes the third KL divergence term, therefore maximizing the log-likelihood of data distribution.
\begin{restatable}{prop}{maximizinglowerbound} \label{prop:maximizing_lower_bound}
If the likelihood for every cell $c \in s^{t + 1}$ increases when updating parameters $\theta$ with $\theta'$ by minimizing Equation~\text{(\ref{eq:loss_voxel})}, where $s^{t + 1} \sim q_\theta(\cdot | s^t)$, i.e., $ p_{\theta'}(c | s^t) \ge p_{\theta}(c | s^t), \; \forall c \in s^{t+1}$, then the KL-divergence term in Equation~\text{(\ref{eq:elbo})} decreases, i.e., $D_{KL}(q_{\theta'}(s^{t + 1} | s^t, s^T) \| p_{\theta'}(s^{t + 1} | s^t)) \le D_{KL}(q_{\theta}(s^{t + 1} | s^t, s^T) \| p_{\theta}(s^{t + 1} | s^t))$.
\end{restatable}
\begin{proof}
\vspace{-0.7em}
The proof is found in Appendix~\ref{appendix:prop2}.
\vspace{-0.7em}
\end{proof}
Thus, the lower bound of the marginal log-likelihood is maximized.
\section{Experiments}
\label{sec:experiments}
We test the probabilistic shape completion of cGCA for scenes (Section~\ref{sec:scene_completion}) and single object (Section~\ref{sec:single_object_completion}).
For all the experiments, we use latent code dimension $K=32$, trained with regularization parameter $\beta = 0.001$.
All the results reported are re-trained for each dataset.
Further implementation details are in the Appendix~\ref{app:implementation_details} and effects of mode seeking steps are discussed in Appendix~\ref{app:ablation_mode_seeking}.
\subsection{Scene Completion}
\label{sec:scene_completion}
In this section, we evaluate our method on two datasets: ShapeNet scene~(\cite{peng2020conv_onet}) and 3DFront~(\cite{fu20203dfront}) dataset.
The input point cloud are created by iteratively removing points within a fixed distance from a random surface point.
We control the minimum preserved ratio of the original complete surface points for each object in the scene to test varying levels of completeness.
The levels of completeness tested are 0.2, 0.5, and 0.8 with each dataset trained/tested separately.
We evaluate the quality and diversity of the completion results by measuring the Chamfer-L1 distance (CD), total mutual distance (TMD), respectively, as in the previous methods~(\cite{peng2020conv_onet, zhang2021gca}).
For probabilistic methods, five completion results are generated and we report the minimum and average of Chamfer-L1 distance (min. CD, avg. CD).
Note that if the input is severely incomplete, there exist various modes of completion that might be feasible but deviate from its ground truth geometry.
Nonetheless, we still compare CD assuming that plausible reconstructions are likely to be similar to the ground truth.
\input{table/result_synthetic}
\begin{figure}[t]
\vspace{-1em}
\includegraphics[width=\textwidth]{./figures/synthetic_result.pdf}
\vspace{-2em}
\centering
\caption[] {
Qualitative comparison on ShapeNet scene dataset.
Best viewed on screen.
Minimum rate indicates the guaranteed rate of surface points for each object in the scene.
While deterministic methods (ConvOcc, IFNet) produce blurry surfaces since they cannot model multi-modal distribution, probabilistic methods (GCA, cGCA) generate multiple plausible scenes.
cGCA is the only method that can generate multiple plausible scenes without losing the details for each object.
}
\label{fig:synthetic_comparison}
\vspace{-1em}
\end{figure}
\textbf{ShapeNet Scene.}
ShapeNet scene contains synthetic rooms that contain multiple ShapeNet~(\cite{chang2015shapenet}) objects, which have been randomly scaled with randomly scaled floor and random walls.
We compare the performance of cGCA with two deterministic scene completion models that utilize occupancy as the implicit representation: ConvOcc~(\cite{peng2020conv_onet}) and IFNet~(\cite{chibane20ifnet}).
We additionally test the quality of our completion against a probabilistic method of GCA~(\cite{zhang2021gca}).
Both GCA and cGCA use $64^3$ voxel resolution with $T=15$ transitions.
cGCA (w/ cond.) indicates a variant of our model, where each transition is conditioned on the initial state $s^0$ obtained by the encoder $g_\phi$ as discussed in Sec.~\ref{sec:sampling}.
Table \ref{table:result_synthetic} contains the quantitative comparison on ShapeNet scene.
Both versions of cGCA outperform all other methods on min. CD for all level of completeness.
The performance gap is especially prominent for highly incomplete input, which can be visually verified from Fig.~\ref{fig:synthetic_comparison}.
The deterministic models generate blurry objects given high uncertainty, while our method consistently generates detailed reconstructions for inputs with different levels of completeness.
Our result coincides with the well-known phenomena of generative models, where the deterministic models fail to generate crisp outputs of multi-modal distribution~(\cite{goodfellow2017gan_tutorial}).
Considering practical scenarios with irregular real-world scans, our probabilistic formulation is crucial for accurate 3D scene completion.
When conditioned with the initial state, the completion results of cGCA stay faithful to the input data, achieving lower CDs.
Further analysis on varying completeness of input is discussed in Appendix~\ref{app:ablation_completeness}.
\input{table/result_3dfront}
\begin{figure}[t]
\vspace{-1em}
\centering
\includegraphics[width=0.97\textwidth]{./figures/3dfront_result.pdf}
\vspace{-1em}
\caption[] {
Qualitative comparison on 3DFront dataset with 0.5 object minimum rate.
Best viewed on screen.
Since the raw inputs of furniture are highly incomplete, there exist multiple plausible reconstructions.
Probabilistic approaches produce diverse yet detailed scene geometry.
GCA suffers from artifacts due to the discrete voxel resolution.
}
\label{fig:3dfront_comparison}
\vspace{-1em}
\end{figure}
\textbf{3DFront.}
3DFront is a large-scale indoor synthetic scene dataset with professionally designed layouts and contains high-quality 3D objects, in contrast to random object placement for ShapeNet scene.
3DFront dataset represents the realistic scenario where the objects are composed of multiple meshes without clear boundaries for inside or outside.
Unless the input is carefully processed to be converted into a watertight mesh, the set-up excludes many of the common choices for implicit representation, such as occupancy or signed distance fields.
However, the formulation of cGCA can be easily adapted for different implicit representation, and we employ \textit{unsigned distance fields} (\cite{chibane2020ndf}) to create the sparse voxel embedding for 3DFront dataset.
We compare the performance of cGCA against GCA, both with voxel resolution of 5cm and $T=20$ transitions.
Table~\ref{table:result_3dfront} shows that cGCA outperforms GCA by a large margin in CD, generating high-fidelity completions with unsigned distance fields.
While both GCA and cGCA are capable of generating multiple plausible results, GCA suffers from discretization artifacts due to voxelized representation, as shown in Fig.~\ref{fig:3dfront_comparison}.
cGCA not only overcomes the limitation of the resolution, but also is scalable to process the entire rooms at once during both training and test time.
In contrast, previous methods for scene completion (\cite{siddiqui2021retrievalfuse, peng2020conv_onet}) divide the scene into small sections and separately complete them.
We analyze the scalability in terms of the network parameters and the GPU usage in Appendix~\ref{app:scalabilty}.
In Appendix~\ref{app:scannet}, we also provide results on ScanNet~(\cite{dai2017scannet}) dataset, which is one of the widely used datasets for real-world indoor environments.
\input{table/result_shapenet}
\begin{figure}[t]
\vspace{-1em}
\includegraphics[width=0.9\textwidth]{./figures/shapenet_result.pdf}
\vspace{-2em}
\centering
\caption[] {
Qualitative comparison on probabilistic shape completion of a single object.
cGCA is the only method that can produce a continuous surface.
}
\label{fig:single_object_comparison}
\vspace{-1em}
\end{figure}
\subsection{Single Object Completion}
\label{sec:single_object_completion}
We analyze various performance metrics of cGCA for a single object completion with chair/sofa/table classes of ShapeNet~(\cite{chang2015shapenet}) dataset.
Given densely sampled points of a normalized object in $[-1, 1]^3$, the incomplete observation is generated by selecting points within the sphere of radius 0.5 centered at one of the surface points.
From the partial observation, we sample 1,024 points which serve as a sparse input, and sample 2,048 points from completion results for testing.
Following the previous method (\cite{wu2020msc}), we generate ten completions and compare MMD (quality), UHD (fidelity), and TMD (diversity).
Our method is compared against other probabilistic shape completion methods: cGAN (\cite{wu2020msc}) which is based on point cloud, and GCA (\cite{zhang2021gca}) which uses voxel representation.
We use $T=30$ transitions.
Quantitative and qualitative results are shown in Table~\ref{table:result_shapenet} and Fig.~\ref{fig:single_object_comparison}.
Our approach exceeds other baselines in all metrics, indicating that cGCA can generate high-quality completions (MMD) that are faithful to input (UHD) while being diverse (TMD).
By using latent codes, the completed continuous surface of cGCA can capture geometry beyond its voxel resolution.
The quality of completed shape in $32^3$ voxel resolution therefore even outperforms in MMD for discrete GCA in higher $64^3$ voxel resolution.
Also, the UHD score of cGCA (w/ cond.) exceeds that of GCA and vanilla cGCA indicating that conditioning latent codes from the input indeed preserves the input partial geometry.
\section{Experiments}
We test the probabilistic shape completion for a single object (Section~\ref{sec:single_object_completion}) and scene (Section~\ref{sec:scene_completion}) dataset.
Details regarding the hyperparameters and baselines are in Appendix~\ref{appendix:implementation_details}.
\subsection{Single Object Completion}
\label{sec:single_object_completion}
The single object probabilistic completion with cGCA is compared against cGAN~\cite{wu2020msc} and GCA~\cite{zhang2021gca}.
We use the shapes of chair, sofa, and table category from ShapeNet~\cite{chang2015shapenet} dataset which are normalized to fit within a unit sphere.
For training cGCA, we obtain the query points with corresponding signed distance values from watertight mesh as in DeepSDF~\cite{park2019deepsdf}.
The partial input is generated by first uniformly sampling the mesh and then extracting points within a sphere of radius 0.5, centered at one of the surface points.
Then we sample 1,024 points to create a sparse input.
For GCA and cGCA, we create a quantized version of the sparse voxel grid by placing the point cloud into a grid and marking cells containing points as occupied.
We generate ten completions for each input and compare the performance in terms of three metrics: minimal matching distance (MMD), unidirectional Hausdorff distance (UHD), and total mutual distance (TMD) by following the work of Wu et al.~\cite{wu2020msc}.
The metrics are calculated using point cloud, which is the output of cGAN.
The output of GCA is a set of occupied voxels which are turned into a mesh with marching cubes~\cite{lorensen1987mcubes} and then sampled.
The points of the implicit representation of cGCA are extracted from positions where the regressed distance value is below half of the grid length.
\input{table/result_shapenet}
\begin{figure}[t]
\vspace{-0.5em}
\includegraphics[width=\textwidth]{./figures/single_object_comparison.pdf}
\centering
\caption[] {
Qualitative comparison on single object completion against cGAN~\cite{wu2020msc} and GCA~\cite{zhang2021gca}
}
\label{fig:single_object_comparison}
\end{figure}
The superior performance of cGCA is reported in Table~\ref{table:single_object_completion_comparison}.
Since the number of generated points in cGAN is fixed to 2,048, we first compare by sampling 2,048 points (top of Table~\ref{table:single_object_completion_comparison}) and once again sampling 16,384 points without cGAN to examine the high resolution surface (bottom).
Our method in signed representation outperforms all other methods in MMD (quality) and UHD (fidelity) in average, while having comparable TMD (diversity) against GCA.
This indicates that cGCA generates shapes close to the data distribution (MMD) while being diverse (TMD) and faithful to the input (UHD).
Figure~\ref{fig:single_object_comparison} shows that cGCA is able to generate a smooth surface while other works are either limited to the number of points or the voxel resolution.
To show that cGCA can generate surfaces beyond the voxel resolution, we compare the results of GCA and cGCA in two different voxel resolutions, $32^3$ and $64^3$.
The voxel resolution affects the estimation accuracy of the true surface of the shape, as compared in Figure~\ref{fig:single_object_ablation}.
Table~\ref{table:single_object_completion_comparison} shows that both GCA and cGCA achieve better performance in all three metrics at $64^3$ voxel resolution.
The difference is less prominent for cGCA, and signed cGCA clearly outperforms GCA of the same resolution in terms of MMD and UHD.
Notably, cGCA of voxel resolution $32^3$ shows better results in MMD than GCA in voxel $64^3$ resolution, thereby showing the expressive power of the implicit representation beyond the voxel resolution.
We also show the results using unsigned distance fields as in NDF~\cite{chibane2020ndf} in place of signed distance fields for the implicit representation.
While signed representation can only represent closed surfaces, the unsigned implicit field is applicable to mesh with various topology.
However, the flexibility comes at the cost of expressive power, achieving higher MMD and UHD compared to the signed counterparts.
Without a clear distinction between the interior/exterior, the unsigned representation also makes it difficult to scrutinize the zero-level isosurface and thus results in comparably thick chair legs as shown in Figure~\ref{fig:single_object_ablation}.
Nonetheless, the unsigned cGCA outperforms MMD scores compared to that of GCA, which indicates that it generates surface beyond the voxel resolution.
\subsection{Scene Completion}
\label{sec:scene_completion}
\input{table/result_synthetic}
\begin{figure}[t]
\includegraphics[width=\textwidth]{./figures/fig3_synthetic}
\centering
\caption[] {
\TODO{Write caption}
}
\label{fig:synthetic_comparison}
\end{figure}
\input{table/result_3dfront}
\begin{figure}[t]
\centering
\includegraphics[width=0.97\textwidth]{./figures/scene_completion_comparison.pdf}
\caption[] {
Qualitative comparison on probabilistic scene completion.
Best viewed on screen.
Since the raw inputs of furniture are highly incomplete, there exists multiple plausible reconstructions.
Probabilistic approaches produce diverse yet detailed scene geometry.
GCA suffers from artifacts due to the discrete voxel resolution.
\TODO{change figure}
\vspace{-1.5em}
}
\label{fig:scene_comparison}
\end{figure}
Probabilistic scene completion is tested with 3DFront~\cite{fu20203dfront} synthetic scene dataset.
We sample 7891 and 1917 rooms for training and validation set, respectively.
The unsigned distance value is sampled in the vicinity of mesh, whose positions are randomly perturbed from the surface with Gaussian noise.
The partial input is created by first sampling surface points and then iteratively removing points within a fixed distance from a random position for each object in the scene.
The partial input for multi-modal scene completion is guaranteed to maintain at least 30\% of initial points sampled from the original surface.
Lastly, the input points are voxelized with size 5cm and 20\% of the voxels are removed for a sparse input.
The sparse and incomplete input encourages diverse outputs as shown in Figure~\ref{fig:scene_comparison}.
Further details about the data preparation are articulated in Appendix~\ref{appendix:data_preparation}.
Note that the scenes are composed of multiple pieces of mesh, and cannot be processed with previous works that utilize signed distance field~\cite{peng2020conv_onet, chibane20ifnet, chiyu2020local}.
To employ signed distance field, it is necessary to convert the input to a watertight mesh with pre-processing, which is especially challenging when the 3D model is obtained from a noisy scan.
Being able to handle unsigned distance fields, our approach can be readily applied in scenes, where the inherent scalability of cGCA can be fully exploited.
We compare our method against NDF~\cite{chibane2020ndf} and GCA~\cite{zhang2021gca}.
NDF uses the implicit representation based on unsigned distance field as ours, but produces deterministic completion.
GCA, on the other hand, is capable of generating multi-modal outcomes, but uses discretized representation with voxels.
For probabilistic methods, we generate five completions and sample points proportional to the size of the room.
We introduce min-Chamfer-\textit{L}1, average-Chamfer-\textit{L}1 metrics, which compare the minimum and average of Chamfer distance of outcomes to the ground truth mesh.
We also evaluate TMD and UHD.
The quantitative results are shown in Table~\ref{table:scene_completion_comparison}.
Our method achieves high TMD scores, implying that we can generate diverse high-quality surfaces.
Despite the diversity, the outcome is still similar to the ground truth with a low Chamfer-\textit{L}1 distance.
As demonstrated in single object completion, the network with unsigned distance function thickens some fine details, and our UHD scores are slightly higher than that of GCA.
However, the visual quality shown in Figure~\ref{fig:scene_comparison} clearly indicates that GCA suffers from discretization artifact of voxels whereas cGCA can generate continuous surfaces.
Our results further verify the discussion by Goodfellow~\cite{goodfellow2017gan_tutorial}: generative models need to capture multi-modal distribution, as deterministic models can merely estimate the mean value of complex multiple possibilities and generate blurry outputs.
As shown in Figure~\ref{fig:scene_comparison} and Table~\ref{table:scene_completion_comparison}, the deterministic method (NDF) fails to generate plausible scenes when the input scene is largely incomplete.
On the other hand, the probabilistic methods (GCA and cGCA) successfully recover a single mode despite the ambiguity of the input.
We further analyze why NDF fails in Appendix~\ref{appendix:baselines}.
Another strength of our method is scalability.
With sparse representation and implicit fields, all of our results process the entire scene as a single input.
On the other hand, previous works~\cite{peng2020conv_onet, chibane20ifnet, chibane2020ndf} process large-scale scenes by cropping them into multiple blocks with sliding window techniques.
This implies that all of the necessary information should be contained in the same block.
Such an assumption fails with severe incompleteness or long-range context.
Being able to observe the rich context of the entire scene, our method produces reasonable completion in various challenging cases.
\section{Related Works}
\paragraph{3D Shape Completion.}
The data-driven completion of 3D shapes demands a large amount of memory and computation.
The memory requirement for voxel-based methods increases cubically to the resolution~(\cite{dai2017complete}) while the counterpart of point cloud based representations~(\cite{yuan2018pcn}) roughly increases linearly with the number of points.
Extensions of scene completion in voxel space utilize hierarchical representation~(\cite{dai2018scancomplete}) or subdivided scenes~(\cite{dai2018scancomplete,dai2020sgnn}) with sparse voxel representations.
Recently, deep implicit representations~(\cite{park2019deepsdf, chen2019implicit_decoder, mescheder2019onet}) suggest a way to overcome the limitation of resolution.
Subsequent works~(\cite{chabra2020deep_local_shapes, chibane20ifnet, chiyu2020local, peng2020conv_onet}) demonstrate methods to extend the representation to large-scale scenes.
However, most works are limited to regressing a single surface from a given observation.
Only a few recent works~(\cite{wu2020msc, zhang2021gca, smith2017iwgan}) generate multiple plausible outcomes by modeling the distribution of surface conditioned on the observation.
cGCA suggests a scalable solution for multi-modal continuous shape completion by employing progressive generation with continuous shape representations.
\paragraph{Diffusion Probabilistic Models.}
One way to capture the complex data distribution by the generative model is to use a diffusion process inspired by nonequilibrium thermodynamics such as~\cite{sohl2015nonequil, ho2020denoising, luo2021diffusion}.
The diffusion process incrementally destroys the data distribution by adding noise, whereas the transition kernel learns to revert the process that restores the data structure.
The learned distribution is flexible and easy to sample from, but it is designed to evolve from a random distribution.
On the other hand, the infusion training by ~\cite{bordes2017infusion} applies a similar technique but creates a forward chain instead of reverting the diffusion process.
Since the infusion training can start from a structured input distribution, it is more suitable to a shape completion that starts from a partial data input.
However, the infusion training approximates the lower bound of variational distribution with Monte Carlo estimates using the reprameterization trick~(\cite{kingma2014vae}).
We modify the training objective and introduce a simple variant of infusion training that can maximize the variational lower bound of the log-likelihood of the data distribution without using Monte Carlo approximation.
\section{Conclusion}
\label{sec:conclusion}
We are the first to tackle the challenging task of probabilistic scene completion, which requires not only the model to generate multiple plausible outcomes but also be scalable to capture the wide-range context of multiple objects.
To this end, we propose continuous Generative Cellular Automata, a scalable generative model for completing multiple plausible continuous surfaces from an incomplete point cloud.
cGCA compresses the implicit field into sparse voxels associated with their latent code named sparse voxel embedding, and
incrementally generates diverse implicit surfaces.
The training objective is proven to maximize the variational lower bound of the likelihood of sparse voxel embeddings, indicating that cGCA is a theoretically valid generative model.
Extensive experiments show that our model is able to faithfully generate multiple plausible surfaces from partial observation.
There are a few interesting future directions.
Our results are trained with synthetic scene datasets where the ground truth data is available.
It would be interesting to see how well the data performs in real data with self-supervised learning.
For example, we can extend our method to real scenes such as ScanNet~\cite{dai2017scannet} or Matterport 3D~(\cite{Matterport3D}) by training the infusion chain with data altered to have different levels of completeness as suggested by ~\cite{dai2020sgnn}.
Also, our work requires two-stage training, where the transition kernel is trained with the ground truth latent codes generated from the pre-trained autoencoder.
It would be less cumbersome if the training could be done in an end-to-end fashion.
Lastly, our work takes a longer inference time compared to previous methods~(\cite{peng2020conv_onet}) since a single completion requires multiple transitions.
Reducing the number of transitions by using a technique similar to \cite{salimans2022progressive} can accelerate the runtime.
\section{Ethics Statement}
\label{sec:code_of_ethics}
The goal of our model is to generate diverse plausible scenes given an observation obtained by sensors. While recovering the detailed geometry of real scenes is crucial for many VR/AR and robotics applications, it might violate proprietary or individual privacy rights when abused. Generating the unseen part can also be regarded as creating fake information that can deceive people as real.
\section{Reproducibility}
\label{sec:reproducibility}
Code to run the experiments is available at \href{https://github.com/96lives/gca}{https://github.com/96lives/gca}.
Appendix~\ref{app:sec:implementation_details} contains the implementation details including the network architecture, hyperparameter settings, and dataset processing.
Proofs and derivations are described in Appendix~\ref{app:sec:proofs}.
\section{Acknowledgement}
We thank Youngjae Lee for helpful discussion and advice on the derivations.
This research was supported by the
National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No.2020R1C1C1008195) and the National Convergence Research of Scientific Challenges through the National Research Foundation of Korea (NRF) funded by Ministry of Science and ICT (NRF2020M3F7A1094300).
Young Min Kim is the corresponding author.
\section{Implementation Details}
\label{app:sec:implementation_details}
We use Pytorch~(\cite{paszke2019pytorch}) and the sparse convolution library MinkowskiEngine~(\cite{choy20194d}) for all implementations.
\subsection{Autoencoder}
\label{app:sec:autoencoder}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{figures/appendix/decoder_architecture.pdf}
\caption{Neural network architecture for the decoder $f_\omega$.
The left side shows the overall architecture for the decoder $f_\omega$ and the right side shows the architecture for sparse convolution layers $f_{\omega_1}$.
The parenthesis denotes the stride of the sparse convolution and every convolution except the feature extracting layer is followed by batch normalization and ReLU activation.
}
\label{fig:decoder_architecture}
\end{figure}
\textbf{Network Architecture.}
The autoencoder is composed of two modules, encoder $g_\phi$ and decoder $f_\omega$ to convert between the sparse voxel embedding and implicit fields, which is depicted in Fig.~\ref{fig:method_overview}.
The encoder $g_\phi$ generates the sparse voxel embedding $s$ from the coordinate-distance pairs $P = \{(p, d_p) | p \in \mathbb{R}^3, d_p \in \mathbb{R} \}$, where $p$ is a 3D coordinate and $d_p$ is the (signed) distance to the surface, $s = g_\phi(P)$.
The decoder $f_\omega$ regresses the implicit field value $d_q$ at the 3D position $q \in \mathbb{R}^3$ given the sparse voxel embedding $s$, $d_q=f_\omega(q, s)$.
The encoder $g_\phi$ is a a local PointNet (\cite{qi2016pointnet}) implemented with MLP of 4 blocks with hidden dimension 128, as in ConvOcc~(\cite{peng2020conv_onet}).
Given a set of coordinate-distance pair $P$, we find the nearest cell $c_p$ for each coordinate-distance pair $(p, d_p)$.
Then we make the representation sparse by removing the coordinate-distance pair if the nearest cell $c_p$ do not contain the surface.
For the remaining pairs, we normalize the coordinate $p$ to the local coordinate $p' \in [-1, 1]^3$ centered at the nearest voxel.
Then we use the local coordinate to build a vector in $\mathbb{R}^4$ for coordinate-distance pair $(p', d_p)$ which serves as the input to the PointNet for the encoder $g_\phi$.
The local PointNet architecture employs average pooling to the point-coordinate pairs that belong to the same cell.
The architecture of decoder $f_\omega$, inspired by IFNet~(\cite{chibane20ifnet}) is depicted in Fig.~\ref{fig:decoder_architecture}.
As shown on the left, the decoder $f_\omega$ can be decomposed into the sparse convolution $f_{\omega_1}$ that extracts hierarchical feature, followed by trilinear interpolation of the extracted feature $\Psi$, and final MLP layers $f_{\omega_2}$.
The sparse convolution layer $f_{\omega_1}$ aggregates the information into multi-level grids in $n$ different resolutions, i.e., $f_{\omega_1}(s) = F_1, F_2, ..., F_{n}$,
Starting from the original resolution $F_1$, $F_k$ contains a grid that is downsampled $k$ times.
However, in contrast to the original IFNet, our grid features are sparse, since they only store the downsampled occupied cells of sparse voxel embedding $s$.
The sparse representation $F_k = \{(c, o_c, e_c) | c \in \mathds{Z}^3, o_c \in \{0, 1\}, e_c \in \mathbb{R}^L\}$ is composed of occupancy $o_c$ and $L$-dimensional feature $e_c$ for each cell, like sparse voxel embedding.
The grid points $c$ that are not occupied are considered as having zero features.
We use $n=3$ levels, with feature dimension $L=32$ for all experiments.
The multi-scale features are then aggregated to define the feature for a query point $q$: $\Psi (f_{\omega_1}(s), q) = \sum_k \psi_k(F_k, q)$.
Here $\psi_k(F_k, q) \in \mathbb{R}^L$ is the trilinear interpolation of features in discrete grid $F_k$ to define the feature at a particular position $q$, similar to IFNet.
We apply trilinear interpolation at each resolution $k$ then combine the features.
Lastly, the MLP layer maps the feature at the query point $q$ to the implicit function value $\hat{d}_q$, $f_{\omega_2}: \mathbb{R}^L \rightarrow \mathbb{R}$.
The MLP is composed of 4 ResNet blocks with hidden dimension 128 with $\tanh$ as the final activation function.
The final value $\hat{d}_q$ is the truncated signed distance within the range of $[-1,1]$ except for 3DFront~(\cite{fu20203dfront}) which does not have a watertight mesh.
We use unsigned distance function for 3DFront as in NDF~(\cite{chibane2020ndf}) by changing the codomain of the decoder to $[0, 1]$ with sigmoid as the final activation function.
Thus, the decoder can be summarized as $f_\omega(s, q) = f_{\omega_2} (\Psi (f_{\omega_1}(s), q))$.
We would like to emphasize that the autoencoder does not use any dense 3D convolutions, and thus the implementation is efficient,
especially when reconstructing the sparse voxel embedding.
Since the decoding architecture only consists of sparse convolutions and MLPs, the memory requirement increases linearly with respect to the number of occupied voxels.
In contrast, the requirement for dense 3D convolution increases cubic with respect to the maximum size of the scene.
\begin{figure}
\centering
\includegraphics[width=0.85\textwidth]{figures/appendix/autoencoder_reconstruction.pdf}
\caption{
Visualizations of reconstructions by autoencoder.
We visualize a reconstruction of a chair using signed distance fields (left) and a scene in 3DFront using unsigned distance fields (right).
Both reconstructions show that sparse voxel embedding is able to reconstruct a continuous surface.
}
\label{fig:autoencoder_reconstruction}
\end{figure}
\textbf{Reconstruction Results.}
Fig.~\ref{fig:autoencoder_reconstruction} shows a reconstruction of a chair (left) using signed distance fields and a scene in 3DFront (right) using unsigned distance fields.
The figure shows that our autoencoder can faithfully reconstruct a continuous surface using the sparse voxel embedding.
However, the usage of unsigned distance fields~(\cite{chibane2020ndf}) representation tends to generate relatively thicker reconstructions compared to that of signed distance fields.
Without a clear distinction between the interior/exterior, the unsigned representation makes it difficult to scrutinize the zero-level isosurface and thus results in comparably thick reconstruction.
Thus, the flexibility of being able to represent any mesh with various topology comes at a cost of expressive power, in contrast to signed distance fields which can only represent closed surfaces.
Our work can be further improved by utilizing more powerful implicit representation techniques, such as SIREN~(\cite{sitzmann2019siren}), but we leave it as future work.
\textbf{Hyperparameters and Implementation Details.}
We train the autoencoder with Adam~(\cite{kingma2015adam}) optimizer with learning rate 5e-4 and use batch size of 3, 1, and 4 for ShapeNet scene~(\cite{peng2020conv_onet}), 3DFront~(\cite{fu20203dfront}), and ShapeNet~(\cite{chang2015shapenet}) dataset, respectively.
The autoencoder for scene completion is pretrained with all classes of ShapeNet assuming that the local geometric details for scenes are represented with the patches from individual objects.
\subsection{Transition Model}
\label{app:implementation_details}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figures/appendix/transition_archietcture.pdf}
\caption{Neural network architecture for the transition model $p_\theta$. We employ the architecture of U-Net.
The paranthesis indicates the stride of convolutions and each convolution is followed by batch normalization.
}
\label{fig:transition_architecture}
\end{figure}
\textbf{Network Architecture.}
Following GCA~(\cite{zhang2021gca}), we implement the transition model by using a variant of U-Net~(\cite{ronneberger2015unet}).
The input to the sparse convolution network is the sparse voxel embedding $s^t$, and the output feature size of the last convolution is $K + 1$, where we use generalized sparse convolution for the last layer to obtain the occupancy (with dimension 1) and the latent code (with dimension $K$) of the neighborhood.
Unlike GCA, our implementation does not need an additional cell-wise aggregation step.
\textbf{$s^0$ Conditioned Model.}
The $s^0$ conditioned model is a variant of our model, where we employ the use of encoder $g_\phi$ to estimate the initial latent codes of initial state $s^0$ and condition every transition kernel on $s^0$.
Given incomplete point cloud $R = \{c_1, c_2, ..., c_M\}$ with $c_i \in \mathbb{R}^3$, the variant constructs the initial $s_0$ by encoding a set of coordinated-distance pair $P_R = \{(c, 0)| c\in R\}$, where the distance is set to 0 for every coordinate in $R$.
Since the point cloud is sampled from the surface of the original geometry, this is a reasonable estimate of the latent code.
The transition kernel $p_\theta$ is conditioned on $s^0$ by taking a sparse voxel embedding $\hat{s}^{t}$ of latent code dimension $2K$ that concatenates the current state $s^t$ and $s^0$ as the input.
Specifically, we define $\hat{s^t}$ to be $\hat{s}^t=\{(c, o_c, z_c)| o_c = o^{s^t}_c \vee o^{s^0}_c, z_c = [z^{s^0}_c, z^{s^t}_c] \in \mathbb{R}^{2K}, c \in \mathbb{Z}^3 \}$, where the $o^{s^t}_c, o^{s^0}_c, z^{s^t}_c, z^{s^0}_c$ are the occupancies and latent codes of $s^t$ and $s^0$.
The occupancy is set to one if either $s^t$ or $s^0$ is occupied and the latent code is set to be the concatenation of latent codes of $s^0$ and $s^t$.
Thus, the only modification to the neural network $p_\theta$ is the size of input feature dimension, which is increased from $K$ to $2K$.
\textbf{Hyperparameters and Implementation Details}
We train our model with Adam~(\cite{kingma2015adam}) optimizer with learning rate 5e-4.
The transition kernel is trained using infusion schedule $\alpha^t = 0.1 + 0.005t$ and standard deviation schedule $\sigma^t = 10^{-1 - 0.01t}$.
We use the accelerated training scheme introduced in GCA~(\cite{zhang2021gca}) by managing the time steps based on the current step.
We use neighborhood size $r=2$, with metric $d$ induced by $L_1$ norm, except for the ShapeNet~(\cite{chang2015shapenet}) experiment on $64^3$ voxel resolution which uses $r=3$.
For transition model training, we utilize maximum batch size of 32, 16 and 6, for ShapeNet scene, 3DFront, ShapeNet dataset, respectively.
We use adaptive batch size for the transition model, which is determined by the total number of cells in the minibatch for efficient GPU memory usage.
Lastly, we use mode seeking steps $T'=5$ for all experiments.
\textbf{Visualizations and Testing.}
All surface visualizations are made by upsampling the original voxel resolution 4 times, where we efficiently query from the upsampled grid only near the occupied cells of the final state $s^{T + T'}$.
Then, the mesh is extracted with marching cubes~(\cite{lorensen1987marching}) with no other postprocessing applied.
From the $4\times$ upsampled high resolution grid, we extract cells with the distance value below 0.5 which serve as the point cloud to compute the metrics for testing.
\subsection{Baselines}
We use the official implementation or the code provided by the authors by contact with default hyperparameters unless stated otherwise.
All results are obtained by training on our dataset.
The output point cloud used for testing, is obtained by sampling from mesh with the same resolution if possible, unless stated otherwise.
For GCA~(\cite{zhang2021gca}) we use the same hyperparameters as ours for fair comparison.
The mesh is created by using marching cubes~(\cite{lorensen1987marching}) with the voxel resolution.
We use the volume $64^3$ encoder for ConvOcc~(\cite{peng2020conv_onet}) which achieves the best results in the original ShapeNet scene dataset~(\cite{peng2020conv_onet}).
For IFNet~(\cite{chibane20ifnet}), we use the ShapeNet128Vox model as used in the work of \cite{siddiqui2021retrievalfuse} with occupancy implicit representation.
cGAN~(\cite{wu2020msc}) is tested by 2,048 generated point cloud.
\subsection{Datasets}
\textbf{ShapeNet Scene.}
ShapeNet scene~(\cite{peng2020conv_onet}) contains synthetic rooms that contain multiple ShapeNet~(\cite{chang2015shapenet}) objects, which has randomly scaled floor and random walls, normalized so that the longest length of the bounding box has length of 1.
The number of rooms for train/val/test split are 3750/250/995.
We sample SDF points for training our model.
The input is created by iteratively sampling points from the surface and removing points within distance 0.1 for each instance.
We run 20 iterations and each removal is revoked if the aggregated removal exceeds that of the guaranteed rate.
Since we are more interested in the reconstruction of objects, the walls and floors are guaranteed to have minimum rate of 0.8, regardless of the defined minimum rate.
Lastly, we sample 10,000 points and add normal noise with standard deviation 0.005.
For computing the metrics at test time, we sample 100,000 points.
\textbf{3DFront.}
3DFront~(\cite{fu20203dfront}) is a large-scale indoor synthetic scene dataset with professionally designed layouts and contains high quality 3D objects.
We collect rooms that contain multiple objects, where 10\% of the biggest rooms are removed from the data to discard rooms with immense size for stable training.
Thus, 7044/904/887 rooms are collected as train/val/test rooms.
The incomplete point cloud is obtained by randomly sampling with density 219 points/$\text{m}^2$ and then removed and noised as in ShapeNet scene, but with removal sphere distance proportional to the object size and noise of standard deviation of 1cm.
During evaluation, we sample points proportionally to the area for each ground truth surface with density of 875 points/$\text{m}^2$ to compute the metrics.
\textbf{ShapeNet.}
We use chair, sofa, table classes of ShapeNet~(\cite{chang2015shapenet}), where the signed distance values of the query points are obtained by using the code of DeepSDF~(\cite{park2019deepsdf}).
The shapes are normalized to fit in a bounding box of $[-1, 1]^3$.
The partial observation is generated by extracting points within a sphere of radius 0.5, centered at one of the surface points. Then we sample 1,024 points to create a sparse input
and compare the metrics by sampling 2,048 points for a fair comparison against cGAN~(\cite{wu2020msc}).
\newpage
\section{Proofs and Detailed Derivations}
\label{app:sec:proofs}
\subsection{Variational Bound Derivation}
\label{app:elbo_derivation}
We derive the variational lower bound for cGCA (Eq.~(\ref{eq:elbo})) below.
The lower bound is derived by the work of \cite{sohl2015nonequil} and we include it for completeness.
\begin{align*}
\log p_\theta (x)
&= \log \sum_{s^{0:T - 1}} p_\theta (s^{0:T - 1}, x) \\
&= \log \sum_{s^{0:T - 1}} q_\theta (s^{0:T-1} | x) \frac{p_\theta (s^{0:T - 1}, x)}{q_\theta (s^{0:T-1} | x)} \\
&\ge \sum_{s^{0:T - 1}} q_\theta (s^{0:T-1} | x) \log \frac{p_\theta (s^{0: T - 1}, x)}{q_\theta (s^{0:T-1} | x)} \quad (\because \text{Jensen's inequality}) \\
&= \sum_{s^{0:T - 1}} q_\theta (s^{0:T-1} | x) (
\log \frac{p (s^0)}{q (s^0)}
+ \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)}
+ \log p (x | s^{T -1})
)\\
&=
\log \frac{p (s^0)}{q (s^0)} +
\sum_{s^{0:T - 1}} q_\theta (s^{0:T-1} | x) \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)}
+\mathbb{E}_{q_\theta}[\log p_\theta(x | s^{T - 1})]
\end{align*}
The second term of the right hand side can be converted into KL divergence as following:
\begin{align*}
&\sum_{s^{0:T - 1}} q_\theta (s^{0:T-1} | x) \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)} \\
&= \sum_{s^{0:T - 1}} \prod_{0 \leq i < T - 1} q_\theta (s^{i + 1} |s^i, x) \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)} \\
&= \sum_{s^{0:T - 1}} \sum_{0 \leq t < T - 1} \prod_{\substack{0 \leq i < T - 1 \\ i\neq t}} q_\theta (s^{i + 1} |s^i, x) q_\theta (s^{t + 1} |s^t, s^T) \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)} \\
&= \sum_{0 \leq t < T - 1} \sum_{s^{t + 1}} (
\sum_{s^{-(t + 1)}}
\prod_{\substack{0 \leq i < T - 1 \\ i\neq t}} q_\theta (s^{i + 1} |s^i, x)
)
q_\theta (s^{t + 1} |s^t, x) \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)} \\
& \quad (s^{-(t + 1)} \text{ denotes variables $s^{0:T - 1}$ except } s^{t + 1})\\
&= \sum_{0 \leq t < T - 1} \sum_{s^{t + 1}}
q_\theta (s^{t + 1} |s^t, x) \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\theta (s^{t + 1}|s^t, x)}\\
&= \sum_{0 \leq t < T - 1} -\sum_{s^{t + 1}}
q_\theta (s^{t + 1} |s^t, x) \log \frac{q_\theta (s^{t + 1}|s^t, x)}{p_\theta (s^{t + 1}|s^t)}\\
&= \sum_{0 \leq t < T - 1} -D_{KL}(q_\theta (s^{t + 1}|s^t, x) \| p_\theta (s^{t + 1}|s^t))
\end{align*}
Thus,
\begin{equation*}
\log p_\theta(x) \ge \log \frac{p(s^0)}{q(s^0)} + \sum_{0 \le t < T - 1} -D_{KL}(q_\theta(s^{t + 1} | s^t, x) \| p_\theta(s^{t + 1} | s^t))) + \mathbb{E}_{q_\theta}[\log p_\theta(x | s^{T - 1})]
\end{equation*}
is derived.
\subsection{KL Divergence Decomposition}
\label{app:kl_decomposition}
We derive the decomposition of KL divergence (Eq.~(\ref{eq:transition_loss}) ) below.
\begin{align*}
\mathcal{L}_t
&= D_{KL}(q_\theta(s^{t + 1} | s^t, x) \| p_\theta(s^{t + 1} | s^t)) \\
&= \sum_{c \in \mathcal{N} (s^t)} D_{KL} (q_\theta(o_c, z_c | s^t, x) \| p_\theta(o_c, z_c | s^t)) \quad (\because \text{Eq.~(\ref{eq:transition_kernel}), Eq.(~\ref{eq:infusion_kernel}}) ) \\
\begin{split}
& =
\sum_{c \in \mathcal{N} (s^t)}
D_{KL} (q_\theta (o_c | s^t, x) \| p_\theta (o_c | s^t) ) \\
& \quad + \sum_{o_c \in \{0, 1\}} q_\theta (o_c| s^t, x) D_{KL} (q_\theta (z_c | s^t, x, o_c) \| p_\theta (z_c | s^t, o_c) ) \\
& \quad (\because \text{Eq.~(\ref{eq:transition_kernel}), Eq.~(\ref{eq:infusion_kernel}) and chain rule for conditional KL divergence})
\end{split} \\
\begin{split}
& =
\sum_{c \in \mathcal{N} (s^t)}
D_{KL} (q_\theta (o_c | s^t, x) \| p_\theta (o_c | s^t) ) \\
& \quad + q_\theta (o_c = 1| s^t, x) D_{KL} (q_\theta (z_c | s^t, x, o_c = 1) \| p_\theta (z_c | s^t, o_c = 1) ) \\
& \quad (\because p_\theta (z_c | s^t, o_c = 0) = q_\theta (z_c | s^t, x, o_c = 0) = \delta_{0}(z_c))
\end{split}
\end{align*}
Thus, Eq.~(\ref{eq:transition_loss}) is derived.
\subsection{Approximating $\mathcal{L}_{\mathrm{final}}$ by $\mathcal{L}_{T - 1}$}
In this section, we claim that $\mathcal{L}_{\mathrm{final}}$ can be replaced by $\mathcal{L}_{T - 1}$, which allows us to train using a simplified objective.
First, we show that $\mathcal{L}_{\mathrm{final}} = \mathbb{E}_{q_\theta}[\log p_\theta(x | s^{T - 1})] = \infty$ due to the usage of Dirac distribution and introduce a non-diverging likelihood by replacing the Dirac function $\delta_0 (z_c)$ with indicator function $\mathds{1}[z_c = 0]$.
Recall,
\begin{align*}
\mathcal{L}_{\mathrm{final}}
&= \mathbb{E}_{q_\theta}[\log p_\theta(x | s^{T - 1})] \\
&= \sum_{c \in \mathcal{N} (s^{T-1})} \log \{
(1 - \lambda_{\theta, c} ) \mathds{1}[o_c^x=0] \delta_{0}(z_c^x) + \lambda_{\theta, c} \mathds{1}[o_c^x=1] {N}(z^x_c; \; \mu_{\theta, c}, \sigma^{T-1} {\bm{I}})
\}
\end{align*}
Thus, if any cell $c \in \mathcal{N} (s^{T - 1})$ is unoccupied, the likelihood diverges to $\infty$ since $z^x_c=0$.
This disables us to use negative log-likelihood as a loss function, since all the gradients will diverge.
This problem can be easily resolved by substituting $\delta_0 (z^x_c)$ with indicator function $\mathds{1}[z^x_c = 0]$.
Both functions have non-zero value only at $z_c \neq 0$, but the former has value of $\infty$ while the latter has value $1$.
While $\mathds{1}[z^x_c = 0]$ is not a probability measure, i.e. $\int_{z^x_c} \mathds{1}[z^x_c = 0] \neq 1$, replacing $\delta_0 (z^x_c)$ will have the effect of reweighting the likelihood at $z^x_c=0$ from $\infty$ to $1$, with the likelihood at other values $z^x_c \neq 0$ unchanged.
Thus, $\mathds{1}[z^x_c = 0]$ is a natural replacement of $\delta_0 (z^x_c)$ for computing a valid likelihood.
\label{app:final_state_loss}
\auxloss*
\begin{proof}
For brevity of notations, we define
\begin{equation*}
\sigma = \sigma^t,
\end{equation*}
\begin{equation*}
\lambda_{q, c} = (1 - \alpha^t) \lambda_{\theta, c} + \alpha^t \mathds{1}[c \in G_{x}(s^t)],
\end{equation*}
\begin{equation*}
\mu_{q, c} = (1 - \alpha^t) \mu_{\theta, c} + \alpha^t z^x_c,
\end{equation*}
where $\lambda_{q, c}, \mu_{q, c}$ is the occupancy probability, and mean of the infusion kernel $q$ at cell $c$.
With $T \gg 1$, there exists $T_1 < T$, such that $\alpha^{T_1} = 1$, since $\alpha^t = \max(\alpha_1t + \alpha_0, 1)$.
Since $d(c, c') = 0$ if and only if $c = c'$, $G_{x}(s^t) = \{\mathrm{argmin}_{c \in \mathcal{N}(s^t)} d(c, c') \mid c' \in x\}$ must converge to a set of occupied coordinates of $x$ since the distance decreases for each cell $c \in x$ at every step.
This indicates that $\lambda_{q, c} = \mathds{1}[c \in G_{x}(s^{T - 1})] = o^x_c$.
Also, $\mu_{q, c} = z^x_c$ holds for all $t > T_1$.
Then, $\mathcal{L}_{T - 1}$ and $\mathcal{L}_{\mathrm{final}}$ can be expressed as the following:
\begin{align*}
\mathcal{L}_{T - 1}
&= - D_{KL} (q_\theta(s^T | s^{T - 1}, x) \| p_\theta(s^T | s^{T - 1}) \\
\begin{split}
& = - \sum_{c \in \mathcal{N} (s^{T - 1})}
D_{KL} (q_\theta (o^T_c | s^{T - 1}, x) \| p_\theta (o^T_c | s^{T - 1}) ) \\
& \quad + q_\theta (o^T_c = 1| s^{T - 1}, x) D_{KL} (q_\theta (z_c | s^{T - 1}, x, o^T_c = 1) \| p_\theta (z_c | s^{T - 1}, o^T_c = 1) ) \quad (\because \text{Eq.~(\ref{eq:transition_loss})}) \\
\end{split} \\
&= - \sum_{c \in \mathcal{N} (s^{T - 1})}
\lambda_{q, c} \log \frac{\lambda_{q, c} }{\lambda_{\theta, c} } + (1 - \lambda_{q, c} ) \log \frac{ 1 - \lambda_{q, c}} { 1 - \lambda_{\theta, c}} + \lambda_{q, c} \frac{1}{2\sigma^2} \|\mu_{\theta, c} - \mu_{q, c}\|^2 \\
&= - \sum_{c \in \mathcal{N} (s^{T - 1})}
o^x_c \log \frac{o^x_c }{\lambda_{\theta, c} } + (1 - o^x_c) \log \frac{ 1 - o^x_c} { 1 - \lambda_{\theta, c}} + o^x_c \frac{1}{2\sigma^2} \|z^x_c - \mu_{q, c}\|^2 \\
& \quad (\because \lambda_{\theta, c} = o^x_c, \mu_{q, c} = z^x_c)
\end{align*}
\begin{align*}
\mathcal{L}_{\mathrm{final}}
&= \mathbb{E}_{q_\theta}[\log p_\theta(x | s^{T - 1})] \\
&= \sum_{c \in \mathcal{N} (s^{T-1})} \log \{
(1 - \lambda_{\theta, c} ) \mathds{1}[o_c^x=0] \delta_{0}(z_c^x) + \lambda_{\theta, c} \mathds{1}[o_c^x=1] {N}(z^x_c; \; \mu_{\theta, c}, \sigma^{T-1} {\bm{I}})
\} \\
&= \sum_{c \in \mathcal{N} (s^{T - 1})} \log \{
(1 - \lambda_{\theta, c} ) \mathds{1}[o^x_c=0] \mathds{1}[z^x_c = 0] \\
& \quad+ \lambda_{\theta, c} \mathds{1}[o^x_c=1] \exp(-\frac{1}{2\sigma^2} \| z^x_c - \mu_{\theta, c}\|^2_2 ) (2 \pi \sigma)^{-\frac{K}{2}}
\}
\end{align*}
We divide cases for $o^x_c = 0$ and $o^x_c = 1$.
When $o^x_c = 0$,
\begin{equation*}
\mathcal{L}_{T - 1} = \mathcal{L}_{\mathrm{final}} = \sum_{c \in \mathcal{N} (s^{T - 1})} \log (1 - \lambda_{\theta, c}) ,
\end{equation*}
where we set $y\log y = 0$ for $y=0$, since $\lim_{y\rightarrow 0} y \log y = 0$.
Analogously, when $o^x_c = 1$,
\begin{equation*}
\mathcal{L}_{T - 1} = \sum_{c \in \mathcal{N} (s^{T - 1})} \log \lambda_{\theta, c} - \frac{1}{2\sigma^2} \|z^x_c - \mu_{\theta, c}\|^2
\end{equation*}
\begin{equation*}
\mathcal{L}_{\mathrm{final}} = \sum_{c \in \mathcal{N} (s^{T - 1})} \log \lambda_{\theta, c} - \frac{1}{2\sigma^2} \|z^x_c - \mu_{\theta, c}\|^2 - \frac{K}{2} \log (2 \pi \sigma).
\end{equation*}
Thus, if $o^x_c = 1$, $\mathcal{L}_{T - 1} = \mathcal{L}_{\mathrm{final}} - \frac{K}{2} \log (2 \pi \sigma)$, where $ - \frac{K}{2} \log (2 \pi \sigma)$ is a constant.
We can conclude that $\nabla_\theta \mathcal{L}_{T - 1} = \nabla_\theta \mathcal{L}_{\mathrm{final}}$ for all $o^x_c$.
\end{proof}
\newpage
\section{Difference of Formulation compared to GCA}
\label{app:gca_difference}
We elaborate the difference of formulations compared to GCA~(\cite{zhang2021gca}).
\textbf{Usage of Latent Code.}
The biggest difference between GCA and our model is the usage of local latent codes~(\cite{chiyu2020local, chabra2020deep_local_shapes}) to generate continuous surface.
Although the formulation of GCA allows scalable sampling of high resolution voxel, it cannot produce continuous surface due to predefined voxel resolution.
However, we define a new state named sparse voxel embedding by appending local latent codes to the voxel occupancy.
Sparse voxel embedding can be decoded to produce continuous geometry, outperforming GCA in accuracy in all the evaluated datasets in Sec.~\ref{sec:experiments}.
\textbf{Training Disconnected Objects.}
During training, GCA assumes a property named partial connectivity between an intermediate state $s^t$ and ground truth state $x$ to guarantee the convergence of infusion sequences to ground truth $x$ (Property 1 and Proposition 1 of \cite{zhang2021gca}).
The connectivity means that any cell in $s^t$ can reach any cell in $x$ following the path of cells in $x$ with local transitions bounded by neighborhood size $r$.
The assumption is required since the infusion kernel of single cell $q^t(o_c | s^t)$ for GCA, defined as
\begin{equation}
q^t(o_c | s^t) = Ber((1 - \alpha^t) \lambda_{\theta, c} + \alpha^t \mathds{1}[c \in x]),
\end{equation}
inserts the ground truth cell $c \in x$ if the cell is within the neighborhood of current state $\mathcal{N}(s^t)$, i.e. $c \in \mathcal{N}(s^t) \cap x$.
However, we eliminate the assumption by defining an auxiliary function
\begin{equation}
G_x(s) = \{\argmin_{c \in \mathcal{N}(s)} d(c, c')| c' \in x\},
\end{equation}
and modifying the formulation as
\begin{equation}
q^t(o_c | s^t) = Ber((1 - \alpha^t) \lambda_{\theta, c} + \alpha^t \mathds{1}[c \in G_x(s^t)]),
\end{equation}
by switching $x$ in the infusion term to $G_x(s^t)$ from the original infusion kernel.
The usage of $G_x(s^t)$ guarantees infusing a temporary ground truth cell $c$ that assures reaching all cells in $x$ regardless of the connectivity.
The assumption is no longer required and therefore our training procedure is a general version of GCA.
\textbf{Training Procedure for Maximizing Log-likelihood.}
Both GCA and our model utilizes the infusion training~(\cite{bordes2017infusion}), by emulating the sequence of states that we aim to learn.
The objective of GCA does not explicitly show the relationship between the training objective and maximizing the log-likelihood, which is what we aim to achieve.
In Sec.~\ref{sec:training}, we give a detailed explanation of how our training objective approximates maximizing the lower bound of log-likelihood of data distribution.
Thus, our training procedure completes the heuristics used in original work of GCA.
\section{Analysis on Scalability}
\label{app:scalabilty}
\input{table/appendix/result_scalability}
In this section, we investigate the scalability of our approach.
Scene completion needs to process the 3D geometry of large-scale scenes with multiple plausible outputs, and sparse representation is crucial to find the diverse yet detailed solutions.
As the scale of the 3D geometry increases, the approaches utilizing the dense convolutional network cannot observe the entire scene at the same time.
As a simple remedy, previous methods employ sliding window techniques for 3D scene completion~(\cite{siddiqui2021retrievalfuse, peng2020conv_onet}).
Basically they divide the input scene into small segments, complete each segment, and then fuse them.
The formulation can only observe the information that is within the same segment for completion, and assumes that the divisions contain all the information necessary to complete the geometry.
The size of the maximum 3D volume that recent works utilize amounts to a 3D box whose longest length is 3.2m for $64^3$ resolution voxel, where individual cells are about $5$~cm in each dimension.
Note that the size is comparable to ordinary furniture, such as sofa or dining table, whose spatial context might not be contained within the same segment.
The relative positions between relevant objects can be ignored, and the missing portion of the same object can accidentally be separated into different segments.
With our sparse representation, the receptive field is large enough to observe the entire room, and we can stably complete the challenging 3D scenes with multiple objects capturing the mutual context.
We further provide the numerical comparison of our sparse representation against the dense convolution of ConvOcc~(\cite{peng2020conv_onet}).
Table~\ref{table:result_scalabilty} shows the number of neural network parameters and the GPU usage for processing single rooms of different sizes in 3DFront~(\cite{fu20203dfront}).
We used 5~cm voxel resolution and the grid size indicates the number of cells required to cover the longest length of the bounding box for the room.
The maximum GPU usage during 25 transitions of cGCA is provided. %
We did not include the memory used for the final reconstruction which regresses the actual distance values at dense query points, since the memory usage differs depending on the number of query points.
Our neural network is powerful, employing about 10 times more parameters than ConvOcc.
With the efficient sparse representation, we can design the network to capture larger receptive fields with deeper network, enjoying more expressive power.
This is reflected in the detailed reconstruction results of our approach.
The memory usage is similar for smaller rooms for both approaches, but cGCA becomes much more efficient for larger rooms.
When the grid size for room reaches 235, our method consumes 7 times less GPU memory compared to ConvOcc.
As widely known, the memory and computation for 3D convolution increases cubic to the grid resolution with dense representation, while those for sparse representations increase with respect to the number of occupied voxels, which is much more efficient.
Therefore our sparse representation is scalable and crucial for scene completion.
\section{Effects of Mode Seeking Steps}
\label{app:ablation_mode_seeking}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{figures/appendix/mode_seeking.pdf}
\caption{
Ablation study on the effects of mode seeking step $T'$.
The left shows an example of chair generation, where the bottom row are sparse voxel embeddings and the top row is the mesh reconstruction of each state.
The right is the result on the effects of mode seeking step $T'$ tested with probabilistic shape completion on sofa dataset.
}
\label{fig:app:mode_seeking}
\end{figure}
We perform an ablation study on the effects of mode seeking steps, which removes the noisy voxels with low-occupancy probability as discussed in Sec.~\ref{sec:sampling}.
We test with ShapeNet dataset~(\cite{chang2015shapenet}) with voxel size $64^3$ and neighborhood radius $r=3$, which tends to show the most intense effect of mode seeking phase due to large neighborhood size.
The right side of Fig.~\ref{fig:app:mode_seeking} shows the graph of metrics on quality (MMD), fidelity (UHD), and diversity (TMD) with respect to mode seeking steps.
There is a dramatic decrease in all metrics on the first step of mode seeking step.
The effect is visualized in the left side of Fig.~\ref{fig:app:mode_seeking}, where the unlikely voxels are removed in a single mode seeking step.
After the first mode seeking step, UHD and TMD metrics increase slightly, but overall, the metrics remain stable.
We choose $T'=5$ for all the conducted experiments to ensure that shape reaches a stable mode.
\section{Analysis on Varying Completeness of Input}
\label{app:ablation_completeness}
In this section, we further investigate the effects on the level of completeness of input.
Sec.~\ref{app:sparse_input} shows results of completion models on sparse inputs and Sec.~\ref{app:non_ambiguous_input} provides how our model behaves when a non-ambiguous input is given.
\subsection{Results on Sparse Input}
\label{app:sparse_input}
\input{table/appendix/result_sparse_input}
\begin{figure}[h!]
\centering
\includegraphics[width=0.97\textwidth]{figures/appendix/sparsity.pdf}
\caption{
Qualitative results on ShapeNet scenes with varying level of density.
Note that all the models are trained on dataset containing 10,000 points (rightmost column), but only tested with different density.
While the probabilistic methods (GCA, cGCA) tries to generate the learned shapes (e.g. shades of lamp) with only 500 points, deterministic methods (ConvOcc, IFNet) tend to fill the gaps between the points of the partial observation.
}
\label{fig:app:sparsity}
\end{figure}
For the experiments of ShapeNet scene dataset, the models are trained and evaluated on the input where the ambiguity lies on the missing parts, but the remaining parts of input pointcloud are relatively clear by sampling 10,000 points, following the experiment setting of ConvOcc~(\cite{peng2020conv_onet}).
In this section, we further provide ambiguity to the remaining parts by sampling less points.
For the models trained on 10,000 point samples with 0.5 completeness, we evaluate the results on sparse input by sampling 500, 1,000, 5,000 and 10,000 points after applying the same iterative removal process as done in training.
As in Sec.~\ref{sec:scene_completion}, we measure the accuracy (CD) and diversity (TMD) metrics compared with ConvOcc (\cite{peng2020conv_onet}), IFNet (\cite{chibane20ifnet}), and GCA (\cite{zhang2021gca}).
Table~\ref{table:result_sparse_input} contains the quantitative comparison results.
For all the models, accuracy (CD) degrades as the input gets sparser.
However, cGCA (w/ cond.) achieves best accuracy at all times, indicating that our model is the most robust to the sparsity of input.
For probabilistic models, the diversity (TMD) increases as the density decreases.
This implies that the diversity of the reconstructions is dependent to the ambiguity of the input.
This is intuitively a natural response, since the multi-modality of plausible outcomes increases as the partial observations become less informative.
Fig.~\ref{fig:app:sparsity} visualizes the results for various sparsity.
All of the models produce less plausible reconstructions with 500 points (leftmost column of Fig.~\ref{fig:app:sparsity}) compared to that of higher density inputs.
However, we observe a clear distinction between the completions of probabilistic models (GCA, cGCA) and deterministic models (ConvOcc, IFNet).
While probabilistic models try to generate the learned shapes from the observations, such as a shade of the lamp, deterministic methods tend to fill the gaps between points.
As discussed in Sec.~\ref{sec:scene_completion}, we hypothesize that this consequence stems from the fact that deterministic methods produce blurry results since it is forced to generate a single result along out of multiple plausible shapes which are the modes of the multi-modal distribution~(\cite{goodfellow2017gan_tutorial}).
Therefore, we claim that a generative model capable of modeling multi-modal distribution should be used for shape completion task.
\subsection{Results on Non-ambiguous Input}
\label{app:non_ambiguous_input}
\input{table/appendix/result_non_ambiguous_input}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{figures/appendix/non_ambiguous.pdf}
\caption{
Qualitative results of cGCA tested on ShapeNet scenes where the input is non-ambiguous.
The model is trained on completeness of 0.2.
}
\label{fig:app:non_ambiguous}
\end{figure}
We present experiments where cGCA is given a non-ambiguous input.
For the models trained on various level of completeness on ShapeNet scene dataset, we evaluate each model on an input presenting distinct geometry.
The input is generated by sampling 100,000 points from the mesh without any procedure of removal.
Table~\ref{table:result_non_ambiguous_input} shows the quantitative results of the experiments.
The first row shows the accuracy (CD) and diversity (TMD) scores of the models operated on test set when the removal procedure is same as that of the training set.
The second row shows the metrics when the input of the test set is sampled densely (100,000 points) without any removal.
For all models trained with varying level of completeness, accuracy increases with decreasing diversity given a clear shape compared to an ambiguous shape.
The result indicates that the diversity of the trained models is dependent on the ambiguity of the input.
This coincides with the result of Sec.~\ref{app:sparse_input} that the reconstructions of our model is dependent on the multi-modality of plausible outcomes of input.
We also observe that the diversity of reconstructions is associated with the training data.
Comparing the quantitative results on non-ambiguous input, the accuracy is low while the diversity is high when the model is trained on relatively incomplete data. %
We hypothesize that models trained with higher level of incompleteness are more likely to generate shapes from the input.
However, Fig.~\ref{fig:app:non_ambiguous} shows that while the completions of cGCA trained on highly incomplete data (min. rate 0.2) are diverse, they are quite plausible.
\section{Probabilistic Scene Completion on Real Dataset}
\label{app:scannet}
We investigate how our model behaves on indoor real-world data.
We test on the ScanNet indoor scene dataset~(\cite{dai2017scannet}), which is highly incomplete compared to other datasets, such as Matterport~(\cite{Matterport3D}).
The input is sampled by collecting 500 points/$\text{m}^2$ from each scene.
ScanNet dataset does not have a complete ground truth mesh, so we test our model trained on 3DFront~(\cite{fu20203dfront}) dataset with completness of 0.8.
We align the walls of ScanNet data to xy-axis since the scenes of 3DFront are axis aligned.
We compare our result with GCA~(\cite{zhang2021gca}).
The results are visualized in Fig.~\ref{fig:app:result_scannet}.
Our model shows diverse results (e.g. chairs, closets) and generalizes well to the real data, which has significantly different statistics compared to the training data.
Especially, the conditioned variant of our model shows better results by generalizing well to new data (e.g. tree), not found in 3DFront training dataset.
We emphasize that our model is trained on unsigned distance fields and does not require the sliding-window technique, unlike the previous methods~(\cite{peng2020conv_onet, chibane20ifnet}).
\begin{figure}[h!]
\centering
\includegraphics[width=0.97\textwidth]{figures/appendix/scannet_appendix.pdf}
\caption{
Qualitative results on ScanNet dataset with models trained on 3DFront dataset.
Best viewed on screen.
Multiple plausible reconstructions are shown (pink and blue box).
cGCA (w/ cond.) shows better results for reconstructing a tree (green box) compared to that of vanilla cGCA, where a tree is never found in the training dataset.
This allows us to infer that the conditioned variant tends to help generalize to unseen data better compared to the vanilla cGCA.
}
\label{fig:app:result_scannet}
\end{figure}
\newpage
\section{Additional Results for Scene Completion}
\subsection{Additional Results for ShapeNet Scene}
\label{app:result_synthetic}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{figures/appendix/synthetic_appendix_1.pdf}
\caption{
Additional qualitative results on ShapeNet scene, with varying level of completeness.
Best viewed on screen.
}
\label{fig:app:result_synthetic_1}
\end{figure}
\newpage
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{figures/appendix/synthetic_appendix_2.pdf}
\caption{
Additional qualitative results on ShapeNet scene, with varying level of completeness.
Best viewed on screen.
}
\label{fig:app:result_synthetic_2}
\end{figure}
\newpage
\subsection{Additional Results for 3DFront}
\label{app:result_3dfront}
\begin{figure}[h!]
\centering
\includegraphics[width=0.95\textwidth]{figures/appendix/3dfront_appendix_1.pdf}
\caption{
Additional qualitative results on 3DFront, with varying level of completeness.
Best viewed on screen.
}
\label{fig:app:result_3dfront_1}
\end{figure}
\newpage
\begin{figure}[h!]
\centering
\includegraphics[width=0.95\textwidth]{figures/appendix/3dfront_appendix_2.pdf}
\caption{
Additional qualitative results on 3DFront, with varying level of completeness.
Best viewed on screen.
}
\label{fig:app:result_3dfront_2}
\end{figure}
\newpage
\section{Additional Results for Single Object Completion}
\begin{figure}[h!]
\centering
\includegraphics[width=0.95\textwidth]{figures/appendix/shapenet_sofa.png}
\caption{Additional qualitative results on ShapeNet sofa.}
\label{fig:app:shapenet_sofa}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.95\textwidth]{figures/appendix/shapenet_chair.png}
\caption{Additional qualitative results on ShapeNet chair.}
\label{fig:app:shapenet_sofa}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.95\textwidth]{figures/appendix/shapenet_table.png}
\caption{Additional qualitative results on ShapeNet table.}
\label{fig:app:shapenet_table}
\end{figure}
\section{Proofs and Derivations of Maximizing Variational Lower Bound}
\label{appendix:proofs_and_derivations}
\subsection{Proof of Proposition 1} \label{appendix:prop1}
Recall the proposition~\ref{prop:infusion_convergence}.
\infusionconvergence*
\begin{proof}
Since $\alpha^t$ is increasing, there exists $T_0$ such that $\alpha^{T_0} = 1$.
Then the infusion transition kernel $q$ becomes a deterministic function with next state defined by $s^{t + 1} = G_x(s^t),\; \forall t \ge T_0$.
We now prove that the sequence $s^{t + 1} = G_x(s^t)$ converges to $x$ in a finite number of steps.
x
We first find $T_1$ such that $x \subset s^{T_0 + T_1}$ for $t \ge T_0 + T_1$.
Let $c \in x$.
We define a sequence of coordinates $c_t = argmin_{c' \in s^{T_0 +t}} d(c, c')$ and sequence of distances $d_t = \min_{c' \in s^{T_0 +t}}$, $d(c, c')$.
Since $s^{t + 1} = G_x(s^t) = \{\mathrm{argmin}_{c \in \mathcal{N}(s^t)} d(c, c') \mid c' \in x\}$ and $s^t \subsetneq \mathcal{N}(s^t)$, $d_t$ is a decreasing function bounded by $0$.
Thus, there exists $T_1$ such that $d_{T_1} = 0$, which implies that corresponding coordinate $c_{T_1} = c$.
Now, consider a state $s^{T_0 + T_1 + 1} = G_x(s^{T_0 + T_1}) = \{\mathrm{argmin}_{c \in \mathcal{N}(s^{T_0 + T_1})} d(c, c') \mid c' \in x\}$.
$x \subset s^{T_0 + T_1} \subset \mathcal{N}(s^{T_0 + T_1})$.
This implies $\min_{c \in \mathcal{N}(s^{T_0 + T_1})} d(c, c') = 0$ for all $c' \in x$.
Thus, $x = s^{T_0 + T_1 + 1}$.
Setting $T = T_0 + T_1 + 1$ concludes the proof.x
\end{proof}
\vspace{-1em}
\subsection{Detailed Derivation of Eq~\ref{eq:elbo}}
\label{section:detailed_derivation}
\begin{align*}
\log p_\theta (s^T)
&= \log \sum_{s^{0:T - 1}} p_\theta (s^{0:T}) \\
&= \log \sum_{s^{0:T - 1}} q_\phi (s^{0:T-1} | s^T) \frac{p_\theta (s^{0:T})}{q_\phi (s^{0:T-1} | s^T)} \\
&\ge \sum_{s^{0:T - 1}} q_\phi (s^{0:T-1} | s^T) \log \frac{p_\theta (s^{0:T})}{q_\phi (s^{0:T-1} | s^T)} \quad (\because \text{Jensen's inequality}) \\
&= \sum_{s^{0:T - 1}} q_\phi (s^{0:T-1} | s^T) (
\log \frac{p (s^0)}{q (s^0)}
+ \log p (s^T | s^{T -1 })
+ \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\phi (s^{t + 1}|s^t, s^T)}
)
\end{align*}
The third term of right hand side can be converted into KL divergence as following:
\begin{align*}
&\sum_{s^{0:T - 1}} q_\phi (s^{0:T-1} | s^T) \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\phi (s^{t + 1}|s^t, s^T)} \\
&= \sum_{s^{0:T - 1}} \prod_{0 \leq i < T - 1} q_\phi (s^{i + 1} |s^i, s^T) \sum_{0 \leq t < T - 1} \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\phi (s^{t + 1}|s^t, s^T)} \\
&= \sum_{s^{0:T - 1}} \sum_{0 \leq t < T - 1} \prod_{\substack{0 \leq i < T - 1 \\ i\neq t}} q_\phi (s^{i + 1} |s^i, s^T) q_\phi (s^{t + 1} |s^t, s^T) \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\phi (s^{t + 1}|s^t, s^T)} \\
&= \sum_{0 \leq t < T - 1} \sum_{s^{t + 1}} (
\sum_{s^{-(t + 1)}}
\prod_{\substack{0 \leq i < T - 1 \\ i\neq t}} q_\phi (s^{i + 1} |s^i, s^T)
)
q_\phi (s^{t + 1} |s^t, s^T) \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\phi (s^{t + 1}|s^t, s^T)} \\
& \quad (s^{-(t + 1)} \text{ denotes variables $s^{0:T - 1}$ except } s^{t + 1})\\
&= \sum_{0 \leq t < T - 1} \sum_{s^{t + 1}}
q_\phi (s^{t + 1} |s^t, s^T) \log \frac{p_\theta (s^{t + 1}|s^t)}{q_\phi (s^{t + 1}|s^t, s^T)}\\
&= -\sum_{0 \leq t < T - 1} \sum_{s^{t + 1}}
q_\phi (s^{t + 1} |s^t, s^T) \log \frac{q_\phi (s^{t + 1}|s^t, s^T)}{p_\theta (s^{t + 1}|s^t)}\\
&= -\sum_{0 \leq t < T - 1} D_{KL}(q_\phi (s^{t + 1}|s^t, s^T) || p_\theta (s^{t + 1}|s^t))
\end{align*}
Thus,
\begin{equation*}
\log p_\theta(s^T) = \log \frac{p(s^0)}{q(s^0)} + \mathbb{E}_{q_\phi}[\log p_\theta(s^T | s^{T - 1})] - \sum_{0 \le t < T - 1} D_{KL}(q_\phi(s^{t + 1} | s^t, s^T) || p_\theta(s^{t + 1} | s^t)))
\end{equation*}
is derived.
\subsection{Proof of Proposition 2} \label{appendix:prop2}
Recall the proposition~\ref{prop:maximizing_lower_bound}.
\maximizinglowerbound*
We first show a lemma that the KL divergence of factorized distribution can be decomposed as
sum of factorization.
\begin{lemma} \label{lemma:kl_factorization}
Suppose the two distribution $q, p$ can be factorized as following:
\begin{align*}
q(z_1, z_2, ..., z_N) = \prod_{0 \le i \le N} q(z_i) \\
p(z_1, z_2, ..., z_N) = \prod_{0 \le i \le N} p(z_i)
\end{align*}
Then,
\begin{equation*}
D_{KL} (q(z_1, z_2, ..., z_N) || p(z_1, z_2, ..., z_N)) = \sum_{0 \le i \le N} D_{KL} (q(z_i) || p(z_i))
\end{equation*}
\begin{proof}
We first show the case where there are $N=2$ variables.
\begin{align*}
D_{KL}(q(z_1, z_2) || p(z_1, z_2)))
&=\sum_{z_1, z_2} q(z_1, z_2) \log \frac{q(z_1, z_2)}{p(z_1, z_2)} \\
&=\sum_{z_1, z_2} q(z_1)q( z_2) \log \frac{q(z_1) q(z_2)}{p(z_1) p(z_2)} \\
&=\sum_{z_1, z_2} q(z_1)q( z_2) ( \log \frac{q(z_1)}{p(z_1)} + \log \frac{q(z_2)}{p(z_2)}) \\
&=\sum_{z_1, z_2} q(z_1)q( z_2) \log \frac{q(z_1)}{p(z_1)} + \sum_{z_1, z_2} q(z_1)q( z_2) \log \frac{q(z_2)}{p(z_2)} \\
&= \sum_{z_1} q(z_1) \log \frac{q(z_1)}{p(z_1)} (\sum_{z_2} q( z_2))
+ \sum_{z_2}q( z_2) \log \frac{q(z_2)}{p(z_2)} (\sum_{z_1} q( z_1)) \\
&=\sum_{z_1} q(z_1) \log \frac{q(z_1)}{p(z_1)} + \sum_{z_2}q( z_2) \log \frac{q(z_2)}{p(z_2)} \\
&= D_{KL} (q(z_1) || p(z_1)) + D_{KL} (q(z_2) || p(z_2))
\end{align*}
Using the above $N$ times iteratively by decomposing single variable at a time provides a direct proof for $N \ge 2$ case.
\end{proof}
\end{lemma}
\begin{proof}
We first factorize the KL divergence of state into cells and apply the definition of $q$ distribution.
\begin{align*}
D_{KL}(q_\theta(s^{t + 1} | s^t, s^T) || p_\theta(s^{t + 1} | s^t))
&= \sum_ {c \in s^{t + 1}}D_{KL}(q_\theta(c | s^t, s^T) || p_\theta(c| s^t)) \quad(\because \text{Lemma~\ref{lemma:kl_factorization}}) \\
&= \sum_ {c \in s^{t + 1}} q_\theta(c | s^t, s^T) \log \frac{q_\theta(c | s^t, s^T)}{p_\theta(c | s^t)} \\
&= \sum_ {c \in s^{t + 1}} ((1 - \alpha^t)p_\theta(c | s^t) + \alpha^t \delta_{G_x(s^t)}(c)) \log \frac{(1 - \alpha^t)p_\theta(c | s^t) + \alpha^t \delta_{G_x(s^t)}(c)}{p_\theta(c | s^t)} \\
&= \sum_ {c \in s^{t + 1}} h_{s^t}(p_\theta(c | s^t))
\end{align*}
where $h_{s^t}$ is defined as
\begin{equation*}
h_{s^t}(\mu) = ((1 - \alpha^t)\mu + \alpha^t \delta_{G_x(s^t)}(c)) \log \frac{(1 - \alpha^t)\mu + \alpha^t \delta_{G_x(s^t)}(c)} {\mu}.
\end{equation*}
We now show that $h_{s^t}(\mu)$ is a decreasing function with respect to $\mu$.
The cases are divided whether $\delta_{G_x(s^t)}(c) = 0$ or $\delta_{G_x(s^t)}(c) = 1$.
If $\delta_{G_x(s^t)}(c) = 0$,
\begin{equation*}
h_{s^t}(\mu) = (1 - \alpha^t) \mu \log (1 - \alpha^t) \\
\end{equation*}
\begin{align*}
\frac{d}{d\mu}h_{s^t}(\mu)
&= (1 - \alpha^t) \log (1 - \alpha^t) \\
&\leq 0 \quad (\because 0 < \alpha^t < 1)
\end{align*}
If $\delta_{G_x(s^t)}(c) = 1$,
\begin{equation}
h_{s^t}(\mu) = ((1 - \alpha^t)\mu + \alpha^t ) \log \frac{(1 - \alpha^t)\mu + \alpha^t} {\mu}
\end{equation}
\begin{align*}
\frac{d}{d\mu}h_{s^t}(\mu)
&= (1 - \alpha^t) \log (1 - \alpha^t + \frac{\alpha^t}{\mu}) - \frac{1}{\mu} \\
&= (1 - \alpha^t) (\log (1 - \alpha^t ) + \log (1 + \frac{\alpha^t}{(1 - \alpha^ t)\mu})) - \frac{1}{\mu} \\
&\leq (1 - \alpha^t) \log (1 - \alpha^t ) + \frac{\alpha^t}{\mu} - \frac{1}{\mu} \quad (\because \log (1 + x) \leq x, \; \forall x > 0) \\
&\leq (1 - \alpha^t) (\log (1 - \alpha^t ) - \frac{1}{\mu}) \\
&\leq 0 \quad (\because 0 < \alpha^t < 1)
\end{align*}
Thus, $h_{s^t}$ is a decreasing function concluding that $D_{KL}(q_\theta(s^{t + 1} | s^t, s^T) || p_\theta(s^{t + 1} | s^t)) \le D_{KL}(q_{\theta'}(s^{t + 1} | s^t, s^T) || p_{\theta'}(s^{t + 1} | s^t))$ for $ p_{\theta'}(c | s^t) \ge p_{\theta}(c | s^t), \; \forall c \in s^{t+1}$.
\end{proof}
\section{Comparison with GCA}
\label{appendix:gca_comparison}
We emphasize the stochastic voxel generation loss in cGCA models 1) theoretically valid loss while GCA does not, 2) can generate any shapes without connetivity assumption.
\section{Effects of Mode Seeking Steps}
\label{appendix:mode_seeking_steps}
\section{Implementation Details}
\label{appendix:implementation_details}
\subsection{Neural Network Architecture and Implementation Details}
\todo{checklist: hyperparameter}
\todo{checklist: computing resources}
\subsection{Data Preparation}
\label{appendix:data_preparation}
\todo{checklist: data split, both}\\
\textbf{ShapeNet.} \todo{include data splits, how they are chosen}
\textbf{3DFront.} \todo{include data splits, how they are chosen}
\label{appendix:baselines}
\subsection{Baselines}
\textbf{GCA~\cite{zhang2021gca}.} \todo{include hyperparameters}
\textbf{NDF~\cite{chibane2020ndf}.} \todo{include hyperparameters}
\subsection{Evaluation Metrics}
|
\section*{Introduction}
The paper is about the regularity of the value function and quantitative propagation of chaos for mean field control (MFC for short) problems obtained as the limit of optimal control problems for large particle systems {with forcing and terminal data which are not assumed to be convex.}. The value function of MFC problems is known to be a Lipschitz continuous but, in general, not $C^1-$function in the space of time and probability measures. Our first result is that there exists an open and dense subset of time and probability measures where
the value function is actually smooth. The second result is a new quantitative propagation of chaos-type property for the optimal solutions of the particle system starting from this open and dense set.
\vskip.075in
\subsection*{The background} In order to state the results it is necessary to introduce the general set-up.
\vskip.075in
We consider the problem of controlling optimally $N$ particles in order to minimize a criterion of the form
\begin{equation}\label{defJN}
J^N(t_0,{\bf x}_0, \alpha) :=\mathbb E\left[ \frac{1}{N} \int_{t_0}^T\sum_{i=1}^N L(X^i_t, \alpha^i_t)dt + \int_{t_0}^T \mathcal F(m^N_{{\bf X}_t})dt + \mathcal G(m^N_{{\bf X}_T})\right].
\end{equation}
Here $T>0$ is a fixed time horizon, $t_0\in [0,T]$ and ${\bf x}_0=(x^1_0,\dots, x^N_0)\in (\mathbb R^d)^N$ are respectively the initial time and the initial position of the system at time $t_0$. The minimization is over the set $d A^N$ of admissible controls $\alpha= (\alpha^k)_{k = 1}^N$ in $L^2([0,T] \times \Omega ; (\mathbb R^d)^N)$ which are adapted to the filtration generated by the independent $d-$dimensional Brownian motions $(B^i)_{i=1, \dots, N}$, and the trajectories
${\bf X}= (X^1, \dots, X^N)$ satisfy, for each $k\in \{1, \dots, N\}$,
$$
X^k_t = x^k_0+\int_{t_0}^t \alpha^k_s ds +\sqrt{2}(B^k_t - B^k_{t_0}) \ \ \text {for} \ \ t\in [t_0,T].
$$
In \eqref{defJN}, $m^N_{{\bf X}_t}$ is the empirical measure of the process ${\bf X}_t$ given by
\begin{equation}\label{m}
m^N_{{\bf X}_t}:= \dfrac{1}{N}\sum_{k=1}^N \delta_{X^k_t},
\end{equation}
where $\delta_x$ is the Dirac mass at $x$.
\vskip.075in
The maps $\mathcal F$ and $\mathcal G$, which are defined on (suitable subsets of) the set of Borel probability measures on $\mathbb R^d$, couple all the particles $X^i$. Finally, the cost function $L=L(x,\alpha):\mathbb R^d\times \mathbb R^d\to \mathbb R$ is, typically, convex and grows quadratically in the second variable.
\vskip.075in
The value function for this optimization problem reads
\begin{equation} \label{def.VN}
\begin{split}
\mathcal V^N(t_0, {\bf x}_0)&:= \inf_{\alpha \in \mathcal A^N} J^N(t_0,{\bf x}_0, \alpha)\\
&=\inf_{\alpha \in \mathcal A^N}
\mathbb E\left[ \int_{t_0}^T (\frac{1}{N}\sum_{k=1}^N L(X^k_t,\alpha^k_t) +\mathcal F(m^N_{{\bf X}_t}))dt +
\mathcal G(m^N_{{\bf X}_T}) \right].
\end{split}
\end{equation}
\vskip.075in
In a more general framework and under slightly different conditions on the data, Lacker \cite{La17} proved that the empirical measure $m^N_{{\bf X}^N_t}$ associated to the optimal trajectories of \eqref{def.VN} converges in a suitable sense to the (weak) optimal solution of the mean field control problem (written here in a strong sense) consisting in minimizing the quantity
\begin{equation}\label{defJinfty}
J^\infty(t_0, m_0, \alpha) = \mathbb E\left[ \int_{t_0}^T L(X_t, \alpha_t)dt + \int_{t_0}^T \mathcal F({\mathcal L}(X_t))dt + \mathcal G(\mathcal L (X_T))\right],
\end{equation}
where $m_0$ is an initial distribution of the particles at time $t_0$, $\alpha \in d A$, the set of admissible controls consisting of square integrable $\mathbb R^d-$valued processes adapted to a Brownian motion $B$ and to an initial condition $\bar X_0$, which is independent of $B$ and of law $m_0$, and the process $(X_t)_{t\in [t_0,T]} $ satisfies
$$
X_t= \bar X_0 +\int_{t_0}^t \alpha_sds + \sqrt{2}(B_t-B_{t_0}) \ \ \text{for} \ \ t\in [t_0,T],
$$
and $d L(X_t)$ denotes the law of $X_t$.
\vskip.075in
The value function $d U$ of the last optimization problem is given (heuristically at this stage) by
\begin{equation}\label{takis1}
\begin{split}
\mathcal U(t_0,m_0):&= \inf_{\alpha \in d A} J^\infty(t_0, m_0, \alpha) \\
&=\inf_{\alpha\in d A} \mathbb E[\int_{t_0}^T\big( L(X_t, \alpha_t) + d F( d L(X_t)\big) dt + d G(d L(X_T)].
\end{split}
\end{equation}
\vskip.075in
In addition, \cite{La17} points out that there is a propagation of chaos-type property (an easy consequence of Sznitman characterization of propagation of chaos), if the minimization problem \eqref{takis1} has a unique weak minimizer.
Note, however, that such uniqueness is known only when the maps $(x,\alpha)\to L(x,\alpha)$, $m\to \mathcal F(m)$ and $m\to \mathcal G(m)$ are globally convex. The conclusions of \cite{La17} were extended to problems with common noise in Djete, Possama\"i and Tan \cite{DPT20}. Several other results on the convergence of MFC problems without diffusion were obtained in Cavagnari, Lisini, Orrieri and Savar\'{e} \cite{CaLiOrSa} and Gangbo, Mayorga and Swiech \cite{GMS}. A quantitative convergence rate for the value function $\mathcal V^N$ to $\mathcal U$ was given, for problems on a finite state space, in Kolokoltsov \cite{Ko12} and Cecchin \cite{Ce21} and, for problems on the continuous state space, in Baryaktar and Chakraborty \cite{barcha21} under a certain structural dependence of the data on the measure variable, in Germain, Pham and Warin \cite{gpw21} under the assumption that the limit value is smooth, and in Cardaliaguet, Daudin, Jackson and Souganidis \cite{CDJS} under a decoupling assumption on the Hamiltonian. In addition, a propagation of chaos is proved in \cite{barcha21, Ce21, gpw21} assuming, however, that the limit value function is smooth.
\vskip.075in
\subsection*{The results}
In this paper we study nonconvex MFC problems for which the limit value function is not expected to be globally smooth and show that for a large class (a dense and open set) of initial times and measures the value function $d U$ is smooth (Theorem~\ref{thm.mainreguU0}) and the propagation of chaos holds with a rate (Theorem~\ref{thm.main20}).
\vskip.075in
To write the results in this introduction requires considerable notation. Thus we postpone stating the precise theorems to section~2.
\vskip.075in
We explain, however, the very general idea of proof. We identify the open and dense in time and space of probabilities set $\mathcal O$ where $d U$ is smooth as the set of initial conditions $(t_0,m_0)$ from which starts a unique (in a strong sense) and stable (in a suitable linear sense) minimizer of $J^\infty$. This step, which is reminiscent of ideas from standard optimal control (see, for instance, the book of Cannarsa and Sinestrari \cite{CaSiBook}), is similar to what was obtained by Briani and Cardaliaguet \cite{BrCa} for MFC problems in a different framework, namely, the state space is the torus and the initial measures have smooth densities. Here, the state space is all of $\mathbb R^d$ and the initial conditions are arbitrary probability measures. To show that $\mathcal U$ is smooth in $\mathcal O$, we adapt ideas used in the construction of a solution to the master equation in mean field games given in Cardaliaguet, Delarue, Lasry and Lions \cite{CDLL} but we argue without the convexity, which translates to monotonicity for general mean field games, assumption used extensively there.
Then, using the regularity of $\mathcal U$, we derive the propagation of chaos property for the optimal solution of the particle system when starting from the set $\mathcal O$.
The key argument is the fact that the optimal trajectories of \eqref{takis1} that start in $\mathcal O$ remain there, while $\mathcal U$ is smooth and satisfies (almost) the same Hamilton-Jacobi equation as $\mathcal V^N$ in $\mathcal O$.
\vskip.075in
We finally comment about possible extensions to problems with common noise { emphasizing once more that we do not assume any monotonicity/convexity. }The convergence (with algebraic rate) of $\mathcal V^N$ to $\mathcal U$ for MFC problems with common noise was established in \cite{CDJS}.
However, the generalization of the results of the present paper to such a setting is far from clear.
For example, one of the basic tools we use to prove that the set $\mathcal O$ is open and dense is a result of Lions-Malgrange-type, which yields the uniqueness of the mean field game system characterizing the optimal solution (see system \eqref{MFG} below), but with initial conditions both in $u$ and $m$: see the proof of Lemma \ref{lem.Odense}.
At this time, we do not know if there is a counterpart of this argument for problems with a common noise, where the mean field system becomes a system of forward-backward stochastic partial differential equations (see \cite{CDLL}).
\vskip.075in
\noindent {\bf Organization of the paper.} We conclude the introduction with the notation we use throughout the paper. In section \ref{sec.1}, we introduce the standing assumptions, present the main results and recall some preliminary facts that are needed for the rest of the paper.
Section \ref{sec.proofThm1} is about establishing the smoothness of $d U$ (Theorem \ref{thm.mainreguU0}). Section~\ref{sec.proofThm2} is devoted to showing the propagation of chaos property (Theorem~\ref{thm.main20}).
\subsection*{Notations}
We work on $\mathbb R^d$ and write $B_R$ for the open ball centered at the origin with radius $R$ and $I_d$ for the identity matrix.
\vskip.075in
We denote by $\mathcal P(\mathbb R^d)$ the set of Borel probability measures on $\mathbb R^d$. Given $m \in \mathcal P(\mathbb R^d)$ and $p \geq 1$, we write $M_p(m)$ for the $p^{th}-$moment of $m$, that is, $M_p(m) = \int_{\mathbb R^d} |x|^p dm$. Then $\mathcal P_p(\mathbb R^d)$ is the set of $m \in \mathcal P(\mathbb R^d)$ such that $M_p(\mathbb R^d) < \infty$. We endow $\mathcal P_p(\mathbb R^d)$ with the Wasserstein metric ${\bf d}_p$, defined by
\begin{align*}
{\bf d}^p_p(m, m') : = \inf_{\pi \in \Pi(m,m')} \int_{\mathbb R^d} |x - y|^p d\pi(x,y),
\end{align*}
where $\Pi(m,m')$ is the set of all $\pi \in \mathcal P(\mathbb R^d \times \mathbb R^d)$ with marginals $m$ and $m'$. For $p=1$, we recall the duality formula
\begin{align*}
{\bf d}_1(m,m') = \sup_{\phi} \int_{\mathbb R^d} \phi d(m- m'),
\end{align*}
where the supremum is taken over all 1-Lipschitz maps $\phi : \mathbb R^d \to \mathbb R$.
\vskip.075in
For ${\bf x} = (x^1,...,x^N) \in (\mathbb R^d)^N$, $m_{{\bf x}}^N \in \mathcal P(\mathbb R^d)$ is the empirical measure of ${\bf x}$, that is, $m_{{\bf x}}^N = \frac{1}{N} \sum_{i = 1}^N \delta_{x^i}$.
\vskip.075in
Given a map $U:\mathcal P_1(\R^d)\to \mathbb R$, we denote by $\delta U/\delta m$ its flat derivative and by $D_mU$ its Lions-derivative when these derivatives exist, and we use the corresponding notations for second order derivatives. We refer to \cite{CDLL} and the books of Carmona and Delarue \cite{CaDeBook} for definitions and properties.
\vskip.075in
We write $C^k_{loc}$ and $C^k$ for the sets of maps on $\mathbb R^d$ with continuous and continuous and bounded derivatives up to order $k$. For $r>0$, $r\notin \mathbb N$, we denote by $C^{r,2r}$ the standard parabolic H\"{o}lder spaces
and by $C^{r,2r}_c$ the subset of functions of $C^{r,2r}$ with a compact support.
\vskip.075in
Given a topological vector space, we write by $E'$ for its dual space.
\vskip.075in
We often need to compare continuous maps defined on different intervals of $[0,T]$. For this, we simply extend the maps continuously on $[0,T]$ by a constant. For instance, if $t_0\in (0,T]$, $E$ is a topological space and $f:[t_0,T]\to E$ is continuous, we set $f(s)=f(t_0)$ for $s\in [0,t_0]$.
\vskip.075in
{Finally, throughout the proofs $C$ denotes a positive constants, which, unless otherwise said, depends on the data and may change from line to line. }
\subsection*{Acknowlegments} Cardaliaguet was partially supported by the Air Force Office for Scientific Research grant FA9550-18-1-0494 and IMSI, the Institute for Mathematical and Statistical Innovation. Souganidis was partially supported by the National Science Foundation grant DMS-1900599, the Office for Naval Research grant N000141712095 and the Air Force Office for Scientific Research grant FA9550-18-1-0494. Both authors would like to thank the IMSI for its hospitality during the Fall 2021 program.
\section{The assumptions, the main results and some preliminary facts}\label{sec.1}
\subsection*{The standing assumptions} We state our standing assumptions on the maps $H:\mathbb R^d\times \mathbb R^d\to \mathbb R$, $\mathcal F:\mathcal P_1(\R^d)\to \mathbb R$ and $\mathcal G:\mathcal P_1(\R^d)\to \mathbb R$, which constitute the data of our problem. We recall that $L:\mathbb R^d\times \mathbb R^d\to \mathbb R$ is the Legendre transform of $H$ with respect to the second variable:
$$
L(x,a)= \sup_{p\in \mathbb R^d} -a\cdot p-H(x,p).
$$
We assume that
\begin{equation}\label{convex}
\begin{cases}
\text {$H=H(x,p):\mathbb R^d\times \mathbb R^d\to \mathbb R$ is of class $C^4_{loc}$ and strictly convex with respect to the}\\[1.2mm]
\text{second variable, that is, for each $R>0$ there exists}\\[1.2mm]
\text{ $c_R, C_R>0$ such that, for all $(x,p)\in \mathbb R^d\times \overline{B_R}$,}\\[1.2mm]
\qquad \qquad D^2_{pp}H(x,p)\geq c_R I_d \ \ \text{and} \ \ |D^2_{xx}H(x,p)| + |D^2_{xp}H(x,p)|\leq C_R,
\end{cases}
\end{equation}
\begin{equation}\label{hyp.growth_1}
\begin{cases}
\text{there exists a constant $C>0$ such that, for all $ (x,p)\in \mathbb R^d\times \mathbb R^d $,}\\[1.2mm]
-C+C^{-1}|p|^2\leq H(x,p)\leq C(1+|p|^2)\ \ \text{and} \ \ |D_xH(x,p)| \leq C(|p|+1),
\end{cases}
\end{equation}
\begin{equation}\label{F}
\text{$\mathcal F:\mathcal P_1(\R^d)\to \mathbb R$ is of class $C^2$ with $\mathcal F$, $D_m\mathcal F$, $D^2_{ym}\mathcal F$ and $D^2_{mm}\mathcal F$ uniformly bounded,}
\end{equation}
\begin{equation}\label{G}
\text{$\mathcal G:\mathcal P_1(\R^d)\to \mathbb R$ is of class $C^4$ with all derivatives up to order $4$ uniformly bounded.}
\end{equation}
\vskip.075in
For simplicity, in what follows we put together all the assumptions above in
\begin{equation}\label{ass.main}
\text{assume that \eqref{convex}, \eqref{hyp.growth_1}, \eqref{F} and \eqref{G}
hold.}
\end{equation}
The assumptions on the Hamiltonian $H$ are fairly standard, although a little restrictive, and are used in \cite{CDJS} to obtain, independent of $N$, Lipschitz estimates on the value function $\mathcal V^N$; see Lemma \ref{lem.estiVN}. An example satisfying \eqref{ass.main} is a Hamiltonian of the form $H(x,p)= |p|^2+ V(x)\cdot p$ for some smooth and globally Lipschitz continuous vector field $V:\mathbb R^d\to \mathbb R^d$. The regularity conditions on $\mathcal F$ and $\mathcal G$ are also important to obtain the estimates of Lemma \ref{lem.estiVN} and to prove the regularity of $\mathcal U$.
\subsection*{The results}
Given $(t_0, {\bf x})\in [0,T)\times (\mathbb R^d)^N$, $\mathcal V^N(t_0, {\bf x})$ is the value function of the optimal control of the $N-$particle problem given by \eqref{def.VN}.
\vskip.075in
We now define in a rigorous way the value function $\mathcal U$ of the MFC, which was informally introduced in \eqref{takis1}.
For each initial point $(t_0,m_0)\in [0,T)\times\mathcal P_1(\R^d) $, we use the set $d M(t_0,m_0)$ of controls given by
\begin{equation}\label{takis10}
d M(t_0,m_0):=\left\{\begin{array}{l}(m,\alpha)\in C^0([t_0,T], \mathcal P_1(\R^d))\times L^0([t_0,T]\times \mathbb R^d; \mathbb R^d ) \;: \\[1.5mm]
\int_0^T\int_{\mathbb R^d}|\alpha|^2m<\infty \ \ \text{and} \ \ \\[1.5mm]
\partial_t m-\Delta m +{\rm div}(m\alpha) = 0\; \; {\rm in}\; \; (t_0,T)\times \mathbb R^d \; \;\text{and} \; \; m(0)=m_0\; {\rm in}\; \mathbb R^d\end{array}\right\},
\end{equation}
with the equation in \eqref{takis10} understood in the sense of distributions.
\vskip.075in
Then the value function $\mathcal U(t_0,m_0)$ of the MFC problem is given by
\begin{equation}\label{takis19}
\mathcal U(t_0,m_0):=\underset{(m,\alpha)\in d M(t_0,m_0)} \inf\left\{ \int_{t_0}^T ((\int_{\mathbb R^d} L(x, \alpha(t,x))m(t,dx))+\mathcal F(m(t)) dt +\mathcal G(m(T))\right\}.
\end{equation}
Our first result is about the regularity of $\mathcal U$ in the set $\mathcal O$ defined by
\begin{equation}\label{takis20}
\mathcal O:=\left\{\begin{array}{l} (t_0,m_0)\in [0,T)\times \mathcal P_2(\R^d) \: : \mbox{\rm there exists a unique stable minimizer}\\[1.5mm]
\text{ in the definition of $\mathcal U(t_0,m_0)$}
\end{array}\right\}.
\end{equation}
The notion of stability is defined in terms of the linearized MFG system and is introduced in section~\ref{sec.proofThm1}.
\vskip.075in
The first main result is stated next.
\begin{thm}\label{thm.mainreguU0} Assume \eqref{ass.main}. The value function $\mathcal U$ is globally Lipschitz continuous on $[0,T]\times \mathcal P_1(\R^d)$ and of class $C^1$ in the set $\mathcal O$, which is open and dense in $[0,T]\times \mathcal P_2(\R^d)$. Moreover, $\mathcal U$ is a classical solution in $\mathcal O$ of the master Hamilton-Jacobi equation
\begin{equation}\label{HJdsO0}
\begin{array}{l}
\displaystyle -\partial_t \mathcal U(t,m)-\int_{\mathbb R^d}{\rm div}( D_m\mathcal U(t,m,y))m(dy) +\int_{\mathbb R^d} H(y, D_m\mathcal U(t,m,y))m(dy)=\mathcal F(m).
\end{array}
\end{equation}
In addition, for any $(t_0,m_0)$ in $\mathcal O$, there exists $\varepsilon>0$ and a constant $C>0$, which depends on $(t_0,m_0)$ and is such that, for any $t\in [0,T]$, $x,y \in \mathbb R^d$ and $m^1,m^2\in \mathcal P_2(\R^d)$ with $|t-t_0|<\varepsilon$, ${\bf d_2}(m_0,m^1)<\varepsilon$ and ${\bf d_2}(m_0,m^2)<\varepsilon$,
\begin{equation}\label{takis50}
\left| D_m\mathcal U(t,m^1,x)-D_m\mathcal U(t,m^2,y)\right| \leq C(|x-y|+ {\bf d}_2(m^1,m^2)).
\end{equation}
\end{thm}
\vskip.075in
By a classical solution of \eqref{HJdsO0}, we mean that the derivatives of $\mathcal U$ involved in the equation exist and are continuous.
\vskip.075in
Our second main result is a quantitative propagation of chaos property about the optimal trajectories of the underlying $N-$particle system.
\begin{thm} \label{thm.main20} Assume \eqref{ass.main}. There exists a constant $\gamma\in (0,1)$ depending only on the dimension $d$ such that, for every $(t_0,m_0)\in \mathcal O$ with $M_{d+5}(m_0)<+\infty$, there is $C=C((t_0,m_0))>0$
such that, if ${\bf Z}=(Z^k)_{k=1, \dots, N}$ is a sequence of independent random variables with law $m_0$, ${\bf B_.}=(B^k_.)_{k=1, \dots, N}$ is a sequence of independent Brownian motions independent of ${\bf Z}$,
and ${\bf Y_.}^N=(Y^{N,k})_{k=1, \dots, N}$ is the optimal trajectory for $\mathcal V^N(t_0, (Z^k)_{k=1, \dots, N})$, that is, for each $k=1,\ldots,N$ and $t\in [t_0,T]$,
\begin{equation}\label{takis51}
Y^{N,k}_t= Z^k-\int_{t_0}^tH_p(Y^k_s, D\mathcal V^N(s, {\bf Y}^N_s))ds +\sqrt{2}(B^k_t-B^k_{t_0}) ,
\end{equation}
then
$$
\mathbb E\left[ \sup_{t\in [t_0,T]}{\bf d_1}(m^N_{{\bf Y}^N_t}, m(t))\right] \leq CN^{-\gamma}.
$$
\end{thm}
\subsection*{Some preliminary facts}\label{subsec.preliminaries} We recall here some well known facts about MFC that we use in the paper.
\vskip.075in
We begin with some regularity properties of the underlying backward-forward MFG system.
Fix $(t_0, m_0)\in [0,T]\times \mathcal P_2(\R^d)$. We recall (see, for example, \cite{LL06cr2, LLJapan} for the original statement or \cite{daudin21}, in a framework closer to our setting) that there exists at least one minimizer for $\mathcal U(t_0,m_0)$ and that, if $(m, \alpha)\in d M(t_0,m_0)$ is a minimizer, then there exists a multiplier $u:[t_0,T]\times \mathbb R^d\to \mathbb R$ such that $\alpha= -H_p(x,Du)$ and the pair $(u,m)$ solves the MFG-system
\begin{equation}\label{MFG}
\left\{\begin{array}{l}
\displaystyle -\partial_t u -\Delta u +H(x,Du)= F(x,m(t)) \ \ {\rm in} \ \ (t_0,T)\times \mathbb R^d,\\[1.2mm]
\displaystyle \partial_t m -\Delta m -{\rm div}(H_p(x,Du)m)=0 \ \ {\rm in} \ \ (t_0,T)\times \mathbb R^d, \\[1.2mm]
\displaystyle m(t_0)=m_0, \; u(T, x)=G(x,m(T)) \ \ {\rm in} \ \ \mathbb R^d,
\end{array}\right.
\end{equation}
where
$$
F(x,m)=\frac{\delta \mathcal F}{\delta m}(m,x) \ \ \text{and} \ \ G(x,m)=\frac{\delta \mathcal G}{\delta m}(m,x).
$$
Note that, in view of the assumed strict convexity of $H$, given $(m,\alpha)$, the relation $\alpha= -H_p(x,Du)$ defines uniquely $Du$.
\begin{lem} \label{lem.reguum} Assume \eqref{ass.main} and let $(u,m)$ be a the solution of \eqref{MFG}. Then, for any $\delta\in (0,1)$, there exists $C>0$, which is independent of $(t_0,m_0)$, such that
\begin{equation}\label{reguum}
\|u\|_{C^{2+\delta/2,4+\delta}} +\sup_{t\neq t'} \frac{{\bf d_2}(m(t), m(t'))}{|t-t'|^{1/2}} \leq C\ \ \ {\rm and }\ \ \ \sup_{t\in [t_0,T]} \int_{\mathbb R^d}|x|^2 m(t,dx) \leq C
\int_{\mathbb R^d}|x|^2 m_0(dx).
\end{equation}
\end{lem}
\begin{rmk} Note that, under our standing assumptions, we do not have, in general, uniqueness of the solution to \eqref{MFG}. Indeed, the problem defining $\mathcal U(t_0,m_0)$ may have several minimizers and/or solutions of \eqref{MFG} may not necessarily be associated with a minimizer of $\mathcal U(t_0,m_0)$.
\end{rmk}
\begin{proof} The estimates of $m$ and the local regularity of $u$ are standard. Indeed, the uniform bound on $Du$ follows as in the proof of Lemma 3.3 in \cite{CDJS}, and then the estimate on $m$ are immediate. Moreover, the local regularity of $u$ is a consequence of the classical parabolic regularity theory.
\vskip.075in
The only point is to explain why this regularity holds globally in space. For this, we first note that the assumptions on $\mathcal F$ and $\mathcal G$ yield a $C>0$ such that, for all $x,y\in \mathbb R^d$,
$$
\sup_{t\in [t_0,T]} \left| F(x,m(t))-F(y,m(t))\right| + \left| G(x,m(t))-G(y,m(t)\right|\leq C|x-y|.
$$
It then follows from the maximum principle that $u$ is, uniformly in $(t_0,m_0)$ and in $t$, Lipschitz continuous in the space variable. The same argument applied to the equation satisfied by $u_{x_i}$ for each $i=1,\ldots,d$, implies that
$Du$ is also uniformly Lipschitz continuous in the space variable.
\vskip.075in
The general conclusion can be established similarly, using the maximum principle for the global estimates and the parabolic regularity for the local one.
\end{proof}
In view of the uniform estimates in \eqref{reguum}, we have the following stability for minimizers in \eqref{takis19} when they are unique.
\begin{lem}\label{lem.stabilo} Assume \eqref{ass.main}, fix $(t_0,m_0)\in [0,T)\times \mathcal P_2(\R^d)$, suppose that $\mathcal U(t_0,m_0)$ has a unique minimizer $(m, \alpha)$, and let $u$ be the associated multiplier. If $(t_0^n, m_0^n)$ converges to $(t_0,m_0)$ and if $(m^n,\alpha^n)$ is a minimizer for $\mathcal U(t_0^n,m_0^n)$ with associated multiplier $u^n$, then $u^n$, $Du^n$ and $D^2u^n$ converge respectively to $u$, $Du$ and $D^2u$ in $C^{\delta/2,\delta}$. In addition, if $t_0^n=t_0$ for all $n$, the convergence of $(u^n)$ holds in $C^{(2+\delta)/2,2+\delta}$.
\end{lem}
\begin{proof} It easily follows from the regularity of $u$ (see \eqref{reguum}) that, without loss of generality, we can assume that $t^n_0=t_0$. Moreover, again in view of \eqref{reguum} and the continuity of $\mathcal U$ (see Lemma~\ref{lem.estiVN} below),
the minimizer $(m^n,\alpha^n)$ converge along subsequences in $C^0([t_0,T]\times \mathcal P_1(\R^d))\times C^{1,2}_{loc}([t_0,T]\times \mathbb R^d)$ to minimizers for $\mathcal U(t_0,m_0)$.
Since the latter is assumed to have a unique minimizer $(m,\alpha)$, the convergence holds along the whole sequence. Arguing as in Lemma \ref{lem.reguum}, we can also check that the convergence of the $u^n$'s holds in $C^{(2+\delta)/2,2+\delta}$, because $z^n=u^n-u$ solves a linearized equation of the form
$$
-\partial_t z^n-\Delta z^n+V^n\cdot Dz^n= F(x,m^n(t))-F(x,m(t)) \ \ \ z^n(T,x)= G(x,m^n(T))-G(x,m(T)),
$$
with $V^n(t,x)= \int_0^1 H_p((1-s)Du^n+sDu,x)ds$, and where, in view of the regularity of $\mathcal F$ and $\mathcal G$ and the convergence of $m^n$ to $m$, $F(\cdot,m^n(\cdot))-F(\cdot,m(\cdot))$ and $G(\cdot,m^n(T))-G(\cdot,m(T))$ converge as $n\to \infty$ to $0$.
\end{proof}
The following second-order optimality condition is used several times in the proofs of the main results.
\begin{lem}\label{lem.secondordercond} Assume \eqref{ass.main}, fix $(t_0,m_0)\in [0,T)\times \mathcal P_2(\R^d)$ and let $(m, \alpha)$ be a minimizer for $\mathcal U(t_0,m_0)$. Fix $\beta\in C^0([t_0,T]\times \mathbb R^d;\mathbb R^d)$ or $\beta \in L^\infty([t_0,T]\times \mathbb R^d; \mathbb R^d)$ with $\beta=0$ in a neighborhood of $t_0$ and let $\rho\in C^0([t_0,T], (C^{2+\delta}(\mathbb R^d))')$ be the solution in the sense of distributions to
\begin{equation}\label{eq.forsecondorder}
\left\{\begin{array}{l}
\partial_t \rho-\Delta \rho +{\rm div} (\rho \alpha)+ {\rm div}(m\beta)= 0 \ \ {\rm in} \ \ (t_0,T)\times \mathbb R^d,\\[1.2mm]
\rho(t_0)= 0 \ \ {\rm in} \ \ \mathbb R^d.
\end{array}\right.
\end{equation}
Then
\begin{equation}\label{ineq.SOC}
\begin{split}
& \int_{t_0}^T \Bigl( \int_{\mathbb R^d} L_{\alpha,\alpha}(x,\alpha(t,x)) \beta(t,x)\cdot \beta(t,x)m(t,dx) +\langle \frac{\delta F}{\delta m}(\cdot,m(t),\cdot), \rho(t)\otimes \rho(t)\rangle\Bigr)dt\\
& \qquad + \langle \frac{\delta G}{\delta m}(\cdot,m(T),\cdot), \rho(T,\cdot)\otimes \rho(T,\cdot)\rangle\geq 0.
\end{split}
\end{equation}
\end{lem}
This statement is an adaptation of an analogous result in \cite{BrCa}. The existence of the solution to \eqref{eq.forsecondorder} and the proof of \eqref{ineq.SOC} are given in the Appendix. \\
It is well-known that the map $\mathcal V^N$ defined in \eqref{def.VN} solves the uniformly parabolic Hamilton-Jacobi-Bellman (HJB for short) equation
$$
\left\{\begin{array}{l}
\displaystyle -\partial_t \mathcal V^N(t,{\bf x}) -\sum_{j=1}^N \Delta_{x^j}\mathcal V^N(t,{\bf x}) +\frac1N\sum_{j=1}^N H(x^j, N D_{x^j}\mathcal V^N(t, {\bf x}))=\mathcal F(m^N_{{\bf x}})\ \ {\rm in}\ \ (0,T)\times (\mathbb R^d)^N \\[1.5mm]
\displaystyle \mathcal V^N(T,{\bf x})= \mathcal G(m^N_{{\bf x}})\ \ {\rm in}\ \ (\mathbb R^d)^N,
\end{array}\right.
$$
and, therefore $\mathcal V^N$ is smooth for any $N$. This is in contrast with the limit $\mathcal U$, which might not be $C^1$. The following result, proved in \cite{CDJS}, states however that both maps are uniformly Lipschitz continuous.
\begin{lem}[Regularity of $\mathcal V^N$ and of $\mathcal U$] \label{lem.estiVN} Assume \eqref{ass.main}. There exists constant $C$ depending on the data such that
$$
\|\mathcal V^N\|_\infty+N \sup_{j=1,\ldots,N} \|D_{x^j}\mathcal V^N\|_\infty \leq C,
$$
and, for all $(t,m), (t',m')\in [0,T]\times \mathcal P_1(\R^d),$
$$
\left|\mathcal U(t,m)-\mathcal U(t',m')\right| \leq C(|t-t'|+{\bf d}_1(m,m')).
$$
\end{lem}
\vskip.075in
The following convergence rate is the main result of \cite{CDJS}.
\begin{prop}[Quantified convergence of $\mathcal V^N$ to $\mathcal U$] \label{prop.quantifCV} Assume \eqref{ass.main}. There exists $\gamma\in (0,1)$ depending on dimension only and $C>0$ depending on the smoothness of the data such that, for any $(t,{\bf x})\in [0,T]\times (\mathbb R^d)^N$,
$$
\left| \mathcal V^N(t,{\bf x})-\mathcal U(t,m^N_{{\bf x}})\right| \leq C \dfrac{1}{N^{\gamma}} \left( 1+ M_2(m^N_{{\bf x}})\right).
$$
\end{prop}
\section{The regularity of $\mathcal U$}\label{sec.proofThm1}
We prove here Theorem \ref{thm.mainreguU0}. A crucial step is the analysis of a linearized system, which is reminiscent of a linearized system studied in \cite{CDLL} and \cite{BrCa} for MFG problems. The main and important difference from \cite{CDLL} is that here we do not assume that $d F$ and $d G$ are convex, while
\cite{BrCa} deals with problems on the torus. We go around the lack of monotonicity by using the notions of
stability and strong stability of a solution, which are introduced next using the linearized system. Finally, stability is also used to define and analyze the open, dense set $\mathcal O$ on which the map $\mathcal U$ will eventually be smooth.
\subsection*{The linearized system} We fix $t_0\in [0,T)$, a constant $C_0$, and, for $m_0\in \mathcal P_2(\R^d)$ and $V:[t_0,T]\times \mathbb R^d\to \mathbb R^d \ \ \text{with} \ \ \|V\|_{C^{1,3}}\leq C_0, $ let $m$ be the solution to
\begin{equation}\label{eq.m-Appen}
\left\{\begin{array}{l}
\partial_t m -\Delta m -{\rm div}(Vm)=0\ \ {\rm in}\ \ (t_0,T)\times \mathbb R^d \\[1.2mm]
m(t_0)=m_0\ \ {\rm in}\ \ \mathbb R^d.
\end{array}\right.
\end{equation}
We analyze the inhomogeneous linearized system
\begin{equation}\label{eq.LS}
\left\{ \begin{array}{l}
\displaystyle -\partial_t z -\Delta z +V(t,x)\cdot Dz = \frac{\delta F}{\delta m}(x,m(t))(\rho(t))+R^1(t,x)\ \ {\rm in}\ \ (t_0,T)\times \mathbb R^d, \\[3mm]
\displaystyle \partial_t\rho-\Delta \rho -{\rm div}(V\rho) -\sigma {\rm div}(m\Gamma Dz) = \sigma {\rm div}(R^2)\ \ {\rm in}\ \ (t_0,T)\times \mathbb R^d,\\[1.2mm]
\displaystyle \rho(t_0)=\sigma \xi \ \ \text{and} \ \ z(T,x)= \frac{\delta G}{\delta m}(x,m(T))(\rho(T))+R^3\ \ {\rm in}\ \ \mathbb R^d,
\end{array}\right.
\end{equation}
where
\begin{equation}\label{takis30}
\begin{split}
& \sigma\in [0,1] \ \ \text{and} \ \ \delta\in (0,1),\\[1.2mm]
& \Gamma\in C^0([0,T]\times \mathbb R^d; \mathbb R^{d\times d}) \ \ \text{ with} \ \ \|\Gamma\|_\infty\leq C_0, \\[1.2mm]
& R^1\in C^{\delta/2,\delta}, \ \ R^2\in L^\infty([t_0,T], (W^{1,\infty})'(\mathbb R^d,\mathbb R^d)), \ \ R^3\in C^{2+\delta} \ \ \text{ and} \ \ \xi\in (W^{1, \infty})'.
\end{split}
\end{equation}
The pair $(z,\rho) \in C^{0,1}([t_0,T]\times\mathbb R^d) \times C^0([0,T], (C^{2+\delta})')$ is a solution to \eqref{eq.LS} if $z$
and $\rho$ satisfy respectively the first and second equation in the sense of distributions.
\vskip.075in
Note that, because of the regularity of $\rho$ and the assumptions on $\mathcal F$ and $\mathcal G$, the maps $(t,x)\to \frac{\delta F}{\delta m}(x,m(t))(\rho(t))$ and $x\to \frac{\delta G}{\delta m}(x,m(T))(\rho(T))$ are continuous and bounded. \
\vskip.075in
We will often use system \eqref{eq.LS} in which $V(t,x)= H_p(x,Du(t,x))$ and $\Gamma(t,x)=H_{pp}(x,Du(t,x))$, where $(u,m)$ is a classical solution to \eqref{MFG}. In this case, $V$, $\Gamma$ and $m$ satisfy the conditions above.
\vskip.075in
Next we introduce the notion of strong stability for the homogeneous version of \eqref{eq.LS}, that is the system
\begin{equation}\label{eq.LSH}
\left\{ \begin{array}{l}
-\partial_t z -\Delta z +V(t,x)\cdot D z = \frac{\delta F}{\delta m}(x,m(t))( \rho(t)) \ \ {\rm in}\ \ (t_0,T)\times \mathbb R^d, \\[1.2mm]
\partial_t \rho-\Delta \rho -{\rm div}(V \rho) -\sigma{\rm div}(m\Gamma D z) =0\ \ {\rm in}\ \ (t_0,T)\times \mathbb R^d, \\[1.2mm]
\rho(t_0)= 0 \ \ \text{and} \ \ z(T,x)= \frac{\delta G}{\delta m}(x,m(T))( \rho(T))\ \ {\rm in}\ \ \mathbb R^d.
\end{array}\right.
\end{equation}
We say that
\begin{equation}\label{takis31}
\text{the system \eqref{eq.LSH} is strongly stable if, for any $\sigma\in [0,1]$, its unique solution is $(z,\rho)=(0,0)$.}
\end{equation}
The main result of the subsection is a uniqueness and regularity result for the solution to \eqref{eq.LS}.
\begin{lem} \label{lem.estLS-Appen} Assume \eqref{ass.main} and \eqref{takis31}.
There exist a neighborhood $\mathcal V$ of $(V,\Gamma)$ in the topology of locally uniform convergence, and $\eta, C>0$ such that, for any $(V', t_0', m_0', \Gamma', R^{1,'}, R^{2,'}, R^{3,'}, \xi',\sigma')$ with
\begin{equation}\label{cond.data-Appen}
\begin{split}
(V',\Gamma')\in \mathcal V,\quad |t_0'-t_0|+{\bf d}_2(m_0',m_0) \leq \eta, \ \ \|V'\|_{C^{1, 3}}+\|\Gamma'\|_\infty \leq 2 C_0, \ \ \sigma'\in [0,1],\\[1.2mm]
R^{1,'}\in C^{\delta/2,\delta}, \; R^{2,'}\in L^\infty([t_0,T], (W^{1,\infty})'(\mathbb R^d,\mathbb R^d)), \; R^{3,'}\in C^{2+\delta},
\; \xi'\in (W^{1, \infty})',
\end{split}
\end{equation}
any solution $(z',\rho')$ to \eqref{eq.LS} associated with these data on $[t_0',T]$ and $m'$ the solution to \eqref{eq.m-Appen} with drift $V'$ and initial condition $m_0'$ at time $t_0'$ satisfies
\begin{equation}\label{takis32}
\|z'\|_{C^{(2+\delta)/2,2+\delta}}+ \sup_{t\in [t_0',T]} \|\rho'(t,\cdot)\|_{(C^{2+\delta})'} + \sup_{t'\neq t} \frac{\|\rho'(t',\cdot)-\rho'(t,\cdot)\|_{(C^{2+\delta})'}}{|t'-t|^{1/2}} \leq CM' ,
\end{equation}
where
\begin{equation}\label{def.Mprime-Appen}
M':= \|\xi'\|_{(W^{1,\infty})'}+\|R^{1,'}\|_{C^{\delta/2, \delta}}+\|R^{3,'}\|_{C^{2+\delta}}+ \sup_{t\in [t_0',T]} \|R^{2,'}(t)\|_{(W^{1,\infty})'}.
\end{equation}
\end{lem}
An immediate consequence is the following corollary.
\begin{cor}
Assume \eqref{ass.main} and \eqref{takis31}. Then, for any $(V', m_0', \Gamma')$ satisfying \eqref{cond.data-Appen},
the corresponding homogeneous linearized system is strongly stable.
\end{cor}
The proof of Lemma~\ref{lem.estLS-Appen} follows some of the ideas of \cite{CDLL}, where a similar system is studied. The main differences are that, here, we use the stability condition instead of the monotonicity assumption of \cite{CDLL} and work in an unbounded space.
\vskip.075in
In what follows, we need a preliminary result which we state and prove next. The difference between the estimate below and the one of Lemma~\ref{lem.estLS-Appen} is the right hand side of the former which depends on the solution itself.
\begin{lem}\label{lem.apriori-Appen} Assume \eqref{ass.main} and let $(z,\rho)$ be a solution to \eqref{eq.LS}. There is a constant $C$, depending only on the regularity of $\mathcal F$, $\mathcal G$ and on $\|V\|_{C^{1,3}}+ \|\Gamma\|_\infty$, such that
$$
\|z\|_{C^{(2+\delta)/2,2+\delta}}+ \sup_{t\in [t_0,T]} \sup_{t'\neq t} \frac{\|\rho(t',\cdot)-\rho(t,\cdot)\|_{(C^{2+\delta})'}}{|t'-t|^{\delta/2}} \leq C\left(M+\sup_{t\in [t_0,T]} \|\rho(t)\|_{(C^{2+\delta})'}\right)
$$
where
$$
M:=\|\xi\|_{(W^{1,\infty})'}+ \|R^{1}\|_{C^{\delta/2, \delta}} + \sup_{t\in [t_0,T]}\|R^{2}(t)\|_{(W^{1,\infty})'}+ \|R^{3}\|_{C^{2+\delta}}.
$$
\end{lem}
\begin{proof} Throughout the proof, $C$ denotes a constant that depends only on the data and may change from line to line.
\vskip.075in
Set $R= \sup_{t\in [t_0,T]} \|\rho(t)\|_{(C^{2+\delta})'}<\infty$.
It follows that the maps $(t,x)\to \frac{\delta F}{\delta m}(x,m(t))(\rho(t))$ and $x\to \frac{\delta G}{\delta m}(x,m(T))(\rho(T))$ are bounded by $CR$, the latter in
$C^{2+\delta}$.
\vskip.075in
Then, standard parabolic regularity gives that $z$ is bounded in $C^{(1+\delta)/2, 1+\delta}$ by $CR$.
\vskip.075in
The main step of the proof is to show that
$$
\sup_{t\in [t_0,T]} \sup_{t'\neq t} \frac{\|\rho(t')-\rho(t)\|_{(C^{2+\delta})'}}{|t'-t|^{\delta/2}} \leq C(M+R).
$$
Arguing by duality, we fix $t_0\leq t_1<t_2\leq T$ and, with $\bar \psi\in C^{2+\delta}$, we consider the solution $\psi^i$ for $i=1,2$ to
\begin{equation}\label{eq.dualeqTER}
\left\{\begin{array}{l}
-\partial_t \psi^i -\Delta \psi^i + V(t,x) \cdot D\psi^i = 0\ \ {\rm on}\ \ (t_0,t_i)\times \mathbb R^d,\\[1.5mm]
\psi(t_i)= \bar \psi \ \ {\rm in}\ \ \mathbb R^d,
\end{array}\right.
\end{equation}
which, in view of the assumption on $V$ and parabolic regularity, satisfies, for $i=1,2$, the bound
\[\|\psi^i\|_{C^{(2+\delta)/2,2+\delta}}\leq C\|\bar \psi\|_{C^{2+\delta}}.\]
It is immediate that, for $i=1, 2$,
$$
\langle \rho(t_i), \bar \psi\rangle = \langle \xi, \psi^i(t_0)\rangle-\sigma \int_{t_0}^{t_i} \int_{\mathbb R^d} \Gamma Dz\cdot D\psi^i m dx dt-\sigma \int_{t_0}^{t_i}\langle R^{2}(t), D\psi^i(t)\rangle dt.
$$
Thus
\begin{equation}\label{takis35}
\begin{split}
& \langle \rho(t_2)-\rho(t_1), \bar \psi\rangle = \langle \xi, \psi^2(t_0)-\psi^1(t_0)\rangle-\sigma \int_{t_0}^{t_1} \int_{\mathbb R^d} \Gamma Dz\cdot (D\psi^2-D\psi^1) m \\[3mm]
& - \sigma \int_{t_1}^{t_2} \int_{\mathbb R^d} \Gamma Dz \cdot D\psi^2 m -\sigma \int_{t_0}^{t_1} \langle R^{2}(t), (D\psi^2-D\psi^1)(t) \rangle dt - \sigma\int_{t_1}^{t_2} \langle R^{2}(t), D\psi^2(t)\rangle dt.
\end{split}
\end{equation}
Note that $\psi^2-\psi^1$ solves \eqref{eq.dualeqTER} on $(t_0,t_1)$ with a terminal condition at $t_1$ given by $(\psi^2-\psi^2)( t_1,\cdot)=\psi^2(t_1,\cdot)-\psi^2 (t_2,\cdot)$, which, in view of the regularity of $\psi^2$, is bounded in $W^{2,\infty}$ by $C(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}$.
\vskip.075in
It then follows follows from the maximum principle that $(\psi^2-\psi^1)( t,\cdot)$ is bounded in $W^{2,\infty}$ by $C(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}$ for any $t$, and, hence,
\begin{align*}
&\left|\langle \xi, \psi^2( t_0,\cdot)-\psi^1( t_0,\cdot)\rangle\right| \leq \|\xi\|_{(W^{1, \infty})'}\|\psi^2( t_0,\cdot)-\psi^1( t_0,\cdot)\|_{W^{1,\infty}} \\[2mm]
& \leq C(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}\|\xi\|_{(W^{1,\infty})'} \leq C(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}M,
\end{align*}
$$
\left| \int_{t_0}^{t_1} \int_{\mathbb R^d} \Gamma Dz \cdot (D\psi^2-D\psi^1) m dx dt\right| \leq C\|Dz\|_\infty(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}\leq C(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}R
$$
and
\begin{align*}
\left|\int_{t_0}^{t_1} \langle R^{2}(t), (D\psi^2-D\psi^1)(t) \rangle dt\right|& \leq \sup_t \|R^2(t)\|_{(W^{1,\infty})'}\sup_t \|D\psi^2(t)-D\psi^1(t)\|_{W^{1,\infty}}\\
& \leq C(t_2-t_1)^{\delta/2}\|\bar \psi\|_{C^{2+\delta}}M.
\end{align*}
Note that the regularity of $\psi^2$ yields
$$
\left| \sigma \int_{t_1}^{t_2} \int_{\mathbb R^d} \Gamma Dz \cdot D\psi^2 m dx dt \right| \leq C(t_2-t_1) \|Dz\|_\infty\|D\psi^2\|_\infty \leq C(t_2-t_1)^{1/2} \|\bar\psi\|_{C^{2+\delta}}R
$$
and
$$
\left|\int_{t_1}^{t_2} \langle R^{2}(t), D\psi^2(t) \rangle dt\right|\leq (t_2-t_1)\sup_{t} \|R^{2}(t)\|_{(W^{1,\infty})'} \sup_{t}\|D\psi^2(t)\|_{W^{1,\infty}} \leq C(t_2-t_1) \|\bar \psi\|_{C^{2+\delta}}M.
$$
\vskip.075in
Since $\bar \psi$ is arbitrary, it follows from \eqref{takis35} that
$$
\|\rho(t_2)-\rho(t_1)\|_{(C^{2+\delta})'} \leq C(t_2-t_1)^{\delta/2}(M+R).
$$
This regularity of $\rho$ implies that the maps $(t,x)\to \frac{\delta F}{\delta m}(x,m(t))(\rho(t))$ and $x\to \frac{\delta G}{\delta m}(x,m(T))(\rho(T))$ are bounded in $C^{\delta/2,\delta}$ and $C^{2+\delta}$ respectively by $C(M+R)$. Thus $z$ is also bounded in $C^{(2+\delta)/2,2+\delta}$ by $C(M+R)$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem.estLS-Appen}] The first (and main) part of the proof consists in showing the existence of $\mathcal V$, $\eta>0$ and $C>0$ such that, for any $(V', t_0', m_0', \Gamma', R^{1,'}, R^{2,'}, R^{3,'}, \xi',\sigma')$ as in \eqref{cond.data-Appen}, any solution $(z',\rho')$ to \eqref{eq.LS} associated with these data on $[t_0',T]$ satisfies, with $M'$ is defined by \eqref{def.Mprime-Appen},
$$
\sup_{t\in [t_0',T]} \|\rho'(t)\|_{(C^{2+\delta})'} \leq CM'.
$$
\vskip.075in
We prove this claim by contradiction, assuming the existence of sequences $(V^n)_{n\in \mathbb N}$, $(t_0^n, m_0^n)_{n\in \mathbb N}$, $(m^n)_{n\in \mathbb N}$, $(\Gamma^n)_{n\in \mathbb N}$, $(R^{i,n})_{n\in \mathbb N}$, $(\xi^n)_{n\in \mathbb N}$, and $(\sigma^n)_{n\in \mathbb N}$ such that $(V^n,\Gamma^n)$ converges locally uniformly to $(V, \Gamma)$,
$$
|t_0^n-t_0|+{\bf d}_2(m_0^n,m_0)\leq 1/n, \ \ \|V^n\|_{C^{1, 3}}+\|\Gamma^n\|_\infty \leq 2 C_0, \ \ \sigma^n\in [0,1],
$$
and
$$
\|\xi^n\|_{(W^{1,\infty})'}+\|R^{3,n}\|_{C^{2+\delta}}+\|R^{1,n}\|_{C^{\delta/2,\delta}} + \sup_{t\in [t_0,T]} \|R^{2,n}(t)\|_{(W^{1,\infty})'} \leq 1,
$$
and, for each $n$, a solution $(z^n,\rho^n)$ to \eqref{eq.LS} associated with the data above, such that
$$
\lambda^n=\sup_{t\in [t_0^n,T]} \|\rho^n(t)\|_{(C^{2+\delta})'}\geq n.
$$
It follows that $(\tilde z^n, \tilde \rho^n)= \frac{1}{\lambda^n}(z^n, \rho^n)$ solves the system
\begin{equation}\label{pierre}
\left\{ \begin{array}{l}
-\partial_t \tilde z^n -\Delta \tilde z^n +V^n(t,x)\cdot D\tilde z^n = \dfrac{\delta F}{\delta m}(x,m^n(t))(\tilde \rho^n(t))+\dfrac{1}{\lambda^n}R_1^n(t,x), \\[2.5mm]
\partial_t\tilde \rho^n-\Delta \tilde \rho^n -{\rm div}(V^n\tilde \rho^n) -\sigma^n{\rm div}(m\Gamma^n D\tilde z^n) = \dfrac{1}{\lambda^n}\sigma^n {\rm div}(R_2^n),\\[2.5mm]
\tilde \rho^n(t_0^n)= \dfrac{1}{\lambda^n}\sigma^n\xi^n \ \ \text{and} \ \ \tilde z^n(T,x)= \dfrac{\delta G}{\delta m}(x,m^n(T))(\tilde \rho^n(T))+\dfrac{1}{\lambda^n}R_3^n.
\end{array}\right.
\end{equation}
Since, by definition, $\sup_t \|\tilde \rho^n(t)\|_{(C^{2+\delta})'}= 1$, Lemma \ref{lem.apriori-Appen} implies that the $(\tilde z^n)$'s and $\tilde \rho^n(t,\cdot)$'s are bounded in $C^{(2+\delta)/2,2+\delta}$ and $C^{\delta/2}([0,T], (C^{2+\delta})')$ respectively.
\vskip.075in
Hence, we may assume that, up to a subsequence, the sequences $(\sigma^n)_{n\in \mathbb N}$, $(\tilde z^n)_{n\in \mathbb N}$, $(\tilde \rho^n)_{n\in \mathbb N}$ and $(V^n)_{n\in \mathbb N}$ converge respectively to some $\sigma\in [0,1]$, $\tilde z \in C^{1,2}_{loc}$, $\tilde \rho \in C^0([t_0,T], (C^{2+\delta}_c)')$, where $(C^{2+\delta}_c)'$ is endowed with the weak-$\ast$ topology, and $V \in C^0([t_0,T], C^{2+\delta}_{loc})$.
\vskip.075in
The goal is to show that $(\tilde z,\tilde \rho)$ is a nonzero solution to the homogenous equation \eqref{eq.LSH}, which will contradict the strong stability of the system.
\vskip.075in
There are two difficulties that need to be addressed both caused by the above claimed weak convergence of the $\tilde \rho^n$'s to $\tilde \rho$.
\vskip.075in
The first is to prove that, as $n\to \infty$,
$$\dfrac{\delta F}{\delta m}(x,m^n(t))(\tilde \rho^n(t)) \to \dfrac{\delta F}{\delta m}(x,m(t))(\tilde \rho(t)) \;\text{and} \; \dfrac{\delta G}{\delta m}(x,m^n(T))(\tilde \rho^n(T)) \to \dfrac{\delta G}{\delta m}(x,m(T))(\tilde \rho(T),$$
\vskip.075in
and the second is to show that, since $\sup_t \|\tilde \rho^n(t)\|_{(C^{2+\delta})'}=1$, we must have $(\tilde z,\tilde \rho)$ is nonzero.
\vskip.075in
To overcome these two issues it is necessary to upgrade the convergence of the $\tilde \rho^n$'s to $\tilde \rho$ in \\$C^0([t_0,T], (C^{2+\delta})')$ from weak to strong.
\vskip.075in
We first note that $\tilde \rho^n=\tilde \rho^{n,1}+\tilde \rho^{n,2}$ with $\tilde \rho^{n,1}$ and $\tilde \rho^{n,2}$ solving respectively
$$
\partial_t\tilde \rho^{n,1}-\Delta \tilde \rho^{n,1} -{\rm div}(V^n\tilde \rho^{n,1}) = \frac{1}{\lambda^n}\sigma^n {\rm div}(R_2^n) \ \ \text{and} \ \
\rho^{n,1}(t_0)= \frac{1}{\lambda^n}\sigma^n\xi^n,
$$
and
\begin{equation}\label{takis38}
\partial_t\tilde \rho^{n,2}-\Delta \tilde \rho^{n,2} -{\rm div}(V^n\tilde \rho^{n,2}) -\sigma^n{\rm div}(\Gamma^n D\tilde z^nm^n) = 0 \ \ \text{and} \ \
\tilde \rho^{n,2}(t_0)= 0.
\end{equation}
\vskip.075in
We show next that $\sup_t \|\tilde \rho^{n,1}(t)\|_{(C^{2+\delta})'} \to 0$. Indeed, for fixed $\bar t\in (t_0^n,T]$ and $\bar \psi\in C^{2+\delta}$, let $\psi^n$ be the solution to the dual problem
\begin{equation}\label{eq.dualeqBIS}
\left\{\begin{array}{l}
-\partial_t \psi^n -\Delta \psi^n + V^n(t,x) \cdot D\psi^n = 0\ \ {\rm in}\ \ (t_0^n,\bar t)\times \mathbb R^d,\\
\psi^n(\bar t)= \bar \psi\ \ {\rm in}\ \ \mathbb R^d.
\end{array}\right.
\end{equation}
It follows from the standard parabolic regularity that $\psi^n$ is bounded in $C^{(2+\delta)/2,2+\delta}$ by $C\|\bar \psi\|_{C^{2+\delta}}$ with $C$ is independent of $n$ and $\bar t$, and since, in view of the duality, we have
$$
\langle \tilde \rho^{n,1}(\bar t), \bar \psi\rangle = \frac{1}{\lambda^n} \big( \sigma^n \langle \xi^n, \psi^n(t_0^n) \rangle -\sigma^n \int_{t_0^n}^{\bar t}\langle R^{2,n}, D\psi^n\rangle dt \big),
$$
we obtain
$$
\sup_{t\in [t_0^n,T]} \|\tilde \rho^{n,1}(t)\|_{(C^{2+\delta})'} \leq \frac{C}{\lambda^n} \big (\|\xi^n\|_{(W^{1,\infty})'} + \sup_{t\in [t_0^n,T]}\|R^{2,n}(t)\|_{(W^{1,\infty})'}\big ).
$$
Hence, $\sup_{t\in [t_0^n,T]} \|\tilde \rho^{n,1}(t)\|_{(C^{2+\delta})'} \to 0$ as $\dfrac{1}{\lambda^n} (\|\xi^n\|_{(W^{1,\infty})'} + \sup_{t\in [t_0^n,T]} \|R^{2,n}(t)\|_{(W^{1,\infty})'}) \to 0$.
\vskip.075in
It follows from \eqref{takis38} that, for any $\bar t^n\in [t_0^n,T]$ and ${\bar \psi}^n\in C^{2+\delta}$, if $\psi^n$ is the solution to \eqref{eq.dualeqBIS} on $[t_0^n,\bar t^n]$ with terminal condition $\psi^n(\bar t^n)={\bar \psi}^n$, then
\begin{equation}\label{repeho2n}
\langle \tilde \rho^{n,2}(\bar t^n), {\bar \psi}^n\rangle = - \sigma^n \int_{t_0^n}^{\bar t^n}\int_{\mathbb R^d} \Gamma^n D\tilde z^n\cdot D\psi^n m^n.
\end{equation}
In order to prove the uniform convergence of the $ \tilde \rho^{n,2}$'s in $(C^{2+\delta})'$, we assume that the $\|{\bar \psi}^n\|_{C^{2+\delta}}$'s are bounded and, without loss of generality, that the $\bar t^n$'s and ${\bar \psi}^n$'s converge respectively to $\bar t\in [t_0,T]$ and $\bar \psi\in C^{2+\delta}$, the last convergence being in $C^{2+\delta_1}_{loc}$ for any $\delta_1\in (0,\delta)$. We need to prove that the $(\langle \tilde \rho^{n,2}(\bar t^n), {\bar \psi}^n\rangle)$'s converge to $\langle \tilde \rho^{2}(\bar t), {\bar \psi}\rangle$, where $\tilde \rho^2$ is the solution to
$$
\partial_t\tilde \rho^2-\Delta \tilde \rho^2 -{\rm div}(V\tilde \rho^2) -\sigma{\rm div}(\Gamma D\tilde zm) = 0 \ \ \text{and} \ \
\tilde \rho^2(t_0)= 0.
$$
Note that
$$
\langle \tilde \rho^{2}(\bar t), {\bar \psi}\rangle = - \sigma \int_{t_0}^{\bar t}\int_{\mathbb R^d} \Gamma D\tilde z\cdot D\psi m,
$$
where $\psi$ is the solution to
\[
-\partial_t \psi -\Delta \psi + V(t,x) \cdot D\psi = 0\ \ {\rm in}\ \ (t_0,\bar t)\times \mathbb R^d \ \ \text{and} \ \ \psi(\bar t)= \bar \psi\ \ {\rm in}\ \ \mathbb R^d,
\]
and recall that the $\Gamma^n$'s and $D\tilde z^n$'s are bounded and converge locally uniformly to $\Gamma$ and to $D\tilde z$ respectively.
\vskip.075in
Similarly, due to the parabolic regularity, the $\psi^n$'s are bounded in $C^{(2+\delta)/2,2+\delta}$ and the $D\psi^n$'s converge locally uniformly to $D\psi$.
\vskip.075in
Moreover, since $m^n$ is the solution to
$$
\partial_t m^n-\Delta m^n+{\rm div}(V^nm^n)=0 \ \ \text{in} \ \ (t_0^n, T] \ \ \text{and} \ \ m^n( t_0^n,\cdot)=m^n_0 \ \ \text{in} \ \mathbb R^d,
$$
with $V^n$ uniformly bounded, we know that the $m^n$'s converge uniformly to $m$ in $\mathcal P_1(\R^d)$, and we have the second-order moment estimate
$$
\sup_{t,n} \int_{\mathbb R^d}|x|^2 m^n(t,dx)\leq C\sup_{n} \int_{\mathbb R^d}|x|^2 m^n_0(dx) \leq C.
$$
In addition, using that $\Gamma D\tilde z\cdot D\psi$ is globally Lipschitz, for any $R\geq 1$ we find
\begin{align*}
\left| \langle \tilde \rho^{n,2}(\bar t^n), {\bar \psi}^n\rangle- \langle \tilde \rho^{2}(\bar t), {\bar \psi}\rangle\right| & \leq
C(|\sigma^n-\sigma|+|\bar t^n-\bar t|+|t_0^n-t_0|) \\
&\qquad + \sigma \left| \int_{t_0\vee t_0^n}^{\bar t\wedge \bar t^n}\int_{\mathbb R^d} (\Gamma D\tilde z\cdot D\psi m-
\Gamma^n D\tilde z^n\cdot D\psi^n m^n)\right| \\
&\leq C(|\sigma^n-\sigma|+|\bar t^n-\bar t|+|t_0^n-t_0|+\sup_t {\bf d}_1(m^n(t),m(t))+R^{-2}) \\
&\qquad + \sigma \left| \int_{t_0\vee t_0^n}^{\bar t\wedge \bar t^n}\int_{B_R} (\Gamma D\tilde z\cdot D\psi -
\Gamma^n D\tilde z^n\cdot D\psi^n) m^n\right| .
\end{align*}
Letting $n\to \infty$ and then $R\to \infty$ proves the convergence of the $\langle \tilde \rho^{n,2}(\bar t^n), {\bar \psi}^n\rangle $'s to $\langle \tilde \rho^{2}(\bar t), {\bar \psi}\rangle$. It follows that the sequence $(\tilde \rho^n)_{n\in \mathbb N}$
converges to $\tilde \rho=\tilde \rho^2$ strongly in $C^0([t_0,T], (C^{2+\delta})')$.
\vskip.075in
To summarize the above, we know that the sequences $(\sigma^n)_{n\in \mathbb N}$, $(\tilde z^n)_{n\in \mathbb N}$, $(\tilde \rho^n)_{n\in \mathbb N}$ and $(V^n)_{n\in \mathbb N}$ converge respectively to $\sigma\in [0,1]$, $\tilde z$ in $C^{1,2}_{loc}$, $\tilde \rho$ in $C^0([0,T], (C^{2+\delta})')$ and $V$ in $C^0([t_0,T], C^{2+\delta}_{loc})$.
\vskip.075in
Passing to the limit in \eqref{pierre} we infer that $(\tilde z,\tilde \rho)$ is a solution to the homogenous equation \eqref{eq.LSH}.
\vskip.075in
Since $\sup_{t\in [t_0^n,T]} \|\rho^n(t)\|_{(C^{2+\delta})'}=1$ for any $n$, it follows that $\sup_{t\in [t_0,T]} \|\rho(t)\|_{(C^{2+\delta})'}=1$. Thus $(\tilde z,\tilde \rho)$ is a nonzero solution to the homogenous equation \eqref{eq.LSH} which contradicts the strong stability assumption \eqref{takis31}.
\vskip.075in
The second part of the proof consists in upscaling the regularity obtained in the first part. For this, we let
$(V', t_0',m_0', \Gamma', R^{1,'}, R^{2,'}, R^{3,'}, \xi',\sigma')$ be such that \eqref{cond.data-Appen} holds and $(z',\rho')$ be a solution to \eqref{eq.LS} associated with these data, where $m'$ is the solution to \eqref{eq.m-Appen} with drift $V'$ and initial condition $m_0'$ at time $t_0'$.
\vskip.075in
We have already established that, for the $M'$ in \eqref{def.Mprime-Appen},
$$
\sup_{t\in [t_0',T]} \|\rho'(t)\|_{(C^{2+\delta})'}\leq CM'.
$$
It then follows from Lemma \ref{lem.apriori-Appen} that
$$
\|z'\|_{C^{(2+\delta)/2,2+\delta}}+ \sup_{t'\neq t} \frac{\|\rho'(t',\cdot)-\rho'(t,\cdot)\|_{C^{2+\delta}}}{|t'-t|^{\delta/2}} \leq C(M'+ \sup_{t\in [t_0',T]} \|\rho'(t)\|_{(C^{2+\delta})'})\leq CM'.
$$
\end{proof}
\subsection*{The stability property} We discuss here the notion of stabllity of a solution $(u,m)$ of the MFG-system arising in MFC.
\vskip.075in
Let $(t_0, m_0)\in [0,T]\times \mathcal P_2(\R^d)$ and $(m, \alpha)$ be a minimizer for $\mathcal U(t_0,m_0)$ with associated multiplier $u$, that is, the pair $(u,m)$ solves \eqref{MFG} and $\alpha(t,x)=-H_p(x,Du(t,x))$.
\begin{defn} The solution $(u,m)$ is strongly stable (resp. stable), if for all $\sigma\in [0,1]$ (resp. $\sigma=1$) the only solution $(z,\mu)\in C^{(1+\delta)/2, 1+\delta} \times C^0([t_0,T]; (C^{2+\delta}(\mathbb R^d))')$ to
the linearized system
\begin{equation}\label{MFGL}
\left\{\begin{array}{l}
\displaystyle -\partial_t z -\Delta z +H_p(x,Du)\cdot Dz= \frac{\delta F}{\delta m}(x,m(t))(\mu(t)) \ \ {\rm in} \ \ (t_0,T)\times \mathbb R^d,\\[2mm]
\displaystyle \partial_t \mu -\Delta \mu - {\rm div}(H_p(x,Du)\mu)-\sigma{\rm div}(H_{pp}(x,Du)Dz m)=0 \ \ {\rm in} \ \ (t_0,T)\times \mathbb R^d,\\[2mm]
\displaystyle \mu(t_0)=0 \ \ \text{and} \ \ z(T, x)=\frac{\delta G}{\delta m}(x,m(T))(\mu(T)) \ \ {\rm in} \ \ \mathbb R^d,
\end{array}\right.
\end{equation}
is the pair $(z,\mu)=(0,0)$.
\end{defn}
Since, given a minimizer $(m,\alpha)$, the relation $\alpha= -H_p(x,Du)$ defines $Du$ uniquely, the stability condition depends on $(m,\alpha)$ only. We say that the minimizer $(m,\alpha)$ is strongly stable (resp. stable) if $(u,m)$ is strongly stable (resp. stable).
\vskip.075in
The above makes also clear the definition of the regularity set $d O$ in \eqref{takis20}. We remark that at this point we do not know whether $d O$ is a nonempty set. This will follow from Lemma~\ref{thm.Oopendense} below.
\vskip.075in
We also note that \eqref{MFGL} is the linearized system studied in the previous subsection for the particular choice of vector field $V(t,x)= H_p(x,Du(t,x))$ and matrix $\Gamma(t,x)=H_{pp}(x,Du(t,x))$. To emphasize that we are working with this particular system and also be consistent with other references, heretofore we use the notation $(z,\mu)$ instead of $(z,\rho).$
\vskip.075in
The following lemma asserts that the minimizers starting from an initial condition in $\mathcal O$ are actually strongly stable.
\begin{lem}\label{lem.StrongStab} Assume \eqref{ass.main}, fix $(t_0,m_0)\in \mathcal O$ and let $(m,\alpha)$ be the unique stable minimizer associated to $\mathcal U(t_0,m_0)$. Then $(m,\alpha)$ is strongly stable.
\end{lem}
\begin{proof} Since, if $\sigma=1$, the claim is just the assumed stability of $(m,\alpha)$, in what follows we assume that $$\sigma\in [0,1).$$
It follows from Lemma \ref{lem.estLS-Appen} that $z \in C^{(2+\delta)/2,2+\delta}$, while using the duality, we have, for any $t\in [t_0,T]$,
$$
\langle z(t,\cdot),\mu(t)\rangle= -\int_{t_0}^t (\int_{\mathbb R^d} \Bigl(\sigma H_{pp}(x,Du) Dz\cdot dz m dx) + \langle \frac{\delta F}{\delta m}(\cdot,m(t))(\mu(t)),\mu(t)\rangle\Bigr)dt,
$$
and, in particular, for $t=T$, we get
\begin{equation}\label{ukzajenrsd}
\begin{split}
&\int_{t_0}^T \Bigl(\int_{\mathbb R^d} \sigma H_{pp}(x,Du) Dz\cdot Dz m dx + \langle \frac{\delta F}{\delta m}(\cdot,m(t))(\mu(t)),\mu(t)\rangle \Bigr)dt\\
& \qquad +
\langle \frac{\delta G}{\delta m}(\cdot,m(T))(\mu(T)),\mu(T)\rangle =0.
\end{split}
\end{equation}
Using that $(m,\alpha)$ is a minimizer as well as the second-order condition \eqref{ineq.SOC} with $(\rho, \beta)= (\mu, \sigma H_{pp}(x,Du)Dz)$ and recalling that $L_{\alpha,\alpha}(x,\alpha(t,x))H_{pp}(x,Du(t,x))= Id$, we get
\begin{align*}
& \int_{t_0}^T \int_{\mathbb R^d} \Bigl(\sigma^2 H_{pp}(x,Du) Dz\cdot Dzm \\
& \qquad +\int_{\mathbb R^d} \frac{\delta F}{\delta m}(x,m(t),y)\mu(t,x)\mu(t,y)dy\Bigr)dxdt + \int_{\mathbb R^{2d}} \frac{\delta G}{\delta m}(x,m(T),y)\mu(T,x)\mu(T,y)dydx\geq 0,
\end{align*}
and, in view of \eqref{ukzajenrsd},
$$
(\sigma -\sigma^2) \int_{t_0}^T \int_{\mathbb R^d} H_{pp}(x,Du) Dz\cdot Dz\ m\leq 0.
$$
Since $H_{pp}>0$ and $\sigma -\sigma^2>0$, the last inequality yields $Dz\ m=0,$ from which we easily conclude, going back to the equations satisfied by $\mu$ and by $z$, that $(z, \mu)=(0,0)$.
\end{proof}
We turn next to $d O$. The next lemma establishes an important property together with the fact $\mathcal O$ is not empty.
A similar statement is proved in \cite{BrCa} when the state space is the torus. The adaptation to the whole space is given here for the sake of completeness.
\begin{lem}\label{lem.Odense} Assume \eqref{ass.main}. Fix $(t_0,m_0)\in [0,T)\times \mathcal P_2(\R^d)$ and let $(m,\alpha)$ be a minimizer for $\mathcal U(t_0,m_0)$. Then $(t,m(t))$ belongs to $\mathcal O$ for any $t\in (t_0,T)$.
\end{lem}
\begin{proof} Fix $(t_0,m_0)\in [0,T)\times \mathcal P_2(\R^d)$, and let $(m,\alpha)$ be a minimizer for $\mathcal U(t_0,m_0)$ and $u$ its associated multiplier.
\vskip.075in
For $t_1\in (t_0,T)$, set $m_1=m(t_1)$ and let $(\tilde m, \tilde \alpha)$ be an optimal solution for $\mathcal U(t_1,m_1)$ with associated multiplier $\tilde u$. Since, in view of the dynamic programming principle,
$$
(\hat m, \hat \alpha) = \left\{\begin{array}{ll}
(m,\alpha) & {\rm on }\; [t_0,t_1)\times \mathbb R^d,\\[1.5mm]
(\tilde m, \tilde \alpha) & {\rm on}\; [t_1,T]\times \mathbb R^d,
\end{array}
\right.
$$
is optimal for $\mathcal U(t_0,m_0)$, we know from Lemma~\ref{lem.reguum} that $\hat \alpha \in C^{(1+\delta)/2, 1+\delta}$. It follows that $\alpha(t_1,\cdot)= \tilde \alpha(t_1,\cdot)$ and thus that $Du(t_1,\cdot)= D\tilde u(t_1,\cdot)$. Thus, the pair $((z^k)_{k=1, \dots, d},\mu)= ((\partial_{x_k}(u-\tilde u))_{k=1, \dots, d}, m-\tilde m)$ solves the system
\begin{equation}\label{eq.systLM}
\begin{split}
& -\partial z^k -\Delta z^k +g^k(t,x)= 0 \ \ {\rm in}\ \ (t_1,T)\times \mathbb R^d,\\
& \partial \mu -\Delta \mu +{\rm div}(h)= 0 \ \ {\rm in} \ \ (t_1,T)\times \mathbb R^d,\\
& \mu(t_1)=0 \ \ z^k(t_1, \cdot)= 0 \ \ {\rm in}\ \ \mathbb R^d,
\end{split}
\end{equation}
where
\begin{align*}
g^k(t,x)& = H_{x_k}(x,Du)-H_{x_k}(x,D\tilde u)+H_p(x,Du)\cdot D(\partial_{x_k} u)-H_p(x,D\tilde u)\cdot D(\partial_{x_k}\tilde u)\\
& \qquad\qquad\qquad -F_{x_k}(x,m(t))+F_{x_k}(x,m^1(t)), \\
h&=H_p(Du)m-H_p(D\tilde u)\tilde m.
\end{align*}
In order to estimate $g^k$ and $h$, we note that, since $t_1>t_0$, $m, \tilde m \in C^{1,2}([t_1,T]\times \mathbb R^d)$ and $m(t,\cdot)$ and $\tilde m(t,\cdot)$ are bounded in $L^2$. It follows that
\begin{align} \label{kuqyesrdnfgc}
&\sum_{k=1}^d |g^k(t,x)|^2 \leq C(|z(t,x)|^2+|Dz(t,x)|^2+\|\mu(t)\|_{L^2}^2), \notag\\
&|h(t,x)|^2\leq C (|z(t,x)|^2+|\mu(t,x)|^2),\\[1.2mm]
& |{\rm div}(h(t,x))|^2\leq C (|z(t,x)|^2+|Dz(t,x)|^2+|\mu(t,x)|^2+|D\mu(t,x)|^2). \notag
\end{align}
Then a Lions-Malgrange-type argument shows that $z_k=\mu=0$, and, hence, the solution starting from $(t_1,m_1)$ is unique. We refer to Lions and Malgrange \cite{LiMa} for the original argument and Cannarsa and Tessitore \cite{CaTe} and \cite{BrCa} for its adaptation to forward-backward equations.
\vskip.075in
Next we check that this solution is stable. Let $(z,\mu)$ be a solution to \eqref{MFGL}
in $[t_1,T]\times \mathbb R^d$ with $\sigma=1$, which by the standard parabolic regularity is actually classical. An elementary calculation yields
\begin{equation}\label{tildeJDz}
\begin{split}
&\int_{t_1}^T\left(\int_{\mathbb R^d} H_{pp}(x,Du(t,x))Dz\cdot Dz\ m dx+\langle \frac{\delta F}{\delta m}(\cdot,m(t)), \mu(t),\mu(t)\rangle \right)dt\\[1.2mm]
& \hskip1in +\langle \frac{\delta G}{\delta m}(\cdot,m(t))(\mu(t)),\mu(t) \rangle =0.
\end{split}
\end{equation}
Using Lemma~\ref{lem.secondordercond}, we know that, for any $\beta\in L^\infty([t_0,T]\times \mathbb R^d;\mathbb R^d)$ vanishing near $t=t_0$, if $\rho$ is the solution in the sense of distributions to \eqref{eq.forsecondorder} in $[t_0,T]\times \mathbb R^d$, then
\begin{align*}
\tilde J(\beta) = & \int_{t_0}^T \Bigl( \int_{\mathbb R^d} L_{\alpha,\alpha}(x,\alpha(t,x)) \beta(t,x)\cdot \beta(t,x)m(t,dx) +\langle \frac{\delta F}{\delta m}(\cdot,m(t),\cdot), \rho(t)\otimes \rho(t)\rangle\Bigr)dt \notag\\
& \qquad + \langle \frac{\delta G}{\delta m}(\cdot,m(T),\cdot), \rho(T,\cdot)\otimes \rho(T,\cdot)\rangle\geq 0.
\end{align*}
The solution $\bar \rho$ to \eqref{eq.forsecondorder} associated to the map $\bar \beta$ defined by $\bar \beta =0$ on $[t_0,t_1)$ and $\bar \beta= -H_{pp}(x,Du)Dz$ on $[t_1,T]$ is given by $\bar \rho(t)=0$ on $[t_0,t_1)\times \mathbb R^d$ and $\bar \rho(t)= \mu(t)$ on $[t_1,T]\times \mathbb R^d$.
\vskip.075in
It then follows from \eqref{tildeJDz} that $\tilde J(\bar \beta)=0$, and, hence, $\bar \beta$ is a minimizer for $\tilde J$, a fact that, by standard arguments (see, for example, \cite{BrCa})), implies that $\beta$ is a continuous map. Thus $Dz(t_1,\cdot)=0$.
\vskip.075in
We differentiate with respect to space variable the first equation in \eqref{MFGL} and obtain that \\$((\partial_{x_k}z)_{k=1,\dots, d}, \mu)$ solves a system of the form \eqref{eq.systLM} with zero initial condition and data $g$ and $h$ satisfying \eqref{kuqyesrdnfgc}.
This implies, as before, that $((\partial_{x_k}z)_{k=1,\dots, d}, \mu)=(0,0)$. Coming back to \eqref{MFGL}, we obtain $z=\mu=0$. Therefore the solution is stable.
\end{proof}
The next theorem establishes the key properties of $d O$.
\begin{thm}\label{thm.Oopendense} Assume \eqref{ass.main}. The set $\mathcal O$ is open and dense in $[0,T)\times \mathcal P_2(\R^d)$.
\end{thm}
\begin{proof}
Lemma \ref{lem.Odense} implies that the set $\mathcal O$ is a nonempty, dense subset of $[0,T)\times \mathcal P_2(\R^d)$.
\vskip.075in
Next we show that $\mathcal O$ is open arguing by contradiction. For this, we fix $(t_0,m_0)\in \mathcal O$ and assume that there are initial positions $(t^n,m^n_0)\notin \mathcal O$ which converge to $(t_0, m_0)$ (in $\mathcal P_2(\R^d)$ for $(m^n_0)_{n\in \mathbb N}$).
Let $(m,\alpha)$ be the unique and stable minimizer for $(t_0,m_0)$ and $u$ be the associated multiplier, that is, $\alpha= -H_p(x,Du)$ and the pair $(u,m)$ solves \eqref{MFG}.
\vskip.075in
Since $(t^n,m^n_0)\notin \mathcal O$, there are two cases (up to subsequences): either, for all $n$, there exist several minimizers for $\mathcal U(t^n, m^n_0)$ or, for all $n$, there exists a unique minimizer which is not stable. This latter case is ruled out by Lemma \ref{lem.estLS-Appen} and the strong stability of $(m,\alpha)$.
\vskip.075in
It remains to consider the first case and we argue as follows. Let $(m^{n,1},\alpha^{n,1})$ and $(m^{n,2},\alpha^{n,2})$ be two distinct minimizers starting from $(t^n, m^n_0)$ with associated multipliers $u^{n,1}$ and $u^{n,2}$ respectively.
\vskip.075in
Since the problem with initial condition $(t_0,m_0)$ has a unique minimizer, it follows from Lemma \ref{lem.stabilo} that, for $i=1,2$, the $(m^{n,i},\alpha^{n,i})$'s converge to $(m,\alpha)$ in $C^0([0,T]; \mathcal P_1(\R^d))\times C^{\delta/2\delta}$ while the $u^{n,i}$'s, $Du^{n,i}$'s and $D^2u^{n,i}$'s converge to $u$, $Du$ and $D^2u$ respectively in $C^{\delta/2,\delta}$.
\vskip.075in
Set
$$
\theta^n = \|Du^{n,1}-Du^{n,2}\|_{C^{\delta/2,\delta}}+\sup_{t\in [t_0,T]} {\bf d}_1(m^{n,1}(t),m^{n,2}(t))
$$ and note that, since $(m^{n,1},\alpha^{n,1})$ and $(m^{n,2},\alpha^{n,2})$ are distinct, $\theta^n>0$ and, in view of the previous discussion, $\theta_n\to 0$, and,
finally,
\begin{equation}\label{lizekjsrdfkekrf}
\theta^n \leq C \|Du^{n,1}-Du^{n,2}\|_{C^{\delta/2,\delta}}.
\end{equation}
This last estimate follows from the fact that again the uniform parabolicity implies that the $D^2u^{n,i}$'s are uniformly bounded. Applying Gronwall's inequality to the stochastic differential equations associated with the Kolmogorov equations satisfied by $m^{n,1}$ and $m^{n,2}$, we find the distance \\
$\sup_{[t_n, T]} {\bf d}_1(m^{n,1}(t),m^{n,2}(t))$ is controlled by $ C \|Du^{n,1}-Du^{n,2}\|_\infty$ and thus by $C \|Du^{n,1}-Du^{n,2}\|_{C^{\delta/2,\delta}}$.
\vskip.075in
Next we introduce the differences $z^n= (u^{n,1}-u^{n,2})/\theta^n$, $\mu^n= (m^{n,1}-m^{n,2})/\theta^n$ and observe that
$$
\left\{\begin{array}{l}
\displaystyle -\partial_t z^n -\Delta z^n +H_p(x,Du^{n,1})\cdot Dz^n= \frac{\delta F}{\delta m}(x,m^{n,1}(t))(\mu^n(t))+R^{n,1},\\[1.5mm]
\displaystyle \partial_t \mu^n -\Delta \mu^n -{\rm div}(H_p(x,Du^{n,1})\mu^n)-{\rm div}(H_{pp}(x,Du^{n,1})Dz^n m^{n,1})={\rm div}(R^{2,n}),\\[1.3mm]
\displaystyle \mu^n(t_0)=0, \; z^n(T, x)=\frac{\delta G}{\delta m}(x,m^{n,1}(T))(\mu^n(T))+R^{3,n}
\end{array}\right.
$$
with
\begin{align*}
R^{n,1} = & (\theta^n)^{-1} \Bigl[H(x,Du^{n,2})-H(x,Du^{n,1})-H_p(x,Du^{n,1})\cdot (Du^{n,2}-Du^{n,1})\\
& \qquad\qquad -\Bigl( F(x,m^{n,2}(t))-F(x,m^{n,1}(t))- \frac{\delta F}{\delta m}(x,m^{n,1}(t))(m^{n,2}(t)-m^{n,1}(t)) \Bigr)\bigr],
\end{align*}
\begin{align*}
R^{n,2}= & -(\theta^n)^{-1} \Bigl[ H_p(x,Du^{n,2})m^{n,2}- H_p(x,Du^{n,1})m^{n,1}-H_p(x,Du^{n,1})(m^{n,2}-m^{n,1})\\
& \qquad \qquad -H_{pp}(x,Du^{n,1})(Du^{n,2}-Du^{n,1})m^{n,1})\Bigr],
\end{align*}
and
$$
R^{n,3}=-(\theta^n)^{-1} \Bigl[ G(x,m^{n,2}(T))-G(x,m^{n,1}(T))- \frac{\delta G}{\delta m}(x,m^{n,1}(T))(m^{n,2}(T)-m^{n,1}(T)) \bigr].
$$
\vskip.075in
It follows from the regularity of $F$, $G$ and $H$ and the definition of $\theta^n$ that
\begin{equation}\label{estiR1R2}
\|R^{n,1}\|_{C^{\delta/2,\delta}}+\|R^{n,3}\|_{C^{2+\delta}} + \sup_{t\in [t_0,T]} \|R^{n,2}(t)\|_{(W^{1,\infty})'} \leq C\theta^n.
\end{equation}
However, Lemma \ref{lem.estLS-Appen} yields that the sequence $(z^n)_{n\in \mathbb N}$ tends to $0$ in $C^{(1+\delta)/2,1+ \delta}$, which contradicts \eqref{lizekjsrdfkekrf}.
\end{proof}
\subsection*{The smoothness of $\mathcal U$ in $\mathcal O$} We prove here Theorem~\ref{thm.mainreguU0}.
\vskip.075in
Before we present the arguments, we state below as lemma a preliminary fact that is needed to establish the regularity of $d U$. It is about a stability property in the appropriate norms for the multipliers associated with minimizers starting in $d O$. In turn, this will allow us to compute the derivative of $d U$ with respect to $m$. Its proof is presented at the end of this subsection.
\vskip.075in
\begin{lem}\label{lem.Lip} Assume \eqref{ass.main} and fix $(t_0,m_0)\in \mathcal O$. There exists $\delta, C>0$ such that, for any $t_0'$, $m_0^1,m_0^2$ satisfying $(t_0',m_0^i)\in \mathcal O$, $|t_0'-t_0|<\delta$, ${\bf d}_2(m_0,m_0^i)<\delta$, if $(m^i,\alpha^i)$ is the unique minimizer starting from $(t_0', m_0^i)$ with associated multiplier $u^i$ for $i=1$ and $i=2$, then
$$
\|u^1-u^2\|_{C^{(2+\delta)/2,2+\delta}}+ \sup_{t\in [t_0',T]} {\bf d}_2(m^1(t),m^2(t)) \leq C {\bf d}_2(m_0^1,m^2_0).
$$
\end{lem}
We continue with the proof of Theorem~\ref{thm.mainreguU0} which consists of three parts. In the first, we establish the regularity of $d U$ in $d O$. In the second, we show that the infinite dimensional Hamilton-Jacobi equation \eqref{HJdsO0} is satisfied in $d O$. Finally, the third part is about \eqref{takis50}.
\begin{proof}[Proof of Theorem \ref{thm.mainreguU0}] {\it Part 1: The regularity of $d U$.}
Lemma \ref{lem.estiVN} yields the Lipschitz continuity of $\mathcal U$.
\vskip.075in
We establish that $\mathcal U$ is differentiable at any $(t_0,m_0)\in \mathcal O$.
We fix such a $(t_0,m_0)$. Let $(m,\alpha)$ be the unique stable minimizer for $\mathcal U(t_0,m_0)$ and $u$ its associated multiplier. We check that $D_m\mathcal U(t_0,m_0,\cdot)$ exists and is given by $Du(t_0,\cdot)$.
\vskip.075in
Let $\delta>0$ and $\delta'\in (0,\delta)$ be such that the $\delta-$neighborhood of the $[0,T]\times \mathcal P_2(\R^d)$-compact set $\{(t,m(t)): t\in [t_0,T]\}$ is contained in $d O$, and,
for any $m^1_0\in B(m_0,\delta')$, $\sup_{t\in [t_0,T]} {\bf d_1}(m(t),m^1(t))<\delta$, where $(m^1,\alpha^1)$ is the minimizer for $\mathcal U(t_0, m_0^1)$.
\vskip.075in
Fix $m^1_0\in B(m_0,\delta')$. Let $(m^1,\alpha^1)$ be the minimizer for $\mathcal U(t_0, m_0^1)$, $u^1$ its associated multiplier, $(z,\mu)$ the solution of the linearized system \eqref{MFGL} with initial condition $\mu (0)= m^1_0-m_0$, set
$$
(w,\rho)= (u^1-u-z, m^1-m-\mu)
$$
and note that $(w,\rho)$ satisfies the linearized system \eqref{eq.LS} with $\xi= 0$,
$$
R^1= -(H(x,Du^1)-H(x,Du)-H_p(x,Du)\cdot (Du^1-Du)) +F(x,m^1)-F(x, m) -\frac{\delta F}{\delta m}(x,m(t))(m^1(t)-m(t)),
$$
\begin{align*}
R^2=& H_p(x,Du^1)m^1-H_p(x,Du)m -H_p(x,Du)(m^1-m)-H_{pp}(x,Du)\cdot (Du^1-Du)m,
\end{align*}
and
$$
R^3=G(x,m^1)-G(x, m) -\frac{\delta G}{\delta m}(x,m(t))(m^1(t)-m(t)).
$$
Then, using Lemma~\ref{lem.Lip}, we get
$$
\|R^1\|_{C^{\delta/2,\delta}}+\|R^{3}\|_{C^{2+\delta}}+ \sup_{t\in [t_0,T]}\|R^{2}(t)\|_{(W^{1,\infty})'} \leq C{\bf d}_1^2(m^1_0,m_0)
$$
and, in view of Lemma~\ref{lem.estLS-Appen},
$$
\|u^1-u-w\|_{C^{(2+\delta)/2,2+\delta}}+\sup_{t\in [t_0,T]} \|m^1(t)-m(t)-\mu(t)\|_{(C^{2+\delta})'} \leq C{\bf d}_1^2(m^1_0,m_0).
$$
Recall that $\alpha^1= -H_p(x, Du^1)$. Thus
$$
\alpha^1= \alpha - H_{pp}(x,Du). Dw+ o({\bf d}_1(m^1_0,m_0)).
$$
where $o(\cdot)$ is small in uniform norm.
\vskip.075in
It follows that
\begin{align*}
&\mathcal U(t_0,m^1_0) = \int_{t_0}^T \Bigl( \int_{\mathbb R^d} L(x,\alpha^1) m^1 +\mathcal F(m^1)\Bigr)dt + \mathcal G(m^1(T)) \\
&\qquad = \mathcal U(t_0,m_0) + \int_{t_0}^T \int_{\mathbb R^d} \Bigl(D_\alpha L(x,\alpha)\cdot(-H_{pp}(x,Du)Dw m +L(x,\alpha)\mu(t,x) +F(x,m(t))\mu(t,x)\Bigr)dxdt \\
& \qquad \qquad + \int_{\mathbb R^d} G(x,m(T))\mu(T,x)dx + o({\bf d}_1(m^1_0,m_0)).
\end{align*}
On the other hand, recalling the equations satisfied by $u$ and $\mu$ and using duality we find
\begin{align*}
& \int_{\mathbb R^d}G(x,m(T))\mu(T,x)dx - \int_{\mathbb R^d} u(t_0,x)(m^1_0-m_0)(dx)\\
& \qquad = \int_{t_0}^T \int_{\mathbb R^d} \Bigl( (H(x,Du) -F(x,m(t))-H_p(x,Du)\cdot Du)\mu -H_{pp}(x,Du)Du\cdot Dw m \Bigr).
\end{align*}
Thus,
\begin{align*}
&\mathcal U(t_0,m^1_0) = \mathcal U(t_0,m_0) + \int_{\mathbb R^d} u(t_0,x)(m^1_0-m_0)(dx) \\
& \qquad + \int_{t_0}^T \int_{\mathbb R^d} \Bigl((H(x,Du)-H_p(x,Du)\cdot Du +L(x,\alpha))\mu \\
& \qquad -(H_{pp}(x,Du)Du\cdot Dw + D_\alpha L(x,\alpha)\cdot(H_{pp}(x,Du)Dw)) m \Bigr)dxdt+ o({\bf d}_1(m^1_0,m_0))
\end{align*}
In view of the relationship (convex duality) between $H$ and $L$ and the fact that $\alpha=-H_p(x,Du)$, we have $D_\alpha L(x,\alpha)=-Du$ and, therefore,
$$
H(x,Du)-H_p(x,Du)\cdot Du +L(x,\alpha)=0
$$
and
$$
H_{pp}(x,Du)Du\cdot Dw + D_\alpha L(x,\alpha)\cdot(H_{pp}(x,Du)Dw)=0.
$$
Thus,
\begin{align*}
&\mathcal U(t_0,m^1_0) = \mathcal U(t_0,m_0) + \int_{\mathbb R^d} u(t_0,x)(m^1_0-m_0)(dx) + o({\bf d}_1(m^1_0,m_0)).
\end{align*}
It follows that $\mathcal U(t_0,\cdot)$ has a linear derivative at $m_0$ given by $u(t_0,\cdot)$ and, hence, $$D_mU(t_0,m_0,x)= Du(t_0,x).$$
\vskip.075in
Recalling the stability of the map $m_0\to (u,m)$ proved in Lemma \ref{lem.stabilo}, we actually have that $(t_0,m_0)\to D_mU(t_0,m_0, \cdot)$ is continuous in $\mathcal O$ with the respect to the ${\bf d}_2-$distance for the \\ measure variable into $C^2$.
\vskip.075in
{\it Part 2: The Hamilton-Jacobi equation.}
Next we show that $d U$ is a classical solution to \eqref{HJdsO0}.
\vskip.075in
Using the notation of Part 1 and the dynamic programming principle with $h>0$ small, we find
$$
\mathcal U(t_0,m_0)= \int_{t_0}^{t_0+h}\Bigl( \int_{\mathbb R^d} L(x,\alpha(t,x))m(t,x)dx+ \mathcal F(m(t))\Bigr)dt + \mathcal U(t_0+h, m(t_0+h)),
$$
and, in view of the $C^1-$regularity of $\mathcal U$,
\begin{align*}
\mathcal U(t_0+h, m(t_0+h))- \mathcal U(t_0+h, m_0) = &\int_{t_0}^{t_0+h} \int_{\mathbb R^d} \Bigl( {\rm Tr}D^2_{ym} \mathcal U(t_0+h, m(t),y)\\
&\;
+
D_m\mathcal U(t_0+h, m(t),y)\cdot H_p(t,Du(t,y))\Bigr) m(t,dy)dy dt.
\end{align*}
It follows that $\partial_t \mathcal U(t_0,m_0)$ exists and is given by
\begin{align*}
\partial_t \mathcal U(t_0,m_0)
=& - \int_{\mathbb R^d} L(x,\alpha(t_0,x))m_0(dx)- \mathcal F(m_0) \\
& -
\int_{\mathbb R^d} \Bigl( {\rm Tr}D^2_{ym} \mathcal U(t_0, m_0,y)
+
D_m\mathcal U(t_0, m_0,y)\cdot H_p(t_0,Du(t_0,y))\Bigr) m_0(dy).
\end{align*}
Since $Du(t_0,x)= D_m\mathcal U(t_0,m_0,x)$ and $\alpha(t_0,x)=-H_p(x,Du(t_0,x))$, \eqref{HJdsO0} is then satisfied.
\vskip.075in
{\it Part 3: The regularity of $D_md U$.} We prove that \eqref{takis50} holds.
\vskip.075in
Let $\delta,C>0$ be given by Lemma \ref{lem.Lip}. For any $t\in (t_0-\delta, t_0+\delta)$ and $m^2_0,m^2_0\in \mathcal P_2(\R^d)$ with ${\bf d}_2(m_0,m^1_0)<\delta$, ${\bf d}_2(m_0,m^2_0)<\delta$, $(t,m^1_0)\in \mathcal O$ and $(t,m^2_0)\in \mathcal O$, and for any $x^1,x^2\in \mathbb R^d$, let $u^1$ (respectively $u^2$) be the multiplier associated with the unique minimizer $(m^1,\alpha^1)$ for $\mathcal U(t,m^1_0)$ (respectively with the unique minimizer $(m^2,\alpha^2)$ for $\mathcal U(t, m^2_0)$).
\vskip.075in
Since, as already established, $D_m\mathcal U(t,m^1_0, x^1)=Du^1(t,x^1)$ and $D_m\mathcal U(t,m^2_0, x^2)=Du^2(t,x^2)$, we find, using Lemma~\ref{lem.reguum} and Lemma~\ref{lem.Lip}, that
\begin{align*}
\left| D_m\mathcal U(t,m^1_0, x^1)- D_m\mathcal U(t,m^2_0, x^2)\right| & \leq \left| Du^1(t, x^2)- Du^2(t,x^2)\right| +\|D^2u^1\|_\infty|x^1-x^2| \\
& \leq C ({\bf d}_2(m^1_0,m^2_0)+|x^1-x^2|).
\end{align*}
\end{proof}
We conclude with the remaining proof.
\begin{proof}[Proof of Lemma~\ref{lem.Lip}] Let $(m,\alpha)$ be the unique stable minimizer starting from $(t_0,m_0)$ with multiplier $u$.
It follows from Lemma \ref{lem.StrongStab} that the associated linear system \eqref{MFGL} is strongly stable.
\vskip.075in
We set $V= -H_p(x,Du)$ and $\Gamma= -H_{pp}(x,Du)$, consider the neighborhood $\mathcal V$ (in the local uniform convergence) of $(V,\Gamma)$ given in Lemma \ref{lem.estLS-Appen}, and choose $\delta>0$ so that, for any $m_0^1$ such that ${\bf d}_2(m_0,m^1_0)<\delta$, we have that $$(V^1, \Gamma^1)\in \mathcal V,$$ with $V^1=-H_p(x,Du^1)$ and $\Gamma_1= H_{pp}(x,Du^1)$, $u^1$ being the multiplier associated with the optimal solution $(m^1,\alpha^1)$ for $\mathcal U(t_0,m_0^1)$. The above is possible since, if $\delta$ is small, then, in view of Lemma \ref{lem.stabilo}, $u^1$ is close to $u$ in $C^{\delta/2,\delta}$.
\vskip.075in
Furthermore, choosing, if necessary, $\delta$ even smaller, we have that, for some $\eta>0$ to be chosen below, and, for any $t_0'$, $m_0^i$, $(m^i, \alpha^i)$ and $u^i$ as above such that $|t_0'-t_0|<\delta$ and ${\bf d}_2(m_0,m_0^i)<\delta$ for $i=1$ and $i=2$,
\begin{equation}\label{lzkej:snrdf,gc}
\|u^1-u^2\|_{C^{(2+\delta)/2,2+\delta}} + \sup_{t\in [t_0',T]} {\bf d}_2(m^1(t),m^2(t)) \leq \eta.
\end{equation}
Classical estimates on the Kolmogorov equation (see, for instance, \cite{CDLL}) yield that, for $t_0'$, $m_0^i$, $(m^i, \alpha^i)$ and $u^i$ as above,
we have
\begin{equation}\label{zuesdzmesdf}
\sup_{t\in [t_0',T]} {\bf d}_2(m^1(t),m^2(t))\leq C( {\bf d}_2(m^1_0,m^2_0)+ \|Du^1-Du^2\|_\infty),
\end{equation}
for a constant $C$ depending on $T$, $H$ and $\|D^2u^i\|_\infty$, which is uniformly bounded by Lemma \ref{lem.apriori-Appen}.
\vskip.075in
Then the pair
$$
(z,\mu)= (u^1-u^2,m^1-m^2)
$$
satisfies the linearized system \eqref{eq.LS} with
\begin{align*}
V(t,x)= -H_p(x,Du^2), \ \ \Gamma= -H_{pp}(x,Du^2), \ \ \xi= m^1_0-m_0^2,
\end{align*}
\begin{align*}
R^1&= -(H(x,Du^1)-H(x,Du^2)-H_p(x,Du^2)\cdot (Du^1-Du^2)) +F(x,m^1(t))-F(x, m^2(t)) \\
& \qquad \qquad -\frac{\delta F}{\delta m}(x,m^2(t))(m^1(t)-m^2(t)),
\end{align*}
\begin{align*}
R^2&= H_p(x,Du^1)m^1-H_p(x,Du^2)m^2 -H_p(x,Du^2)(m^1-m^2)-H_{pp}(x,Du^2)\cdot (Du^1-Du^2)m^2\\
&=(H_p(x,Du^1)-H_p(x,Du^2))(m^1-m^2)\\
&\qquad +(H_p(x,Du^1)-H_p(x,Du^2)-H_{pp}(x,Du^2)\cdot (Du^1-Du^2))m^2,
\end{align*}
and
$$
R^3=G(x,m^1)-G(x, m^2) -\frac{\delta G}{\delta m}(x,m^2(t))(m^1(t)-m^2(t)).
$$
Note that
\begin{align*}
R^2=& (H_p(x,Du^1)-H_p(x,Du^2))(m^1-m^2)\\
& \qquad +(H_p(x,Du^1)-H_p(x,Du^2)-H_{pp}(x,Du^2)\cdot (Du^1-Du^2))m^2.
\end{align*}
Thus
\begin{align*}
M& =\|\xi\|_{(W^{1,\infty})'}+\|R^1\|_{C^{\delta/2,\delta}}+\|R^{3}\|_{C^{2+\delta}}+ \sup_{t\in [t_0',T]}\|R^{2}(t)\|_{(W^{1,\infty})'}\\
& \leq {\bf d}_1(m^1_0,m^2_0)+
C\{\|Du^1-Du^2\|_{C^{\delta/2,\delta}}^2+ \sup_t{\bf d}^2_{2}(m^1(t),m^2(t))\}.
\end{align*}
It follows from Lemma \ref{lem.estLS-Appen} that
$$
\|u^1-u^2\|_{C^{(2+\delta)/2, 2+\delta}}\leq C\{ {\bf d}_1(m^1_0,m^2_0)+ \|Du^1-Du^2\|_{C^{\delta/2,\delta}}^2+ \sup_t{\bf d}^2_{2}(m^1(t),m^2(t))\}.
$$
Hence, choosing $\eta>0$ small enough in \eqref{lzkej:snrdf,gc}, we find
$$
\|u^1-u^2\|_{C^{(2+\delta)/2, 2+\delta}}\leq C\{ {\bf d}_1(m^1_0,m^2_0)+ \sup_t{\bf d}^2_{2}(m^1(t),m^2(t))\},
$$
and inserting the last inequality in \eqref{zuesdzmesdf} we obtain
$$
\sup_t{\bf d}_{2}(m^1(t),m^2(t)) \leq C\{ {\bf d}_2(m^1_0,m^2_0)+ \sup_t{\bf d}^2_{2}(m^1(t),m^2(t))\},
$$
which yields, for $\eta>0$ small enough,
$$
\sup_t{\bf d}_{2}(m^1(t),m^2(t)) \leq C {\bf d}_2(m^1_0,m^2_0).
$$
Going back to the previous inequality on $\|u^2-u^2\|_{C^{(2+\delta)/2, 2+\delta}}$ completes the proof.
\end{proof}
\section{The propagation of chaos} \label{sec.proofThm2}
We present the proof of Theorem \ref{thm.main20}, which consists of several steps each of which is stated below as separate lemma.
\vskip.075in
In preparation, we fix $(t_0,m_0)\in \mathcal O$ with $M_{d+5}(m_0)<+\infty$ and let $(m,\alpha)$ be the unique minimizer for $\mathcal U(t_0,m_0)$.
\vskip.075in
It follows from Theorem~\ref{thm.mainreguU0} and the compactness of the curve $\{(t,m(t)) : t\in [t_0,T]\}$ that there exists
$\delta, C>0$ such that, for any $t_1\in [t_0,T]$, $t\in (t_1-\delta, t_1+\delta)$ and $m^1_0,m^2_0\in \mathcal P_2(\R^d)$ with ${\bf d}_2(m(t_1),m^1_0)<\delta$, ${\bf d}_2(m(t_1),m^2_0)<\delta$, $(t_1,m^1_0)\in \mathcal O$ and $(t_1,m^2_0)\in \mathcal O$, and $x^1,x^2\in \mathbb R^d$,
\begin{equation}\label{DmULipsch}
|D_m\mathcal U(t_1,m^1_0,x^1)-D_m\mathcal U(t_1,m^2_0,x^2)|\leq C (|x^1-x^2|+ {\bf d}_2(m^1_0,m^2_0)).
\end{equation}
For $\sigma\in (0,\delta)$, set
$$
V_\sigma = \{(t,m')\in [t_0, T]\times \mathcal P_2(\R^d) : {\bf d}_2(m',m(t))<\sigma\}
$$
and
$$
V^N_\sigma= \{(t, {\bf x})\in [0,T]\times (\mathbb R^d)^N : (t,m^N_{{\bf x}})\in V_\sigma\}.
$$
We consider the solution $({\bf X}^N_t)_{t\in [t_0,T]}=(X^{N,1}_t, \ldots,X^{N,N}_t )_{t\in [t_0,T]}$ to
\begin{equation}\label{def.XNi}
dX^{N,j}_t= Z^j-\int_{t_0}^t H_p(X^{N,j}_s, D_m\mathcal U(s,m^N_{{\bf X^{N}_s}},X^{N,j}_s))ds +\sqrt{2}(B^j_s-B^j_{t_0}),
\end{equation}
on the time interval $[t_0, \tau^N]$, where the stopping time $\tau^N$ is defined by
$$
\tau^N=\inf\left\{t\in [t_0,T], \; (t,{\bf X}^N_t) \notin V^N_{\delta/2} \right\} \ \ \text{or} \ \ \text{$\tau^N=T$, if there is no such a $t$.}
$$
Note that, in view of \eqref{DmULipsch}, ${\bf X}^N$ is uniquely defined.
\vskip.075in
Set
\begin{equation}\label{def.tildetau}
\tilde \tau^N=\inf\{ t\in [t_0,\tau^N], \; (t,{\bf Y}^N_t)\notin V^N_{\delta}\} \ \ \text{or} \ \ \text{$\tilde \tau^N=\tau^N$ is there is no such a $t$.}
\end{equation}
\begin{lem} \label{lem.aqemzjlrsdk} Assume \eqref{ass.main}. There is a constant $C>0$ depending on $(t_0,m_0)$ such that
\begin{equation}\label{aqemzjlrsdk}
\mathbb E\left[\sup_{t\in [t_0,\tau^N]} {\bf d}_1(m^N_{{\bf X}^N_t}, m(t)) \right]\leq C N^{-1/(d+8)}
\end{equation}
and
$$
\mathbb P\left[\tau^N<T\right] \leq C N^{-1/(d+8)}.
$$
\end{lem}
\begin{proof} The proof is standard and relies on propagation of chaos estimates; see, for instance, the proof of Theorem 5.6 in \cite{DeLaRa20}.
\vskip.075in
Let $\tilde X^{N,i}$ be the i.i.d. solutions to
$$
d\tilde X^{N,i}_t= -H_p(\tilde X^{N,j}_t, D_m\mathcal U(t,m(t),\tilde X^{N,i}_t) )dt +\sqrt{2}dB^j_t \ \ \ \tilde X^{N,i}_{t_0}=Z^i,
$$
which exist on $[t_0,T]$, in view of the global Lipschitz property of $D_m\mathcal U(t,m(t),\cdot)$.
\vskip.075in
Then \eqref{aqemzjlrsdk} is an easy consequence of the inequality
$$
\mathbb E\left[\sup_{t\in [0,T]} {\bf d}^2_2(m^N_{\tilde {\bf X}^N_t}, m(t))\right]\leq CN^{-2/(d+8)},
$$
which follows from Theorem 3.1 in Horowitz and Karandikar \cite{HoKa}, for some $C$ depending only on $(t_0,m_0)$ through the regularity of $D_m\mathcal U$ in $V_\delta$.
\vskip.075in
Then
$$
\mathbb P\left[\tau^N<T\right] \leq \mathbb P\left[ \sup_{t\in [t_0,\tau^N]} {\bf d}_1(m^N_{{\bf X}^N_t}, m(t))\geq \delta/2\right] \leq C \delta^{-1} N^{-1/(d+8)}.
$$
\end{proof}
Next, for $(t,{\bf x}) \in V^N_\delta$, we set
$$
\mathcal U^N(t, {\bf x})= \mathcal U(t, m^N_{{\bf x}}).
$$
It is immediate that $\mathcal U^N$ is $C^{1,1}$ on $V^N_\delta$ and (see, for example, \cite{CDLL})
$$
D_{x^j} \mathcal U^N(t, {\bf x})= \frac1N D_m \mathcal U(t, m^N_{{\bf x}}, x^j) \ \ \text{and} \ \ | D^2_{x^jx^j} \mathcal U^N(t, {\bf x})- \frac1N D^2_{ym} \mathcal U(t, m^N_{{\bf x}}, x^j)|\leq \frac{C}{N^2}.
$$
Finally, $\mathcal U^N$ satisfies
\begin{equation}\label{eq.OON}
\left\{\begin{array}{l}
\displaystyle -\partial_t \mathcal U^N(t,{\bf x}) -\sum_{j=1}^N \Delta_{x^j}\mathcal U^N(t,{\bf x}) +O^N(t,{\bf x})
+\frac1N\sum_{j=1}^N H(x^j, N D_{x^j}\mathcal U^N(t, {\bf x}))=\mathcal F(m^N_{{\bf x}})\ \ {\rm in}\ \ V^N_\delta, \\
\displaystyle \mathcal U^N(T,{\bf x})= \mathcal G(m^N_{{\bf x}})\ \ {\rm on}\ \ (\mathbb R^d)^N,
\end{array}\right.
\end{equation}
where
\begin{equation}\label{evalON}
|O^N(t,{\bf x})|\leq CN^{-1}\qquad a.e. .
\end{equation}
\begin{lem}\label{lem.sldjkfncv} Let ${\bf Y}^N=(Y^{N,i})_{i=1,\ldots, N}$ and $\tilde \tau^N$ be defined by \eqref{takis51} and \eqref{def.tildetau} respectively. Then,
\begin{align*}
& \mathbb E\bigl[ \int_{t_0}^{\tilde \tau^N} N^{-1}\sum_j |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N)-H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)|^2dt \bigr] \leq C(N^{-1}+ R^N),
\end{align*}
where $R^N=\|\mathcal U^N-\mathcal V^N\|_\infty$ and $\mathcal U^N$, $\mathcal V^N$ and their derivatives are evaluated at $(t, {\bf Y}^N_t)$.
\end{lem}
\begin{proof}
For $t\in [t_0,\tilde \tau^N]$ and, in view of \eqref{evalON}, we have
\begin{align*}
&d\mathcal U^N(t, {\bf Y}^N_t)= (\partial_t \mathcal U^N +\sum_j \Delta_{x^j} \mathcal U^N- \sum_j H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)\cdot D_{x^j}\mathcal U^N)dt+\sqrt{2}\sum_j D_{x^j}\mathcal U^N\cdot dB^j_t \\
& = (\frac1N \sum_j H(Y^{N,j}, N D_{x^j}\mathcal U^N(t, {\bf Y}^N_t)) - \sum_j H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)\cdot D_{x^j}\mathcal U^N+ O^N -\mathcal F(m^N_{{\bf Y}^N_t}))dt\\
& \qquad +\sqrt{2}\sum_j D_{x^j}\mathcal U^N\cdot dB^j_t \\
& \geq (\frac1N \sum_j (-L(Y^{N,j}, -H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N))+C^{-1} |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N)-H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)|^2) \\
& \qquad \qquad -CN^{-1} -\mathcal F(m^N_{{\bf Y}^N_t}))dt +\sqrt{2}\sum_j D_{x^j}\mathcal U^N\cdot dB^j_t,
\end{align*}
the inequality following from the uniform convexity of $H$ in bounded sets.
\vskip.075in
We take expectations and integrate between $t_0$ and $\tilde \tau^N$ above to get
\begin{align*}
&\mathbb E\left[\mathcal U^N(\tilde \tau^N, {\bf Y}^N_{\tilde \tau^N})\right] -\mathbb E\left[\mathcal U^N(t_0, {\bf Z}^N)\right] \\
& \geq \mathbb E\bigl[ \int_{t_0}^{\tilde \tau^N} (\frac1N \sum_j \Bigl(-L(Y^{N,j}, -H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N))+C^{-1} |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N)-H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)|^2\Bigr) \\
& \qquad \qquad -CN^{-1} -\mathcal F(m^N_{{\bf Y}^N_t}))dt \bigr].
\end{align*}
Rearranging, using the definition of $R^N$, the dynamic programming principle and the optimality of ${\bf Y}^N$ for $\mathcal V^N(t_0, {\bf Z}^N)$ we find
\begin{align*}
&\mathbb E\left[\mathcal U^N(t_0, {\bf Z}^N)\right] + \mathbb E\bigl[ \int_{t_0}^{\tilde \tau^N} \frac{1}{CN} \sum_j |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N)-H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)|^2dt \bigr] \\
& \leq \mathbb E\bigl[ \int_{t_0}^{\tilde \tau^N} (\frac1N \sum_j L(Y^{N,j}, -H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N))+CN^{-1} +\mathcal F(m^N_{{\bf Y}^N_t}))dt
+\mathcal V^N(\tilde \tau^N, {\bf Y}^N_{\tilde \tau^N})\bigr] + R^N\\
& \leq \mathbb E\left[\mathcal V^N(t_0, {\bf Z}^N)\right]+ CN^{-1}+ R^N,
\end{align*}
and, using once more the definition of $R^N$, we get
\begin{align*}
& \mathbb E\bigl[ \int_{t_0}^{\tilde \tau^N} \frac{1}{CN} \sum_j |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N)-H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N)|^2dt \bigr] \leq CN^{-1}+ 2R^N.
\end{align*}
\end{proof}
\begin{lem}[Convergence of optimal trajectories]\label{lem.CvOptiTraj} For ${\bf X}^N=(X^{N,i})$ and ${\bf Y}^N=(Y^{N,i})$ defined by \eqref{def.XNi} and \eqref{takis51} respectively, we have
$$
\mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} N^{-1}\sum_j | X^{N,j}_{s}-Y^{N,j}_{s}| \right]\leq C(N^{-1}+ R^N)^{1/2}
$$
and
$$
\mathbb P\left[ \tilde \tau^N<T\right]\leq C (N^{-1/(d+8)}+ (R^N)^{1/2}).
$$
\end{lem}
\begin{proof} Lemma~\ref{lem.sldjkfncv}, the regularity of $\mathcal U^N$ in \eqref{DmULipsch} and an application of Gronwall's inequality give the first inequality since, for any $t\in [t_0,T]$,
\begin{align*}
& \mathbb E\left[ \sup_{s\in [t_0,t\wedge \tilde \tau^N]} N^{-1}\sum_j | X^{N,j}_{s}-Y^{N,j}_{s}| \right]\\
& \leq \mathbb E\bigl[ \int_{t_0}^{t\wedge \tilde \tau^N}N^{-1} \sum_j |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N(t, {\bf Y}^N_t))-H_p(Y^{N,j}_t, ND_{x^j}\mathcal V^N(t, {\bf Y}^N_t))|dt \bigr]\\
& \qquad +
\mathbb E\bigl[ \int_{t_0}^{t\wedge \tilde \tau^N}N^{-1} \sum_j |H_p(Y^{N,j}_t, ND_{x^j}\mathcal U^N(t, {\bf Y}^N_t))-H_p(X^{N,j}_t, ND_{x^j}\mathcal U^N(t, {\bf X}^N_t))|dt \bigr] \\
&\leq C(N^{-1}+ R^N)^{1/2} + C N^{-1}\sum_j \mathbb E\left[\int_{t_0}^{t\wedge \tilde \tau^N}
|X^{N,j}_{s}-Y^{N,j}_{s}|ds\right].
\end{align*}
\vskip.075in
Then,
\begin{align*}
\mathbb P\left[ \tilde \tau^N<T \right] & \leq \mathbb P\left[ \tau^N<T\right]+ \mathbb P\left[ \sup_{s\in [t_0, \tilde \tau^N]} N^{-1}\sum_j | X^{N,j}_{s}-Y^{N,j}_{s}| >\delta/2\right] \\
& \leq C N^{-1/(d+8)} + C\delta^{-1} (N^{-1}+ R^N)^{1/2} .
\end{align*}
\end{proof}
We can proceed now with the proof of the propagation of chaos property.
\begin{proof}[Proof of Theorem \ref{thm.main20}] It is immediate that
\begin{equation}\label{takis52}
\begin{split}
& \mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} {\bf d}_1(m^N_{{\bf Y}^N_s}, m(s)) \right] \leq
\mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} {\bf d}_1(m^N_{{\bf Y}^N_s}, m^N_{{\bf X}^N_s}) \right] +
\mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} {\bf d}_1(m^N_{{\bf X}^N_s}, m(s)) \right] \\
& \qquad \leq \mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]}N^{-1}\sum_j | X^{N,j}_{s}-Y^{N,j}_{s}| \right] + \mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} {\bf d}_1(m^N_{{\bf X}^N_s}, m(s)) \right] .
\end{split}
\end{equation}
Lemma~\ref{lem.CvOptiTraj} gives that the first term in the right-hand side of \eqref{takis52} is not larger than $C(N^{-1}+ R^N)^{1/2} $, where $R^N$ can be estimated by Proposition \ref{prop.quantifCV}, while Lemma \ref{lem.aqemzjlrsdk} implies that the second term in the right-hand side of \eqref{takis52} is not larger than $C N^{-1/(d+8)}$.
\vskip.075in
Thus, for some $\gamma'$ depending on $d$ only and $C$ depending on the initial condition $(t_0,m_0)$,
\begin{align*}
\mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} {\bf d}_1(m^N_{{\bf Y}^N_s}, m(s)) \right] & \leq CN^{-\gamma'}.
\end{align*}
Finally, we have
\begin{align*}
& \mathbb E\left[ \sup_{s\in [t_0, T]} {\bf d}_1(m^N_{{\bf Y}^N_s}, m(s)) \right] \\
&\qquad \leq \mathbb E\left[ \sup_{s\in [t_0, \tilde \tau^N]} {\bf d}_1(m^N_{{\bf Y}^N_s}, m(s)) \right]
+ \mathbb E\left[ \sup_{s\in [t_0, T]} {\bf d}_1^2(m^N_{{\bf Y}^N_s}, m(s)) \right]^{1/2} \mathbb P\left[ \tilde \tau^N<T\right]^{1/2} \\
& \qquad \leq CN^{-\gamma'} +C (N^{-1/(d+8)}+ (R^N)^{1/2})^{1/2},
\end{align*}
the last inequality coming from Lemma \ref{lem.CvOptiTraj} and the facts that, since, in view of Lemma \ref{lem.reguum}, $m(s)$ has a uniformly bounded second-order moment and, by Lemma~\ref{lem.estiVN}, the drift of the process ${\bf Y}^N$ is also uniformly bounded,
$$
\mathbb E\left[ \sup_{s\in [t_0, T]} {\bf d}_1^2(m^N_{{\bf Y}^N_s}, m(s)) \right] \leq C
$$
Using once more Proposition \ref{prop.quantifCV}, we get that, for a new constant $\gamma''\in (0,1)$ depending on $d$ only and a new constant $C'$ depending on the initial condition $(t_0,m_0)$,
\begin{align*}
& \mathbb E\left[ \sup_{s\in [t_0, T]} {\bf d}_1(m^N_{{\bf Y}^N_s}, m(s)) \right] \leq C'N^{-\gamma''}.
\end{align*}
\end{proof}
|
\section{Introduction}
\label{sec:1}
Let $(M,\omega)$ be a compact K\"ahler manifold of complex dimension $n$ and $\chi$ be another K\"ahler metric on $M$. In local coordinates,
\[\omega=\sqrt{-1}g_{i\bar{j}}dz^i\wedge d\bar{z}^j\ \text{and}\ \chi=\sqrt{-1}\chi_{i\bar{j}}dz^i\wedge d\bar{z}^j.\]
$\mathcal{H}_\chi$ denotes the set of all $\phi\in C^\infty(M,\mathbb{R})$ satisfying
\[\chi_\varphi:=\chi+\sqrt{-1}\partial\overline{\partial}\varphi>0.\]
The $J$-flow introduced by S. K. Donaldson \cite{do99} and X. Chen \cite{chen00} independently is a parabolic flow of
\begin{equation}\label{1.4}
\frac{\partial\varphi}{\partial t}=c-\frac{\omega\wedge\chi_\varphi^{n-1}}{\chi_\varphi^n}
\end{equation}
with $c$ is the constant given by
\[c=\frac{\int_M\omega\wedge\chi^{n-1}}{\int_M\chi^n}.\]
X. Chen \cite{sw7} proved the $J$-flow \eqref{1.4} exists for all time with any smooth initial data. It is known that the critical point to \eqref{1.4} is given by the Donaldson's equation
\begin{equation}\label{1.5}
\omega\wedge\chi^{n-1}=c\chi^n,
\end{equation}
where the K\"ahler metric $\chi$ is in the cohomology class $[\chi]$ (see \cite{do99,chen00} for details). S. K. Donaldson \cite{do99} found that if \eqref{1.5} has a solution in $[\chi]$ then
\begin{equation}\label{1.6}
[nc\chi]-[\omega]>0,
\end{equation}
and remarked a natural conjecture that the necessary condition \eqref{1.6} is also sufficient for the existence for a solution to \eqref{1.5} in $[\chi]$.
Donaldson's conjecture was proved by X. Chen \cite{sw7} (by adding a curvature assumption) and B. Weinkove \cite{wein04} on K\"ahler surfaces, while by B. Weinkove \cite{wein06} in higher dimensions. Moreover, J. Song and B. Weinkove \cite{sw08} showed that \eqref{1.5} has a smooth solution in $[\chi]$ if and only if there exists a metric $\chi'\in[\chi]$ with
\begin{equation}\label{1.7}
(nc\chi'-(n-1)\omega)\wedge\chi'^{n-2}>0,
\end{equation}
and \eqref{1.7} is also a sufficient and necessary condition for the $J$-flow \eqref{1.4} converges in $C^\infty$ to a function $\phi_\infty\in\mathcal{H}_\chi$ with $\chi_{\phi_\infty}$ satisfies \eqref{1.5}. In 2021, G. Chen \cite{gchen2021} considered the $J$-equation
\begin{equation}\label{1.17}
tr_{\omega_\varphi}\chi=c
\end{equation}
and proved that there exists a K\"ahler metric $\omega_\varphi=\omega+\sqrt{-1}\partial\bar{\partial}\varphi>0$ satisfying the $J$-equation \eqref{1.17} if and only if $(M,[\omega],[\chi])$ is uniformly $J$-stable.
On a compact Hermitian manifold, V. Tosatti and B. Weinkove \cite{tw10} obtained a uniform $C^\infty$ estimate for a smooth solution $\varphi$ to the complex Monge-Amp\'ere equation
\begin{equation}\label{1.0}
(\omega+\sqrt{-1}\partial\overline{\partial}\varphi)^n=e^F\omega^n
\end{equation}
with $\omega+\sqrt{-1}\partial\overline{\partial}\varphi>0$ and proved the Hermitian version of Calabi conjecture that every representative of the first Bott-Chern class can be
represented as the first Chern form of a Hermitian metric of the form $\omega+\sqrt{-1}\partial\overline{\partial}\varphi$. In 2011, M. Gill \cite{gill11} proved long-time existence to the parabolic complex Monge-Amp\'ere equation
\begin{equation}\label{1.8}
\frac{\partial\varphi}{\partial t}=\log\frac{(\omega+\sqrt{-1}\partial\overline{\partial}\varphi)^n}{\omega^n}-F
\end{equation}
with $\omega+\sqrt{-1}\partial\overline{\partial}\varphi>0$ and the initial condition $\varphi(\cdot,0)=0$ on a compact Hermitian manifold $(M,\omega)$ of complex dimension $n$, and provided a parabolic proof of the main theorem in
V. Tosatti and B. Weinkove \cite{tw10}. More generally, W. Sun \cite{sun15} derived uniform $C^\infty$ estimates for the normalized solution and the $C^\infty$ convergence of the parabolic complex Monge-Amp\'ere type equations
\begin{equation}\label{1.9}
\frac{\partial u}{\partial t}=\log\frac{\chi_u^n}{\chi_u^{n-\alpha}\wedge\omega^\alpha}-\log\psi\ (1\leq \alpha\leq n)
\end{equation}
with $\chi$ is a smooth positive $(1,1)$ form on a compact Hermitian manifold $(M^n,\omega)$.
The $C^2$ estimate for the complex Monge-Amp\'ere equation \eqref{1.0} on a compact K\"ahler manifold plays a connecting link role in S. T. Yau's original proof of the Calabi conjecture (see \cite{yau} for details). It is noted that the $C^2$ estimate in S. T. Yau \cite{yau} is directly obtained once the $C^0$ estimate is known and does not make use of gradient estimates. However, there has been an increasing interest for researchers to study gradient estimates on the complex Monge-Amp\'{e}re equation on compact manifolds. Z. B{\l}ocki \cite{blocki09} derived a gradient estimate for the complex Monge-Amp\'{e}re equation on a compact K\"ahler manifold directly from its $C^0$ estimate, without using the $C^2$ estimates. B. Guan and Q. Li \cite{GL09} generalized this estimate to a complex Monge-Amp\'ere equation satisfying a Dirichlet condition on a compact Hermitian manifold with smooth boundary. Moreover, X. Chen and W. He \cite{ch12} improved the gradient estimate by Z. B{\l}ocki \cite{blocki09} that the right-hand-side of \eqref{1.0} has rather weaker regularity. More recently, X. Chen and J. Cheng \cite{cc19} showed $L^p$ ($p<\infty$) and $L^\infty$ estimates for the gradient of the solutions to the complex Monge-Amp\'ere equation \eqref{1.0} in terms of the continuity of $F$ and without assuming any bounds on its derivatives. In 2021, B. Guo, D. H. Phong and F. Tong \cite{gpt21} proved a gradient estimate for the complex Monge-Amp\'ere equation \eqref{1.0} on a compact K\"ahler manifold with boundary by using the Alexandrov-Bakelman-Pucci (ABP) maximum principle.
X. Zhang \cite{xzhang10} obtained gradient estimates for geometric solutions of the complex Hessian equations
\begin{equation}\label{1.11}
\sigma_k(g^{-1}(g_{i\bar{j}}+\phi_{i\bar{j}}))=f
\end{equation}
and
\begin{equation}\label{1.12}
\sigma_k(g_{i\bar{j}}+\phi_{i\bar{j}}+\mu\phi_i\phi_{\bar{j}})=f
\end{equation}
on a compact Hermitian manifold $(M,g_{i\bar{j}})$ by maximum principle arguments, where $\sigma_k$ is the $k$-th $(1\leq k\leq n)$ elementary symmetric function and $\mu$, $f$ $(f>0)$ are smooth functions on $M$. It is clear that \eqref{1.11} and \eqref{1.12} are natural generalizations of the complex Monge-Amp\'ere equation. In 2018, R. Yuan \cite{ryuan18} derived gradient estimates from the Bernstein
method and other ordered apriori estimates on a class of second order fully nonlinear elliptic equations containing gradient terms on compact Hermitian manifolds under some assumptions. See also \cite{twwy23,sjdg20,sjdg21,sjdg23,sjdg25,sjdg35,sjdg36,sjdg38,sjdg,twwy64,ryuan22} for related work on apriori estimates for fully nonlinear equations on manifolds.
In this paper, we consider gradient estimates on smooth solutions $\varphi\in\mathcal{H}_\chi$ to the Donaldson's equation
\begin{equation}\label{1.1}
\omega\wedge\chi_\varphi^{n-1}=e^F\cdot\chi_\varphi^n
\end{equation}
and the parabolic Donladson's equation
\begin{equation}\label{1.10}
\frac{\partial \varphi}{\partial t}=\log\frac{\chi_\varphi^n}{\omega\wedge\chi_\varphi^{n-1}}+F,
\end{equation}
respectively, on an $n$-dimensional compact K\"ahler manifold $(M,\omega)$ with $\chi$ is another Hermitian metric on $M$ and $F\in C^\infty(M,\mathbb{R})$ is a given smooth function. It is noted that the Donaldson's equation \eqref{1.1} is a generalization of the complex Monge-Amp\'ere equation \eqref{1.0} while the parabolic Donladson's equation \eqref{1.10} is a special case of \eqref{1.9}. The main analytic tools we apply are ABP maximum principles. We do not use any known apriori estimates other than $C^2$ estimates.
In 2014, Y. Li \cite{yli14} proved $C^\infty$ estimate for the Donaldson's equation \eqref{1.1} when $\omega$ is a Hermitian metric and $\chi-\frac{n-1}{ne^F}\omega>0$. The first goal of this paper is to complete apriori estimates of \eqref{1.1} in the case of $\omega$ is K\"ahler. To be precise, we establish a gradient estimate from the $C^2$ estimate by Y. Li \cite{yli14}.
\begin{theorem}\label{thm1.1}
Let $(M,\omega)$ be an $n$-dimensional compact K\"ahler manifold (without boundary) and $\chi$ be another Hermitian metric on $M$. Assume that
\begin{equation}\label{1.2}
\chi-\frac{n-1}{ne^F}\omega>0.
\end{equation}
If $\varphi\in\mathcal{H}_\chi$ is a solution to the Donaldson's equation \eqref{1.1} on $M$, then there exists a uniform constant $C$ depending on $M$, $\omega$, $\chi$ and $F$, and a positive constant $\lambda$ depending only on $\omega$ and $\chi$ such that
\begin{equation}\label{1.3}
|\nabla\varphi|_\omega^2\leq Ce^{\lambda(\varphi-\inf_M\varphi)}
\end{equation}
on $M$.
\end{theorem}
Following the definition given in \cite{sun15}, we say the Hermitian metric $\chi$ satisfies the cone condition if $[\chi]\in\mathcal{C}(F)$, where
\[\mathcal{C}(F):=\{[\chi]:\ \text{there exists}\ \chi'\in[\chi]^+,\ n\chi'^{n-1}>(n-1)e^{-F}\chi'^{n-2}\wedge\omega\}\]
with $[\chi]^+:=\{\chi'\in[\chi]:\chi'>0\}$. The second goal in this paper is to prove a gradient estimate for the parabolic Donaldson's equation \eqref{1.10} by only using the $C^2$ estimate in W. Sun \cite{sun15} and the ABP maximum principle of parabolic version.
\begin{theorem}\label{thm1.2}
Let $(M,\omega)$ be an $n$-dimensional compact K\"ahler manifold (without boundary) and $\chi$ be another Hermitian metric on $M$. Assume that $[\chi]\in\mathcal{C}(F)$ and
\[e^F\chi^n\leq\omega\wedge\chi^{n-1}.\]
If $\varphi$ is a solution to the parabolic Donaldson's equation \eqref{1.10} on $M\times[0,T]$ for $0<T<\infty$, then there exists a uniform constant $\bar{C}$ depending only on the initial geometric data and $T$, and a positive constant $\lambda$ depending only on $\omega$ and $\chi$ such that
\begin{equation}\label{1.3}
|\nabla\varphi|_\omega\leq \bar{C} e^{\lambda(\varphi-\inf_{M\times[0,T]}\varphi)}
\end{equation}
on $M\times[0,T]$.
\end{theorem}
In the rest of this paper, we prove Theorem \ref{thm1.1} and Theorem \ref{thm1.2} in Section \ref{sec:2} and Section \ref{sec:3}, respectively.
\section{Gradient estimate for Donaldson's equation}
\label{sec:2}
Motivated by the work of Guo-Phong-Tong \cite{gpt21}, we estimate upper bounds for $|\nabla\varphi|_\omega$ with $\varphi\in\mathcal{H}_\varphi$ solving \eqref{1.1}. As in \cite{yli14}, we define a Hermitian metric $h$ on $M$ with its inverse matrix is locally given by
\[h^{i\bar{j}}=\chi_\varphi^{i\bar{l}}\chi_\varphi^{k\bar{j}}g_{k\bar{l}}.\]
Let $R(\omega)_{i\bar{j}k\bar{l}}$ be the curvature tensor of $\omega$. $\Delta_\omega$ and $\Delta_h$ denote the Laplacian operator associated to the K\"ahler metric $\omega$ and the Hermitian metric $h$, respectively.
First of all, we calculate a formula for $\Delta_h|\nabla\varphi|_\omega^2$.
\begin{proposition}\label{prp2.1}
The following formula holds on $M$.
\begin{eqnarray}\label{2.1}
\Delta_h|\nabla\varphi|_\omega^2&=&-2ne^F\nabla F\cdot_\omega \nabla\varphi-2tr_h\nabla\chi\cdot_\omega\nabla\varphi\notag\\
&&+2h^{i\bar{j}}g^{k\bar{q}}g^{p\bar{l}}R(\omega)_{i\bar{j}k\bar{l}}\varphi_p\varphi_{\bar{q}}\notag\\
&&+h^{i\bar{j}}g^{k\bar{l}}(\varphi_{ki}\varphi_{\bar{j}\bar{l}}+\varphi_{k\bar{j}}\varphi_{i\bar{l}}).
\end{eqnarray}
\end{proposition}
\begin{proof}
Without loss of generality, we calculate in a holomorphic normal coordinate system for the K\"ahler metric $\omega$ at an arbitrary point $x$. By direct computation, we have
\begin{eqnarray}\label{2.2}
\Delta_h|\nabla\varphi|_\omega^2&=&h^{i\bar{j}}(g^{k\bar{l}}\varphi_k\varphi_{\bar{l}})_{\bar{j}i}\notag\\
&=&h^{i\bar{j}}(g^{k\bar{l}},_{\bar{j}}\varphi_k\varphi_{\bar{l}}+g^{k\bar{l}}\varphi_{k\bar{j}}\varphi_{\bar{l}}+g^{k\bar{l}}\varphi_k\varphi_{\bar{l}\bar{j}})_i\notag\\
&=&h^{i\bar{j}}(g^{k\bar{l}},_{\bar{j}i}\varphi_k\varphi_{\bar{l}}+\varphi_{k\bar{j}i}\varphi_{\bar{k}}+\varphi_{k\bar{j}}\varphi_{\bar{k}i}\notag\\
&&+\varphi_{ki}\varphi_{\bar{k}\bar{j}}+\varphi_k\varphi_{\bar{k}\bar{j}i})\notag\\
\end{eqnarray}
At such a point, it is clear that
\begin{equation}\label{2.3}
g^{k\bar{l}},_{\bar{j}i}=R(\omega)_{i\bar{j}k\bar{l}},
\end{equation}
and
\begin{equation}\label{2.4}
\varphi_{\bar{k}\bar{j}i}=\varphi_{\bar{k}i\bar{j}}+R(\omega)_{i\bar{j}l\bar{k}}\varphi_{\bar{l}}=\varphi_{i\bar{j}\bar{k}}+R(\omega)_{i\bar{j}l\bar{k}}\varphi_{\bar{l}}.
\end{equation}
Applying \eqref{2.3} and \eqref{2.4} to \eqref{2.2} yields
\begin{eqnarray}\label{2.5}
\Delta_h|\nabla\varphi|_\omega^2&=&2h^{i\bar{j}}R(\omega)_{i\bar{j}k\bar{l}}\varphi_{\bar{k}}\varphi_l+h^{i\bar{j}}(\varphi_{i\bar{j}k}\varphi_{\bar{k}}+\varphi_k\varphi_{i\bar{j}\bar{k}})\notag\\
&&+h^{i\bar{j}}(\varphi_{k\bar{j}}\varphi_{\bar{k}i}+\varphi_{ki}\varphi_{\bar{k}\bar{j}}).
\end{eqnarray}
From \eqref{1.1}, we know that
\begin{equation}\label{2.5x}
ne^F=n\frac{\omega\wedge\chi_\varphi^{n-1}}{\chi_\varphi^n}=tr_{\chi_\varphi}\omega:=\chi_\varphi^{i\bar{j}}g_{i\bar{j}}.
\end{equation}
Hence,
\[ne^F\cdot F_k=-h^{i\bar{j}}\varphi_{i\bar{j}k}-h^{i\bar{j}}\chi_{i\bar{j},k}\ \text{and}\ ne^F\cdot F_{\bar{k}}=-h^{i\bar{j}}\varphi_{i\bar{j}\bar{k}}-h^{i\bar{j}}\chi_{i\bar{j},\bar{k}}.\]
Moreover, we obtain
\begin{eqnarray}\label{2.6}
h^{i\bar{j}}(\varphi_{i\bar{j}k}\varphi_{\bar{k}}+\varphi_k\varphi_{i\bar{j}\bar{k}})&=&-ne^F(F_k\varphi_{\bar{k}}+F_{\bar{k}}\varphi_k)-h^{i\bar{j}}(\chi_{i\bar{j},k}\varphi_{\bar{k}}+\chi_{i\bar{j},\bar{k}}\varphi_k)\notag\\
&=&-2ne^F\nabla F\cdot_\omega\nabla\varphi-2tr_h\nabla\chi\cdot_\omega\nabla\varphi.
\end{eqnarray}
\eqref{2.1} follows by applying \eqref{2.6} to \eqref{2.5} since it is independent of the choice of holomorphic normal coordinate systems.
\end{proof}
Let $-K$ be a uniform lower bound of the bisectional curvature of $\omega$ and $\lambda$ be a positive constant (depending only on $\omega$ and $\chi$) to be determined later.
\begin{lemma}\label{lem2.1}
Define $G:=e^{-\lambda(\varphi-\inf_M\varphi)}|\nabla\varphi|_\omega^2$, then there exists a uniform constant $C$ that depends only on $M$, $\omega$, $\chi$ and $F$ so that
\begin{equation}\label{2.7}
\Delta_hG\geq-CG^\frac{1}{2}+(\lambda tr_h\chi-2Ktr_h\omega)G-3\lambda ne^FG.
\end{equation}
\end{lemma}
\begin{proof}
Note that
\begin{equation}\label{2.9}
h^{i\bar{j}}(\chi_\varphi)_{i\bar{j}}=\chi_\varphi^{i\bar{l}}\chi_\varphi^{k\bar{j}}g_{k\bar{l}}(\chi_\varphi)_{i\bar{j}}=\chi_\varphi^{k\bar{l}} g_{k\bar{l}}=ne^F,\notag
\end{equation}
where we used \eqref{2.5x}. Then we have
\begin{eqnarray}\label{2.10}
-h^{i\bar{j}}\varphi_{i\bar{j}}&=&-h^{i\bar{j}}(\chi_\varphi)_{i\bar{j}}+h^{i\bar{j}}\chi_{i\bar{j}}\notag\\
&=&-ne^F+tr_h\chi.
\end{eqnarray}
Moreover,
\begin{eqnarray}\label{2.8}
\Delta_he^{-\lambda(\varphi-\inf_M\varphi)}&=&h^{i\bar{j}}(e^{-\lambda(\varphi-\inf_M\varphi)})_{\bar{j}i}\notag\\
&=&h^{i\bar{j}}(-\lambda\varphi_{\bar{j}}e^{-\lambda(\varphi-\inf_M\varphi)})_{i}\notag\\
&=&h^{i\bar{j}}(-\lambda\varphi_{i\bar{j}}+\lambda^2\varphi_i\varphi_{\bar{j}})e^{-\lambda(\varphi-\inf_M\varphi)}\notag\\
&=&(-\lambda ne^F+\lambda tr_h\chi+\lambda^2|\nabla\varphi|_h^2)e^{-\lambda(\varphi-\inf_M\varphi)}
\end{eqnarray}
By direct computation, we obtain
\begin{eqnarray}\label{2.10}
\Delta_hG&=&\Delta_h(e^{-\lambda(\varphi-\inf_M\varphi)}|\nabla\varphi|_\omega^2)\notag\\
&=&e^{-\lambda(\varphi-\inf_M\varphi)}\Delta_h|\nabla\varphi|_\omega^2+(\Delta_he^{-\lambda(\varphi-\inf_M\varphi)})|\nabla\varphi|_\omega^2\notag\\
&&+2\nabla e^{-\lambda(\varphi-\inf_M\varphi)}\cdot_h\nabla|\nabla\varphi|_\omega^2\notag\\
&=&e^{-\lambda(\varphi-\inf_M\varphi)}[-2ne^F\nabla F\cdot_\omega \nabla\varphi-2tr_h\nabla\chi\cdot_\omega\nabla\varphi\notag\\
&&+2h^{i\bar{j}}g^{k\bar{q}}g^{p\bar{l}}R(\omega)_{i\bar{j}k\bar{l}}\varphi_p\varphi_{\bar{q}}+h^{i\bar{j}}g^{k\bar{l}}(\varphi_{ki}\varphi_{\bar{j}\bar{l}}+\varphi_{k\bar{j}}\varphi_{i\bar{l}})]\notag\\
&&+G(-\lambda ne^F+\lambda tr_h\chi+\lambda^2|\nabla\varphi|_h^2)\notag\\
&&-2\lambda e^{-\lambda(\varphi-\inf_M\varphi)}\nabla\varphi\cdot_h\nabla|\nabla\varphi|_\omega^2.
\end{eqnarray}
Next, we deal with the right-hand-side of \eqref{2.10} in a holomorphic normal coordinate system under K\"ahler metric $\omega$ so that
\[g_{i\bar{j}}=\delta_{ij},\ \chi_{i\bar{j}}=\tau_i\delta_{ij},\ \text{and}\ (\chi_\varphi)_{i\bar{j}}=\rho_i\delta_{ij}\]
for some $\tau_1$,...,$\tau_n>0$ and $\rho_1$,...,$\rho_n>0$. Hence, $h^{i\bar{j}}$ can be diagonalized that
\begin{equation}\label{2.x}
h^{i\bar{i}}=(\chi_\varphi^{i\bar{i}})^2=\frac{1}{\rho_i^2}\leq(\sum_{i=1}^n\frac{1}{\rho_i})^2=(\chi^{i\bar{j}}_\varphi g_{i\bar{j}})^2=n^2e^{2F},
\end{equation}
where we used \eqref{2.5x}.
It follows from Cauchy's inequality and \eqref{2.x} that
\begin{eqnarray}\label{2.20}
&&e^{-\lambda(\varphi-\inf_M\varphi)}(-2ne^F\nabla F\cdot_\omega \nabla\varphi-2tr_h\nabla\chi\cdot_\omega\nabla\varphi)\notag\\
&\geq&(-2ne^F|\nabla F|_\omega-2n^2e^{2F}|\nabla\chi|_\omega)\cdot e^{-\frac{\lambda}{2}(\varphi-\inf_M\varphi)}|\nabla\varphi|_\omega\notag\\
&\geq&-CG^\frac{1}{2}
\end{eqnarray}
for a uniform constant depends only on $M$, $\omega$, $\chi$ and $F$.
\begin{eqnarray*}
-2\lambda\nabla\varphi\cdot_h\nabla|\nabla\varphi|_\omega^2&=&-2\lambda Re(h^{i\bar{j}}\varphi_i(\varphi_k\varphi_{\bar{k}})_{\bar{j}})\notag\\
&=&-2\lambda Re(h^{i\bar{j}}\varphi_i\varphi_{k\bar{j}}\varphi_{\bar{k}}+h^{i\bar{j}}\varphi_i\varphi_k\varphi_{\bar{k}\bar{j}})\notag\\
&\geq&-2\lambda h^{i\bar{j}}\varphi_i\varphi_{\bar{k}}(\chi_\varphi)_{k\bar{j}}+2\lambda h^{i\bar{j}}\varphi_i\varphi_{\bar{k}}\chi_{k\bar{j}}\notag\\
&&-h^{i\bar{j}}\varphi_{ki}\varphi_{\bar{k}\bar{j}}-\lambda^2|\nabla\varphi|_h^2|\nabla\varphi|_\omega^2,
\end{eqnarray*}
where we used Cauchy's inequality.
Note that
\begin{eqnarray*}
-h^{i\bar{j}}\varphi_i\varphi_{\bar{k}}(\chi_\varphi)_{k\bar{j}}&=&-\chi_\varphi^{i\bar{p}}\chi_\varphi^{p\bar{j}}(\chi_\varphi)_{k\bar{j}}\varphi_i\varphi_{\bar{k}}\notag\\
&=&-\chi_\varphi^{i\bar{k}}\varphi_i\varphi_{\bar{k}}\notag\\
&\geq&-(tr_{\chi_\varphi}\omega)|\nabla\varphi|_\omega^2\notag\\
&=&-ne^F|\nabla\varphi|_\omega^2,
\end{eqnarray*}
where we used \eqref{2.5x}, and
\[h^{i\bar{j}}\varphi_i\varphi_{\bar{k}}\chi_{k\bar{j}}=\sum_i\frac{\tau_i}{\rho_i^2}\varphi_i\varphi_{\bar{i}}\geq0,\]
we have
\begin{eqnarray}\label{2.11}
&&-2\lambda\nabla\varphi\cdot_h\nabla|\nabla\varphi|_\omega^2\notag\\
&\geq&-2\lambda ne^F|\nabla\varphi|_\omega^2-h^{i\bar{j}}\varphi_{ki}\varphi_{\bar{k}\bar{j}}-\lambda^2|\nabla\varphi|_h^2|\nabla\varphi|_\omega^2.
\end{eqnarray}
Plugging \eqref{2.20} and \eqref{2.11} into \eqref{2.10}, we have
\begin{eqnarray*}
\Delta_hG&\geq&e^{-\lambda(\varphi-\inf_M\varphi)}(-CG^\frac{1}{2}+2h^{i\bar{j}}g^{k\bar{q}}g^{p\bar{l}}R(\omega)_{i\bar{j}k\bar{l}}\varphi_p\varphi_{\bar{q}}\notag\\
&&+h^{i\bar{j}}g^{k\bar{l}}\varphi_{k\bar{j}}\varphi_{i\bar{l}})+G(-3\lambda ne^F+\lambda tr_h\chi)\notag\\
&\geq&-CG^\frac{1}{2}+(\lambda tr_h\chi-2Ktr_h\omega)G-3\lambda ne^FG.
\end{eqnarray*}
\end{proof}
The ABP maximum principle (see e.g. Lemma 9.3 in \cite{elliptic} of Lemma 5.2 in \cite{chencheng1}) is essential to derive a gradient estimate for \eqref{1.1}. For reader's convenience, we present this result below.
\begin{lemma}[ABP maximum principle]
Let $\Omega\subset\mathbb{R}^d$ be a bounded domain. Suppose $u\in C^2(\Omega)\cap C(\overline{\Omega})$. Denote $U=\sup_\Omega u-\sup_{\partial\Omega}u$. Define
\begin{eqnarray*}
\Gamma^-(u,\Omega)&:=&\{x\in \Omega:u(y)\leq u(x)+\nabla u(x)\cdot(y-x),\notag\\
&&\text{for\ any}\ y\in\Omega\ \text{and}\ |\nabla u(x)|\leq\frac{U}{3diam(\Omega)}\}.
\end{eqnarray*}
Then for some dimensional constant $C_d>0$,
\[U\leq C_d\big(\int_{\Gamma^-(u,\Omega)}det(-D^2u)dx\big)^\frac{1}{d}.\]
In particular, suppose $u$ satisfies $a_{ij}\partial_i\partial_ju\geq f$. Here $a_{ij}$ satisfies the ellipticity condition $a_{ij}\xi_i\xi_j\geq0$. Define $D^*= (deta_{ij})^\frac{1}{d}$. Then there exists another dimensional constant $C'_d>0$ so that
\[U\leq C'_ddiam(\Omega)\|\frac{f^-}{D^*}\|_{L^d(\Omega)}.\]
\end{lemma}
Now we are ready to finish the proof of Theorem \ref{thm1.1}.
\\\\\textbf{Proof of Theorem \ref{thm1.1}:} We apply a trick that Chen-Cheng \cite{chencheng1} and Guo-Phong-Tong \cite{gpt21} used. Let $r$ be chosen so that for any $p\in M$, the geodesic ball $B_\omega(p,r):=\{x\in M|d_\omega(x,p)<r\}$ is contained in a single coordinate neighborhood and under this coordinate, we may identify $B_\omega(p,r)$ with a bounded domain in $\mathbb{R}^{2n}$.
Let $0<\theta\leq\frac{1}{2}$ be a given constant to be determined later, and $\eta:M\rightarrow\mathbb{R}_+$ be a cut-off function such that
\[\eta(x)=
\begin{cases}
1& x\in B_\omega(p,r/2),\\
1-\theta& x\in B_\omega(p,r)\backslash B_\omega(p,3r/4),
\end{cases}
\]
and $\frac{1}{2}\leq1-\theta\leq\eta\leq1$ in the annulus between. Moreover,
\begin{equation}\label{2.12}
|\nabla\eta|_\omega^2\leq\frac{c_0\theta^2}{r^2}\ \text{and}\ |\nabla^2\eta|_\omega\leq\frac{c_0\theta}{r^2}
\end{equation}
for some fixed constant $c_0$ that depends only on $\omega$.
By direct computation, we have
\begin{eqnarray}\label{2.13}
\Delta_h(\eta G^2)&=&\eta\Delta_h G^2+4G\nabla\eta\cdot_h\nabla G+G^2\Delta_h\eta\notag\\
&\geq&\eta(2G\Delta_hG+2|\nabla G|_h^2)-|\nabla G|_h^2\notag\\
&&-4G^2|\nabla\eta|_h^2-G^2|\nabla^2\eta|_\omega\cdot tr_h\omega\notag\\
&\geq&-C\eta G^\frac{3}{2}+2(\lambda tr_h\chi-2Ktr_h\omega)\eta G^2\notag\\
&&-6\lambda ne^F\eta G^2+(2\eta-1) |\nabla G|_h^2\notag\\
&&-4G^2|\nabla\eta|_\omega^2\cdot tr_h\omega-G^2|\nabla^2\eta|_\omega\cdot tr_h\omega\notag\\
&\geq&-C\eta G^{\frac{3}{2}}-6\lambda ne^F\eta G^2\notag\\
&&+G^2tr_h[2\eta(\lambda\chi-2K\omega)-\frac{c_0\theta+4c_0\theta^2}{r^2}\omega],
\end{eqnarray}
where we used Cauchy's inequality in the first inequality, Lemma \ref{lem2.1} in the second and \eqref{2.12} in the last.
Choose $\theta=\min\{\frac{1}{2},\frac{r^2}{2c_0},\frac{r}{4\sqrt{c_0}}\}$. There exists a constant $\lambda$ depending only on $\omega$ and $\chi$ with
\[2\eta(\lambda\chi-2K\omega)-\frac{c_0\theta+4c_0\theta^2}{r^2}\omega\geq0.\]
Therefore, \eqref{2.13} reduces to
\begin{equation}\label{2.14}
\Delta_h(\eta G^2)\geq-C\eta G^\frac{3}{2}-C\eta G^2.
\end{equation}
Before we applying the ABP maximum principle to \eqref{2.14}, we only need to estimate $det(h^{i\bar{j}})^{-1}$.
Under the holomorphic coordinates mentioned in Lemma \ref{lem2.1}, we have $h^{i\bar{i}}=\frac{1}{\rho_i^2}$ for $i=1,2,...,n$. Moreover,
\begin{eqnarray}\label{2.15}
det(h^{i\bar{j}})^{-1}&=&\big(\prod_{i=1}^n\frac{1}{\rho_i^2}\big)^{-1}=\prod_{i=1}^n\rho_i^2\leq\big(\frac{1}{n}(\sum_{i=1}^n\rho_i)^2\big)^n\notag\\
&=&\frac{1}{n^n}(tr_\omega\chi_\varphi)^{2n}.
\end{eqnarray}
Under the condition of \eqref{1.2}, the main theorem of \cite{yli14} shows that
\[tr_\omega\chi_\varphi\leq C\]
for some uniform constant $C$ depending only on $M$, $\omega$, $\chi$ and $F$.
It follows that
\begin{equation}\label{2.16}
det(h^{i\bar{j}})^{-1}\leq C,
\end{equation}
and
\begin{equation}\label{2.18}
\Delta_\omega\varphi=tr_\omega\chi_\varphi-tr_\omega\chi\leq C.
\end{equation}
Integration by parts gives
\begin{equation}\label{2.17}
\int_MG\omega^n=\int_Me^{-\lambda(\varphi-\inf_M\varphi)}|\nabla\varphi|^2_\omega\omega^n=\frac{1}{\lambda}\int_Me^{-\lambda(\varphi-\inf_M\varphi)}\Delta_\omega\varphi\omega^n\leq C,
\end{equation}
where we used \eqref{2.18} and the fact of $e^{-\lambda(\varphi-\inf_M\varphi)}\leq 1$.
For simplicity, we denote $V:=\sup_{B_\omega(p,r)}(\eta G^2)$. Applying the ABP maximum principle to $\eta G^2$ on $B_\omega(p,r)$, we obtain that
\begin{eqnarray}\label{2.19}
&&V\notag\\
&\leq&\sup_{\partial B_\omega(p,r)}(\eta G^2)+C\cdot r\big(\int_{B_\omega(p,r)}[\eta G^\frac{3}{2}+\eta G^2]^{2n}\omega^n\big)^{\frac{1}{2n}}\notag\\
&\leq&(1-\theta)V+C\cdot \big[ V^\frac{3}{4}+V^{1-\frac{1}{2n}}(\int_{B_\omega(p,r)}G\omega^n)^\frac{1}{2n}\big]\notag\\
&\leq&(1-\theta)V+C\cdot (V^\frac{3}{4}+V^{1-\frac{1}{2n}}),
\end{eqnarray}
where we used \eqref{2.16} in the first inequality and \eqref{2.17} in the last.
We can conclude from \eqref{2.19} that
\[G^2(p)\leq V\leq C.\]
The arbitrary of $p$ implies
\[|\nabla\varphi|_\omega^2\leq C e^{\lambda(\varphi-\inf_M\varphi)}\]
on $M$.$\hfill\Box$
\section{Gradient estimate for parabolic Donaldson's equation}
\label{sec:3}
We give a proof of Theorem \ref{thm1.2} in this section. Using the notations in Section \ref{sec:2}, we have a similar inequality as in Lemma \ref{lem2.1} for $$H\equiv H(x,t):=\exp\{-\lambda(\varphi(x,t)-\inf_{M\times[0,T]}\varphi)\}|\nabla\varphi|_\omega^2(x,t),$$
where $\lambda$ is a positive constant (depending only on $\omega$ and $\chi$) to be determined later.
\begin{lemma}\label{lem3.1}
There exists a uniform constant $C$ depending only on $M$, $\omega$, $\chi$ and $F$ so that
\begin{eqnarray}\label{3.1}
&&(\Delta_h-n\exp\{F-\frac{\partial\varphi}{\partial t}\}\frac{\partial}{\partial t})H\notag\\
&\geq&-(1+\exp\{F-\frac{\partial\varphi}{\partial t}\})C H^\frac{1}{2}+(\lambda tr_h\chi-2Ktr_h\omega)H\notag\\
&&-3\lambda n\exp\{F-\frac{\partial\varphi}{\partial t}\}H
\end{eqnarray}
on $M\times(0,T]$.
\end{lemma}
\begin{proof}
Calculating in a holomorphic normal coordinate system for the K\"ahler metric $\omega$ at an arbitrary point $x$ as \eqref{2.2} to \eqref{2.5}, we can obtain that
\begin{eqnarray}\label{3.2}
&&(\Delta_h-n\exp\{F-\frac{\partial\varphi}{\partial t}\}\frac{\partial}{\partial t})|\nabla\varphi|_\omega^2\notag\\
&=&2h^{i\bar{j}}R(\omega)_{i\bar{j}k\bar{l}}\varphi_{\bar{k}}\varphi_l+h^{i\bar{j}}(\varphi_{k\bar{j}}\varphi_{\bar{k}i}+\varphi_{ki}\varphi_{\bar{k}\bar{j}})\notag\\
&&+h^{i\bar{j}}(\varphi_{i\bar{j}k}\varphi_{\bar{k}}+\varphi_k\varphi_{i\bar{j}\bar{k}})\notag\\
&&-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}\nabla\frac{\partial\varphi}{\partial t}\cdot_\omega\nabla\varphi.
\end{eqnarray}
It follows from \eqref{1.10} that
\begin{equation}\label{3.3}
n\exp\{F-\frac{\partial\varphi}{\partial t}\}=n\frac{\omega\wedge\chi^{n-1}_\varphi}{\chi^n_\varphi}=tr_{\chi_\varphi}\omega.
\end{equation}
Hence,
\[n\exp\{F-\frac{\partial\varphi}{\partial t}\}\cdot (F_k-\frac{\partial\varphi_k}{\partial t})=-h^{i\bar{j}}\varphi_{i\bar{j}k}-h^{i\bar{j}}\chi_{i\bar{j},k}\]
and
\[n\exp\{F-\frac{\partial\varphi}{\partial t}\}\cdot (F_{\bar{k}}-\frac{\partial\varphi_{\bar{k}}}{\partial t})=-h^{i\bar{j}}\varphi_{i\bar{j}{\bar{k}}}-h^{i\bar{j}}\chi_{i\bar{j},\bar{k}}.\]
Moreover, we can obtain
\begin{eqnarray}\label{3.4}
&&h^{i\bar{j}}(\varphi_{i\bar{j}k}\varphi_{\bar{k}}+\varphi_k\varphi_{i\bar{j}\bar{k}})-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}\nabla\frac{\partial\varphi}{\partial t}\cdot_\omega\nabla\varphi\notag\\
&=&-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}\nabla F\cdot_\omega\nabla\varphi-2tr_h\nabla\chi\cdot_\omega\nabla\varphi.
\end{eqnarray}
Applying \eqref{3.4} to \eqref{3.2}, we have
\begin{eqnarray}\label{3.5}
&&(\Delta_h-n\exp\{F-\frac{\partial\varphi}{\partial t}\}\frac{\partial}{\partial t})|\nabla\varphi|_\omega^2\notag\\
&=&2h^{i\bar{j}}R(\omega)_{i\bar{j}k\bar{l}}\varphi_{\bar{k}}\varphi_l+h^{i\bar{j}}(\varphi_{k\bar{j}}\varphi_{\bar{k}i}+\varphi_{ki}\varphi_{\bar{k}\bar{j}})\notag\\
&&-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}\nabla F\cdot_\omega\nabla\varphi-2tr_h\nabla\chi\cdot_\omega\nabla\varphi.
\end{eqnarray}
The rest computation is follows almost line by line from the proof of Lemma \ref{lem2.1} with minor modifications. Therefore, we can conclude \eqref{3.1}.
\end{proof}
We present the following parabolic version of ABP maximum principle (see \cite{lieb}, Theorem 7.1 or \cite{cc21}, Lemma 3.8) that serves as the main analytic tool in the proof of Theorem \ref{thm1.2}.
\begin{lemma}[Parabolic ABP maximum principle]
Let $\Omega\subset\mathbb{R}^d$ be a bounded domain and $u$ be a smooth function on $\Omega\times[0,T]$, such that
\[b^{ij}\partial_{i}\partial_{j}u-\frac{\partial u}{\partial t}\geq f.\]
Here $b^{ij}$ satisfies the ellipticity condition $b^{ij}\xi_i\xi_j\geq0$. Define $D= (detb^{ij})^\frac{1}{d+1}$. Then there exists a dimensional constant $C_d>0$ so that
\[\sup_{\Omega\times[0,T]}u\leq \sup_{\partial_P(\Omega\times[0,T])}u+C_d(diam(\Omega))^\frac{n}{n+1}\|\frac{f^-}{D}\|_{L^{d+1}(E)}.\]
In the above, $\partial_P(\Omega\times[0,T])$ is the parabolic boundary given by $(\Omega\times\{0\})\cup(\partial\Omega\times[0,T])$, $f^-=\max(-f,0)$ and $E$ is the set of $(x,t)$ on which $\partial_tu\geq0$ and $D_x^2u\leq0$.
\end{lemma}
Similar as the arguments as in the elliptic case, we give a proof of Theorem \ref{thm1.2} in the following.
\\\\\textbf{Proof of Theorem \ref{thm1.2}:} We apply a trick that X. Chen and J. Cheng \cite{cc21} used. Let $(q,s)$ be a maximum point of $H$ on $M\times[0,T]$ and identify the geodesic ball $B_\omega(q,k):=\{x\in M|d_\omega(x,q)<k\}$ with an open domain in $\mathbb{R}^{2n}$ under local coordinates. Let $\beta$ $(0<\beta\leq\frac{1}{2})$ be a given constant to be determined later, and $\xi\in[1-\beta,1]$ be a cut-off function on $M$ satisfying
\[\xi(q)=1,\ \xi=1-\beta\ \text{on}\ M-B_\omega(q,k),\]
and
\[\ |\nabla\xi|^2_\omega\leq\frac{c_1\beta^2}{k^2},\ |\nabla^2\xi|_\omega\leq\frac{c_1\beta}{k^2}\]
for some fixed constant $c_1$ that depends only on $\omega$.
By direct computation, we have
\begin{eqnarray}\label{3.6}
&&(\Delta_h-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}\frac{\partial}{\partial t})(\xi H^2)\notag\\
&=&\xi(\Delta_h-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}) H^2+4H\nabla\xi\cdot_h\nabla H+H^2\Delta_h\xi\notag\\
&\geq&\xi(2H(\Delta_h-2n\exp\{F-\frac{\partial\varphi}{\partial t}\})H+2|\nabla H|_h^2)-|\nabla H|_h^2\notag\\
&&-4H^2|\nabla\xi|_h^2-H^2|\nabla^2\xi|_\omega\cdot tr_h\omega\notag\\
&\geq&-(1+\exp\{F-\frac{\partial\varphi}{\partial t}\})C \xi H^\frac{3}{2}+2(\lambda tr_h\chi-2Ktr_h\omega)\xi H^2\notag\\
&&-6\lambda n\exp\{F-\frac{\partial\varphi}{\partial t}\}\xi H^2+(2\xi-1) |\nabla H|_h^2\notag\\
&&-4H^2|\nabla\xi|_\omega^2\cdot tr_h\omega-H^2|\nabla^2\xi|_\omega\cdot tr_h\omega\notag\\
&\geq&-(1+\exp\{F-\frac{\partial\varphi}{\partial t}\})C \xi H^{\frac{3}{2}}-6\lambda n\exp\{F-\frac{\partial\varphi}{\partial t}\}\xi H^2\notag\\
&&+H^2tr_h[2\xi(\lambda\chi-K\omega)-\frac{4c_1\beta^2+c_1\beta}{k^2}\omega],
\end{eqnarray}
where we used Cauchy's inequality in the first and second inequalities and Lemma \ref{lem3.1} in the second inequality.
Let $\beta=\min\{\frac{1}{2},\frac{k^2}{2c_1},\frac{k}{4\sqrt{c_1}}\}$. We can choose a constant $\lambda$ depending only on $\omega$ and $\chi$ with
\[2\xi(\lambda\chi-K\omega)-\frac{4c_1\beta^2+c_1\beta}{k^2}\omega\geq0.\]
Therefore, \eqref{3.6} reduces to
\begin{eqnarray*}
&&(\Delta_h-2n\exp\{F-\frac{\partial\varphi}{\partial t}\}\frac{\partial}{\partial t})(\xi H^2)\notag\\
&\geq&-(1+\exp\{F-\frac{\partial\varphi}{\partial t}\})C\xi H^{\frac{3}{2}}-6\lambda n\exp\{F-\frac{\partial\varphi}{\partial t}\}\xi H^2,
\end{eqnarray*}
i.e.,
\begin{eqnarray*}
&&(\frac{1}{2n}\exp\{F-\frac{\partial\varphi}{\partial t}\}h^{i\bar{j}}\partial_i\partial_{\bar{j}}-\frac{\partial}{\partial t})(\xi H^2)\notag\\
&\geq&-(\exp\{\frac{\partial\varphi}{\partial t}-F\}+1)C\xi H^{\frac{3}{2}}-3\lambda \xi H^2.
\end{eqnarray*}
Eqt. (27) in W. Sun \cite{sun15} shows that
\begin{equation}\label{3.9}
|\frac{\partial \varphi}{\partial t}|\leq\sup_{M\times\{0\}}|\frac{\partial \varphi}{\partial t}|
\end{equation}
on $M\times[0,T]$. Hence,
\begin{eqnarray}\label{3.7}
&&(\frac{1}{2n}\exp\{F-\frac{\partial\varphi}{\partial t}\}h^{i\bar{j}}\partial_i\partial_{\bar{j}}-\frac{\partial}{\partial t})(\xi H^2)\notag\\
&\geq&-(\exp\{\sup_{M\times\{0\}}|\frac{\partial \varphi}{\partial t}|\}+1)C\xi H^{\frac{3}{2}}-3\lambda \xi H^2.
\end{eqnarray}
Consider in a holomorphic normal coordinate system under K\"ahler metric $\omega$ so that
\[g_{i\bar{j}}=\delta_{ij}\ \text{and}\ (\chi_\varphi)_{i\bar{j}}(\cdot,t)=\rho_i\delta_{ij}(\cdot,t)\]
for any fixed $t\in(0,T]$. Similar as \eqref{2.15}, we have
\[det(h^{i\bar{j}})^{-1}\leq\frac{1}{n^n}(tr_\omega\chi_\varphi)^{2n}=\frac{1}{n^n}(tr_\omega\chi+\Delta_\omega\varphi)^{2n}.\]
From Theorem \ref{thm1.1} and Theorem \ref{thm1.2} in W. Sun \cite{sun15}, we know that
\begin{equation}\label{3.8}
\Delta_\omega \varphi\leq \bar{C}
\end{equation}
on $M\times[0,T]$ for some uniform constant $\bar{C}$ depending only on the initial geometric data and $T$. Therefore,
\begin{equation}\label{3.8}
det(h^{i\bar{j}})^{-1}\leq \bar{C}.
\end{equation}
For any fixed $t\in(0,T]$, integration by parts gives
\begin{eqnarray}\label{3.11}
\int_{M\times[0,T]}G\omega^ndt&=&\int_0^T\int_Me^{-\lambda(\varphi(x,t)-\inf_{M\times[0,T]}\varphi)}|\nabla\varphi|^2_\omega(x,t)\omega^ndt\notag\\
&=&\int_0^T\big(\frac{1}{\lambda}\int_Me^{-\lambda(\varphi(x,t)-\inf_{M\times[0,T]}\varphi)}\Delta_\omega\varphi(x,t)\omega^n\big)dt\notag\\
&\leq& \bar{C}.
\end{eqnarray}
Then we finish the proof by applying the parabolic ABP maximum principle to $\xi H^2$ on $B_\omega(q,k)\times[0,T]$. Note that $H^2(q,s)=\sup_{B_\omega(q,k)\times[0,T]}(\xi H^2)$, we have
\begin{eqnarray}\label{3.12}
H^2(q,s)&\leq&\sup_{\partial_P(B_\omega(q,k)\times[0,T])}(\xi H^2)\notag\\
&&+\bar{C}\cdot k^\frac{2n}{2n+1}\big(\int_{B_\omega(q,k)\times[0,T]}(\xi H^{\frac{3}{2}}+ \xi H^2)^{2n+1}\omega^ndt\big)^{\frac{1}{2n+1}}\notag\\
&\leq&(1-\beta)H^2(q,k)+\bar{C}\cdot \big[ H^\frac{3}{2}(q,s)\notag\\
&&+H^{2-\frac{1}{2n+1}}(q,s)(\int_{B_\omega(q,k)\times[0,T]}H\omega^ndt)^\frac{1}{2n}\big]\notag\\
&\leq&(1-\beta)H^2(q,s)+\bar{C}\cdot (H^\frac{3}{2}(q,s)+H^{2-\frac{1}{2n+1}}(q,s)),
\end{eqnarray}
where we used \eqref{3.8} in the first inequality and \eqref{3.11} in the last.
We can conclude from \eqref{3.12} that
\[(\sup_{M\times[0,T]}H)^2=H^2(q,s)\leq \bar{C}.\]
Moreover,
\[|\nabla\varphi|_\omega^2\leq \bar{C}e^{\lambda(\varphi-\inf_{M\times[0,T]}\varphi)}\]
on $M\times[0,T]$.$\hfill\Box$
\section*{Acknowledgements}
The author would like to thank Prof. Xiaokui Yang for helpful discussions.
|
\section{Introduction}
\label{sec:intro}
\blfootnote{*This work was conducted during an internship at Meta Reality Labs Research.}
3D human digitization has drawn significant attention in recent years, with a wide range of applications such as photo editing, video games and immersive technologies.
To obtain photo-realistic renders of free-viewpoint videos, existing approaches~\cite{collet2015high,shysheya2019textured} require complicated equipment with expensive synchronized cameras, which makes them difficult to be applied to realistic scenarios.
To date, modeling detailed appearance of dynamic clothed humans such as cloth wrinkles and facial details such as eyes from videos remains a challenging problem.
To represent static scenes, neural radiance field-based models (NeRF)~\cite{mildenhall2020nerf} learn an implicit representation using neural networks, which has enabled photo-realistic rendering of shape and appearance from images.
With dense multi-view observations as input, NeRF encodes density and color as a function of 3D coordinates and viewing directions by the MLPs along with a differentiable renderer to synthesize novel views.
While it shows unprecedented visual quality on static scenes, applying it to high quality free-viewpoint rendering of humans in dynamic videos remains a challenging task.
To generalize NeRF from static scenes to dynamic videos, D-NeRF~\cite{pumarola2020d} encodes a time step $t$ to differentiate motions across frames and converts scenes from the observation space to a shared canonical space in order to model the neural radiance field.
As such, they can handle dynamic scenes to some extent but the poses remain uncontrollable by users.
Furthermore, some approaches~\cite{peng2020neural,yariv2020multiview} introduce human pose as an additional input to serve as a geometric guidance for different frames. However, they either cannot generalize to novel poses or need more than one input view.
To overcome these limitations, we propose a novel approach by learning implicit radiance fields based on pose and appearance representations for high fidelity novel view and pose synthesis.
Motivated by~\cite{peng2020neural}, we leverage the human pose extracted from the parametric body model as a geometric prior to model motion information across frames. Shared latent codes anchored to the human poses are optimized, which can integrate information across frames.
However, a model which only formulates latent codes in a shared space will not generalize well to unseen poses without test-time optimization of the latent codes.
To generalize the model to unseen poses, we propose to model the appearance information with the assist of the single-view RGBD image.
Our model learns the appearance code anchored to incomplete point clouds in the 3D space.
Point clouds can be obtained by using single-view depth information to lift the RGB image to the 3D space, which provides partial visible parts of the human body.
The learned implicit representation enables reasoning of the unknown regions and complements the missing details on the human body.
To further leverage the temporal information from multiple frames, we introduce a temporal transformer that aggregates the trackable information.
To achieve this, we use the parametric body model to track points from the query frame to the key frames.
Then, based on the learned implicit representation, we extract the pose code across frames and forward it into the temporal transformer for feature aggregation.
Our method is extensively evaluated against state-of-the-art techniques on several sequences of humans in motion and exhibits significantly higher rendering quality of novel view and novel pose synthesis.
In addition, we are able to reconstruct fine-level details such as cloth wrinkles, face details or logos at a resolution and fidelity that the prior works such as NeuralBody~\cite{peng2020neural} was not able to achieve (see Fig.~\ref{fig:intro}).
The contributions of this work are summarized below:
\begin{enumerate}
\item We introduce a new framework with monocular RGB-D as input and show significant improvement on the unseen poses compared to existing methods, with high-fidelity reconstruction of fine-level facial, cloth and body details.
\item We combine pose and appearance representations by modeling shared information across frames and specific information at each individual frame.
These two representations help the model to generalize better to novel poses compared to only utilizing the pose representation.
\item A temporal transformer is introduced to combine information across frames, which helps to recover non-visible details in the query frame (at unseen views).
\end{enumerate}
\section{Related Work}
\label{sec:relatedwork}
\noindent \textbf{3D Neural Representations}.
Early 3D shape representation works can be classified into three categories: point-based methods~\cite{achlioptas2018learning,qi2017pointnet}, voxel-based methods~\cite{wu20153d,choy20163d} and mesh-based methods~\cite{groueix2018papier,ben2018multi,yang2017foldingnet}.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{latex/fig/framework.png}
\vspace{-2mm}
\caption{\textbf{Overview.} Given a query point as input, our method learns pose and appearance codes which can simultaneously integrate shared information across frames and model the appearance information at each frame. The pose and appearance codes are anchored to the human pose and point clouds, respectively. In addition, we use the body motion to track 3D points from the query frame to the key frames and extract the pose code from the key frames. A transformer fusion module is proposed to combine the pose codes of all frames. Similar to NeRF~\cite{mildenhall2020nerf}, we use spatial location and viewing direction as the extra inputs and train the model to predict the density and color for each 3D point.}
\label{fig:pipeline}
\end{figure}
Implicit representations are then used to represent shapes by reconstructing a continuous surface geometry, which utilizes the spatial coordinates as the input and outputs the signed distances or occupancy values.
With advances in differential rendering methods, geometry and appearance can be learned from multi-view observations.
Related works can be categorized into static~\cite{sitzmann2019scene,liu2020nsvf,mildenhall2020nerf,sitzmann2019deepvoxels,yu2021pixelnerf,yu2021plenoctrees} and dynamic scenes~\cite{pumarola2020d,peng2020neural,park2020deformable,hedman2021snerg,kwon2021neural,gafni2021dynamic,tretschk2021non,raj2021pva,chen2021animatable,peng2021animatable,li2021neural,du2021neural,gao2021dynamic,nguyen2021human}.
\noindent \textbf{Static Scene Representations}.
SRN~\cite{sitzmann2019scene} represents scenes as continuous functions that maps 3D coordinates to a feature representation of local scene properties and formulates the image as a differentiable ray-marching algorithm.
NSVF~\cite{liu2020nsvf} utilizes a sparse voxel octree to represent a set of voxel-bounded
implicit fields. A diffentiable
ray-marching operation is adopted to render views from a set of posed RGB images.
NeRF~\cite{mildenhall2020nerf} achieves state-of-the-art renderings for novel view synthesis by optimizing a neural radiance field for a scene, which maps 3D coordinates and viewing directions to density and color using a neural network.
While NeRF can render photo-realistic images given dense images as input, it is limited mostly to static scenes.
\noindent \textbf{Dynamic Scene Representations}. Dynamic NeRFs~\cite{pumarola2020d, peng2020neural} extend NeRF to dynamic scenes by introducing a latent deformation field or human poses.
D-NeRF~\cite{pumarola2020d} learns a mapping from the observation space to the canonical space at a particular time.
NeuralBody~\cite{peng2020neural} proposes a set
of latent codes shared across all frames anchored to a human body model in order to replay character motions from arbitrary view points under training poses.
These methods, where the deformations are learned by neural networks, allow to handle general deformation, and synthesize novel poses by using interpolation in the latent space.
However, the human poses cannot be controlled by users or the synthesis fails under novel poses.
Human pose based representation can model the body shape at any time step but will fail to capture detailed appearance.
To overcome this problem, we propose to construct the appearance-based representation by utilizing the 2D features anchored to the point clouds as an input.
\noindent \textbf{Dynamic Scene Fusion}.
To model the temporal cues across frames, previous works\cite{dou2016fusion4d, yu2018doublefusion,li2021neural,kwon2021neural} combine motion information.
Li \textit{et al.}~\cite{li2021neural} propose to learn dense scene flow fields that aim to learn 3D offset vectors from a point in time $t$ to the same point in time $t$-1 and time $t$+1.
The offsets are implicitly supervised with 2D optical flow.
Kwon \textit{et al.}~\cite{kwon2021neural} employ a temporal transformer to integrate skeletal features across different frames.
The vertices of the human body are first reprojected to the 2D image plane and then image features are sampled to obtain the skeletal features.
Although both the work by Kwon \textit{et al.}~\cite{kwon2021neural} and our method use the temporal transformer, the way we use the transformer differs from~\cite{kwon2021neural}. Kwon \textit{et al.}~\cite{kwon2021neural} use the transformer to combine pixel-aligned skeletal features. The features are obtained by first projecting the vertices to the 2D image plane and then sampled from 2D images using bilinear interpolation.
In addition, they require multiple views as input as inaccurate features are extracted when projecting the 3D vertices into single view.
Instead of combining the skeletal features, we propose to use the transformer to combine the pose codes for any 3D point and its tracked points. Our method optimizes the 3D feature volumes and does not require multiple views as input.
\section{Methodology}
Given a monocular RGB-D video of a human in motion, we aim to synthesize free-viewpoint videos of the person under novel views and new poses.
We denote the set of video frames as $\{I_t|t=1,...,N_f\}$, where $t$ represents the frame index and $N_f$ is the number of frames.
To avoid the influence of background changes due to the camera movements, we use a mask to remove the background color and only focus on the human in the foreground.
We start by briefly introducing the notation and concept of NeRF for static scenes in Sec.~\ref{nerf_formula}.
In Sec.~\ref{pose_representation}, NeRF is extended to dynamic human performers which are conditioned on the pose code shared across all frames.
In Sec.~\ref{appear_representation}, the appearance representation is introduced by using a single RGB-D image as the cue for each frame.
To integrate the motion information across frames, we propose to employ a temporal transformer as discussed in Sec.~\ref{transformer} to help recover the non-visible parts.
Volume rendering and the corresponding objective functions are introduced in Sec.~\ref{volume_rendering} and Sec.~\ref{loss_function}.
The overview of our approach is illustrated in Fig.~\ref{fig:pipeline}.
\subsection{Neural Radiance Fields} \label{nerf_formula}
Neural Radiance Fields (NeRFs)~\cite{mildenhall2020nerf} represent a static scene as a radiance field and render the color using classical volume rendering~\cite{kajiya1984ray}.
NeRF utilizes the 3D location $\bm{x}=(x,y,z)$ and 2D viewing direction $\bm{d}$ as input and outputs color $\bm{c}$ and volume density $\sigma$ with a neural network for any 3D point,
\begin{equation}
F_{\Theta}:(\gamma_x({\bm x}), \gamma_d({\bm d})) \rightarrow (\bm{c}, \sigma).
\end{equation}
To render the color of an image pixel, NeRF uses the volume rendering integral equation by accumulating volume densities and colors for all sampled points along the camera ray.
Let $\bm{r}$ be the camera ray emitted from the center of projection to a pixel on the image plane.
The expected color of that pixel bounded by $h_n$ and $h_f$ is then given by:
\begin{align}
\Tilde{C}(\bm r) = \int_{h_n}^{h_f} T(h) \sigma(\bm{r}(h))\bm{c}(\bm{r}(h),\bm{d})dh,
\label{eq:nerf_equa}
\end{align}
where $T(h) = \exp (- \int_{h_n}^h \sigma(\bm{r}(s)) ds)$.
The function $T(h)$ denotes the accumulated transmittance along the ray from $h_n$ to $h$.
NeRF is trained on a collection of images for each static scene with known camera parameters, and can render scenes with photo-realistic quality.
To extend NeRF to model the dynamic human body, we propose to learn the implicit representation to represent the shape and appearance information of the human.
Specifically, we introduce a pose-conditioned representation \emph{shared by all frames} and an appearance-conditioned representation \emph{specified by each individual frame}.
\subsection{Pose-conditioned Representation}\label{pose_representation}
For each frame, we assume the 3D human model is given.
We first extract the vertices from the posed 3D mesh and aim to learn a set of pose codes $\bm Z=\{\bm z^1, \bm z^2, \ldots,\bm z^{N_m}\}$ anchored to the vertices of the human body model at frame $t$. Here $N_m$ denotes the number of codes.
Similar to~\cite{peng2020neural}, the dimension of each pose code is set to 16.
Then the implicit representation is learned by forwarding the pose code into a neural network, which aims to represent the geometry and shape of a human performer.
The pose space is shared across all frames, which can be treated as a common canonical space and enables the representation of a dynamic human based on the NeRF.
Finally a neural network is utilized to learn the density and color for any 3D point and volume rendering is used to render the RGB value for each pixel.
The pose codes anchored to the body model are relatively sparse in the 3D space and directly calculating the pose codes using trilinear interpolation would lead to less effective features for most points.
To overcome this challenge, motivated by~\cite{yan2018second,peng2020convolutional,shi2020pv,peng2020neural}, a SparseConvNet is used which propagates the codes defined on the mesh surface to the nearby 3D space.
Specifically, to acquire the pose code for each point sampled along the camera ray, we utilize the trilinear interpolation to query the code at continuous 3D locations.
Here the pose code for point $\bm x_t^i$ at frame $t$ is represented by $\phi(\bm x_t^i,\bm Z)$ and will then be forwarded into a neural network to predict the density and color.
The pose codes learned in the shared space on all frames can model the human shape well in both known and unseen poses.
However, the synthesized views still lack details under novel poses without optimizing the pose codes.
To model the details on each frame, we
propose an appearance-conditioned implicit representation by utilizing the single view RGB-D as the reference input.
\subsection{Appearance-conditioned Representation}\label{appear_representation}
A monocular RGB image can serve as the appearance prior for the the human body under one view.
To learn detailed information at each individual frame, we propose to learn the appearance code anchored to the point clouds.
The points clouds are generated by the 2D image lifted to the 3D space using the depth image.
The point clouds generated in this way model the partial body of the human performer and show details such as wrinkles on the clothes.
Given a 2D pixel $\bm{p}_t^i$ and corresponding depth value $d_t^i$, the point clouds generation process
can be formulated as:
\begin{equation}
\bm{p}^{s_i}_t = F_\varphi(\bm{p}_t^i, d_t^i).
\end{equation}
Here $\bm{p}^{s_i}_t$ is the 3D point generated by the 2D pixel for frame $t$. $F(\dots)$ is the function generating a 3D point given a 2D pixel and camera pose.
Different from the pose-conditioned latent codes that are shared across all frames, here the proposed appearance-conditioned codes are anchored to the point clouds, which are obtained from the pixel-aligned features extracted from the image encoder $E$.
To take advantage of the rich semantic and detailed cues from images, we use a 2D convolution network (ResNet34~\cite{he2016deep}) to encode the image feature map $E(I_t)$ for the given image $I_t$.
Specifically, we first extract features from the ResNet34, and then utilize three convolutional layers to reduce the dimension followed by a SparseConvNet to encode the features anchored to the sparse point clouds.
To obtain the appearance code for each point sampled along the camera ray, we utilize the trilinear interpolation to query the code at the continuous 3D locations.
$\psi(\bm{x}_t^i,E)$ is adopted to represent the appearance code for point $\bm{x}_t^i$.
The appearance code together with the pose code will be forwarded into a neural network to predict the density and color.
The appearance code learned on each single frame can model the details on the human body and help recover some missing pixels in the 3D space.
\subsection{Temporal Fusion Module}\label{transformer}
Frames from different time steps can provide complementary information to the query frame. Given the features extracted from the key frames, a temporal transformer is utilized to effectively integrate the features.
To obtain the corresponding pixels in the key frame, we use the body model extracted from each frame to track the points.
\noindent\textbf{Point Tracking}. First, $N_s$ points on each face of the mesh are randomly sampled, which result in $N_s \times N_m$ points on the whole surface of a human body. Here $N_m$ represents the number of faces.
We calculate the distance between a 3D point sampled on the camera ray and all points on the surface at the query frame $I_t$.
Here we only keep each sample $\bm{x}_t^i$ close to the
surface for rendering the color if $\min_{v\in \mathcal{V}_t}||\bm{x}_t^i-v||_2<\gamma$ and obtain the nearest point $\bm{\hat{x}}_t^i$ on the surface at frame $I_t$, where $\mathcal{V}_t$ is the set of sampled points.
In addition, we can track the points at different frames that match $\bm{\hat{x}}_t^i$ by the body motion, and assign the feature of the tracked points to $\bm{x}_t^i$.
\noindent\textbf{Key Frame Selection}. We automatically select three key frames from the training frames.
We first calculate the distances between all training poses and the pose for the query frame $\bm{S}_t$ by $||\bm{S}_t-\bm{S}_j||_2 (j \in N_f)$ and keep the frames with the $K$-NN distances.
Here $\bm{S}$ are the coordinates of the vertices extracted from the body mesh and $K$ is set to 2.
In addition, we select the first frame as the fixed key frame.
For simplicity, the key frame selection strategy is not trained with the whole model.
\begin{figure}[t]
\centering
\begin{tabular}{c@{\hspace{0.3mm}}c}
\includegraphics[height=0.4\linewidth]{latex/fig/transformer_ns_v2.PNG}&\\%2_horizontal
\end{tabular}
\caption{\textbf{Temporal Fusion Module.} Given the pose codes from the query point and tracked points as input, temporal fusion aims to aggregate the codes by using a transformer-based structure.}
\label{fig:visual_fusion}
\end{figure}
\noindent\textbf{Temporal Fusion}. After obtaining the pose code
from $N$ frames ($K$+1 key frames and one query frame), a transformer based structure~\cite{dosovitskiy2020image} is employed to take the $N$ features as input and utilize a multi-head attention mechanism and feed-forward multi-layer perceptron (MLP) for feature aggregation.
The fusion module is described in Fig.~\ref{fig:visual_fusion}.
Here we introduce the multi-head self-attention mechanism, which generates a trainable associate memory with a query, key and value to an output via linearly transforming the input.
Given the input feature $\phi(\bm x_{t}^i, \bm Z)$, the query, key and value are represented by $f_q(\phi(\bm x_{t}^i, \bm Z))$, $f_k(\phi(\bm x_{t}^i, \bm Z))$ and $f_v(\phi(\bm x_{t}^i, \bm Z))$ respectively.
The query and the key are used to calculate the attention map using the multiplication operation, which represents the correlation between all the features.
The attention map is used to retrieve and combine the feature in the value.
Formally, the attention weight for point $\bm x_t^i$ in frame $t$ and tracked point $\bm x_k^i$ in frame $k$ is calculated by:
\begin{equation}
a_{t,k}^i = \psi\left(\frac{f_q(\phi(\bm x_t^i, \bm Z)) \cdot f_k(\phi(\bm x_k^i, \bm Z))^{\top}}{\sqrt{d}}\right),
\end{equation}
where $\sqrt{d}$ is a scaling factor based on the network depth, and $\psi(\cdot)$ denotes the softmax operation.
The aggregated feature for input $\phi(\bm x_t^i, \bm Z_{t})$ is formulated as:
\begin{equation}
\phi'(\bm x_t^i, \bm Z) = \sum_{k\in \mathcal{K}}\phi(\bm x_t^i, \bm Z) \cdot a_{t,k}^i + f_v(\phi(\bm x_t^i, \bm Z)),
\end{equation}
where $\mathcal{K}$ denotes the index set of the combined frames.
In this work, multi-head self-attention is adopted by running multiple self-attention operations, in parallel. The results from different heads are integrated to obtain the final output.
After the self-attention mechanism, each input feature contains its original information and also takes into account the information from all other frames.
As such, the information from key frames and the query frame are combined together.
Average pooling is then employed to integrate all features, which serves as the output of the temporal fusion module.
In our implementation, we do not adopt any positional encoding on the input feature sequence.
\subsection{Density and Color Regression}
\label{volume_rendering}
Figure~\ref{fig:pipeline} shows the prediction of density and color that are represented by a neural network.
For each frame, the network takes the pose code, appearance code, spatial location and viewing direction as the input and outputs the density and color for each point in the 3D space.
Similar to~\cite{mildenhall2020nerf, rahaman2019spectral}, we apply positional encoding to both the viewing direction $\bm d$ and the spatial location $\bm x$ by mapping the inputs to a higher dimensional space.
For frame $t$, the volume density and color at point $\bm x_t^i$ is predicted as a function of the latent codes, which is defined as:
\begin{equation}
(\sigma_t^i, \bm c_t^i) = M(\phi(\bm x_t^i,\bm Z),\psi(\bm x_t^i,E), \gamma_d(\bm d_t^i), \gamma_x(\bm x_t^i)),
\end{equation}
where $M(\cdot)$ represents a neural network.
$ \gamma_d(\bm d_t^i)$ and $ \gamma_x(\bm x_t^i)$ are the positional encoding functions for viewing direction and spatial location, respectively.
Details of the network architecture can be found in the supplementary material.
\subsection{Objective Functions}
\label{loss_function}
We optimize the proposed model using the following objective function:
\begin{equation}
\mathcal{L}=\mathcal{L}_{c1}+\mathcal{L}_{c2}, \end{equation}
where $\mathcal{L}_{c1}$ and $\mathcal{L}_{c2}$ denote the reconstruction loss for the rendered pixels and image loss for the image decoder network $D$.
The image decoder comprises multiple Conv2D layers behind the ResNet34, which aims to reconstructed the input image.
We render the color of each ray using both the coarse and fine set of samples, and minimize the mean squared error between the rendered pixel color $\Tilde{C}_c(\mathbf{r})$ and ground-truth color $C(\mathbf{r})$ for training:
\begin{equation}
\mathcal{L}_{c1}=\sum_{\mathbf{r}\in R}||\Tilde{C}_c(\mathbf{r}) - C(\mathbf{r})||_2^2 + ||\Tilde{C}_f(\mathbf{r}) - C(\mathbf{r})||_2^2,
\end{equation}
where $R$ is the set of rays. $\Tilde{C}_c$ and $\Tilde{C}_f$ denote the prediction of the coarse and fine networks.
\begin{equation}
\mathcal{L}_{c2}=\sum_{\bm p\in \mathcal{I}}||\Tilde{I}(\bm p) - I(\bm p)||_2^2.
\end{equation}
The symbol $\Tilde{I}(\bm p)$ and ${I}(\bm p)$ represent the reconstructed and ground truth colors for pixel $\bm p$. $\mathcal{I}$ is the set of pixels.
\section{Implementation Details}
\begin{table*}[t]
\setlength{\tabcolsep}{0.4em}
\begin{center}
\small
\scalebox{0.88}{
\begin{tabular}{lccccccccccccc}
\toprule
& &\multicolumn{2}{c}{\emph{Sequence1}} & \multicolumn{2}{c}{\emph{Sequence2}} & \multicolumn{2}{c}{\emph{Sequence3}} & \multicolumn{2}{c}{\emph{Sequence4}} &
\multicolumn{2}{c}{\emph{Sequence5}} \\
\midrule
Models & \begin{tabular}[c]{@{}c@{}}Train \\ Views\end{tabular} & PSNR$\uparrow$ & SSIM$\uparrow$& PSNR$\uparrow$ & SSIM$\uparrow$ &PSNR$\uparrow$ & SSIM$\uparrow$ &PSNR$\uparrow$ & SSIM$\uparrow$
&PSNR$\uparrow$ & SSIM$\uparrow$\\
\midrule
NeRF+pose & \checkmark &16.87&0.59&20.75&0.67&12.76&0.32&19.94&0.54&23.81& 0.77\\
NeRF+pose+depth & \checkmark &17.32&0.66&20.93&0.72&13.08&0.34&20.65&0.61&23.91&0.80\\
NeuralBody & \checkmark &19.08&0.77&22.83&0.79&14.12&0.46&23.84&0.77 & 24.66&0.83\\
NeuralBody+depth & \checkmark &\textbf{19.67} &\textbf{0.77}&23.81&0.80&14.82&0.48&24.01&0.76&24.70&0.83\\
\textbf{Ours} & \checkmark &19.21 & 0.75 & \textbf{24.91} & \textbf{0.81} & \textbf{15.72} & \textbf{0.49} & \textbf{24.96}& \textbf{0.77} &\textbf{24.95} & \textbf{0.84}\\
\midrule
NeRF+pose &&16.88 &0.59 &20.03 & 0.65 &12.05 &0.29 &19.87 &0.52 &22.15 &0.70\\
NeRF+pose + depth & &17.65 &0.71&21.60&0.74&12.25&0.30&21.01&0.54&22.47&0.75\\
NeuralBody &&19.91 & \textbf{0.79} & 22.76 &0.79 & 13.52 &0.41 &23.81 &\textbf{0.77} & 23.17 & 0.78\\
NeuralBody+depth & &\textbf{19.92} &0.78&23.80&0.80&13.86&0.43&23.92&0.75&23.49&0.79\\
\textbf{Ours} &&19.06&0.73&\textbf{24.89} & \textbf{0.82} & \textbf{15.01} &\textbf{0.45} & \textbf{24.63} & 0.74 &\textbf{23.91} &\textbf{0.81} \\
\bottomrule
\end{tabular}
}
\caption{\textbf{Quantitative Comparison} on training (first five rows) and novels views (last five rows) under novel poses across all sequences.}\label{tab:quan_novel_view}
\end{center}
\end{table*}
\noindent \textbf{Network details}. For the encoder $E$, we extract a feature pyramid~\cite{lin2017feature} from each image similar to~\cite{yu2021pixelnerf}.
A ResNet34 backbone pretrained on ImageNet is utilized for our experiments.
The output feature of the decoder has $1/4$ the spatial resolution compared to the input image.
Multi-scale features are extracted prior to the fourth pooling layer.
We extract pixel-aligned features using bilinear interpolation, and then concatenate them to form a latent vector of size 256.
To construct the image decoder $D$, we simply connect several convolutional layers with upsampling operation to reconstruct the input image.
For the transformer network, we utilize three heads for the self-attention module, which has a similar structure as~\cite{dosovitskiy2020image}.
Following NeRF~\cite{mildenhall2020nerf}, we perform hierarchical volume sampling and simultaneously optimize a coarse and fine network with identical network architecture. %
At the coarse scale, we sample a set of $M_c$ points using stratified sampling,. %
With the prediction of the coarse network, we then sample another set of points along each camera ray, where samples are more likely to be located at relevant regions for rendering.
We sample additional $M_f$ locations and use all $M_c+M_f$ locations to render the fine results, where $M_c$ and $M_f$ are set to 64 respectively.
\noindent \textbf{Training details}. We train all layers using Adam~\cite{kingma2014adam}.
The base learning rates for the encoder-decoder network and other layers are $10^{-3}$ and $5\times10^{-4}$, which decay exponentially during the optimization. Additional network architecture as well as implementation details will be provided in the supplementary material.
\section{Experiments}
\setlength{\tabcolsep}{.04cm}
\begin{table}[t]
\centering
\footnotesize
\begin{tabular}{lcc}
\toprule
Model Variant & PSNR$\uparrow$ & SSIM$\uparrow$\\
\midrule
W/o appearance-based representation &23.81&0.80 \\
W/o temporal transformer
&23.74&0.79\\
\textbf{Ours} &\textbf{24.89} &\textbf{0.82} \\
\bottomrule
\end{tabular}
\;\;\;\;\;\;\;
\begin{tabular}{lcc}
\toprule
\# frames & PSNR$\uparrow$ & SSIM$\uparrow$\\
\midrule
1 &24.05&0.80 \\
3 (Ours)
&24.89 &0.82\\
5 &24.92 &0.82 \\
\bottomrule
\end{tabular}
\caption{\textit{\textbf{Left}}: Ablation Study of appearance-based representation and transformer fusion module on \emph{Sequence2} \textit{\textbf{Right}}: Effect of the number of key frames on \emph{Sequence2}}
\label{tab:aba_appearance_code}
\end{table}
\noindent \textbf{Datasets}.
To train our method, we rely on four sequences of real humans in motion that captured with a 3dMD full-body scanner as well as a single sequence of a synthetic human in motion.
The 3dMD body scanner comprises 18 calibrated RGB cameras that can capture a human in motion performing various actions and facial expressions and output a reconstructed 3D geometry and material image file per frame.
These scans tend to be noisy but can capture facial expressions and fine-level details like cloth wrinkles.
The synthetic scan is a high-res animated 3D human model with synthetic clothes (t-shirt and pants) that were simulated.
Unlike the 3dMD scans, this 3D geometry is very clean but it lacks facial expressions.
We render RGB and Depth for all real and synthetic sequences from 31 views at \(2048 \times 2048\) resolution that cover the whole hemisphere (very similar to the way that NeRF data is generated) at 6 fps using Blender Cycles.
The number of video frames in our experiments varies between 200 to 600 depending on the sequence. The image resolution for the training and test is set to 1024$\times$1024.
To train our model, we select the first half of the frames for training and the rest frames for inference.
Both training and test frames contain large variations in terms of the motion and facial expressions.
At training and test stages, a single RGBD image at each frame is used as the input.
All the input RGBD images at different frames share the same camera pose. In addition, 29 more views with different camera poses are used to train the network. The output is a rendered view given any camera pose
(not including the camera pose of the input RGBD image).
\noindent \textbf{Evaluation Metrics}.
Following existing approaches~\cite{mildenhall2020nerf, peng2020neural}, we evaluate the performance on the proposed dataset using two metrics, including the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).
\noindent \textbf{Baselines}. We evaluate the proposed model against state-of-the-art methods on view synthesis (training and novel views) under unseen poses:
\begin{compactitem}
\item NeRF+pose: Similar to~\cite{liu2021neural}, we extend NeRF~\cite{mildenhall2020nerf} from the static scenes to a pose-conditioned NeRF by utilizing the pose vector as an extra input.
\item NeRF+pose+depth: We extend NeRF+pose by using the depth information. Specifically, single RGBD image is first forwarded into the ResNet34 to extract the 2D image feature. Given any 3D point sampled on the camera ray, the pixel-aligned feature is extracted which serves as an extra input.
\item NeuralBody: NeuralBody~\cite{peng2020neural} extends NeRF to dynamic scenes using the latent codes anchored to the human pose as an extra input besides the coordinate and viewing direction.
\item NeuralBody+depth: 2D image feature is first extracted by using the single RGBD image as the input. The vertices on the human body are projected to the 2D space to extract the skeletal feature. The skeletal feature together with the latent codes are integrated, which are anchored to the vertices.
\end{compactitem}
For training and inference, these baselines use the same camera views and human poses as ours for fair comparison.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{latex/fig/nikos_figs/tiantian_ablations_truetony.png}
\caption{\textbf{Ablation Studies.} Effect of the appearance code and temporal transformer on synthesized images. Both components can help recover better details on the human.}
\label{fig:applications_appearance_trans}
\end{figure*}
\subsection{Experimental Results}
We evaluate all methods on unseen new poses.
Results on novel as well as training views are presented.
\noindent \textbf{Novel view and novel pose}.
Table~\ref{tab:quan_novel_view} shows quantitative results.
The proposed method clearly outperforms other competitive approaches on most of the video sequences.
All baseline methods, have a hard time generating realistic renders when the test poses deviate significantly from what was seen during training.
Comparing with baseline models, the proposed method can generate better details on the human body, which indicates that the point cloud-based representation and the temporal transformer can help reason
the missing structure and recover the non-visible parts at the unseen poses.
\noindent \textbf{Training view and novel pose}.
The quantitative results on training views and novel poses are shown in Table~\ref{tab:quan_novel_view}.
On most video sequences, the proposed method performs better compared to all other baselines.
\noindent \textbf{Qualitative Results}. We show visual comparison with other methods under novel poses in Fig.~\ref{fig:applications_vis}.
With the human pose as the geometric guidance, NeuralBody can predict the body shape well but fails to render fine-level details on the human body.
NeuralBody does not generalize well to novel poses because the shared latent codes across all frames are not optimized during the test stage.
More specifically, this model overfits to the training poses and generates blurry results for unseen poses, as one can observe in Fig~\ref{fig:applications_vis} for the details on the body (1st row) and head (2nd row).
\begin{figure}[t]
\centering
\includegraphics[width=0.85\linewidth,height=0.38\linewidth]{latex/fig/fig6/fig6.jpg}
\caption{\textbf{Failure Cases} on \emph{Sequence5}. The proposed method does not render smooth facial details when the unseen poses and training poses have significantly different expressions.}
\label{fig:applications_failure}
\end{figure}
\subsection{Ablation Studies}
We analyze the effects of key components in our method with ablation studies.
\begin{figure*}[ht]
\centering
\includegraphics[width=.99\linewidth]{latex/fig/fig4_v3.png}
\vspace{-2mm}
\caption{\textbf{Qualitative Results.} Our proposed approach achieves significantly better render quality compared to NeuralBody (+Depth) across various poses, viewpoints and subjects. All poses are unseen and have not been used for training.}
\label{fig:applications_vis}
\end{figure*}
\noindent{\textbf{Contribution of appearance-based representation.}} We analyze the effect of the appearance-based representation in Table~\ref{tab:aba_appearance_code} and Fig.~\ref{fig:applications_appearance_trans}.
Using the appearance code brings performance improvement on the fine structures
in different parts of the body, which demonstrates that the appearance code anchored to the points clouds can help recover the missing pixels in the query view.
\noindent{\textbf{Contribution of the temporal transformer}}. As shown in Table~\ref{tab:aba_appearance_code} and Fig.~\ref{fig:applications_appearance_trans}, utilizing the temporal fusion module can help the model generate better rendering performance. We observe that the details like logos on the shirt are finer, the hands are cleaner and the face is significantly more crisp.
\noindent{\textbf{Effect of the number of key frames}}. To evaluate the impact of the number of key frames, we report the
performance in Table~\ref{tab:aba_appearance_code}. The performance increases with more key frames and saturates with 5 frames.
\subsection{Limitations}
Our method leverages a pose code anchored to the human pose and an appearance code anchored to the point clouds that can model the shape of the human and can help fill in the missing parts in the body.
The temporal transformer module can help recover more non-visible pixels in the body. Without encoding the facial expressions, our method can handle humans without substantial facial expression variations.
As shown in Fig.~\ref{fig:applications_failure}, the proposed method predicts blurred expressions on the face of \emph{Sequence5}, in which the query frame has facial expression different from the key frames and combining key frames with the query frame makes the network unable to differentiate the specific facial expression in the query frame.
To overcome this, future work will encode facial expression for each frame as a separate code and thus being able to render such diverse facial expressions under new views.
\section{Conclusion}
In this paper, we build upon recent advances of neural radiance fields pertaining to digital humans and address key challenges that existing human body based methods suffer from, preventing them to generalize to unseen poses. Towards that direction, we propose to integrate a pose code and an appearance code to synthesize humans in novel views and different poses with high fidelity.
The pose code that is anchored to the human pose can help model the human shape whereas the appearance code anchored to the points clouds can help infer the details and recover the missing parts. %
To leverage temporal information, we further propose to utilize the body motion to track points from the query frame to a few automatically selected key frames and adopt a temporal transformer to aggregate information across multiple frames.
The transformer-based fusion module can recover the non-visible part in the query frame.
The proposed method achieves significantly better results against the state-of-the-art methods under novel views and poses with quality that has not been observed in prior work. Fine-level information such as fingers, logos, cloth wrinkles and face details are rendered with high fidelity. |
\section{Introduction}
\label{intro}
Embedded and real-time systems play an important role in modern applications. These systems are utilized in many critical applications such as health care, automation, avionic, communication, and so on. Embedded processors as the main part of these systems should perform correctly and autonomously~\cite{wang2017real,li2003real}. In many applications, the target system is employed in critical or hardly available situations where their correct and self-determined operation is very effective on their final performance. Arithmetic and logic operations are the main part of each processing unit and play a critical role in the system's functionality and reliability. Considering fault tolerance property in arithmetic and logic unit of embedded processors, make their operations more precise and autonomous~\cite{FT-ALU-2009,FAZELI2011-emalu,lizard}.
Along with the technology advances and shrink in size of transistors, the vulnerability of systems against transient and permanent faults increases. These faults affect various aspects of systems and make their results erroneous. Computation and arithmetic units are very sensitive to the faults because their results propagate to the system output, directly. Applying fault tolerance approaches in the computation unit is necessary to meet its correctness and autonomous requirements~\cite{xia2021error,ft-alu-2012,lizard}. Redundancy is the main solution in designing fault-tolerant systems and could be applied in various aspects of the system. Depending on the employed level of redundancy, fault detection or correction is possible in the cost of various performance and area overheads. In the case of transient faults with a very short lifetime, detection could be solely enough that leads to system restarting and recovery. Oppositely, permanent faults are more complicated and remain at the systems for a longer time. Thus, considering more complicated approaches that include correction and recovery along with detection methods are essential for their mitigation~\cite{ft-redundancy-2021,Tay2017,nelson1990fault,abdi2012}.
Time redundancy is an efficient fault-tolerance approach that replicates the system execution in time. In this approach, multiple replicated versions of the system are considered and repeat the same operation, serially. By replicating the execution of operands and comparing the results, transient faults are simply tolerated. To consider permanent faults in these methods, making variations in each replicated version of execution is beneficial and facilitates fault detection by comparing the differentiated results of computation. Detection capability is not enough for making a system fault-tolerant and locating the faulty place and correcting it consequently, are also required. These capabilities are achieved through considering extra controlling schemes such as reconfiguration or voting.
Time redundancy-based methods are efficient in terms of hardware cost compared to parallel execution and are considered as a proper choice for designing fault-tolerant embedded and real-time systems due to their time predictability~\cite{nelson1990fault,wang2017real,abdi2018hystery,alvarez2018mixing}.
Since in embedded and real-time systems the execution time analysis is performed based on worst case estimation, applying some time overhead during system execution could be acceptable. Normal executions are generally shorter than the worst-case estimation, so the saved time could be employed for the replicated executions and making the system fault-tolerant. The main mission of embedded and real-time systems could be summarized in control applications where the exact and correct computations are very important. Computations are mainly performed in the arithmetic and logic unit (ALU) of the processors so designing them fault-tolerant is very important. In this context, applying redundancy to various aspects of the system is applicable but the resulting overhead should be studied precisely based on the requirements of embedded systems~\cite{FTadder-MEJ,berger-Tabriz,acharya2018berger-iop,barredo2020efficiency}.
In this paper, we propose our fault-tolerant time redundancy-based ALU architecture for embedded and real-time systems (FT-EALU). The main purpose of this method is to mitigate the transient and permanent faults that occurred in the arithmetic and logic unit of embedded processors. Since most embedded and real-time systems are autonomous, their self-recovery from potential faults is required. In this context, the passive fault tolerance approaches that perform voting among the replicated versions of the operation and do not require additional management are convenient. Transient faults are detected by comparing the results of replicated versions of computation and generally corrected as the result of performing a simple voting mechanism among them due to their short remaining time in the system.
Reversely, to correct permanent faults, repeating the execution is not enough due to their long lifetime, and mitigation approaches including fault locating and recovery are also required. Moreover, by replicating the same version of the execution, permanent faults will not be detected, and applying variations to each version is necessary to make a diversion in the final results.
In our proposed approach, simple and low overhead modifications of input data such as shift and swap are considered to avoid hardware costs and make sufficient diversity in input data to mitigate permanent faults. By diversifying the various versions of computation, different results, due to the effect of permanent faults, are generated that could be mitigated during the voting process. Traditional majority voting schemes are generally not efficient for mitigating permanent and inclusive faults so employing an appropriate weighted voter that identifies the trusted results and magnifies their effect on final output improves the efficiency.
Experimental results show that our proposed FT-EALU method is capable of correcting $84.93\%$ of injected permanent faults in arithmetic and logic operations.
The main contributions of our proposed fault-tolerant approach could be summarized as follows:
\begin{itemize}
\item Presenting a time redundancy-based fault-tolerant approach to correct transient and permanent faults of computation in arithmetic and logic unit of embedded systems;
\item Presenting an adaptive and software-implemented weighted voting approach based on reward and punishment trust coefficient factor to cover the effect of faults and isolate them from final result;
\item Employing a learning-based approach to define the appropriate weight for each bit of the results in voting process based on various fault injection scenarios to limit the cost and make the proposed method more autonomous;
\item Presenting an adaptive and self-tuned architecture to adjust the system detection and correction capability proportional to its tolerable overhead.
\end{itemize}
The rest of this paper is organized as follows: The related works are reviewed in section 2. In Section 3, the details of our proposed fault tolerance architecture including the detection and correction approaches are explained. The experimental results and evaluation of FT-ALU are provided in section 4. Finally, the conclusion remarks are mentioned in section 5.
\section{Related Work}
\label{related}
The arithmetic and logic unit is one of the most critical elements of embedded systems that directly affect the correctness and reliability of the system output~\cite{lizard,FTadder-MEJ}. The fault-tolerant design of this component is very important and effective on system reliability. Based on the applications of embedded systems, their reliability and autonomous operation are very important. In this context, several studies concentrate on fault-tolerant computation and ALU operation but at different levels and aspects. The realization of fault tolerance in system design is generally reflected by redundancy in its different aspects. In the design of fault-tolerant ALU, mainly hardware and information-based redundancy approaches are proposed~\cite{lizard,berger-Tabriz,akbar2020self-repair-adder}.
Information redundancy-based methods add error detection/correction codes to data to protect them against potential faults. Here, encoding and decoding modules are required because, in ALU, the focus is on computation so the value of input data is important and should be unchanged to have the correct result. Various methods based on parity, residue, and Berger codes are proposed in this category.
In~\cite{acharya2018berger-iop} a fault-tolerant architecture for embedded processors based on employing Berger code is proposed. This method mainly tolerates transient faults and expands the operands by a logarithmic factor of their length. Extracting the raw and intended output from its encoded version require decoding hardware that leads to hardware overhead which is hardly acceptable in embedded and real-time systems due to their limitations. In~\cite{parity-2003-TVLSI}, a fault-tolerant architecture for ALU and adder based on adding parity codes is presented. In this approach, the hardware cost caused by parity codes is low but the fault handling approach requires considering an extra ALU which is not applicable in most embedded systems. Generally, information redundancy-based methods are not appropriate for arithmetic circuits because it is probable that the added codes modify during the computation in ALU and lose their efficiency.
In~\cite{towhidy2019-residu-jcsc}, a fault-tolerant approach based on residue codes is proposed. In this study to overcome the high hardware overhead of previous research, a new residue checker circuit is proposed by considering more compression. However, the final hardware overhead decreases but still it is high for embedded systems. In~\cite{thakral2020novel-parity-id}, to make the ALU fault-tolerant, parity gates are employed which have limited capability in fault tolerance and are suitable to detect single-bit faults. Hardware redundancy-based methods enforce extra hardware to the embedded systems that are not acceptable in many applications where the area and performance of these systems are critical.
In~\cite{berger-Tabriz}, a fault-tolerant design of ALU using enhance Berger code is presented. The focus of this approach is on reducing the hardware and performance overhead of Berger code and its effectiveness in mitigating the permanent faults of ALU is studied.
In~\cite{FAZELI2011-emalu}, a fault-tolerant ALU architecture for embedded systems is proposed. In this approach, the unused sections of the ALU system are utilized for redundant computations and make the system fault-tolerant. This method shows good performance in terms of area and fault correction abilities but still adds extra hardware and delay to the system.
In~\cite{lizard}, a narrow width value-based method for handling hard faults of ALU is presented. In this method, two half-word ALUs are employed to detect system faults. After fault detection, each ALU is partitioned to locate and isolate the occurred fault. This approach uses the unused parts of the system to perform operations in parallel but its detection and isolation capabilities are limited based on the system capacity.
Employing duplication with comparison (DWC) and triple duplex redundancy (TMR) are other effective approaches for fault detection and correction and these methods are employed along with various re-execution approaches. Employing these methods in hardware redundancy-based approaches leads to high overhead, power consumption, and delay which is not acceptable in embedded and real-time systems. Time redundancy techniques decrease the hardware cost but prolong the execution time that could be managed in most real-time applications due to their WCET estimation. Convenient time redundancy methods such as RESO, RESWO, and REDWC~\cite{RESO,RESWO,REDWC} consider various modifications in different versions of computation to make them different and improve the detection capability.
In~\cite{gade2020run}, a TMR-based fault tolerance method based on considering the basic time redundancy approaches such as RESO \cite{RESO} is presented. The considered voting approach of this study is designed based on majority voter.
These methods are effective in fault detection at low cost and overhead but their coverage is improvable. Considering the combinations and enhanced versions of these techniques could be effective in the design of fault-tolerant computing systems.
\section{Proposed Method}
\label{proposed method}
Design of reliable and autonomous embedded and real-time systems is very important due to their wide applications. Fault tolerance approaches are defined at different levels of system design and manage the occurred errors and keep them away from system output. Fault tolerance realizes along with active or passive redundancy in the system. Active fault tolerance approaches have fault detection, localization, and isolation/correction phases. Implementation of active methods make some delays in the system due to the required system reconfiguration. Oppositely, passive methods cover the fault from system output and do not make any interrupt in its operation. Our proposed fault-tolerant architecture is based on passive approaches and voting among multiple versions of operations. As it has been explained in previous sections, time redundancy-based approaches are the proper choice for embedded and real-time systems due to their worst-case time analysis during design that generally leads to some slacks in runtime execution. Our proposed method hardens the ALU against transient and permanent faults by employing enhanced time redundancy-based approaches and performing fair voting among them to isolate the occurred errors from system output and is proper for embedded systems.
Fault detection is the main prerequisite of the fault tolerance process. The better this step is performed, the fault tolerance method will be more efficient. Conventional time redundancy-based fault detection approaches, diversify the operations by simple modifications to enhance their detection capability confronting the permanent faults. These simple modifications include complementing the operands, shifting or swapping them that cause to make difference between original and replicated executions' results~\cite{RESO,RESWO,REDWC}. The fault detection coverage of these methods is further enhanced by modifying their traditional approaches~\cite{ICEEabdi,gade2020run,kermani2016}. In our proposed fault-tolerant approach, we utilize the original and modified versions of input data based on the mentioned enhanced methods as the system's basic blocks.
As it has been explained, considering effective detection approaches and making more simple differences among replicated executions, facilitates the fault correction and recovery process. Our proposed method employs multiple versions of execution with diversifying operands and vote among them. Thus, in our proposed approach basically, three various versions of input data are considered for required serial operations: a) raw operands b) left-shifted operands~\cite{RESO} and c) left-shifted and swapped operands~\cite{ICEEabdi}. Then, the generated results of three versions of executions are adapted. In this way, the result of the second execution is shifted back to the right and the result of the last operation is shifted to the right and swapped again to have the final results correctly.
To improve the detection capability, in our proposed approach the potential carry of arithmetic operations is also considered and adjusted after modifications. Using simple operations such as shift and swap, enforces less overhead to the system but distributes the effect of stuck-at faults and makes their detection very probable. It is possible to consider different variations on input data but based on our experiments, considering the selected three versions of execution leads to the most proper correction capability. Since their final results have less similarity to each other and in this way more shuffling is applied to input data that makes the effect of occurred permanent fault more clear in voting. The effect of other variations of input data on final output is studied in detail and reported in section~\ref{results}.
To cover the effect of permanent faults that are resistant to repeated executions from the final output, voting among the results of defined operations on various input data is required. Majority voting is the simplest choice but its efficiency is limited due to the probability of wrong result repetition and its leading to final output. To overcome this, weighted voting is an appropriate solution but its efficiency is tightly dependent on the determined weights.
To specify the weight of each version of execution in our proposed method, we employ a \textit{trust factor} parameter based on the correction coverage of each version of execution. In this context, we propose two efficient weighting approaches based on reward/punishment, depending on the function of execution operation confronting the injected permanent faults. The first weighting approach is punitive and the second one employs the combination of reward and punishment in the evaluation and determining the \textit{trust factor}.
It should be noted that in our proposed approach, determining the weights is performed statically and through an offline learning approach. During this process, the correction capability of operands facing the injected permanent faults is investigated and based on their efficiency, a score is assigned to them. This score is derived based on several injected faults and determines the final weight of each operand in various execution versions. Since this process is performed statically and during system design, it enforces no complexity and time overhead. Afterward, the derived weights are utilized in the voting process among the results of execution versions.
In the proposed approach, the granularity of assigned weights is considered as fine as single bits to make it more precise.
The architecture of our proposed fault-tolerant approach is shown in figure~\ref{fig-arch}. In this figure, \textit{$W_{ij}$} shows the computed weight of $j^{th}$ version of execution for $i^{th}$ bit of the result (\textit{$result_{ij}$}). The various versions of execution in this figure, perform identical operations but on various operands that apply sequentially in time on a unique hardware platform. Employing various operands improves the permanent fault detection and correction capabilities, respectively.
The details of determining the weights of each execution and the employed voting mechanism are explained in the following subsection.
\begin{figure}
\centering
\includegraphics[width=4.5 in]{Fig-architecture.eps}
\caption{The high-level architecture of our proposed time redundancy-based fault tolerant approach}
\label{fig-arch}
\end{figure}
\subsection{The Proposed Voter Architecture}
\label{proposed-voter}
To determine the weights of various executions in the voting process, we propose and study two strategies based on reward and punishment.
Our first proposed weighted approach determines the weight of each execution (\textit{trust factor}) based on a punitive strategy. Here, the detection capability of considered executions confronting stuck-at faults in different bits of operands is studied. Since the operations are performed on various forms of operands, the effect of the fault and its propagation on different bits of the data are not the same. Thus, we define the weight for each bit of operands in execution versions based on their ability to correct the injected stuck at faults.
In the punitive weighting approach, the score of ideal executions that cover the injected stuck at fault does not change and the one(s) that propagate the error get a negative score.
This score is to each bit of operand and for each fault scenario, the unit negative score ("-1") divides among all versions of operation that are incapable of correcting the injected fault.
This score is a representation of punishment that reduces the \textit{trust factor} and the effect of the unreliable executions in the final output, consequently. In the other words, the score of unreliable execution is more negative and has less impact on deriving the final result of the system.
The total weight of each bit of operands in every version of execution is computed from its assigned scores in various fault injection scenarios. In this context, the total score of every single bit of input data, from all fault injection scenarios, divides by the considered faulty cases and then normalized. The normalization process scales the weights in a specified range and it is possible to consider various normalization approaches for deriving the final weights.
Here, we study three popular normalization approaches that are compared in section~\ref{results} and among them standard normalization scheme had the best results. Since standard normalization provides more scattered weights, it separates the results properly and reflects their unique effect on voting.
The explained weighting approach is performed at design time and through a learning-based scheme. In this context, several injected faults are considered during the training phase to determine the derived weight more efficiently.
Afterward, for each execution version, each bit of the result multiplies to its corresponding weight that is determined through the punitive approach and passes to the voter.
In the voter, the received results from three versions of execution are added together bit by bit and then divides by the total weights of the target bit in various versions. Then the derived result of the defined weighted average scheme should be interpreted. Due to the normalization process and final division, the final output of the voter is in the range of "zero" to "one". To map this bit-level result to binary values, we consider a threshold based on the performed operation. In this context, if the output of voter is greater than or equal to "0.5", it will be considered as one, otherwise, its value will be specified as zero in the binary scale.
Here, the value that is greater than "0.5" represents the case in which the operations with more than half of the total weights of the system indicate that the final result should be one and vice versa.
Thus, based on their domination in weight and the defined trust factor parameter, the most reliable versions of execution determine the final output of the system. It should be noted that other values of threshold are also studied but based on the mentioned explanations, the selected bound leads to the best results.
The defined punitive weighting approach is somehow strict and assigns the scores are greedily in one direction. In this approach, the weights are determined independent of the complexity of occurred fault and the weight of reliable execution is always zero. Thus, the effect of reliable versions is always constant in the final output and is not determinant.
Our second weighting approach is more efficient and defined as a combination of reward and punishment. Here, the assigned scores are negative for faulty cases (punishment) and positive for the corrected ones (reward). Moreover, in this approach, the scores are dynamically adapted based on the complexity and hardness of the faulty scenarios. To this aim, the maximum score of each faulty scenario, "+1", divides among the execution versions that are capable of mitigating it. For instance, if only one of the considered versions of execution is capable of correcting the injected fault, its corresponding score will be the highest, "+1", but in the case of two or three simultaneous fault corrections, this score divides by two or three respectively. Thus, the scores are set proportionally to the reliability of executions and the complexity of faulty scenarios.
The same procedure is also considered for punishment, and the complicated scenarios with hardly correctable errors have less punishment. The positive or negative score of each operation shows its assigned reward or punishment based on its fault correction capability.
Here again, after determining the scores of each bit, their corresponding weights are generated by normalization and division the same as the procedure that has been explained earlier for the first weighting approach.
The weights are determined for each bit of data to make the voting process more precise. These weights are extracted statically and at design time through a learning approach considering various permanent fault scenarios.
Figure \ref{fig_weight} shows the details of our proposed weighting approach. In this figure, the faulty scenarios are saved in fault injection storage and are applied to the input data during the training phase. Moreover, $"R_i"$ represents the intermediate results generated by the execution versions. At the end of the fault injection process, the final integrated scores are applied to the normalization box, and the weights are computed.
\begin{figure}[h]
\centering
\includegraphics[width=5 in]{Fig_weight.eps}
\caption{The details of our proposed weighting process based on a reward/punishment learning scheme through several fault injection scenarios}
\label{fig_weight}
\end{figure}
As this figure shows, based on comparing the results of each execution version to the correct output, its score is updated through a reward/punishment scheme. The reward of corrected results is determined as a positive score based on the complexity of the faulty situation. Similarly, the punishment is applied to the faulty results to decrease their score and affect the final output of the system. After applying all considered faults, the integrated final score of various executions is normalized and specified as the weight of each replicated operation.
It should be noted that the explained voting process is mainly implemented in software and enforces very limited complexity to hardware. Since the execution versions are distributed in time, they are implemented on the same hardware platform. Moreover, the weighted voting process is performed in software. However, if the hardware platform of the target system consists of redundant modules that could be utilized during the voting process, it will be appropriate and reduce the performance overhead. Since the target application of our proposed approach is embedded systems, extra hardware does not apply to them due to the cost and space constraints. Besides, extra hardware modules make the system more vulnerable to potential permanent faults. Thus, the hardware complexity of our proposed FT-EALU is very limited and could be expanded based on the requirements of the target system.
The details of the defined voting process and our proposed fault tolerance approach is presented in algorithm~\ref{alg_1}.
\begin{algorithm}[h]
\scriptsize
\caption{Pseudo-code of our proposed time redundancy-based fault tolerance approach (FT-EALU)}
\label{alg_1}
\textbf{// Replicate any operation three times in three various versions}
- Operation (F) on original input data (V1):
$R\_V1$ = F(A,B)
- Operation (F) on left shifted input data (V2):
AS = A \verb|<<| 1, BS = B \verb|<<| 1
$RS\_V2$ = F(AS,BS)
$R\_V2$ = $RS\_V2$ \verb|>>| 1
- Operation (F) on left shifted and swapped input data (V3):
Halve the input data to AL, AH, BL, BH
ALS = AL \verb|<<| 1, AHS = AH \verb|<<| 1, BLS = BL \verb|<<| 1, BHS = BH \verb|<<| 1
$RHS\_V3$ = F(ALS,BLS) , $RLS\_V3$ = F(AHS,BHS)
$RH\_V3$ = $RLS\_V3$ \verb|>>| 1 , $RL\_V3$ = $RHS\_V3$ \verb|>>| 1
$R\_V3$ = ($RL\_V3 | RH\_V3$)
\textbf{// Perform the proposed weighted voting among the results}
\textit{- Determine the corresponding score of each bit based on the proposed reward-punishment approach}
for t=1:N (number of fault injection scenarios)
\quad for i=1:input data length
\quad \quad if (three results are correct)
\quad \quad \quad $score(t)(V1_i)$ = $score(t)(V2_i)$ = $score(t)(V3_i)$ = $+\frac{1}{3}$;
\quad \quad if (two results are correct (i.e V1,V2) and one is faulty (i,e V3))
\quad \quad \quad $score(t)(V1_i)$ = $score(t)(V2_i)$ = +0.5 , $score(t)(V3_i)$ = -1;
\quad \quad if (two results are faulty (i.e V1,V2) and one is correct (i,e V3))
\quad \quad \quad $score(t)(V1_i)$ = $score(t)(V2_i)$ = -0.5 , $score(t)(V3_i)$ = +1;
\quad \quad else
\quad \quad \quad $score(t)(V1_i)$ = $score(t)(V2_i)$ = $score(t)(V3_i)$ = $-\frac{1}{3}$;
\quad end
end
\textit{- Determine the normalized weight of each bit of input data in various versions of operation}
$W_i(V1)$ = Standardize ( $\Sigma_t (score(t)(V1_i))/ N$ )
$W_i(V2)$ = Standardize ($\Sigma_t (score(t)(V2_i))/ N$ )
$W_i(V3)$ = Standardize ($\Sigma_t (score(t)(V3_i))/ N$ )
\textit{- Performing the weighted voting among various versions of operation to generate final output}
for i=1:input data length
\quad total $weight_i$ = ($W_i(V1)+W_i(V2)+W_i(V2)$)
\quad Result (i) = [$W_i(V1)$ * $R\_V1_i$ + $W_i(V2)$ * $R\_V2_i$ + $W_i(V3)$ * $R\_V3_i$] / total $weight_i$
\quad if (Result (i) $\geq$ 0.5)
\quad \quad Final Result(i) = 1
\quad else
\quad \quad Final Result(i) = 0
end
\end{algorithm}
\subsection{Illustrative Example}
\label{ill-example}
To show the workflow and details of our proposed fault-tolerance architecture (FT-EALU), in this section we apply it to an illustrative example. In this example, we focus on a four-bit input data with "stuck at one" permanent fault and the "add" operation as one of the key functions of ALU. As figure~\ref{fig-example} shows, in this example, the injected fault affects the second bit of data and makes it corrupted.
Here based on the explained approach, input data is applied to three distinct versions of operations: the original data, the shifted data, and the swapped-shifted data. Then the add operation is performed on the prepared input data and the replicated versions are executed in serial. The red bit in this figure represents the faulty bit and the result of each execution is passed to the voter, separately. The corresponding weight of each bit of data is prepared based on its capability in mitigating the injected fault.
In this example, the results of the majority and our proposed weighted voters are computed and compared. In the considered case, the final result of the second and third executions of this example are the same and the majority voter obeys them mistakenly.
Nevertheless in our proposed weighted voter, determining the score of each bit of results based on the defined trust factor leads to the correct result that covers the effect of injected fault, accurately.
In this example, the assigned weight of each bit of result is computed and normalized based on our proposed scoring algorithm and shown as \textit{W} vector in the figure. The final result is computed from performing a bitwise weighted sum on the assigned weights and results of each version of execution.
As this example shows, the overhead of our proposed fault tolerance approach is mainly related to the execution time of various versions of execution. Since the weighting process is mostly performed statically, it does not prolong the final execution time.
If the system hardware allows, modifying the input data for each version of execution could be performed in parallel with other executions. Eventually, this time overhead is managed by utilizing the extra time of the estimated deadline of real-time applications.
\begin{figure}[h]
\centering
\includegraphics[width=5.8 in]{Fig-example.eps}
\caption{An illustrative example showing the details of our proposed fault tolerance approach on a tiny sample case}
\label{fig-example}
\end{figure}
It is to be noted that, Our proposed \textit{FT-EALU} approach could be adaptively extended to five versions of operations and perform them voting among them like a 5MR system. In this case, the correction ability increases in the cost of more time overhead caused by employing more versions of execution. Adjusting the appropriate overhead and fault coverage levels are related to the application requirements and could be adaptively set by the designers based on the target system characteristics.
\section{Experimental Results}
\label{results}
In this section, to demonstrate the effectiveness of our proposed fault tolerance approach in making the ALU of embedded systems more reliable, several experiments and comparisons are presented.
\subsection{Simulation Setup}
To simulate and evaluate the proposed fault mitigation approach, our considered underlying platform is an ALU that implements basic operations such as "and", "or", "xor", "not", "add" and "subtract". The considered input data length is 4 and 16-bit and a comprehensive set of stuck-at faults are injected into them. Since our target underlying platform is an embedded system, the considered data length is appropriate and close to real cases but it could be easily extended to any other input length and fault pattern based on the target application requirements.
Our simulation platform implements the complete system architecture and focuses on permanent stuck-at faults that corrupt one or two bits of operands forever. The transient faults are fully mitigated during the time redundancy-based fault-tolerant approaches due to their short lifespan. Thus, we exclude transient faults from the experiments and focus on permanent ones. Since we study the effect of permanent faults on ALU and at the instructions level, they mainly affect input data or target operation. In both cases, the occurred fault modifies the value of data bits permanently. Stuck-at faults have a persistent effect on data bit and are proper to model the effect of permanent faults as they are employed in related research~\cite{FAZELI2011-emalu,ICEEabdi,gade2020run}.
In the following experiments, all permutations of single and double bit stuck at faults on 4-bit input data are considered. Moreover, all possible stuck-at faults (0 and 1) are applied to evaluate the proposed approach comprehensively.
Since considering all possible 16-bit numbers and applying all single and double bits stuck-at faults on them make the design space very large ($2^{21}$ various cases for single bit stuck at faults) which is not feasible in our simulation platform, we have limited them in our experiments. Here, selecting a limited number of random inputs is not representative so to make our experiments more precise and extended, we have selected one hundred random 16-bit numbers ten times (based on various randomness criteria) and applied all the possible stuck-at faults on them.
Our investigation and study show that in this case, one hundred random numbers is an appropriate choice that reflects the intended properties of the system. Figure \ref{fig-100-16bit} shows that by increasing the number of selected 16-bit numbers, the correction capability pattern of our proposed FT-EALU is not changed, and it nearly converges. The elbow point of this figure is in "one-hundred" point, and we consider it in our experiments. It should be noted that, due to applying all possible stuck-at faults on the considered data set, its size grows dramatically. Thus, the processing power of our underlying platform does not let us extend this study to more than 300 samples. However, the efficiency of our choice could be concluded from the final constant pattern of this figure.
\begin{figure}[h]
\centering
\includegraphics[width=5.8 in]{16-bit.eps}
\caption{The correction capability of FT-EALU for different sizes of input data set selected from all permutations of 16-bit numbers}
\label{fig-100-16bit}
\end{figure}
As this figure shows, selecting one hundred 16-bit samples is representative and reflects the correction capability of our proposed method appropriately.
The proposed approach and the simulation platform are implemented by Python. In this context, a high-level model of ALU that is capable of performing the main logical and arithmetic operations is simulated and various combinational of faulty input data are applied to it. The simulated platform is implemented in Jupyter Notebook environment on an Intel Core i7 CPU with 8 GB RAM. During the simulation, we have implemented the considered operations of ALU, and some built-in modules such as math, random, and copy are utilized.
\subsection{Experiments and discussion}
To demonstrate the effectiveness of our proposed fault mitigation approach and compare it to the related research, in this section three distinct categories of experiments are considered as: (I) selecting the most proper variations of input data to enhance the correction capability of our proposed fault-tolerant approach by providing more variations in executive versions and cover the fault more efficiently, (II) evaluating the efficiency of our proposed voting approach in covering the permanent faults and determining the appropriate weights statically or permanently based on a learning-based strategy, (III) assessing the overhead of our proposed approach and comparing its correction capability to the related fault-tolerant approaches.
\textit{A. Selecting the proper combination of execution versions for the replicas}
The first experiment aims at determining the most proper combinations of input data variations that lead to the best correction coverage.
As it has been explained, in time redundancy-based approaches, it is required to make some variations in the replicated blocks of the system to make appropriate diversity to detect permanent faults.
Low-cost and lightweight operators are the best choice for making the required variations and improving the correction ability. However, the precise selection and combination of replicated versions are very important. Considering the similar operands in various versions of execution leads to repeated operations and analogous results that reduce the correction ability during the voting process.
To study the effect of combining different versions of execution in our proposed time redundancy-based approach, we evaluate it with various combinations of diversified operations. The modifications are mainly based on applying simple and low-cost operators such as shift, swap, and rotate on the input data of the operations.
In this context we consider seven distinct versions of operands for each operation: original input data, shifted input data (RESO)~\cite{RESO}, swapped input data (RESWO)~\cite{RESWO}, enhanced shifted input data (E-RESO)~\cite{ICEEabdi}, shifted and swapped input data (E-RESWO)~\cite{ICEEabdi} and two rotated input data approaches.
These modifications of operands in replicated executions are required to preserve and propagate the effect of injected faults in their corresponding results. In this way, the results of replicated executions are not similar and the occurred fault will be masked during the voting process.
In our TMR-based approach, the first version of execution is performed on original input data and the mentioned variations are applied to the replicated operations. Based on applying the mentioned candidate variations on input data of the replicated execution versions, six meaningful cases are formed. Here, applying similar variations such as shift and rotate in combinations is not considered due to their analogous results that reduce the correction ability.
The correction capability of the defined combinations of executions is presented and compared in table~\ref{table-exp1}.
The reported correction coverage of this table is computed in average for all considered arithmetic and logical operations of our simulated ALU.
\begin{table}[h]
\caption {Investigating the correction capability of different combinations of variation in input data}
\centering
\begin{tabular}{|l|c|c|}
\hline
\multirow{2}{*}{\textbf{Correction Approach}} & \multicolumn{2}{|c|}{\textbf{Correction Coverage}} \\
\cline{2-3}
& \textbf{4-bit} & \textbf{16-bit} \\ \hline
Basic operands $\vert$ RESO $\vert$ RESWO & 79.06\% & 80.17\% \\ \hline
Basic operands $\vert$ E-RESO $\vert$ E-RESWO & 60.28\% & 54.41\% \\ \hline
Basic operands $\vert$ Rotated RESO $\vert$ E-RESWO & 78.68\% & 80.93\% \\ \hline
Basic operands $\vert$ Rotated E-RESO $\vert$ E-RESWO & 43.99\% & 59.76\% \\ \hline
Basic operands $\vert$ Rotated E-RESO $\vert$ RESWO & 37.45\% & 32.33\% \\ \hline
Basic operands $\vert$ RESO $\vert$ E-RESWO & 80.35\% & 82.38\% \\ \hline
\end{tabular}
\label{table-exp1}
\end{table}
As this table shows, combining different versions of executions in replicated versions lead to various correction capability.
Based on this table, combining the basic operands with their shifted (RESO) and shifted-swapped (E-RESWO) versions lead to the best correction coverage. Thus, our proposed FT-EALU considers this combination of execution that provides appropriate diversions in the results. As the result of this diversion, the effect of injected fault is kept in the intermediate results and then covered successfully during the voting process.
Moreover, based on the presented results, the correction coverage for 4 and 16-bit input data are almost matching which confirms the extension capability of this study.
It should be noted that the correction coverage is not the same for all ALU operations. The coverage of arithmetic operations is lower than logical ones due to their complexity.
Figure~\ref{fig-exp1} shows the correction capability of the studied combinations of table~\ref{table-exp1} categorized to different ALU operations. In this figure, the correction approaches are presented in the order of table~\ref{table-exp1}.
\begin{figure}
\centering
\includegraphics[width=5.8 in]{exp_1.eps}
\caption{Correction capability of various combinations of operands categorized by ALU operations}
\label{fig-exp1}
\end{figure}
As this figure shows, the effectiveness of correction approaches in mitigating faults of logical operations is better than the arithmetic functions such as add and subtraction. Since arithmetic operations are more complex and potential of error propagation due to the carry and borrow transfer among data bits.
Our proposed fault-tolerance approach, considered as "Correction Approach 6" in this figure, shows the best coverage capability in all ALU operations due to its proper variations of input data and considering the effect of the carry bit on data swap and shifting.
It should be noted that the presented results of this figure are based on 4-bit data and the same pattern is valid for other data sizes based on our experiments.
Moreover, in this experiment, the basic majority voter is employed to show the effectiveness of various combinations of input data for different ALU operations.
\textit{B. Evaluation of the proposed weighted voter}
The second experiment aims at demonstrating the effectiveness of our proposed weighted voting approach. Until now we have employed majority voting approach in our experiments.
Sometimes in complex cases, various versions of executions propagate the errors to the same position of system output, so considering majority voting is unreliable. Our proposed weighted voting approach evaluates the fault detection capability of various versions of computation and determines a weight for each of them statically. This weight represents the trust coefficient during the voting process and is determined per bit of input data. In this way the corresponding results of various versions of executions are not the same for the voter and their effect on final output is determined based on their trust factor.
In this experiment, the fault tolerance ability of conventional TMR that utilizes majority voters is compared to our proposed weighted voter's schemes. The weights of various executions in our proposed voter are determined statically through a learning-based approach. This weight represents their capability in error detection that is derived from their effectiveness in various cases.
As it has been explained in section~\ref{proposed-voter}, the first weighting approach is designed based on punishment of the incapable executions that propagate the errors and the second scheme employs reward and punishment schemes simultaneously. More capable approaches get higher weights and their corresponding result is more effective in final output due to their higher trustworthiness.
Figure~\ref{fig-exp2} shows the robustness of the considered execution versions in FT-EALU method separated by each bit of data for 4-bit input samples. In this figure, the average effect of single and double bit stuck at faults on input data during the add operation is studied. The effectiveness of various versions of execution in logical operations is very high and close to each other. Thus, to show the different trustworthiness of input bits more clearly, this experiment focuses on add operation which is more complicated.
\begin{figure}
\centering
\includegraphics[width=5 in]{exp_2.eps}
\caption{Robustness of different bits of various versions of input data against single and double stuck at faults during add operation (Bit 0: LSB and Bit 3: MSB)}
\label{fig-exp2}
\end{figure}
Based on this figure, the robustness of different bits of input data against injected faults are not the same. Since some faults are covered during the operation and do not propagate to the output. These faults are not our target in our proposed fault-tolerant approach but they improve our system reliability. We try to focus on uncovered faults but the correction capability of each version of execution could be an appropriate guide for determining our target trust factor.
Based on figure~\ref{fig-exp2}, our considered execution versions in FT-EALU provide various robustness on input data and this could be utilized as the trust factor of each bit of result, laterally.
The reason behind this difference in robustness of input data bits is the effect of their employed variations. For instance, the most vulnerability of the shifted and swapped input data is in bit 2 that has the constant zero value due to shift operation. This constant value makes the data more vulnerable against stuck at faults.
The presented results of this experiment, verify that the reliability of different execution versions is not the same so applying them in the final result with various weights could be beneficial in fault correction.
As it has been explained, our weighted voting considers two different approaches to determine the emphasis of each version of execution in the final result. Moreover, during the weighting process, to have tuned final results, it is required to consider a proper normalization approach.
Linear normalization is the simplest approach for adjusting the computed weights of our proposed voters but we do not limit our study to it. We consider various efficient normalization approaches and study their effects on the correction capability of our proposed method.
In this experiment, the correction capability of our proposed voting approach in both proposed weighting schemes considering three different normalization methods: 1) linear (min-max) normalization, 2) adding the weight of each bit of data to the absolute minimum value of that bit in all experiments and 3) standard normalization are investigated.
The average results of this study for all operations of our considered ALU, are presented in table \ref{table-exp2}.
\begin{table}[h]
\caption {Correction capability of our proposed weighted voting approach considering the proposed punishment-based and punish-reward weighting approaches normalized by various normalization schemes}
\centering
\resizebox{\columnwidth}{!}{\begin{tabular}{l c c c c}
\hline
& \multicolumn{3}{c}{\textbf{Correction Coverage}} \\
\cline{2-4}
& \textbf{Normalization 1} & \textbf{Normalization 2}& \textbf{Normalization 3} \\ \hline
Proposed voting scheme 1 (Punishment-based) & 81.76 \% & 78.22 \% & 82.57 \% \\ \hline
Proposed voting scheme 2 (Reward-Punishment) & 82.37 \% & 80.23\% & 83.61\% \\ \hline
\end{tabular}}
\label{table-exp2}
\end{table}
As this table shows, employing the weighted voter and increasing the effect of more reliable executions in the final output, leads to better correction capability than majority voting.
As has been expected, the proposed punishment-based weighting approach is less efficient than the reward-punishment method. Since the punishment-based approach only reduces the effect of unreliable executions, unidirectional. However, the punish/reward weighting approach, adjust the weights in both incremental and decreasing sides and reinforces the effect of robust results in final outputs.
Until now, the corresponding weights of each execution are determined dynamically according to the system feedback. Since our target application is autonomous and real-time embedded systems, it is important to reduce performance overhead and estimate the weight of each operation statically. As it has been explained earlier, the weighting process could be more efficient and performed statically through a learning-based approach. To this aim, the corresponding and proper weights for each bit of various versions of execution are determined based on several random stuck-at faults.
To show the details of this learning approach and determine the appropriate weights for each bit of operands, in this experiment, a dataset including several randomly stuck at faults along with the errors that aim some vulnerable bits of data are considered.
Vulnerable bits are the ones that are at risk due to our considered variation techniques. For instance, the middle bits of data that are affected by the "swap" operations or the MSB and LSB that are modified by "shift" are considered vulnerable bits.
Propagating the stuck-at faults on these vulnerable bits to the final output is more probable and their correction is harder.
Our considered data set for the learning approach is formed of several random input data and various injected stuck at faults.
During the learning process, the data set is ten-folded and nine parts are employed for training and the remaining one fold is used to test the extracted values.
Table~\ref{table-exp3} shows the correction capability of FT-EALU based on the weights that are extracted through the described learned process. In this experiment, the weights are computed by the reward/punishment scheme due to its efficiency.
It should be noted that in this experiment, several random and specified stuck-at faults on single and double bits of 16-bit input data
are considered.
To perform the learning approach more efficiently, the dataset should be as large as possible and contain several faults. By considering more samples, the training phase of learning is implemented more efficiently. Thus, in this experiment, we expand our dataset to 400 random 16-bit data samples, and all permutations of single and double bits stuck-at faults are applied to it.
Moreover, in table~\ref{table-exp3}, the effect of defined normalization schemes are also studied.
\begin{table}[h]
\small
\caption{Correction capability of FT-EALU based on the static weights of voter determined by a ten-folding cross validation approach}
\centering
\begin{tabular}{c|c|c|c|c|c|c|}
\cline{2-7}
& \multicolumn{2}{c|}{\textbf{Normalization 1}} & \multicolumn{2}{c|}{\textbf{Normalization 2}} & \multicolumn{2}{c|}{\textbf{Normalization 3}} \\
\cline{2-7}
& 1-bit SA & 2-bit SA & 1-bit SA & 2-bit SA & 1-bit SA & 2-bit SA \\ \hline
\multicolumn{1}{|c|}{\textbf{FT-EALU}} & 82.09\% & 68.83\% & 83.42\% & 68.88\% & 84.93\% & 69.71\% \\
\hline
\end{tabular}
\label{table-exp3}
\end{table}
As this table shows, the extracted weights provide appropriate correction coverage and are applicable to utilize as predefined weights in other unknown scenarios. Based on the results, the standard normalization has the best efficiency and could be integrated with our proposed weighting approach. These results are extracted based on various cases and could be simply utilized in other data lengths and various applications.
To demonstrate the efficiency of our proposed learning approach in determining appropriate weights, its correction capability is compared to the results of table~\ref{table-exp2} that is extracted the weights directly from the input data. By applying the extracted weights of learning to the input data of previous experiment, the correction capability differs about $0.51\%$. Thus our learning-based weighting approach could be efficiently employed in any application by its predefined weights. This process reduces the performance overhead of our proposed fault tolerance approach dramatically and makes it more applicable for real-time applications.
\textit{C. Comparison of FT-EALU to related approaches}
The last class of experiments compares the correction capability and performance overhead of our proposed fault tolerance approach to the related methods.
To this aim, the information redundancy-based method of~\cite{berger-Tabriz}, hardware reuse approach of~\cite{FAZELI2011-emalu} and TMR-based method of~\cite{gade2020run} that utilizes the basic time redundancy schemes are considered.
In~\cite{berger-Tabriz} the fault mitigation of ALU is performed through an enhanced version of Berger code that modifies the operations and operands to make the show of fault and tolerate it during the computation. In~\cite{FAZELI2011-emalu} a hardware redundancy-based approach to duplicate the operations is considered and its cost is moderated through the idle module reuse approach. The time redundancy-based approach of~\cite{gade2020run}, protects ALU through replicated operations in time and performs a majority voting among the results.
The comparison results of our proposed FT-EALU to these approaches in terms of correction capability and performance overhead are presented in table~\ref{table-exp4}.
Since the reported results of the considered related studies are mostly based on addition, in this experiment we have limited the ALU operation to "add". Moreover, the overhead of related studies is different due to their various redundancy schemes and. Hence, to have a fair and high-level comparison, we have not reimplemented the related methods and presented their reported correction capability and overhead.
\begin{table}
\caption{Comparison of our proposed FT-EALU to related studies in terms of fault correction capability and performance overhead for "add" operation}
\centering
\resizebox{\columnwidth}{!}{\begin{tabular}{l|c|c}
\hline
Fault Tolerance Approach & Average Correction Coverage & Performance overhead over baseline\\
\hline
Information redundancy approach~\cite{berger-Tabriz} & 46.2\% & 16.3\% \\
Hardware reuse approach~\cite{FAZELI2011-emalu} & 38.4\% & 9.38\% \\
TMR-based time redundancy approach~\cite{gade2020run} & 48.7\% & 35.99\% \\
FT-EALU & 52.5\% & 37.74\% \\
\hline
\end{tabular}}
\label{table-exp4}
\end{table}
As this table shows, the correction capability of our proposed FT-EALU outperforms the related approaches by $9.4\%$ on average. Since our research aims at critical embedded systems, their correction capability and robustness in presence of permanent faults are the main objectives. Besides the system constraints such as application deadline, cost and power consumption should be considered.
As it has been explained in section~\ref{proposed method}, to meet these constraints our proposed weighted voting unit is implemented in software and the weights are determined statically through a learning-based scheme. Moreover, the performance overhead of our proposed approach consists of triplicating each operation and performing a bit-wise voting among the results considering the determined weights of the learning approach.
Based on our implementation, the performance overhead of FT-EALU is about $37.74\%$ in cost of $84.93\%$ correction capability for all ALU operations which is reasonable in critical embedded systems that require high reliability.
This performance overhead is related to the replicated execution of various versions of operations in time and implementing the proposed weighted voter in software. If there exist redundant modules in the hardware platform of the target system, this cost could be reduced by implementing some of the mentioned operations in parallel by hardware.
Moreover, our proposed approach does not enforce extra hardware costs due to its replication in time. Since applying redundant hardware is not applicable in embedded systems and makes the system more vulnerable to faults, our proposed weighted voter is implemented in software.
Finally, the correction capability of our proposed approach could be improved by upgrading it to 5MR-based architecture. If the requirements of the target system in terms of timing constraints allow, it is possible to improve the correction capability of FT-EALU. In this context and based on our experiments, the correction capability of FT-EALU could be improved to $87.3\%$ in cost of $68.59\%$ performance overhead. Thus, the existing trade-off between correction capability and performance overhead of FT-EALU could be adjusted based on the system requirements and the target application.
\section{Conclusion Remarks}
In this paper, a new fault-tolerance approach for ALU of embedded and real-time systems is presented. Due to the expanded applications of embedded systems, their precise and autonomous operation is critical. The arithmetic and logic unit (ALU) as the main part of the processor leads the basic operations of the system and affects its final output directly. Thus, our proposed fault tolerance approach (FT-EALU) aims at making the ALU reliable through a time redundancy-based scheme. FT-EALU mitigates transient faults by replicating the execution versions in time. Moreover, it covers permanent faults by applying variations in input data of replicated computations in time. The mentioned variations are applied on input data by simple modifications such as shift and swap that improve the correction capability along with forcing low-performance overhead to the system.
To mitigate the faults and determine the final result, FT-EALU performs weighted voting among the serial replicated executions. In this way, the effect of the occurred faults will be handled during the operation and autonomously as it is expected in embedded applications.
The efficiency of the considered weighted voting approach is tightly related to determining the weights. In FT-EALU, this weight is specified based on the fault mitigation capability of various versions of execution and through an adaptive reward/punishment approach. The weights are determined statically and during system design based on several fault injection scenarios. The extracted weights are extracted for each bit of data to improve the precision of our proposed approach. Then, these weights are applied to the results of the execution versions separately and the final output is generated by their integration.
Several experiments are performed to demonstrate the effectiveness of our proposed fault tolerance approach considering the main ALU operations and many fault injection cases. Experimental results show that our proposed approach is capable of correcting $84.93\%$ of the injected single bit stuck-at faults on arithmetic and logical operations on average. The performance overhead of FT-EALU is mainly related to the serial executions and could be decreased by reusing the redundant hardware of the system.
As a future trend, applying more low-cost variations on operations and decreasing the time overhead of the proposed method by effective approaches such as approximate computing is proposed. Moreover, considering a dynamic threshold to determine the final result of the proposed weighted voter by evolutionary schemes could be interesting.
\bibliographystyle{unsrt}
|
\section{Introduction and main results}
\setcounter{equation}{0}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
Originally motivated by its application in telecommunications, stochastic filtering has been studied extensively since the early work of Stratonovich \cite{Stratonovic1960,Stratonovic1966} and Kushner \cite{Kushner1964,Kushner1967}.
The celebrated paper Fujisaki $et\; al$ \cite{Fujisaki1972} brings to a culmination
the innovation approach to non-linear filtering of diffusion processes. The
optimal filtering equation is a non-linear stochastic partial differential equation (SPDE), which is usually called the Kushner–Stratonovich equation or
the Kushner–FKK equation. The work of Kallianpur and Striebel \cite{kallianpur1968,kallianpur1969} establishes the representation
of the optimal filter in terms of the unnormalized one, which
was studied in the pioneering doctoral dissertations of Duncan \cite{duncan1967},
Mortensen \cite{mortensen1966} and the important paper of Zakai \cite{zakai1969}. The linear SPDE which the unnormalized filter satisfies is called the Duncan–Mortensen–Zakai equation, or, simply, Zakai’s
equation. We refer the reader to the books of Liptser and Shiryaev (\cite{LS1, LS2}), Kallianpur \cite{K:c}, Xiong \cite{X}, and Bain and Crisan \cite{BC} for more detailed introduction to nonlinear filtering.
Recently, stochastic filtering has found various applications in mathematical finance. The observation
processes are usually the price of stocks or other securities and their derivatives. While the related quantities, such as the appreciation rates, are usually the ``signal'' which need to be estimated. We refer the reader to the papers of Lakner \cite{Lak}, Zeng \cite{Zen}, Brennan and Xia \cite{BX},
Xia \cite{Xia2}, Rogers \cite{Rogers}, Nagai and Peng \cite{NP}, Xiong and Zhou \cite{xiong2007}, Huang $et\; al$ \cite{HWW}, and Xiong $et\; al$ \cite{xiong2020} for some examples. A related topic is the so called optimal control under partial information which has been studied extensively. Here we mention a few works of Huang $et\; al$ \cite{HWX}, {\O}kesandal and Sulem \cite{os}, Wang $et\; al$ \cite{wang2013, wang2015}. We refer to the book of Wang $et\; al$ \cite{wang2018} for a detailed introduction to the afore mentioned topic.
A key assumption of the classical stochastic filtering is that we can perfectly model the signal and the observation processes. However, this assumption is not always true
in many application scenarios. Especially, model ambiguity is very common in mathematical finance,
see, for example, Chen and Epstein \cite{chen2002} and Epstein and Ji \cite{EJ}. The aim of this article is to study the filtering problem with model ambiguity.
For simplicity of notation, we consider the following filtering model with real valued signal and observation processes:
\begin{equation}\label{intro-filter}
\left\{\begin{array}{ccl}
dX_t&=&b(X_t)dt+\si(X_t)dW_t,\quad X_0=x,\\
dY_t&=&h(X_t)dt+dB_t,\quad Y_0=0,
\end{array}\right.
\end{equation}
where $(W_t,B_t)$ is a 2-dimensional standard Brownian motion defined on a filtered probability space $(\Omega,\mathcal{F},\cF_t,P)$, the coefficients $b,\si$ and $h$ are continuous real functions. For simplicity of notation, we assume that $\si(x)\ge 0$ for all $x\in\RR$. The signal process $X_t$, or a function $f(X_t)$ of it, is what we want to estimate and the observation process $Y_t$ provides the information we can use. Namely, if the model is without ambiguity, we look for a $\cG_t\equiv\si(Y_s:\;s\le t)$-adapted process $(u_t)$ such that $\EE\int^T_0|f(X_t)-u_t)|^2dt$ is minimized.
\begin{definition}
A control process $u_t$ is called admissible if it is $\cG_t$-adapted and square integrable in the sense that
\[\EE\int^T_0u_t^2dt<\infty.\]
The set of all admissible controls is denoted by $\cU_{ad}$.
\end{definition}
We denote by $C_b(\RR^d)$ the set of all bounded and continuous mappings on $\RR^d$ and $L_\FF^2(0,T;\RR^d)$ the set of all $\RR^d$-valued $\cF_t$-adapted process $X$ such that
\[
\EE\left[\int_0^T|X_s|^2ds\right]<\infty.
\]
Let $\cP$ be a class of probability measures which is defined as
\begin{equation}\label{def-cP}
\cP=\left\{Q\sim P: \frac{dQ}{dP}=\exp\(\int^T_0\th_sdW_s-\frac12\int^T_0\th_s^2ds\),\;\;|\th_s|\le k\right\}.
\end{equation}
where $k$ is a nonnegative constant. The model ambiguity means that the true probability measure is one taken from $\cP$. This is also equivalent to drift ambiguity because by Girsanov formula, $\widetilde{W}_t\equiv W_t-\int^t_0\th_sds$ is a Brownian motion and, under $Q$, $X_t$ is a diffusion process with drift coefficient $b+\si\th$. For this model ambiguity, a large deviation principle is studied in Chen and Xiong \cite{chen2012}.
Because of the model ambiguity, we consider the square error in the worst case scenario.
\begin{definition}
For $f\in C_b(\RR)$, an admissible control $\bar{u}$ is called the ambiguity filter of $f(X_t)$ if
$J(\bar{u})=\inf_{u\in\cU_{ad}}J(u)$, where
\[J(u)=\sup_{Q\in\cP}\EE^Q\int^T_0|f(X_t)-u_t|^2dt.\]
\end{definition}
Next, we present the main results of this paper. In the first main result, we establish the existence and uniqueness of the
ambiguity filter. To this end, we need to make the following
\begin{hypo*}[H1]
The functions $b,\sigma, h$ are continuously differentiable with respect to $x$ and their partial derivatives $b_x,\si_x,h_x$ are uniformly bounded.
\end{hypo*}
\begin{thm}\label{thm1.1}
Let $f\in C_b(\RR)$ and suppose that (H1) holds. Then, there exists a unique ambiguity filter.
\end{thm}
We proceed to characterizing the ambiguity filter which is the second main result of this article.
For each $Q\in\cP$, we define another probability measure $\widetilde{Q}$ such that $\widetilde{Q}\sim Q$ with Radon-Nikodym derivative given by
\begin{equation}\label{eq0905b}
\left.\frac{d\widetilde{Q}}{dQ} \right|_{\cF_t}=M_t^{-1}\equiv\exp\(-\int^t_0h(X_s)dB_s-\frac12\int^t_0h(X_s)^2ds\),\end{equation}
as long as Novikov's condition holds. Note that, under the probability measure $\widetilde{Q}$, $Y_t$ is a Brownian motion independent of $\widetilde{W}_t$, and
\begin{equation}\label{eq1102b}
dM_t=h(X_t)M_tdY_t,\qquad M_0=1.
\end{equation}
Let $(p^i_t,q^{ij}_t,\;i,j=1,2)$ be the solution of the following backward stochastic differential equation (BSDE):
\begin{equation}\label{eq0905a}
\left\{
\begin{array}{ccl}
dp^1_t
&=&-\Big\{p^1_t(b'(X_t)+\si'(X_t)\th_t)+h'(X_t)M_tq^{12}_t+q^{21}_t\si'(X_t)\\
&&-f'(X_t)M_t(f(X_t)-\frac{Z^1_t}{Z^2_t})\Big\}dt+q^{11}_tdY_t+q^{21}_td\widetilde{W}_t\\
dp^2_t
&=&-\Big\{q^{12}_th(X_t)-\frac12(f(X_t)-\frac{Z^1_t}{Z_t^2})^2\Big\}dt+q^{12}_tdY_t+q^{22}_td\widetilde{W}_t\\[0.5em]
p^1_T&=&0,\quad p^2_T=0.
\end{array}
\right.
\end{equation}
with $Z_t^1=\widetilde{\EE}(f(X_t)M_t|\cG_t)$ and $Z_t^2=\widetilde{\EE}(M_t|\cG_t)$,
where $\tilde{\EE}$ denotes the expectation with respect to probability measure $\tilde{Q}$.
\begin{thm}\label{thm1.2}
For each $\th$ in (\ref{def-cP}) fixed, the forward and backward differential equation (FBSDE) (\ref{intro-filter}, \ref{eq1102b}, \ref{eq0905a}) has a unique solution. Further, the optimal ambiguity filter is given by
\begin{equation}\label{eq-u-frac}
u_t=\frac{{\tilde{\EE}}(f(X_t)M_t|\cG_t)}{{\tilde{\EE}}(M_t|\cG_t)},
\end{equation}
with $\th_t=k \mbox{ sgn}(p^1_t)$.
\end{thm}
The rest of this article is organized as follows. In Section 2, we derive the stochastic maximum principle (SMP) for an optimal control problem for a kind of conditional mean-field FBSDE system. Besides its own interests, this result will be used in characterizing the ambiguity filter. Theorem \ref{thm1.1} is proved in Section 3. Section 4 is devoted to the characterization of the ambiguity filter by converting the filtering problem to a conditional mean field optimal control problem.
\section{Control for conditional mean-field FBSDE}\label{sec2}
\setcounter{equation}{0}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
Mean field optimal control problem has been studied extensively. We mention only a few papers here: Bensoussan et al. \cite{B2015}, Buckdahn et al. \cite{Buckdahn2017,B2017}, Nguyen et al. \cite{Nguyen2020}, Guo and Xiong \cite{guo2018}. In this section, we extend the stochastic maximum principle to the case of conditional mean field problem to suit our purpose. The result in this section is of interest on its own.
Denote $U=[-k,k]$. Let $c,\ \si^1,\ \si^2:\ \RR^{d_1}\times\RR^{d_1}\times U\to\RR^{d_1}$,
$g:\RR^{d_1}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_1}\times U\to\RR^{d_2}$ and $\eta:\ \RR^{d_1}\times\RR^{d_2}\to\RR^{d_1}$ be measurable mappings.
We consider the following conditional mean-field type forward-backward stochastic control system:
\begin{equation}\label{eq-fbsde-xy}
\left\{\begin{array}{ccl}
dX_t&=&c(t,X_t,\EE(X_t|\cF_t^0),u_t)dt+\si^1(t,X_t,\EE(X_t|\cF_t^0),u_t)dW^1_t\\
&&+\si^2(t,X_t,\EE(X_t|\cF_t^0),u_t)dW^2_t,\\
dY_t&=&-g(t,X_t,Y_t,Z^1_t,Z^2_t,\EE(\eta(X_t,Y_t)|\cF^0_t),u_t)dt+Z^1_tdW^1_t+Z_t^2dW^2_t,\\
X_0&=&x, \quad Y_T=\Psi(X_T),
\end{array}\right.
\end{equation}
where $W^i_t$, $i=1,2$, are independent Brownian motions, $x\in\RR^{d_1}$, $\cF^0_t$ is a sub-filtration of $\cF_t$. Let $\cZ_t$ be another sub-filtration of $\cF_t$. We require admissible control $u_t$ to be $\cZ_t$-adapted.
In this section, we seek optimal control $u$ to minimize the cost functional
\begin{eqnarray}\label{eq-j-cmfc}
J(u)&=&\EE\bigg[ \int^T_0 l(t,X_t,Y_t,Z^1_t,Z^2_t,\EE(\eta(X_t,Y_t)|\cF^0_t),u_t)dt+\Gamma(Y_0)\nonumber\\
&&\quad+\Phi\Bigl(X_T,\EE\bigl(\eta(X_T,Y_T)\mid\cF^0_T\bigr)\Bigr) \bigg],
\end{eqnarray}
where $l:[0,T]\times\RR^{d_1}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_1}\times U\to\RR$, $\Ga:\RR^{d_2}\to\RR$, $\Phi:\RR^{d_1}\times\RR^{d_1}\to\RR$.
In the rest of this paper, we use $\<\cdot,\cdot\>$ and $|\cdot|$ to denote the usual inner product and norm in an Euclidean space, respectively, and the matrix $A^*$ for the transpose of the matrix $A$.
\begin{hypo}\label{hypo-1}
The mappings $\phi(t,x,\bar{x},u)$, $\psi(t,x,y,z^1,z^2,\mu,u)$, $\eta(x,y)$, $\Phi(x,\mu)$, $\Psi(x)$ are continuously differentiable in $(t,x,\bar{x},u)$,$(t,x,y,z^1,z^2,\mu,u)$, $(x,y)$,$(x,\mu)$, $x$, respectively, where $\phi=c,\si^1,\si^2$ and $\psi=g,l$. Moreover, their partial derivatives $\phi_x,\phi_{\bar{x}},\phi_u,\psi_x,\psi_y,\psi_{z^1},\psi_{z^2},\psi_\mu, \psi_u, \eta_x,\eta_y,\Phi_x,\Phi_\mu$ and $\Psi_x$ are uniformly bounded.
\end{hypo}
\begin{hypo}\label{hypo-2}
The mappings $\phi(t,x,\bar{x},u)$, $\psi(t,x,y,z^1,z^2,\mu,u)$, $\eta(x,y)$, $\Phi(x,\mu)$, $\Psi(x)$ are bounded by $C(1+|x|+|\bar{x}|+|u|)$, $C(1+|x|+|y|+|z^1|+|z^2|+|\mu|+|u|)$, $C(1+|x|+|y|)$, $C(1+|x|+|\mu|)$ and $C(1+|x|)$, respectively. Here $C$ is a nonnegative constant.
\end{hypo}
\begin{thm}\label{th-uni-exist}
Under Hypotheses \ref{hypo-1}-\ref{hypo-2}, the FBSDE \eqref{eq-fbsde-xy} admits a unique adapted solution $(X_t,Y_t,Z^1_t,Z^2_t)\in L_\FF^2(0,T;\RR^{d_1}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_2})$ for each admissible control $u\in \cU_{ad}$.
\end{thm}
Similar proofs are shown in \cite{B2017,li2018} with coefficients independent of the conditional expectation $\EE[\eta(X_t,Y_t)|\cF_t^0]$. However, the extension to current case is straight forward, so we omit the proof.
Let $v(\cdot)$ be such that $u(\cdot)+v(\cdot)\in\cU_{ad}$. For any $\ep\in(0,1)$, by the convexity of $\cU_{ad}$, we see that $u+\ep v(\cdot)\in\cU_{ad}$. We denote $(X_t^{u+\ep v},Y_t^{u+\ep v},Z_t^{1,u+\ep v},Z_t^{2,u+\ep v})$ as the solution of \eqref{eq-fbsde-xy} along with the control $u_t+\ep v_t$. We assume $u$ to be the optimal control which minimizes the cost functional \eqref{eq-j-cmfc}, and $(X_t^u,Y_t^u,$ $Z_t^{1,u},Z_t^{2,u})$ the corresponding optimal state. For simplicity, we now denote $(X_t,Y_t,Z^1_t,Z^2_t)=(X^u_t,Y^u_t,Z^{1,u}_t,Z^{2,u}_t)$. We proceed to proving the convergence of $(X_t^{u+\ep v},Y_t^{u+\ep v}, Z_t^{1,u+\ep v},Z_t^{2,u+\ep v})$ to $(X_t,Y_t,Z_t^1,Z_t^2)$ and to establishing the rate of convergence.
\begin{lemma}\label{ineq-u+epv}
If the Hypotheses \ref{hypo-1}-\ref{hypo-2} hold, then there is a constant $C>0$ such that
\begin{eqnarray*}
&&\EE\int_0^T\big|X^{u+\ep v}_t-X_t\big|^2dt \le C\ep^2\EE\int_0^T|v_t|^2dt,\\
&&\EE\int_0^T\big|Y^{u+\ep v}_t-Y_t\big|^2dt \le C\ep^2\EE\int_0^T|v_t|^2dt,\\
&&\EE\int_0^T\big|Z^{1,u+\ep v}_t-Z^1_t\big|^2dt \le C\ep^2\EE\int_0^T|v_t|^2dt,\\
&&\EE\int_0^T\big|Z^{2,u+\ep v}_t-Z^2_t\big|^2dt \le C\ep^2\EE\int_0^T|v_t|^2dt.
\end{eqnarray*}
\end{lemma}
Proof: To simplify the notation, we denote
\begin{eqnarray*}
\phi(t)&=&\phi(t,X_t,\EE[X_t|\cF^0_t],u_t),\quad \phi=c,\si^1,\si^2,\\
\phi^{u+\ep v}(t)&=& \phi(t,X_t^{u+\ep v},\EE[X_t^{u+\ep v}|\cF^0_t],u_t+\ep v_t),\\
\psi(t)&=&\psi(t,X_t,Y_t,Z_t^1,Z_t^2,\EE[\eta(X_t,Y_t)|\cF^0_t],u_t),\quad \psi=g,l,\\
\psi^{u+\ep v}(t)&=&\psi(t,X_t^{u+\ep v},Y_t^{u+\ep v},Z_t^{1,u+\ep v},Z_t^{2,u+\ep v},\EE[\eta(X_t^{u+\ep v},Y_t^{u+\ep v})|\cF^0_t],u_t+\ep v_t),\\
\chi_t^\ep&=&\chi_t^{u+\ep v}-\chi_t,\quad \chi_t=X_t,Y_t,Z^1_t,Z^2_t,\\
\eta^\ep_t&=&\eta(X_t^{u+\ep v},Y_t^{u+\ep v})-\eta(X_t,Y_t).
\end{eqnarray*}
It is easy to see from \eqref{eq-fbsde-xy} that
\begin{equation}\label{eq-fbsde-xy-ep}
\left\{\begin{array}{ccl}
dX^\ep_t&=& \big(c^{u+\ep v}(t)-c(t)\big)dt+\big(\si^{1,u+\ep v}(t)-\si^1(t)\big)dW^1_t\\[0.5em]
&&+\big(\si^{2,u+\ep v}(t)-\si^2(t)\big)dW^2_t,\\[0.5em]
dY^\ep_t&=&-\big(g^{u+\ep v}(t)-g(t)\big)dt+Z^{1,\ep}_tdW^1_t+Z_t^{2,\ep}dW^2_t,\\[0.5em]
X^\ep_0&=&0, \quad Y^\ep_T=\Psi(X^{u+\ep v}_T)-\Psi(X_T).
\end{array}\right.
\end{equation}
The Burkholder–Davis–Gundy inequality yields
\begin{align*}
&\;\EE\Big[\sup_{t\in[0,T]}|X^\ep_T|^2\Big]\\
\le&\;3\EE\int_0^T\Big(|c^{u+\ep v}(s)-c(s)|^2+|\si^{1,u+\ep v}(s)-\si^1(s)|^2+|\si^{2,u+\ep v}(s)-\si^2(s)|^2\Big)ds\\
\le&\;9C_1\EE\int_0^T \Big(|X^\ep_s|^2+\EE\big[|X^\ep_s|^2\cF^0_s\big]+\ep^2|v_s|^2\Big)ds\\
\le&\;9C_1\EE\int_0^T \Big(2|X^\ep_s|^2+\ep^2|v_s|^2\Big)ds,
\end{align*}
where $C_1>0$ is a constant. By Gronwall's inequality, we then have
\begin{equation}\label{ineq-x-sup}
\EE\Big[\sup_{t\in[0,T]}|X^\ep_t|^2\Big]\le C_2\ep^2\EE\int_0^T|v_s|^2ds,
\end{equation}
where $C_2>0$ is a constant, which implies that
\begin{equation*}
\EE \int_0^T|X^\ep_t|^2dt\le C_2T\ep^2\EE\int_0^T|v_t|^2dt.
\end{equation*}
Further, applying It\^{o}'s formula to $|Y^\ep_t|^2$, we can derive that
\begin{align}\label{ineq-y-z}
&\EE\Big[|Y^\ep_t|^2\Big]+\EE\int_t^T\Big(|Z^{1,\ep}_s|^2+|Z^{2,\ep}_s|^2\Big)ds\nonumber\\
=&\;2\EE\int_t^T\<Y^\ep_s,g^{u+\ep v}(s)-g(s)\>ds+\EE\Big[\big(\Psi(X^{u+\ep v}_T)-\Psi(X_T)\big)^2\Big]\nonumber\\
\le &\;2C_3\int_t^T\Bigg\{\EE\Big[|Y^\ep_s||X^\ep_s|\Big]+\EE\Big[|Y^\ep_s|^2\Big]+\EE\Big[|Y^\ep_s||Z^{1,\ep}_s|\Big]+\EE\Big[|Y^\ep_s||Z^{2,\ep}_s|\Big]\nonumber\\
&+\EE\Big[|Y^\ep_s|\EE\big[|\eta^\ep_s|\big|\cF^0_s\big]\Big]+\EE\Big[|Y^\ep_s||\ep v_t|\Big]\Bigg\}ds+\EE\Big[|X^\ep_T|^2\Big]\nonumber\\
\le &\; C_3\int_t^T\Bigg\{2\EE\Big[|X^\ep_s|^2\Big]ds+(7+\frac 2\ga+\frac 2\ga)\EE\Big[|Y^\ep_s|^2\Big]+\ga\EE\Big[|Z^{1,\ep}_s|^2\Big]\nonumber\\
&+\ga\EE\Big[|Z^{2,\ep}_s|^2\Big]+\ep^2\EE\Big[|v_s|^2\Big]\Bigg\}ds+\EE\Big[|X^\ep_T|^2\Big],
\end{align}
where $C_3>0$ is a constant. Taking $\ga=\frac 1{2C_3}$ in (\ref{ineq-y-z}), it follows from \eqref{ineq-x-sup} that
\begin{align*}
&\EE\Big[\sup_{t\in[0,T]}|Y^\ep_t|^2\Big]\\
\le&\;C_3\int_0^T\Bigg\{2\EE\Big[|X^\ep_s|^2\Big]ds+(7+\frac 4\ga)\EE\Big[|Y^\ep_s|^2\Big]+\ep^2\EE\Big[|v_s|^2\Big]\Bigg\}ds+\EE\Big[|X^\ep_T|^2\Big]\\
\le&\;C_3\int_0^T\Bigg\{(7+\frac 4\ga)\EE\Big[|Y^\ep_s|^2\Big]+(2C_2T+C_2+1)\ep^2\EE\Big[|v_s|^2\Big]\Bigg\}ds.
\end{align*}
Hence, by Gronwall's inequality we can obtain that
\begin{equation}\label{ineq-y-sup}
\EE\Big[\sup_{t\in[0,T]}|Y^\ep_t|^2\Big]\le C_4\ep^2\EE\int_0^T|v_s|^2ds,
\end{equation}
which yields that
\begin{equation*}
\EE \int_0^T|Y^\ep_t|^2dt\le C_4T\ep^2\EE\int_0^T|v_t|^2dt.
\end{equation*}
Here $C_4>0$ is a constant. Plugging \eqref{ineq-y-sup} into \eqref{ineq-y-z}, it follows that
\begin{equation*}
\EE \int_0^T\Big(|Z^{1,\ep}_s|^2+|Z^{2,\ep}_s|^2\Big)ds\le C_4\ep^2\EE\int_0^T|v_t|^2dt.
\end{equation*}
\qed
We use the following notations for simplification which are denoted as the partial derivates of $\phi=c,\si^1,\si^2$ with respect to $\chi=x,\bar{x},u$, and $\psi=g,l$ with respect to $\la=x,y,z^1,z^2,\mu,u$ by
\begin{eqnarray*}
\phi_\chi(t)&=&\phi_\chi(t,X_t,\EE[X_t|\cF^0_t],u_t),\\
\psi_\la(t)&=&\psi_\la(t,X_t,Y_t,Z_t^{1},Z_t^{2},\EE[\eta(X_t,Y_t)|\cF^0_t],u_t).
\end{eqnarray*}
We introduce the variational equations:
\begin{equation}\label{var-x}
\left\{\begin{array}{lll}
dX^1_t&=&\Big(c_x(t)X^1_t+c_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+c_u(t)v_t\Big)dt\\[0.5em]
&&+\Big(\si^1_x(t)X^1_t+\si^1_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+\si^1_u(t)v_t\Big)dW^1_t\\[0.5em]
&&+\Big(\si^2_x(t)X^1_t+\si^2_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+\si^1_u(t)v_t\Big)dW^2_t,\\[0.5em]
dY^1_t&=&-\Big(g_x(t)X^1_t+g_y(t)Y^1_t+g_{z^1}(t)Z_t^{1,1}+g_{z^2}(t)Z_t^{2,1}+g_u(t)v_t\\[0.5em]
&&+g_\mu(t)\EE[\eta_x(t)X^1_t|\cF^0_t]+g_\mu(t)\EE[\eta_y(t)Y^1_t|\cF^0_t]\Big)dt\\[0.5em]
&&+Z^{1,1}_tdW^1_t+Z^{2,1}_tdW^2_t,\\[0.5em]
X^1_0&=&0,\quad Y^1_T=\Psi_x(X(T))X^1(T).
\end{array}\right.
\end{equation}
For $v$ being fixed as before, by Hypotheses \ref{hypo-1}-\ref{hypo-2}, it is easy to see that \eqref{var-x} admits a unique solution.
\begin{lemma}\label{lemma-x-ep}
If Hypotheses \ref{hypo-1}-\ref{hypo-2} hold and
\[\hat{\chi}^\ep_t=\frac{\chi^{u+\ep v}_t-\chi_t}{\ep}-\chi^1_t,\]
where $\chi=X,Y,Z^1,Z^2$, then,
\[\lim_{\ep\to 0}\EE\int_0^T \Big(|\hat{X}^\ep_t|^2+|\hat{Y}^\ep_t|^2+|\hat{Z}^{1,\ep}_t|^2+|\hat{Z}^{2,\ep}_t|^2\Big)dt=0. \]
\end{lemma}
Proof: It follows from \eqref{eq-fbsde-xy-ep} and \eqref{var-x} that
\begin{align}\label{eq:x-i*3}
&d\hat{X}^\ep_t=\Bigg\{\frac{1}{\ep}\Big(c^{u+\ep v}(t)-c(t)\Big)-\Big(c_x(t)X^1_t+c_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+c_u(t)v_t\Big)\Bigg\}dt\nonumber\\
&\;+\Bigg\{\frac{1}{\ep}\Big(\si^{1,u+\ep v}(t)-\si^1(t)\Big)-\Big(\si^1_x(t)X^1_t+\si^1_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+\si^1_u(t)v_t\Big)\Bigg\}dW^1_t\nonumber\\
&\;+\Bigg\{\frac{1}{\ep}\Big(\si^{2,u+\ep v}(t)-\si^2(t)\Big)-\Big(\si^2_x(t)X^1_t+\si^2_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+\si^1_u(t)v_t\Big)\Bigg\}dW^2_t,
\end{align}
with the initial $\hat{X}^\ep_0=0$ and
\begin{equation}\label{eq:y-i*3}
\left\{
\begin{aligned}
d\hat{Y}^\ep_t=&-\Bigg\{\frac{1}{\ep}\Big(g^{u+\ep v}(t)-g(t)\Big)-\Big(g_x(t)X^1_t+g_y(t)Y^1_t+g_{z^1}(t)Z_t^{1,1}+g_{z^2}(t)Z_t^{2,1}\\
&+g_\mu(t)\EE[\eta_x(t)X^1_t|\cF^0_t]+g_\mu(t)\EE[\eta_y(t)Y^1_t|\cF^0_t]+g_u(t)v_t\Big)\Bigg\}dt\\
&+\hat{Z}^{1,\ep}_tdW^1_t+\hat{Z}^{2,\ep}_tdW^2_t,\\
\hat{Y}^\ep_T=&\frac{1}{\ep}\left(\Psi(X^{u+\ep v}_T)-\Psi(X_T)\right)-\Psi_x(X(T))X^1_T.
\end{aligned}\right.
\end{equation}
Note that for $\chi=X,Y,Z^1,Z^2$ and a constant $\la\in[0,1]$,
\[
\chi^{u+\ep\la v}_t=\chi_t+\ep\la(\hat{\chi}^\ep_t+\chi^1_t).
\]
Then for $\phi=c,\si^1,\si^2$,
\begin{align}\label{eq-x-u+eplav}
&\frac{1}{\ep}(\phi^{u+\ep v}(t)-\phi(t))\nonumber\\
=&\int_0^1\Bigg\{\phi_x^{u+\ep\la v}(t)(\hat{X}^\ep_t+X^1_t)+\phi_{\bar{x}}^{u+\ep\la v}(t)\EE\big[(\hat{X}^\ep_t+X^1_t)|\cF^0_t\big]+\phi_u^{u+\ep\la v}(t)v_t\Bigg\}d\la,
\end{align}
\begin{align}\label{eq-y-u+eplav}
&\frac{1}{\ep}(g^{u+\ep v}(t)-g(t))\nonumber\\
=&\int_0^1\Bigg\{g_x^{u+\ep\la v}(t)(\hat{X}^\ep_t+X^1_t)+g_y^{u+\ep\la v}(t)(\hat{Y}^\ep_t+Y^1_t)+g_{z^1}^{u+\ep\la v}(t)(\hat{Z}^{1,\ep}_t+Z^{1,1}_t)\nonumber\\
&+g_{z^2}^{u+\ep\la v}(t)(\hat{Z}^{2,\ep}_t+Z^{2,1}_t)+g_\mu^{u+\ep\la v}(t)\EE[\eta_x^{u+\ep\la v}(t)(\hat{X}^\ep_t+X^1_t)|\cF^0_t]\nonumber\\
&+g_\mu^{u+\ep\la v}(t)\EE[\eta_y^{u+\ep\la v}(t)(\hat{Y}^\ep_t+Y^1_t)|\cF^0_t]+g_u^{u+\ep\la v}(t)v_t\Bigg\}d\la,
\end{align}
and
\begin{equation}\label{eq-Psi-u+eplav}
\frac{1}{\ep}\left(\Psi(X^{u+\ep v}_T)-\Psi(X_T)\right)=\int_0^1\Big[\Psi_x(X_T^{u+\ep\la v})(\hat{X}^\ep_T+X^1_T)\Big]d\la.
\end{equation}
Thus, an immediate consequence of some simple rearrangements of \eqref{eq:x-i*3}, \eqref{eq:y-i*3}, \eqref{eq-x-u+eplav}, \eqref{eq-y-u+eplav} and \eqref{eq-Psi-u+eplav} gives us that
\begin{equation*}
\left\{\begin{array}{ccl}
d\hat{X}^\ep_t&=&\Big(\hat{c}^{1}(t)\hat{X}^\ep_t+\hat{c}^{2}(t)X^1_t+\hat{c}^{3}(t)v_t\Big)dt\\
&&+\Big(\hat{\si}^{1,1}(t)\hat{X}^\ep_t+\hat{\si}^{1,2}(t)X^1_t+\hat{\si}^{1,3}(t)v_t\Big)dW^1_t\\
&&+\Big(\hat{\si}^{2,1}(t)\hat{X}^\ep_t+\hat{\si}^{2,2}(t)X^1_t+\hat{\si}^{2,3}(t)v_t\Big)dW^2_t,\\
d\hat{Y}^\ep_t&=&-\Big(\hat{g}^{1,1}(t)\hat{X}^\ep_t+\hat{g}^{1,2}(t)\hat{Y}^\ep_t+\hat{g}^{1,3}(t)\hat{Z}^{1,\ep}_t+\hat{g}^{1,4}(t)\hat{Z}^{2,\ep}_t+\hat{g}^{5}(t)v_t\\
&&+\hat{g}^{2,1}(t)X^1_t+\hat{g}^{2,2}(t)Y^1_t+\hat{g}^{2,3}(t)Z^{1,1}_t+\hat{g}^{2,4}(t)Z^{2,1}_t\Big)dt\\
&&+\hat{Z}^{1,\ep}_tdW^1_t+\hat{Z}^{2,\ep}_tdW^2_t,\\[0.5em]
\hat{X}^\ep_0&=&0,\quad \hat{Y}^\ep_T=\Psi^{1,\ep}_T\hat{X}^\ep_T+\Psi^{2,\ep}_TX^1_T,
\end{array}\right.
\end{equation*}
where for $\phi=c,\si^1,\si^2$,
\begin{eqnarray*}
\hat{\phi}^1(t)\hat{X}^\ep_t&=&\Big(\int_0^1\phi_x^{u+\ep\la v}(t)d\la\Big)\hat{X}^\ep_t+\Big(\int_0^1\phi_{\bar{x}}^{u+\ep\la v}(t)d\la\Big)\EE\big[\hat{X}^\ep_t|\cF^0_t\big],\\
\hat{\phi}^2(t)X^1_t&=&\Big(\int_0^1\phi_x^{u+\ep\la v}(t)d\la-\phi_x(t)\Big)X^1_t+\Big(\int_0^1\phi_{\bar{x}}^{u+\ep\la v}(t)d\la-\phi_{\bar{x}}(t)\Big)\EE\big[X^1_t|\cF^0_t\big],\\
\hat{\phi}^3(t)&=&\int_0^1\phi_u^{u+\ep\la v}(t)d\la-\phi_u(t),\\
\hat{g}^{1,i}_t\hat{A}^\ep_t&=&\Big(\int_0^1g_{A}^{u+\ep\la v}(t)d\la\Big)\hat{A}^\ep_t+\int_0^1\Big(g_{\mu}^{u+\ep\la v}(t)\EE[\eta_A^{(u+\ep\la v)}(t)\hat{A}^\ep_t|\cF^0_t]\Big)d\la,\\
\hat{g}^{2,i}_tA^1_t&=&\Big(\int_0^1g_{A}^{u+\ep\la v}(t)d\la-g_A(t)\Big)A^1_t+\int_0^1\Big(g_{\mu}^{u+\ep\la v}(t)\EE[\eta_A^{(u+\ep\la v)}(t)A^1_t|\cF^0_t]\Big)d\la\\
&&-g_\mu\EE[\eta_A(t)A^1_t|\cF^0_t)],\;\;i=1,2, \mbox{when } i=1, A=X; \mbox{when } i=2, A=Y,\\
\hat{g}^{1,j}(t)&=&\int_0^1g_{B}^{u+\ep\la v}(t)d\la,\;\; j=3,4,\mbox{ when } j=3, B=Z^1; \mbox{ when } j=4, B=Z^2,\\
\hat{g}^{2,j}(t)&=&\int_0^1g_{B}^{u+\ep\la v}(t)d\la-g_B(t),\\
\hat{g}^{5}(t)&=&\int_0^1g_{u}^{u+\ep\la v}(t)d\la-g_u(t),\\
\Psi_T^{1,\ep}&=&\int_0^1\Psi_x(X^{u+\ep\la v}_T)d\la,\quad \Psi_T^{2,\ep}=\int_0^1\Psi_x(X^{u+\ep\la v}_T)d\la-\Psi_x(X(T)).
\end{eqnarray*}
Combining with the above estimations and applying It\^{o}'s formula to $|\hat{X}^\ep_t|^2$, it
follows from the Burkholder–Davis–Gundy inequality that
\begin{eqnarray*}
\EE\Big[\sup_{t\in[0,T]}|\hat{X}^\ep_t|^2\Big]
&\le&K_1\EE\int_0^T |\hat{X}^\ep_t|^2dt\\
&&+K_1\EE\int_0^T\Big(|\hat{c}^{2}(t)|^2+|\hat{\si}^{1,2}(t)|^2+|\hat{\si}^{2,2}(t)|^2\Big)|X^1_t|^2dt\\
&&+K_1\EE\int_0^T\Big(|\hat{c}^{3}(t)|^2+|\hat{\si}^{1,3}(t)|^2+|\hat{\si}^{2,3}(t)|^2\Big)|v_t|^2dt,
\end{eqnarray*}
where $K_1>0$. By virtue of the continuity boundness of $c_x, c_{\bar{x}},c_u,\si^1_x,\si^1_{\bar{x}},\si^1_u,\si^2_x,\si^2_{\bar{x}}, \si^2_u$ and the Gronwall's inequality,
\begin{equation}\label{eq-x-hat-ep}
\lim\limits_{\ep\goto0}\EE\Big[\sup_{t\in[0,T]}|\hat{X}^\ep_t|^2\Big]=0.
\end{equation}
Next, using similar arguments to $|\hat{Y}^\ep_t|^2$, we can obtain
\begin{align*}
&\EE\Big[|\hat{Y}^\ep_t|^2\Big]+\EE\int_t^T\Big(|\hat{Z}^{1,\ep}_s|^2|+|\hat{Z}^{2,\ep}_s|^2\Big)\\
=&\;2\EE\int_t^T\Big<\hat{g}^{1,1}(s)\hat{X}^\ep_s+\hat{g}^{1,2}(s)\hat{Y}^\ep_s+\hat{g}^{1,3}(s)\hat{Z}^{1,\ep}_s+\hat{g}^{1,4}(s)\hat{Z}^{2,\ep}_s+\hat{g}^{5}(s)v_s\\
&\;+\hat{g}^{2,1}(s)X^1_s+\hat{g}^{2,2}(s)Y^1_s+\hat{g}^{2,3}(s)Z^{1,1}_s+\hat{g}^{2,4}(s)Z^{2,1}_s,\hat{Y}^\ep_s\Big>ds\\
&\;+\EE\Big[|\Psi^{1,\ep}_T\hat{X}^\ep_T+\Psi^{2,\ep}_TX^1_T|^2\Big]\\
\le&\;K_2\int_t^T\Big(\EE\big[|\hat{X}^\ep_t|^2\big]+\Big(4+\frac2{\ga}+\frac2{\ga}\Big)\EE\big[|\hat{Y}^\ep_t|^2\big]+\ga\EE\big[|\hat{Z}^{1,\ep}_t|^2\big]+\ga\EE\big[|\hat{Z}^{2,\ep}_t|^2\big]\Big)dt\\
&\;+K_2\EE\int_t^T\Big(|\hat{g}^{2,1}(s)|^2|X^1_s|^2+|\hat{g}^{2,2}(s)|^2|Y^1_s|^2+|\hat{g}^{2,3}(s)|^2|Z^{1,1}_s|^2\\
&\;+|\hat{g}^{2,4}(s)|^2|Z^{2,1}_s|^2+|\hat{g}^{5}(s)|^2|v_s|^2\Big)ds,
\end{align*}
where $K_2$ is a nonnegative constant. Taking $\ga=\frac 1{2K_2}$, by \eqref{eq-x-hat-ep}, the Gronwall inequality, the boundedness and the continuity of $g_x,g_y,g_{z^1},g_{z^2},g_\mu,g_u$, we then have
\begin{eqnarray*}
\lim_{\ep\to 0}\EE\Big[\sup_{t\in[0,T]}|\hat{Y}^\ep_t|^2\Big]=0,
\end{eqnarray*}
which yields that
\begin{equation*}
\lim_{\ep\to 0}\EE\int_0^T\Big(|\hat{Z}^{1,\ep}_t|^2+|\hat{Z}^{2,\ep}_t|^2\Big)dt=0.
\end{equation*}
Hence, combining with above estimates, we arrive at the desired conclusion.
\qed
To proceed, the next lemma is concerned with the perturbation of the cost functional $J(\cdot)$ with respect to the parameter $\ep$. The proof follows from the same arguments as those in last lemma so we omit it.
\begin{lemma}
If Hypotheses \eqref{hypo-1}-\eqref{hypo-2} hold, then
\begin{align}\label{eq-j-ep}
&\frac{d}{d\ep}J(u_t+\ep v_t)\Big|_{\ep=0}\nonumber\\
=&\EE\Bigg\lbrace\int_0^T\Big(l_x(t)X^1_t+l_y(t)Y^1_t+l_{z^1}(t)Z^{1,1}_t+l_{z^2}(t)Z^{2,1}_t+l_u(t)v_t\nonumber\\
&+l_\mu(t)\EE\big[\eta_x(t)X^1_t\big|\cF^0_t\big]+l_\mu(t)\EE\big[\eta_y(t)Y^1_t\big|\cF^0_t\big]\Big)dt\nonumber\\
&+\Ga_y(Y_0)Y^1_0+\Phi_x(T)X^1_T+\Phi_\mu(T)\EE\big[\eta_x(T)X^1_T\big|\cF^0_T\big]\nonumber\\
&+\Phi_\mu(T)\EE\big[\eta_y(T)\Psi_x(X_T)X^1_T\big|\cF^0_T\big]\Bigg\rbrace.
\end{align}
\end{lemma}
Next we introduce the adjoint equations
\begin{equation}\label{eq-p}
\left\{
\begin{aligned}
dp_t=&-\Big(c_x(t)^*p_t+\si^{1}_x(t)^*q^1_t+\si^{2}_x(t)^*q^2_t-g_x(t)^*\tilde{p}_t+l_x(t)\\
&+\EE\big[c_{\bar{x}}(t)^*p_t+\si^1_{\bar{x}}(t)^*q^1_t+\si^2_{\bar{x}}(t)^*q^2_t\big|\cF^0_t\big]\\
&+\eta_x(t)^*\EE\big[l_\mu(t)-g_\mu(t)^*\tilde{p}_t\big|\cF^0_t\big]\Big)dt+q^1_tdW^1_t+q^2_tdW_t^2,\\
d\tilde{p}_t=&\,\Big(g_y(t)^*\tilde{p}_t+\eta_y(t)^*\EE\big[g_\mu(t)^*\tilde{p}_t-l_\mu(t)\big|\cF_t^0\big]-l_y(t)\Big)dt\\
&+\Big(g_{z^1}(t)^*\tilde{p}_t-l_{z^1}(t)\Big)dW^1_t+\Big(g_{z^2}(t)^*\tilde{p}_t-l_{z^2}(t)\Big)dW^2_t,\\
p_T=&\,\Phi_x(T)+\Big(\eta_x(T)^*+\Psi_x(X_T)^*\eta_y(T)^*\Big)\EE\big[\Phi_\mu(T)|\cF^0_T\big]-\Psi_x(X_T)^*\tilde{p}_T,\\
\tilde{p}_0=&-\Ga_y(Y_0).
\end{aligned}
\right.
\end{equation}
where the notation $A^*$ stands for the transpose of matrix $A$. In view of Hypotheses \ref{hypo-1}-\ref{hypo-2}, \eqref{eq-p} admits a unique solution for each admissible control $u$.
We are now ready to present the main theorem in this section.
\begin{thm}[Stochastic Maximum Principle]
Suppose that Hypotheses \ref{hypo-1}-\ref{hypo-2} hold. Suppose that $u\in\cU_{ad}$ is a local minimum for $J(\cdot)$ in the sense that for all $v(\cdot)$ satisfying $v(\cdot)+u(\cdot)\in\cU_{ad}$, there exists an $\ga>0$ such that $u(\cdot)+\ep v(\cdot)\in\cU_{ad}$ for any $\ep\in(-\ga,\ga)$ and $J(u(\cdot)+\ep v(\cdot))$ attains its minimun at $\ep=0$. Then we have
\[\EE(H_u(t)|\cZ_t)=\left\{\begin{array}{ll}
=0,&\mbox{ if }|u_t|< k,\\[0.5em]
\le 0,&\mbox{ if }u_t= k,\\[0.5em]
\ge 0,&\mbox{ if }u_t=-k,
\end{array}\right.\]
where the Hamiltonian function $H: [0,T]\times\RR^{d_1}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_2}\times\RR^{d_1}\times U\times\RR^{d_1}\times\RR^{d_2} \times\RR^{d_1}\times\RR^{d_1}\goto \RR$ is defined by
\begin{align}\label{eq-hamilton}
H(t,x,y,z^1,z^2,\mu,u,p,\tilde{p},q^1,q^2) =& \big<p,c(t)\big>+\big<q^1,\si^1(t)\big>+\big<q^2,\si^2(t)\big> \nonumber\\
&+\big<\tilde{p},g(t)\big>+l(t).
\end{align}
\end{thm}
Proof:
Applying It\^o's formula to $\<p_t,X^1_t\>$ and $\<\tilde{p}_t,Y^1_t\>$ along with \eqref{eq-p} and \eqref{var-x}, we derive
\begin{align*}
dp^*_tX_t^1
=&\Bigg\{-\Big(p^*_tc_x(t)+(q^1_t)^*\si^1_x(t)+(q^2_t)^*\si^2_x(t)-\tilde{p}_t^*g_x(t)+l^*_x(t,u)\\
&+\EE\big[p_t^*c_{\bar{x}}(t)+(q^1_t)^*\si^1_{\bar{x}}(t)+(q^2_t)^*\si^2_{\bar{x}}(t)\big|\cF^0_t\big]\\
&+\EE\big[l_\mu(t)^*-(\tilde{p}_t)^*g_\mu(t)\big|\cF^0_t\big]\eta_x(t)\Big)X^1_t\\
&+p^*_t\Big(c_x(t)X^1_t+c_{\bar{x}}(t)\EE\big[X^1_t\big|\cF^0_t\big]+c_u(t)v_t\Big)\\
&+(q^1_t)^*\Big(\si^1_x(t)X^1_t+\si^1_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+\si^1_u(t)v_t\Big)\\
&+(q^2_t)^*\Big(\si^2_x(t)X^1_t+\si^2_{\bar{x}}(t)\EE[X^1_t|\cF_t^0]+\si^1_u(t)v_t\Big)\Bigg\}dt+d(martingale),
\end{align*}
and
\begin{align*}
d\tilde{p}^*_tY^1_t
=&\Bigg\{\Big(\tilde{p}_t^*g_y(t)+\EE\big[\tilde{p}_t^*g_\mu(t)-l_\mu(t)^*\big|\cF_t^0\big]\eta_y(t)-l_y(t)^*\Big)Y^1_t\\
&-\tilde{p}_t^*\Big(g_x(t)X^1_t+g_y(t)Y^1_t+g_{z^1}(t)Z_t^{1,1}+g_{z^2}(t)Z_t^{2,1}+g_u(t)v_t\\
&+g_\mu(t)\EE[\eta_x(t)X^1_t|\cF^0_t]+g_\mu(t)\EE[\eta_y(t)Y^1_t|\cF^0_t]\Big)\\
&\,+\Big(\tilde{p}_t^*g_{z^1}(t)-l_{z^1}(t)\Big)Z^{1,1}_t+\Big(\tilde{p}_t^*g_{z^2}(t)-l_{z^2}(t)\Big)Z^{2,1}_t\Bigg\}dt\\
&\,+d(martingale).
\end{align*}
Since
\[\EE\Big[\Phi_\mu(T)\EE\big[\eta_x(T)X^1_T\big|\cF^0_t\big]\Big]=\EE\Big[\EE\big[\Phi_\mu(T)\big|\cF^0_t\big]\eta_x(T)X^1_T\Big],\]
taking the integral on both sides of the above differential equations, we can obtian that
\begin{eqnarray*}
&&\EE\Big(p_T^*X^1_T+\tilde{p}^*_TY_T^1-p_0^*X^1_0-\tilde{p}^*_0Y_T^0\Big)\\
&=&\EE\Big[\Phi_x(T)X^1_T+\Phi_\mu(T)\EE\big[\eta_x(T)X^1_T+\eta_y(T)\Psi_x(X_T)X^1_T\big|\cF^0_T\big]\\
&&+\Ga_y(Y_0)Y^1_0\Big]\\
&=&\EE\int_0^T\Bigg\{-\Big(l_x(t)X^1_t+l_y(t)Y^1_t+l_{z^1}(t)Z^{1,1}_t+l_{z^2}(t)Z^{2,1}_t\Big)\\
&&-\EE\big[l_\mu(t)^*|\cF^0_t\big]\Big(\eta_x(t)X^1_t+\eta_y(t)Y^1_t\Big)\\
&&+\Big(p^*_tc_u(t)+(q^1_t)^*\si^1_u(t)+(q^2_t)^*\si^2_u(t)-\tilde{p}^*_tg_u(t)\Big)v_t\Bigg\}dt.
\end{eqnarray*}
Plugging it back into \eqref{eq-j-ep}, we get
\begin{align*}
&\EE\int^T_0\Big(l_u(t)+p^*_tc_u(t)+(q^1_t)^*\si^1_u(t)+(q^2_t)^*\si^2_u(t)+\tilde{p}^*_tg_u(t)\Big)v_tdt\\
=&\EE\int^T_0\big<\EE[H_u(t)|\cZ_t],v_t\big>dt\ge 0.
\end{align*}
Note that $v_t=u^0_t-u_t$ for $u^0_t\in\cU_{ad}$. Thus,
\[
\EE\int^T_0\big<\EE[H_u(t)|\cZ_t],u^0_t-u_t\big>dt\ge 0.
\]
Therefore, to ensure that the above inequality holds, we have $\EE(H_u(t)|\cZ_t)=0$ if $|u_t|<k$. If $u_t=k$, $u^0_t-u_t\le 0$. In this case, we must have $\EE(H_u(t)|\cZ_t)\le 0$. Similarly, if $u_t=-k$, $\EE(H_u(t)|\cZ_t)\ge 0$.
\qed
\section{Existence and uniqueness of the ambiguity filter}
\setcounter{equation}{0}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
In this section, we proceed to proving Theorem \ref{thm1.1}.
Denote by $V^Q_t$ as the conditional expectation of the total square error in time interval
$[t,T]$ with respect to an admissible measure $Q$:
\[V^Q_t=\EE^Q\bigg[\int^T_t|f(X_s)-u_s|^2ds\;\Big|\;\mathcal{F}_t\bigg],\quad 0\le t\le T.\]
It is easy to show that $V^Q_t+\int^t_0|f(X_s)-u_s|^2ds$ is a martingale under the probability measure $Q$. The martingale representation theorem implies that $V^Q_t$ is a solution to the backward stochastic differential equation (BSDE):
\[
\left\{\begin{array}{lll}
dV^Q_t&=& z_tdW_t+\tilde{z}_tdB_t-|f(X_t)-u_t|^2dt-z_t\th_tdt,\\[0.5em]
V^Q_T&=&0,
\end{array}\right.\]
where $z_t$ and $\tilde{z}_t$ are predictable processes. Note that the probability uncertainty only reflects on the drift, and hence, $z_t$ and $\tilde{z}_t$ do not depend on the probability measure $Q$.
Let
\[y_t=\sup_{Q\in\cP}V^Q_t.\]
It is easy to see that $J(u)=y_0$.
\begin{thm}
The process $y_t$ is the unique solution to the BSDE
\begin{equation}\label{eq:y_t}
\left\{\begin{array}{lll}
dy_t&=&\big(-|f(X_t)-u_t|^2+k|z_t|\big)dt+z_tdW_t+\tilde{z}_tdB_t, \\[0.5em]
y_T&=&0.
\end{array}\right.
\end{equation}
\end{thm}
Proof:
Note that
\[\max_{|\th_t|\le k}\th_tz_t=k|z_t|.\]
Then,
\begin{eqnarray*}
y_t &=& \sup_{Q\in\cP}V^Q_t=\sup_{Q\in\cP}\bigg\{\int_t^T \big(|f(X_s)-u_s|^2+z_s\th_s\big)ds-\int_t^T z_sdW_s-\int_t^T \tilde{z}_sdB_s\bigg\} \\
&\le&\sup_{Q\in\cP}\bigg\{\int_t^T \big(|f(X_s)-u_s|^2+k|z_s|\big)ds-\int_t^T z_sdW_s-\int_t^T \tilde{z}_sdB_s\bigg\}\\
&=&\int_t^T \Big(|f(X_s)-u_s|^2+k|z_s|\Big)ds-\int_t^T z_sdW_s-\int_0^t \tilde{z}_sdB_s.
\end{eqnarray*}
On the other hand, by Lemma B.1(b) in \cite{chen2002}, there exists $\th^*_t\in[-k,k]$ such that
\[
\th^*_tz_t=\max_{\th_t}\th_tz_t=k|z_t|.
\]
Hence,
\begin{eqnarray*}
y_t &=& \sup_{Q\in\cP}V^Q_t=\sup_{Q\in\cP}\bigg\{\int_t^T \big(|f(X_s)-u_s|^2+z_s\th_s\big)ds-\int_t^T z_sdW_s-\int_t^T \tilde{z}_sdB_s\bigg\} \\
&\ge& \int_t^T \Big(|f(X_s)-u_s|^2+\th^*_tz_t\Big)ds-\int_t^T z_sdW_s-\int_t^T \tilde{z}_sdB_s\\
&=&\int_t^T \Big(|f(X_s)-u_s|^2+k|z_t|\Big)ds-\int_t^T z_sdW_s-\int_t^T \tilde{z}_sdB_s,
\end{eqnarray*}
which implies $y_t$ is a solution to (\ref{eq:y_t}). The uniqueness follows from standard result for BSDE since the coefficients satisfy Lipschitz continuity.
\qed
Note that $y_t$ can also be expressed as the unique solution to the following BSDE
\[\left\{\begin{array}{ccl}
dy_t&=&\big(-|f(X_t)-u_t|^2-h(X_t)\tilde{z}_t+k|z_t|\big)dt+z_tdW_t+\tilde{z}_tdY_t,\\[0.5em]
y_T&=&0.
\end{array}\right.\]
Before we can prove Theorem \ref{thm1.1} we need the following preparation.
\begin{lemma}\label{lem1011a}
We can restrict the admissible control to those $u$ with $\|u\|_\infty\le\|f\|_\infty$, where $\|\cdot\|_\infty$ denotes the supremum norm.
\end{lemma}
Proof: For any $u\in\cU_{ad}$ we define
\[\tilde{u}_t=\left\{\begin{array}{cl}
u_t,&\quad\mbox{ if } |u_t|\le\|f\|_\infty,\\[0.5em]
\|f\|_\infty,&\quad\mbox{ if } u_t>\|f\|_\infty,\\[0.5em]
-\|f\|_\infty,&\quad\mbox{ if } u_t<-\|f\|_\infty.
\end{array}\right.\]
It is easy to show that
\[|f(X_t)-\tilde{u}_t|\le |f(X_t)-u_t|,\]
and hence, $J(\tilde{u},Q)\le J(u,Q)$. This implies that $J(\tilde{u})\le J(u)$.
\qed
{\it Proof of Theorem \ref{thm1.1}}: Let $u^n$ be such that $J(u^n)\to J_0$. By Lemma \ref{lem1011a}, without loss of generality, we may and will assume that $\|u^n\|_\infty\le\|f\|_\infty$. Then, $\{u^n\}$ is bounded in $\HH\equiv L^2([0,T]\times\Om)$ and hence, it is compact in weak topology of $\HH$. Without loss of generality, we assume that
$u^n\to u$ in weak topology. By Mazur's theorem, there is a sequence of convex combinations
\[\hat{u}^n=\sum_j\la^n_j u^{n+j}\to u\mbox{ in strong topology,}\]
where $\la^n_j\ge 0$ with $\sum_j\la^n_j =1$.
Consider BSDEs
\[\left\{\begin{array}{ccl}
dy^n_t&=&\(k|z_t^n|-|f(X_t)-u^n_t|^2-h(X_t)\tilde{z}^n_t\)dt+z^n_tdW_t+\tilde{z}^n_tdY_t,\\[0.5em]
y_T&=&0,
\end{array}\right.\]
and
\[\left\{\begin{array}{ccl}
d\hat{y}^n_t&=&\(k|\hat{z}^n_t|-|f(X_t)-\hat{u}^n_t|^2-h(X_t)\hat{\tilde{z}}^n_t\)dt+\hat{z}^n_tdW_t+\hat{\tilde{z}}^n_tdY_t,\\[0.5em]
y_T&=&0.
\end{array}\right.\]
Note that
\[|f(X_t)-\hat{u}^n_t|^2\le\sum_j\la^n_j|f(X_t)-u^{n+j}_t|^2,\]
and $\tilde{y}^n_t\equiv\sum_j\la^n_jy^{n+j}_t$ satisfies
\[\left\{\begin{array}{ccl}
d\tilde{y}^n_t&=&\(k|z_t^n|-\sum_j\la^n_j|f(X_t)-u^{n+j}_t|^2-h(X_t)\tilde{z}^n_t\)dt+z^n_tdW_t+\tilde{z}^n_tdY_t,\\[0.5em]
y_T&=&0.
\end{array}\right.\]
By comparison theorem, we have $\hat{y}^n_t\le \sum_j\la^n_jy^{n+j}_t$, and hence,
\begin{equation}\label{eq0903a}
J(\hat{u}^n,Q)\le \sum_j\la^n_j J(u^{n+j},Q).\end{equation}
So,
\[J(\hat{u}^n)\le \sum_j\la^n_j J(u^{n+j}).\]
For any $\ep>0$, let $N>0$ be such that $J(u^n)<J_0+\ep$ for all $n\ge N$. Thus,
\[J_0\le J(\hat{u}^n)\le \sum_j\la^n_j (J_0+\ep)=J_0+\ep.\]
By the continuity dependence of the BSDE on the generator, $y_t^{ \hat{u}^n}\to y_t^u$. Therefore,
$J(u)=J_0$ and hence, $u$ is an optimal ambiguity filter.
The uniqueness follows from the convexity directly, while the convexity is obtained by comparison similar to (\ref{eq0903a}).
\qed
\section{Characterization of the ambiguity filter}
\setcounter{equation}{0}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
In this section, we use the conditional mean-field approach of Section 2 to establish a necessary condition for the ambiguity filter. Namely, we proceed to presenting the proof of Theorem \ref{thm1.2}.
Theorem 2.1 in Chen and Epstein \cite{chen2002} has proved that the probability set $\cP$ defined in \eqref{def-cP} is convex, which allows us to apply the minimax theorem (see Theorem B.1.2 in Pham \cite{pham2009}) to the ambiguity filtering problem. That is,
\begin{equation}\label{eq-minimax}
\min_{v(\cdot)\in\cG}\sup_{Q\in\cP}J(v(\cdot),Q)=\sup_{Q\in\cP}\min_{v(\cdot)\in\cG}J(v(\cdot),Q),
\end{equation}
where
\[J(v(\cdot),Q)=\EE^Q\int^T_0|f(X_t)-v_t|^2dt.\]
Recall that the probability measure $\widetilde{Q}$ is absolutely continuous with respect to $Q$ and the Radon-Nikodym derivative $M_t^{-1}$ satisfies the following equation
\begin{equation}\label{eq1102bb}
dM_t=h(X_t)M_tdY_t,\quad M_0=1.
\end{equation}
We can first fix $\th$ with $|\th_t|\le k$, and search for the optimal filter. Under the probability measure $\tilde{Q}$ defined in \eqref{eq0905b}, $\widetilde{W}_t$
is still a Brownian motion. The signal equation can be rewritten as
\begin{equation}\label{eq1102a}
dX_t=(b(X_t)+\si(X_t)\th_t)dt+\si(X_t)d\widetilde{W}_t,\qquad X_0=x.
\end{equation}
Applying filtering theory (see Chapter 5 in Xiong \cite{X}) with $\th$ fixed, the solution $u_t$ to the minimal problem on the right side in \eqref{eq-minimax} satisfies (\ref{eq-u-frac}).
Now the problem is converted to a conditional mean-field optimal control problem
of Section 2 with $\cZ_t=\cF^0_t=\cG_t$, the state equations (\ref{eq1102bb}, \ref{eq1102a}) and the cost function
\begin{equation}\label{model-j}
J(\th)=-\frac12\widetilde{\EE}\int^T_0\left|f(X^\th_t)-\frac{\widetilde{\EE}(f(X_t)M_t|\cG_t)}{\widetilde{\EE}(M_t|\cG_t)}\right|^2M_tdt.
\end{equation}
We use the result of Section \ref{sec2} with control variable $\th$, coefficients
\[c(x,m,\th)=\left(\begin{array}{c}
b(x)+\si(x)\th\\ 0\end{array}\right),\;
\si^1(x,m)=\left(\begin{array}{c}
0 \\ h(x)m\end{array}\right),\;
\si^2(x)=\left(\begin{array}{c}
\si(x) \\ 0\end{array}\right),\]
and
\[ l(x,m,z)=-\frac12\(f(x)-\frac{z^1}{z^2}\)^2m,\quad \Gamma=0,\quad \Phi=0.\]
Here for simplifying the notations, we denote
\[Z_t^1=\widetilde{\EE}(f(X_t)M_t|\cG_t)\mbox{ and }Z^2_t=\widetilde{\EE}(M_t|\cG_t).\]
The Hamiltonian
\begin{eqnarray*}
&&H(x,m,z,\th,p^1,p^2,q^{11},q^{12},q^{21},q^{22})\\
&=&p^1(b(x)+\si(x)\th)+q^{12}h(x)m+q^{21}\si(x)-\frac12\left|f(x)-\frac{z^1}{z^2}\right|^2m,
\end{eqnarray*}
It is clear that
\[H_\th(t)=p^1_t\si(X_t),\]
and the adjoint process $p_t$ satisfies the BSDE (\ref{eq0905a}).
For $(\th_t)$ fixed, SDE (\ref{eq1102bb},\ref{eq1102a}) has a unique solution $(X,M)$. The BSDE \eqref{eq0905a} is linear with random coefficients. According to Theorem 3.6 in \cite{mo2019}, BSDE \eqref{eq0905a} admits a unique solution, which implies that the FBSDE (\ref{eq1102bb},\ref{eq1102a},\ref{eq0905a}) has a unique solution.
By SMP obtained in Section 2, we get $\th_t=-k \mbox{ sgn}(p^1_t)$. This finishes the proof of Theorem \ref{thm1.2}.
|
\section{Introduction}
\subsection{Nuclear magnetic quadrupole moments produced by T,P-odd nuclear forces and atomic and molecular EDM produced by magnetic field of these moments}
Measurements of atomic and molecular time reversal (T) and parity (P) violating electric dipole moments are used to test unification theories predicting CP-violation. Such measurements have already excluded a number of models and significantly reduced the parametric space of other popular models including supersymmetry~\cite{PR,ERK}. Another motivation is related to the baryogenesis problem, the matter-antimatter asymmetry in the universe which is produced by an unknown CP-violating interaction. The expected magnitude of an EDM is very small, therefore, it is advantageous to search for mechanisms that enhance the effects - see e.g.~\cite{Khriplovich,KL,GF}.
Schiff demonstrated that the nuclear EDM is completely screened in neutral atoms and molecules, and noted that a nonzero atomic EDM still may be produced if the distribution of EDM and charge in a nucleus are not proportional to each other~\cite{Schiff}. Further works~\cite{Sandars,Hinds,SFK,FKS1985,FKS1986} introduced and calculated the so called Schiff moment, a vector moment producing an electric field inside the nucleus after taking into account the screening of the nuclear EDM by electrons. The electric field produced by this Schiff moment polarizes the atom and produces an atomic EDM directed along the axis of nuclear spin. Refs.~\cite{Sandars,Hinds} calculated the Schiff moment due to the proton EDM. Refs.~\cite{SFK,FKS1985,FKS1986} calculated (and named) the nuclear Schiff moment produced by the P,T-odd nuclear forces.
It was shown in~\cite{SFK} that the contribution of the T,P-odd forces to the nuclear EDM and Schiff moment is larger than the contribution of a nucleon EDM. In Ref.~\cite{FG} an accurate expression for the Schiff moment electrostatic potential has been derived and the finite nuclear size corrections to the Schiff moment operator introduced (see also \cite{AKozlovC,AKozlovA}). The Schiff moment is proportional to the third power of the nuclear size which is very small on an atomic scale.
The magnetic interaction between nuclear MQM and atomic electrons mixes electron orbitals of opposite parity and produces an atomic EDM and T,P-violating nuclear spin - molecular axis interaction constants for molecules in electronic states with non-zero electron angular momentum \cite{SFK}. This magnetic interaction is not screened, so generically (without special enhancement factors) the atomic EDM produced by the interaction between nuclear MQM and electrons is expected to be an order of magnitude bigger than the EDM produced by the Schiff moment \cite{SFK} and two orders of magnitude bigger than the EDM produced by the electric octupole moment \cite{Octupole}. The nuclear EDM, Schiff, electric octupole and magnetic quadrupole produced by the T,P-odd nuclear forces are enhanced due to an opposite parity level with the same spin close to the ground state \cite{HH,SFK,Octupole}. Collective enhancement of the magnetic quadrupole moments in nuclei with a quadrupole deformation has been demonstrated in \cite{F1994} (see also \cite{FDK,Lackenby2018}).
\subsection{Enhancement of nuclear magnetic quadrupole moment due to nuclear octupole deformation and soft octupole vibration mode}
The largest enhancement ($\sim 10^2 - 10^3$ times) of the nuclear Schiff moment and electric octupole moment occurs in nuclei with an intrinsic octupole deformation, where both the small energy difference of nuclear levels with opposite parity and the collective effect work together \cite{Auerbach,Spevak,Octupole}. According to Refs.~\cite{Auerbach,Spevak} this happens in some isotopes of Fr, Rn, Ra and other actinide atoms. The atomic and molecular EDMs produced by the Schiff moment, electric octupole moment and MQM increase with nuclear charge $Z$ at a faster rate than $Z^2$ \cite{SFK,Octupole}. This again explains why the EDMs in actinide atoms and their molecules are expected to be significantly larger than in other systems.
The T,P-violating electric octupole moment and Schiff moment in the laboratory frame are proportional to the squared octupole deformation parameter $(\beta_3)^2$ \cite{Spevak,Octupole}. In nuclei with an octupole deformation $(\beta_3)^2 \sim (0.1)^2$. According to Ref. \cite{Engel2000}, in nuclei with a soft octupole vibration mode the squared dynamical octupole deformation parameter $\ev{(\beta_3)^2} \sim (0.1)^2$, i.e. it is the same as the static octupole deformation. This implies that a similar enhancement of the Schiff moment and electric octupole moment may be due to the dynamical octupole effect \cite{Engel2000,FZ,soft2} in nuclei where $\ev{\beta_3}=0$ \footnote{Recall an ordinary oscillator where $\ev{x}=0$ while $\ev{x^2} \neq 0$ .}. This observation significantly increases the list of nuclei where the Schiff moment and electric octupole moment are enhanced.
In the papers \cite{Auerbach,Spevak,EngelRa,Jacek2018}, numerical calculations of the Schiff moments and estimates of the atomic EDM produced by the electrostatic interaction between electrons and these moments have been done for $^{223}$Ra, $^{225}$Ra, $^{223}$Rn, $^{221}$Fr, $^{223}$Fr, $^{225}$Ac and $^{229}$Pa. Unfortunately, these nuclei have a short lifetime. Several experimental groups have considered experiments with $^{225}$Ra and $^{223}$Rn \cite{RaEDM,RaEDM2,RnEDM}. The only published EDM measurements \cite{RaEDM,RaEDM2} have been done for $^{225}$Ra, which has a half-life of 15 days. In spite of the Schiff moment enhancement in $^{225}$Ra, EDM measurements have not yet reached the sensitivity of the T,P-odd interaction comparable to the Hg EDM experiment \cite{HgEDM}. These experiments continue, despite the problems caused by the instability of $^{225}$Ra and the relatively small number of atoms available. In Ref. \cite{Th} the
nuclear Schiff moment of $^{229}$Th has been estimated, as this nucleus has a much longer lifetime (7917 years). Ref. \cite{FF19} extended the list of the candidates for the enhanced Schiff moments to include the following
stable isotopes: $^{153}$Eu, $^{161}$Dy, $^{163}$Dy, $^{155}$Gd, and long lifetime nuclei $^{235}$U, $^{237}$Np, $^{233}$U, $^{229}$Th, $^{153}$Sm, $^{165}$Er, $^{225}$Ac, $^{227}$Ac, $^{231}$Pa, $^{239}$Pu.
In this paper we estimate the nuclear MQM in nuclei where we expect $\ev{(\beta_3)^2}\, > (0.05)^2$ and express the MQM in terms of the CP-violating $\pi$-meson - nucleon interaction constants ${\bar g}_0$, ${\bar g}_1$, ${\bar g}_2$, QCD parameter ${\bar \theta}$ and quark chromo-EDMs.
Potential candidate nuclei include $^{153}$Eu $(I^{P} = 5/2^{+})$, $^{161}$Dy $(5/2^{+})$, $^{221}$Fr $(5/2^{-})$, $^{223}$Fr $(3/2^{-})$, $^{223}$Ra $(3/2^{+})$ , $^{223}$Rn $(7/2^{+})$, $^{225}$Ac $(3/2^{-})$, $^{227}$Ac $(3/2^{-})$, $^{229}$Th $(5/2^{+})$, $^{229}$Pa $(5/2^{+})$, $^{233}$U $(5/2^{+})$, $^{235}$U $(7/2^{-})$ and $^{237}$Np $(5/2^{+})$. We then estimate the T,P-odd effects in molecules containing these nuclei (EuO, EuN$^+$, RaF, AcO, AcN$^{+}$, AcF$^{+}$, ThO and ThF$^+$).
\subsection{Oscillating nuclear magnetic quadrupole moments and atomic and molecular electric dipole moments produced by axion dark matter}
The CP-violating neutron EDM may be due to the QCD $\theta$-term \cite{Witten}. It was noted in Ref.~\cite{Graham} that axion dark matter produces an oscillating neutron EDM, as the axion field is equivalent to the oscillating ${\bar \theta}$. The QCD $\theta$-term also produces T,P-odd nuclear forces, creating T,P-odd nuclear moments. Correspondingly, the axion field also produces oscillating nuclear T,P-odd moments \cite{Stadnik} which are enhanced by the octupole mechanism. To obtain results for the oscillating T,P-odd moments it is sufficient to replace the constant ${\bar \theta}$ by ${\bar \theta}(t)= a(t)/f_a$, where $f_a$ is the axion decay constant, $a(t) =a_0 \cos{m_a t}$, $(a_0)^2= 2 \rho /(m_a)^2$, where $\rho$ is the axion dark matter energy density \cite{Graham,Stadnik}. Moreover, in the case of a resonance between the frequency of the axion field oscillations and molecular transition frequency there may be an enormous resonance enhancement of the oscillating nuclear T,P-odd moment effect \cite{OscillatingEDM}. Since oscillating nuclear T,P-odd moments may be produced by axion dark matter, corresponding measurements may be used to search for this dark matter. This research is in progress, and the first results have been published in Ref. \cite{nEDM}, in which the oscillating neutron EDM and oscillating $^{199}$Hg Schiff moment have been measured. The effect produced by the oscillating axion-induced Pb Schiff moment in solid state materials has been measured by the CASPEr collaboration in Ref. \cite{CasperNew}.
The effect of oscillating T,P-odd nuclear polarizability has been measured in Ref. \cite{Cornell} (see theory in \cite{pol1,pol2,pol3}). Oscillating MQMs produce resonance transitions in molecules \cite{MQMtransitions}.
\section{Estimates of nuclear magnetic quadrupole moments} \label{Schiff}
\subsection{Calculation of MQM in nuclei with octupole deformation or soft octupole mode}
The magnetic quadrupole moment of a nucleus due to the electromagnetic current of a single nucleon with mass $m$ is defined by the second order tensor operator \cite{SFK},
\begin{align} \label{eq:MQMTensor}
\begin{split}
\hat{M}_{kn}^{\nu} = \dfrac{e}{2m}\biggl[3\mu_{\nu}\left(r_k\sigma_n + \sigma_kr_n - \dfrac{2}{3}\delta_{kn}\hat{\boldsymbol{\sigma}}\textbf{r}\right) \\
+ 2q_{\nu}\left(r_kl_n + l_kr_n\right)\biggr]
\end{split}
\end{align}
where $\nu = p,n$ for protons and neutrons respectively and m, $\mu_{\nu}$ and $q_{\nu}$ are the mass, magnetic moment in nuclear magnetons $\dfrac{e}{2m}$ and charge of the nucleon respectively. In the case of an axially symmetric nucleus for an orbital with definite spin projection on the nuclear axis $\Sigma$ and orbital angular momentum projection $\Lambda$, we may give the following estimate for the expectation value of the MQM
\begin{align} \label{eq:Mzz}
M_{zz}^{\nu} = 4 \dfrac{e}{2m}\mu_z \ev{r_z},
\end{align}
where $\mu_z= 2 \Sigma \mu_{\nu} + q_{\nu} \Lambda$ is the projection of the nucleon magnetic moment and $\ev{ r_z}$ is the expectation value of the radius vector ${\bf r}$. The latter vanishes in the absence of an octupole deformation, therefore $\ev{ r_z} \sim \beta_3 R$, where $R\approx A^{1/3} 1.2$ fm is the nuclear radius.
Above, we have presented the MQM in the intrinsic frame which rotates with the nucleus. We must now find the MQM in the laboratory frame. A nucleus with an octupole deformation and non-zero nucleon angular momentum has a doublet of close opposite parity rotational states $\ket{I^{\pm}}$ with the same angular momentum $I$
\begin{align} \label{doublet}
\ket{ I^{\pm} }=\frac{1}{\sqrt{2}} (\ket{\Omega} \pm \ket{-\Omega}),
\end{align}
where $\Omega=\Sigma +\Lambda$ is the projection of $I$ on to the nuclear axis). In the case of the ordinary electric quadrupole moment $Q$, which conserves $T$ and $P$ symmetries, we have the relation $\mel{\Omega }{ Q_{zz} }{ \Omega}= \mel{ - \Omega }{ Q_{zz} }{ - \Omega}$, and the following relation between the intrinsic value $Q_{zz}$ and laboratory value $Q$ in the ground rotational state \cite{BohrMott}:
\begin{align} \label{eq:RotationalFactor}
Q = \dfrac{I\left(2I - 1\right)}{\left(I + 1 \right)\left(2I + 3\right)}Q_{zz},
\end{align}
where $I=I_z= \left|\Omega\right|$ is the projection of total nuclear angular momentum (nuclear spin $I$) on the symmetry axis, $\Omega={\bf I \cdot n}$. This expression for $Q$ shows that we can only detect these second order tensor properties in nuclei with spin $I > 1/2$. In the case of the MQM we have $\mel{\Omega }{ M_{zz} }{ \Omega}= - \mel{ - \Omega }{ M_{zz} }{ - \Omega}$, and the laboratory value of $M$ vanishes in the states of definite parity (\ref{doublet}) which have equal weights of the $\Omega$ and $-\Omega$ components. This is a consequence of $T$ and $P$ conservation.
However, the states of this doublet are mixed by the $T,P-$violating interaction $W$, with mixing coefficient:
\begin{equation}\label{alpha}
\alpha_{\pm}=\frac{\mel{I^-}{ W}{ I^+}}{E_+ - E_-} .
\end{equation}
This mixing produces non-equal weights of $\Omega$ and $-\Omega$ , $(1+ \alpha_{\pm})^2/2$ and $(1 - \alpha_{\pm})^2/2$ respectively, and leads to a non-zero expectation value of $\ev{{\bf I \cdot n}}$, i.e. the mixing polarises the nuclear axis ${\bf n}$ along the nuclear spin ${\bf I}$ \cite{Auerbach,Spevak}:
\begin{equation}\label{n}
\ket{n_z}= 2 \alpha_{\pm} \frac{I_z}{I+1}\,.
\end{equation}
As a result all intrinsic T,P-odd nuclear moments appear in the laboratory frame. Using Eqs. (\ref{eq:Mzz}--\ref{eq:RotationalFactor}), we obtain the following for MQMs in the ground nuclear state
\begin{align} \label{eq:MQMRotationalFactor}
\begin{split}
M & = 2 \alpha_{\pm} \dfrac{I\left(2I - 1\right)}{\left(I + 1 \right)\left(2I + 3\right)} M_{zz} \ , \\
& = 8 \alpha_{\pm} \dfrac{I\left(2I - 1\right)}{\left(I + 1 \right)\left(2I + 3\right)} \dfrac{e}{2m}\mu_z \ev{ r_z}\,.
\end{split}
\end{align}
According to Ref. \cite{Spevak} the T,P-violating matrix element is approximately equal to
\begin{equation}\label{W}
\mel{I^-}{ W}{ I^+} \approx \frac{\beta_3 \eta}{A^{1/3}} [ \textrm{eV}].
\end{equation}
Here $\eta$ is the dimensionless strength constant of the nuclear $T,P$- violating potential $W$:
\begin{equation}\label{eta}
W= \frac{G}{\sqrt{2}} \frac{\eta}{2m} ({\bf \sigma \nabla}) \rho ,
\end{equation}
where $G$ is the Fermi constant, $m$ is the nucleon mass and $\rho$ is the nuclear number density. The nuclear magnetic moment in the intrinsic frame is related to $\mu$ in the laboratory frame by
\begin{equation}\label{mu}
\mu= \frac{I}{I+1}\mu_z\,.
\end{equation}
The value of $\mu$ has been measured in all nuclei of experimental interest. Using $\ev{r_z} \approx 1.2 \beta_3 A^{1/3} $ fm we obtain nuclear MQM
\begin{align} \label{eq:Mlab}
M\approx \dfrac{2I - 1}{2I + 3} \ev{(\beta_3)^2} \frac{ \textrm{eV}} {E_+ - E_-} \mu \eta \, e \,\textrm{fm}^2\,.
\end{align}
The typical energy interval of the doublet $E_+ - E_-$ is between 25 keV and 100 keV, which is much smaller than the interval between the opposite parity orbitals in spherical nuclei ($\sim$ 8 MeV). Therefore, value of the MQM in nuclei with an octupole deformation may be 1 - 2 orders of magnitude bigger than the MQM of a spherical nucleus estimated in Ref. \cite{SFK}. To avoid misunderstanding, we should mention that this enhancement is not as big as the enhancement of the nuclear Schiff moment (2-3 orders of magnitude) which has a collective nature in the nuclei with an octupole deformation \cite{Auerbach,Spevak,EngelRa,Jacek2018}.
Note that the MQM in Eq. (\ref{eq:Mlab}) is quadratic in the octupole deformation parameter. Therefore, it is sufficient to have a soft octupole deformation mode, i.e. dynamical deformation, which actually gives values of $ \ev{(\beta_3)^2} $ comparable to that for the static octupole deformation. In this case, the situation is similar to the Schiff moment calculations, see Refs. \cite{Engel2000,FZ}.
Within the meson exchange theory, the $\pi$-meson exchange gives the dominating contribution to the T,P-violating nuclear forces \cite{SFK}. According to Ref. \cite{FDK} the neutron and proton constants in the T,P-odd potential (\ref{eta}) may be presented as
\begin{align}
\eta_n=-\eta_p = (- g \bar{g}_{0} + 5 g \bar{g}_{1} + 2 g \bar{g}_{2} ) 10^{6},
\end{align}
where $g$ is the strong $\pi$-meson - nucleon interaction constant and ${\bar g}_0$, ${\bar g}_1$, ${\bar g}_2$ are the $\pi$-meson - nucleon CP-violating interaction constants in the isotopic channels $T=0,1,2$. The numerical coefficient comes from $(Gm_{\pi}^2/2^{1/2})^{-1}= 6.7 \cdot 10^6$ times the factor 0.7 corresponding to the zero range reduction of the finite range interaction due to the $\pi_0$ exchange \cite{SFK,FKS1985,FKS1986}.
We can also express $\eta$ in terms of the QCD $\theta$-term constant. Using the results presented in Refs. ~\cite{Yamanaka2017,Vries2015,Bsaisou2015}
\begin{align}
g \bar{g}_{0} &=-0.21 \, \bar{\theta}, \\
g \bar{g}_{1} &= 0.046 \, \bar{\theta},
\end{align}
we obtain
\begin{align} \label{etawrttheta}
\eta_n=-\eta_p = 4 \times 10^{5} \ \bar{\theta},
\end{align}
Further, we can express $\eta$ via the quark chromo-EDMs ${\tilde d_u}$ and ${\tilde d_d}$. Using relations $g {\bar g}_0 = 0.8 \times 10^{15}({\tilde d_u} +{\tilde d_d})$/cm, $g {\bar g}_1 = 4 \times 10^{15}({\tilde d_u} - {\tilde d_d})$/cm~\cite{PR} we obtain:
\begin{align}
\eta_n=-\eta_p = ( 2 (\tilde{d_{u}} - \tilde{d_{d}}) - 0.1 (\tilde{d_{u}} + \tilde{d_{d}}) ) 10^{22}/\text{cm}.
\end{align}
In the expressions above, the interaction constants have opposite signs for a valence proton or valence neutron (which in this case is an unparied nucleon, defining a non-zero value of $\Omega$). However, the magnetic moments of protons and neutrons have an opposing sign, meaning the overall sign of the product $\mu \eta$ for valence protons and neutrons is the same. As we are interested in the limits on these interaction constants obtained from the limits on the measured values of MQM measurements, which are presented as absolute values, this sign of these constants is not of importance. Therefore, in our estimates of MQM for specific nuclei we are only interested in their absolute values. Using Equation (\ref{eq:Mlab}) in conjunction with the above calculations, we may now write the MQM in terms of the interaction constants
\begin{align}
M(\eta) & \approx 10^{-31}M_0 \eta\,,
\label{Meta} \\
M(\theta) & \approx 4 \cdot 10^{-26} M_0 \bar{\theta}\,,
\label{Mtheta} \\
M(g) & \approx 10^{-25} M_0 \,
\label{Mg} (- g \bar{g}_{0} + 5 g \bar{g}_{1} + 2 g \bar{g}_{2} ) \,, \\
M(d) & \approx 10^{-9} M_0 \,
\label{Md} ( 2 (\tilde{d_{u}} - \tilde{d_{d}})-0.1 (\tilde{d_{u}} + \tilde{d_{d}}) )/\textrm{ cm} \,,
\end{align}
where
\begin{align}
M_0 & = \dfrac{2I - 1}{2I + 3} (\beta_3)^2 \frac{ 100 \, \textrm{keV}} {E_+ - E_-} \mu \, e \, \textrm{ cm}^2.
\end{align}
Note that a comparable contribution to the MQM is given by a spin hedgehog mechanism which requires quadrupole deformation (but does not require octupole deformation) \cite{F1994}.
Magnetic quadrupoles in this case have collective nature, somewhat similar to collective electric quadrupoles in deformed nuclei. Collective contributions to MQM have been calculated in Refs. \cite{F1994,FDK,Lackenby2018}.
\subsection{MQM for specific nuclei}
We may now present our calculations of the MQM for specific nuclei with an octupole deformation or soft octupole vibration mode. As an example, let us first consider the $^{237}$Np nucleus, which has a half-life of 2.14 million years, and is produced in macroscopic quantities in nuclear reactors. The interval between opposite parity levels which are mixed by the T,P-odd interaction is $E(5/2^{-}) - E(5/2^{+}) = 59.5$ keV. The experimental nuclear excitation spectra for this nucleus satisfies the criteria for the existence of an octupole deformation. Thus, given $^{237}$Np has one additional proton above $^{236}$U, it's octupole deformation parameter may be interpolated between the values for $^{234}$U and $^{238}$U (see Ref.~\cite{Afanasjev2016}) to be $\beta_{3} = 0.12$. Taking the value of the nuclear magnetic moment from Ref.~\cite{StoneTable} to be $\mu = 3.1(4)$, we may calculate explicitly the MQM for $^{237}$Np via substitution into Equations (\ref{Meta} - \ref{Md}) above
\begin{align} \label{MQMNp}
\begin{split}
M(\eta) & \approx 3.8 \cdot 10^{-33} \eta \, e \, \textrm{ cm}^2, \\
M(\theta) & \approx 1.5 \cdot 10^{-27} \bar{\theta} \, e\, \textrm{ cm}^2, \\
M(g) & \approx 3.8 \cdot 10^{-27}
(- g \bar{g}_{0} + 5 g \bar{g}_{1} + 2 g \bar{g}_{2} ) \, e \, \textrm{ cm}^2, \\
M(d) & \approx 3.8 \cdot 10^{-11} ( 2 (\tilde{d_{u}} - \tilde{d_{d}})-0.1 (\tilde{d_{u}} + \tilde{d_{d}}) ) \, e \, \textrm{ cm}.
\end{split}
\end{align}
Using a similar method, we performed these calculations for a range of nuclei which exhibit a nuclear excitation spectra consistent with that of an octupole deformation: $^{153}$Eu, $^{161}$Dy, $^{221}$Fr, $^{223}$Fr, $^{223}$Ra, $^{223}$Rn, $^{225}$Ac, $^{227}$Ac, $^{229}$Th, $^{229}$Pa, $^{233}$U and $^{235}$U. The results are presented in Table \ref{Nuclei1}. Values for the nuclear magnetic moments are taken from Ref.~\cite{StoneTable}, unless otherwise indicated.
\begin{widetext}
\begin{center}
\begin{table}[ht]
\begin{ruledtabular}
\begin{tabular}{ccccccccc}
& & & & & \multicolumn{4}{c}{$M$} \\
\cmidrule{6-9}
Nucleus & $I^{P}$ & $\mu$ (nm) & $\Delta E$ (keV) & $\beta_{3}$ & $ 10^{-34} \ \eta \ e \ \text{cm}^{2}$ & $ 10^{-28} \bar{\theta} \ e \ \text{cm}^{2} $ & $ 10^{-28} \ \bar{G} \ e \ \text{cm}^{2}$~\footnote{For brevity, we make the substitution $\bar{G} \equiv (- g \bar{g_{0}} + 5 g \bar{g_{1}} + 2 g \bar{g_{2}})$. } & $ 10^{-12} \ \tilde{D} \ e \ \text{cm} $~\footnote{Similarly, we denote $\tilde{D} \equiv 2(\tilde{d}_{u} - \tilde{d}_{d}) - 0.1 (\tilde{d}_{u} + \tilde{d}_{d}))$.} \\
\midrule
$^{153}$Eu & 5/2$^{+}$ & 1.5(3) & 97.4 & 0.15~\cite{FF19} & 18 & 7.1 & 18 & 18 \\
$^{161}$Dy & 5/2$^{+}$ & -0.48(3) & 25.7 & $\sim 0.1$ & -9.4 & -3.7 & -9.4 & -9.4 \\
$^{221}$Fr & 5/2$^{-}$ & 1.6(3) & 234 & 0.10~\cite{Spevak} & 3.4 & 1.4 & 3.4 & 3.4 \\
$^{223}$Fr & 3/2$^{-}$ & 1.2(2) & 160 & 0.090~\cite{Spevak} & 2.0 & 0.79 & 2.0 & 2.0 \\
$^{223}$Ra & 3/2$^{+}$ & 0.27(2) & 50.1 & 0.10~\cite{Spevak} & 1.8 & 0.72 & 1.8 & 1.8 \\
$^{223}$Rn & 7/2$^{+}$ & -0.78(8) & 130 & 0.081~\cite{Spevak} & -2.3 & -0.94 & -2.3 & -2.3 \\
$^{225}$Ac & 3/2$^{-}$ & 1.1(1)~\footnote{Using the value for $^{227}$Ac, which has identical spin and parity to $^{225}$Ac.} & 40.1 & 0.10~\cite{Spevak} & 9.9 & 4.0 & 9.9 & 9.9 \\
$^{227}$Ac & 3/2$^{-}$ & 1.1(1) & 23.4 & 0.12~\cite{FF19} & 13 & 5.0 & 13 & 13 \\
$^{229}$Th & 5/2$^{+}$ & 0.46(4) & 133 & 0.12~\cite{Minkov} & 2.3 & 0.91 & 2.3 & 2.3 \\
$^{229}$Pa & 5/2$^{+}$ & 1.96~\footnote{Calculated using the Schmidt model.} & 0.220 & 0.082~\cite{Spevak} & 3000& 1200 & 3000 & 3000 \\
$^{233}$U & 5/2$^{+}$ & 0.59(5) & 299 & 0.17~\cite{Afanasjev2016} & 4.8 & 1.9 & 4.8 & 4.8 \\
$^{235}$U & 7/2$^{-}$ & -0.38(3) & 81.7 & 0.17~\cite{FF19} & -8.1 & -3.2 & -8.1 & -8.1 \\
$^{237}$Np & 5/2$^{+}$ & 3.1(4) & 59.5 & 0.12~\cite{Afanasjev2016} & 38 & 15 & 38 & 38 \\
\end{tabular}
\end{ruledtabular}
\caption{Nuclear spin $I$, magnetic moment $\mu$, experimental interval between opposite parity levels mixed by the T,P-odd interaction $\Delta E = E^{+} - E^{-}$ and octupole deformation parameter $\beta_{3}$. Calculations using these quantities allow for the MQM of various octupole deformed nuclei to be expressed in terms of different fundamental constants.}
\label{Nuclei1}
\end{table}
\end{center}
\end{widetext}
\section{MQM energy shift in diatomic molecules} \label{sec:MQMmolecule}
The direct measurement of nuclear MQMs in an external magnetic field is unfeasible. As mentioned above the use of molecular systems is promising, as the nuclear MQM will interact with the internal magnetic field. Molecules in particular present the most attractive option due to existence of very close paired levels of opposite parity, the $\Omega$-doublet - see e.g. \cite{SFK}. For polar molecules consisting of a heavy and light nucleus (for example, Th and O) the effect of MQM is $\sim Z^2$, and thus it is calculated for the heavier nucleus. The Hamiltonian of diatomic paramagnetic molecule including the $T, P-$ odd nuclear moment effects is given by \cite{SFK,Kozlov1995}:
\begin{align}
H = W_d d_e \mathbf{S}\cdot\mathbf{n} + W_{Q}\dfrac{Q_s}{I}\mathbf{I}\cdot\mathbf{n} - \dfrac{W_{M}M}{2I(2I -1)}\mathbf{S}\hat{\mathbf{T}}\mathbf{n},
\end{align}
where $d_e$ is the electron EDM, $Q_s$ is the nuclear Schiff moment, $M$ is the nuclear MQM, $\mathbf{S}$ is the effective electron spin, $\mathbf{n}$ is the symmetry axis of the molecule, $\hat{\mathbf{T}} $ is the second rank tensor operator characterised by the nuclear spins $T_{ij} = I_iI_j + I_jI_i - \tfrac{2}{3}\delta_{ij}I(I + 1)$ and $W_d$, $W_Q$ and $W_M$ are fundamental parameters for each interaction which are dependent on the particular molecule. We have omitted the $T$-,$P$- odd electron-nucleon interaction terms which are presented e.g. in the review \cite{GF}.
Parameters $W_d$, $W_Q$ and $W_M$
are related to the electronic molecular structure of the state.
These parameters are calculated using relativistic many-body methods (see e.g. \cite{FDK}). For the nuclear MQM we are interested only in $W_{M}$, which has been calculated for various molecules of our interest: RaF~\cite{Talukdar}, AcO~\cite{Oleynichenko2022}, AcN$^{+}$~\cite{Oleynichenko2022}, AcF~\cite{Oleynichenko2022}, ThO \cite{Skripnikov2014} and ThF$^{+}$\cite{Skripnikov2017}. For other molecules of interest, where accurate many-body calculations are absent (such as EuO and EuN$^{+}$), we may use the $Z^2$ scaling for the electron factor in the energy shifts $W_{M} \Omega$, and thus estimate the value of $W_{M}$. We may thus express the energy shifts induced by MQMs in terms of the $CP$-violating $\pi$-meson - nucleon interaction constants $\bar{g}_{0}$, $\bar{g}_{1}$ and $\bar{g}_{2}$, the QCD parameter $\theta$ and the quark chromo-EDMs. The results of these calculations are shown in Table \ref{MoleculeTable}.
Using the current limits on the $CP$-violating parameters $|\bar{\theta}| < 2.4 \times 10^{-10}$, $|\tilde{d}_{u} - \tilde{d}_{d} | < 6 \times 10^{-27}$cm and $|\frac{1}{2}\tilde{d}_{u} + \tilde{d}_{d} | < 3 \times 10^{-26}$cm \cite{Swallows}, the MQM energy shifts $(|W_{M} M S|)$ in $^{229}$ThO are $< 25 \ \mu$Hz and $< 23 \ \mu$Hz respectively. The $^{232}$ThO molecule has recently been used
to set new limits on the electron EDM, with an accuracy for the energy shift of $80 \ \mu$Hz \cite{ACMECollaboration}, which is a factor of 12 improvement in accuracy compared to their previous work \cite{TheEDM}.
$^{232}$Th has zero nuclear spin and no MQM. The possibility of performing a similar experiment with $^{229}$ThO is promising, as in principle, such measurements would improve constraints on nuclear $CP$-violating interactions. The results for the other molecules presented in Table \ref{MoleculeTable} are similar, and given the expected enhancement of the MQM in their nuclei, these may be useful candidates in the search for new physics in the hadronic sector.
\begin{widetext}
\begin{center}
\begin{table}[!ht]
\begin{ruledtabular}
\begin{tabular}{cccccccc}
& & & $| W_{M} | $ & \multicolumn{4}{c}{$|W_{M} MS| $ ($\mu$Hz)} \\
\cmidrule{4-8}
Molecule & $I^{P}$ & State & $10^{39} \ \mu \text{Hz}/ e \cdot \text{cm}^{2}$ & $10^{5} \ \eta$ & $10^{11} \ \bar{\theta}$ & $10^{11} \bar{G}$ & $10^{27} \ \tilde{D} / \text{cm}$ \\
\midrule
$^{153}$EuO,$^{153}$EuN$^{+}$ & $5/2^{+}$ & --~\footnote{At this time, the $\Omega$ value of these Eu based molecules are unknown.} & 0.60~\footnote{Estimated using $Z^{2}$ scaling.} & 11 & 4.2 & 11 & 11 \\
$^{223}$RaF & $3/2^{+}$ & $^{2}\Sigma_{1/2}$ & 1.2~\cite{Talukdar} & 1.1 & 0.44 & 1.1 & 1.1 \\
$^{225}$AcO, $^{225}$AcN$^{+}$, $^{225}$AcF$^{+}$ & $3/2^{-}$ & $^{2}\Sigma_{1/2}$ & 1.7~\cite{Oleynichenko2022} & 8.2 & 3.3 & 8.2 & 8.2 \\
$^{227}$AcO, $^{227}$AcN$^{+}$, $^{227}$AcF$^{+}$ & $3/2^{-}$ & $^{2}\Sigma_{1/2}$ & 1.7~\cite{Oleynichenko2022} & 11 & 4.4 & 11 & 11 \\
$^{229}$ThO & $5/2^{+}$ & $^{3}\Delta_{1}$ & 1.1~\cite{Skripnikov2014} & 2.5 & 1.0 & 2.5 & 2.5 \\
$^{229}$ThF$^{+}$ & $5/2^{+}$ & $^{3}\Delta_{1}$ & 0.59~\cite{Skripnikov2017} & 1.3 & 0.53 & 1.3 & 1.3 \\
\end{tabular}
\end{ruledtabular}
\caption{Frequency shifts due to the MQM interaction with the electron magnetic field of the molecules. The energy shifts are presented in terms of the CP-violating parameters $\eta$, $\bar{\theta}$, $\bar{G} \equiv (- g \bar{g_{0}} + 5 g \bar{g_{1}} + 2 g \bar{g_{2}})$ and $\tilde{D} \equiv 2(\tilde{d}_{u} - \tilde{d}_{d}) - 0.1 (\tilde{d}_{u} + \tilde{d}_{d})$.}
\label{MoleculeTable}
\end{table}
\end{center}
\end{widetext}
\section{Conclusion}
In this publication we have presented a method to estimate the magnetic quadrupole moment in various isotopes with nuclear spin $I \ge 1$ and an experimental nuclear excitation spectra indicating an octupole deformation. Such nuclei exhibit an enhanced MQM due to the mixing of close, opposite parity rotational states with identical spin, and their study may provide an opportunity for detection of T,P-odd effects in the hadronic sector. Paramagnetic molecules containing heavy atoms are promising candidates for measuring the nuclear MQM. For several molecules of our interest, there exist accurate many-body calculations of the electronic factors $W_{M}$ before the nuclear magnetic quadrupole moment. Using these values, we calculated the energy shift induced by MQM in these molecules. In molecules where these calculations are absent, we have used the scaling $W_{M} \propto Z^{2}$ and accurate results for other molecules to estimate $W_{M}$. The calculations presented in this work may allow experimentalists to measure the values of fundamental $CP$-violating $\pi$-meson - nucleon interaction constants $\bar{g}_{0}$, $\bar{g}_{1}$ and $\bar{g}_{2}$, the QCD parameter $\theta$ and the quark chromo-EDMs.
We should note that for planning and interpretation of future experiments one needs dedicated many-body calculations of the nuclear MQM and their effects in molecules. Molecular calculations have already achieved few percent accuracy in diatomic molecules \cite{Oleynichenko2022,Skripnikov2014,Skripnikov2017}. Recently high accuracy molecular calculations of MQM effects have been extended to triatomic molecules \cite{Maison2020,Maison2019}. There are no
accurate nuclear many body calculations of MQM. However, the methods of the calculations of the nuclear Schiff moments presented in Refs.\cite{EngelRa,Jacek2018} are applicable for the calculations of MQM. Moreover, we may expect that calculations of the nuclear MQM may be more accurate since in the formula for the Schiff moment there are two terms of opposite sign while for MQM we do not expect any cancellations.
\section{Acknowledgements}
This work was supported by the Australian Research Council grants DP190100974 and DP200100150, and the JGU Gutenberg Fellowship.
\bibliographystyle{apsrev}
|
\subsection{Artifacts}\label{subsec_artifacts}
In this subsection, we describe the artifacts arising in the reconstruction from the local data, when there are conjugate points.
For convenience, we assume the weight $w = 1$ and use the notation $I$ instead of $I_w$ in the following.
First, we consider the backprojection ${{I}}^*{{I}}$ to reconstruct $f$, in the presence of conjugate covectors.
Suppose $\gamma$ is the $\lambda$-geodesic in Theorem \ref{thm_C_diff} with conjugate covectors $(x, \xi)$ in $V^1$ and $(y, \eta)$ in $V^2$.
Let $f = f_1 + f_2$ with $f_j$ singular in $V^j$, $j = 1,2 $.
In a small neighborhood of $\gamma$, we have
\begin{equation*
{{I}}^*{{I}}f = I_1^*I_1 f_1+I_1^*I_2 f_2+I_2^*I_1 f_1+I_2^*I_2 f_2.
\end{equation*}
Recall $I_1$ and $I_2$ are defined microlocally and are elliptic FIOs of order $-1$ with canonical relations ${C}_1$ and ${C}_2$, which are diffeomorphisms.
Then by Theorem \ref{thm_N},
$I_1^*I_1$ and $I_2^*I_2$ are elliptic $\Psi$DOs of order $-1$ with principal symbol
${4\pi}/{|\xi|}$.
For $I_1^* I_2$ and $I_2^*I_1$,
by \cite[Theorem 25.2.2]{Hoermander2009} and the transversal composition calculus,
they are FIOs of order $-1$ associated with canonical relations $C^{-1}_1 \circ C_2 $ and $C^{-1}_2 \circ C_1$ respectively.
Let $-\Delta_g$ be the Laplacian operator in $M$ and its square root $\sqrt{-\Delta_g}$ is a $\Psi$DO of order $1$ with principal symbol $|\xi|_g$.
Let $\Lambda = \sqrt{-\Delta_g}/{(4\pi)}$.
Then module lower order operators, one has
\[
\Lambda I_1^*I_1 \equiv \text{Id}, \quad \Lambda I_1^*I_1 \equiv \text{Id},
\quad \Lambda I_2^*I_1 \equiv F_{21}, \quad \Lambda I_1^*I_2 \equiv F_{12}.
\]
It follows that
\begin{align}\label{eq_fb}
\Lambda I^*If = f_1 + F_{21} f_1 + f_2 + F_{12} f_2
\end{align}
up to lower order terms.
This implies that we recover the singularities of $f_1 + f_2$ together with $ F_{12} f_1 + F_{21} f_2$ from the backprojection.
The later are artifacts.
If we write a distribution singular in $V^1 \cup V^2$ as a vector-valued function with the first component equal to its restriction to $V^1$ and the second component equal to its restriction to $V^2$, then
\begin{equation*
\Lambda{{I}}^*{{I}}f =
\begin{bmatrix}
\begin{array}{ll}
\text{Id} & F_{12}\\
F_{21}& \text{Id}
\end{array}
\end{bmatrix}
\begin{bmatrix}
\begin{array}{l}
f_1\\
f_2
\end{array}
\end{bmatrix}
\equiv M
\begin{bmatrix}
\begin{array}{l}
f_1\\
f_2
\end{array}
\end{bmatrix}
\end{equation*}
up to lower-order terms.
Especially when $f_2 = 0$, from the filtered backprojection $\Lambda{{I}}^*{{I}}f$, we recover $f_1 + F_{21}f_1$ and therefore the artifacts equal to $F_{21}f_1$ arises in the reconstruction.
Next, we consider the numerical reconstruction by using the Landweber iteration as in \cite{Holman2017}.
For more details of the method, see \cite{STEFANOV2017}.
We follow the same argument in \cite{Monard2015,Holman2017},
see also \cite{Zhang2020}.
Let $\chi$ be a smooth cutoff in $\mathbb{R}^2$ with $\chi = 1$ in a small neighborhood of $M$ to avoid dealing with non-local operators.
We set $\mathcal{L} = \Lambda \chi I^*I$ and compute
\[
\mathcal{L}^* \mathcal{L} =I^*I \chi \frac{1}{4\pi}({-\Delta_g})\chi I^*I
=
\begin{bmatrix}
\begin{array}{ll}
\text{Id} + F^*_{21}F_{21}& F^*_{21} + F_{12}\\
F_{12}^* + F_{21}& \text{Id} + F_{12}^*F_{12}
\end{array}
\end{bmatrix}.
\]
Let $g$ be the local data and it is assumed be in the range of ${{I}}$.
Now we use the Landweber iteration to solve the equation ${{I}} f = g$.
We write
\begin{equation}\label{landweber}
(\text{Id} - (\text{Id}- \gamma \mathcal{L}^* \mathcal{L})) f = \gamma \mathcal{L}^*\Lambda\chi I^* g.
\end{equation}
Then with a small enough and suitable $\gamma > 0$, it can be solved by the Neumann series and we have the truncated scheme
\[
f^{(N)} = \sum_{k=0}^{N}(\text{Id}- \gamma \mathcal{L}^*\mathcal{L})^k
\gamma \mathcal{L}^*\Lambda\chi I^* g.
\]
This series converge to the minimal norm solution to $\mathcal{L} f = \Lambda\chi I^* g$.
Suppose the original function is $f = f_1 + f_2$ with $f_2 = 0$.
The analysis in \cite[Section 3.2.3]{Holman2017} shows that
\begin{align*}
\text{Landweber solution} =
[f_1 - (\text{Id} + F_{21}^*F_{21})^{-1}f_1] + [F_{21}(\text{Id} + F^*_{21}F_{21})^{-1}f_1],
\end{align*}
where the first square brackets are terms microlocally supported in $V^1$ and the second term in $V^2$.
The artifacts arising in the reconstruction is
\begin{align*}
f - \text{Landweber solution} =
[(\text{Id} + F_{21}^*F_{21})^{-1}f_1] - [F_{21}(\text{Id} + F^*_{21}F_{21})^{-1}f_1].
\end{align*}
We emphasize that the artifacts above arises in the reconstruction from the local data.
If we consider the recovery from the global data, i.e., with the knowledge of the integral transform over all curves in $\Gamma$,
then the singularities of $f$ might be recoverable.
This is because the singularities can be probed by more than one smooth curves in $\Gamma$.
In some cases, the recovery of certain singularity depends on a discrete dynamical system, i.e., a sequence of conjugate covectors, inside $M$, see (\ref{eq_M}).
If this sequence goes out of $M$, then we can resolve the corresponding singularity as is discussed in \cite{MR3080199, Zhang2020}.
For more details, see Proposition \ref{pp_m}.
\section{Conjugate points and Jacobi fields}
\subsection{Conjugate points.}
We define the \textit{exponential map} as $\exp_p(t,v) = \gamma_{p,v}(t)$ for $(p,v) \in SM$, see \cite{MR2970707}.
This definition uses polar coordinates and is independent of a change of the parameterization for curves in $\Gamma$.
For $v \in S_pM$, $t \in \mathbb{R}$, the map $V = t v \mapsto \exp_p(t,v)$ may not be smooth near $V = 0 \in T_pM$, see \cite[A.3]{MR2351370} for more details and properties.
By $d_{v} \exp_p(t,v)$, we always mean the differential of $\exp_p(t,v)$ w.r.t $v \in S_p M$.
This notation is different from the one $d_{v} \exp_p(v)$ for $v \in T_p M$ used in the case of geodesics.
We say a point $q = \gamma_{p,v}(t_0)$ is \textit{conjugate} to $p = \gamma_{p,v}(0)$ if the differential $d_{t,v} \exp_p(t,v)$ of the exponential map in polar coordinates has the rank less than the maximal one at $(t_0, v)$.
To further describe the conjugate points, we consider the flow $\phi_t: SM \rightarrow SM$ given by
$
\phi_t(x,v) = (\gamma_{x,v}(t),\dot{\gamma}_{x,v}(t)).
$
For a fixed curve $\gamma_{p_0,v_0}(t)$, we set
\[
(p_1,v_1) = \phi_t(p_0,v_0), \quad (p_2,v_2) = \phi_{t+s}(p_0,v_0),
\]
where $(p_j,v_j) \in SM$, for $j = 0,1,2$.
Note that $\phi_t: SM \rightarrow SM$ is a local diffeomorphism, see \cite[Chapter 7]{Merry2011}.
Its pushforward $\diff \phi_t: T_{(p,v)} SM \rightarrow T_{\phi_t(p,v)} SM$ is invertible.
Let $\pi$ be the natural projection from $SM$ to $M$, which induces a map $\diff \pi$ from $T_{(p,v)} SM$ to $T_{p} M$.
We have the following proposition that relates the conjugate points with the pushforward of the flow $\phi_t$.
\begin{pp}\label{conjugate_J}
A point $p_2 =\gamma_{p_0,v_0}(t+s) $ is conjugate to $p_1 =\gamma_{p_0,v_0}(t)$ if and only if there exists
$\zeta_0 \in T_{(p_0, v_0)} SM$
and $c \in \mathbb{R}$ satisfying
\begin{equation}\label{zeta_0}
\diff \pi (\diff \phi_t)(\zeta_0) = 0,\qquad
\diff \pi (\diff \phi_{t+s})(\zeta_0) + c \dot{\gamma}_{p_0,v_0}(t+s) = 0.
\end{equation}
\end{pp}
\begin{proof}
Let $\zeta_j \in T_{(p_j,v_j)} SM$ for $j = 0, 1, 2$ such that
\[
\zeta_2 = \diff \phi_s (\zeta_1) = \diff \phi_{t+s} (\zeta_0),
\]
which implies $\zeta_1 = \diff \phi_t (\zeta_0)$ by the group property of $\phi_t$.
Let $\{p_j^i, v_j^i\}$ be the local charts near $(p_j, v_j)$
and they induce a natural local basis $\{\partial/\partial p_j^i, \partial/\partial v_j^i\}$ in $T_{(p_j, v_j)} SM$, for $j = 0,1,2$.
We write $\zeta_j = (\alpha_j, \beta_j)$ in this local coordinates, where $\alpha_j$ can be identified as an element in $T_{p_j}M$ and
$\beta_j$ in $T_{v_j} S_{p_j} M$.
Note that $\diff \pi (\zeta_j) = \alpha_j$ for $j = 0,1,2$.
Consider the representation of $\diff \phi_t$ in these local coordinates near $(p_j,v_j)$ given by the Jacobian matrix
\[
J_{(p_j,v_j,t)} =
\begin{bmatrix}
\begin{array}{cc}
\diff_p\exp_{p}(t,v) & \diff_v\exp_{p}(t,v)\\
\diff_p {\dot{\exp}_{p}(t,v)}& \diff_v {\dot{\exp}_{p}(t,v)}
\end{array}
\end{bmatrix}_{(p_j,v_j,t)}.
\]
More explicitly, we have
\begin{equation}\label{jacobieq1}
\begin{bmatrix}
\begin{array}{c}
\alpha_2\\
\beta_2
\end{array}
\end{bmatrix}
= J_{(p_1,v_1,s)}
\begin{bmatrix}
\begin{array}{c}
\alpha_1\\
\beta_1
\end{array}
\end{bmatrix}
=
J_{(p_0,v_0,t+s)}
\begin{bmatrix}
\begin{array}{c}
\alpha_0\\
\beta_0
\end{array}
\end{bmatrix},
\end{equation}
and
\begin{equation}\label{jacobieq1_2}
\begin{bmatrix}
\begin{array}{c}
\alpha_1\\
\beta_1
\end{array}
\end{bmatrix}
= J_{(p_0,v_0,t)}
\begin{bmatrix}
\begin{array}{c}
\alpha_0\\
\beta_0
\end{array}
\end{bmatrix}.
\end{equation}
Recall that $p_2$ is conjugate to $p_1$ if only if there exists
$c \in \mathbb{R}$ and $\beta_1 \in T_{v_1} S_{p_1} M$ such that
\begin{align*}
d_{s,v} \exp_{p_1}(s,v_1)(c, \beta_1) = 0 \Rightarrow d_v \exp_{p_1}(s,v_1)(\beta_1) = -c \dot{\gamma}_{p_1,v_1}(s),
\end{align*}
On the one hand, if such $c, \beta_1$ exist, then we pick $\zeta_1 = (0, \beta_1)$ and
let $\zeta_0 = J_{(p_0,v_0,t)}^{-1} \zeta_1$.
In this case,
by equation (\ref{jacobieq1}) and (\ref{jacobieq1_2}) we have
\begin{align*}
& 0 = \alpha_1 = \diff \pi(\diff \phi_t)(\zeta_0) = d_p\exp_{p_0}(t,v_0)(\alpha_0) + d_v\exp_{p_0}(t,v_0)(\beta_0),\\
& \alpha_2 = \diff \pi(\diff \phi_{t+s})(\zeta_0) = d_p\exp_{p_1}(s,v_1)(\alpha_1) + d_v\exp_{p_1}(s,v_1)(\beta_1)= -c \dot{\gamma}_{p_1,v_1}(s),
\end{align*}
which proves (\ref{zeta_0}).
On the other hand, if there is $\zeta_0$ such that (\ref{zeta_0}) is true, then we take $(\alpha_1, \beta_1) = J_{(p_0,v_0,t)}(\zeta_0)$.
By (\ref{zeta_0}) we must have $\alpha_1 =0$ and then $d_v \exp_{p_1}(s,v_1)(\beta_1) = -c \dot{\gamma}_{p_1,v_1}(s)$, which implies that $p_2$ is conjugate to $p_1$.
\end{proof}
In the following, recall some results in \cite[Chapter 4, 7]{Merry2011}.
We consider a moving frame $\{F(x,v), H(x,v), V(x,v)\}$ on $SM$, where $F(x,v)$ is the infinitesimal generator of $\phi_t$,
$H(x,v)$ is the horizontal vector field, and $V(x,v)$ is the vertical vector field.
For a detailed definition of $H$ and $V$, see \cite[Section 4.1]{Merry2011}.
One can show that
\begin{align}\label{projection_frame}
\diff \pi(F(x,v)) = v, \quad \diff \pi(H(x,v)) = v^\perp, \quad \diff \pi(V(x,v)) =0.
\end{align}
The $\lambda$-geodesic vector field $F$ is related to the geodesic vector field $X$ by
\[
F = X + \lambda V.
\]
One can show the commutators between them are given by
\begin{align*}
&[H, V] = X = F- \lambda V,\\
&[V, F] = H + V(\lambda) V,\\
&[F, H] = (K - H(\lambda) + \lambda^2) V - \lambda F,
\end{align*}
where $K$ is the Gaussian curvature.
\begin{claim}\label{cl_1}
For any constant $c$, we can find $\zeta_0' \in T_{(p_0, v_0)} SM$ such that
\[
\diff \pi(\diff \phi_{t})(\zeta_0') = c \dot{\gamma}_{p_0,v_0}(t).
\]
\end{claim}
\begin{proof}
By \cite[Section 7.1]{Merry2011}, we consider the infinitesimal generator $F(x,v)$ of the flow $\phi_t$ and define
$$Y(t) = c (\diff \phi_{-t})(F(\phi_t(x,v))).$$
Differentiating both sides with respect to $t$ implies
$$
\dot{Y}(t) = c[F(x,v),F(x,v)] = 0.
$$
Thus $Y(t)$ is a constant vector field on $SM$. By denoting it as $Y = \zeta_0'$, we have $ d\phi_{t}(\zeta_0') = cF(\phi_t(x,v))$.
With $\diff \pi (F(x,v)) = v$, applying $\diff \pi$ to $ \diff \phi_{t}(\zeta_0')$ claims what we need.
\end{proof}
\subsection{Jacobi Fields}Inspired by Proposition \ref{conjugate_J}, we say a vector field $J_{(p_0,v_0)}(t)$ is a \textit{Jacobi field} along $\gamma_{p_0,v_0}$ if it can be written as
\begin{align}\label{jacobifield_rep}
J_{(p_0,v_0)}(t) = \diff \pi(\diff \phi_t)(\zeta) + (c_1 + c_2t) \dot{\gamma}_{p_0,v_0}(t),
\end{align}
for some $\zeta \in T_{(p_0, v_0)} SM$ and $c_1, c_2 \in \mathbb{R}$.
We emphasize that the representation of a Jacobi field in form of (\ref{jacobifield_rep}) is not unique, which can be seen from Claim \ref{cl_1} and Proposition \ref{jocabifield_simplied}.
Moreover, this definition is different from the usual one that the Jacobi fields are defined as the variation field along $\gamma_{p_0,v_0}$,
for example see \cite{Assylbekov2017}.
However, they are essentially the same and the definition we choose here is more straightforward for the purpose of this work.
To understand the last term $t \dot{\gamma}_{p_0,v_0}(t)$, recall we extend $\Gamma$ in Section \ref{subsection_extend},
such that for any $V = rv \in T_pM\setminus 0$ with $r>0$ and $v\in S_pM$, there exists a unique curve $\gamma_{x,V}(t)$ belonging to $\Gamma$.
\begin{pp}
A point $p_2 = \gamma_{p_0,v_0}(t_2)$ is conjugate to $p_1 = \gamma_{p_0,v_0}(t_1)$ if and only if there exists a nonvanishing Jacobi field $J(t)$ along $\gamma_{p_0,v_0}$ such that $J(t_1) = J(t_2) = 0$.
\end{pp}
\begin{proof}
If $p_2 = \gamma_{p_0,v_0}(t_2)$ is conjugate to $p_1 = \gamma_{p_0,v_0}(t_1)$,
then by Proposition \ref{conjugate_J}, there exists $\zeta_0$ and $c$ such that
\begin{equation*}
\diff \pi (\diff \phi_{t_1})(\zeta_0) = 0,\qquad
\diff \pi (\phi_{t_2})(\zeta_0) + c \dot{\gamma}_{p_0,v_0}(t_2) = 0.
\end{equation*}
Let $J_{(p_0,v_0)}(t) = \diff \pi(\diff \phi_t) (\zeta_0) + c \frac{t-t_1}{t_2 - t_1} \dot{\gamma}_{p_0,v_0}(t)$.
Then we have $J(t_1) = J(t_2) = 0$.
Conversely, suppose there is a nonzero Jacobi field with $J_{(p_0,v_0)}(t_1)=J_{(p_0,v_0)}(t_2)=0$. More explicitly, we have
\begin{align}\label{conjugate equation}
\begin{split}
&\diff \pi(\diff \phi_{t_1})(\zeta_0) + \lambda_1\dot{\gamma}_{p_0,v_0}(t_1) = 0, \\
&\diff \pi(\diff \phi_{t_2})(\zeta_0) + \lambda_2\dot{\gamma}_{p_0,v_0}(t_2) = 0,
\end{split}
\end{align}
where $\lambda_1 = c_1+c_2t_1,\lambda_2 = c_1+c_2t_2$.
It suffices to find $\zeta_0'$ such that $\diff \pi(\diff \phi_{t_1})(\zeta_0') = \lambda_1 \dot{\gamma}_{p_0,v_0}(t)$.
Such $\zeta_0'$ can be found by Claim \ref{cl_1}.
In this way, the term with $\dot{\gamma}_{p_0,v_0}(t_1)$ in the first equation can replaced by considering $\tilde{\zeta}_0 = \zeta_0 + \zeta_0'$.
By Proposition \ref{conjugate_J} we have $p_2$ is conjugate to $p_1$.
\end{proof}
The same conclusion is proved in \cite[Theorem 5.3]{Assylbekov2017} based on the definition of the Jacobi fields as the variational fields.
\subsection{The Jacobi equation.}
In this subsection, we show the following proposition in the case of $\lambda$-geodesic flow, analogous to \cite[Proposition 4.14 ]{Merry2011} in the case of the geodesic flow.
\begin{pp}[{\cite[Lemma 4.6]{Assylbekov2017}}]\label{dphi_t}
For any $\zeta \in T_{(p,v)} SM$, if we write $\diff \phi_t(\xi)$
\[
\diff \phi_t(\zeta) = x(t) F(\phi_t(x,v)) + y(t) H(\phi_t(x,v)) + z(t)V(\phi_t(x, v)),
\]
in the moving frame $\{F, H, V\}$,
then the smooth functions $x,y,z$ should satisfy
\begin{align*}
&\dot{x} = \lambda y,\\
&\dot{y} = z,\\
& \dot{z} = -({K-H(\lambda) + \lambda^2}) y + V(\lambda) z.
\end{align*}
\end{pp}
Combining the last two equations above, one can see that $y$ should satisfy a second-order ODE.
Notice $\dot{\gamma}(t)^{\perp}, \dot{\gamma}(t)$ give us a moving frame along $\gamma_{p_0,v_0}(t)$.
This proves the following proposition analogous to \cite[Lemma 5.2]{Assylbekov2017}.
\begin{pp} \label{jacobi_eq}
If $J(t)$ is a Jacobi field along $\gamma_{p_0,v_0}$ expressed as
\[
J(t) = x(t) \dot{\gamma}_{p_0,v_0}(t) + y(t)\dot{\gamma}_{p_0,v_0}^\perp(t),
\]
then $x(t), y(t)$ satisfy the Jacobi equations
\begin{align}
&\dot{x} = \lambda y + c, \label{Jeq_x}\\
&\ddot{y} - V(\lambda) \dot{y} + (K-H(\lambda) + \lambda^2)y = 0, \label{Jeq_y}
\end{align}
where $c$ is a constant.
\end{pp}
\begin{remark}
If $\lambda=0$, we have the case of geodesics.
Compared to \cite[Lemma 5.2]{Assylbekov2017}, we have an extra constant $c$ in (\ref{Jeq_x}), since by extending $\Gamma$ we include $t \dot{\gamma}(t)$ in our definition of the Jacobi field.
\end{remark}
\begin{proof}
Recall $J(t)$ has the representation
\[
J_{(p_0,v_0)}(t) = \diff \pi(\diff \phi_t)(\zeta) + (c_1 + c_2t) \dot{\gamma}_{p_0,v_0}(t).
\]
Applying $\diff \pi$ to Proposition \ref{dphi_t}, we have the desired result by (\ref{projection_frame}).
\end{proof}
\section{Examples of $\lambda$-geodesics with conjugate points}\label{sec_example}
In this section, we present several examples of the family of curves that we study.
These curves are different from geodesics and we show the conjugate points exist.
\begin{example}\label{ex_c}
The first example comes from \cite{MR2970707}.
Let $\Gamma_1$ consist all the unit circles in $\mathbb{R}^2$ with a fixed orientation.
These circles are actually the magnetic geodesics w.r.t the Euclidean metric and a constant non-zero magnetic field by \cite{MR2351370}.
Suppose they have a fixed orientation and we parameterize a unit circle through point $p =(p^1, p^2)$ in the direction of $v = (\cos \theta, \sin \theta)$ by
\[
\gamma_{p, \theta}(t) = (p^1, p^2) + (\cos(t + \theta), \sin (t+ \theta)) - (\cos \theta, \sin \theta).
\]
We have
\[
\dot{\gamma}_{p, \theta}(t) = (-\sin(t + \theta), \cos (t+ \theta)), \quad \ddot{\gamma}_{p, \theta}(t) = - (\cos(t + \theta), \sin (t+ \theta)),
\]
and therefore
\[
\ddot{\gamma}_{p, \theta}(t) =
\begin{pmatrix}
0 & -1 \\
1 & 0
\end{pmatrix}
\dot{\gamma}_{p, \theta}(t).
\]
One can check that the properties (A1) and (A2) are satisfied.
As is shown in \cite{MR2970707}, since
\begin{align}\label{eq_dexp}
\det \diff_{s, v} \exp_p(t,v) =\det \frac{\partial \gamma_{p, \theta}(t)}{\partial (t, \theta)} = \sin t,
\end{align}
for each point $p$, it has conjugate points corresponding to $t = \pi$ and any $\theta \in \mathbb{R}$.
Thus, the conjugate locus of $p$ is
\[
\Sigma(p) = \{\gamma_{p, \theta}(\pi), \ \theta \in \mathbb{R}\} = \{y:\ |y-p| = 2\},
\]
which is the circle centered at $p$ with radius equal to $2$.
Now let $(p_1, \xi^1) \in T^*M$
and suppose $f$ is singular near $(p_1, \xi^1)$.
We would like to find out the conjugate covector $(p_2, \xi^2)$ of $(p_1, \xi^1)$.
To parameterize $\Gamma_1$, we assume $p_1$ is near the origin and $\xi^1$ is in a conic neighborhood of $\xi^0 = (1,0)$.
Consider the line $H = \{(y, 0), y \in \mathbb{R}\}$.
By the notations used in Section \ref{sec_microlocal}, a unit circle passing the point $(y, 0) \in \mathbb{R}^2$ in the direction of $v = (\cos \eta, \sin\eta)$ is now parameterized by
\[
\gamma(t, y,\eta) = (y,0) + (\cos(t + \eta), \sin (t+ \eta)) - (\cos \eta, \sin \eta).
\]
Suppose $p_1 = (y_1, 0)$ and
$
\xi^1 = (\cos \eta_1, \sin \eta_1).
$
By Theorem \ref{thm_C_diff}, $(p_2, \xi^2)$ and $(p_1, \xi^1)$ are conjugate covectors when
$p_2$ is conjugate to $p_1$ and
\[
\xi^1 = \mu a_1(t_2) \dot{\gamma}^\perp (t_1), \quad \xi^2 = \mu a_1(t_1) \dot{\gamma}^\perp (t_2)
\] for some nonzero $\mu$.
Then by equation (\ref{eq_dexp}), we write
\[
p_2 = \gamma(\pi, y_1, \eta_1), \quad \text{or } p_2 = \gamma(\pi, y_1, -\eta_1)
\]
with $t_1 = 0, t_2 = \pi$, which depends on the different sign of $\mu$.
We denote the first case by $(p_2^{(1)}, \xi^2_{(1)})$ and the second case by $(p_2^{(2)}, \xi^2_{(2)})$.
We compute
\[
p^{(1)}_2 = (y_1, 0) - 2(\cos \eta_1, \sin \eta_1), \quad p^{(2)}_2 = (y_1, 0) + 2(\cos \eta_1, \sin \eta_1).
\]
Since $a_1(t) = (\dot{\gamma}^\perp(t), \frac{\partial \gamma(t,y,\eta)}{\partial y})$,
we can find
\[
\xi^2_{(1)} = \xi^1, \quad \xi^2_{(2)} = \xi^1.
\]
Therefore, there are two covectors that are conjugate to $(p_1, \xi^1)$ given by
\begin{equation}\label{eq_cc}
(p_2^{(1)}, \xi^2_{(1)}) = (p_1 - 2(\cos \eta_1, \sin \eta_1), \xi^1), \quad (p_2^{(2)}, \xi^2_{(2)}) = (p_1 + 2(\cos \eta_1, \sin \eta_1), \xi^1).
\end{equation}
\begin{example}
In this example, let $\Gamma_2$ consist all identical ellipses with a fixed orientation starting from different points
\[
\gamma_{p}(t) = (p^1, p^2) + (a\cos t -a , b\sin t),
\]
where $a, b$ are constants. Different values of $a, b$ give us different $\Gamma_2$.
We have
\[
\dot{\gamma}_{p}(t) = (-a\sin t, b\cos t), \quad \ddot{\gamma}_{p}(t) = -(a\cos t, b\sin t).
\]
One can verify that $\Gamma$ satisfies the properties (A1) and (A2).
Indeed, first we have
\[
\ddot{\gamma}_{p}(t) =
\begin{pmatrix}
0 & -{a}/{b}\\
{b}/{a} & 0
\end{pmatrix}
\dot{\gamma}_{p}(t).
\]
Then, for any point $x = (x_1, x_2)$ and $v = (\cos \theta, \sin \theta)$,
there is a curve $\gamma_p$ passing $x$ in the direction of $v$ if we can find some $p_0, t_0$ such that
\[
-a \sin t_0 = \mu \cos \theta, \quad b \cos t_0 = \mu \sin \theta, \quad {\gamma}_{p}(t) = x,
\]
for some $\mu >0$. This is true since we can solve $t_0$ from
\[
\cos t_0 = \frac{-\mu }{a} \cos \theta,
\quad \sin t_0 = \frac{\mu }{b}\sin \theta,
\]
where $ \mu = \frac{ab}{\sqrt{b^2 \cos^2 \theta + a^2 \sin^2 \theta}}$ and then
$p_0= x - (a\cos t_0 -a, b \sin t_0) $.
One can see that $t_0$ is unique in the sense of modulo $2\pi$ and it depends on $\theta$ in a smooth way.
We denote it by $t_0(\theta)$ and compute
\begin{gather*}
\frac{\diff t_0}{\diff \theta} = \frac{b^2 \cos^2 \theta + a^2 \sin^2 \theta}{ab}. \\
\end{gather*}
To parametrize $\Gamma_2$, we pick $H = {(0, x^2)}$.
Notice for every point $x = (0, x^2) \in H$ and any direction $v = (\cos \theta, \sin \theta)$,
there is a unique $\gamma_p(t)$ passing $x$ in the direction of $v$.
From the analysis above, we reparameterize this ellipse as
\[
\rho_{x^2, \theta}(t) = (a\cos(t+t_0 )- a \cos t_0, b \sin(t + t_0)-b \sin t_0 + x^2).
\]
Thus, for any $x = (0, x^2)$ we have
\[
\det \diff_{s, v} \exp_x(t,v) =\det \frac{\partial \rho_{x^2, \theta}(t)}{\partial (t, \theta)} = \frac{\diff t_0}{\diff \theta} ab \sin t,
\]
and its conjugate points are corresponding to $t = \pi$ and any $\theta$.
This gives us the conjugate locus $\Sigma(x) = \{\rho_{x^2, \theta}(\pi), \ \theta \in \mathbb{R}\} = \{x - 2(a\cos t_0, b\sin t_0), t_0 \in \mathbb{R}\}$,
which is a larger ellipse centered at $x$.
\end{example}
\begin{example}
In this example, we choose the a family of smooth curves that are locally defined.
For $p = (p^1, p^2)$, define
\[
\gamma_p(t) = e^{at} (\cos(t - \varphi) - \cos (-\varphi), \sin (t - \varphi) - \sin (-\varphi)) + p,
\]
where $a$ is a nonzero constant and $\varphi$ satisfies
\[
\cos \varphi = \frac{a}{\sqrt{a^2 + 1}}, \quad \sin \varphi =\frac{1}{\sqrt{a^2 + 1}}.
\]
We compute
\[
\dot{\gamma}_p(t) = e^{at} \sqrt{a^2 + 1}(\cos t, \sin t), \quad \ddot{\gamma}_p(t) = e^{at} \sqrt{a^2 + 1}(a\cos t - \sin t , a\sin t + \cos t),
\]
and one can show that
\[
\ddot{\gamma}_{p}(t) =
\begin{pmatrix}
a & -1\\
1 & a
\end{pmatrix}
\dot{\gamma}_{p}(t),
\]
which implies the property (A2).
To satisfy the property (A1) at least locally,
we fix a curve
\[
\gamma_0(t) = e^{at} (\cos(t - \varphi), \sin (t - \varphi)),\text{ for } t \in (-\frac{\pi}{4}, \frac{7 \pi}{4})
\]
which passes the point $(0,0)$ at $t=0$ in the direction $\theta = 0$.
Now let
\[
\Gamma_3 = \{ \gamma_0(t) + p, \text{ where } t \in (-\frac{\pi}{4}, \frac{7 \pi}{4}) \}.
\]
Then $\Gamma_3$ is an open family of smooth curves.
For each point $x = (x_1, x_2)$ and direction $v = (\cos \theta, \sin \theta)$ with $\theta \neq -\frac{\pi}{4}$,
there is a unique
\[
t_0 = \theta, \quad p_0 = x - e^{a \theta}(\cos(\theta - \varphi), \sin(\theta - \varphi) + (\cos \varphi, -\sin \varphi),
\]
such that $\gamma_{p_0}(t_0)$ through $x$ in the direction of $v$.
We reparameterize $\Gamma_3$ using $H = {(0, x^2)}$ by writing these curves as
\begin{align*}
\rho_{x^2, \theta}(t) &= e^{a(t+\theta)} (\cos(t +\theta - \varphi) , \sin (t +\theta - \varphi) )\\
&- e^{a\theta}(\cos (\theta-\varphi), \sin (\theta-\varphi))) + (0,x^2),
\text{ for } t \in (-\frac{\pi}{4}, \frac{7 \pi}{4}).
\end{align*}
Notice that $\rho_{x^2, \theta}(t)$ passes $(0,x^2)$ at $t=0$ in the direction of $\theta$.
We mention that $\Gamma_3$ forms a neighborhood of $\rho_{0, 0}(t)$
and we compute
\[
\det \diff_{s, v} \exp_x(t,v) =\det \frac{\partial \rho_{x^2, \theta}(t)}{\partial (t, \theta)} = e^{2a + \theta}\sin t.
\]
Thus, for the point $x = (0 ,0)$, there are conjugate points corresponding to $t = \pm \pi$ and $\theta \in (-{\pi}/{4}, {7 \pi}/{4})$.
The conjugate locus is
\begin{align*}
\Sigma(x) &= \{\rho_{0, \theta}(\pm\pi), \ \theta \in \mathbb{R}\} \\
&= \{ - (e^{\pm a\pi} + 1)e^{a\theta}(\cos (\theta - \varphi), \sin (\theta - \varphi)),
\text{ for } \theta \in (-\frac{\pi}{4}, \frac{7 \pi}{4}) \}.
\end{align*}
In Figure \ref{picref}, we choose $a = -1/4$ (this implies $\varphi = \arccos(1/\sqrt{17})$) and we plot all $\rho_{0, \theta}(t)$ for $\theta \in (-{\pi}/{4}, {\pi}/{4})$. The light green curve is part of the conjugate locus $\Sigma(x)$.
\begin{figure}[h]
\includegraphics[height=0.3\textwidth]{fig/newconjugate}
\caption{ Light green curve: part of conjugate locus $\Sigma(x)$. }
\label{picref}
\end{figure}
\end{example}
\end{example}
\section{Introduction}
In this work, we consider the integral transform over a generic family of smooth curves with conjugate points.
Let $M \subset \mathbb{R}^2$ be a bounded domain with a Riemannian metric $g$.
We consider a family $\Gamma$ of smooth curves satisfying the following properties.
\begin{itemize}
\item[(A1)] For each $(x,v) \in TM\setminus0$, there is exactly one unique curve $\gamma(t) \in \Gamma$ passing $x$ in the direction of $v$. We can assume $\gamma(0) = x$ and $\dot{\gamma}(0) = \mu(x,v) v$ with a smooth function $\mu(x,v) >0$ and denote it by $\gamma_{x,v}(t)$.
\item[(A2)] Suppose $\gamma_{x,v}(t)$ depends on $(x,v)$ smoothly, and thus we assume
it solves a second order ODE
\begin{equation*
D_t\dot{\gamma}_{x,v}(t) = G(\gamma_{x,v},\dot{\gamma}_{x,v}),
\end{equation*}
where $D_t$ is the covariant derivative along $\dot{\gamma}$ and the generator $G$ is a smooth map invariantly defined on $TM$
with the transformation law
\[
\widetilde{G}^i(\tilde{x}^i, \tilde{y}^i) = \frac{\partial \tilde{x}^i}{\partial x^j} G^j(x^i,y^i), \quad \text{ where }\tilde{y}^i = \frac{\partial \tilde{x}^i}{\partial x^j} y^j,
\]
under the coordinate change $x^i \mapsto \tilde{x}^i$ of $M$.
\end{itemize}
We emphasize that the second property describes a large family of curves, including geodesics in Riemannian surfaces, geodesics in Finsler spaces, and magnetic geodesics, for more details see \cite[Chapter 4]{Bucataru2007}, \cite[Chapter 3]{Shen2013}, \cite{MR2351370}, and \cite{douglas1927general}.
For such a curve, we can freely shift the parameter but rescaling it may give us a different one, since the trajectory depends on the energy level.
This implies the first property is necessary for us to have a geodesic-like family of curves.
We define the integral transform along curves in $\Gamma$ as
\[
I_w f (\gamma)= \int w(\gamma(t),\dot{\gamma}(t)) f(\gamma(t)) dt, \quad \gamma \in \Gamma,
\]
for any distribution $f$ compactly supported in $M$.
This integral transform has been studied in \cite{Frigyik2007}.
It is shown that if $\Gamma$ is analytic and regular, i.e., if $G$ is analytic and $N^*\Gamma$ covers $T^*M$ without conjugate points,
then $I_w$ with an analytic and nonvanishing weight function is injective. A stability result for the localized normal operator is established there.
In \cite{Assylbekov2017}, the integral transform of functions and $1$-forms over such curves in an oriented Finsler surface is studied and the injectivity is proved if there are no conjugate points.
However, when conjugate points exist, in many cases the microlocal stability are lost.
The phenomenon of cancellation of singularities due to the existence of conjugate points are fully studied in the case of synthetic aperture radar imaging in \cite{MR3080199}, the X-ray transforms over a family of geodesic-like curves in \cite{MR2970707}, the geodesic ray transforms in \cite{Holman2015,Monard2015, Holman2017}. In particular, \cite{Monard2015} shows
that regardless of the type of the conjugate points, the geodesic ray transforms on Riemannian surfaces are
always unstable and we have loss of all derivatives, which leads to the artifacts in the reconstruction near
pairs of conjugate points.
The same phenomenon occur in the inversion of broken ray transforms with conjugate points, see \cite{Zhang2020}.
For more references about the integral transform over geodesics, magnetic geodesics, or a general family of curves, see \cite{Ainsworth, homan2017injectivity, kunstmann2021, Geometric,zhou2018local}.
This work is inspired by \cite{Monard2015} and we prove an analog of cancellation of singularities for $I_w$ in Theorem \ref{thm_C_diff} and \ref{thm_microkernel}.
We prove that the local problem is ill-posed if there are conjugate points, i.e., singularities
conormal to smooth curves in $\Gamma$ cannot be recovered uniquely.
The microlocal kernel is described in Theorem \ref{thm_microkernel}.
In Section \ref{lambda}, we introduce $\lambda$-geodesics to describe this generic family of smooth curves and we define the conjugate points.
In Section \ref{sec_microlocal}, we consider the recovery of $f$ from the integral transform $I_w$ using the local data.
Here by the local data, we mean the integral transform is only known in a small neighborhood of a fixed curve $\gamma(t) \in \Gamma$.
The transform $I_w$ is an Fourier integral operator (FIO) with its canonical relation described in Proposition \ref{pp_C}.
We microlocalize the problem and prove Theorem \ref{thm_C_diff} and Theorem \ref{thm_microkernel}.
In Section \ref{sec_example}, we present several examples of such family of curves and show the existence of conjugate points.
In Section \ref{sec_num}, we illustrate the artifacts arising
in the reconstruction by numerical experiments.
In particular, we show for the family of smooth curves in Example \ref{ex_c}, the singularities of $f$ can be recovered without artifacts if we consider the global data and have the prior knowledge that $f$ is a distribution with compact support.
\subsection*{Acknowledgments}
The author would like to thank Plamen Stefanov for suggesting this problem and for numerous helpful discussions with him throughout this project, and to thank Gabriel Paternain and Gunther Uhlmann for helpful suggestions.
Part of this research was supported by NSF Grant DMS-1600327, the Simons Travel Grants, and was performed while the author was visiting IPAM.
\section{The $\lambda$-geodesics}\label{lambda}
\subsection{The generator $G$}
Locally we can rewrite the second order ODE as
\[
\ddot{\gamma}^i_{x,v}(t) = G^i(\gamma_{x,v}, \dot{\gamma}_{x,v}) - \dot{\gamma}_{x,v}^k\dot{\gamma}_{x,v}^l\Gamma_{kl}^i, \quad i = 1, 2,
\]
where $\Gamma_{kl}^i$ is the Christoffel symbol.
The generator $G$ induces a vector field $\mathbf{G}$ on $TM$ given by
\[
\mathbf{G} = y^i \frac{\partial }{\partial x^i} + (G^i(x, y) - y^ky^l \Gamma_{kl}^i)\frac{\partial }{\partial y^i}
\]
in a local chart $\{x^i, y^i\}$ of $TM$.
It preserves the form under coordinate changes by the following arguments and thus is invariantly defined, see also (30) in \cite{Frigyik2007}.
Indeed, consider a local coordinate chart $\{x^i, y^i\}$ near a fixed point $(x, y) \in TM$.
Under the coordinate change $x^i \mapsto \tilde{x}^i$ of $M$, we have $\tilde{y}^i = ({\partial \tilde{x}^i}/{\partial x^j}) y^j$.
The double tangent bundle $T_{(x,y)} TM$ has a natural basis $\{{\partial }/{\partial {x}^i},{\partial }/{\partial {y}^i}\}$, which transforms as
\begin{align*}
&\frac{\partial }{\partial {x}^i} = \frac{\partial \tilde{x}^j}{\partial x^i} \frac{\partial }{\partial \tilde{x}^j} + \frac{\partial \tilde{y}^j}{\partial x^i} \frac{\partial }{\partial \tilde{y}^j}
= \frac{\partial \tilde{x}^j}{\partial x^i} \frac{\partial }{\partial \tilde{x}^j} + y^k \frac{\partial^2 \tilde{x}^j}{\partial x^i\partial x^k} \frac{\partial }{\partial \tilde{y}^j}, \\
&\frac{\partial }{\partial {y}^i} = \frac{\partial \tilde{x}^j}{\partial x^i} \frac{\partial }{\partial \tilde{y}^j}.
\end{align*}
In addition, the Christoffel symbol transforms as
\[
{\Gamma}_{kl}^i = \frac{\partial x^i}{\partial \tilde{x}^m}\frac{\partial \tilde{x}^n}{\partial x^k} \frac{\partial \tilde{x}^p}{\partial x^l} \tilde{\Gamma}_{np}^m + \frac{\partial^2 \tilde{x}^m}{\partial x^k \partial x^l} \frac{\partial x^i}{\partial \tilde{x}^m}.
\]
We plug these transformation laws above in $\textbf{G}$ to have
\begin{align*}
\textbf{G} &= y^i \frac{\partial }{\partial x^i} + (G^i(x, y) - y^ky^l \Gamma_{kl}^i)\frac{\partial }{\partial y^i} \\
&= y^i\frac{\partial \tilde{x}^j}{\partial x^i} \frac{\partial }{\partial \tilde{x}^j}
+ \big(y^iy^k \frac{\partial^2 \tilde{x}^j}{\partial x^i\partial x^k} + (G^i - y^ky^l (\frac{\partial x^i}{\partial \tilde{x}^m}\frac{\partial \tilde{x}^n}{\partial x^k} \frac{\partial \tilde{x}^p}{\partial x^l} \tilde{\Gamma}_{np}^m + \frac{\partial^2 \tilde{x}^m}{\partial x^k \partial x^l} \frac{\partial x^i}{\partial \tilde{x}^m})) \frac{\partial \tilde{x}^j}{\partial x^i}\big) \frac{\partial }{\partial \tilde{y}^j}\\
&= \tilde{y}^j \frac{\partial }{\partial \tilde{x}^j} + (\tilde{G}^j - \tilde{y}^k\tilde{y}^l \tilde{\Gamma}_{kl}^j)
\frac{\partial }{\partial \tilde{y}^j}.
\end{align*}
\subsection{Reparameterize $\gamma$}
Since $\dot{\gamma}$ never vanishes, we reparameterize these curves such that they have unit speeds w.r.t. the Riemannian metric $g$.
Note that this reparameterization will also change the generator $G(\gamma, \dot{\gamma})$ and the weight function in the integral transform.
In the following, we abuse the notation and use $\gamma_{x,v}(t)$ to denote the smooth curve with arc length passing $(x,v) \in SM$, where $SM$ is the unit circle bundle.
The family of smooth curves with this arc length parameterization solves a new ODE
\[
D_t \dot{\gamma} = \lambda(\gamma, \dot{\gamma}) \dot{\gamma}^{\perp},
\]
and are referred to as the $\lambda$-geodesics in \cite[Chapter 7]{Merry2011},
where $ \lambda(\gamma, \dot{\gamma}) \in C^{\infty}(SM)$ and $\dot{\gamma}^{\perp}$ is the vector normal to ${\gamma}^{\perp}$ with the same length and the rotation of $\pi/2$.
When $\lambda$ is a smooth function on $M$, the corresponding $\lambda$-geodesic flow is called the magnetic flow, see \cite{ginzburg1996closed,MR2351370}.
When $\lambda$ is a smooth function on $SM$, it is called the thermostats, see \cite{dairbekov2007entropy}.
\subsection{Extend $\Gamma$}\label{subsection_extend}
Moreover, it is convenient to extend $\Gamma$,
such that for any $V = rv \in T_pM\setminus 0$, with $r>0$ and $v\in S_pM$, there exists a unique curve $\gamma_{x,V}(t)$ belonging to $\Gamma$.
This can be done by defining $\gamma_{x,V}(t) = \gamma_{x,v}(rt)$.
We note that $\gamma_{x,V}(t)$ satisfies the following ODE
$$
D_t \dot{\gamma}_{x,V}(t) = |V|_g \lambda({\gamma}_{x,V}(t),\frac{1}{|V|_g}\dot{\gamma}_{x,V}(t)) \dot{\gamma}^\perp_{x,V}(t),
$$
which is equivalent to extend $\lambda(x, v)$ to $C^{\infty}(TM\setminus0)$.
As we mentioned before, for a curve satisfying the equation in Property (A2) in general, rescaling it might not give us the same one, i.e.,
$\gamma_{p,v}$ and $\gamma_{p,rv}$ are different curves.
However, with the reparameterization and the extension above, the rescaling will cause no problem.
\section{Microlocal analysis of the local problem}\label{sec_microlocal}
\subsection{Parameterization of $\mathcal{M}$}
Since the group action of $\mathbb{R}$ on $M$ by $\phi_t$ is free and proper,
the family of curves form a smooth manifold $\mathcal{M} = SM/\phi_t$ of dimension $2\times 2-2=2$.
We parameterize $\mathcal{M}$ below near a fixed curve $\gamma \in \Gamma $, in the same way for the case of the geodesics in \cite[Section 3.1]{Monard2015}.
This parameterization gives us a local charts of the manifold of curves.
For this purpose, for fixed $\gamma$ we choose a hypersurface $H$ such that $\gamma$ hits $H$ transversally at $p_0 = \gamma(0)$ and $v_0 = \dot{\gamma}(0)$.
Let $\mathcal{H}$ consist of all $(p, v) \in SM$ in a small neighborhood of $(p_0, v_0)$,
such that $p \in H$ and $v$ is transversal to $H$.
Then we can parameterize all curves near fixed $\gamma$
by the intersection point $p$ and the direction $v$, i.e., by an element $(p,v) \in \mathcal{H}$.
Suppose $H$ is locally given by $x^1=0$ and has coordinate $y$.
We can write each element in $\mathcal{H}$ using $(y, \eta)$ instead, where $\eta$ is the parameterization of $v \in S_{(0,y)}M$.
In the following, we use the new notation $\gamma(y, \eta, t)$ to denote the smooth curve starting from $p = (0,y)$ in the direction $v$.
In some cases, we write $\gamma(t), \frac{\partial \gamma}{\partial y}(t), \frac{\partial \gamma}{\partial \eta}(t)$ and omit the variables $y, \eta$ for simplification, if there is no confusion caused.
\begin{pp}\label{jocabifield_simplied}
With this parameterization, the Jacobi field $J(t)$ along $\gamma(y, \eta, t)$ can be simplified as
\[
J(t) = \kappa_1 \frac{\partial \gamma}{\partial y}(t)
+ \kappa_2 \frac{\partial \gamma}{\partial \eta}z(t) + c_1\dot{\gamma}(t) + c_2t \dot{\gamma}(t),
\]
where $\kappa_1, \kappa_2, c_1, c_2 \in \mathbb{R}$.
\end{pp}
\begin{proof}
It is shown in Claim \ref{cl_1} that for any $c \in \mathbb{R}$, there exists $\zeta \in T_{(p,v)}SM $ such that
$$
S(t) = \diff \pi(\diff \phi_t)(\zeta) + c \dot{\gamma}(y, \eta, t) \equiv 0,
$$
where in local coordinates $p =(0,y)$ and we assume $c \neq 0$.
On the other hand, in the local coordinates $(x^1,y, \eta)$ of $SM$,
the pushforward $\diff \phi_t$ has a representation and
we can write
\[
\diff \pi(\diff \phi_t)(\zeta) = \kappa_1 \frac{\partial \gamma}{\partial y} +
\kappa_2 \frac{\partial \gamma}{\partial x^1} + \kappa_3 \frac{\partial \gamma}{\partial \eta}.
\]
Then it suffices to show that $\kappa_2 \neq 0$ and therefore $\frac{\partial \gamma}{\partial x}$ can be written as a linear combination of $\frac{\partial \gamma}{\partial y},\frac{\partial \gamma}{\partial \eta},\dot{\gamma}$.
Indeed, we have the following initial conditions
\begin{equation}\label{eq_initial}
\frac{\partial \gamma}{\partial \eta}(0) = 0, \quad (\frac{\partial \gamma}{\partial y}(0),\dot{\gamma}^\perp(0))_g \neq 0,
\end{equation}
where the first one comes from $\gamma(y, \eta, 0) \equiv p$ and the second one is from the assumption that $\gamma(y, \eta, t)$ hit $H$ transversally at $t=0$.
Now we compute
\[
0 \equiv (S(0),\dot{\gamma}^\perp(0))
=\kappa_1 (\frac{\partial \gamma}{\partial y}(0),\dot{\gamma}^\perp(0))_g +
\kappa_2 (\frac{\partial \gamma}{\partial x^1}(0),\dot{\gamma}^\perp(0))_g.
\]
It follows that $\kappa_2 \neq 0$, otherwise one has $\kappa_2 = \kappa_1 = 0$,
which leads to \[
\kappa_3 \frac{\partial \gamma}{\partial \eta} + c \dot{\gamma}(y, \eta, t) = 0.
\]
This contradicts with the initial condition in (\ref{eq_initial}) and that $\dot{\gamma}(y,\eta,t)$ is nonvanishing.
\end{proof}
\subsection{$I_w$ as an FIO}
The same analysis for the geodesic transform in \cite[Section 3.1]{Monard2015} works for $I_w$.
We briefly state these arguments for $I_w$ in this subsection.
For more details see \cite[Section 3.1]{Monard2015}.
We consider the curve-point relation
\[
Z = \{(\gamma(y,\eta, t), x) \in \mathcal{M} \times M: \ x \in \gamma \},
\]
from the double fibration point of view of \cite{Gelextquotesinglefand1969,Helgason2013}.
Note that $Z$ is a smooth manifold
with the conormal bundle
\[
N^* Z = \{(\gamma, \hat{\gamma}, x, \xi)\in T^*(\mathcal{M} \times M)\setminus 0: \ (\hat{\gamma}, \xi) \text{ vanishes on } T_{(\gamma, x)} Z_0\}.
\]
The integral transform along curves in $\Gamma$ defined by
\[
I_w f (\gamma)= \int w(\gamma(t),\dot{\gamma}(t)) f(\gamma(t)) \diff t, \quad f \in \mathcal{E}'(M),\ \gamma \in \Gamma.
\]
has a delta-type Schwartz kernel $w \delta_{Z}$, which is a conormal distribution supported in $N^*Z$.
Therefore, it is an FIO of order $-{1}/{2}$ and associated with the canonical relation $C = N^*Z'$, the twisted version of $N^*Z$ given by
\[
C = \{(\gamma, \hat{\gamma}, x, -\xi): \ \{(\gamma, \hat{\gamma}, x, \xi)\in N^*Z\}.
\]
By the local parameterization of $\mathcal{M}$, we can show $(y,\eta,\hat{y},\hat{\eta},x,\xi) \in C$ if and only if there exists $t$ with $x = \gamma(y,\eta,t)$ and
\begin{align}\label{conditioncanonical}
\xi_i \dot{\gamma}^i = 0,
\quad \xi_i \frac{\partial \gamma^i}{\partial y}= \hat{y},
\quad \xi_i \frac{\partial \gamma^i}{\partial \eta}= \hat{\eta},
\end{align}
where the Einstein summation convention is used.
\subsection{The description of $C$}
Inspired by \cite[Section 3.2]{Monard2015}, in this subsection, we write $C$ in a more explicit form and describe its properties.
Let \[
e_1(t) = \dot{\gamma}(t)^{\perp} , \quad e_2(t) = \dot{\gamma}(t)
\] be a moving frame.
We have the corresponding dual basis $\{e^1(t), e^2(t)\}$.
If we regard $\xi$ as a function of $t$, then the first condition in (\ref{conditioncanonical}) means
\[
\xi(t) = h(t) e^1(t),
\]
for some nonvanishing function $h(t)$.
Suppose we have the following expansion w.r.t the frame
\begin{align}
\frac{\partial \gamma}{\partial y} = a_1(t)e_1(t) + a_2(t) e_2(t), \quad
\frac{\partial \gamma}{\partial \eta} = b_1(t)e_1(t) + b_2(t) e_2(t).
\end{align}
The second and third conditions in (\ref{conditioncanonical}) imply
\begin{align*}\label{secondthirdcon}
h(t) a_1(t) = \hat{y}, \quad h(t) b_1(t) = \hat{\eta}.
\end{align*}
Therefore, the canonical relation has the local representation
\begin{align*}
C = \{ (y,\eta,
\underbrace{\mu a_1(t,y,\eta)}_{\hat{y}},
\underbrace{\mu b_1(t,y,\eta)}_{\hat{\eta}},
\gamma(t,y,\eta),
\mu \dot{\gamma}^\perp (t,y,\eta) ): \
(y, \eta) \in \mathcal{H}, \ \mu \neq 0 \},
\end{align*}
where
\begin{equation}
a_1(t,y,\eta) = (e_1(t), \frac{\partial \gamma(t,y,\eta)}{\partial y})_g, \quad
b_1(t,y,\eta) = ( e_1(t), \frac{\partial \gamma(t,y,\eta)}{\partial \eta})_g \label{jeq_ab}
\end{equation}
are projections along $\dot{\gamma}^\perp(t,y,\eta)$.
The canonical relation $C$ can be parameterized by $(y,\eta, t, \mu )$ and therefore we have $\dim C = 4$.
\begin{pp}\label{pp_C}
Let $\pi_M: C \rightarrow T^*M\setminus 0 $ and $\pi_\mathcal{M}: C \rightarrow T^*\mathcal{M}\setminus 0 $ be the natural projections.
Then
\begin{itemize}
\item[(a)] $\pi_M: (y, \eta, \hat{y}, \hat{\eta}, x, \xi) \mapsto (x, \xi) $ is a diffeomorphism,
\item[(b)] $\pi_\mathcal{M}: (y, \eta, \hat{y}, \hat{\eta}, x, \xi) \mapsto (y, \eta, \hat{y}, \hat{\eta})$ is a local diffeomorphism.
\end{itemize}
\end{pp}
\begin{proof}
We prove (a) first. For each $(x,\xi)\in T^*M\setminus 0$, with
the property (A1), there is a unique curve $\gamma$ passing $x$ at time $t$ and conormal to $\xi$.
Suppose $\gamma$ hits $H$ transversally at $y$ in the direction parameterized by $\eta$, at $t = 0$.
Then $(y,\eta)$ can be derived from the flow $\phi_{-t}(x,v)$, composed with
the restriction on $H$, where $v$ is the unit vector corresponds to $\xi$.
Thus, in a small neighborhood $(y,\eta)$ depends on $(x,\xi)$ smoothly and $(\hat{y},\hat{\eta})=(\mu a_1(t,y,\eta), \mu b_1(t,y,\eta))$ with $\mu = |\xi|$ also depends on $(x,\xi)$ in a smooth way, if we are away from the zero section. This implies one can regard $(x, \xi)$ as a parameterization of $C$ and $\pi_M$ is a diffeomorphism.
For (b), in local parameterization we have $\pi_\mathcal{M}: (y, \eta, t, \mu) \mapsto (y, \eta, \mu a_1, \mu b_1)$.
We compute $\diff \pi_\mathcal{M}$ to have
\[
\det \diff \pi_\mathcal{M} = \mu (\dot{a}_1 b_1 - \dot{b}_1 a_1) \equiv \mu W(t),
\]
where $W(t) $ is the Wronskian and it satisfies $\dot{W}(t) = V(\lambda) W(t)$.
Indeed,
recall $a_1, b_1$ are the projections of Jacobi fields onto $\dot{\gamma}^\perp(t)$
and therefore both of they should satisfy the equation in (\ref{Jeq_y}).
Solving the equation of the Wronskian, we have $W(t) = W(0) e^{\int V(\lambda) \diff t}$,
which indicates $W(t)$ is zero at any $t$, if it vanishes at some point.
When $W(t)$ vanishes, we have $a_1(t) = c b_1(t)$ for some constant $c$.
However, since
$a_1(0) \neq 0$ and
$b_1(0) = 0$ by the initial condition (\ref{eq_initial}),
it is impossible
and therefore we have $W(t) \neq 0$.
\end{proof}
We say the integral transform $I_w$ satisfies the Bolker condition, see \cite{Guillemin1979}, if the projection $\pi_\mathcal{M}$ is an injective immersion.
\begin{thm}\label{thm_C_diff}
The map $\mathcal{C}(x,\xi) \equiv \pi_\mathcal{M} \circ \pi_M^{-1}(x,\xi)$
is a local diffeomorphism from $T^*M\setminus 0$ to $T^*\mathcal{M}\setminus 0$.
Moreover, we have $\mathcal{C}(p_1, \xi^1) = \mathcal{C}(p_2, \xi^2)$ if and only if there is a curve $\gamma(t,y,\eta)$ joining $p_1$ at $t_1$ and $p_2$ at $t_2$, with $t_2 > t_1$, such that
\begin{itemize}
\item[(a)] $p_2$ is conjugate to $p_1$.
\item[(b)] $\xi^1 = \mu a_1(t_2) \dot{\gamma}^\perp (t_1)$ and $\xi^2 = \mu a_1(t_1) \dot{\gamma}^\perp (t_2)$ for some nonzero $\mu$, where $a_1(t)$ is defined in (\ref{jeq_ab}).
\end{itemize}
In particular, if there are no conjugate points, then the Bolker condition is satisfied and $\mathcal{C}$ is a diffeomorphism.
\end{thm}
\begin{proof}
If there exist different $(p_1,\xi^1)$ at $t_1$ and $(p_2,\xi^2)$ at $t_2$
such that $\mathcal{C}(p_1, \xi^1) = \mathcal{C}(p_2, \xi^2) = (y,\eta, \hat{y}, \hat{\eta})$,
we must have $\hat{y} b_1(t) - \hat{\eta}a_1(t) =0$ is true for both $ t= t_1, t_2$.
Consider the vector field
\[
c_0(y,\eta,t)
= \hat{y} \frac{\partial \gamma}{\partial \eta} - \hat{\eta} \frac{\partial \gamma}{\partial \eta}
= (\hat{y} b_1(t) - \hat{\eta}a_1(t)) e_1(t) + (\hat{y} b_2(t) - \hat{\eta}a_2(t))e_2(t).
\]
It satisfies
\begin{align*}
&c_0(y,\eta,t_1) = (\hat{y} b_2(t_1) - \hat{\eta}a_2(t_1))e_2(t_1) \equiv \mu_1 \dot{\gamma}(t_1).\\
&c_0(y,\eta,t_2) = (\hat{y} b_2(t_2) - \hat{\eta}a_2(t_2))e_2(t_2) \equiv \mu_2 \dot{\gamma}(t_2).
\end{align*}
It follows that we can define the Jacobi field
\begin{equation*}\label{jacobifield}
c(y,\eta,t)
= \hat{y} \frac{\partial \gamma}{\partial \eta} - \hat{\eta} \frac{\partial \gamma}{\partial \eta} - (\mu_1 \frac{t - t_2}{t_1 -t_2} +\mu_2 \frac{t - t_1}{t_2 -t_1} ) \dot{\gamma}(t),
\end{equation*}
which satisfies $c(y,\eta,t_1)=c(y,\eta,t_2)=0$. This implies $p_2$ is conjugate to $p_1$.
Suppose
\[
\xi^1 = \mu_1 \dot{\gamma}^\perp (t_1), \quad \xi^2 = \mu_2 \dot{\gamma}^\perp (t_2)
\]
By (\ref{conditioncanonical}), they satisfy
\begin{equation*}
\hat{y} = \mu_1 a_1(t_1) = \mu_2 a_1(t_2).
\end{equation*}
This indicates we can find nonzero $\mu$ such that $\mu_1 = \mu a_1(t_2)$ and $\mu_2 = \mu a_1(t_1)$.
Conversely, if $p_2$ is conjugate to $p_1$, then there is a nonzero Jacobi field $J(t)$
so that $J(t_1)=J(t_2)=0$.
More precisely, we have
\[
\kappa_1 \frac{\partial \gamma}{\partial y} + \kappa_2 \frac{\partial \gamma}{\partial \eta} + c_1 \dot{\gamma}(t) +c_2 t\dot{\gamma}(t) = 0,
\quad \text{for } t = t_1, t_2.
\]
The projection on $\dot{\gamma}^\perp(t)$ shows
\begin{equation}\label{singularmatrix}
\begin{cases}
\kappa_1 a_1(t_1) + \kappa_2 b_1(t_1) =0\\
\kappa_1 a_1(t_2) + \kappa_2 b_1(t_2) =0
\end{cases}
\iff
\left[
\begin{array}{cc}
a_1(t_1) & b_1(t_1)\\
a_1(t_2) & b_1(t_2)
\end{array}
\right]
\left[
\begin{array}{cc}
\kappa_1\\
\kappa_2
\end{array}
\right]
= 0.
\end{equation}
Notice $\kappa_1, \kappa_2$ cannot be both zero, otherwise we have the trivial Jocabi field tangent to $\gamma$.
This implies the matrix in (\ref{singularmatrix}) is singular.
Now since $\xi^1 = \mu a_1(t_2) \dot{\gamma}^\perp (t_1)$ and $\xi^2 = \mu a_1(t_1) \dot{\gamma}^\perp (t_2)$ for some nonzero $\mu$,
by (\ref{conditioncanonical}) one have
\[
\hat{y} = \mu a_1(t_2) a_1(t_1) = \mu a_1(t_1)a_1(t_2), \quad
\hat{\eta} = \mu a_1(t_2) b_1(t_1) = \mu a_1(t_1)b_1(t_2),
\]
where the last equality is from the zero determinant of the matrix in (\ref{singularmatrix}).
This proves $\mathcal{C}(p_1, \xi^1) = \mathcal{C}(p_2, \xi^2)$.
\end{proof}
For $(p_1, \xi^1)$ and $(p_2, \xi^2)$ satisfying Theorem \ref{thm_C_diff}, we call them the \textit{conjugate covectors}.
\begin{thm}[{\cite[Proposition 2]{Frigyik2007}}]\label{thm_N}
If there are no conjugate points, the normal operator $N = I_w^* I_w$ is a $\Psi$DO of order $-1$ with principal symbol
\[
\sigma_p(N)(x, \xi) = \frac{2\pi}{|\xi|}(|w(x, \xi_\perp)|^2 +|w(x, -\xi_\perp)|^2 ),
\]
where $\xi_\perp$ is the vector conormal to the covector $\xi$ with the same length and the rotation of $\pi/2$.
\end{thm}
\begin{proof}
This theorem is proved in {\cite[Section 4]{Frigyik2007}}. By the proof of {\cite[Proposition 2]{Frigyik2007}}, the normal operator is a $\Psi$DO with the principal symbol
\[
\sigma_p(N)(x,\xi) = 2\pi\int_{S^1} A(x,0,\omega)\delta(\xi(\omega)) \diff \omega,
\]
where $\delta$ is the Dirac delta function and in local coordinates we have
\[
A(x, 0, \omega) = J^{-1}(x, 0, \omega)|w(x, \omega)|^2.
\]
Here we write
\[
\exp_x(t,v) - x = t m(t,v;x)
\]
and introduce the change of variables $(r, w) \in \mathbb{R} \times S^1$ by
\[
r = t|m(t,v;x)|, \quad \omega = m(t,v;x)/|m(t,v;x)|,
\]
with the Jacobian
\[
J(x, t, v) \equiv \det \frac{\partial (r, \omega)}{\partial (t, v)}.
\]
For more details, see equations (32) - (34) in \cite{Frigyik2007}.
Note that
\[
m(0,v;x) = \dot{\gamma}_{x,v}(0) = v
\] since we use arc length parameterization.
Thus, we have that $J(x, 0, v) = 1$ and the principal symbol of $N$ is given as claimed.
See also \cite[Theorem 5.2]{MR2970707} and its proof.
\end{proof}
\subsection{The local problem}
In this following, we present the cancellation of singularities arising in the local inverse problem for the integral transform $I_w$, if there are conjugate covectors.
This is the analog to the case of geodesic ray transforms in \cite{Monard2015}.
Consider a fixed curve $\gamma \in \Gamma$ with conjugate covectors $(x_1, \xi^1)$ at $t_1$ and $(x_2, \xi^2)$ at $t_2$.
Let $V^j$ be small conic neighborhoods of $(x_j, \xi^j)$, with base $U_j$ as a small neighborhood of $x_j$, for $j=1,2$.
With $\mathcal{C}$ being a local diffeomorphism, it maps a small conic neighborhood of $(x_j, \xi^j)$ to one of $\mathcal{C}(x_j, \xi^j)$, for $j = 1, 2$.
Notice $V^j$ should have two disjoint components, for more details see \cite{Monard2015}.
By shrinking those neighborhoods a bit, one can assume that $\mathcal{C}(V^1) = \mathcal{C}(V^1) \equiv \mathcal{V}$.
We define the restriction $\mathcal{C}_j \equiv \mathcal{C}|_{V^j}$ for $j=1,2$.
Note that $\mathcal{C}_j$ are diffeomorphisms.
It follows that
\[
\mathcal{C}_{12} \equiv \mathcal{C}_1^{-1} \mathcal{C}_2: V^2 \rightarrow V^1, \quad \quad \mathcal{C}_{21} \equiv \mathcal{C}_2^{-1} \mathcal{C}_1: V^1 \rightarrow V^2
\]
are also diffeomorphisms.
Let $I_j$ be $I_w$ restricted to distributions with wave front sets supported in $V^j$, for $j =1,2$.
Then $I_j$ are FIOs with canonical relations ${C}_j$, where $C_j$ are restriction of the canonical relation $C$ to $V^j \times \mathcal{V}$.
When the weights are nonvanishing, the restriction $I_j$ are elliptic FIOs, and therefore we can define
\begin{align}\label{def_F12}
F_{21} = I_2^{-1}I_1, \quad F_{12} = I_1^{-1}I_2.
\end{align}
Note that they are FIOs with canonical relations
$C_2^{-1}\circ C_1$ and $C_1^{-1}\circ C_2$ respectively.
We can show the following result by the same arguments in \cite{Monard2015}.
\begin{thm}\label{thm_microkernel}
Suppose $w(x_1, \xi^1_\perp), w(x_2, \xi^2_\perp) \neq 0$.
Let $f_j \in \mathcal{E}'(U_j)$ with $\text{\( \WF \)}(f_j) \subset V^j$, for $j = 1,2$.
Then the local data
\[
{{I_w}} (f_1+f_2) \in H^s(\mathcal{V})
\]
if and only if
\[
f_1 + F_{12} f_2 \in H^{s-1/2}(V^1) \Leftrightarrow
F_{21} f_1 + f_2 \in H^{s-1/2}(V^2),
\]
where $F_{12}$ and $F_{21}$ are elliptic FIOs defined in (\ref{def_F12}).
\end{thm}
This theorem indicates that
given a distribution $f_1$ singular in $ V^1$, there exists a distribution $f_2 = -F_{21}f_1$ singular in $V^2$ such that the transform is smooth.
In other words, if we suppose $f=f_1$, the singularities of $f$ cannot be resolved from the singularities of the transform $I_w f$.
Indeed, the singularities of $f$ can only be recovered up to an error in the microlocal kernel, i.e., an error in form of $(\text{Id} - F_{21})h_1$ with some $h_1$ singular in $V^1$,
since $I_w ((\text{Id} - F_{21})h_1)$ is always smooth.
For a more detailed description, see \cite{Holman2017}.
\section{Numerical experiments}\label{sec_num}
This section aims to illustrate the artifacts arising
in the reconstruction by numerical experiments.
We consider the family of unit circles with a fixed orientation,
i.e., the magnetic geodesics w.r.t. the Euclidean metric and a constant nonzero magnetic field.
See Example \ref{ex_c} in Section \ref{sec_example} and \cite{MR2970707} for more details.
More explicitly, let $M$ be a bounded domain without boundary in $\mathbb{R}^2$, for example,
the open disk of radius $R=3$ centered at the origin.
Suppose $f$ is a smooth function supported in $M$. We define
$$
I f(x_1,x_2) = \int_{|w|=1} f((x_1,x_2) + w) d l_w = \int_0^{2\pi} f(x_1+\cos \alpha, x_2 + \sin \alpha) d\alpha
$$
as the integral transform performed over unit circles with radius $1$.
These circles can be parameterized by their centers.
Then the adjoint operator is
$$
I^* g(y_1,y_2) = \int_{|w|=1} g((y_1+y_2) - w) d l_w = \int_0^{2\pi} g(y_1-\cos \alpha, y_2 - \sin \alpha) d\alpha
$$
which coincides with $I$ itself.
Numerically we compute $I f$ by the following steps.
\begin{itemize}
\item[(1)] Discretization. We introduce $6 N \times 6N$ equispaced points in the square domain $[-3,3]\times[-3,3]$ so the grid spacing is ${1}/{N}$.
We discretize the input function $f$ over the grids.
\item[(2)] We compute the integral by the Trapezoidal method
$$
I f(x_1,x_2) = \sum_{k=1}^{n} \frac{f(z_{k-1},w_{k-1}) + f(z_{k},w_{k})}{2} \Delta \alpha + O(\frac{1}{n^2}),
$$
where $z_{k} = x_1 + \cos{{2 \pi k}/{n}}, w_{k} = x_2 + \sin{{2 \pi k}/{n}}$ and $\Delta \alpha = \frac{2 \pi}{n}$.
When the input function is only given over grids, we can use the bilinear interpolation to approximate its values at $\{(z_k,w_k), k = 1, \ldots, n\}$ in Step (2) before the numerical integration.
\end{itemize}
Then we use the same method to compute $I^*$.
\subsection{Backprojection}
First, we consider the reconstruction of $f$ from the transform by the backprojection $I^*If$.
To check the numerical implementation of the backprojection, one can compare the numerical result with the analytical one given by the formula
\[
I^*If(x_1,x_2) = \int_{0}^{2} \int_{0}^{2\pi} \frac{4}{4 - r^2} f(x_1+r \cos \theta, x_2+ r\sin \theta) d\theta \diff r.
\]
in \cite{MR2970707}.
Let $f_1$
be a truncated Gaussian concentrated near the origin, as an approximation of a delta function at the origin, and let $f_2 = 0$.
We choose $N = 40$ and the relative error between the numerical result $R_N$ and the analytical one
\[
\frac{\|{R_N - I^*If }\|_2}{\|I^*If\|_2} = 0.0140
\]
is relatively small.
In this case, $f$ is discretized as a matrix of $240\times240$ and we integrate it over unit circles.
From Figure \ref{fig_1}, we can see
the artifacts appear exactly in the location of conjugate points.
Indeed, by the analysis in Example \ref{ex_c}, the conjugate locus of the origin is the circle centered at the origin of radius $2$.
With $f$ as an approximation of the delta function, the singularities of $f$ are located near the origin in all directions.
The conjugate covectors of the singularities of the delta function are described by the equation (\ref{eq_cc}), which are covectors conormal to the circle of radius $2$.
Then by (\ref{eq_fb}),
from the backprojection we recover both the singularities of $f = f_1$ and the singularities of $ F_{12} f_1$.
Note here we can omit $\Lambda$ in (\ref{eq_fb}) if we only consider the singularities, since $\Lambda$ is an elliptic $\Psi$DO.
\begin{figure}[h]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth]{figcir/fun1_orig.png}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth]{figcir/fun1_recon.png}
\end{subfigure}
\caption{The true $f$ and the reconstruction from the backprojection.}\label{fig_1}
\end{figure}
\subsection{Landweber iteration}
In the following, we choose $f = f_1$ to be a modified Gaussian with singularities located both in
certain space and in certain direction, that is, a coherent state, as is shown in Figure \ref{fig_2} (a).
Note the singularities of $f$ are actually semi-classical and they are located near the origin.
The same analysis works (see \cite{Stefanov2020} for FIOs and semi-classical wave front sets) and here we use this coherent state to illustrate the artifacts in the analysis before.
If we use the backprojection to reconstruct $f$, the artifacts appear in the location of conjugate points,
see Figure \ref{fig_2} (b).
This is described by (\ref{eq_cc}) and (\ref{eq_fb}).
Next, we use the Landweber iteration to reconstruct $f$.
The analysis in Section \ref{subsec_artifacts} explains the artifacts in the reconstruction from the local data.
However, if we use the global data and have the prior knowledge that $f$ is supported in a compact set, then we can recover the singularities of $f$ without artifacts.
The following proposition is an analog to \cite[Corrollary 3]{Zhang2020}.
\begin{pp}\label{pp_m}
Suppose $f \in \mathcal{E}'(\mathbb{R}^2)$ and $If \in C^\infty(\mathcal{M})$. Then $f$ is smooth.
\end{pp}
\begin{proof}
Let $(x_0,\xi^0) \in T^*M$ and assume it is in the wave front set of $f$.
This singularity can be canceled by its conjugate covectors, if they exist.
As in \cite[Section 5]{Zhang2020}, we define
\begin{align}\label{eq_M}
\mathcal{M}(x_0,\xi^0) = \{ (x_k,\xi^k),
&\text{ if it exists and is conjugate to $(x_{k'},\xi^{k'})$}, \\
& \text{where } k'= k-\operatorname{sgn} k, \text{for }k = \pm 1, \pm 2, \ldots \} \nonumber
\end{align}
as the set of all conjugate covectors related to $(x_0,\xi^0)$.
We can assume $\xi^0 = (\cos \eta_0, \sin \eta_0)$ is in a conic neighborhood of the covector $\xi = (1,0)$.
By (\ref{eq_cc}), we have
\begin{align*}
\mathcal{M}(x_0,\xi^0) = \{ (x_k,\xi^0): x_k = x_0 + 2k(\cos \eta_0, \sin \eta_0), \text{ for } k = \pm 1, \pm 2, \ldots
\}.
\end{align*}
Let $V^k$ be a small conic neighborhoods of $(x_k ,\xi^k) \in \mathcal{M}(x_0,\xi^0)$.
Let $f_k$ be $f$ microlocally restricted to $V^k$ and $I_k$ be $I$ restricted to distributions singular in $V^k$.
For each $k$, near the curve $\gamma_k$ where $(x_{k-1} ,\xi^{k-1})$ and $(x_k ,\xi^k)$ are conjugate covectors,
we have the equation of cancellation of singularities
\begin{equation}\label{eq_cs}
I_{k-1}f_{k-1} + I_k f_k = 0 \mod C^\infty,
\end{equation}
if we shrink $V^k$ such that $C(V^k)= V^{k-1}$.
Since $f$ has compact support, there exist $k_+>0$ and $k_-<0$ such that $f$ is smooth near all $(x_k, \xi^k)$ with $k < k_+$ or $k< k_-$.
Then we have
\[
I_k f_k = 0 \mod C^\infty, \text{ for any } k < k_+ \text{ or } k< k_-.
\]
By (\ref{eq_cs}), it follows that $I_k f_k = 0 \mod C^\infty$ for all $k$.
Thus, we have $f$ is smooth.
\end{proof}
\begin{figure}[h]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{figcir/gl_orig20}
\subcaption{true $f$}
\end{subfigure}%
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{figcir/gl_recon2001}
\subcaption{backprojection $f^{(1)}$}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{figcir/gl_recon20}
\subcaption{$f^{(100)}$}
\end{subfigure}
\caption{Reconstruction for a modified Gaussian function $f$, where $e = \frac{\|f - f^{(100)}\|_2}{\|f\|_2}$ is the relative error.}\label{fig_2}
\end{figure}
In the numerical experiment, suppose $f$ is compactly supported in the disk $D$ that is centered at the origin with radius $3$.
When we use Landweber iteration,
after performing the backprojection operator in each step, we smoothly cut the function such that it is still supported in radius $3$.
More explicitly, we consider the operator
$
\mathcal{L}_\varphi = \varphi \mathcal{L} =\varphi \Lambda \chi I^*I,
$
where $\varphi$ is a smooth cutoff function with $\varphi = 1$ in $D$ and supported in a slightly larger disk $D'$.
With $f$ compacted supported in $D$, we have
$
\mathcal{L}_\varphi: L^2(D') \rightarrow L^2(D').
$
In this case, to solve the equation ${{I}} f = g$, we write
\begin{equation*
(\text{Id} - (\text{Id}- \gamma \mathcal{L}_\varphi^* \mathcal{L}_\varphi)) f = \gamma \mathcal{L}_\varphi^*(\varphi\Lambda\chi I^* g)
\end{equation*}
and use the truncated Neumann series
\[
f^{(N)} = \sum_{k=0}^{N}(\text{Id}- \gamma \mathcal{L}^*_\varphi \mathcal{L}_\varphi)^k
\gamma \mathcal{L}^*_\varphi (\varphi\Lambda\chi I^* g).
\]
This series converges to the minimal norm solution to $\mathcal{L}_\varphi f = \varphi\Lambda\chi I^* g$ in $L^2(D')$.
After 100 steps of iteration, we get a quite good reconstruction (with the relative error $e = 0.055$).
This illustrates the result of Proposition \ref{pp_m}.
|
\section{Introduction}
Several crucial processing steps for virtual communication have immensely profited from deep learning, e.g., speech enhancement (SE) \cite{perceptnet_dns, dtln_dns} and acoustic echo cancellation (AEC) \cite{perceptnet_aec, dtln_aec}.
The underlying models can be created with deep-learning tools alone or based on a combination with ''classic`` signal-processing approaches such as adaptive filters and with deep-learning models \cite{perceptnet_aec, aec_braunschweig}.
Another important challenge is to improve communication systems with effective packet loss concealment (PLC) using deep learning.
The task of PLC is to conceal or predict lost packets during packet-based transmissions. A well-performing PLC system is especially important for poor network connections, since more packets are lost or overly delayed and therefore discarded. A robust PLC system could also decrease listening fatigue and effort in situations with poor audio quality and therefore result in overall improved communication channels. PLC systems should work efficiently so they can be used for many device classes without fully utilizing the processing power and sacrificing battery run-time on mobile devices.
Packet loss concealment is a sequence problem since missing frames from an audio sequence must be predicted from past information.
An early deep approach \cite{dnn_packet_loss} utilized feed-forward neural networks for predicting the magnitude and phase from a number of past frames using mean-squared-error based loss in time-frequency domain.
Another approach is a causal sequence to sequence model such as RNNs as used for SE \cite{nsnet2, dtln_dns, cruse}. In related research, RNN-based models have been applied to PLC \cite{mohamed2020concealnet, plc_lstm, facebook_plc}. In \cite{plc_lstm} a recurrent model with online adaptation is proposed by exploiting the recent past of the signal. \cite{mohamed2020concealnet} used an LSTM-based PLC model to increase the performance of speech emotion recognition trained with a concordance correlation coefficient as cost function. A convolutional recurrent structure for time-domain PLC was introduced in \cite{facebook_plc} trained with a mean absolute error cost in time domain.
Alternative deep approaches utilizing generative adversarial networks (GAN) \cite{plc_gan_1, inpainting_gan, plc_gan_2}, which were shown in the past to be able to predict speech frames.
A comparison between these studies is difficult because of the lack of a common training and test set, and therefore Microsoft proposed a deep PLC Challenge for Interspeech 2022 \cite{diener2022interspeech}.
The authors provided a large amount of speech data, traces of lost frames from real calls and a
deep model for predicting mean opinion scores (MOS), indicating PLC performance of enhanced audio. The challenge evaluations will be performed as in previous challenges organized by Microsoft \cite{dns_challenge, AEC_challenge_ver1, dns_2021_interspeech}
with a crowd-sourcing approach on the Amazon Mechanical Turk.
For the first time, to the best of our knowledge, we propose in the context of the PLC Challenge a time-domain sequence-to-one (seq2one) RNN model for PLC trained with a combined magnitude and complex mean absolute error (MAE) loss in time-frequency domain. In contrast to previous RNN approaches \cite{mohamed2020concealnet, plc_lstm, facebook_plc}, the model predicts only lost frames from a short context buffer similar as in \cite{dnn_packet_loss} in a non-continuous way instead of running inference for each incoming audio frame to update internal states. This seq2one approach can decrease, depending on the complexity of the model, the computational load of the host system
and probably increase the battery run-time of a mobile device.
\section{Methods}
\subsection{Deep PLC Strategy}
The proposed PLC seq2one system is based on a context buffer of length $n$ that contains non-overlapping frames of length $m$.
For $m$ 10\,ms (160 samples at 16\,kHz sampling rate) are chosen. It is therefore possible to use one look-ahead frame and still comply with the PLC-Challenge rules, which specify a maximum delay of 20\,ms \cite{diener2022interspeech}. The use of look-ahead frames has been shown to be beneficial in \cite{facebook_plc} to connect predicted frames better to available audio. When no packet loss is detected, the current and the look-ahead frames are concatenated to a frame with 320 samples and subsequently hann windowed.
When a packet loss is detected, the dPLC model predicts a 320 sample long frame that is windowed as well.
With an overlap-add procedure the current 10\,ms output frame is predicted which is time-aligned to the current input frame.
The current output frame is written back to the context buffer.
This procedure, especially the windowing and overlap-add, was chosen to reduce artifacts on the edges of lost packets and connect the predicted content better to the available frames and so pursues a similar goal as the cross-fading shown in \cite{dnn_packet_loss}. The proposed PLC system is also illustrated in \autoref{fig:plc-system}.
\begin{figure}[t]
\centering
\includegraphics[width=167pt]{PLC_system.pdf}
\caption{Illustration of the proposed PLC system. Incoming audio is written to a context buffer. When no frame is lost, the current frame and the look-ahead frame are concatenated and hann-windowed. If the the frame on position $x$ or $x+1$ is lost, the dPLC predicts an audio frame which is also hann-windowed. The resulting windowed frames are processed by an overlap-add procedure returning the current frame $y$ time aligned to input frame $x$.}
\label{fig:plc-system}
\end{figure}
\subsection{Model Architectures}
The model is build as seq2one model predicting one time domain frame from a context buffer. The general architecture is inspired by the NSnet2 architecture \cite{nsnet2} and an architecture proposed in \cite{facebook_plc}. It was adapted to seq2one task and working completely in the time domain.
First, the frames of the context buffer are processed by a fully-connected (FC) layer with relu activation that maps the time-domain frames to a non negative feature space, as has been proposed with the TasNet approach \cite{tasnet_orig,conv_tasnet, dua-path}.
Second, all frames are processed by another FC layer with leaky-relu activation. The resulting intermediate representation sequence is processed by two consecutive Conv1D layers with leaky relu-activation and kernel size 4 and 2, to extract local patterns from the sequence. The sequence is padded so that the resulting sequence has still the same length.
The sequence is subsequently processed by two bidirectional Gated Recurrent Unit (GRU) \cite{chung2014empirical} layers, which enables the use of information about the whole sequence. The final hidden state of the second BGRU is forwarded to two FC mapping layers with leaky-relu activation. A last FC layer without activation predicts the time-domain frame.
To construct a feed-forward baseline (FF-Baseline) the Conv1Ds and BGRUs layers are replaced by a flatten operation to remove the sequence dimension and 3 FC layers with 512 units and leaky-relu activation.
In total, three different size configurations are tested for the proposed tPLCnet to explore the scaling properties of the model (\emph{Small, Medium and Large}). The number of units or filters are listed in \autoref{tab:units}. Additionally, \autoref{tab:results} shows the number of Multiply Accumulates (MACs) and execution time for predicting one 10\,ms frame on an Intel i5-7267U dual-core processor clocked at 3.1\,GHz.
\begin{table}[th]
\caption{Number of units/filters of each layer for the configurations \emph{Small, Medium and Large} of the proposed deep PLC model.}
\label{tab:units}
\centering
\begin{tabular}{ l c c c}
\toprule
\textbf{Layer} & \textbf{Small} & \textbf{Medium} & \textbf{Large} \\
\midrule
FC encoding & \multicolumn{3}{c}{512} \\
\midrule
FC embedding & 128 & 256 & 512 \\
\midrule
Conv1x4 & 128 & 256 & 512 \\
\midrule
Conv1x2 & 128 & 256 & 512 \\
\midrule
BGRU 1 & 64 & 128 & 256 \\
\midrule
BGRU 2 & 64 & 128 & 256 \\
\midrule
FC mapping 1 & \multicolumn{3}{c}{512} \\
\midrule
FC mapping 2 & \multicolumn{3}{c}{512} \\
\midrule
FC decoding & \multicolumn{3}{c}{320} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Loss Functions}
Recent studies have shown that the loss function is crucial to perceived audio quality in speech enhancement \cite{loss_functions_1}, but at the same time, relatively simple loss functions are surpassing more complex ones \cite{loss_functions_2}.
To quantify the effect loss functions for PLC, our study compared three loss functions, which are all applied to the whole enhanced utterance.
\cite{facebook_plc} proposed to use a simple mean absolute error (MAE) in time domain which shown promising results.
\begin{equation}
\mathrm{MAE_{time}} = |\hat{x} - x|
\end{equation}
where the loss calculated as the mean absolute difference of the clean audio $x$ and the predicted audio $\hat{x}$.
Due to its definition in time domain, MAE is phase sensitive and is - in comparison to the mean squared error - less sensitive to outliers.
In \cite{loss_functions_1}, it was shown that the MAE in time-frequency domain (TF-domain) produces competitive results for speech enhancement.
MAE has also been explored for speech synthesis \cite{wang17n_interspeech}, which in some aspects is related to PLC.
To calculate the MAE in TF-domain, the clean and predicted utterances are subject to an STFT which results in the complex TF representations $X(t,f)$ and $\hat{X}(t,f)$.
The magnitude MAE loss is given by
\begin{equation}
\mathrm{MAE_{mag}} = ||\hat{X}| - |X||.
\end{equation}
Results of related studies indicate that integrating a phase component in the loss function can be beneficial \cite{psa_1, loss_functions_1}, which can be implemented by using a complex distance. The combined TF-domain MAE loss is given by
\begin{equation}
\mathrm{MAE_{comb}} = (1-\alpha) ||\hat{X}| - |X|| + \alpha |\hat{X} - X|,
\end{equation}
where $\alpha$ is a weighting factor for the amount of phase sensitivity. In this study, $\alpha = 0.1$ is chosen, which was found to be a good trade-off between some phase sensitivity and a good magnitude reconstruction in preliminary experiments.
\begin{table*}[!th]
\caption{Results on the known test-set in terms of PESQ [MOS], PLC-MOS [MOS] and DNS-MOS P.835. For PESQ and PLC-MOS the data is also shown for the three loss subsets low (up to 120\,ms), med. (120 to 320\,ms) and high (320 to 1000\,ms). For DNS-MOS all three metrics are shown. Additionally the number Multiply Accumulates (MACs) and the execution time [ms] for one 10\,ms frame are shown.}
\centering
\begin{tabular}{@{}l l r r r r r r r r r r r r r r@{}}
\toprule
& & \multicolumn{1}{c}{\textbf{exec.}} & \multicolumn{4}{c}{\textbf{PESQ}} & & \multicolumn{4}{c}{\textbf{PLC-MOS}} & & \multicolumn{3}{c}{\textbf{DNS-MOS}}
\\
\midrule
\textbf{Method} & \textbf{\# MACs} & \textbf{Time} & low & med. & high & ovr. &
& low & med. & high & ovr. & & \textbf{nSIG} & \textbf{nBAK} & \textbf{nOVR}
\\
\midrule
Baseline & & & 2.59 & 1.75 & 1.73 & 2.19 & & 3.15 & 2.46 & 2.77 & 2.87 & & 3.47 & 3.61 & 2.97 \\
\midrule
FF baseline & 2.49\,M & 0.47 & 3.24 & 2.13 & 1.91 & 2.68 & & 4.05 & 3.52 & 3.63 & 3.81 & & 4.15 & 4.23 & 3.75 \\
$\mathrm{tPLCnet_{S}}$ & 2.96\,M & 0.57 & 3.28 & 2.14 & 1.97 & 2.71 & & 4.09 & 3.57 & 3.72 & 3.86 & & 4.14 & 4.23 & 3.74 \\
$\mathrm{tPLCnet_{M}}$ & 7.85\,M & 1.41 & 3.31 & 2.15 & 1.98 & 2.74 & & 4.12 & 3.61 & 3.79 & 3.90 & & 4.17 & 4.25 & 3.78 \\
$\mathrm{tPLCnet_{L}}$ & 26.18\,M & 3.95 & \textbf{3.36} & \textbf{2.18} & \textbf{1.99} & \textbf{2.77} & & \textbf{4.16} & \textbf{3.66} & \textbf{3.82} & \textbf{3.95} & & \textbf{4.18} & \textbf{4.27} & \textbf{3.79} \\
\bottomrule
\end{tabular}
\label{tab:results}
\end{table*}
\subsection{Datasets and Augmentation}
The challenge organizers are providing a training set, a known test set and a blind test set. The training set contains around 64\,h of clean speech and packet loss traces from real calls.
The traces are vectors of ones and zeros indicating for each 20\,ms frame/packet of an audio file if it is lost (one) or not (zero).
The speech is sampled from public available podcast audio. Each utterance is around 10\,s long.
During training, the speech files and PLC traces are cut to 8\,s with a random start and end. The order of speech files and traces are shuffled for each epoch. The traces are time-reversed with 50\,\% probability to further increase the variance and increase robustness. To create the degraded audio the lost segments are set to zero according to the corresponding trace. The level of the speech is sampled from a normal distribution with a mean of -26\,dB and a standard deviation 10\,dB.
The known test set contains around 2.7\,h of clean speech, degraded speech and PLC traces. As in the training set, each file is around 10\,s long. The data is divided in three subsets by the organizers depending on the maximum burst loss length of the traces. The categories are: low loss up to 120\,ms (52\,\% of the data), medium loss from 120 to 320\,ms (32\,\% of the data) and high loss from 320 to 1000\,ms (16\,\% of the data).
The blind test set has the same properties, but does not contain the clean speech files. Further details on the data can be found in the challenge description \cite{diener2022interspeech}.
\subsection{Training Setup}
The models are trained over 200 epochs with a batch size of 16 and a learning rate of 5e-4. The learning rate is multiplied by 0.8 when the loss on the test set does not decrease for three consecutive epochs. The Adam optimizer is used in combination with a gradient norm clipping of 3.
The context buffer size was set to a value of 6 as a trade-off between complexity and performance.
During training, the four most recent frames of the context buffer are sampled from the degraded signal, as well as the last two frames from the clean signal. We found that this strategy improves performance compared to training on the degraded audio only. The training for the medium-size model takes around 6\,h on an Nvidia RTX A5000.
\subsection{Objective and Challenge Evaluation}
The first objective metric used is the wide band perceptual evaluation of speech quality (PESQ) \cite{pesq}, which was originally built to evaluate the quality of transmitted speech. The second one is DNS-MOS P.835 \cite{reddy2022dnsmos} which is a single-ended/blind measure based on a deep-learning approach for estimating Mean Opinions Scores (MOS) of speech quality (nSIG), background noise quality (nBAK) and an overall quality (nOVR). DNS-MOS was trained with ratings from previous deep-noise suppression challenges organized by Microsoft \cite{dns_2021_interspeech}. The last measure PLC-MOS \cite{diener2022interspeech} is provided by the challenge organizers. PLC-MOS utilizes a similar approach as DNS-MOS but can be used as double-ended or single-ended measure. It predicts MOS scores indicating the quality of the PLC system.
The challenge organizers evaluate the improved audio created from blind test with four metrics: Double-ended PLC-MOS, second, the DNS-MOS P.808 \cite{reddy2021dnsmos}, which returns an overall quality rating. Third, a crowd-sourced Comparison Category Rating (CCR MOS) is used, for which the participants are asked to rate the enhanced audio compared to the clean reference. This evaluation is run on the Amazon Mechanical Turk. Fourth, the word accuracy calculated by the Microsoft Azure cognitive services speech recognition is used. The final score is calculated from CCR MOS and word accuracy.
\section{Experiments and Results}
In the following the experiments and results are described: \\
\textbf{Objective results on complexity comparison}:
\autoref{tab:results} shows the objective results for the zero filling baseline, the feed-forward baseline trained on the same setup, as well as three complexity configurations of the tPLCnet (cf. \autoref{tab:units}).
The FF baseline already reaches a good performance increase for all metrics.
Performance increases with model size for all metrics.
The total improvement in terms of PLC-MOS for $\mathrm{tPLCnet_{Large}}$ is 1.08. The PESQ improvement over the zero filling baseline is 0.58 and for DNS-MOS nOVR 0.82. The PESQ and PLC-MOS improvements are constant over all subsets.
\\
\textbf{Objective results on loss functions and loss function configurations }: We compared $\mathrm{MAE_time}$, $\mathrm{MAE_mag}$ and $\mathrm{MAE_comb}$; for the latter, three different window sizes with 50\,\% overlap were tested to identify an optimal configuration for the PLC task. All results were acquired with $\mathrm{tPLCnet_{Medium}}$. The results are shown in \autoref{tab:loss_results}. The lowest performance in all metrics is achieved with the time-domain $\mathrm{MAE_time}$ loss. The highest PESQ (2.76) and DNS-MOS (3.85) values are reached with $\mathrm{MAE_comb}$ loss with 20\,ms window size. The highest PLC loss (3.90) is achieved with $\mathrm{MAE_comb}$ and 32\,ms window size.
\begin{table}[th]
\caption{Objective scores for different loss functions. For DNS-MOS, only nOVR is shown.}
\label{tab:loss_results}
\centering
\begin{tabular}{l r r r }
\toprule
& & \textbf{PLC-} & \textbf{DNS-} \\
\textbf{loss} & \textbf{PESQ} & \textbf{MOS} & \textbf{MOS}\\
\midrule
$\mathrm{MAE_{time}}$ & 2.43 & 3.30 & 3.54 \\
$\mathrm{MAE_{mag}}$ (32 ms, 50\%) & 2.75 & 3.88 & 3.79\\
$\mathrm{MAE_{comb}}$ (64 ms, 50\%) & 2.61 & 3.53 & 3.66 \\
$\mathrm{MAE_{comb}}$ (32 ms, 50\%) & 2.73 & \textbf{3.90} & 3.78 \\
$\mathrm{MAE_{comb}}$ (20 ms, 50\%) & \textbf{2.76} & 3.71 & \textbf{3.85} \\
\bottomrule
\end{tabular}
\end{table}
\\
\textbf{Comparison of context buffer lengths}:
The length of the context buffer is a major factor for complexity in terms of operations for our seq2one model, since the whole buffer must be processed each time by the convolutional and the BGRU layers.
Hence, we explore the effect of different buffer lengths on performance. \autoref{fig:buffer_len} shows the performance in terms of PLC-MOS and DNS-MOS against the buffer length.
The optimal lengths for PLC-MOS and DNS-MOS are 7 and 6, respectively, while the lowest performance is reached with 20 frames (PLC-MOS) or 15 frames (DNS-MOS).
\begin{figure}[th]
\centering
\includegraphics[trim={4cm 12.3cm 4cm 12cm}, clip, width=1.0\linewidth]{buffer_len_emb.pdf}
\caption{Performance in terms of PLC-MOS and OVR DNS-MOS P.835 versus the number of 10\,ms frames in the context buffer length.}
\label{fig:buffer_len}
\end{figure}
\\
\textbf{Results of the challenge evaluations}:
The challenge results for the zero-filling baseline and all approaches submitted to the challenge are shown in \autoref{tab:ch_results}.
$\mathrm{tPLCnet_{Large}}$ with $\mathrm{MAE_{comb}}$ using a loss window length of 32\,ms ranked 3rd in the challenge. It improves on all metrics except for word accuracy. The 1st and 2nd ranked entries improve on all metrics (including word accuracy). Our tPLCnet achieves the second-best score in terms of PLC-MOS.
\begin{table}[th]
\caption{2022 INTERSPEECH Audio Deep PLC Challenge Results. Teams marked as "tied": Difference not significant ($p \geq 0.05$).}
\label{tab:ch_results}
\centering
\begin{tabular}{@{}l r r r r r@{}}
\toprule
& \textbf{PLC-} & \textbf{DNS-} & \textbf{CCR-} & \textbf{Word} & \textbf{Final}\\
\textbf{Rank} & \textbf{MOS} & \textbf{MOS} & \textbf{MOS} & \textbf{acc.} & \textbf{score}\\
\midrule
7 & 2.896 & 3.474 & -1.309 & 0.860 & 0.712 \\
Baseline & 2.904 & 3.444 & -1.231 & 0.861 & 0.725 \\
6 & 3.478 & 3.738 & -1.042 & 0.825 & 0.739 \\
5 & 3.277 & 3.513 & -1.102 & 0.863 & 0.748 \\
\textbf{3} (tied) \textbf{(ours)} & \textbf{3.976} & \textbf{3.688} & \textbf{-0.838} & \textbf{0.859} & \textbf{0.790} \\
3 (tied) & 3.829 & 3.684 & -0.812 & 0.868 & 0.798 \\
2nd & 3.744 & 3.788 & -0.638 & 0.882 & 0.835 \\
1st & 4.282 & 3.797 & -0.552 & 0.875 & 0.845 \\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion}
In the following some aspects of the results are discussed. While an increase in complexity of tPLCnet also results in a performance increase, the gain seems to saturate compared to the negative aspects of increased complexity.
The FF baseline already produces a robust PLC performance, indicating that a large portion of the performance comes from our general approach in combination with the training setup.
This implies that for real on-device applications the PLC model can be chosen reasonable depending on the available resources and it must not be the largest possible model.
In \cite{cruse} a similar saturation was found for NSNET2 architecture for SE, i.e., a larger amount of RNN units heavily increases the complexity, but only slightly increases the noise reduction performance.
Regarding the comparison of loss functions, the spectral MAE emerged as the best choice, with the time-domain MAE achieving clearly lower scores.
The magnitude-only loss reaches comparable performance to the combined version which is in line with observations in \cite{loss_functions_1}.
Interestingly, the highest PLC-MOS is observed for a window size of 32\,ms, while the best choice for other metrics was a value of 20\,ms.
Also,
the higher spectral resolution achieved with the 64\,ms window seems not to be helpful for PLC.
The evaluation of the length of the context buffer shows that a short context buffer with only 4 frames already reaches a good PLC performance while very long context buffers even seem detrimental.
A similar effect for longer input buffers was observed in \cite{dnn_packet_loss}. This suggests that the main information required for robust PLC is temporally localized around the lost frame and long-term dependencies only weakly influence overall performance. However, the observed performance differences between the context buffer lengths are not very large, which also means a short buffer length can be chosen without sacrificing performance.
One observation from the challenge result is that a high PLC-MOS does not necessarily result in a high CCR-MOS. Our model reached a higher PLC-MOS than the 2nd place, but the the 2nd ranked model reached a significantly better result in the crowd-sourced CCR-MOS. We assume this can be caused by addtional speech enhancement used by the 2nd-ranked model, which modifies the spectrum such that PLC-MOS is decreased but perceived audio quality is increased.
Word accuracies for the baseline and the tPLCnet are very close to each other (0.861 and 0.859, respectively), i.e., the PLC improvement perceived by the raters does not help the speech recognition system. We assume this is due to a slight deterioration of clean speech in the vicinity of the edges of lost frames, since our algorithm smooths the transition between these areas.
It also seems that the speech recognition system used by challenge organizers can already handle missing packets relatively well.
\section{Conclusions}
This study introduced tPLCnet, a straightforward real-time approach for predicting lost packets from a short context buffer in the time-domain. It achieves state-of-the-art performance and ranks 3rd in the deep PLC Challenge. The training was performed with open-source data which resulted in a robust and reproducible approach which can be easily applied in the real-world.
\section{Acknowledgements}
This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy – EXC 2177/1 - Project ID 390895286 and by - Project-ID 352015383 - SFB 1330.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
With the introduction of hardware with dedicated triangle rasterization units, hand-crafting rasterization routines in software became largely obsolete. Such custom-built rasterizers have nevertheless remained an ongoing topic of research in order to develop and study new rasterization approaches. Some of them eventually managed to beat hardware rasterization in specific scenarios~\cite{10.1145/1730804.1730817}, but in general, dedicated hardware remains the fastest approach. Nanite is the first framework that promises practical improvements for 3D games via hybrid software and hardware rasterization~\cite{karis2021nanite}. They found that directly rasterizing the fragments of a small triangle with atomic min-max operations can be faster than pushing the triangle through the hardware rendering pipeline. Therefore, only larger triangles are rasterized via hardware.
Point-cloud models offer additional opportunities for efficient software rasterization, as the hardware rendering pipeline is largely dedicated to the rasterization of triangles and not points. In this paper, we consider point clouds as 3D models made of colored vertices, where each vertex is projected to exactly one pixel. Although this is a fairly strict limitation, it allows us to device algorithms that compete with graphics APIs that also only support one-pixel points, such as DirectX (POINTLIST primitive) and all backends that use it (WebGL, WebGPU, ANGLE, MS Windows games and applications, ...). We intent to support larger point-sprites in the future and use the evaluated performances of one-pixel points as a baseline for comparisons. Point clouds have no connectivity, so index buffers or vertex duplication are not required. The lack of a connected surface also makes uv-maps and textures irrelevant, which is why colors are typically directly stored on a per-vertex basis. Furthermore, point clouds acquired by laser scanners do not contain surface normals. Normals could be computed in a pre-processing step, but computed normals are not robust in sparsely sampled regions with high-frequency details such as vegetation, strings/cables/wires or even noise. We will therefore not consider normals in this paper, either.
Our approach builds on~\cite{SCHUETZ-2021-PCC} to further optimize several aspects of software rasterization of points, which leads to an up to 3x higher brute-force performance. Specifically, our contributions to the state of the art of software rasterization of point clouds are:
\begin{itemize}
\item Assigning larger workloads to batches to enable efficient batch-level optimizations.
\item Adaptive coordinate precision coupled with visibility-buffer rendering for three times higher brute-force performance.
\item Fine-grained frustum culling on batches of about 10'240 points, directly on the GPU.
\item Support for state-of-the-art level-of-detail structures for point clouds.
\end{itemize}
\section{Discussion}
In this section we'd like to discuss several issues and limitations, as well as potential improvements that we have not evaluated, yet.
First, we believe that this is a significant improvement for point cloud rendering, but it's not useful for games in its current state. Point clouds require a large amount of colored vertices to represent geometry without holes and sufficient color-detail, while meshes can use a single textured triangle in place of thousands of points. However, massive amounts of point cloud data sets exist as a result of scanning the real world, and this paper provides tools to render them faster without the need to process the data into LOD structures or meshes. Still, we hope that the presented approach might provide useful insights in future developments of hybrid hardware+software rasterization approaches for triangle meshes. For example, we could imagine that an approach like adaptive coordinate precision could be used to treat smaller batches like a point cloud and only load data that is relevant for triangles for larger batches.
The quality for virtual reality currently suffers from lack of proper color filtering. Although high-quality shading is applied and improves the results via blending, the issue is that the LOD structure removes most of the overlapping points, thus the blended result is not representative of all points, including the discarded ones. The results can be improved by applying color filtering (e.g. computing averages) to points in lower levels of detail during the construction of the LOD structure~\cite{QSplat,Wand2008,schuetz-2019-CLOD} (similar to mip mapping). Furthermore, implementing continuous LOD could improve the visual quality through a subtle transition in point density between LODs, and by eliminating popping artifacts as details are added and removed while navigating through the scene \cite{schuetz-2019-CLOD, LOM2020}.
The adaptive coordinate precision approach leads to significant performance improvements through a reduction in memory bandwidth usage, but it does not compress the points -- coordinates still use up 12 bytes of GPU memory. At this time, we deliberately did not employ sophisticated compression approaches such as delta and entropy encoding~\cite{deering1995geometry, ISENBURG:LAZ} or hierarchical encoding~\cite{10.5555/581896.581904} due to their additional computational overhead and because they make it impossible to decode coordinates individually. But in theory they could work. Delta and entropy encoding require to decode the points sequentially, which could work on a per-thread basis as each thread renders about 80 points sequentially. Generally, we think that compression could work on a per-batch (10'240 points) basis, a per thread (80 points) basis and/or a per-subgroup (32 or 64 cooperating threads) basis.
\section{Conclusions}
We have shown that software rasterization using OpenGL compute shaders is capable of rendering up to 144.7 billion points per second (Section~\ref{sec:raster_performance}), which translates to 2.3 billion points at 60 frames per second (16 ms per frame). The data structure is simple and generated on-the-fly during loading for unstructured point clouds, but LOD structures may also be generated in a preprocessing step for further performance improvements. Peak performances were observed in Morton ordered data sets, but many other orderings, for example aerial lidar scans that are sorted by timestamp and split into tiles, also provide substantial performance improvements by exploiting the spatial locality between consecutive points in memory. Data sets without sufficient locality (e.g., terrestrial laser scans) can simply be sorted by Morton order.
The source code to this paper is available at~\url{https://github.com/m-schuetz/compute_rasterizer}.
\section{Acknowledgements}
The authors wish to thank \emph{Schloss Schönbrunn Kultur- und Betriebs GmbH, Schloss Niederweiden} and \emph{Riegl Laser Measurement Systems} for providing the data set of Schloss Niederweiden, the \emph{TU Wien, Institute of History of Art, Building Archaeology and Restoration} for the Candi Banyunibo data set~\cite{isprs-archives-XLII-2-W15-555-2019}, \emph{Open Topography} and \emph{PG\&E} for the Morro Bay (CA13) data set~\cite{CA13}, \emph{Pix4D} for the Eclepens quarry data set, Sketchfab user \emph{nedo} for the \emph{old tyres} data set (CC BY 4.0), and the\emph{Stanford University Computer Graphics Laboratory} for the \emph{Stanford Bunny} data set.
This research has been funded by the FFG project \textit{LargeClouds2BIM} and the Research Cluster “Smart Communities and Technologies (Smart CT)” at TU Wien.
\section{Related Work}
\subsection{Software Rasterization of Triangle Meshes}
Early CPU-side solutions for triangle rasterization in software were largely made obsolete in the 2000s by GPUs and their high-performance rasterization components.
However, their (partly hardwired) pipeline lacks the flexibility to perform custom hierarchical or context-dependent optimizations between individual stages.
The continuously advancing programmability of GPUs has given software rasterization its second wind:
\emph{Freepipe} demonstrated that for scenes containing many, small triangles, GPU software rasterization with one thread per triangle can outperform the hardware pipeline \cite{10.1145/1730804.1730817}.
CudaRaster and Piko expanded on this idea, introducing optimizations for hierarchical triangle rasterization, achieving competitive performance with hardware rasterization even for larger triangles \cite{10.1145/2018323.2018337, 10.1145/2766973}.
Complete software implementations of OpenGL-style streaming pipelines, including sort-middle binning and hierarchical rasterization, have been presented for NVIDIA CUDA and OpenCL \cite{Kenzel:2018, Kim2021}: rather than attempting to outperform hardware rasterization, these solutions aim to provide an environment for experimenting with extensions and optimizations that may be suitable for future hardware pipelines.
A comprehensive analysis of previous software rasterization approaches and the challenges they tackle is found in \cite{frolov20comparative}.
Most recently, software rasterization has received increased attention due to the launch of the Unreal Engine 5 and its virtual geometry feature, \emph{Nanite} \cite{karis2021nanite}.
\emph{Nanite} provides both a hardware and a software pipeline for rasterization geometry and selects the proper route for rendered geometry dynamically.
In scenes with mostly pixel-sized triangles, their software pipeline reportedly achieves more than $3\times$ speedup. Its striking success begs the question whether high-performance software rasterization has not been overlooked as a viable method for other 3D representations as well.
\subsection{Software Raserization of Point Clouds}
Günther et al. proposed a GPGPU-based approach that renders points up to an order of magnitude faster than native OpenGL points primitives -- a feat that is possible because the fixed-function rendering pipeline is mainly targeted towards triangles~\cite{Gnther2013AGP}. Whenever a point modifies a pixel, their busy-loop approach locks that pixel, updates depth and color buffers, and then unlocks the pixel. Marrs et al. use atomic min/max to reproject a depth buffer to different views~\cite{Marrs2018Shadows}. Since only depth values are needed, 32 bit atomic operations are sufficient. Schütz et al. render colored point clouds via 64 bit atomic min-max operations by encoding depth and color values of each point into 64 bit integers, and using atomicMin to compute the points with the lowest depth value for each pixel inside an interleaved depth and color buffer~\cite{SCHUETZ-2021-PCC}. Our paper is based on this approach and the published source code\footnote{\url{https://github.com/m-schuetz/compute_rasterizer/releases/tag/compute_rasterizer_2021}}, and makes it several times faster for brute-forcing, but also adds support for frustum culling and LOD rendering. Rückert et al. also adapt this approach into a differentiable novel-view synthesis renderer that uses one-pixel point rendering to draw multiple resolutions of a point cloud, and neural networks to fill holes and refine the results~\cite{ruckert2021adop}. They also significantly reduce overdraw by stochastically discarding points whose assumed world-space size is considerably smaller than the pixel they are projected to.
\subsection{Level-of-Detail for Point Clouds}
Rusinkiewicz and Levoy introduced QSplat, a point-based level-of-detail data structure, as a means to interactively render large meshes~\cite{QSplat}. They use a bounding-sphere hierarchy that is traversed until a sphere with a sufficiently small radius is encountered, which is then drawn on screen. Sequential Point Trees are a more GPU-friendly approach that sequentializes a hierachical point-based representation of the model into a non-hierarchical list of points, sorted by their level of detail. From a distance, only a small continuous subset representing a lower level of detail needs to be rendered, without the need for costly traversal through a dense hierarchical structure~\cite{Dachsbacher2003}.
Layered point clouds~\cite{LPC} were one of the most impactful improvements to LOD rendering of point clouds, and variations of it still serve as today's state of the art. The original LPC constitutes a binary tree that splits the 3D space in half at each level of the hierachy. The key difference to the bounding-sphere hierarchy of QSPLATs is that each node itself is not a sphere, but a smaller point cloud comprising thousands of randomly selected points of the full point cloud. The large amount of geometry in each node reduces the amount of nodes that are required to store the full data set, which allows utilizing GPUs that are efficient at rendering hundreds of batches comprising thousands of points, each. Further research improved several aspects of layered point clouds, such as the utilized tree-structure, LOD generation times, and using density-based subsampling to properly support data sets without uniform density~\cite{Wand2008, Goswami2010, scheiblauer2011, Elseberg2013, MartinezRubi2015, Kang2019, Bormann:PCI}. Section~\ref{sec:lod} describes how our rasterization approach supports layered point clouds, using the \cite{Potree} structure as a specific example.
\subsection{Coordinate Quantization}
Quantization is the conversion of a continuous signal to discrete samples. In case of coordinates, quantization typically refers to the conversion of floating point or double coordinates (the pseudo-continuous signal) to fixed-precision integer coordinates with a carefully chosen amount of bits and a uniform coordinate precision over the whole range. The uniform precision and control over the supported coordinate range, precision, and amount of bits makes quantization a frequently used part of coordinate compression schemes, complemented by delta and entropy encoding ~\cite{deering1995geometry, ISENBURG:LAZ} or hierarchical encoding~\cite{10.5555/581896.581904}. In this paper, we use quantization to encode coordinate bits such that they can be adaptively loaded, i.e., to load just a few bits if low precision is sufficient, or optionally load additional bits to refine the previously loaded low-precision coordinate. We do not, however, apply delta, entropy, hierarchical or similar encodings because they add additional computational overhead, and they make the decoding of one point dependant on the decoding of others, i.e., they can't be decoded individually.
\section{Evaluation}
\label{sec:evaluation}
The proposed method was evaluated with the test data sets shown in Figure~\ref{fig:test_data_sets}. The smaller data sets, Eclepens and Morro Bay, are relatively easy to handle due to their low depth complexity, i.e., the number of hidden surfaces is typically small. Niederweiden poses a bigger challenge due to the higher point density and an interior room that is either occluded when viewed from the outside, or it occludes everything else when viewed from the inside, but occluded points are still processed due to the lack of occlusion culling.
The performance was computed through OpenGL timestamp queries at the start and end of each frame. All durations represent the average time of all frames over the course of one second. The evaluation was conducted on the following test systems:
\begin{itemize}
\item NVIDIA RTX 3090 24GB, AMD Ryzen 7 2700X (8 cores), 32GB RAM, running Microsoft Windows 10.
\item NVIDIA GTX 1060 3GB, AMD Ryzen 5 1600X (6 cores), 32GB RAM, running Microsoft Windows 10.
\end{itemize}
Unless specified otherwise, all reported timings are from the RTX 3090 system. The GTX 1060 (3GB) was only capable of keeping the smallest test data set, Eclepens (68M points), in memory.
\subsection{Rasterization Performance}
\label{sec:raster_performance}
Table~\ref{tab:our_performance} shows the results of rendering various data sets with our proposed basic (with frustum culling and adaptive precision; Sections ~\ref{sec:basic},~\ref{sec:vertex_compression}), prefetch (basic + each thread prefetches 4 points at a time; Section~\ref{sec:prefetch}) and LOD (Section~\ref{sec:lod}) approaches, and compares it with a previous one~\cite{SCHUETZ-2021-PCC} and with GL\_POINTS. For unstructured point-cloud data, our approach with the prefetch optimization performs the fastest in all cases -- up to three times faster than previous work in overview scenarios, and five times faster for the inside viewpoint that benefits from frustum culling.
Table~\ref{tab:performance_details} shows how many nodes and points were rendered in a frame from the given viewpoint. Processed nodes include all batches/nodes of the data set since we spawn one workgroup per node. Rendered nodes are those that pass the frustum and LOD culling steps. Frustum culling can reduce the amount of rendered nodes to slightly less than half for unstructured point clouds (Banyunibo inside), or down to several thousand out of hundreds of thousands of nodes in conjunction with LOD structures. The throughput is computed by taking the number of processed points in Table~\ref{tab:performance_details} and dividing it by the rendering time in Table~\ref{tab:our_performance}. On the RTX 3090 system and with the prefetch method for unstructued data sets, we get throughputs of 69, 126.8, 144.7, 97.5 and 142.3 billion points per second for the five scenarios -- all of them larger than the peak throughput of 50 billion points per second reported in prior work~\cite{SCHUETZ-2021-PCC}. We can also look at the throughput in terms of how many points we are able to render in real-time (60fps) by mapping the results from points per second to points per $\frac{1000}{60} \approx 16$ milliseconds, which results in 1.1, 2.0, 2.3, 1.6, 2.3 billion points per 16 milliseconds. Thus, in three out of five scenarios we were able to render two billion points in real-time.
\begingroup
\begin{table*}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} |l|rr|l|r|rrr|rr|rr|}
\hline
& \multicolumn{2}{c|}{} & & & \multicolumn{3}{c|}{ours (unstructured)} & \multicolumn{2}{c|}{ours (LOD)} & \multicolumn{2}{c|}{\cite{SCHUETZ-2021-PCC}} \\
Data Set & points & size & GPU & GL\_POINTS & basic & prefetch & hqs & basic & hqs & dedup & hqs \\
\hline
Eclepens & 69M & 1.1GB & RTX 3090 & 34.9 & 1.1 & 1.0 & 2.8 & 0.7 & 1.4 & 1.9 & 7.6 \\
Eclepens & 69M & 1.1GB & GTX 1060 & 72.2 & 6.5 & 5.1 & 16.7 & 2.1 & 5.3 & 9.5 & 26.2 \\
Morro Bay & 279M & 4.4GB & RTX 3090 & 231.7 & 2.7 & 2.2 & 9.6 & 0.8 & 1.9 & 6.0 & 33.5 \\
Banyunibo (outside) & 529M & 8.5GB & RTX 3090 & 198.3 & 4.2 & 3.3 & 9.0 & 1.3 & 3.3 & 10.7 & 25.4 \\
Banyunibo (inside) & 529M & 8.5GB & RTX 3090 & 67.9 & 2.6 & 2.2 & 6.1 & 2.1 & 4.6 & 11.1 & 24.4 \\
Niederweiden & 1000M & 16GB & RTX 3090 & 873.9 & 8.2 & 6.4 & 20.9 & 1.9 & 4.5 & 19.8 & 69.1 \\
\hline
\end{tabular*}
\caption{Comparing rendering times (in milliseconds) of GL\_POINTS with our new approach and the old one by \protect\cite{SCHUETZ-2021-PCC}. Framebuffer size: 2560 x 1140 (2.9MP). All point clouds sorted by Morton code.}
\label{tab:our_performance}
\end{table*}
\endgroup
\begingroup
\begin{table*}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} |l|r|r|r|r|r|r|r|r|}
\hline
& \multicolumn{4}{c|}{unstructured} & \multicolumn{4}{c|}{LOD} \\
\hline
& \multicolumn{2}{c|}{nodes} & \multicolumn{2}{c|}{points} & \multicolumn{2}{c|}{nodes} & \multicolumn{2}{c|}{points} \\
\hline
Data Set & processed & rendered & processed & rendered & processed & rendered & processed & rendered \\
\hline
Eclepens & 6.7k & 6.7k & 68.7M & 13.0M & 23.5k & 422 & 3.9M & 1.4M \\
Morro Bay & 27.2k & 27.2k & 278.5M & 12.2M & 93.5k & 577 & 5.6M & 1.9M \\
Banyunibo (outside) & 51.7k & 46.6k & 477.6M & 13.7M & 195.7k & 3.3k & 26.9M & 4.4M \\
Banyunibo (inside) & 51.7k & 20.9k & 214.5M & 19.1M & 195.7k & 8.2k & 67.1M & 12.5M \\
Niederweiden & 97.7k & 88.9k & 910.7M & 51.6M & 346.6k & 2.1k & 27.3M & 6.4M \\
\hline
\end{tabular*}
\caption{Detailed stats about the processed and rendered nodes and points during a frame. Processed nodes: All fixed-size batches (unstructured) or variable-sized octree nodes (LOD). Rendered Nodes: Nodes that passed frustum and LOD culling. Processed points: All points that were loaded. Rendered points: Points that pass point-wise frustum culling and early-depth, i.e., all points for which atomicMin is called. }
\label{tab:performance_details}
\end{table*}
\endgroup
\begin{figure*}
\centering
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/eclepens_w600.jpg}
\caption{Eclepens}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/morro_bay_w600.jpg}
\caption{Morro Bay }
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/candi_banyunibo_outside_w600.jpg}
\caption{Banyunibo outside}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/candi_banyunibo_inside_w600.jpg}
\caption{Banyunibo inside}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/niederweiden_w600.jpg}
\caption{Niederweiden}
\end{subfigure}
\hfill
\caption{Test data sets. (a) A quary captured with photogrammetry. (b) A coastal city, captured with aerial LIDAR. (c+d) A candi in indonesia, captured with photogrammetry and a terrestrial laser scanner. (e) A manor captured with terrestrial lasser scanning. }
\label{fig:test_data_sets}
\end{figure*}
\subsection{The Impact of Vertex Ordering}
The disadvantage of our naïve approach of generating batches by grouping 10'240 consecutive points is that the resulting performances depend on the vertex order of the data set. Figure~\ref{fig:ordering} illustrates the vertex ordering of a terrestrial laser scanner that scans on two rotational axis, first top-bottom (pitch) and then right-left (yaw). Due to this, 10'240 consecutive points usually form a 3-dimensional curve along the surface of the scanned object. The resulting batch has a large extent with mostly empty space. The next batch is formed by the next "scan-line", with an almost identical bounding box. The bottom row of Figure~\ref{fig:ordering} demonstrates vertex order and the resulting batches after the points are sorted by Morton order. The resulting batches are more compact with less empty space, and therefore more suitable to frustum culling and rendering with lower coordinate precision.
We evaluated the performance differences between scan-order and Morton-order on a subset of the Banyunibo data set comprising only of the laser scans. From an outside viewpoint, scan-order requires 5.5ms to render a frame and Morton-order requires 3.9ms, which is mostly attributed to the lower coordinate precision requirements of the compact batches. From an inside viewpoint, the scan-order requires 7.8ms to render a frame and Morton-order requires 2.1ms. In this case, the even greater performance differences can further be attributed to frustum culling, which culls about 68\% of all nodes of the Morton-ordered data set, but only 35\% nodes of the scan-ordered data set.
\begin{figure}
\centering
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/radial_batchcolors.jpg}
\caption{Batches - Scan Order}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/radial_batch.jpg}
\caption{Batch - Scan Order}
\end{subfigure}
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/morton_batchcolors.jpg}
\caption{Batches - Morton Order}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/morton_batch.jpg}
\caption{Batch - Morton Order}
\end{subfigure}
\caption{Terrestrial laser scanners typically scan along two rotation axes. (a+b) Poor locality in scan order results in huge but mostly empty batches. (c+d) Morton order efficiently establishes locality and results in compact batches. Both batches in (b) and (d) contain the same amount of points. }
\label{fig:ordering}
\end{figure}
\subsection{VR Performance}
VR rendering requires several times higher performance since we need to render data sets with higher frame rates, twice per frame, and in high quality. We evaluated the VR performance of our approach on a Valve Index HMD (1440×1600 pixels per eye) with an RTX 3090 GPU. The targeted framerate is 90fps, i.e., each frame needs to be finished in 11.1 milliseconds, or closer to about 9ms to account for additional computations and overhead in the VR pipeline. The targeted resolution is 2478 x 2740 (6.8MP) per eye, mandated by the 150\% resolution setting in SteamVR. The high resolution alleviates some aliasing issues, but high-quality shading is also needed and used to avoid severe flickering artifacts.
We evaluated the VR performance with an outside-in view of the Candi Banyunibo data set, comprising 529 million points in 195k octree nodes. 21.4 million points in 2k nodes passed the frustum and LOD culling routines. After early-depth testing, 8 million points were drawn with atomicMin. The total frame time was 8.3 milliseconds, which provides sufficient reserves for additional computations and overhead of the VR rendering pipeline. Rendering the depth buffer for the HQS shader took 1.7 ms for both eyes, or 1.3 ms when rendering just a single eye, which demonstrates the benefits of drawing each point to both render targets within a single compute shader invocation. Similarly, drawing the color buffer of the HQS approach took 2.5 ms for both eyes and 1.5 ms for a single eye. The resolve pass, which enlarges the more sparsely rendered points in the periphery and stores the colors into an OpenGL texture, takes about 2.7 ms per frame for VR rendering, which is mainly attributed to the large and dynamic dilation kernels of 3 x 3 (center) to 7 x 7 (periphery) pixels. Finally, clearing the relatively large depth and color buffers (2468 x 2740 per eye) takes about 0.6 ms per frame.
\section{Method}
The core aspect of our rasterization method is that we assign larger batches of points to be rendered by each workgroup (e.g., 128 threads) instead of only one point per workgroup thread. These larger batches (e.g., 10k points -- 80 per thread) enable several optimizations that would be too costly on small batches (e.g., 128 points -- 1 per thread), but whose costs amortize with each additionally rendered point. Using larger batches with varying amounts of points also allows us to support several widely used level-of-detail structures, as discussed in Section~\ref{sec:lod}. Figure~\ref{fig:pipeline} gives an overview of the rendering steps within a workgroup.
We will first describe our basic rendering pipeline, which we will then gradually expand by additional features and optimizations that ultimately allow us to render point clouds several times faster than the state of the art.
\begin{figure}
\centering
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/flowchart.png}
\caption{Flowchart of a single workgroup invocation.}
\end{subfigure}
\caption{Each workgroup renders one batch or octree node. If its projected bounding box is small, fewer coordinate bits are loaded, reducing memory bandwidth usage and boosting render performance accordingly.}
\label{fig:pipeline}
\end{figure}
\subsection{Data Structure}
The data structure consists of a list of batches and a list of points. Each batch represents a number of consecutive points in the point list (see Figure~\ref{fig:batches_memory}), so for each batch we store the index of the first point of that batch, the number of points in that batch, and their bounding box. Each point consists of 4 attributes: low, medium, and high precision parts of the coordinates (details in Section~\ref{sec:vertex_compression}), and a color value. The 4 attributes are stored in a struct-of-arrays fashion, i.e., in separate buffers for low-precision part of the coordinate, another for medium precision part, etc., so that we only need to load components from memory during rendering that are actually needed. For the majority of points, this will just be the low-precision coordinate bits (4 bytes out of 16 bytes per point).
For regular, unstructured point-cloud data sets, we suggest to simply group about 10'240 consecutive points into batches and compute their bounding box while loading the points. In practice, the majority of data sets we've encountered already provide sufficient locality, especially data sets generated through aerial LIDAR or photogrammetry. But not all of them do and almost all can benefit from sorting by Morton code (z-order) -- an easy to implement and efficient approach to create data sets with good locality~\cite{10.1145/588011.588037, 1532800, 10.1007/3-540-45033-5_3, LOM2020, SCHUETZ-2021-PCC}. For the remainder of this section, we will assume that data sets exhibit sufficient locality, either by default or through sorting by Morton-code, and we will further evaluate and discuss the impact of poor locality in Section~\ref{sec:evaluation}.
Figure~\ref{fig:teaser} and Figure~\ref{fig:batches_boxes_colors} show the results of grouping 10'240 consecutive points of a Morton-ordered point cloud into batches. Although the locality isn't perfect, the majority of batches are sufficiently compact with only a couple of outliers that experience extremely large jumps between different clusters of points. The main advantages of this approach are that it's trivial to implement, requires no preprocessing, and can be done while a data set is loaded with negligible impact on performance.
\begin{figure}
\centering
\hfill
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/batches_unstructured.png}
\caption{Unstructured Point Clouds}
\end{subfigure}
\par\bigskip
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/batches_lod.png}
\caption{LOD Batches / Nodes}
\end{subfigure}
\caption{(a) For unstructured, Morton-code-ordered data sets, we group 10'240 consecutive points into a batch. (b) For LOD data, each octree node equals a batch and batches additionally store memory location and number of points inside the node/batch.}
\label{fig:batches_memory}
\end{figure}
\begin{figure*}
\centering
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_0.jpg}
\caption{Point Cloud Data Set}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_1.jpg}
\caption{Batches}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_2.jpg}
\caption{Bounding Boxes}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_3.jpg}
\caption{A single batch}
\end{subfigure}
\caption{A Morton-code ordered point cloud, grouped into batches of 10'240 consecutive points. The locality that is provided by the Morton order is sufficient for frustum culling and a bounding-box-based vertex compression/quantization scheme. }
\label{fig:batches_boxes_colors}
\end{figure*}
\subsection{Basic Rendering Pipeline}
\label{sec:basic}
The basic pipeline spawns one compute workgroup for each batch to render its points. Each workgroup comprises 128 threads and each thread renders $n$ points. In practice, we found $n$ between 60 to 200 to perform equally well on an RTX 3090, and will therefore assume $n = 80$ points per thread (128 * 80 = 10'240 points per batch) throughout the paper. The rasterization process of each individual point is the same as in previous work \cite{SCHUETZ-2021-PCC}: Each point is projected to pixel coordinates, its depth and color value are encoded into a single 64 bit integer, and atomicMin is used to compute the point with the smallest depth value for each pixel, as shown in Listing \ref{lst:basicRasterization}. An early-depth test (a non-atomic depth comparison before an expensive atomic depth-compare and write) ensures that the majority of occluded points are rejected before invoking an expensive atomic update.
\begin{lstlisting}[language=Java,label={lst:basicRasterization},caption={Point rasterization, including an early-depth test.},captionpos=b]
vec4 pos = worldViewProj * position;
int pixelID = toPixelID(pos);
int64_t depth = floatBitsToInt(pos.w);
int64_t newPoint = (depth << 32) | pointIndex;
// fetch previously written point
uint64_t oldPoint = framebuffer[pixelID];
// Early-depth test
if(newPoint < oldPoint){
atomicMin(framebuffer[pixelID], newPoint);
}
\end{lstlisting}
The first improvement over \cite{SCHUETZ-2021-PCC} is that the larger workloads per workgroup allow implementing efficient workgroup-wise frustum culling with a granularity of 10'240 points. At the start of the workgroup invocation, we first load the bounding box of the current batch and abort right away if it does not intersect with the view frustum. Another improvement is that spawning fewer workgroups and giving each of them larger tasks reduces the scheduling overhead of the GPU.
\subsection{Adaptive Vertex Precision}
\label{sec:vertex_compression}
One of the main bottlenecks in previous work \cite{SCHUETZ-2021-PCC} is memory bandwidth usage. They reported a peak performance of 50 billion points per second using 16 bytes per point, which utilizes 85\% of the GPU's memory bandwidth. Since this speed is limited by bandwidth, any significant improvement of brute-force rendering performance requires some form of vertex compression.
We propose an adaptive precision scheme that allows us to load as many bits for coordinates as necessary for a given viewpoint. We achieve this by splitting the bits of a coordinate into three separate buffers storing the low, medium, and high precision parts, as shown in Figure~\ref{fig:memory_layout}. The low precision part always needs to be loaded. The medium precision part can be optionally loaded to recover some of the bits that we removed, and the high precision part is used to optionally recover the remaining bits. The different precision levels are established via coordinate quantization, i.e., by converting coordinates to fixed-precision integers with specific amounts of bits, and then splitting the quantized bits. By quantizing the point coordinates with respect to the bounding box of the batch instead of the bounding box of the whole 3D model, we can achieve a high coordinate precision with few bits. Figure~\ref{fig:quantization} illustrates that the resolution of the coordinates quickly approaches the resolution of the pixel grid. In our case, we first convert the coordinates within a batch to 30 bit fixed-precision integer coordinates that indicate each point's position within its batch. The X-coordinate is computed as follows, for example:
\begin{equation}
X_{30} = min( \lfloor 2^{30} * \frac{x - boxMin.x}{boxSize.x} \rfloor , 2^{30} - 1)
\label{eq:targetIndexPermute}
\end{equation}
These 30 bits are then split into three 10-bit components representing low, medium and high precision parts. The 10 most significant bits (bit indices 20 to 29) of this 30 bit integer are the low-precision bits. They tell us the coordinate of a point within a batch with a precision of $\frac{1}{2^{10}} = \frac{1}{1024}$ of its size. Considering that the majority of rendered batches in any given viewpoint are typically smaller than a couple of hundreds of pixels, and that 10 bits grant us 1024 different coordinate values, we find that 10 bits per axis are sufficient to render most points with sub-pixel coordinate precision. Due to buffer alignment recommendations, simplicity, and because the smallest accessible primitive values on GPUs are typically 32 bits, we then pack the 10-bit x, y and z components into a single 32 bit integer with the remaining 2 bits used as padding, as shown in Listing~\ref{lst:encoding}. The result is a 32-bit integer buffer where each 4-byte element contains the 10 lowest precision bits of the three coordinate axes of a single point.
\lstset{
mathescape,
literate={->}{$\rightarrow$}{2}
{ε}{$\varepsilon$}{1}
}
\begin{lstlisting}[language=Java,label={lst:encoding},caption={Encoding the most significant 10 bits of each axis into a single 32 bit integer. },captionpos=b]
uint32_t X$_{low}$ = (X$_{30}$ >> 20) & 1023;
uint32_t Y$_{low}$ = (Y$_{30}$ >> 20) & 1023;
uint32_t Z$_{low}$ = (Z$_{30}$ >> 20) & 1023;
uint32_t encoded = X$_{low}$ | (Y$_{low}$ << 10) | (Z$_{low}$ << 20)
\end{lstlisting}
Likewise, we generate two more buffers for the medium (bits 10 to 19, middle) and high precision bits (bits 0 to 9, least-significant). During rendering, each workgroup can now choose to load a single 4-byte integer containing just the low precision bits that grants 1'024 coordinate values, 2 such integers (low+medium precision) granting 1'048'576 coordinate values, or 3 such integers (low+medium+high precision), which theoretically allows for one billion different coordinate values. However, we should note that the usefulness of 30 bit precision is limited in practice because we still convert the integer coordinates to floating-point coordinates during rendering, where the conversion from a 30-bit fixed-precision integer to a 32-bit floating point value will incur a loss of precision due to rounding errors and the non-uniform distribution of precision in floats (high precision for small values, low precision for large values). This issue is not specific to our approach, however, as point clouds are typically already stored in integer coordinates on disk and the conversion to floats for rendering has always been an issue for point clouds with a large extent. Handling this issue is out of scope of this paper, but we would like to note that storing coordinates as 30 bit fixed-precision integers would allow us to convert them to highly accurate double-precision coordinates (e.g., for measurements), while traditional floating point storage causes an irrecoverable loss of precision.
\begin{figure}
\centering
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_0.jpg}
\caption{2 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_1.jpg}
\caption{3 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_2.jpg}
\caption{4 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_3.jpg}
\caption{5 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_4.jpg}
\caption{6 bits}
\end{subfigure}
\caption{Coordinate quantization provides a trade-off between precision and storage size. The amount of possible coordinate values per axis is given by $2^{bits}$.}
\label{fig:quantization}
\end{figure}
The coordinate precision during rendering is determined through the projected size of a batch. If a batch projects to less than 500 pixels, we use 10 bit coordinates. The reason for choosing 500 pixels as the threshold even though 10 bits can represent 1024 different coordinates values is that quantization changes the distance between any two points by up to the size of a quantization grid cell, i.e., points that were one grid cell's size apart might now be twice as far apart. Using half the size of the quantization grid as the pixel threshold ensures that we do not introduce additional holes between rasterized points.
Compared to previous work~\cite{SCHUETZ-2021-PCC}, this approach reduces the required memory bandwidth for the majority of rendered points from 16 bytes down to 8, since only 4 bytes instead of 12 are needed for most coordinates, and another 4 for colors. However, we can further cut this into half by using a visibility-buffer (item-buffer) approach~\cite{Burns2013Visibility, 10.1145/357332.357335}, i.e., we render point indices rather than colors into the framebuffer during the geometry pass, and transform the point indices to vertex colors in a post-processing step. Doing so reduces the amount of memory accesses to color values down from the number of processed points to the number of visible points.
\begin{figure}
\centering
\hfill
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/points_aos.png}
\caption{Common point-cloud memory layout with 16 bytes per point.}
\end{subfigure}
\hfill
\par\bigskip
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/points_soa.png}
\caption{Splitting coordinate bits into a low, medium and high-precision buffer. Each buffer stores 10 bits per coordinate axis, encoded into 4 bytes per point. }
\end{subfigure}
\caption{(a) Point cloud renderers typically load at least 16 bytes per point during rendering. (b) Splitting coordinate bits into separate buffers allows loading just the required bits, depending on the viewpoint. 10-bit coordinates typically suffice for the majority of points. }
\label{fig:memory_layout}
\end{figure}
\subsection{Vertex Pre-Fetching}
\label{sec:prefetch}
To identify performance bottlenecks of the above approach, we performed a direct port of its GLSL shader code to NVIDIA CUDA.
Doing so enabled the use of the proprietary Nsight Compute tool to pinpoint suboptimal behavior in our routines.
Our investigation into dominant stall reasons revealed that performance is---unsurprisingly---governed by memory operations, with approximately $70\%$ of the total kernel run time spent on them.
Since we choose the batch size to be a multiple of the work group size, threads persist and process multiple points in a loop before returning.
Naively, the corresponding point data is fetched and rasterized to the framebuffer in each iteration.
Batches simply define a linear range in memory, thus their contents can be transferred by workgroups with perfect coalescence.
Furthermore, the compression scheme outlined in Section \ref{sec:vertex_compression} economizes on available bandwidth when fetching point data.
However, loading each processed point individually at the start of its associated iteration still incurs a direct dependency of the steps in Listing \ref{lst:basicRasterization} on global device memory latency.
Due to the early-depth test in its body and the parameterizable iteration count, the compiler cannot trivially unroll the loop without creating secondary issues associated with complex program flow (e.g., frequent instruction cache misses).
Hence, we perform manual pre-fetching of point data. For the lowest precision (using 32 bits per point), we execute a vectorized 128-bit load in each thread, yielding enough data for four successive iterations; for medium-precision points, a 128-bit load fetches the data for two iterations.
This policy significantly alleviates reported stalls due to memory latency and simultaneously reduces the number of memory instructions, without compromising on coalesced accesses.
Overall, we found pre-fetching to cause $\approx 30\%$ performance gain.
A welcome side effect, though less impactful, is the reduction of updates to the frame buffer by this policy, if points are spatially sorted. Since threads now load and process consecutive points in memory, the early-depth test has a higher chance of failing: threads are now more likely to query (and find) information in the L1 cache for pixels that they wrote to in a previous iteration.
We note that, with pre-fetching in place, the main remaining bottleneck is the code block for early-depth testing and framebuffer updates. Although these accesses are somewhat localized if points are spatially ordered, a residual irregularity remains in their access pattern.
The coarse-grained memory transfer policy of the GPU consequently causes more than $2\times$ the amount of actually accessed information to be transferred.
However, this update pattern is inherent to the overall routine design, and any further attempts to improve on it without extensive revision led to diminishing returns.
\subsection{Level-of-Detail Rendering}
\label{sec:lod}
In this section we will describe how support for some of the most popular LOD structures for point clouds -- Layered Point Clouds (LPC)~\cite{LPC} and its variations -- can be added to our point rasterization approach. Layered point clouds are a hierarchical, spatial acceleration structure where points with varying density are stored in the nodes of the tree. Lower levels contain coarse low-density subsets of the whole point cloud. With each level, the size of the nodes decreases, leading to an increase of the the density of points as the number of points in each node stays roughly the same. The structures are often additive -- meaning that higher LODs increase detail by rendering more points in addition to lower LODs instead of replacing them -- but replacing schemes are also possible.
We implement and evaluate our support for LPC structures based on the Potree format, which constitutes a variation of LPC that uses an octree and populates nodes with subsamples of the point cloud with a specific, level-dependant minimum distance. Each octree node comprises about 1k to 50k points, and Potree itself typically renders about 100 to 1000 nodes for any given viewpoint. Up until now we have assumed that each batch renders exactly 10'240 points, but we can easily adapt our rasterization approach to support the Potree structure by allowing varying amounts of points per batch, as shown in Figure~\ref{fig:batches_memory}. In addition to the bounding box, we will also need to store the number of points as well as the memory location of the first point for each batch. The workgroup size of 128 threads remains the same, but each thread will now render $\lceil \frac{batch.numPoints}{128} \rceil$ points instead of exactly 80.
The Potree format is structured such that octree nodes whose projected bounding boxes are small can be entirely ignored, because the points stored in their parents will already provide sufficient detail. This means that the procedure in Section~\ref{sec:vertex_compression} that is used to determine the coordinate precision during rendering based on the screen-size of the bounding box can now be used to entirely cull the node. We suggest to cull the nodes with following conditions in mind: Each Potree node is cubic and stores a subsample of points with a resolution that approximately matches a grid with $128^3$ cells, and our rasterizer maps each point to exactly 1 pixel. To avoid holes between rendered points, we therefore suggest to cull those nodes that are smaller than 100 pixels on screen. However, it is also viable to cull larger octree nodes on lower end GPUs to improve performance and cover up the resulting holes in a post-processing pass, e.g., via depth-aware dilation or more sophisticated hole-filling approaches~\cite{ruckert2021adop, AutoSplats, Rosenthal2008, grossman1998point, Pintus2011, Marroquim:2007:pbg}.
\subsection{High-Quality Shading}
LOD rendering works well in conjunction with high-quality splatting (point-sprites or surfels~\cite{Botsch:HQS}) or shading (one-pixel points~\cite{SCHUETZ-2021-PCC}), a form of color filtering for point clouds that blends overlapping points together. Colors and amount of points inside a pixel within a certain depth range (e.g., 1\% behind the closest point) are summed up during the geometry processing stage, and a post-processing shader then divides the sum of colors by the counters to compute the average. Schütz et al. \cite{SCHUETZ-2021-PCC} suggest two variations: one that uses two 64-bit atomicAdd instructions per point into four 32 bit integers to sum up color values and counters, and another variation that uses a single 64 bit atomic instruction per point to compute the sum of up to 255 points, with a fallback that uses two 64-bit atomics if more than 255 points contribute to the average. However, when using an LOD structure, the amount of overlapping points with similar depth is essentially guaranteed to be lower than 255, so we can safely use high-quality shading with just a single 64-bit atomic add instruction per point. This limit can even be raised to up to 1023 points by using the ``non-robust" variation without overflow protection. For unstructured data sets, we suggest to use the variation with overflow protection due to the potentially large amount of overlaps.
\subsection{Virtual Reality Rendering}
Virtual reality (stereo) rendering greatly increases the performance requirements -- even more so for point clouds as they typically suffer from aliasing artifacts that need to be addressed to provide an acceptable VR experience. In addition to a specific configuration of our pipeline (use LOD rendering for performance, high-quality shading to reduce aliasing and large framebuffers for additional anti-aliasing via supersampling), we can exploit specific properties of VR rendering in our approach.
First, the scene needs to be drawn twice -- once for each eye. Instead of duplicating the entire rasterization pipeline by calling the compute shaders twice, we suggest to modify the shader to simply draw each point into both framebuffers. While this doesn't double the performance, it provides a significant improvement, as discussed in Section~\ref{sec:evaluation}.
Second, in a VR setup, details in the periphery aren't as noticeable as details in the center of the view. This is partially because most details are perceived in the gaze direction, i.e., mostly straight ahead in VR, but also because the rendered image will be distorted before it is shown inside the HMD to counter the lens distortion~\cite{Vlachos:2015:AVR}. The applied distortion compresses peripheral regions, thus reducing the amount of detail in the image. We therefore suggest to render peripheral regions of the framebuffer with reduced geometric complexity by raising the threshold for LOD culling, e.g., culling nodes smaller than 300 pixels in the periphery, nodes smaller than 100 pixels in the center of the view, and 200 pixels in between. The resulting holes between points are then filled in post-processing, in our case via a simple depth-aware dilation shader that increases point sizes to up to 7x7 in the periphery and 3x3 in the center. Depth-aware means that the closest point within the pixel ranges are expanded.
\section{Acknowledgements}
The authors wish to thank \emph{Schloss Schönbrunn Kultur- und Betriebs GmbH, Schloss Niederweiden} and \emph{Riegl Laser Measurement Systems} for providing the data set of Schloss Niederweiden, the \emph{TU Wien, Institute of History of Art, Building Archaeology and Restoration} for the Candi Banyunibo data set~\cite{isprs-archives-XLII-2-W15-555-2019}, \emph{Open Topography} and \emph{PG\&E} for the Morro Bay (CA13) data set~\cite{CA13}, \emph{Pix4D} for the Eclepens quarry data set, Sketchfab user \emph{nedo} for the \emph{old tyres} data set (CC BY 4.0), and the\emph{Stanford University Computer Graphics Laboratory} for the \emph{Stanford Bunny} data set.
This research has been funded by the FFG project \textit{LargeClouds2BIM} and the Research Cluster “Smart Communities and Technologies (Smart CT)” at TU Wien.
\section{Discussion}
In this section we'd like to discuss several issues and limitations, as well as potential improvements that we have not evaluated, yet.
First, we believe that this is a significant improvement for point cloud rendering, but it's not useful for games in its current state. Point clouds require a large amount of colored vertices to represent geometry without holes and sufficient color-detail, while meshes can use a single textured triangle in place of thousands of points. However, massive amounts of point cloud data sets exist as a result of scanning the real world, and this paper provides tools to render them faster without the need to process the data into LOD structures or meshes. Still, we hope that the presented approach might provide useful insights in future developments of hybrid hardware+software rasterization approaches for triangle meshes. For example, we could imagine that an approach like adaptive coordinate precision could be used to treat smaller batches like a point cloud and only load data that is relevant for triangles for larger batches.
The quality for virtual reality currently suffers from lack of proper color filtering. Although high-quality shading is applied and improves the results via blending, the issue is that the LOD structure removes most of the overlapping points, thus the blended result is not representative of all points, including the discarded ones. The results can be improved by applying color filtering (e.g. computing averages) to points in lower levels of detail during the construction of the LOD structure~\cite{QSplat,Wand2008,schuetz-2019-CLOD} (similar to mip mapping). Furthermore, implementing continuous LOD could improve the visual quality through a subtle transition in point density between LODs, and by eliminating popping artifacts as details are added and removed while navigating through the scene \cite{schuetz-2019-CLOD, LOM2020}.
The adaptive coordinate precision approach leads to significant performance improvements through a reduction in memory bandwidth usage, but it does not compress the points -- coordinates still use up 12 bytes of GPU memory. At this time, we deliberately did not employ sophisticated compression approaches such as delta and entropy encoding~\cite{deering1995geometry, ISENBURG:LAZ} or hierarchical encoding~\cite{10.5555/581896.581904} due to their additional computational overhead and because they make it impossible to decode coordinates individually. But in theory they could work. Delta and entropy encoding require to decode the points sequentially, which could work on a per-thread basis as each thread renders about 80 points sequentially. Generally, we think that compression could work on a per-batch (10'240 points) basis, a per thread (80 points) basis and/or a per-subgroup (32 or 64 cooperating threads) basis.
\section{Conclusions}
We have shown that software rasterization using OpenGL compute shaders is capable of rendering up to 144.7 billion points per second (Section~\ref{sec:raster_performance}), which translates to 2.3 billion points at 60 frames per second (16 ms per frame). The data structure is simple and generated on-the-fly during loading for unstructured point clouds, but LOD structures may also be generated in a preprocessing step for further performance improvements. Peak performances were observed in Morton ordered data sets, but many other orderings, for example aerial lidar scans that are sorted by timestamp and split into tiles, also provide substantial performance improvements by exploiting the spatial locality between consecutive points in memory. Data sets without sufficient locality (e.g., terrestrial laser scans) can simply be sorted by Morton order.
The source code to this paper is available at~\url{https://github.com/m-schuetz/compute_rasterizer}.
\section{Evaluation}
\label{sec:evaluation}
The proposed method was evaluated with the test data sets shown in Figure~\ref{fig:test_data_sets}. The smaller data sets, Eclepens and Morro Bay, are relatively easy to handle due to their low depth complexity, i.e., the number of hidden surfaces is typically small. Niederweiden poses a bigger challenge due to the higher point density and an interior room that is either occluded when viewed from the outside, or it occludes everything else when viewed from the inside, but occluded points are still processed due to the lack of occlusion culling.
The performance was computed through OpenGL timestamp queries at the start and end of each frame. All durations represent the average time of all frames over the course of one second. The evaluation was conducted on the following test systems:
\begin{itemize}
\item NVIDIA RTX 3090 24GB, AMD Ryzen 7 2700X (8 cores), 32GB RAM, running Microsoft Windows 10.
\item NVIDIA GTX 1060 3GB, AMD Ryzen 5 1600X (6 cores), 32GB RAM, running Microsoft Windows 10.
\end{itemize}
Unless specified otherwise, all reported timings are from the RTX 3090 system. The GTX 1060 (3GB) was only capable of keeping the smallest test data set, Eclepens (68M points), in memory.
\subsection{Rasterization Performance}
\label{sec:raster_performance}
Table~\ref{tab:our_performance} shows the results of rendering various data sets with our proposed basic (with frustum culling and adaptive precision; Sections ~\ref{sec:basic},~\ref{sec:vertex_compression}), prefetch (basic + each thread prefetches 4 points at a time; Section~\ref{sec:prefetch}) and LOD (Section~\ref{sec:lod}) approaches, and compares it with a previous one~\cite{SCHUETZ-2021-PCC} and with GL\_POINTS. For unstructured point-cloud data, our approach with the prefetch optimization performs the fastest in all cases -- up to three times faster than previous work in overview scenarios, and five times faster for the inside viewpoint that benefits from frustum culling.
Table~\ref{tab:performance_details} shows how many nodes and points were rendered in a frame from the given viewpoint. Processed nodes include all batches/nodes of the data set since we spawn one workgroup per node. Rendered nodes are those that pass the frustum and LOD culling steps. Frustum culling can reduce the amount of rendered nodes to slightly less than half for unstructured point clouds (Banyunibo inside), or down to several thousand out of hundreds of thousands of nodes in conjunction with LOD structures. The throughput is computed by taking the number of processed points in Table~\ref{tab:performance_details} and dividing it by the rendering time in Table~\ref{tab:our_performance}. On the RTX 3090 system and with the prefetch method for unstructued data sets, we get throughputs of 69, 126.8, 144.7, 97.5 and 142.3 billion points per second for the five scenarios -- all of them larger than the peak throughput of 50 billion points per second reported in prior work~\cite{SCHUETZ-2021-PCC}. We can also look at the throughput in terms of how many points we are able to render in real-time (60fps) by mapping the results from points per second to points per $\frac{1000}{60} \approx 16$ milliseconds, which results in 1.1, 2.0, 2.3, 1.6, 2.3 billion points per 16 milliseconds. Thus, in three out of five scenarios we were able to render two billion points in real-time.
\begingroup
\begin{table*}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} |l|rr|l|r|rrr|rr|rr|}
\hline
& \multicolumn{2}{c|}{} & & & \multicolumn{3}{c|}{ours (unstructured)} & \multicolumn{2}{c|}{ours (LOD)} & \multicolumn{2}{c|}{\cite{SCHUETZ-2021-PCC}} \\
Data Set & points & size & GPU & GL\_POINTS & basic & prefetch & hqs & basic & hqs & dedup & hqs \\
\hline
Eclepens & 69M & 1.1GB & RTX 3090 & 34.9 & 1.1 & 1.0 & 2.8 & 0.7 & 1.4 & 1.9 & 7.6 \\
Eclepens & 69M & 1.1GB & GTX 1060 & 72.2 & 6.5 & 5.1 & 16.7 & 2.1 & 5.3 & 9.5 & 26.2 \\
Morro Bay & 279M & 4.4GB & RTX 3090 & 231.7 & 2.7 & 2.2 & 9.6 & 0.8 & 1.9 & 6.0 & 33.5 \\
Banyunibo (outside) & 529M & 8.5GB & RTX 3090 & 198.3 & 4.2 & 3.3 & 9.0 & 1.3 & 3.3 & 10.7 & 25.4 \\
Banyunibo (inside) & 529M & 8.5GB & RTX 3090 & 67.9 & 2.6 & 2.2 & 6.1 & 2.1 & 4.6 & 11.1 & 24.4 \\
Niederweiden & 1000M & 16GB & RTX 3090 & 873.9 & 8.2 & 6.4 & 20.9 & 1.9 & 4.5 & 19.8 & 69.1 \\
\hline
\end{tabular*}
\caption{Comparing rendering times (in milliseconds) of GL\_POINTS with our new approach and the old one by \protect\cite{SCHUETZ-2021-PCC}. Framebuffer size: 2560 x 1140 (2.9MP). All point clouds sorted by Morton code.}
\label{tab:our_performance}
\end{table*}
\endgroup
\begingroup
\begin{table*}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} |l|r|r|r|r|r|r|r|r|}
\hline
& \multicolumn{4}{c|}{unstructured} & \multicolumn{4}{c|}{LOD} \\
\hline
& \multicolumn{2}{c|}{nodes} & \multicolumn{2}{c|}{points} & \multicolumn{2}{c|}{nodes} & \multicolumn{2}{c|}{points} \\
\hline
Data Set & processed & rendered & processed & rendered & processed & rendered & processed & rendered \\
\hline
Eclepens & 6.7k & 6.7k & 68.7M & 13.0M & 23.5k & 422 & 3.9M & 1.4M \\
Morro Bay & 27.2k & 27.2k & 278.5M & 12.2M & 93.5k & 577 & 5.6M & 1.9M \\
Banyunibo (outside) & 51.7k & 46.6k & 477.6M & 13.7M & 195.7k & 3.3k & 26.9M & 4.4M \\
Banyunibo (inside) & 51.7k & 20.9k & 214.5M & 19.1M & 195.7k & 8.2k & 67.1M & 12.5M \\
Niederweiden & 97.7k & 88.9k & 910.7M & 51.6M & 346.6k & 2.1k & 27.3M & 6.4M \\
\hline
\end{tabular*}
\caption{Detailed stats about the processed and rendered nodes and points during a frame. Processed nodes: All fixed-size batches (unstructured) or variable-sized octree nodes (LOD). Rendered Nodes: Nodes that passed frustum and LOD culling. Processed points: All points that were loaded. Rendered points: Points that pass point-wise frustum culling and early-depth, i.e., all points for which atomicMin is called. }
\label{tab:performance_details}
\end{table*}
\endgroup
\begin{figure*}
\centering
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/eclepens_w600.jpg}
\caption{Eclepens}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/morro_bay_w600.jpg}
\caption{Morro Bay }
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/candi_banyunibo_outside_w600.jpg}
\caption{Banyunibo outside}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/candi_banyunibo_inside_w600.jpg}
\caption{Banyunibo inside}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\textwidth}
\includegraphics[width=\textwidth]{images/niederweiden_w600.jpg}
\caption{Niederweiden}
\end{subfigure}
\hfill
\caption{Test data sets. (a) A quary captured with photogrammetry. (b) A coastal city, captured with aerial LIDAR. (c+d) A candi in indonesia, captured with photogrammetry and a terrestrial laser scanner. (e) A manor captured with terrestrial lasser scanning. }
\label{fig:test_data_sets}
\end{figure*}
\subsection{The Impact of Vertex Ordering}
The disadvantage of our naïve approach of generating batches by grouping 10'240 consecutive points is that the resulting performances depend on the vertex order of the data set. Figure~\ref{fig:ordering} illustrates the vertex ordering of a terrestrial laser scanner that scans on two rotational axis, first top-bottom (pitch) and then right-left (yaw). Due to this, 10'240 consecutive points usually form a 3-dimensional curve along the surface of the scanned object. The resulting batch has a large extent with mostly empty space. The next batch is formed by the next "scan-line", with an almost identical bounding box. The bottom row of Figure~\ref{fig:ordering} demonstrates vertex order and the resulting batches after the points are sorted by Morton order. The resulting batches are more compact with less empty space, and therefore more suitable to frustum culling and rendering with lower coordinate precision.
We evaluated the performance differences between scan-order and Morton-order on a subset of the Banyunibo data set comprising only of the laser scans. From an outside viewpoint, scan-order requires 5.5ms to render a frame and Morton-order requires 3.9ms, which is mostly attributed to the lower coordinate precision requirements of the compact batches. From an inside viewpoint, the scan-order requires 7.8ms to render a frame and Morton-order requires 2.1ms. In this case, the even greater performance differences can further be attributed to frustum culling, which culls about 68\% of all nodes of the Morton-ordered data set, but only 35\% nodes of the scan-ordered data set.
\begin{figure}
\centering
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/radial_batchcolors.jpg}
\caption{Batches - Scan Order}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/radial_batch.jpg}
\caption{Batch - Scan Order}
\end{subfigure}
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/morton_batchcolors.jpg}
\caption{Batches - Morton Order}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\columnwidth}
\includegraphics[width=\textwidth]{images/orderings/morton_batch.jpg}
\caption{Batch - Morton Order}
\end{subfigure}
\caption{Terrestrial laser scanners typically scan along two rotation axes. (a+b) Poor locality in scan order results in huge but mostly empty batches. (c+d) Morton order efficiently establishes locality and results in compact batches. Both batches in (b) and (d) contain the same amount of points. }
\label{fig:ordering}
\end{figure}
\subsection{VR Performance}
VR rendering requires several times higher performance since we need to render data sets with higher frame rates, twice per frame, and in high quality. We evaluated the VR performance of our approach on a Valve Index HMD (1440×1600 pixels per eye) with an RTX 3090 GPU. The targeted framerate is 90fps, i.e., each frame needs to be finished in 11.1 milliseconds, or closer to about 9ms to account for additional computations and overhead in the VR pipeline. The targeted resolution is 2478 x 2740 (6.8MP) per eye, mandated by the 150\% resolution setting in SteamVR. The high resolution alleviates some aliasing issues, but high-quality shading is also needed and used to avoid severe flickering artifacts.
We evaluated the VR performance with an outside-in view of the Candi Banyunibo data set, comprising 529 million points in 195k octree nodes. 21.4 million points in 2k nodes passed the frustum and LOD culling routines. After early-depth testing, 8 million points were drawn with atomicMin. The total frame time was 8.3 milliseconds, which provides sufficient reserves for additional computations and overhead of the VR rendering pipeline. Rendering the depth buffer for the HQS shader took 1.7 ms for both eyes, or 1.3 ms when rendering just a single eye, which demonstrates the benefits of drawing each point to both render targets within a single compute shader invocation. Similarly, drawing the color buffer of the HQS approach took 2.5 ms for both eyes and 1.5 ms for a single eye. The resolve pass, which enlarges the more sparsely rendered points in the periphery and stores the colors into an OpenGL texture, takes about 2.7 ms per frame for VR rendering, which is mainly attributed to the large and dynamic dilation kernels of 3 x 3 (center) to 7 x 7 (periphery) pixels. Finally, clearing the relatively large depth and color buffers (2468 x 2740 per eye) takes about 0.6 ms per frame.
\section{Introduction}
With the introduction of hardware with dedicated triangle rasterization units, hand-crafting rasterization routines in software became largely obsolete. Such custom-built rasterizers have nevertheless remained an ongoing topic of research in order to develop and study new rasterization approaches. Some of them eventually managed to beat hardware rasterization in specific scenarios~\cite{10.1145/1730804.1730817}, but in general, dedicated hardware remains the fastest approach. Nanite is the first framework that promises practical improvements for 3D games via hybrid software and hardware rasterization~\cite{karis2021nanite}. They found that directly rasterizing the fragments of a small triangle with atomic min-max operations can be faster than pushing the triangle through the hardware rendering pipeline. Therefore, only larger triangles are rasterized via hardware.
Point-cloud models offer additional opportunities for efficient software rasterization, as the hardware rendering pipeline is largely dedicated to the rasterization of triangles and not points. In this paper, we consider point clouds as 3D models made of colored vertices, where each vertex is projected to exactly one pixel. Although this is a fairly strict limitation, it allows us to device algorithms that compete with graphics APIs that also only support one-pixel points, such as DirectX (POINTLIST primitive) and all backends that use it (WebGL, WebGPU, ANGLE, MS Windows games and applications, ...). We intent to support larger point-sprites in the future and use the evaluated performances of one-pixel points as a baseline for comparisons. Point clouds have no connectivity, so index buffers or vertex duplication are not required. The lack of a connected surface also makes uv-maps and textures irrelevant, which is why colors are typically directly stored on a per-vertex basis. Furthermore, point clouds acquired by laser scanners do not contain surface normals. Normals could be computed in a pre-processing step, but computed normals are not robust in sparsely sampled regions with high-frequency details such as vegetation, strings/cables/wires or even noise. We will therefore not consider normals in this paper, either.
Our approach builds on~\cite{SCHUETZ-2021-PCC} to further optimize several aspects of software rasterization of points, which leads to an up to 3x higher brute-force performance. Specifically, our contributions to the state of the art of software rasterization of point clouds are:
\begin{itemize}
\item Assigning larger workloads to batches to enable efficient batch-level optimizations.
\item Adaptive coordinate precision coupled with visibility-buffer rendering for three times higher brute-force performance.
\item Fine-grained frustum culling on batches of about 10'240 points, directly on the GPU.
\item Support for state-of-the-art level-of-detail structures for point clouds.
\end{itemize}
\section{Method}
The core aspect of our rasterization method is that we assign larger batches of points to be rendered by each workgroup (e.g., 128 threads) instead of only one point per workgroup thread. These larger batches (e.g., 10k points -- 80 per thread) enable several optimizations that would be too costly on small batches (e.g., 128 points -- 1 per thread), but whose costs amortize with each additionally rendered point. Using larger batches with varying amounts of points also allows us to support several widely used level-of-detail structures, as discussed in Section~\ref{sec:lod}. Figure~\ref{fig:pipeline} gives an overview of the rendering steps within a workgroup.
We will first describe our basic rendering pipeline, which we will then gradually expand by additional features and optimizations that ultimately allow us to render point clouds several times faster than the state of the art.
\begin{figure}
\centering
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/flowchart.png}
\caption{Flowchart of a single workgroup invocation.}
\end{subfigure}
\caption{Each workgroup renders one batch or octree node. If its projected bounding box is small, fewer coordinate bits are loaded, reducing memory bandwidth usage and boosting render performance accordingly.}
\label{fig:pipeline}
\end{figure}
\subsection{Data Structure}
The data structure consists of a list of batches and a list of points. Each batch represents a number of consecutive points in the point list (see Figure~\ref{fig:batches_memory}), so for each batch we store the index of the first point of that batch, the number of points in that batch, and their bounding box. Each point consists of 4 attributes: low, medium, and high precision parts of the coordinates (details in Section~\ref{sec:vertex_compression}), and a color value. The 4 attributes are stored in a struct-of-arrays fashion, i.e., in separate buffers for low-precision part of the coordinate, another for medium precision part, etc., so that we only need to load components from memory during rendering that are actually needed. For the majority of points, this will just be the low-precision coordinate bits (4 bytes out of 16 bytes per point).
For regular, unstructured point-cloud data sets, we suggest to simply group about 10'240 consecutive points into batches and compute their bounding box while loading the points. In practice, the majority of data sets we've encountered already provide sufficient locality, especially data sets generated through aerial LIDAR or photogrammetry. But not all of them do and almost all can benefit from sorting by Morton code (z-order) -- an easy to implement and efficient approach to create data sets with good locality~\cite{10.1145/588011.588037, 1532800, 10.1007/3-540-45033-5_3, LOM2020, SCHUETZ-2021-PCC}. For the remainder of this section, we will assume that data sets exhibit sufficient locality, either by default or through sorting by Morton-code, and we will further evaluate and discuss the impact of poor locality in Section~\ref{sec:evaluation}.
Figure~\ref{fig:teaser} and Figure~\ref{fig:batches_boxes_colors} show the results of grouping 10'240 consecutive points of a Morton-ordered point cloud into batches. Although the locality isn't perfect, the majority of batches are sufficiently compact with only a couple of outliers that experience extremely large jumps between different clusters of points. The main advantages of this approach are that it's trivial to implement, requires no preprocessing, and can be done while a data set is loaded with negligible impact on performance.
\begin{figure}
\centering
\hfill
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/batches_unstructured.png}
\caption{Unstructured Point Clouds}
\end{subfigure}
\par\bigskip
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/batches_lod.png}
\caption{LOD Batches / Nodes}
\end{subfigure}
\caption{(a) For unstructured, Morton-code-ordered data sets, we group 10'240 consecutive points into a batch. (b) For LOD data, each octree node equals a batch and batches additionally store memory location and number of points inside the node/batch.}
\label{fig:batches_memory}
\end{figure}
\begin{figure*}
\centering
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_0.jpg}
\caption{Point Cloud Data Set}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_1.jpg}
\caption{Batches}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_2.jpg}
\caption{Bounding Boxes}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth]{images/bunny_3.jpg}
\caption{A single batch}
\end{subfigure}
\caption{A Morton-code ordered point cloud, grouped into batches of 10'240 consecutive points. The locality that is provided by the Morton order is sufficient for frustum culling and a bounding-box-based vertex compression/quantization scheme. }
\label{fig:batches_boxes_colors}
\end{figure*}
\subsection{Basic Rendering Pipeline}
\label{sec:basic}
The basic pipeline spawns one compute workgroup for each batch to render its points. Each workgroup comprises 128 threads and each thread renders $n$ points. In practice, we found $n$ between 60 to 200 to perform equally well on an RTX 3090, and will therefore assume $n = 80$ points per thread (128 * 80 = 10'240 points per batch) throughout the paper. The rasterization process of each individual point is the same as in previous work \cite{SCHUETZ-2021-PCC}: Each point is projected to pixel coordinates, its depth and color value are encoded into a single 64 bit integer, and atomicMin is used to compute the point with the smallest depth value for each pixel, as shown in Listing \ref{lst:basicRasterization}. An early-depth test (a non-atomic depth comparison before an expensive atomic depth-compare and write) ensures that the majority of occluded points are rejected before invoking an expensive atomic update.
\begin{lstlisting}[language=Java,label={lst:basicRasterization},caption={Point rasterization, including an early-depth test.},captionpos=b]
vec4 pos = worldViewProj * position;
int pixelID = toPixelID(pos);
int64_t depth = floatBitsToInt(pos.w);
int64_t newPoint = (depth << 32) | pointIndex;
// fetch previously written point
uint64_t oldPoint = framebuffer[pixelID];
// Early-depth test
if(newPoint < oldPoint){
atomicMin(framebuffer[pixelID], newPoint);
}
\end{lstlisting}
The first improvement over \cite{SCHUETZ-2021-PCC} is that the larger workloads per workgroup allow implementing efficient workgroup-wise frustum culling with a granularity of 10'240 points. At the start of the workgroup invocation, we first load the bounding box of the current batch and abort right away if it does not intersect with the view frustum. Another improvement is that spawning fewer workgroups and giving each of them larger tasks reduces the scheduling overhead of the GPU.
\subsection{Adaptive Vertex Precision}
\label{sec:vertex_compression}
One of the main bottlenecks in previous work \cite{SCHUETZ-2021-PCC} is memory bandwidth usage. They reported a peak performance of 50 billion points per second using 16 bytes per point, which utilizes 85\% of the GPU's memory bandwidth. Since this speed is limited by bandwidth, any significant improvement of brute-force rendering performance requires some form of vertex compression.
We propose an adaptive precision scheme that allows us to load as many bits for coordinates as necessary for a given viewpoint. We achieve this by splitting the bits of a coordinate into three separate buffers storing the low, medium, and high precision parts, as shown in Figure~\ref{fig:memory_layout}. The low precision part always needs to be loaded. The medium precision part can be optionally loaded to recover some of the bits that we removed, and the high precision part is used to optionally recover the remaining bits. The different precision levels are established via coordinate quantization, i.e., by converting coordinates to fixed-precision integers with specific amounts of bits, and then splitting the quantized bits. By quantizing the point coordinates with respect to the bounding box of the batch instead of the bounding box of the whole 3D model, we can achieve a high coordinate precision with few bits. Figure~\ref{fig:quantization} illustrates that the resolution of the coordinates quickly approaches the resolution of the pixel grid. In our case, we first convert the coordinates within a batch to 30 bit fixed-precision integer coordinates that indicate each point's position within its batch. The X-coordinate is computed as follows, for example:
\begin{equation}
X_{30} = min( \lfloor 2^{30} * \frac{x - boxMin.x}{boxSize.x} \rfloor , 2^{30} - 1)
\label{eq:targetIndexPermute}
\end{equation}
These 30 bits are then split into three 10-bit components representing low, medium and high precision parts. The 10 most significant bits (bit indices 20 to 29) of this 30 bit integer are the low-precision bits. They tell us the coordinate of a point within a batch with a precision of $\frac{1}{2^{10}} = \frac{1}{1024}$ of its size. Considering that the majority of rendered batches in any given viewpoint are typically smaller than a couple of hundreds of pixels, and that 10 bits grant us 1024 different coordinate values, we find that 10 bits per axis are sufficient to render most points with sub-pixel coordinate precision. Due to buffer alignment recommendations, simplicity, and because the smallest accessible primitive values on GPUs are typically 32 bits, we then pack the 10-bit x, y and z components into a single 32 bit integer with the remaining 2 bits used as padding, as shown in Listing~\ref{lst:encoding}. The result is a 32-bit integer buffer where each 4-byte element contains the 10 lowest precision bits of the three coordinate axes of a single point.
\lstset{
mathescape,
literate={->}{$\rightarrow$}{2}
{ε}{$\varepsilon$}{1}
}
\begin{lstlisting}[language=Java,label={lst:encoding},caption={Encoding the most significant 10 bits of each axis into a single 32 bit integer. },captionpos=b]
uint32_t X$_{low}$ = (X$_{30}$ >> 20) & 1023;
uint32_t Y$_{low}$ = (Y$_{30}$ >> 20) & 1023;
uint32_t Z$_{low}$ = (Z$_{30}$ >> 20) & 1023;
uint32_t encoded = X$_{low}$ | (Y$_{low}$ << 10) | (Z$_{low}$ << 20)
\end{lstlisting}
Likewise, we generate two more buffers for the medium (bits 10 to 19, middle) and high precision bits (bits 0 to 9, least-significant). During rendering, each workgroup can now choose to load a single 4-byte integer containing just the low precision bits that grants 1'024 coordinate values, 2 such integers (low+medium precision) granting 1'048'576 coordinate values, or 3 such integers (low+medium+high precision), which theoretically allows for one billion different coordinate values. However, we should note that the usefulness of 30 bit precision is limited in practice because we still convert the integer coordinates to floating-point coordinates during rendering, where the conversion from a 30-bit fixed-precision integer to a 32-bit floating point value will incur a loss of precision due to rounding errors and the non-uniform distribution of precision in floats (high precision for small values, low precision for large values). This issue is not specific to our approach, however, as point clouds are typically already stored in integer coordinates on disk and the conversion to floats for rendering has always been an issue for point clouds with a large extent. Handling this issue is out of scope of this paper, but we would like to note that storing coordinates as 30 bit fixed-precision integers would allow us to convert them to highly accurate double-precision coordinates (e.g., for measurements), while traditional floating point storage causes an irrecoverable loss of precision.
\begin{figure}
\centering
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_0.jpg}
\caption{2 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_1.jpg}
\caption{3 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_2.jpg}
\caption{4 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_3.jpg}
\caption{5 bits}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.19\columnwidth}
\includegraphics[width=\textwidth]{images/quantization_4.jpg}
\caption{6 bits}
\end{subfigure}
\caption{Coordinate quantization provides a trade-off between precision and storage size. The amount of possible coordinate values per axis is given by $2^{bits}$.}
\label{fig:quantization}
\end{figure}
The coordinate precision during rendering is determined through the projected size of a batch. If a batch projects to less than 500 pixels, we use 10 bit coordinates. The reason for choosing 500 pixels as the threshold even though 10 bits can represent 1024 different coordinates values is that quantization changes the distance between any two points by up to the size of a quantization grid cell, i.e., points that were one grid cell's size apart might now be twice as far apart. Using half the size of the quantization grid as the pixel threshold ensures that we do not introduce additional holes between rasterized points.
Compared to previous work~\cite{SCHUETZ-2021-PCC}, this approach reduces the required memory bandwidth for the majority of rendered points from 16 bytes down to 8, since only 4 bytes instead of 12 are needed for most coordinates, and another 4 for colors. However, we can further cut this into half by using a visibility-buffer (item-buffer) approach~\cite{Burns2013Visibility, 10.1145/357332.357335}, i.e., we render point indices rather than colors into the framebuffer during the geometry pass, and transform the point indices to vertex colors in a post-processing step. Doing so reduces the amount of memory accesses to color values down from the number of processed points to the number of visible points.
\begin{figure}
\centering
\hfill
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/points_aos.png}
\caption{Common point-cloud memory layout with 16 bytes per point.}
\end{subfigure}
\hfill
\par\bigskip
\begin{subfigure}[t]{\columnwidth}
\includegraphics[width=\textwidth]{images/points_soa.png}
\caption{Splitting coordinate bits into a low, medium and high-precision buffer. Each buffer stores 10 bits per coordinate axis, encoded into 4 bytes per point. }
\end{subfigure}
\caption{(a) Point cloud renderers typically load at least 16 bytes per point during rendering. (b) Splitting coordinate bits into separate buffers allows loading just the required bits, depending on the viewpoint. 10-bit coordinates typically suffice for the majority of points. }
\label{fig:memory_layout}
\end{figure}
\subsection{Vertex Pre-Fetching}
\label{sec:prefetch}
To identify performance bottlenecks of the above approach, we performed a direct port of its GLSL shader code to NVIDIA CUDA.
Doing so enabled the use of the proprietary Nsight Compute tool to pinpoint suboptimal behavior in our routines.
Our investigation into dominant stall reasons revealed that performance is---unsurprisingly---governed by memory operations, with approximately $70\%$ of the total kernel run time spent on them.
Since we choose the batch size to be a multiple of the work group size, threads persist and process multiple points in a loop before returning.
Naively, the corresponding point data is fetched and rasterized to the framebuffer in each iteration.
Batches simply define a linear range in memory, thus their contents can be transferred by workgroups with perfect coalescence.
Furthermore, the compression scheme outlined in Section \ref{sec:vertex_compression} economizes on available bandwidth when fetching point data.
However, loading each processed point individually at the start of its associated iteration still incurs a direct dependency of the steps in Listing \ref{lst:basicRasterization} on global device memory latency.
Due to the early-depth test in its body and the parameterizable iteration count, the compiler cannot trivially unroll the loop without creating secondary issues associated with complex program flow (e.g., frequent instruction cache misses).
Hence, we perform manual pre-fetching of point data. For the lowest precision (using 32 bits per point), we execute a vectorized 128-bit load in each thread, yielding enough data for four successive iterations; for medium-precision points, a 128-bit load fetches the data for two iterations.
This policy significantly alleviates reported stalls due to memory latency and simultaneously reduces the number of memory instructions, without compromising on coalesced accesses.
Overall, we found pre-fetching to cause $\approx 30\%$ performance gain.
A welcome side effect, though less impactful, is the reduction of updates to the frame buffer by this policy, if points are spatially sorted. Since threads now load and process consecutive points in memory, the early-depth test has a higher chance of failing: threads are now more likely to query (and find) information in the L1 cache for pixels that they wrote to in a previous iteration.
We note that, with pre-fetching in place, the main remaining bottleneck is the code block for early-depth testing and framebuffer updates. Although these accesses are somewhat localized if points are spatially ordered, a residual irregularity remains in their access pattern.
The coarse-grained memory transfer policy of the GPU consequently causes more than $2\times$ the amount of actually accessed information to be transferred.
However, this update pattern is inherent to the overall routine design, and any further attempts to improve on it without extensive revision led to diminishing returns.
\subsection{Level-of-Detail Rendering}
\label{sec:lod}
In this section we will describe how support for some of the most popular LOD structures for point clouds -- Layered Point Clouds (LPC)~\cite{LPC} and its variations -- can be added to our point rasterization approach. Layered point clouds are a hierarchical, spatial acceleration structure where points with varying density are stored in the nodes of the tree. Lower levels contain coarse low-density subsets of the whole point cloud. With each level, the size of the nodes decreases, leading to an increase of the the density of points as the number of points in each node stays roughly the same. The structures are often additive -- meaning that higher LODs increase detail by rendering more points in addition to lower LODs instead of replacing them -- but replacing schemes are also possible.
We implement and evaluate our support for LPC structures based on the Potree format, which constitutes a variation of LPC that uses an octree and populates nodes with subsamples of the point cloud with a specific, level-dependant minimum distance. Each octree node comprises about 1k to 50k points, and Potree itself typically renders about 100 to 1000 nodes for any given viewpoint. Up until now we have assumed that each batch renders exactly 10'240 points, but we can easily adapt our rasterization approach to support the Potree structure by allowing varying amounts of points per batch, as shown in Figure~\ref{fig:batches_memory}. In addition to the bounding box, we will also need to store the number of points as well as the memory location of the first point for each batch. The workgroup size of 128 threads remains the same, but each thread will now render $\lceil \frac{batch.numPoints}{128} \rceil$ points instead of exactly 80.
The Potree format is structured such that octree nodes whose projected bounding boxes are small can be entirely ignored, because the points stored in their parents will already provide sufficient detail. This means that the procedure in Section~\ref{sec:vertex_compression} that is used to determine the coordinate precision during rendering based on the screen-size of the bounding box can now be used to entirely cull the node. We suggest to cull the nodes with following conditions in mind: Each Potree node is cubic and stores a subsample of points with a resolution that approximately matches a grid with $128^3$ cells, and our rasterizer maps each point to exactly 1 pixel. To avoid holes between rendered points, we therefore suggest to cull those nodes that are smaller than 100 pixels on screen. However, it is also viable to cull larger octree nodes on lower end GPUs to improve performance and cover up the resulting holes in a post-processing pass, e.g., via depth-aware dilation or more sophisticated hole-filling approaches~\cite{ruckert2021adop, AutoSplats, Rosenthal2008, grossman1998point, Pintus2011, Marroquim:2007:pbg}.
\subsection{High-Quality Shading}
LOD rendering works well in conjunction with high-quality splatting (point-sprites or surfels~\cite{Botsch:HQS}) or shading (one-pixel points~\cite{SCHUETZ-2021-PCC}), a form of color filtering for point clouds that blends overlapping points together. Colors and amount of points inside a pixel within a certain depth range (e.g., 1\% behind the closest point) are summed up during the geometry processing stage, and a post-processing shader then divides the sum of colors by the counters to compute the average. Schütz et al. \cite{SCHUETZ-2021-PCC} suggest two variations: one that uses two 64-bit atomicAdd instructions per point into four 32 bit integers to sum up color values and counters, and another variation that uses a single 64 bit atomic instruction per point to compute the sum of up to 255 points, with a fallback that uses two 64-bit atomics if more than 255 points contribute to the average. However, when using an LOD structure, the amount of overlapping points with similar depth is essentially guaranteed to be lower than 255, so we can safely use high-quality shading with just a single 64-bit atomic add instruction per point. This limit can even be raised to up to 1023 points by using the ``non-robust" variation without overflow protection. For unstructured data sets, we suggest to use the variation with overflow protection due to the potentially large amount of overlaps.
\subsection{Virtual Reality Rendering}
Virtual reality (stereo) rendering greatly increases the performance requirements -- even more so for point clouds as they typically suffer from aliasing artifacts that need to be addressed to provide an acceptable VR experience. In addition to a specific configuration of our pipeline (use LOD rendering for performance, high-quality shading to reduce aliasing and large framebuffers for additional anti-aliasing via supersampling), we can exploit specific properties of VR rendering in our approach.
First, the scene needs to be drawn twice -- once for each eye. Instead of duplicating the entire rasterization pipeline by calling the compute shaders twice, we suggest to modify the shader to simply draw each point into both framebuffers. While this doesn't double the performance, it provides a significant improvement, as discussed in Section~\ref{sec:evaluation}.
Second, in a VR setup, details in the periphery aren't as noticeable as details in the center of the view. This is partially because most details are perceived in the gaze direction, i.e., mostly straight ahead in VR, but also because the rendered image will be distorted before it is shown inside the HMD to counter the lens distortion~\cite{Vlachos:2015:AVR}. The applied distortion compresses peripheral regions, thus reducing the amount of detail in the image. We therefore suggest to render peripheral regions of the framebuffer with reduced geometric complexity by raising the threshold for LOD culling, e.g., culling nodes smaller than 300 pixels in the periphery, nodes smaller than 100 pixels in the center of the view, and 200 pixels in between. The resulting holes between points are then filled in post-processing, in our case via a simple depth-aware dilation shader that increases point sizes to up to 7x7 in the periphery and 3x3 in the center. Depth-aware means that the closest point within the pixel ranges are expanded.
\section{Related Work}
\subsection{Software Rasterization of Triangle Meshes}
Early CPU-side solutions for triangle rasterization in software were largely made obsolete in the 2000s by GPUs and their high-performance rasterization components.
However, their (partly hardwired) pipeline lacks the flexibility to perform custom hierarchical or context-dependent optimizations between individual stages.
The continuously advancing programmability of GPUs has given software rasterization its second wind:
\emph{Freepipe} demonstrated that for scenes containing many, small triangles, GPU software rasterization with one thread per triangle can outperform the hardware pipeline \cite{10.1145/1730804.1730817}.
CudaRaster and Piko expanded on this idea, introducing optimizations for hierarchical triangle rasterization, achieving competitive performance with hardware rasterization even for larger triangles \cite{10.1145/2018323.2018337, 10.1145/2766973}.
Complete software implementations of OpenGL-style streaming pipelines, including sort-middle binning and hierarchical rasterization, have been presented for NVIDIA CUDA and OpenCL \cite{Kenzel:2018, Kim2021}: rather than attempting to outperform hardware rasterization, these solutions aim to provide an environment for experimenting with extensions and optimizations that may be suitable for future hardware pipelines.
A comprehensive analysis of previous software rasterization approaches and the challenges they tackle is found in \cite{frolov20comparative}.
Most recently, software rasterization has received increased attention due to the launch of the Unreal Engine 5 and its virtual geometry feature, \emph{Nanite} \cite{karis2021nanite}.
\emph{Nanite} provides both a hardware and a software pipeline for rasterization geometry and selects the proper route for rendered geometry dynamically.
In scenes with mostly pixel-sized triangles, their software pipeline reportedly achieves more than $3\times$ speedup. Its striking success begs the question whether high-performance software rasterization has not been overlooked as a viable method for other 3D representations as well.
\subsection{Software Raserization of Point Clouds}
Günther et al. proposed a GPGPU-based approach that renders points up to an order of magnitude faster than native OpenGL points primitives -- a feat that is possible because the fixed-function rendering pipeline is mainly targeted towards triangles~\cite{Gnther2013AGP}. Whenever a point modifies a pixel, their busy-loop approach locks that pixel, updates depth and color buffers, and then unlocks the pixel. Marrs et al. use atomic min/max to reproject a depth buffer to different views~\cite{Marrs2018Shadows}. Since only depth values are needed, 32 bit atomic operations are sufficient. Schütz et al. render colored point clouds via 64 bit atomic min-max operations by encoding depth and color values of each point into 64 bit integers, and using atomicMin to compute the points with the lowest depth value for each pixel inside an interleaved depth and color buffer~\cite{SCHUETZ-2021-PCC}. Our paper is based on this approach and the published source code\footnote{\url{https://github.com/m-schuetz/compute_rasterizer/releases/tag/compute_rasterizer_2021}}, and makes it several times faster for brute-forcing, but also adds support for frustum culling and LOD rendering. Rückert et al. also adapt this approach into a differentiable novel-view synthesis renderer that uses one-pixel point rendering to draw multiple resolutions of a point cloud, and neural networks to fill holes and refine the results~\cite{ruckert2021adop}. They also significantly reduce overdraw by stochastically discarding points whose assumed world-space size is considerably smaller than the pixel they are projected to.
\subsection{Level-of-Detail for Point Clouds}
Rusinkiewicz and Levoy introduced QSplat, a point-based level-of-detail data structure, as a means to interactively render large meshes~\cite{QSplat}. They use a bounding-sphere hierarchy that is traversed until a sphere with a sufficiently small radius is encountered, which is then drawn on screen. Sequential Point Trees are a more GPU-friendly approach that sequentializes a hierachical point-based representation of the model into a non-hierarchical list of points, sorted by their level of detail. From a distance, only a small continuous subset representing a lower level of detail needs to be rendered, without the need for costly traversal through a dense hierarchical structure~\cite{Dachsbacher2003}.
Layered point clouds~\cite{LPC} were one of the most impactful improvements to LOD rendering of point clouds, and variations of it still serve as today's state of the art. The original LPC constitutes a binary tree that splits the 3D space in half at each level of the hierachy. The key difference to the bounding-sphere hierarchy of QSPLATs is that each node itself is not a sphere, but a smaller point cloud comprising thousands of randomly selected points of the full point cloud. The large amount of geometry in each node reduces the amount of nodes that are required to store the full data set, which allows utilizing GPUs that are efficient at rendering hundreds of batches comprising thousands of points, each. Further research improved several aspects of layered point clouds, such as the utilized tree-structure, LOD generation times, and using density-based subsampling to properly support data sets without uniform density~\cite{Wand2008, Goswami2010, scheiblauer2011, Elseberg2013, MartinezRubi2015, Kang2019, Bormann:PCI}. Section~\ref{sec:lod} describes how our rasterization approach supports layered point clouds, using the \cite{Potree} structure as a specific example.
\subsection{Coordinate Quantization}
Quantization is the conversion of a continuous signal to discrete samples. In case of coordinates, quantization typically refers to the conversion of floating point or double coordinates (the pseudo-continuous signal) to fixed-precision integer coordinates with a carefully chosen amount of bits and a uniform coordinate precision over the whole range. The uniform precision and control over the supported coordinate range, precision, and amount of bits makes quantization a frequently used part of coordinate compression schemes, complemented by delta and entropy encoding ~\cite{deering1995geometry, ISENBURG:LAZ} or hierarchical encoding~\cite{10.5555/581896.581904}. In this paper, we use quantization to encode coordinate bits such that they can be adaptively loaded, i.e., to load just a few bits if low precision is sufficient, or optionally load additional bits to refine the previously loaded low-precision coordinate. We do not, however, apply delta, entropy, hierarchical or similar encodings because they add additional computational overhead, and they make the decoding of one point dependant on the decoding of others, i.e., they can't be decoded individually.
|
\section{Introduction}
\label{secI:introduction}
\IEEEPARstart{S}{Semantic} segmentation is one of the most challenging problems in computer vision. It aims to categorize each pixel of an image into a particular semantic class. Accurate semantic segmentation can be applied to wide-ranging applications in real-world scenarios such as autonomous driving \cite{janai2020computer}, video surveillance, and robot sensing \cite{9023474}. Driven by the recent advancement of deep convolutional neural networks (DCNNs), the pixel-wise semantic segmentation frameworks have witnessed unprecedented progress \cite{long2015fully,chen2018encoder,ronneberger2015u}. Existing CNNs based methods are usually constructed of stacked convolutional and downsampling layers. The shallower layers consist of spatial resolution feature maps, while the deeper stages capture the global context features.\\
Inspired by these observations, many new methods have followed fully convolutional network (FCN) \cite{long2015fully} based structures to perform pixel-wise semantic segmentation tasks \cite{pohlen2017full,ronneberger2015u,badrinarayanan2017segnet}. The encoder-decoder models \cite{badrinarayanan2017segnet,chen2018encoder} first capture the high-level semantic information by employing the original FCN and gradual upsampling to retrieve the original spatial information. Meanwhile, skip connections compensate for the fine information loss in the downsampling process. However, U-shape frameworks have achieved notable progress and received much attention in many computer vision applications \cite{pohlen2017full,badrinarayanan2017segnet}, but still less effective in extracting sufficient contextual information due to the following issues. \\
The first issue is dealing with objects that exist at multiple scales and have complex structures. A typical CNNs \cite{long2015fully} have a weak ability to extract sufficient multi-scale information for accurate objects prediction. Another issue is that in a typical encoder-decoder framework (\ref{fig:networks_comparison}.\subref{fig1:encoder_decoder}), the encoder encodes the global context information in the deeper stages. The encoded information may progressively be diluted and cannot be fully recovered. Also, architectures based on UNet use skip-connections to aggregate features of the same scale by connecting encoder-decoder layers in parallel manners, but this simple implementation of skip-connections neglect global information, which results in pixel labels misclassification.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{accuracy_speed_spfnet.png}
\caption{Accuracy-Speed performance comparison on the Cityscapes test set.Our methods are presented in red dots while other methods are presented in blue dots. Our approaches achieve state-of-the-art speed-accuracy trade-off}
\label{fig:accuracy_speed}
\end{figure}
In recent years, methods with larger receptive fields have been proposed to solve these problems \cite{zhao2017pyramid,chen2017deeplab} (\ref{fig:networks_comparison}.\subref{fig2:atrous_convolution}). Semantic segmentation performance can be improved significantly by utilizing multi-scale context information. However, how to integrate such information effectively is still an open research question.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=5.8cm,height=6cm]{Encoder_decoder.png}
\caption{Encoder-Decoder Structure}
\label{fig1:encoder_decoder}
\end{subfigure} \hspace{1mm}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=5.8cm,height=6cm]{Encoder_decoder_atrous.png}
\caption{Encoder-Decoder with Atrous Conv}
\label{fig2:atrous_convolution}
\end{subfigure}\hspace{1mm}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=5.8cm,height=6cm]{Encoder_decoder_spfm.png}
\caption{Encoder-Decoder with SPFM}
\label{fig3:fpn}
\end{subfigure}\hspace{1mm}
\caption{A comparison of an important semantic segmentation architectures. (a) encoder-decoder structure.(b) encoder-decoder structure with atrous convolutions.(c) our network, encoder-decoder structure with atrous convolutions and context aggregation module.}
\label{fig:networks_comparison}
\end{figure*}
Multi-scale information is the most crucial problem that should be considered to improve the performance of semantic segmentation networks. Typically, a class of targets may exist at different scales in the image, and a well-designed model should be able to extract this property, which leads to better semantic segmentation accuracy. In this paper, we propose a Subspace Pyramid Fusion Module (SPFM) to improve the multi-scale feature learning in semantic segmentation. In particular, the SPFM generates global contextual information in each feature subspace based on parallel Reduced Pyramid Pooling submodule (RPP). RPP submodule constructed from three components: dilated convolution or atrous convolution, subpixel convolution, and average pooling. Two dilated convolutions were used for each reduced pyramid pooling level to capture object features with various receptive fields, followed by subpixel convolution instead of the original upsampling method to unify the feature resolution. To address the problems of resolution loss via downsampling in the encoding part, we develop the efficient global context aggregation module (EGCA) with two branches. Through the lower branch, i.e., EGCA integrates the shuffle attention block to facilitate cross-channel information communication for feature maps from different encoder levels and reduces channel redundancy. Finally, a Decoder-based Subpixel Convolution (DSC) is proposed to improve category pixel localization and recover object details. The SPFM module is embedded in the center between the encoder and the decoder to help select the suitable receptive field for objects with different scales by fusing the multi-scale context information. Based on these descriptions, we design the Subspace Pyramid Fusion Network (SPFNet), see Figure\ref{spfnet}. The proposed SPFNet is validated with two challenging semantic segmentation benchmarks, Camvid \cite{brostow2009semantic} and Cityscapes \cite{cordts2016cityscapes}. Figure. \ref{fig:accuracy_speed} shows the accuracy and the speed comparison of different methods on Cityscapes dataset.
Our main contributions are summarized as follows:\\
\begin{enumerate}
\item A novel Subspace Pyramid Fusion (SPFM) module is proposed to learn fused multi-scale and global context information for each feature subspace.
\item We introduce the Efficient Global Context Aggregation (EGCA) module, which utilizes channel shuffle operation to integrate the complementary global context
\item We propose decoder based on subpixel convolution to retrieve the high-resolution feature maps.
\item We proposed an architecture based on SPFM and EGCA modules that can be easily applied for road scene understanding. Furthermore, these modules can be plugged in any network
\item Extensive experimental results on Camvid and Cityscapes semantic segmentation benchmarks show that the proposed method has good generalization ability.
\end{enumerate}
\begin{figure*}
\centering
\includegraphics[width=0.7\linewidth]{SPFNet.png}
\caption{An overview of the proposed Subspace Pyramid Fusion Network (SPFNet).}
\label{spfnet}
\end{figure*}
\section{Related Work}
\subsection{Semantic Segmentation}
Semantic segmentation has been widely studied in the past few years, and convolutional neural networks are at the center of this progress. Most available semantic segmentation architectures are motivated by the seminal fully convolutional network \cite{long2015fully} or U-Net \cite{ronneberger2015u}. FCN was the first architecture to adopt standard classification CNNs for dense pixel prediction by replacing the fully connected layers with fully convolutional layers. In FCN, to retrieve the low-level representations of the input image, the output spatial map is upsampled in a single step using deconvolution \cite{zeiler2011adaptive, zeiler2014visualizing}. Further studies have utilized skip-connections which combine the semantic information from coarse layers with context information from shallow layers. On the other hand, U-Net comprises two parts. A contracting path is created using convolution layers with pooling to extract context information and expanding path to recover the original resolution. Skip-connections are used to propagate the features from contracting to expanding path layers. Various variants of these networks have been developed to solve different pixel-wise segmentation tasks in a wide range of applications \cite{zhou2018unet++,badrinarayanan2017segnet}
\subsection{Multi-scale and Context Aggregation }
\label{subsec:lit_rev_multi_scale}
Multi-scale feature fusion is given great attention by the deep learning research community \cite{chen2017deeplab,fourure2017residual,chen2017deeplab,chen2016attention} and has proven to be powerful in enhancing semantic segmentation performance. The straightforward way is to integrate low-level and high-level features to extract patterns of different granularity \cite{badrinarayanan2017segnet,ronneberger2015u}. DeepLab \cite{chen2017deeplab} applied atrous Spatial Pyramid Pooling (ASPP) in which a dilated convolution is used to increase the receptive field while maintaining the feature map resolution, leading to global context aggregation. PSPNet accomplished a similar goal by the spatial pyramid pooling \cite{zhao2017pyramid}. In PAN \cite{li2018pyramid} a feature pyramid attention is introduced to extract context information; in this approach, convolutions with large kernel sizes are used instead of using atrous convolution to construct the pyramid. Feature Pyramid Network \cite{lin2017feature} aggregates the multi-scale feature maps in a top-down fashion with progressive upsampling. Other network such as BiSeNet \cite{yu2021bisenet}, ContextNet \cite{poudel2018contextnet}, GUN \cite{mazzini2018guided}, and DSANet \cite{elhassan2021dsanet} utilized detail branch to capture low-level details in shallow layers.
\subsection{Attention Mechanism }
Self-attention mechanism has been extensively studied in the past few years, first in the domain of machine translation \cite{vaswani2017attention}. Recently, computer vision related tasks such as classification, detection, and segmentation have integrated different forms of attention mechanism \cite{yin2020disentangled,zhang2021sa}. Some works utilize attention to update features with learned weights. For instance, SENet \cite{hu2018squeeze} highlights the channel maps through global pooling features. In \cite{zhang2019deep} an attention mechanism is adopted to highlight the important region of interest at multiple spatial scales. CBAM \cite{woo2018cbam} emphasizes the important regions, it expands the \cite{hu2018squeeze} to spatial dimension. SKNet \cite{li2019selective} further proposed a dynamic kernel selection mechanism with small parameters overhead to improve the classification performance DANet \cite{fu2019dual} proposed adaptive method to integrate local and global features for semantic inter-dependencies modeling in spatial and channel dimensions. \cite{fu2020scene} proposed an attention mechanism that guides the feature update and helps to enrich relevant features adaptively. \cite{li2019spatial} combines channel attention and spatial attention consecutively. \cite{wang2020eca} adopted a 1-D convolution filter to avoid dimensionality reduction and increase cross-channel interaction, which reduced SENet model complexity. In this work, we constructed an attention mechanism based on channel shuffle operation to capture discriminative multi-level features. \cite{saini2020ulsam} introduced a lightweight attention module to improve the efficiency of compact CNNs architectures. The proposed module can be plugged into a classification network such as MobileNet to help learn multi-frequency and multi-scale feature.
\section{Methodology}
In this section, we first present an overview of the SPFNet, and then elaborate the mechanism of different modules that used to construct this network.
\subsection{Overview}
As illustrated in Figure \ref{spfnet}, the SPFNet is constructed under a general encoder-decoder framework, where the encoder downsamples the input data to learn the rich multi-scale features and the decoder performs the reconstruction of the high-level features from the low-level features for pixel-level semantic prediction. The SPFNet first takes an input image and encodes its feature maps with a general multi-scale backbone, i.e., the ResNet34 \cite{he2016deep}. The SPFNet utilizes $[F_{1}^{b},F_{2}^{b},F_{3}^{b},F_{4}^{b},F_{5}^{b}]$ to denote feature hierarchy from the backbone (encoder), b refers to backbone. Note that the feature maps have a stride of $[2,4,8,16,32]$. Unlike other methods such as \cite{chen2018encoder}, which replace stride with atrous convolution in the feature extraction part, we have set a stride of 1 for $F_{2}^{b}$ to train SPFNet34H. Inspired by the discussion in the second problem, we further propose Subspace Pyramid Fusion Module (SPFM) (subsection. \ref{sec:subspace_pyramid}), which generates multi-scale global context information. $[F_{2}^{egca},F_{3}^{egca},F_{4}^{egca},F_{5}^{egca}]$ are the rich context information that extracted from the feature hierarchy in the Efficient Global Context Aggregation (EGCA) module (section. \ref{sec:shuffle_attention}). $[F_{1}^{d},F_{2}^{d},F_{3}^{d},F_{4}^{d}]$ are the feature from Decoder-based Subpixel Convolution (DSC) (section. \ref{sec:decoder_sub_pixel}), in which a gradual features upsampling and concatenation are performed to integrate the low-level and high-level feature representations.
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=8.2cm,height=5.4cm]{SPFM2.png}
\caption{Subspace pyramid fusion module (SPFM).}
\label{subfig:spfm}
\end{subfigure} \hspace{1.5mm}
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=8cm,height=5.4cm]{RPP.png}
\caption{Reduced pyramid pooling module (RPP).}
\label{subfig:rpp}
\end{subfigure}\hspace{1.5mm}
\caption{Illustration of the (a) Reduced pyramid pooling module (RPP). (b) Subspace Pyramid Fusion (SPFM) module. SPFM extracts the multi-scale information in each sub-features using RPP module, then transmit the information to decoder.}
\label{fig:sub_pyramid}
\end{figure*}
\subsection{Subspace Pyramid Fusion Module (SPFM)}
\label{sec:subspace_pyramid}
It has been demonstrated in semantic segmentation literature \cite{chen2017deeplab, gu2019net,li2018pyramid} that increasing the receptive fields is beneficial to the semantic segmentation task. In a typical encoder-decoder architecture, the encoder extracts global context information from the input image, including the adjacent and class characteristics of the object. However, transmitting information to shallower layers will weaken the extraction of the context information due to the downsampling processes. In this subsection, motivated by this observation and inspired by \cite{chen2017deeplab} and \cite{saini2020ulsam}, we proposed a novel subspace pyramid fusion module. For each subspace feature in SPFM, we use only one reduced pyramid pooling (RPP). Unlike \cite{chen2017deeplab} and \cite{xie2018vortex}, SPFM (as shown in Figure. \ref{fig:sub_pyramid}) collects multi-scale contextual information differently. Instead of fusing the dilated convolution directly, SPFM first generates multi-scale features for each split using reduced pyramid pooling module. RPP constructed with only two parallel dilated convolutions, with dilation rates of (4,8). Next, a pixel shuffle is employed to increase the feature resolution. In addition, adaptive average pooling is also fused with the output from two parallel convolutions to capture complex context information. We then concatenate all the RPP to construct the SPFM. Compared with SPP and ASPP, our module can acquire more information using the subpixel convolution and the stacked RPP blocks. Unlike pyramid-based multi-scale learning such as SPP and ASPP, our module SPFM acquires more multi-scale information by learning multiple Reduced Pyramid Pooling blocks (RPP) for each feature map. Then, it combines these RPP blocks in one Subspace Pyramid Fusion Module. The main difference between the proposed SPFM and \cite{saini2020ulsam} is the way we aggregate information. \cite{saini2020ulsam} integrate attention across different stages in a pre-existing compact backbones to help the model learn the global information. In contrast, SPFM increases the receptive fields with dilated convolutions and adaptive average pooling to capture global context.
Given the feature map $F_{5}^{b}\in \mathbb{R}^{H\times W\times C}$ from the backbone, where H and W are the spatial dimensions of the feature maps and C is the number of channels. Our objective is to learn to capture multi-scale features efficiently. In Figure \ref{fig:sub_pyramid}.\subref{subfig:spfm}, we present the proposed SPFM module and its submodule RPP in Figure. \ref{fig:sub_pyramid}.\subref{subfig:rpp}. SPFM divides the input feature map $F_{5}^{b}$ into s splits along the channel dimension, i.e., $F_{5}^{b} ={\{F_{i}^{p}\}}_{i=1}^{s}$, where $F_{i}^{p}$ is input feature maps for a single Reduced Pyramid Pooling. Each subspace has $F_{i}^{p}\in \mathbb{R}^{H\times W\times G}$ feature maps, where $G = \frac{C}{s}$.
Here are the details of SPFM and RPP modules.\\
\textbf{Reduced Pyramid Pooling (RPP)}: Each RPP block takes one split from SPFM. First, the input features $F_{i}^{p}$ enter into two identical pathways, each path consists of dilated convolution with kernel size of 3 followed by batch normalization(BN) and Parametric Rectified Linear Unit (PReLU) to generate multi-scale context information from these feature maps. Subsequently, we apply pixel shuffle to obtain $P_{i,1}^{p}$ as in Eq. \ref{p1} and $P_{i,2}^{p}$ Eq. \ref{p2}. Then, we concatenate the output features $P_{i,c}^{p}$ Eq. \ref{spconc1} and refined it with three consecutive convolutions with kernel sizes ($1\times1$), ($3\times3$), and ($1\times1$) respectively, to reduce aliasing effect. Finally, the output is added with feature map of adaptive average pooling and convolution Eq. \ref{avgp} to produce the final reduced pyramid pooling $P_{\tilde{s},i}^{p}$ as shown in Eq \ref{rpp}.
\begin{equation}\label{p1}
P_{i,1}^{p} = \mathbb{PS}(\mathbb{D}_{conv}@4(F_{i}^{p}))
\end{equation}
\begin{equation}\label{p2}
P_{i,2}^{p} = \mathbb{PS}(\mathbb{D}_{conv}@8(F_{i}^{p}))
\end{equation}
\begin{equation}\label{spconc1}
P_{i,c}^{p} = Concat([P_{i,}^{p},P_{i,2}^{p}])
\end{equation}
\begin{equation}\label{avgp}
A_{i}^{p} = f^{1}(APool(F_{i}^{p}))
\end{equation}
where Concat([.]) is the element-wise concatenation.
The reduced pyramid fusion for each split $P_{\tilde{s},i}^{p}$ is computed by the following equation.
\begin{equation}\label{rpp
P_{\tilde{s},i}^{p} = A_{i}^{p} \oplus P_{i,c}^{p}
\end{equation}
In Eq. \ref{p1}, and Eq. \ref{p2}, $\mathbb{D}_{conv}@$ refers to dilated convolution layer with kernel size of $3\times3$. In Eq\ref{avgp} APool(.), is the adaptive average pooling, $f^{1}$ represents convolution layer with kernel size of $1\times1$. In Eq. \ref{p1} and Eq. \ref{p2}, $\mathbb{PS}(.)$ is the PixelShuffle with upsample scale factor of 2.\\
\textbf{Subspace Pyramid Fusion Module (SPFM)}: The final output for SPFM module $\hat{P}_{out}$ is obtained by concatenating RPP of all splits (Eq. \ref{pf}).
\begin{equation}\label{pf}
\hat{P}_{out} = f_{1}( Cancat([\hat{P}_{\tilde{s},1},\hat{P}_{\tilde{s},2},...,\hat{P}_{\tilde{s},s}]))
\end{equation}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth]{global_context.png}
\caption{An overview of the proposed Global Context Aggregation module.
It utilize channel split/shuffle to process features in each group. }
\label{Figure:egca}
\end{figure*}
\subsection{Efficient Global Context Aggregation}
\label{sec:shuffle_attention}
Local and global context information are known to be useful for semantic segmentation and other computer vision tasks in even traditional machine learning era \cite{arbelaez2010contour}. Deep learning methods have demonstrated astounding performance when integrating multi-scale features \cite{chen2016attention}. In this work, we propose an efficient global context aggregation module (Figure. \ref{Figure:egca}) to make use of both local and global context. EGCA integrates channel attention through the shuffle unit operator to increase cross-feature interaction between all sub-features. In this setting, the input features from backbone are denoted as $F_{j}^{b}$, where j indicates the level in encoder (Figure. \ref{spfnet}). Since the features in different levels have different semantic, we employ the EGCA before upsampling them to a common resolution using bilinear interpolation. Then features from all encoder scales and the last decoder-based subpixel convolution outputs are concatenated to from multi-scale feature map. Thus, fuse information from shallow layers and deeper layers.
Given the $j^{th}$ feature $F_{j}^{b}\in \mathbb{R}^{H\times W\times C}$ from the backbone, where H and W are spatial dimensions and $C$ is the channel dimension. The output $F_{j}^{egca}$ of efficient global context aggregation module is computed using the following equation:\\
\begin{equation}\label{eq:egca}
{F}_{j}^{egca} = \widehat{F}_{j} \oplus F_{j}^{u}
\end{equation}
Where $\oplus$ represents the element-wise summation, $\widehat{F}_{j}$ is the output of shuffle attention part, and $F_{j}^{u}$ represents the output feature of the upper part. At first, the input feature $F_{j}^{b}$ is used to process information into upper branch and the shuffle attention branch (see Figure. \ref{Figure:egca}). More details are given as follow:\\
\textbf{Part I}: Upper branch\\
in this branch $F_{j}^{b}$ is passed through depthwise separable convolutions, leading to feature map of $F_{j}^{u}\in \mathbb{R}^{H\times W\times C}$ as shown in Eq. \ref{fu_upper}.
\begin{equation}\label{fu_upper}
F_{j}^{u} = f_{DW}^{3}(f^{1}(F_{j}^{b}))
\end{equation}
where $f_{DW}^{3}(.)$ represents depthwise convolution with kernel $3\times3$.\\
\textbf{Part II}: Lower branch- shuffle attention\\
The second branch uses shuffle channel attention to generate sub-features that gradually capture multi-scale semantic responses. In this branch, the input $F_{j}^{b}$ is divided into two groups $F_{j,1}\in \mathbb{R}^{H\times W\times C/2}$, and $F_{j,2}\in \mathbb{R}^{H\times W\times C/2}$ along the channel dimensions.
\begin{equation}\label{sa1}
A_{j,1} = f_{PW}^{1}(\mathbb{MP}(f_{DW}^{1}(F_{j,1})))
\end{equation}
\begin{equation}\label{sa2}
A_{j,2} = f_{PW}^{1}(\mathbb{MP}(f_{DW}^{1}(F_{j,2})))
\end{equation}
$A_{j,1}\in \mathbb{R}^{H\times W\times C}$ ,and $A_{j,1}\in \mathbb{R}^{H\times W\times C}$, have the same operations. Each of them starts with depth-wise convolution followed by Max Pooling and Point-wise. The two part are combined to form an attention map $A_{j}$ for each sub-group to captures the long-range dependencies.
\begin{equation}\label{sa3}
A_{j} = Softmax(A_{j,1}\oplus A_{j,2})
\end{equation}
The output $\widehat{F}_{j}$ is computed using Eq. \ref{2nd_branch}.
\begin{equation}\label{2nd_branch}
\widehat{F}_{j} = Shuffle((A_{j}\odot F_{j}^{b})\oplus F_{j}^{b})
\end{equation}
Where $\odot$ is the element-wise multiplication. $f_{DW}^{1}(.)$ represents a depth-with convolution with kernel size of $1\times1$, $\mathbb{MP}(.)$ is a Max Pooling and Point-wise $f_{PW}^{1}(.)$.
The output is normalized with softmax function in Eq. \ref{sa3}. After that, an element-wise multiplication and summation are applied in residual-like connections with input feature $A_{j}$. Furthermore, we adopt a channel shuffle operator similar to ShuffleNet \cite{ma2018shufflenet} to enable cross-group information communication. Finally, the channel attention output Eq. \ref{2nd_branch} is added to the upper branch to form the Efficient Global Context Aggregation Module Eq. \ref{eq:egca}. The final output of the EGCA module is the same size as $F_{i}^{b}$, making EGCA quite simple and can be plugged in all the stages of the encoder, helping the module to learn multi-scale feature with different semantic. Thus, the EGCA module can be integrated easily into modern semantic segmentation architecture and other similar tasks. Table \ref{tab:egca_stages} shows the effect of the proposed efficient context aggregation module, in terms of floating-point operations per second (FLOPS), parameters, speed, and the mean intersection over union.
\subsection{Decoder-based subpixel Convolution (DSC)}
\label{sec:decoder_sub_pixel}
Many works have suggest a decoder structure in semantic segmentation methods. In some networks a naive decoder with direct bilinear upsampling is proposed \cite{chen2017rethinking,zhao2017pyramid}. PAN proposed a module in which the low level feature is guided with global context. Different from mentioned works and in order to retrieve the high resolution feature maps efficiently, we exploit subpixel convolution \cite{shi2016real} concept and technique. Multiple decoder-based subpixel convolution blocks are utilized in the DSC path to retrieve the spatial information with high level semantic feature generated by the SPFM module, and fuse the global context information gradually. The main component of the proposed DSC is shown in Figure \ref{sub_decoder} In details, we perform bilinear upsample to the high-level feature maps followed by $3\times3$ convolution, to reduce aliasing artifacts. $1\times1$ convolution is applied to the low level feature. Then we concatenate both low high feature and refine the output with $3\times3$ convolution. Finally, subpixel convolution is applied to generate a DSC with higher resolution. The following equation explains the relation between the different components of DSC.
\begin{equation}
F_{DSC} = \mathbb{PS}(f^{3}(Concat([up(f^{3},f^{1}])))
\end{equation}
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{subpixel_decoder.png}
\caption{Illustration of a decoder-based on subpixel convolution.}
\label{sub_decoder}
\end{figure}
\begin{table*}[hpt]
\centering
\caption{\MakeUppercase{The Per-class, class, and category IoU evaluation on the Cityscapes Test set. List of classes from left to right: Road, Side walk, Building, wall, Fence, Pole, Traffic light, Traffic sign, Vegetation, Terrain, Sky, Pedestrian, Rider, Car,Truck,Bus,Train, Motorbike, and Bicycle."cla->mIoU". ”-” indicates the corresponding result is not reported by the methods}}
\label{Table:Tab8}
\vspace{1ex}
\begin{center}
\begin{adjustbox}{width=1\textwidth}
\small
\begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c}
\hline Method&Road&Sidewalk&Building&Wall&Fence&Pole&Traffic light&Traffic sign&Vegetation&Terrain&Sky&Person&Rider&Car&Truck&Bus&Train&Motor&Bicyclist&mIoU\\
\hline
\hline
CRF-RNN\cite{zheng2015conditional}&96.3&73.9&88.2&47.6&41.3&35.2&49.5&59.7&90.6&66.1&93.5&70.4&34.7&90.1&39.2&57.5&55.4&43.9&54.6&62.5\\
FCN\cite{long2015fully}&97.4&78.4&89.2&34.9&44.2&47.4&60.1.5&65.0&91.4&69.3&93.9&77.1&51.4&92.6&35.3&48.6&46.5&51.6&66.8&65.3\\
DeepLabv2\cite{chen2017deeplab}&-&-&-&-&-&-&-&-&-& -&-&-&-&-&-&-&-&-&-&70.4\\
Dilation10\cite{yu2016multi}&97.6&79.2&89.9&37.3&47.6&53.2&58.6&65.2&91.8&69.4&93.7&78.9&55.0&93.3&45.5&53.4&47.7&52.2&66.0&67.1\\
PSPNet\cite{zhao2017pyramid}&\textbf{98.6}&\textbf{86.2}&\textbf{92.9}&50.8&\textbf{58.8}&64.0&\textbf{75.6}&\textbf{79.0}&\textbf{93.4}&\textbf{72.3}&\textbf{95.4}&\textbf{86.5}&\textbf{71.3}&\textbf{95.9}&\textbf{68.2}&\textbf{79.5}&\textbf{73.8}&\textbf{69.5}&\textbf{77.2}&\textbf{78.4}\\
AGLNet\cite{zhou2020aglnet}&97.8& 80.1& 91.0&\textbf{51.3}&50.6&58.3&63.0&68.5&92.3&71.3&94.2&80.1&59.6&93.8&48.4&68.1&42.1&52.4& 67.8&70.1\\
BiSeNetV2\slash BiSeNetV2\_L\cite{yu2021bisenet}&-&-&-&-&-&-&-&-&-& -&-&-&-&-&-&-&-&-&-&73.2\\
LBN-AA\cite{dong2020real}&98.2&84.0&91.6&50.7&49.5&60.9&69.0&73.6&92.6& 70.3&94.4&83.0&65.7&94.9&62.0&70.9&53.3&62.5&71.8-&73.6\\
\hline
Our(SPFNet)&98.5&85.2&92.6&48.4&55.8&\textbf{67.0}&74.5&77.8&\textbf{93.4}&70.7&95.2&85.7&68.8&95.2&53.3&73.0&59.7&67.2&75.4&75.7\\
Our(SPFNet)&98.4&84.9&92.2&47.6&54.1&65.9&73.6&76.2&93.0&66.4&94.8&85.3&68.5&94.6&47.6&53.9&28.8&65.7&75.0&71.9\\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table*}
\section{Experiments}
In this section, a comprehensive experiments have been carried out to evaluate the effectiveness of our proposed method, these experiments are conducted on Camvid \cite{brostow2009semantic} and Cityscapes \cite{cordts2016cityscapes} datasets. Experimental results demonstrate that SPFNet achieves state-of-the-art performance on Camvid dataset and comparable results to other methods on Cityscapes dataset. In the following we present the dataset and implementation details, then we demonstrate a series of ablation study on Camvid and Cityscapes datasets to explore the effect of each component in SPFNet. Finally, list the comparison of our method with the other state-of-the-arts networks on the two datasets.
\subsection{Experiments Settings}
\subsubsection{Camvid} The Camvid dataset \cite{brostow2009semantic} consists of 376 training examples, 101 evaluating images, and 233 testing images. For a fair comparison with the other state-of-the-art models, we evaluated our model in 11 classes such as building, sky, tree, car, and road and the class 12th was marked as ignore class to hold the unlabelled data. The small size of the data and unbalance distribution of its labels makes the dataset more challenging.\\
\subsubsection{Cityscapes} The Cityscapes dataset \cite{cordts2016cityscapes} is an urban scene understanding benchmark. It has 5000 high-resolution of 2048,1024pixels images with fine annotations captured from different cities. The dataset is divided into 2,975 images for training,500 images for validation, and the remaining 1525 images are for testing. We evaluate for 19 semantic segmentation classes.
\subsubsection{Implementation Details}
To implement our model, we use Adam optimizer \cite{kingma2014adam} with weight decay $5\times10^{-6}$, power 0.9, and the starting learning rate $lr_{init}$ for Camvid and Cityscapes datasets is set to $3\times10^{-4}$. We optimize the network by adopting the poly learning rate policy Eq. \ref{eq:scheduler} similar to the previous works \cite{chen2017deeplab,zhao2017pyramid}. All experiments are trained on PyTorch \cite{pytorch} on NVIDIA 3090 RTX GPU for 150 epochs and 500 epochs for Camvid and Cityscapes, respectively. The BatchNorm layers in SPFNet (except for the encoder) are replaced with InplaceBN-Sync \cite{ bulo2018place} to reduce memory usage, while the speed, and FLOPS analysis are measured using Nvidia GTX1080 GPU. \\
All the experiments are implemented with the same data augmentation. Specifically, the input images are randomly scaled between 0.75 and 2.0, random horizontal flip, and random cropping of $512 \times 1024$ and $360\times480$ image patches for training Cityscapes and Camvid, respectively. We trained both datasets with weighted cross-entropy loss. The compare the experimental results of our model on the Cityscapes validation and testing sets with the current state-of-the-art segmentation models. Further analysis elaborated in subsequent sections.
\begin{equation}\label{eq:scheduler}
lr = lr_{init}\times (1-\frac{iter}{max\_iter})^{power}
\end{equation}
Where ${max\_iter}$ is the maximum number of iterations.
\subsection{Ablation Study}
In this subsection, we run series of experiments to evaluate the impact of each component in the proposed SPFNet. First, an encoder-decoder based network that consists ResNet-based encoder and decoder-based sub\_pixel convolution was employed as the baseline network for the ablation study.
\subsubsection{Ablation for efficient global aggregation(EGCA)module}
\label{subsec:Ablation for efficient global aggregation(EGCA)module}
To test the influence of the efficient global context aggregation module in each stage, we set the integrate our proposed EGCA into a baseline of U-shape network ResNet34H (encoder) and DSC (decoder). Table. \ref{tab:egca_stages} illustrates the experimental results in terms of FLOPS, parameters, speed, and mean intersection over union. This process helps the model to learn multi-level cross-feature communication and reduce the semantic gap of feature from different level. The SPFM is not integrated into the network at this point. Using EGCA increases the computation by a small margin, but as we can see from the results it improves the segmentation accuracy. Table. \ref{tab:egca_stages} shows that integrating EGCA the second stage increases the FLOPS a little higher than other stages. Based on the configuration that uses EGCA in $stages{2}$ to $stage_{5}$, the mIoU improves over $2.84\%$ than the baseline. EGCA dropped the speed by 5.7 FPS than the baseline.\\
\begin{table}[hpt]
\caption{\MakeUppercase{Evaluation of EGCA effect at different stages on Cityscapes validation set.FLOPS, SPEED ARE ESTIMATED FOR AN INPUT SIZE OF 512,1024.}}
\label{tab:egca_stages}
\begin{center}
\begin{adjustbox}{width=0.48\textwidth}
\small
\begin{tabular}{l|c|c|c|c|c|c|c|c}
\hline
Baseline&$Stage_{2}$&$Stage_{3}$&$Stage_{4}$&$Stage_{5}$&FLOPS (G)&Params( M)&Speed (FPS)&mIoU(\%)\\
\hline
\hline
\checkmark&&&&&268.6&37.7&19.3&72.16\\
\checkmark&\checkmark&&&&280.3&37.8&16.2&72.8\\
\checkmark&\checkmark& \checkmark&&&287.0&37.9&15.2&73.19\\
\checkmark&\checkmark& \checkmark & \checkmark&&293.6&38.0&14.2&74.01\\
\checkmark&\checkmark& \checkmark& \checkmark& \checkmark&300.2&38.7&13.6&75.0\\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table}
\subsubsection{Ablation on Subspace Pyramid Fusion Module (SPFM)}
Hyperparameter analysis on SPFM Module on the Camvid dataset:
We investigate the impact of the hyperparameter s in the SPFM in terms of efficiency and accuracy on Camvid and Cityscapes datasets using Nvidia GTX1080 GPU. The parameter s represents the number of splits that used to learn each RPP module in the SPFM. We compared four cases for using ResNet18 (SPFNet18L) as backbone and another four cases with ResNet34 (SPFNet23L). s = \{2,4,8,16\}. At first, to test the baseline we replace the SPFM with 1x1 Conv and report the results in Table. \ref{tab:compare_spfm_camvid} for Camvid dataset. When s = 2, the SPFM fuses two reduced pyramid fusion modules to generate the multi-scale features. As shown in Table. \ref{tab:compare_spfm_camvid}, using SPFM with different number of split show higher accuracy (mIoU) than the baseline, and the model with higher s shows segmentation accuracy improvement. This implies that using higher s increase the network ability to extract complex multi-scale features for both SPFNet34L and SPFNet18, but at the expense of the speed (FPS). AS shown in Table. \ref{tab:compare_spfm_camvid} and Figure. \ref{fig:hyperparameter_s_camvid}.\subref{fig:hyperparameter_s_speed_camvid}. The performance of SPFM drops when using s=8 for both backbones as compared to the baseline. Using 2 split adds more parameters and FLOPS than using bigger s with faster inference. We also compared the SPFM with four splits in terms of FLOPS, number of parameters and speed. As illustrated in Table. \ref{tab:compare_spfm_camvid}, a larger s adds less FLOPS and parameters, but it decreases the speed. Based on this analysis, we finally chose s=4 for SPFM considering both efficiency and the segmentation accuracy.
\begin{table}
\centering
\caption{\MakeUppercase{The Evaluation and analysis of the Hyperparameter s on the SPFM Module. The input size for the baseline on the Camvid test set is $360\times840$.the input size for SPFM is $512\times12\times15$.}}
\label{tab:compare_spfm_camvid}
\vspace{1ex}
\begin{adjustbox}{width=0.48\textwidth}
\begin{tabular}{l|cc|cc|cc|c}
\hline
\multicolumn{8}{c}{Baseline SPFNet-34L(ResNet34 as backbone)}\\
\hline
\multirow{2}{*}{Network}&\multicolumn{2}{c|}{FLOPS (G)} & \multicolumn{2}{c|}{Params (M)}&\multicolumn{2}{c|}{Speed (FPS)} & \multirow{2}{*}{IoU (\%)} \\\cline{2-7}
\multicolumn{1}{c|}{}&Overall&SPFM& Overall& SPFM & Overall & SPFM& \\
\hline\hline
Baseline &22.8 & &37.7& &122.8& & 69.2\\
SPFM (s=2) &22.8 &3.0&37.7&6.7 &122.8&553.2&70.0\\
SPFM (s=4) &22.8 &1.6&37.7&3.4&122.8&398.7&70.3\\
SPFM (s=8) &22.8 &0.896&37.7&1.9&122.8&214.0&70.1\\
SPFM (s=16)&22.8 &0.544&37.7&1.0&122.8&109.0&70.6\\
\hline
\multicolumn{8}{c}{Baseline SPFNet-18L(ResNet18+DSC Module)}\\
\hline
Baseline &16.4& &27.6& &163.9&&68.5 \\
SPFM (s=2) &16.4&3.0&27.6&6.7&163.9&553.2&69.7\\
SPFM (s=4) &16.4&1.6&27.6&3.4&163.9&398.7&69.8\\
SPFM (s=8) &16.4&0.896&27.6&1.9&163.9&214.0&69.1 \\
SPFM (s=16)&16.4&0.544&27.6&1.0&163.9&109.0&69.9\\
\hline
\end{tabular}
\end{adjustbox}
\end{table}
\begin{figure}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=4cm,height=4cm]{sub_params_accuracy_camvid.png}
\caption{Hyperparameter vs accuracy.}
\label{fig:hyperparameter_s_accuracy_camvid}
\end{subfigure} \hspace{1.5mm}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=4cm,height=4cm]{sub_params_inference_camvid.png}
\caption{Hyperparameter vs FPS.}
\label{fig:hyperparameter_s_speed_camvid}
\end{subfigure}\hspace{1.5mm}
\caption{(a) The accuracy obtained by the proposed SPFM with different
values of the parameter s on the Camvid test set. (b) The inference speed with different s values.}
\label{fig:hyperparameter_s_camvid}
\end{figure}
Hyperparameter analysis on SPFM Module on the Cityscapes dataset:
We conducted similar analysis to examine the hyperparameter s of the SPFM on Cityscapes dataset in terms of the efficiency and accuracy. For this analysis we run the experiments for 150 epochs. From Table. \ref{tab:compare_spfm_cityscapes} and Figure. \ref{fig:hyperparameter_s_cityscapes} Experimental results and evaluation shows that the hyperparameter s follows the same patterns as on Camvid dataset.
\begin{table}
\centering
\caption{\MakeUppercase{The Evaluation and analysis of the Hyperparameter s on the SPFM Module.The input size for the baseline on the Cityscapes test set is $512\times1024$.input for the SPFM is $512\times16\times32$.}}
\label{tab:compare_spfm_cityscapes}
\vspace{1ex}
\begin{adjustbox}{width=0.48\textwidth}
\begin{tabular}{l|cc|cc|cc|c}
\hline
\multicolumn{8}{c}{Baseline SPFNet-34L(ResNet34 as backbone)}\\
\hline
\multirow{2}{*}{Network}&\multicolumn{2}{c|}{FLOPS (G)} & \multicolumn{2}{c|}{Params (M)}&\multicolumn{2}{c|}{peed (FPS)} & \multirow{2}{*}{mIoU (\%)} \\\cline{2-7}
\multicolumn{1}{c|}{}&Overall&SPFM& Overall& SPFM & Overall & SPFM& \\
\hline\hline
Baseline &68.1 & &37.7& &65.2& & 71.4\\
SPFM (s=2) &68.1 &8.6&37.7&6.7 &65.2&355.6&72.5\\
SPFM (s=4) &68.1 &4.6&37.7&3.4&65.2&266.2&73.7\\
SPFM (s=8) &68.1 &2.5&37.7&1.9&65.2&204.4&73.1\\
SPFM (s=16)&68.1 &1.5&37.7&1.0&65.2&107.6&73.7\\
\hline
\multicolumn{8}{c}{Baseline SPFNet-18L(ResNet18+DSC Module)}\\
\hline
Baseline &48.7& &27.6& &83.7&&71.1 \\
SPFM (s=2) &48.7&8.6&27.6&6.7&83.7&355.6&72.1\\
SPFM (s=4) &48.7&4.6&27.6&3.4&83.7&266.2&73.4\\
SPFM (s=8) &48.7&2.56&27.6&1.9&83.7&204.4&72.9\\
SPFM (s=16)&48.7&1.5&27.6&1.0&83.7&109.0&73.6\\
\hline
\end{tabular}
\end{adjustbox}
\end{table}
\noindent
\begin{figure}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=4cm,height=4cm]{sub_params_accuracy_cityscapes.png}
\caption{Hyperparameter vs accuracy.}
\label{fig:hyperparameter_s_accuracy_cityscapes}
\end{subfigure} \hspace{1.5mm}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=4cm,height=4cm]{sub_params_inference_cityscapes1.png}
\caption{Hyperparameter vs FPS.}
\label{fig:hyperparameter_s_speed_cityscapes}
\end{subfigure}\hspace{1.5mm}
\caption{(a) The accuracy obtained by the proposed SPFM with different
values of the parameter s on the Cityscapes validation set. (b) The inference speed with different s values.}
\label{fig:hyperparameter_s_cityscapes}
\end{figure}
\subsubsection{Ablation of SPFM and Other Multi-scale Modules}
To further investigate the effect of different modules in our proposed network, we first took a modified ResNet34 by not downsampling the second stage of the encoder. Then we integrated DSC module, and EGCA module as a baseline to test different multi-scale feature extractor modules. For this ablation setting the experiments run for 500 epochs. In detail, we conducted a number of experiments to compare the proposed SPFM with the multi-scale pyramid modules, i.e., ASPP, Vortex pooling, and DenseASPP, proposed in Deeplab \cite{chen2017deeplab}, and Vortex Pooling \cite{xie2018vortex}, and DenseASPP \cite{yang2018denseaspp}, respectively. In these experiments, we used somewhat lighter backbone, the aforementioned ResNet18, and modified ResNet34. As shown in Table. \ref{tab:mult_scale_spfm}, all the multi-scale modules have improved the baseline. The baseline model with the proposed SPFM outperforms those with ASPP (75.64\% mIoU), Vortex Pooling (75.72\% mIoU), and DenseASPP (75.21\% mIoU), showing the highest accuracy 78.04 mIoU with parameters and speed closer to ASPP. Besides, SPFM requires less parameters and FLOPS than DenseASPP and Vortex Pooling. This could be attributed to the capability of SPFM module to extract complex multi-scale semantic information more efficiently. SPFM learns multiple RPP at the same time, also it uses small dilation rate to obtain different-resolution feature maps, which introduces less noise.
\begin{table}
\centering
\caption{\MakeUppercase{Performance comparison of SPFM with state-of-the-art multi-scale extractor methods such as ASPP, DenseASPP, and Vortex Pooling on the Cityscapes Validation. FLOPS, SPEED ARE ESTIMATED FOR AN INPUT SIZE OF 512,1024. Under the SPFNet34H setting.}}
\label{tab:mult_scale_spfm}
\vspace{1ex}
\begin{adjustbox}{width=0.48\textwidth}
\small
\begin{tabular}{l|c|c|c|c}
\hline
Method&FLOPS&Params&FPS&mIoU(\%)\\
\hline
\hline
ResNet34+DSC baseline&268.6&37.7&19.3&72.16\\
ResNet34+DSC+EGCA&300.0&38.7&13.6&75.0\\
ResNet34+DSC+SPFM&286.6 &40.8&16.3&75.8\\
ResNet34+DSC+EGCA+DenseASPP&380.7&77.9&8.7&75.21\\
ResNet34+DSC+EGCA+ASPP&304.8&41.0&12.3&75.64\\
ResNet34+DSC+EGCA+Vortex Pooling& 322.0&49.4&11.1&75.72\\
ResNet34+DSC+EGCA+SPFM& 317.0&41.8&12.4&78.04\\
\hline
\end{tabular}
\end{adjustbox}
\end{table}
\subsection{Results on Cityscapes Dataset}
As shown in Table \ref{tab:cityscape_result_sota}, we compared the performance of the proposed model SPFNet against the other state-of-the-art models in terms of FLOPS, parameters, speed. We run the experiments with the training set and validation set. The input resolution is reduced from 2048 × 1024 into half to train our models. We evaluate the model segmentation accuracy on the test set, then submit the results to Cityscapes dataset online server \footnote{\url{https://www.cityscapes-dataset.com/submit/}} to get the results on the Cityscapes benchmark. Here, the network is compared against small networks such ICNet \cite{zhao2018icnet}, BiSeNet \cite{bilinski2018dense}, and larger networks PSPNet \cite{zhao2017pyramid}, RefineNet \cite{lin2017refinenet}, and Deeplab \cite{chen2014semantic}. The proposed SPFNet34H achieves 75.7\% mean IoU, with 41.9M parameters and inference speed of 12.7 FPS, while SPFNet18L obtains 71.9\% mIoU, with 31.7M parameters and inference speed of 46.5. Most of the larger models in Table\ref{tab:cityscape_result_sota} incorporate very deep feature extraction architectures with a larger number of parameters, whereas the proposed model used ResNet34 and ResNet18, which have fewer parameters but still obtain comparable results. We do not use multi-scale testing or multi-crop evaluation, two techniques that been used by many practitioners to help improve the accuracy. We illustrate some visual examples of SPFNet on Cityscapes validation set in Figure. \ref{cityscapes_fig}.
\begin{table*}
\centering
\caption{\MakeUppercase{Comparison between the proposed method SPFNet and other state-of-the-arts methods on the Cityscapes test dataset. Flops, speed are estimated for an Input size of 512,1024. ”-” indicates the corresponding result is not reported by the methods.}}
\label{tab:cityscape_result_sota}
\begin{adjustbox}{width=0.9\textwidth}
\small
\begin{tabular}{l|c|c|c|c|c|c|c}
\hline
Method&Backbone&Resolution&Parameters(M)&FLOPS (G)&Speed(FPS)&test set&mIoU\\
\hline
\hline
CRF-RNN\cite{zheng2015conditional}&VGG16&512$\times$1024&-&- &1.4&\checkmark&62.5\\
Deeplab\cite{chen2014semantic}&VGG16&512$\times$1024&262.1&457.8&0.25& \checkmark&63.1\\
FCN\cite{long2015fully}&VGG16 &512$\times$1024&134.5&136&2&\checkmark &65.3\\
Dilation10\cite{yu2016multi}&VGG16&512$\times$1024&140.8&-&-&\checkmark &67.1\\
RefineNet\cite{lin2017refinenet}&ResNet101 &512$\times$1024&118.1&526&9.1&\checkmark &73.6\\
PSPNet\cite{zhao2017pyramid}&ResNet101 &713$\times$713&250&412.2&0.78&\checkmark &78.4\\
\hline
SegNet\cite{badrinarayanan2017segnet}&BiseNet18&640$\times$360&29.5&286&16.7&\checkmark &57\\
TD4-Bise18\cite{hu2020temporally}&BiseNet18&1024$\times$2048&-&&&\checkmark &74.9\\
BiSeNet1\cite{yu2018bisenet}&Xception39&768$\times$1536&5.8&14.8&72.3&\checkmark &68.4\\
BiSeNet2\cite{yu2018bisenet}&ResNet18&768$\times$1536&49.0&54.0&45.7&\checkmark &74.7\\
ICNet\cite{zhao2018icnet}&PSPNet50&1024$\times$2048&26.5&28.3&30.3&\checkmark &69.5\\
LBN-AA\cite{dong2020real}&LBN-AA+MobileNetV2&488$\times$896&6.2&49.5&51.0&\checkmark &74.4\\
BiSeNetV2\-L\cite{yu2021bisenet}&No&512$\times$1024&&&47.3&\checkmark &75.3\\
\hline
SPFNet34H&ResNet34M&512$\times$1024&41.9&317&12.7&\checkmark &75.7\\
SPFNet18L&ResNet18&512$\times$1024&31.7&61.0&46.5&\checkmark &71.9\\
\hline
\end{tabular}
\end{adjustbox}
\end{table*}
\begin{figure*}
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.5cm,height=2.5cm]{frankfurt_000000_000576_leftImg8bit.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{frankfurt_000000_000576_leftImg8bit_gt.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{frankfurt_000000_000576_leftImg8bit_color.png}
\end{subfigure}
\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{frankfurt_000000_011074_leftImg8bit.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{frankfurt_000000_011074_leftImg8bit_gt.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{frankfurt_000000_011074_leftImg8bit_color.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{lindau_000012_000019_leftImg8bit.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{lindau_000012_000019_leftImg8bit_gt.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{lindau_000012_000019_leftImg8bit_color.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{lindau_000024_000019_leftImg8bit.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{lindau_000024_000019_leftImg8bit_gt.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{lindau_000024_000019_leftImg8bit_color.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{munster_000013_000019_leftImg8bit.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{munster_000013_000019_leftImg8bit_gt.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{munster_000013_000019_leftImg8bit_color.png}
\end{subfigure}
\centering
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{munster_000167_000019_leftImg8bit.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{munster_000167_000019_leftImg8bit_gt.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2.5cm]{munster_000167_000019_leftImg8bit_color.png}
\end{subfigure}
\centering
\caption{Visual results of our method SPFNet on Cityscapes dataset. The first row is the original image,the second row is the ground truth while the last row is represents the model performance.}
\label{cityscapes_fig}
\end{figure*}
\subsection{Results on Camvid Dataset}
Table \ref{tab:individual_camvid_s} presents the segmentation accuracy of our proposed method on the Camvid test set and its comparison with other state-of-the-arts methods. The model trained with an input resolution of 360x480 pixels with no use of external data we use the training samples and validation samples to train and validate our models while the testing samples are used to get the result to compare with other state-of-the-art models. Overall, our model has better accuracy. Our SPFNet34H with ResNet34 as a backbone achieves a significantly 75.1\% mIoU on the test set and inference speed 33.0 FPS, which is trade-off between speed and accuracy. Our SPFNet43L uses downsampling at stage 2 achieves 71.4\% mIoU with 79.3 FPS. Furthermore, we test our methods with two configurations of ResNet18. The SPFNet18H obtains 72.2\%mIoU with inference speed of 39.7. SPFNet18L obtains 70.3\%mIoU with inference speed of 109 FPS. These experiments validate the proposed architecture design. As demonstrated in Table\ref{tab:individual_camvid_s}, training the model with extra data increases the performance, for instance. BiSeNetV2* \cite{yu2021bisenet} performance improves from 72.4\% mIoU, when trained without extra data, to 76.7\% with model first train on Cityscapes data, while BiSeNetV2-L* performance improves from 73.2\% mIoU, when trained without extra data, to 78.5\% with th same settings. We report the individual category results on Camvid test set in Table. \ref{tab:individual_camvid_s}. Finally, we show some of the visual results of our method SPFNet on Camvid test set.
\begin{table}
\centering
\caption{\MakeUppercase{The accuracy, speed and parameters comparison of the proposed method against other semantic segmentation methods on the Camvid test set. the speed of our models are estimated for an Input size of 360,480. $^*$ indicates the model pre-trained on Cityscapes. ”-” indicates the corresponding result is not reported by the methods.}}
\label{tab:sota_camvid_s}
\vspace{1ex}
\begin{adjustbox}{width=0.48\textwidth}
\small
\begin{tabular}{l|c|c|c|c|c}
\hline
Method&Year&Resolution&Params (M)&Speed (FPS)&mIoU\\
\hline\hline
Deeplabv2\cite{chen2017deeplab}&2017&$720\times960$&262.1&4.9&61.6\\
PSPNet\cite{zhao2017pyramid}&2017&$720\times960$&250&5.4&69.1\\
DenseDecoder\cite{bilinski2018dense}&2018&$720\times960$&-&-&70.9\\
Dilation8\cite{yu2015multi}&2016&$720\times960$&-&4.4&65.3\\
\hline
SegNet\cite{badrinarayanan2017segnet}&2015&$360\times480$&29.5&-&55.6\\
ENet\cite{paszke2016enet}&2016&$360\times480$&0.37&-&51.3\\
DFANet-A\cite{li2019dfanet}&2019&$720\times960$&7.8&120&64.7\\
DFANet-B\cite{li2019dfanet}&2019&$720\times960$&4.8&160&59.3\\
BiSeNet1\cite{yu2018bisenet}&2018&$720\times960$&5.8&175&65.7\\
BiSeNet2\cite{yu2018bisenet}&2018&$720\times960$&49.0&116.3&68.7\\
ICNet\cite{zhao2018icnet}&2018&$720\times960$&26.5&27.8&67.1\\
DABNet\cite{li2019dabnet}&2019&$360\times480$&0.76&&66.4\\
CAS\cite{zhang2019customizable}&2020&$720\times960$&-&169&71.2\\
GAS\cite{lin2020graph}&2020&$720\times960$&-&153.1&72.8\\
AGLNet\cite{zhou2020aglnet}&2020&$360\times480$&1.12&90.1&69.4\\
CGNet\cite{9292449cgnet}&2020&$360\times480$&0.5&-&65.6\\
NDNet45-FCN8-LF\cite{yang2020small}&2020&$360\times480$&1.1&-&57.5\\
LBN-AA\cite{dong2020real}&2020&$720\times960$&6.2&39.3&68.0\\
BiSeNetV2/BiSeNetV2L\cite{yu2021bisenet}&2021&$720\times960$&-&124.5/32.7&72.4/73.2\\
BiSeNetV2$^*$/BiSeNetV2L$^*$\cite{yu2021bisenet}&2021&$720\times960$&-&124.5/32.7&76.7/78.5\\
\hline
SPFNet34H(ours) & &$360\times480$&41.8&33&75.1\\
SPFNet34L(ours) & &$360\times480$&41.8&79.3&71.4\\%\hline
SPFNet18H(ours) & &$360\times480$&31.7 &39.7&72.2\\
SPFNet18L(ours) & &$360\times480$&31.7&109&70.3\\
\hline
\end{tabular}
\end{adjustbox}
\end{table}
\begin{table*}
\centering
\caption{\MakeUppercase{Individual category results on Camvid test set in terms of mIoU for 11 classes.
”-” indicates the corresponding result is not reported by the methods.}}
\label{tab:individual_camvid_s}
\vspace{1ex}
\begin{tabu}{l|c|c|c|c|c|c|c|c|c|c|c|c}
\hline
Method&Building&Tree&Sky&Car&Sign&Road&Ped&Fence&Pole&Sidewalk&Bicyclist&mIoU\\
\hline
\hline
SegNet\cite{badrinarayanan2017segnet}&\textbf{88.8}&\textbf{87.3}&92.4&82.1&20.5&97.2&57.1&49.3&27.5&84.4&30.7&65.2\\
BiSeNet1\cite{yu2018bisenet}&82.2&74.4&91.9&80.8&42.8&93.3&53.8&49.7&25.4&77.3&50.0&65.6\\
BiSeNet2\cite{yu2018bisenet}&83.0&75.8&92.0&83.7&46.5&94.6&58.8&53.6&31.9&81.4&54.0&68.7\\
AGLNet\cite{zhou2020aglnet}&82.6&76.1&91.8&87.0&45.3&95.4&61.5&39.5&39.0&83.1&62.7&69.4\\
LBN-AA\cite{dong2020real}&83.2&70.5&92.5&81.7&51.6&93.0&55.6&53.2&36.3&82.1&47.9&68.0\\
BiSeNetV2\slash BiSeNetV2L\cite{yu2021bisenet}&-&-&-&-&-&-&-&-&-&-&-&72.4\slash 73.2\\ \hline
SPFNet34H(ours)&86.4&79.2&\textbf{92.6}&\textbf{90.3}&\textbf{56.3}&\textbf{95.9}&\textbf{68.1}&\textbf{54.5}&\textbf{42.3}&85.4&\textbf{74.2}&\textbf{75.1}\\
SPFNet34L(ours)&84.6&77.2&91.6&90.2&49.9&95.6&61.5&47.9&35.4&85.0&66.1&71.4\\
SPFNet18H(ours)&84.7&78.0&92.5&89.9&52.0&95.4&66.0&43.1&40.7&84.1&67.7&72.2\\
SPFNet18L(ours)&83.7&76.7&91.8&89.6&49&95.4&59.7&45.2&34.4&83.5&63.7&70.3\\
\hline
\end{tabu}
\end{table*}
\begin{figure*}
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{0001TP_008730.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{0001TP_008730_color.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{0001TP_008730_gt.png}
\end{subfigure}
\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{0001TP_008910.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{0001TP_008910_color.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{0001TP_008910_gt.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f00210.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f00210_color.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f00210_gt.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f01290.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f01290_color.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f01290_gt.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f03150.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f03150_color.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f03150_gt.png}
\end{subfigure}
\\
\centering
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f04590.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f04590_color.png}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=4.8cm,height=2cm]{Seq05VD_f04590_gt.png}
\end{subfigure}
\caption{Visual results of our method SPFNet on Camvid test set. The first row is the image,the second row is the prediction while the last row is ground truth.}
\label{Figure:Fig6}
\end{figure*}
\section{Conclusion}
This paper proposed subspace pyramid fusion module SPFM, which learns multi-scale context information by dividing the input feature maps into multiple subspaces. Furthermore, we have introduced the efficient global context aggregation module EGCA, which uses the channel-shuffle operator to enhance the information communication across different sub-features. We further propose a semantic segmentation structure based on our SPFM and EGCA modules, called SPFNet, to address the multi-scale feature fusion. The ablation studies in the Cityscapes dataset show the effectiveness of the proposed SPFM and EGCA. SPFNet34H, SPFNet34L, SPFNet18H, and SPFNet18L achieve 75.1\% mIoU, 71.4\% mIoU, 72.2\% mIoU, and 70.3\% mIoU , respectively on Camvid dataset. Our SPFNet34H and SPFNet18L obtain 75.7\% mIoU and 71.9\% mIoU , respectively on Cityscapes test sets.
\ifCLASSOPTIONcaptionsoff
\clearpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:1}
Robotics is one of the most rapidly emerging branches in the world of technology. From heavy-duty applications to the most sensitive microscopic application, robotic systems are present everywhere. Therefore, with the rise of applications of robotic systems, grew the demand for facilitating critical tasks and requirements of high dexterity applications. To fulfill these demands, a separate branch of robotics emerged coined as flexible robotic. A field in which manipulators and actuators of hyper redundant mechanisms were developed. Hyper redundant systems are robotic systems having a very high number of degree of freedom \cite{yesh_1}. These flexible robotic systems are very versatile in their applications. These flexible robotic systems are generally employed at a place with convoluted workspace and restricted maneuverability \cite{degani}. Such kind of flexible robotic systems can be abundantly seen as probing and inspection drones, surgical robots, military UGVs, etc \cite{wang_1,hwang_1}.
The flexible robotic systems also called snake/worm-like robotic systems can be divided into two groups, based on their mechanical structure, discrete systems, and continuous systems. Discrete systems have finite discrete rigid elements joint together \cite{yesh_2} while the continuous systems have soft or compliant materials used in the robotic system \cite{suzumori}. But in recent times, the line between these two categories has been blurred as newer designs being developed have both of the features. Another classification of hyper redundant robotic systems can be made based on how the robot is used. It can be classified either as a manipulator or a locomotor. If a robotic system is attached to a fixed base, then it can be called a manipulator \cite{xu_1}, but if the system can move freely, (such as a work or a snake), then it is to be called a locomotor \cite{loco}.\\
The flexible robotic systems are generally tendon actuated, pneumatically, or hydraulically actuated or are based on the shape memory feature of the material. Tendon-based mechanisms are one of the most used mechanisms for flexible manipulators. The manipulators are actuated by pushing or pulling a tendon to facilitate bending or curvature \cite{li_1}. These tendons-driven mechanisms are sometimes also referred to as tensegrity mechanisms because of the design principle wherein compression and expansion of lengths of opposite tendons takes place simultaneously \cite{furet r1, wenger r2}. A study was conducted to analyze and compare the kinematic properties of discrete tendon-driven flexible robots, continuous tendon-driven flexible robots, and concentric tube manipulators \cite{li_2}. The results of this study indicated that in the case of dexterity, the discrete system was the finest followed by continuous and concentric tube type manipulators. From a scaling point of view, concentric tube manipulators are better for miniaturization whereas discrete tendon-driven systems were more suitable for scaling up \cite{le_1}.\\
This flexible robotic system has found some very interesting applications in the field of medical sciences \cite{shang}. Even before the conceptualization of robotic surgeries, endoscopes were used to investigate the insides of the human gastrointestinal tract, and it was done manually \cite{haber,abbott}. With the medical science turning into minimally invasive surgeries (MIS), several surgical tools were required to be redesigned for MIS such as clippers, scalpels, cauterizers, catheters, etc \cite{roh,hu_1}. One such essential surgical tool is the surgical stapler \cite{Akopov2021}, which is generally used to avoid any spillage of intestine content onto the wounds of the patient and prevent any infection during the surgery. Commonly five kinds of staplers are used, namely TA, Thoracic-Abdominal; GIA, Gastro-Intestinal Anastomosis; Endo GIA, Endoscopic Gastro-Intestinal Anastomosis; EEA, End-End Anastomosis and Skin Stapler. The staplers are generally equipped with a separating knife to cut the tissues after the stapling is done. Endo GIA staplers are most commonly used for MIS. The Endo GIA staplers are available in three forms: passive articulated wrist type (PAW) \cite{Bolanos1997}, active articulated wrist type (AAW) \cite{Milliman1999}, and radial reload type staplers (RR). The RR has a fixed ``U'' shaped jaw which requires a huge incision into the abdomen for insertion. Therefore there is a dire need for a surgical stapler that can enter through the laparoscopic openings but can work as an RR-type stapler inside the body. Our study aims to design a stapler for laparoscopic rectal cancer surgery where conventional tools cannot be easily accessed. Two mechanisms will have to be positioned on each side of the colon to insert two lines of staples and then carry out its cutting.
The article is organized as follows. In section 2, the kinematics of the studied mechanism is presented. Section 3 reports the different computational results to study the singularities associated with the mechanism, then in section 4, the study of stability based on the principle of minimum energy, and finally, in section 5, an optimization of the dimensions of the mechanism is presented.
\section{Mechanism Design}
\label{sec:2}
In this article, the robotic system designed is a multi-segment planar robotic system. As depicted in Fig.~\ref{fig:robot_general}, the robotic system comprises of 3 trapezoidal segments A, B, and C stacked one above the other. Each segment may have a larger base plate and a smaller moving plate. Both the plates are connected by a central serial spine linkage, comprising of 3 links. The system is only composed of 3 sections to facilitate the position control of the end effector. Since the system is being controlled by only 2 tendons, hence it is under-actuated. Therefore, not more than 3 segments can be controlled effectively.
For example, in section A, the central spinal linkages are $A_{0}B_{0}$, $B_{0}C_{0}$ and $C_{0}D_{0}$. The link $A_{0}B_{0}$ and $C_{0}D_{0}$ are rigidly fixed and are perpendicular to their respective plate $A_{1}A_{2}$ and $D_{1}D_{2}$. These central spinal links are connected by two revolute joints with identical rotation angles. To achieve this coupling, the joints can have a sliding surface, similar to a knee joint in humans, or can also have an X-shape tensegrity module \cite{furet r1, wenger r2}, or a gear train to couple the movement of the two revolute joints.
The base plate and the moving plates are also attached with the help of two cables $\rho_1$ and $\rho_2$, present on either side of the central spine and two springs of stiffness $k_1$ and $k_2$ between ($A_1$ $B_1$) and ($A_2$ $B_2$), respectively.
The contraction of the cables will stimulate angular displacement in the revolute joints of the central spine and replication of this phenomenon in each of the three segments will eventually facilitate the bending of the robotic structure. The length of the cables is measured as $\rho_i$ and the angular deviation of the revolute joints is measured as $\alpha_i$ where
\begin{equation}
\alpha_{1} = \alpha_{2},
\quad \alpha_{3} = \alpha_{4},
\quad\alpha_{5} = \alpha_{6},
\end{equation}
because of the coupling inside each segment.
\begin{figure}
\centering
\includegraphics[width=12cm]{figures/Robot_cas_general.eps}
\caption{The tensegrity mechanism understudy with three segments stacked, named A, B, C for $\alpha_i=0$ (home pose) in the left and $\alpha_i=0.3$ in the right}
\label{fig:robot_general}
\end{figure}
The fixed coordinate frame of the base is represented by $\Sigma_0$ with the origin at $A_0$. The distances between the coordinates of each points are $||{\bf a}_1 - \bf{a}_0||=l_1$, $||{\bf a}_2 - \bf{a}_0||=l_1$, $||{\bf b}_1 - \bf{a}_0||=h_1$, $||{\bf c}_0 - \bf{b}_0||=h_2$, $||{\bf d}_0 - \bf{c}_0||=h_3$, $||{\bf d}_1 - \bf{d}_0||=l_2$ and $||{\bf d}_2 - \bf{d}_0||=l_2$.
The vector co-ordinates for the base mounting points are given by
\begin{equation}
{\bf a}_1=[-l_1,0] \quad {\bf a}_2=[l_1,0]
\end{equation}
The kinematic chain between the base and the platform of the first segment is defined by the following points
\begin{equation}
{\bf b}_0= \left[\begin{array}{c}
0 \\
h_{1}
\end{array}\right],~
{\bf c}_0= \left[\begin{array}{c}
-h_{2} \sin \! \left(\alpha_{1}\right) \\
h_{1}+h_{2} \cos \! \left(\alpha_{1}\right)
\end{array}\right],~
{\bf d}_0= \left[\begin{array}{c}
-h_{2} \sin \! \left(\alpha_{1}\right)-h_{3} \sin \! \left(2 \alpha_{1}\right) \\
h_{1}+h_{2} \cos \! \left(\alpha_{1}\right)+h_{3} \cos \! \left(2 \alpha_{1}\right)
\end{array}\right]
\end{equation}
The moving co-ordinate frame of the first segment is represented by $\Sigma_1$ with its origin at $D_0$. The spring mounting points are represented by $D_1$ and $D_2$.
\begin{eqnarray}
{\bf d}_1&=&\left[\begin{array}{c}
\left(-2 h_{3} \cos \! \left(\alpha_{1}\right)-h_{2}\right) \sin \! \left(\alpha_{1}\right)-2 l_{2} \left(\cos^{2}\left(\alpha_{1}\right)\right)+l_{2}
\\
2 h_{3} \left(\cos^{2}\left(\alpha_{1}\right)\right)+\left(-2 l_{2} \sin \! \left(\alpha_{1}\right)+h_{2}\right) \cos \! \left(\alpha_{1}\right)+h_{1}-h_{3}
\end{array}\right] \\
{\bf d}_2&=&\left[\begin{array}{c}
\left(-2 h_{3} \cos \! \left(\alpha_{1}\right)-h_{2}\right) \sin \! \left(\alpha_{1}\right)+2 l_{2} \left(\cos^{2}\left(\alpha_{1}\right)\right)-l_{2}
\\
2 h_{3} \left(\cos^{2}\left(\alpha_{1}\right)\right)+\left(2 l_{2} \sin \! \left(\alpha_{1}\right)+h_{2}\right) \cos \! \left(\alpha_{1}\right)+h_{1}-h_{3}
\end{array}\right]
\end{eqnarray}
The inverse kinematic model for a segment is used to determine the length of the springs and the cables between the base and the moving platform and the moving platform. The equations are given by
\begin{equation}
||{\bf a}_{1}-{\bf d}_{1}||=\rho_{1}, \quad
||{\bf a}_{2}-{\bf d}_{2}||=\rho_{2}
\end{equation}
These two equations can also be written as,
\begin{eqnarray}
\left(\left(-2 h_{3} \cos \! \left(\alpha_{1}\right)-h_{2}\right) \sin \! \left(\alpha_{1}\right)-2 l_{2} \left(\cos^{2}\left(\alpha_{1}\right)\right)+l_{2}+l_{1}\right)^{2} \nonumber\\
+\left(2 h_{3} \left(\cos^{2}\left(\alpha_{1}\right)\right)+\left(-2 l_{2} \sin \! \left(\alpha_{1}\right)+h_{2}\right) \cos \! \left(\alpha_{1}\right)+h_{1}-h_{3}\right)^{2}= \rho_1^2 \label{closed_loop_1} \\
\left(\left(-2 h_{3} \cos \! \left(\alpha_{1}\right)-h_{2}\right) \sin \! \left(\alpha_{1}\right)+2 l_{2} \left(\cos^{2}\left(\alpha_{1}\right)\right)-l_{2}-l_{1}\right)^{2} \nonumber\\
+\left(2 h_{3} \left(\cos^{2}\left(\alpha_{1}\right)\right)+\left(2 l_{2} \sin \! \left(\alpha_{1}\right)+h_{2}\right) \cos \! \left(\alpha_{1}\right)+h_{1}-h_{3}\right)^{2}= \rho_2^2
\end{eqnarray}
\section{Singularity Analysis}
For determining the singularities of the multi-segment planar mechanism, only section A was taken into consideration. As all other sections are similarly constructed, analysis of singularities in one section will lead to similar results in all other sections, and sections B and C will also have similar singularities.
Two closed-loop mechanisms can be described by ($A_1$, $A_0$, $B_0$, $C_0$, $D_0$, $D_1$) and ($A_2$, $A_0$, $B_0$, $C_0$, $D_0$, $D_2$). For the first closed-loop mechanism, the singular configurations with $\alpha_1$ as Cartesian values and $\rho_1$ as the input values are computed by differentiating with respect to time the Eq.~\ref{closed_loop_1}, as follow,
\begin{eqnarray}
&& -8 h_{3}^{2} \cos^{3}\left(\alpha_{1}\right) \sin \! \left(\alpha_{1}\right)+8 h_{3}^{2} \cos \! \left(\alpha_{1}\right) \sin \! \left(\alpha_{1}\right)-8 l_{2}^{2} \cos^{3}\left(\alpha_{1}\right) \sin \! \left(\alpha_{1}\right)\nonumber\\
&& +8 l_{2}^{2} \cos \!\left(\alpha_{1}\right) \sin \! \left(\alpha_{1}\right)-4 h_{3} \sin \! \left(\alpha_{1}\right) \cos^{2}\left(\alpha_{1}\right) h_{2}-4 h_{3} \sin^{3}\left(\alpha_{1}\right) h_{2}\nonumber\\
&& -4 h_{3} \cos^{2}\left(\alpha_{1}\right) l_{1}+4 h_{3} \sin^{2}\left(\alpha_{1}\right) l_{1}-4 l_{2} \cos^{2}\left(\alpha_{1}\right) h_{1}+4 l_{2} \sin^{2}\left(\alpha_{1}\right) h_{1}\nonumber\\
&& -8 h_{3}^{2} \sin^{3}\left(\alpha_{1}\right) \cos \! \left(\alpha_{1}\right)-2 h_{2} \cos \! \left(\alpha_{1}\right) l_{2}-2 h_{2} \cos \! \left(\alpha_{1}\right) l_{1}+8 l_{2} \cos \! \left(\alpha_{1}\right) l_{1} \sin \! \left(\alpha_{1}\right)\nonumber\\
&& -8 l_{2}^{2}\sin^{3}\left(\alpha_{1}\right) \cos \! \left(\alpha_{1}\right)-8 h_{3} \cos \! \left(\alpha_{1}\right) h_{1} \sin \! \left(\alpha_{1}\right)-2 h_{2} \sin \! \left(\alpha_{1}\right) h_{1}\nonumber\\
&& +2 h_{2} \sin \! \left(\alpha_{1}\right) h_{3}=0
\end{eqnarray}
The singularities are the roots of a 4th-degree equation. When $l_1 \neq l_2$, only a numerical method allows us to calculate them. In our case, we use the ``RootFinding:-Isolate'' function of Maple which computes all the roots after a substitution by the half-angle of $\alpha_1$ to obtain an algebraic equation \cite{rouillier}.
For the second closed-loop, the singularity locus is for opposite values of $\alpha_1$. For each closed loop, there can be up to four singular positions. The absolute value of the smallest angle $\alpha_i$ is called $\alpha_{sing}$ and it represents the largest travel that the mechanism can achieve. This value must be maximized.
As an example, for $h_1=1$, $h_2=1$, $h_3=1$, $l_1=1$, $l_2=1$, the singularity locus are
\begin{equation}
\alpha_1=-\frac{\pi}{4}, \alpha_1=\frac{3 \pi}{4},
\alpha_1=\arctan \! \left(\frac{\frac{1}{4}+\frac{\sqrt{7}}{4}}{-\frac{1}{4}+\frac{\sqrt{7}}{4}}\right), \alpha_1=\arctan \! \left(\frac{\frac{1}{4}-\frac{\sqrt{7}}{4}}{-\frac{1}{4}-\frac{\sqrt{7}}{4}}\right)-\pi
\end{equation}
Only two singular configurations are close to the home pose as shown in Fig.~\ref{fig:robot_sing}. The smallest absolute value of $\alpha_1$ defines the range of motion of the segment in both directions.
\begin{figure}
\centering
\includegraphics[width=10cm]{figures/Robot_sing.eps}
\caption{Four singular configurations for $h_{1} = 1, h_{2} = 1, h_{3} = 1, l_{1} = 1$ and $l_{2} = 1$}
\label{fig:robot_sing}
\end{figure}
\section{Stability Analysis}
\label{sec:3}
The stability of the robotic system is one of the major components in its performance evaluation \cite{behzadipour st1}. Generally, when a system is in equilibrium and is displaced by an external force if the system comes back to its equilibrium state, it is called a stable equilibrium. If the system does not return to equilibrium, it is considered an unstable system and if the system does not affect itself, then the system is known as a neutral system \cite{chadefaux st2}.
A robotic system is always desired to be a stable system in its operational capacity. If a robotic system becomes unstable at any point in its trajectory, then it may cause substantial errors during its control \cite{wanda}.
For this multi-segment planar robotic system, the stability is evaluated by using the principle of minimum internal energy.
For section A, two springs are connected between $A_1D_1$ and $A_2D_2$ with a no-load length equal to $l_0$. This length is chosen as 40\% of the length in the home pose. This value will be proportional to the size of the second and third segments if we want to characterize the stability of the complete mechanism.
For segment A, the energy $E$ for a given $\alpha$ inside the mechanism is
\begin{equation}
E= \frac{1}{2}\left(k_1(\rho_1-l_0)^2+k_2(\rho_2-l_0)^2 \right)
\end{equation}
To compare two mechanisms, we introduce the total energy as $E_t=\int_{-\alpha_{sing}}^{\alpha_{sing}} E$.
Depending on the sizes of the robot, several stability schemes can appear either stable around $\alpha=0$ up to the singularities (Fig.~\ref{fig:robot_energy}(left)) or unstable in the position for $\alpha=0$ (Fig.~\ref{fig:robot_energy}(right)).
\begin{figure}
\centering
\includegraphics[width=12cm]{figures/Robot_Energy.eps}
\caption{Variation of the energy as a function of $\alpha_1$ for a stable home pose for $h_{1} = 0$, $h_{2} = 1$, $h_{3} = 0$, $l_{1} = 1$, $l_{2} = 1$, $k = 1$ (left) and unstable home pose for $h_{1} = 1$, $h_{2} = 1$, $h_{3} = 1$, $l_{1} = 1$, $l_{2} = \frac{1}{2}$, $k = 1$}
\label{fig:robot_energy}
\end{figure}
\section{Optimization of One Segment}
The objective of our optimization is to find the design parameters that allow the greatest angular deflection and then to study the stability of the solutions found. A ratio $\lambda$ between the base and the platform is also studied to see if elephant trunk structures, with sections of decreasing sizes, are relevant \cite{yang_1}.
So, the objective function is defined as
\begin{equation}
f({\bf x})= \alpha_{sing} \rightarrow \max
\end{equation}
where ${\bf x}$ is the set of the design parameters $[h_1, h_2, h_3, l_1, \lambda]$. To simplify the optimization, we set $h_3=h_1$, $l_2= \lambda l_1$ and we defined the following constraints:
\begin{eqnarray}
0<l_1<4.5 \quad 0 \leq h_1 \leq 1 \quad 0 \leq h_2 \leq 2 \quad 1/20 \leq \lambda \leq 1
\end{eqnarray}
The solution to this optimization problem can be achieved by several methods. As the dimension of the problem is small, a discretization of the parameter space has been performed to find the optimal solutions.
The results of the optimization show that the maximum value of $\alpha$ is $\pi/2$ and this for any value of $\lambda$ and for $l_2=2$, i.e. the maximum bound, and for $h_1=0$ for the minimum bound. Figure~\ref{fig:robot_l} shows the evolution of $l_1$ and $l_2$ as a function of $\lambda$.
\begin{figure}
\centering
\includegraphics[width=5.5cm]{figures/Robot_opt_l.eps}
\includegraphics[width=6cm]{figures/Robot_opt_energy.eps}
\caption{Variation of $l_1$ in black and $l_2$ in blue as a function of $\lambda$ (left) and the stability analysis of the optimal solutions as a function of $\lambda$}
\label{fig:robot_l}
\end{figure}
For all solutions, the energy has its maximum value for $\alpha_{sing}$ and has a local minimum for $\alpha=0$.
If we observe the value of $E(\alpha=0)$ and $E(\alpha=\pi/2)$, the difference is always the same. The total energy decreases when $\lambda$ tends to 1. This means that the actuation forces will be smaller when $\lambda=1$. Conversely, the smaller $\lambda$ is, the greater the external forces must be to move the mechanism from its home position. Moreover, to have a regular stack of modules, module C is smaller than module B which is smaller than module A if $\lambda$ is different from 1.
\section{Conclusions and Future Work}
\label{sec:4}
To summarise, in this article a multi-segmented planar tensegrity mechanism was designed consisting of a stack of three tensegrity mechanisms. The mechanism was mechanically formulated as having linear springs and cables for actuation. The singularities of the parallel mechanism were analyzed as well as the stability by using the principle of minimum energy. Optimization was done to obtain the greatest angular deflection according to a ratio between the size of the base and the platform.
We observe that the angle $\alpha$ can vary between $-\pi/2$ and $\pi/2$, that all optimal solutions are stable up to the singularity, and that when $\lambda=1$, the total energy in a module is minimal.
Further research is in progress to control the three modules with only two cables going from the base to the third segment and to know the stability of the mechanism against external forces. We also need to estimate the radii of curvature required to encompass the patients' colon from CT scan studies and encapsulate the mechanism with an elastomer with the clamps and knife to section the colon.
\input{references}
\end{document}
|
\section{Conclusion and Future Work}
This work studies filter bubbles in interactive recommendation, where overexposure effect can be easily tracked and estimated. Accordingly, we propose a counterfactual interactive recommender system (CIRS), with leveraging causal inference in offline RL to deduce users' varying satisfaction.
To conduct evaluations, we innovatively create a faithful RL environment, KuaiEnv, based on a real-world fully observed user rating dataset.
Extensive experiments demonstrate that the proposed method can burst filter bubbles and increase users' cumulative satisfaction.
In the future, it is interesting to explore other types of biases in interactive recommendation.
This work illustrates filter bubble would evolve during the process of user-system interaction. We believe other biases may also have this nature. It will be valuable to transfer the experience of this work to tackle other biases in a dynamic environment. The proposed causality-enhanced IRS framework could be adapted for other biases.
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
\label{sec:intro}
\setlength{\epigraphwidth}{.9\columnwidth}
\renewcommand{\epigraphflush}{center}
\renewcommand{\textflush}{flushepinormal}
\epigraph{\textit{``Personalization filters serve up a kind of invisible autopropaganda, indoctrinating us with our own ideas, amplifying our desire for things that are familiar and leaving us oblivious to the dangers lurking in the dark territory of the unknown."} }
{{\footnotesize{\textit{––Eli Pariser, The Filter Bubble \cite{pariser2011filter}}}}}
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\linewidth]{figs/intrograph.pdf}
\caption {Disentangling causal effects on user satisfaction.}
\label{fig:intrograph}
\end{figure}
Recommender systems have deeply affected our lives. They change the way of retrieving information from searching strenuously, to obtaining conveniently via the precise personalization. The system usually achieves personalization by learning on the collected behavior data of users and selecting the products that users potentially like \citep{MFmodel,DeepFM,lightgcn,gao2019bloma}. With time evolving and data accumulating, the recommender gradually becomes a mirror reflecting each user's interest and narrows down the recommendation lists to the items with the maximum user interest.
However, this comes at a price. While enjoying the precise personalized recommendations, users have to face the fact that the variety of information shrinks. When users become isolated from the information that varies from their dominant preferences, they are stuck in the so-called ``filter bubbles''.
Filter bubbles are common in recommender systems. Recent studies conducted extensive experiments in large-scale recommender systems and found there are two primary causes of filter bubbles \cite{WWW21fb,fb2014,Measuring_misinfo,how_youtube,Auditing_youtube,recsys21best}. From users' perspective, those who have less diverse preferences are more liable to be stuck in the bubbles. From the system's perspective, learning can lead to emphasizing the dominant interest of a user.
Moreover, the system usually assumes that user satisfaction equals intrinsic interest --- even though the interested items have been overexposed, it assumes that the user satisfaction remains unchanged, which is improper. We believe that such an overexposure effect is the main cause of filter bubbles.
Overexposing items has a pernicious effect on a user's satisfaction, even though the user is interested in the recommended item. For example, a user who likes dancing can be satisfied when she receives a recommendation about a dancing song. However, she may be bored after dozens of times of incessant recommendations about dancing songs, and thus refuses to choose it. Therefore, it is of great significance to burst filter bubbles for a recommender system for the purpose of maximizing user satisfaction. The key is to disentangle the causal effects on user satisfaction, i.e., modeling how a user's intrinsic interest together with the overexposure effect affect the user's final satisfaction (\myfig{intrograph}).
Existing methods address filter bubble with two strategies: (1) helping users improve their awareness of diverse social opinions \cite{fb_awreness,echo_1}, and (2) making the model enhance diversity \cite{liu2020diversified,Deconstructing_fb,echo_2}, serendipity \cite{combating_fb,neural_serendipity}, or fairness \cite{AAAI2020burstingfb} of the recommendation results. However, most of these strategies are heuristic and focus on the static recommendation setting.
They cannot fundamentally address the problem since they do not capture the main cause of filter bubble, i.e., the overexposure effect.
\begin{figure}[!t]
\centering
\includegraphics[width=0.85\linewidth]{figs/intro.pdf}
\vspace{-2mm}
\caption {Illustration of interactive recommendation and the formation of the filter bubble.}
\label{fig:intro}
\end{figure}
\medskip
In this work, we focus on the interactive recommender system (IRS) in the dynamic environment. An IRS is formulated as a sequential decision making process, which allows the tracking and modeling of the dynamic and real-time overexposure effect.
\myfig{intro}(a) draws an example of interactive recommendation, where a model recommends items (i.e., makes an action $a$) to a user based on the interaction context (i.e., state $s$ reflecting the user's information and interaction history), then it receives user feedback (i.e., reward $r$ representing user satisfaction). The interaction process is repeated until the user quits. The model will update its policy $\pi_\theta$ with the goal to maximize the cumulative satisfaction over the whole interaction process. To achieve this goal, the IRS should avoid overexposing items because users will get bored and their satisfaction will drop in filter bubbles (\myfig{intro}(b)).
Though appealing, this idea faces an inevitable challenge:
it is difficult to learn an IRS online with real-time feedback. The reason is twofold:
(1) for the model, training the policy online increases the learning time and the deploying complexity \cite{xu2015infrastructure};
(2) for the user, interacting with a half-baked system can hurt satisfaction \cite{schnabel2018short,gao2021advances}.
Hence, it is necessary to train the IRS offline on historical logs before serving users online.
To this end, we need to conduct causal inference \cite{pearl2009causality} on the offline data to disentangle the causal effect of user interest and overexposure. This provides an opportunity to answer a counterfactual question in the serving phase: ``\emph{Would the user still be satisfied with the interested item if it had been overexposed?}'' --- If the answer is no, then the filter bubble will occur once recommending the item,
so we should not recommend it.
\smallskip
We propose a counterfactual interactive recommender system (CIRS) to achieve this goal. It enhances offline reinforcement learning (offline RL) \cite{levine2020offline} with causal inference \cite{pearl2009causality}.
\myfig{intro}(c) shows its learning framework, which contains three recurrent steps:
1) learning a causal user model to capture both user interest and item overexposure effect,
2) using the learned user model to provide \emph{counterfactual satisfaction} (i.e., the rewards given by the causal user model instead of immediate users' feedback) for planning an RL policy,
and 3) evaluating the RL policy in terms of the cumulative user satisfaction in real environments.
In addition, we propose an interactive recommendation environment, Kuai\-Env, for evaluating the problem.
It is created based on the KuaiRec dataset that contains a fully filled user–item matrix (density: 99.6\%) collected from the Kuaishou App\footnote{\url{https://www.kuaishou.com/cn}} \cite{gao2022kuairec}. In this matrix, IRSs can be evaluated more faithfully since there are no missing values.
To effectively reflect the effect of filter bubbles, we further add a "feel bored then quit" exit mechanism in both VirtualTaobao and KuaiEnv to simulate the reactions of real users.
The experiments show that the proposed CIRS can capture the overexposure effect in offline learning and achieve a larger cumulative satisfaction compared with other baselines.
Our contributions are summarized as follows:
\begin{itemize}
\item To the best of our knowledge, this is the first work to address filter bubbles in interactive recommendation, where filter bubbles can be more naturally observed and evaluated via modeling the overexposure effect on user satisfaction.
\item We solve the problem by integrating causal inference into offline RL. We are the first to combine causal inference and offline RL in interactive recommender systems.
\item We conduct empirical studies on Kuaishou App to verify that overexposing items indeed hurts user experience, and we demonstrate that the proposed method can burst filter bubbles and increase the cumulative satisfaction.
\end{itemize}
\section{Related Work}
\label{sec:related}
We briefly review the related works from three perspectives: filter bubbles, causality inference, and offline RL in recommendation.
\subsection{Filter Bubbles in Recommendation}
The term filter bubble was coined by internet activist Eli Pariser \cite{pariser2011filter} and used for describing the state of intellectual isolation that results from algorithm's personalization process.
When a user is stuck in a filter bubble, the algorithm will show the user limited information that conforms with their preexisting belief or interest.
In recommendation, many works systematically analysed the causes of filter bubbles. There are in general two causes. The first cause may be users' personal features \cite{how_youtube,Auditing_youtube,Measuring_misinfo,WWW21fb}.
For example, \citet{WWW21fb} conducted simulation studies on news recommendations and found that users with more extreme preferences are more liable to be stuck in filter bubbles. Additionally, \citet{Measuring_misinfo} studied misinformation filter bubbles on YouTube, the most popular video-sharing platform, and found that personalized information, e.g., age, gender, and watch history, can indeed have a significant effect on amplifying misinformation filter bubbles.
The second cause is the overexposure effect. Recent empirical studies verified that the formation of filter bubbles is mainly due to that the model emphasizes certain items with dominant user interest \cite{recsys21best,WWW21fb}. This kind of filter bubbles can have a pernicious effect on user satisfaction. In this work, we focus on modeling and alleviating the overexposure effect to address the filter bubble issue.
Existing methods that aim to combat filter bubbles focus on improving users awareness of diverse social opinions \cite{fb_awreness,echo_1}, enhancing models' diversity \cite{liu2020diversified,Deconstructing_fb,echo_2}, serendipity \cite{combating_fb,neural_serendipity}, fairness \cite{AAAI2020burstingfb} of recommendation results, correcting model behavior by watching debunking content \cite{recsys21best}. However, these strategies focus on the static recommendation setting, where the dynamic nature of filter bubbles is hard to model. In contrast, we capture the overexposure effect, i.e., the main cause of filter bubbles, and solve the problem in interactive recommendation.
\subsection{Causality-enhanced Recommendation}
Recently, causal inference (CI) has drawn a lot of attention in neural language processing (NLP) \cite{CIinNLP}, computer vision (CV) \cite{lopez2017discovering,wang2020visual} and recommender system (RS) \cite{expomf,rs_treatment,Causalembedding_recsys18,zhangyang,yang2021top,Counterfactual-Data-Augmented}. Instead of exploiting the correlation relationships between input and output by feeding data to the black-box neural networks, CI explicitly models the causal mechanism among variables \cite{pearl2009causality,rubin2020book}.
In recommender systems, CI can be a powerful tool for addressing various biases in the data or the learning process, e.g., selection bias and popularity bias \cite{chen2020bias}.
For example, many studies tried to estimate the unbiased user preference based on inverse propensity scoring (IPS) \cite{saito_unbiased,rs_treatment}, which is an effective CI-based method. Intuitively, when a user has a lower probability of seeing an item, we should assign increased importance to this sample and vice versa. However, IPS-based causal methods suffer from the high variance issue and it is difficult to estimate the propensity score \cite{saito_unbiased}.
Recently, researchers have followed Pearl's causal inference framework \cite{pearl2009causality} in RS \cite{zhangyang,yang2021top,xu2021causal,wenjiekdd21}. We generally summarize the procedure of these studies as three steps:
1) Constructing a causal graph, i.e., a representation of the structure causal model (SCM) that describes the causal relationship among the related variables.
2) In the learning stage, fitting an unbiased model (e.g., implemented as a neural network) on the training dataset based on the proposed causal graph.
(3) In the inference stage, actively changing certain variables (called \emph{intervention}) according to the certain requirement, then predicting the unbiased result of the target variable.
In this work, we use this framework because it enables the model to remove the effect of confounding factors \cite{wenjiekdd21,xu2021causal}, or answer the counterfactual question \cite{zhangyang} in recommendation.
\subsection{Offline RL for Recommendation}
Recommender systems are designed to enhance user satisfaction or increase sales when serving online users. We usually consider recommendation as a sequential decision making process, where the recommender policy decides to make recommendations according to previous user feedback. Common studies on static recommendation models (e.g., DeepFM \cite{DeepFM} and LightGCN \cite{lightgcn}) only pay attention to improving the performance in single-round recommendation. To directly gain the long-term success in the multi-round decision making problem, we can adopt the interactive recommender system (IRS) based on reinforcement learning (RL) \cite{wang2022best,liu2018deep,sigir20_interactive,zhang2019reward,NICF,lixin_kdd,liu2020diversified,gao2021advances}. Though effective, learning an IRS policy with online users is impractical as it is too slow and will hurt user experience \cite{gilotte2018offline}. On the other hand, the recorded recommender logs and offline user feedback are
easier to obtain \cite{kiyohara2021accelerating}. Therefore, it is natural to think of learning a policy on the offline data, which is the core idea of offline RL \cite{levine2020offline}.
Commonly used offline RL strategies include off-policy evaluation (OPE) learning \citep{swaminathan2015counterfactual} and model-based RL methods \cite{Pseudo-Dyna-Q,chen2019generative}. However, the former one has the high-variance problem and the latter suffers from bias issues \cite{levine2020offline}.
As mentioned above, CI is a powerful tool to address the bias problem in recommender system. Hence, we combine the model-based offline RL and the CI technique to develop an unbiased IRS that can recognize and address the filter bubble problem on the offline data.
Here, we briefly summarized in \mytable{summary} six types of recommenders with respect to three dimensions:
(1) whether the system explicitly builds a user model trying to capture real user preference,
(2) whether the system considers debiasing, and
(3) whether the system has an RL-based policy.
Note that there are two other works lie in the same category as our CIRS, but they are not designed for the filter bubble problem.
\begin{table}[t]
\caption{Six Types of Recommender Systems}
\label{tab:summary}
\vspace{-2mm}
\tabcolsep=1pt
\footnotesize
\renewcommand\arraystretch{1.1}
\begin{tabular}{@{}lcccc@{}}
\toprule
& \textbf{User Model} & \textbf{Debiasing} & \textbf{RL-based} & \textbf{Publications} \\ \toprule
Static RS & \green{\small{\Checkmark}} & & &\citep{MFmodel,DeepFM,lightgcn,yu2019generating} \\ \midrule
Unbiased static RS & \green{\small{\Checkmark}} & \green{\small{\Checkmark}} & &\citep{saito_unbiased,rs_treatment,KDD21popularitybias,li2021causal,wenjiekdd21,zhangyang,DICE,bias_general} \\ \midrule
Traditional IRS & & & \green{\small{\Checkmark}} & \citep{liu2018deep,Xinxin,sigir20_interactive,zhang2019reward,NICF,lixin_kdd,embedding_IRS,liu2020diversified} \\ \midrule
Model-based IRS & \green{\small{\Checkmark}} & & \green{\small{\Checkmark}} & \citep{Pseudo-Dyna-Q,xiangyu_model_based,xiangyuwww21,nips_model_bai,chen2019generative,wang2021AAAI} \\ \midrule
OPE-based IRS & & \green{\small{\Checkmark}} & \green{\small{\Checkmark}} & \citep{swaminathan2015counterfactual,minmin_topK,ie2019youtube,off_two_stage,xiao2021general,jagerman-when-2019,KDD20_counterfactual,Pessimistic2021recsys}\\ \midrule
\begin{tabular}[c]{@{}c@{}}Unbiased model-\\ based IRS\end{tabular} & \green{\small{\Checkmark}} & \green{\small{\Checkmark}} & \green{\small{\Checkmark}} & \begin{tabular}[c]{@{}c@{}}\citep{Keeping-recsys,minmin_user_model}\\\textbf{CIRS (Ours)}\end{tabular} \\ \bottomrule
\end{tabular}
\vspace{-3mm}
\end{table}
\section{Prerequisites}
\label{sec:pre}
In this section, we introduce the problem definition and the empirical analyses of the real-world data from Kuaishou.
\subsection{Problem Definition}
We denote the user set as $\mathcal{U}$ and item set as $\mathcal{I}$.
The set of all interaction sequences of a user $u\in\mathcal{U}$ can be denoted as $\mathcal{D}_u = \{\mathcal{S}_u^1,\mathcal{S}_u^2,\cdots,\mathcal{S}_u^{\left|\mathcal{D}_u\right|}\}$. Each $\mathcal{S}_u^k\in\mathcal{D}_u$ is the $k$-th interaction sequence (i.e., trajectory) recording a complete interaction process: $\mathcal{S}_u^k=\{(u, i_l, t_l)\}_{1\leq l \leq \left|\mathcal{S}_u^k\right|}$, where the user $u$ begins to interact with the system at time $t_1$ and quits at time $t_{|\mathcal{S}_u^k|}$, and $i_l \in \mathcal{I}$ is the recommended item at time $t_l$. Let $\mathbf{e}_u \in \mathbb{R}^{d_u}$ and $\mathbf{e}_i \in \mathbb{R}^{d_i}$ be the feature representation vectors of user $u$ and item $i$, respectively.
For the system, the task is to recommend items to users based on their preferences and interaction history. This process can be cast as a reinforcement learning problem, whose key components are summarized as follows:
\begin{itemize}
\item \textbf{State}. The system maintains a state $\mathbf{s}_t \in \mathbb{R}^{d_s}$ at time $t$ is regarded as a vector representing information of all historical interactions between user $u$ and the system prior to $t$. In this paper, we obtain the $\mathbf{s}_t$ by using the Transformer model \cite{transformer}.
\item \textbf{Action}. The system makes an action $a_t$ at time $t$ is to recommend items to user $u$. Let $\mathbf{e}_{a_t} \in \mathbb{R}^{d_a}$ denote the the representation vector of action $a_t$. In this paper, each action $a_t$ recommends only one item $i$. Hence, we have $\mathbf{e}_{a_t} = \mathbf{e}_i$.
\item \textbf{Reward}. The user $u$ returns feedback as a reward score $r_t$ reflecting its satisfaction after receiving a recommended item $i$.
The reward can also be the \emph{counterfactual satisfaction} predicted by the causal user model $\phi_M$ instead of real users' feedback.
\item \textbf{Policy network}. The policy network optimizes the target policy $\pi_\theta = \pi_\theta(a_t|\mathbf{s}_t)$ that represents the probability of making an action $a_t$ conditioned on the state $\mathbf{s}_t$. It decides how to generate an item $i$ to recommend and is implemented as a fully-connected neural network.
\end{itemize}
\medskip
To make full use of the historical data, we base our method CIRS on the offline RL framework. Learning CIRS requires three recurrent steps (as illustrated by three color blocks in \myfig{method}):
\begin{enumerate}
\item Training a causal user model $\phi_M$ on historical interaction data $\{(u, i, r)\}$ to estimate not only user interest but also the effect of item overexposure.
\item Using the learned causal user model $\phi_M$ (instead of real users) to train policy $\pi_\theta$. In each interaction loop, $\phi_M$ samples a user $u$ to interact with $\pi_\theta$. When $\pi_\theta$ makes an action $a_t$ (recommends $i$), $\phi_M$ provides a \emph{counterfactual satisfaction} as the reward $r$. Intuitively, if $\pi_\theta$ has made similar recommendations before, $\phi_M$ shrinks the reward $r$.
\item Serving the learned policy $\pi_\theta$ to real users and evaluating the results in the interactive environment. When the interaction ends, we save the log $\{u, i, r, t\}$ to historical data for the purpose of future learning.
\end{enumerate}
The three steps can be conducted repeatedly to continuously improve $\phi_M$ and $\pi_\theta$.
\subsection{Field Study of Overexposure Effect on User Satisfaction}
\label{sec:verification}
As mentioned above, besides users' responsibility, the cause of filter bubbles is that the model incessantly recommends similar items to users. We suppose this is pernicious as users may not feel comfortable under such circumstances. Thus, we present a hypothesis as follows:
\begin{hyp}
\label{hyp:hyp}
Users' satisfaction will drop if the recommended items (or similar items) are repeatedly exposed and recommended to them in a short time.
\end{hyp}
To verify this hypothesis, we conduct empirical studies on real data from Kuaishou, a video-sharing mobile App. We chose $7,176$ users from the video-watching user pool of the platform. We filter and collect their interaction history from August 5th, 2020 to August 11st, 2020. There are $34,215,294$ views in total and $3,110,886$ videos were watched. Each item is tagged with at least one and at most four category tags, and there are $31$ category tags in total.
During watching a video, users can choose to quit watching at anytime by scrolling down to the next video or leaving the video-playing interface. Each video has a comments section where users can enter by clicking the ``comment'' button. If users are interested in a video, they will stay watching it for a longer time or enter its comments section. Therefore, we design two key metric indicators to reflect user satisfaction: One is the time duration staying in comments section, and the other is the video watching ratio, which is the ratio of viewing time to the total video length.
To show how item exposure affects user satisfaction, we study how the two indicators change with the degree of overexposure effect. Specifically, we group all collected views with respect to (a) the number of videos with the same tag watched in one hour, or (b) the time interval between now (watching this video) and the last time viewing a video with the same tag. Then we compute the average values of the mentioned indicators in every group. The results are shown in \myfig{statistics}. Two observations are found:
\begin{obs}
User satisfaction towards a recommended item drops when the system increases the number of similar items in recent recommendations.
\end{obs}
\begin{obs}
User satisfaction towards a recommended item drops as the time interval between two similar items is shortened.
\end{obs}
\begin{figure}[!t]
\centering
\includegraphics[width=1\linewidth]{figs/Statistics_kuaishou.pdf}
\vspace{-2mm}
\caption {Empirical studies of exposure effect on Kuaishou. Statistics of two metrics vary with the exposure effect.}
\vspace{-2mm}
\label{fig:statistics}
\end{figure}
The result is statistically significant since even the smallest group contains enough points: $31,277$ points for the group at $x=50$ in \myfig{statistics} (a) and $76,303$ points for the group at $x=(7h\sim8h)$ in \myfig{statistics} (b).
Therefore, \myhyp{hyp} is empirically proved. This suggests that the filter bubble, i.e., the overexposure effect of the recommendation algorithm, does have a pernicious impact on user satisfaction. Consequently, we can design a ``feel bored then quit'' exit mechanism in the constructed environments as a reflection of user behavior, which enables us to simulate the effect of filter bubbles effectively. We will illustrate it in \mysec{exp}.
\section{Proposed Methods}
\label{sec:method}
In this section, based on the observations obtained in the field studies, we propose a counterfactual interactive recommender system (CIRS) that leveraging causal inference in offline RL.
We first introduce CIRS's three main modules in the offline RL framework, then we describe how to leverage causal inference to disentangle the causal effects on user satisfaction.
\subsection{Offline RL-based Framework}
We base our CIRS model on offline RL, where we can utilize the large amount of offline data to train the interactive recommender system. The framework of CIRS is illustrated in \myfig{method}. It contains three stages (shown in three color blocks): pre-learning stage, RL planning stage, and the RL evaluation stage. The functions of these three stages correspond to the three:
(1) pre-learning the user model $\phi_M$ via supervised learning,
(2) using the learned user model $\phi_M$ to learn RL policy $\pi_\theta$ by providing \emph{counterfactual satisfaction} as the reward, and
(3) evaluating policy $\pi_\theta$ in the real environment.
Specially, the real environment can either be the real-world online environment or the simulation environment that can reflect real user behavior.
Next, we separately introduce the three main components in CIRS: causal user model $\phi_M$, the state tracker module, and RL-based interactive policy $\pi_\theta$.
\subsubsection{Causal User Model}
The causal user model $\phi_M$ learns user interest based on the historical data and provide the counterfactual satisfaction $r$ for the RL planning stage. It aims to explicitly disentangle the causal effect by correctly modeling the effect of item overexposure on user satisfaction. There are two sub-modules in the user model: an interest estimation module designed for computing the user' intrinsic interest $y$, and a counterfactual satisfaction estimation module capturing how the overexposure effect affects user satisfaction $r$.
The details of this component and will be reported in \mysec{causal},
where we will formally introduce the causal view of the recommender.
\begin{figure}[!t]
\centering
\includegraphics[width=1\linewidth]{figs/method.pdf}
\vspace{-2mm}
\caption {Learning Framework of Counterfactual IRS.}
\vspace{-2mm}
\label{fig:method}
\end{figure}
\subsubsection{Transformer-based State Tracker}
The state $\mathbf{s}_t$ in the interactive recommendation should include all critical information at time $t$ for policy $\pi_\theta$ to make decisions. That includes: the feature vector $\mathbf{e}_u$ of the user $u$, the feature vectors of the recently recommended items, i.e., actions of the system in the interaction loop $\{\mathbf{e}_{a_1},\cdots,\mathbf{e}_{a_t}\}$, and the user's feedback towards them $\{r_1,\cdots,r_t\}$.
In order to automatically extract key information from these vectors, we use the Transformer model \cite{transformer} to derive $\mathbf{s}_t$ as illustrated in \myfig{method}.
Transformer is a state-of-the-art sequence-to-sequence model with an attention mechanism that can capture the dependence between current input and previous input sequences \cite{zhang2022mixhead}. In this work, we use only a two-layer encoder of Transformer.
Since the input is generated sequentially, we need to add a mask to prevent future information leaking into each state of the sequence.
We further use a gate mechanism to filter information from the action $\mathbf{e}_{a_t}$ and user feedback $r_t$. Hence, the input for Transformer at time $t$ is $\mathbf{e}^\prime_{a_t}:=\boldsymbol{g}_{t}\odot\mathbf{e}_{a_t}$. Where `$\odot$' denotes the element-wise product, and the gating vector $\boldsymbol{g}_{t}$ is computed as:
\begin{equation}
\boldsymbol{g}_{t}=\sigma\left(\mathbf{W} \cdot \operatorname{Concat}\left(\mathbf{r}_{t}, \mathbf{e}_{a_t}\right)+\mathbf{b}\right),
\end{equation}
where $\mathbf{W} \in \mathbb{R}^{d_s \times (1+ d_a)}$ and $\mathbf{b} \in \mathbb{R}^{d_s}$ refers to the weight matrix and bias vector, respectively. $\operatorname{Concat}(\cdot)$ is the vector concatenation operator. Besides, we use a feed-forward network ($\operatorname{FFN}$) to transform the representation vector of user $u$ from $\mathbf{e}_u \in \mathbb{R}^{d_u}$ to $\mathbf{e}^\prime_u \in \mathbb{R}^{d_s}$ to let it lie in the same space
of $\mathbf{e}^\prime_{a_t}$.
The feature vectors, the parameters in Transformer and the gate are initialized randomly and trained in the end-to-end manner using the gradients passed from the RL model.
\subsubsection{RL-based Interactive Recommendation Policy}
We implement our interactive recommendation policy $\pi_\theta$ as the PPO algorithm \cite{ppo}. PPO is a powerful on-policy reinforcement learning algorithm based on actor-critic framework \cite{a2c}.
It can work in both discrete and continuous state space and action space.
We aim to maximize the cumulative user satisfaction of the long-term interaction, which can be realized by maximizing the objective function of PPO:
\begin{equation}
\label{eq:L_rl}
\mathbb{E}_{t}\left[\min \left(\frac{\pi_{\theta}\left(a_{t} \mid \mathbf{s}_{t}\right)}{\pi_{\theta_{\text{old}}}\left(a_{t} \mid \mathbf{s}_{t}\right)} \hat{A}_t, \operatorname{clip}\left(\frac{\pi_{\theta}\left(a_{t} \mid \mathbf{s}_{t}\right)}{\pi_{\theta_{\text{old}}}\left(a_{t} \mid \mathbf{s}_{t}\right)}, 1-\epsilon, 1+\epsilon\right) \hat{A}_t\right)\right],
\end{equation}
where $\epsilon$ is the hyperparameter that controls the maximum percentage of change that can be updated at one time. The function $\operatorname{clip}(x,a,b)$ clips the variable $x$ in the range of $[a,b]$. $\theta_{\text{old}}$ is the policy before updating, i.e., the interaction data are generated under policy $\theta_{\text{old}}$. And the advantage function $\hat{A}_t$ is implemented as the generalized advantage estimator (GAE) \cite{gae}
which can be generally understood as a quantity that is proportional the normalized cumulative rewards of a sequence $\sum_{l=0}^{|\mathcal{S}|} \gamma^{l} r_{t+l}$. For more details, we refer the audiences to the original PPO algorithm \cite{ppo}.
At last, to learn a good policy, we need the counterfactual satisfaction given by $\phi_M$ to be as correct and constructive as possible. Next, we will introduce how to build a causal user model $\phi_M$ to capture the overexposure effect thus avoiding the filter bubble.
\begin{figure}[!t]
\centering
\includegraphics[width=1\linewidth]{figs/causal.pdf}
\caption {Causal graphs of traditional IRS models (a) and the CIRS model (b-c). $U$: a user, $I$: an item, $R$: user satisfaction (quantified by feedback such as clicks or viewing time), $Y$: intrinsic interest, $E_t$ and $e^*_t$: the overexposure effect. Latent variables are shaded.}
\vspace{-2mm}
\label{fig:causal}
\end{figure}
\subsection{Causal Inference-based User Satisfaction Disentanglement}
\label{sec:causal}
We illustrate the causal graphs of the traditional recommender systems and our CIRS model in \myfig{causal}.
\begin{itemize}
\item Node $U$ represents a certain user $u$, e.g., an ID or the profile feature that can represent the user.
\item Node $I$ represents an item $i$ that is recommended to user $u$.
\item Node $R$ represents user $u$'s real-time satisfaction for the recommended item $i$. It is the feedback such as a click or the video watching ratio.
\item Node $Y$ represents the user's intrinsic interest that is static regardless of item overexposure.
\item Node $E_t$ and $e^*_t$ represent the overexposure effect of item $i$ on user $u$. $E_t$ is a random variable and $e^*_t$ is the value of $E_t$ computed in the inference stage (i.e., the RL planning stage).
\end{itemize}
Traditional recommender systems will fit user satisfaction $R$ based on solely the feature information of user $U$ and item $I$ (\myfig{causal} (a)). This assumes that user satisfaction equals to intrinsic interest, which is improper as stated before.
Our proposed CIRS model innovatively takes into account the overexposure effect $E_t$, and disentangles the causal effect on user satisfaction $R$. Concretely, $R$ is generated from two causal paths:
\begin{enumerate}
\item $(U, I) \rightarrow Y \rightarrow R$: This path projects user $u$ and item $i$ to their corresponding intrinsic interest ${y}_{ui}$. Then user satisfaction ${r}_{ui}$ is proportional to ${y}_{ui}$.
\item $I \rightarrow E_t \rightarrow R$: This path captures the real-time overexposure effect $e_t(u,i)$ of an item $i$ on user $u$'s satisfaction $r$.
\end{enumerate}
\smallskip\noindent\textbf{Intrinsic Interest Estimation.}
We estimate the user's intrinsic interest ${y}_{ui}$ as $\hat{y}_{ui} = f_\theta (u, i)$. The estimation model $f_\theta (u, i)$ can be implemented by various established recommendation models, such as DeepFM \cite{DeepFM} used in this work. We illustrated this part in the interest estimation module in \myfig{method}.
\smallskip\noindent\textbf{Overexposure Effect Definition}
Considering that overexposure effect negatively affects user satisfaction,
we define the overexposure effect $e_t$ of recommending an item $i$ to user $u$ at time $t$ as:
\begin{equation}
\label{eq:ee}
e_t:= e_t(u,i):= \alpha_u \beta_i \hspace{-10pt}\sum_{\left(u, i_{l}, t_{l}\right) \in \mathcal{S}_{u}^{k},t_{l}<t} \hspace{-10pt} \exp\left(-\frac{t-t_{l}}{\tau}\times \operatorname{dist}(i,i_l)\right),
\end{equation}
where $\operatorname{dist}(i,i_l)$ is distance between two items $i$ and $i_l$. $\alpha_u$ represents the \textit{sensitivity} of user $u$ to the overexposure effect, e.g., a user with a large $\alpha_u$ is more likely to feel bored when overexposed to similar content. Likewise, $\beta_i$ represents the \textit{unendurableness} of item $i$. For example, a classical music might be more endurable than a pop song, so $\beta_i$ of the classical music is smaller. $\tau$ is a temperature hyperparameter.
\smallskip\noindent\textbf{User Satisfaction Estimation.}
Generally, a similar item $i_l$ (i.e., with smaller $\operatorname{dist}(i,i_l)$) that was recommended recently (i.e., with smaller $t-t_l$) contributes a larger overexposure effect to item $i$. And $e_t$ is the sum of the effect of all items recommended to the user $u$ before time $t$.
After obtaining the overexposure effect $e_t$ via \myeq{ee} and intrinsic interest $\hat{y}_{u i}$ via DeepFM, we estimate user $u$'s satisfaction on item $i$ as:
\begin{equation}
\hat{r}^{t}:=\hat{r}_{u i}^{t}=\frac{\hat{y}_{u i}}{1+e_t(u,i)}.
\end{equation}
Therefore, a large $e_t(u,i)$ will diminish user satisfaction $\hat{r}_{u i}^{t}$ even with unchanged intrinsic interest $y_{ui}$.
In the training stage of causal user model $\phi_M$, we minimize the objective function in common recommendation models. In experiments, we use the MSE loss for the VirtualTaobao and BPR loss for KuaiEnv:
\begin{equation}
L_{\mathrm{MSE}}=\hspace{-10pt} \sum_{(u, i, t) \in \mathcal{D}}\hspace{-10pt}
\left(\hat{r}_{u i}^{t}-r_{u i}^{t}\right)^2,
\hspace{5pt}
L_{\mathrm{BPR}}=- \hspace{-15pt} \sum_{(u, i, t) \in \mathcal{D}, j \sim p_{n}}\hspace{-15pt}
\log \left(\sigma\left(\hat{r}_{u i}^{t}-\hat{r}_{u j}^{t}\right)\right).
\end{equation}
Where $\sigma(x)=\frac{1}{1+e^{-x}}$ is the Sigmoid function. The item $j$ is a negative instance sampled from the distribution $p_n$.
\smallskip\noindent\textbf{Counterfactual Satisfaction Estimation.}
In the RL planning stage,
when the learned causal user model $\phi_M$ interacts with the policy $\pi_\theta$,
the overexposure effect $e^*_t$ now is different to $e_t$ in the pre-learning stage. Therefore, we perform the causal intervention $do(E_t=e^*_t)$ \cite{pearl2009causality} by cutting off the path $I \rightarrow E_t \rightarrow R$ as shown in \myfig{causal}(c). Unlike traditional causal methods aiming to remove the effect of confounders \cite{wang2020visual,wenjiekdd21}, we still need to model the correct overexposure effect $e^*_t$ in this stage. Note that we use the asterisk to mark out all values in this intervention stage. We compute $e^*_t$ as:
\begin{equation}
\label{eq:eestar}
e^*_t(u,i) = \gamma^*\cdot \alpha_u \beta_i \hspace{-10pt}\sum_{\left(u, i_{l}^*, t_{l}^*\right) \in \mathcal{S}_{u}^{*},t_{l}^*<t} \hspace{-10pt} \exp\left(-\frac{t-t_{l}^*}{\tau^*}\times \operatorname{dist}(i,i_l^*)\right),
\end{equation}
where $\mathcal{S}_{u}^{*}$ is the new interaction trajectory produced in the RL plan ning stage. $\gamma^*$ is a hyper parameter introduced to adjust the scale of the overexposure effect. We fix $\gamma$ to be $10$ throughout experiments. $\tau^*$ is the temperature hyperparmerter in the intervention stage and can have a different value with $\tau$ in \myeq{ee}.
We estimate the \emph{counterfactual satisfaction} as:
\begin{equation}
\hat{r}^{t*}_{u i}=\frac{\hat{y}_{u i}}{1+ e^*_t(u,i)}.
\label{eq:cs}
\end{equation}
By now, we can use the estimated \emph{counterfactual satisfaction} as the reward signal to update the RL policy by optimizing \myeq{L_rl}. The whole process is illustrated in the pre-learning and RL planning stage in \myfig{method}.
We use the Adam optimizer \cite{adam} in learning the causal user model $\phi_M$, the policy $\pi_\theta$, and the state tracker.
\smallskip
At last, by innovatively enhancing the offline RL framework with causal inference, we obtain a policy that can guarantee large user satisfaction by preventing filter bubbles, i.e., overexposing items.
\section{Experiments}
\label{sec:exp}
In this section, we conduct experiments to evaluate the IRS.
We aim to investigate the following research questions:
\textbf{(RQ1)} How does CIRS perform compared with SOTA static recommendation methods and RL-based interactive recommendation policies?
\textbf{(RQ2)} How does CIRS perform in a limited number of interactive rounds?
\textbf{(RQ3)} How does CIRS perform in different environments with varying user tolerance of filter bubble?
\textbf{(RQ4)} What is the effect of the key parameters in CIRS?
\subsection{Experimental Setup}
We introduce the experimental settings with regards to the environments, evaluation metrics, and state-of-the-art methods.
\subsubsection{Recommendation Environments}
\label{seq:env}
Traditional recommendation datasets are too sparse to evaluate the interactive recommender systems.
We use two recommendation environments, VirtualTaobao\footnote{\url{https://github.com/eyounx/VirtualTaobao}} \cite{virtualtaobao} and KuaiEnv. The two environments can play the same role as the online real users. For the recommenders, an environment is like a black box as shown in the upper right corner of \myfig{method}.
\begin{figure*}[t!]
\centering
\includegraphics[width=1\linewidth]{figs/main_result.pdf}
\vspace{-6mm}
\caption{Results of all methods with large interaction rounds (\mysec{large_round}) and limited interaction rounds (\mysec{limited_round}).}
\vspace{-3mm}
\label{fig:main_result}
\end{figure*}
\smallskip
\textbf{VirtualTaobao} is a benchmark RL environment for recommendation. It is created by simulating the behaviors of real users on Taobao, one of the largest online retail platforms. In VirtualTaobao, a user is represented as an $88$-dimensional vector $\mathbf{e}_{u} \in \{0,1\}^{88}$, and a recommendation is represented as a $27$-dimensional vector $\mathbf{e}_{i} \in \mathbb{R}^{27}, \mathbf{0}\leq \mathbf{e}_{i} \leq\mathbf{1}$. When a model makes a recommendation $\mathbf{e}_{i}$, the environment will immediately return a reward signal representing user interest, i.e., a scalar $r\in \{0,1,\cdots, 10\}$.
It provides $100,000$ logged interactions for training the offline RL policy. In VirtualTaobao, we use Euclidean distance for the $\operatorname{dist}(i,i_l)$ term in \myeq{ee} and \myeq{eestar}.
\smallskip
\textbf{KuaiEnv} is created by us on the KuaiRec dataset\footnote{\url{https://chongminggao.github.io/KuaiRec/}} \cite{gao2022kuairec}.
KuaiRec is a real-world dataset that contains a fully observed user-item interaction matrix, which means each user has viewed each video and then left feedback.
Therefore, unlike VirtualTaobao that simulates real users by training a model on Taobao data, KuaiEnv uses real user historical feedback. We define the reward signal as the video watching ratio which is the ratio of viewing time to the total video length. Without loss of generality, we use this floating point number to indicate users' intrinsic interest.
We use the fully-observed matrix, i.e., \emph{small matrix}, to evaluate the policy $\pi_\theta$. For pre-learning the user model $\phi_M$, we use the additional sparse user–video interactions in the \emph{big matrix}. For the details of the data, please refer to the KuaiRec dataset \cite{gao2022kuairec}.
Each video in KuaiRec has at least one and no more than four categorical attributes, e.g., Food or Sports. Hence we use the Hamming distance for computing the $\operatorname{dist}(i,i_l)$ term in \myeq{ee} and \myeq{eestar}.
\smallskip
\textbf{Exit Mechanism.}
By now, VirtualTaobao and KuaiEnv can provide users' intrinsic interest as the reward signal. However, they cannot reflect users' responses to the overexposure effect. To this end, we introduce the ``feel bored then quit'' mechanism in two environments to penalize filter bubbles.
Concretely, in VirtualTaobao, we compute the Euclidean distance between recommended target and the most recent $N$ recommended items. If any of them is lower than the threshold $d_Q$, the environment will quit the interaction process as the real users can feel bored and quit under such the monotonous recommendation.
In KuaiEnv, similarly, for the most recent $N$ recommended items, if there are more than $n_Q$ items in the $N$ items have at least one attribute of the current recommended target, then the environment ends the interaction process.
\subsubsection{Evaluation Metrics}
We aim to evaluate the model performance with regard to cumulative satisfaction over the whole interaction trajectory $\mathcal{S}$, i.e., $\sum_{l=0}^{|\mathcal{S}|} r_{t+l}$,
where $r_t$ is the reward signal returned by VirtualTaobao or KuaiEnv. Note that in this setting, user satisfaction is set as:
\begin{equation*}
\vspace{-1mm}
\text{satisfaction}=\left\{
\begin{array}{rcl}
\text{interest}, & & \text{if no filter bubble ever occurs}, \\
0, & & \text{otherwise}.
\end{array} \right.
\end{equation*}
I.e., if the recommendation does not trigger the exit mechanism, we can accumulate the rewards represent intrinsic interest. But whenever an overexposed item triggers the exit mechanism, the interaction is interrupted and no reward can be added anymore.
Intuitively, to pursue long-term success, the recommender policy must find a trade-off between pursuing a higher single-round satisfaction and maintaining a long interaction sequence.
We report the average cumulative satisfaction over $100$ interaction sequences.
\subsubsection{Baselines}
We use the commonly used static recommendation models plus some straightforward policies as baselines.
For KuaiEnv, the static recommendation modules in the user model are implemented as the powerful baseline \textbf{DeepFM} \cite{DeepFM} and three debiasing methods: \textbf{IPS} \cite{swaminathan2015counterfactual}, \textbf{PD} \cite{zhangyang}, and \textbf{DICE} \cite{DICE}. We implement the debiasing modules of the three methods on the basic DeepFM model. Since these static methods do not have a strategy for varying the recommendation results, we add a Softmax layer on their predicting results and randomly sample items from the distribution. For comparing with the other policies, we add several effective policies include \textbf{Random}, \textbf{$\mathbf{\epsilon}$-greedy}, \textbf{UCB} \cite{UCB} on the results of DeepFM.
For VirtualTaobao, since the users and items are given by feature vectors, we can only implement a multilayer perceptron (\textbf{MLP}) with sampling on the Softmax results and the $\epsilon$-greedy strategy.
Besides, we implement the powerful RL baseline, PPO \cite{ppo}, on the same offline RL setting, i.e., it is learned by interacting with the user model but without the causal inference module. For comparison, we denoted this baseline as $\textbf{CIRS w/o CI}$.
\subsection{Overall Performance Comparison}
\label{sec:large_round}
We evaluate the proposed CIRS and baselines in two environments. For general comparison, we do not limit the length of the interaction and set the max round to be large enough (but feasible to implement). We set the max round to be $50$ and $100$ for VirtualTaobao and KuaiEnv, respectively. For the parameters in the environment setting, we set the window size, i.e., the number of the most recent recommendations, to be $N=5$ and exit threshold to be $d_Q = 3.0$ for VirtualTaobao and $N=1$, $n_Q=1$ for KuaiEnv.
The results are shown in \myfig{main_result} (A-B). From (A1) and (B1) we can see the proposed CIRS achieves the maximal average cumulative satisfaction after certain epochs' learning. In the first few epochs in VirtualTaobao, the performances of both CIRS and CIRS w/o CI improve because the RL policy gradually finds the right user preference so the satisfaction in each round increases (A3). Interestingly, the increasing of the single-round satisfaction compromises the length of trajectory in the beginning (A2). And the policy of CIRS eventually finds a balance point between length and single-round satisfaction, thus achieving the maximal cumulative satisfaction. However, without the causal inference module, the policy becomes unstable and the performance degenerates with epoch increasing.
In KuaiEnv, CIRS also achieves the largest cumulative satisfaction (B1) after enough epochs.
For both VirtualTaobao and KuaiEnv, CIRS beats the counterpart method CIRS w/o CI, which demonstrates the effectiveness of the causal module in CIRS.
For other baselines, we can see that all other methods except Random can achieve better single-round performance (A3 and B3). However, their recommendation results are too limited and narrow even with the randomness introduced by the basic policies (i.e., Random sampling, Softmax-based sampling, and $\epsilon$-greedy). Note that in VirtualTaobao, even the random sampling cannot bring longer interaction sequence because of the curse of dimensionality: The action space has $88$ dimensions so the Euclidean distance of any two random points becomes indiscriminate. IPS has the maximal high variance in terms of the single-round performance (B3), which has been widely discussed \cite{swaminathan2015counterfactual}. UCB is a policy that can automatically balance exploration and exploitation, it has the best performance in the beginning. However, after several epochs of exploration, the policy enhances its belief on certain items and thus leads to the filter bubble. Therefore, UCB ends up with the lowest interaction length.
To conclude, except the deep RL policy-based methods (i.e., CIRS and CIRS w/o CI), static recommendation models with heuristic policies (i.e., Softmax-based sampling, and $\epsilon$-greedy, and UCB) cannot overcome the overexposure effect, thus results in filter bubbles and low user satisfaction. Furthermore, by comparing CIRS with CIRS w/o CI, we show the effectiveness of causal inference in the offline RL framework.
\subsection{Results with Limited Interaction Rounds}
\label{sec:limited_round}
In real-world recommendation scenarios, users have limited energy and will not spend too many rounds for interacting with the recommender. Therefore, we limit the max round to be $10$ and $30$ in VirtualTaobao and KuaiEnv, respectively.
We aim to investigate whether the policy can exploit to improve the single-round satisfaction under such a situation.
We alter the exit threshold $d_Q = 1.0$ in VirtualTaobao for better demonstration.
From the results in \myfig{main_result} (C-D), we can see that CIRS outperforms CIRS w/o CI in KuaiEnv, and tie it in VirtualTaobao. In VirtualTaobao, both of the two policies achieve the same level of single-round performance (greater than $4.0$) that is similar as the static methods (C3 and A3). In KuaiEnv, both the two policies attain higher single-round performances (D3) compared with that in the former setting (B3). Specially, CIRS has a great improvement on single-round performance (D3 and B3), which means that it is suitable for real-world interactive recommendation scenarios with limited interaction rounds. Actually, in (B1 and B3), we can also see that the performance of CIRS continues increasing at $epoch=200$. This means CIRS has potential even under huge rounds, given enough training time.
Again, in this setting, we conclude that by integrating with causal inference, we let CIRS outperform its counterpart greatly.
\begin{figure}[!t]
\centering
\includegraphics[width=1\linewidth]{figs/leave.pdf}
\caption{Results under different user sensitivity}
\vspace{-4mm}
\label{fig:leave}
\end{figure}
\subsection{Results with Different User Sensitivity}
\label{sec:vary_env}
To validate the generality of CRS, we vary the parameters $d_Q$ and $N$ in VirtualTaobao and KuaiEnv, respectively. A large $d_Q$ or $N$ means that users get more sensitive to filter bubbles and become easier to quit the interaction. The results in \myfig{leave} show that CIRS outperforms all baselines when users are less sensitive. However, the performance inevitably decreases when users become more sensitive. At $d_Q \geq 4$ or $N \geq 4$, CIRS and CIRS w/o CI can only achieve the same performance with other baselines. This means, facing extremely picky users, even the model enhanced by causal inference cannot
alleviate the dissatisfaction caused by filter bubbles.
Meanwhile, other baselines have similar performance under different user sensitivity --- the recommendations make users feel bored and quit even through the user is more tolerant to filter bubble (i.e., less sensitive to item overexposure). This also demonstrates they are not suitable for addressing the filter bubble issue.
\begin{figure}[!t]
\tabcolsep=-1pt
\centering
\begin{tabular}{cc}
\includegraphics[width=0.5\linewidth]{figs/alpha_popularity.pdf} & \includegraphics[width=0.5\linewidth]{figs/beta_popularity.pdf}
\end{tabular}
\vspace{-4mm}
\caption {Relationship between the learned $\alpha_u, \beta_i$ and the data statistics (user activity and item popularity).}
\label{fig:ab}
\end{figure}
\begin{figure}[!t]
\centering
\vspace{-4mm}
\includegraphics[width=0.9\linewidth]{figs/tau.pdf}
\begin{tabular*}{0.7\linewidth}{@{\extracolsep{\fill}}cc}
(a) VirtualTaobao & (b) KuaiEnv
\end{tabular*}
\caption{Cumulative satisfaction with different $\tau$-$\tau^*$ pairs.}
\label{fig:tau}
\end{figure}
\subsection{Effect of Key Parameters}
\label{sec:key_p}
We investigate the effect of the key parameters of CIRS in KuaiEnv. We compare the learned $\alpha_u$ and $\beta_i$ in the user model with two statistics of data, i.e., the activity of the users and the popularity of items, which are derived by summing the rows and columns of the \emph{big matrix}. \myfig{ab} shows the reasonable results: user sensitivity, i.e., $\alpha_u$, is proportional to the user activity, i.e., an active user is easier to get bored when viewing overexposed videos. In addition, item unendurableness, i.e., $\beta_i$, is proportional to the item popularity, i.e., popular videos are less endurable when they are overexposed.
Furthermore, we investigate the effect of different combinations of $\tau$ (in \myeq{ee}) and $\tau^*$ (in \myeq{eestar}) on the cumulative satisfactions. CIRS with $\tau=0, \tau^*=0$ degenerates to CIRS w/o CI since both $e_t(u,i)$ and $e_t^*(u,i)$ become $0$, i.e., the modeling of user satisfaction will not take into account the overexposure effect. The results in \myfig{tau} demonstrate that suitable $\tau$-$\tau^*$ pairs indeed improve the performance compared to CIRS w/o . Note that the orders of magnitude of $\tau$ and $\tau^*$ differ greatly in KuaiEnv, because the unit of time in \myeq{ee} and \myeq{eestar} are different. The former uses second(s) in log data and the latter uses step(s) in the RL planning and evaluation stages.
|
\section{Introduction}\label{sec:intro}
This paper continues our study of the \emph{local Gross-Prasad conjecture} (\cite{MR1186476}\cite{MR1295124}). Let $(W,V)$ be a pair of quadratic spaces defined over a local field $F$ of characteristic zero, where $W$ is a subspace of $V$ and the orthogonal complement $W^\perp$ of $W$ in $V$ is split of odd dimension. Let $H={\mathrm{SO}}(W)\ltimes N$ be the subgroup of $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$ where $N$ is the unipotent radical of ${\mathrm{SO}}(V)$ defined by a full isotropic flag attached to $W^\perp$. Fix a generic character $\xi$ of $N(F)$ which uniquely extends to $H(F)$. For any irreducible admissible representation $\pi$ of $G(F)$, set
\begin{equation}\label{eq:intro:1}
m(\pi) = \dim {\mathrm{Hom}}_H(\pi,\xi).
\end{equation}
From \cite{agrsmut1}\cite{ggporiginal}\cite{MR3202559} over $p$-adic fields, and \cite{szmut1}\cite{MR2720228} over archimedean fields, it is known that
$$
m(\pi)\leq 1.
$$
The local Gross-Prasad conjecture speculates a refinement for the behavior of $m(\pi)$. The pure inner forms of ${\mathrm{SO}}(W)$ are parametrized by $H^1(F,{\mathrm{SO}}(W))\simeq H^1(F,H)$, which classifies quadratic spaces of the same dimension and discriminant as $W$ over $F$. For any ${\alpha}\in H^1(F,H)$, there is a pair of quadratic spaces $(W_{\alpha}, V_{\alpha} = W_{\alpha}\oplus^\perp W^\perp)$ that enjoys parallel properties as $(W,V)$. Moreover, the Langlands dual group of $G_{\alpha} = {\mathrm{SO}}(W_{\alpha})\times {\mathrm{SO}}(V_{\alpha})$, which is a pure inner form of $G$, is isomorphic to the Langlands dual group ${}^{L}G$ of $G$. For any local $L$-parameter $\varphi:{\mathcal {L}}_F\to {}^{L}G$ where ${\mathcal {L}}_F$ is the local Langlands (Weil-Deligne) group of $F$, let $\Pi^{G_{\alpha}}(\varphi)$ be the corresponding local $L$-packet of $G_{\alpha}$ (\cite{langlandsproblems}).
Following the work of D. Vogan (\cite{MR1216197}), one can introduce the so called \emph{Vogan $L$-packet} attached to the parameter $\varphi$,
$$
\Pi^{\mathrm{Vogan}}(\varphi) =
\bigsqcup_{{\alpha}\in H^1(F,G)}\Pi^{G_{\alpha}}(\varphi).
$$
It was conjectured by Vogan (\cite{MR1216197}), which is known over archimedean fields (\cite[Thm.~6.3]{MR1216197}), that after fixing a Whittaker datum for $\{G_{\alpha}\}_{{\alpha}\in H^1(F,G)}$, there is a non-degenerate pairing
$$
\Pi^\mathrm{Vogan}(\varphi)\times {\mathcal {S}}_\varphi \to \{\pm 1\}.
$$
Here ${\mathcal {S}}_\varphi$ is the component group of $S_\varphi$, where $S_\varphi$ is the centralizer of the image $\Im(\varphi)$ in the dual group $\widehat{G}$. In particular, based on the conjecture, every $\pi\in \Pi^{\mathrm{Vogan}}(\varphi)$ corresponds uniquely to a character $\chi_\pi:{\mathcal {S}}_\varphi\to \{\pm 1\}$. On the other hand, it was suggested by Gross and Prasad that one may consider the following relevant Vogan $L$-packet attached to $\varphi$ and pairs of quadratic spaces $(W_{\alpha},V_{\alpha})_{{\alpha}\in H^1(F,H)}$,
$$
\Pi^{\mathrm{Vogan}}_\mathrm{rel}(\varphi) = \bigsqcup_{{\alpha}\in H^1(F,H)}\Pi^{G_{\alpha}}(\varphi).
$$
The following conjecture was formulated by B. Gross and D. Prasad.
\begin{conjec}[\cite{MR1186476}\cite{MR1295124}]\label{conjec:ggp:intro}
The following two statements hold.
\begin{enumerate}
\item
The following identity holds for $\varphi$ a generic local $L$-parameter,
$$
\sum_{\pi\in \Pi^{\mathrm{Vogan}}_{\mathrm{rel}}(\varphi)}m(\pi) = 1.
$$
\item
Fix the Whittaker datum for $\{G_{\alpha}\}_{{\alpha}\in H^1(F,G)}$ as in \cite[(6.3)]{MR1295124}. Based on (1), the unique representation $\pi_\varphi\in \Pi^\mathrm{Vogan}_\mathrm{rel}(\varphi)$ with the property that $m(\pi_\varphi) = 1$ enjoys the following characterization:
$$
\chi_{\pi_\varphi} =\chi_\varphi
$$
where $\chi_\varphi$ is defined in \eqref{eq:chivphicharacter:1}.
\end{enumerate}
\end{conjec}
When $F$ is $p$-adic, Conjecture \ref{conjec:ggp:intro} for tempered local $L$-parameters was proved by Waldspurger (\cite{waldspurger10}\cite{MR3155344}\cite{MR3155345}\cite{waldspurgertemperedggp}\cite{MR3202559}). C. Moeglin and Waldspurger proved the conjecture for generic local $L$-parameters based on the tempered case (\cite{MR3155346}).
When $F$ is archimedean, Part (1) of Conjecture \ref{conjec:ggp:intro} for tempered local $L$-parameters was proved by the second author in his thesis (\cite{thesis_zhilin}), which follows the strategy of Waldspurger and R. Beuzart-Plessis (\cite{beuzart2015local}). In \cite{chen2021local}, the first author reduced the proof of Conjecture \ref{conjec:ggp:intro} for generic local $L$-parameters to tempered local $L$-parameters over ${\mathbb {R}}$, following the work of Moeglin and Waldspurger, and H. Xue (\cite{xue1besselgeneric}). In particular over ${\mathbb {C}}$, Conjecture \ref{conjec:ggp:intro} was fully established in \cite{chen2021local}. Note that Part (1) of Conjecture \ref{conjec:ggp:intro} over ${\mathbb {C}}$ was also established in \cite{mollers2017symmetry} when the codimension of $W$ in $V$ is one.
Hence to finish the proof of Conjecture \ref{conjec:ggp:intro}, it remains to establish the following theorem, which is the main result of this paper:
\begin{thm}\label{thm:toprove}
Over ${\mathbb {R}}$, Part (2) of Conjecture \ref{conjec:ggp:intro} holds for tempered local $L$-parameters.
\end{thm}
The proof follows the strategy of \cite{MR3155345}. First let us briefly recall the proof of \cite{MR3155345} over $p$-adic fields. By Part (1) of Conjecture \ref{conjec:ggp:intro}, to prove Theorem \ref{thm:toprove}, it suffices to show that for any $s\in {\mathcal {S}}_\varphi$,
\begin{equation}\label{eq:intro:2}
\sum_{\pi\in \Pi^\mathrm{Vogan}_{\mathrm{rel}}(\varphi)}
\chi_{\pi}(s)\cdot m(\pi) = \chi_{\varphi}(s).
\end{equation}
In \cite{waldspurger10}\cite{waldspurgertemperedggp}, Waldspurger established a geometric multiplicity formula for any tempered representation $\pi$ of $G(F)$, which expresses the multiplicity $m(\pi)$ via the germ expansion of the distribution character of $\pi$ (\cite{SDPS99}). In particular \eqref{eq:intro:2} can be replaced as
\begin{equation}\label{eq:intro:3}
\sum_{\pi\in \Pi^\mathrm{Vogan}_{\mathrm{rel}}(\varphi)}
\chi_{\pi}(s)\cdot m_\mathrm{geom}(\pi) = \chi_{\varphi}(s).
\end{equation}
By ordinary endoscopy (\cite{ar13}, see also \cite[\S1.6]{MR3155345}), for any $s\in {\mathcal {S}}_\varphi$ there is an endoscopic group $G_{1,s}\times G_{2,s}$ of $G$ with $G_{i,s}$ coming from a pair of quadratic spaces $(W_i,V_i)$ in Gross-Prasad conjecture, and $\varphi|_{{}^L(G_1\times G_2)} = \varphi_1\times \varphi_2$. Based on the endoscopic character relation,
\begin{equation}\label{eq:intro:3}
\sum_{\pi\in \Pi^\mathrm{Vogan}_{\mathrm{rel}}(\varphi)}
\chi_{\pi}(s)\cdot m_\mathrm{geom}(\pi) =
m^S_\mathrm{geom}(\varphi_1)\cdot m^S_\mathrm{geom}(\varphi_2)
\end{equation}
which was established in \cite[Prop.~3.3]{MR3155345}. Here $m^S_\mathrm{geom}$ is a stable variant of $m_\mathrm{geom}$ (\cite[\S 3.2]{MR3155345}). Then from the theory of twisted endoscopy and the twisted character relation (\cite{MR1687096}\cite{ar13}, see also \cite[\S 1.6,\S 1.8]{MR3155345}), $m^S_\mathrm{geom}$ is connected with $\widetilde{m}_\mathrm{geom}$ (\cite[Prop.~3.4]{MR3155345}) where the latter is the geometric multiplicity attached to a pair of twisted general linear groups of Gross-Prasad type, which turns out to be equal to $\chi_\varphi(s)$ by \cite{MR3155344}.
It turns out that over ${\mathbb {R}}$, by Theorem \ref{thm:classifyparameters}, besides the situation when the underlying quadratic space is of dimension $\leq 3$, the tempered local $L$-parameters of special orthogonal groups are either of parabolic induction type, or of endoscopic induction type, which is simplier than $p$-adic fields. Hence we do not need to use twisted endoscopy. In particular, we do not need to establish the archimedean analogue of \cite{MR3155344}. However, there are the following two issues: First, there are only two pairs $(W_{\alpha},V_{\alpha})_{{\alpha}\in H^1(F,H)}$ over $p$-adic fields, which is not the case over ${\mathbb {R}}$. Second, the proof of \cite[Lem.~13.4~(ii)]{waldspurger10}, which is a necessary ingredient in the work of Waldspurger, relies on \cite[Conj.~1.2]{waldtransfert}. The parallel result seems unknown over ${\mathbb {R}}$. We solve the first problem through working with families of pairs $(W_{\alpha},V_{\alpha})_{{\alpha}\in H^1(F,H)}$ indexed by the Kottwitz sign of $G_{\alpha}$ (\cite{kottsign}), and the second by a formula of W. Rossmann (\cite{MR508985}).
Finally it is worth pointing out that as formulated by W. Gan, Gross and Prasad in \cite{ggporiginal}, there are similar conjectures for unitary groups, which was treated by Beuzart-Plessis (\cite{raphaelpadic}\cite{MR3251763}) over $p$-adic fields for tempered local $L$-parameters following the work of Waldspurger, and Gan and A. Ichino over $p$-adic fields for generic local $L$-parameters (\cite{MR3573972}) following the work of Moeglin-Waldspurger. Beuzart-Plessis solved Part (1) of Conjecture \ref{conjec:ggp:intro} for temperd local $L$-parameters of unitary groups over ${\mathbb {R}}$ in \cite{beuzart2015local}, generalizing the work of Waldspurger. Xue solved Part (2) of Conjecture \ref{conjec:ggp:intro} for tempered local $L$-parameters of unitary groups over ${\mathbb {R}}$ in \cite{xue1bessel}, and reduced Conjecture \ref{conjec:ggp:intro} from generic local $L$-parameters to tempered local $L$-parameters of unitary groups over ${\mathbb {R}}$ in \cite{xue1besselgeneric}. The special orthogonal and unitary groups cases discussed above are usually called the \emph{Bessel case}. There are also parallel conjectures called the \emph{Fourier-Jacobi case} formulated in \cite{ggporiginal}, which involve Weil representations and treat (skew-hermitian) unitary groups and symplectic-metaplectic groups. Over $p$-adic fields, they were resolved by Gan and Ichino for (skew-hermitian) unitary groups (\cite{MR3573972}), and H. Atobe for symplectic-metaplectic groups (\cite{MR3788848}), using techniques from theta correspondence reducing Fourier-Jacobi case to Bessel case.
\subsubsection{Organization}
Throughout the paper, we will always work over ${\mathbb {R}}$.
In section \ref{sec:localgp}, we fix notation and conventions used in the paper, and recall the statement of local Gross-Prasad conjecture formulated in \cite{MR1186476}\cite{MR1295124}.
In section \ref{sec:regconj}, we first review the parameterization of certain regular semi-simple conjugacy classes in special orthogonal groups following \cite[\S 1.3]{MR2672539} and \cite[\S1.3~\S 1.4]{MR3155345}, then establish Proposition \ref{pro:unionpureinnerconjclassof}, which is analogous to the description of the fiber of $p_G$ in \cite[\S 1.4]{MR3155345} giving a characterization of the union of these regular semi-simple conjugacy classes over pure inner forms of special orthogonal groups with fixed Kottwitz sign.
In section \ref{sec:geomultiplicity}, we first review the geometric multiplicity formula established in \cite{thesis_zhilin}. Then we establish Lemma \ref{lem:thecalculationofj}, which is analogous to \cite[Lem.~13.4~(ii)]{waldspurger10} over ${\mathbb {R}}$. As a corollary we establish Corollary \ref{cor:germfunaslimitofquasichar}, which is analogous to \cite[\S 13.6]{waldspurger10} expressing the germs of a distribution character by the distribution character itself. Then we recall the stable variant of geometric multiplicity formula introduced in \cite[\S 3.2]{MR3155345}, and establish Lemma \ref{lem:fiberoverkappappunionoverpure} describing the union of the support of the geometric multiplicity formula over pure inner forms of Gross-Prasad triples with fixed Kottwitz sign.
In section \ref{sec:redtoendoscopypf}, we first classify the tempered local $L$-parameters of special orthogonal groups over ${\mathbb {R}}$, then we complete the proof of Theorem \ref{thm:toprove} via mathematical induction and \cite[Prop.~7.4]{MR1295124}.
\newline
\newline
\textbf{Acknowledgement.}
We would like to thank D. Jiang suggesting the problem and helpful comments. We would also like to thank C. Wan for helpful discussion and crucial suggestions on Theorem \ref{thm:classifyparameters}. The work of this paper is supported in part by the Research Assistantship from the NSF grant DMS-1901802. Z. Luo would also like to thank the Department of Mathematics, University of Chicago, for the research fund through the
Dickson Instructorship. Finally we would like to thank the anonymous referee for helpful comments and suggestions.
\section{Local Gross-Prasad conjecture}\label{sec:localgp}
In this section, we review the local Gross-Prasad conjecture following \cite{MR1186476}\cite{MR1295124}.
\subsection{Gross-Prasad triples}\label{subsec:gptriples}
In this subsection, we recall the notion of Gross-Prasad triples following \cite[\S 6]{thesis_zhilin}.
Let $(W,V)$ be a pair of non-degenerate quadratic spaces over ${\mathbb {R}}$. The pair $(W,V)$ is called \textbf{admissible} if there exists an anisotropic line $D$ and a split non-degenerate quadratic space $Z$ of dimension $2r$ over ${\mathbb {R}}$ such that
$
V\simeq W\oplus^\perp D\oplus^\perp Z.
$
Let $q$ be the quadratic form on $V$. There exists a basis $\{z_i\}_{i=\pm 1}^{\pm r}$ of $Z$, such that
$$
q(z_i,z_j) = {\delta}_{i,-j},\quad \text{$\forall$ $i,j\in \{\pm 1,...,\pm r$\}}
$$
Let $N$ be the unipotent radical of the parabolic subgroup of ${\mathrm{SO}}(V)$ stabilizing the following totally isotropic flag
$
\langle z_r\rangle
\subset \langle z_r,z_{r-1}\rangle \subset...
\subset \langle z_r,...,z_1\rangle.
$
Set $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$. The group ${\mathrm{SO}}(W)$ can be identified as an algebraic subgroup of $G$ via diagonal embedding. In particular ${\mathrm{SO}}(W)$ acts on $N$ via adjoint action. Set
$
H={\mathrm{SO}}(W)\ltimes N.
$
Define a morphism ${\lambda}:N\to {\mathbb {G}}_a$ via
$$
{\lambda}(n) = \sum_{i=0}^{r-1}q(z_{-i-1},nz_i),\quad n\in N.
$$
Then ${\lambda}$ is ${\mathrm{SO}}(W)$-conjugation invariant and hence admits a unique extension to $H$ trivial on ${\mathrm{SO}}(W)$, which is still denoted as ${\lambda}$. Let ${\lambda}_{\mathbb {R}}:H({\mathbb {R}})\to {\mathbb {R}}$ be the induced morphism on ${\mathbb {R}}$-rational points. Fix an additive character $\psi$ of ${\mathbb {R}}$ and set
$$
\xi(h) = \psi({\lambda}_{\mathbb {R}}(h)),\quad h\in H({\mathbb {R}}).
$$
\begin{defin}\label{defin:ggptriple}
The triple $(G,H,\xi)$ is called the \textbf{Gross-Prasad triple} associated to the admissible pair $(W,V)$.
\end{defin}
\subsection{Vogan $L$-packets}\label{subsec:vgLpacket}
In this subsection, we recall the notion of Vogan $L$-packets of special orthogonal groups over ${\mathbb {R}}$ following \cite[\S 3]{MR1186476} and \cite{MR1216197}.
First let us recall the local Langlands correspondence over ${\mathbb {R}}$ following the exposition of \cite{knapp55local}. Let ${\mathcal {L}}_{\mathbb {R}}$ be the Weil group of ${\mathbb {R}}$. From \cite{langlandsclassify}, for any reductive algebraic group $G$ over ${\mathbb {R}}$, there is a bijective correspondence between local $L$-parameters $\varphi$ for $G$ and local $L$-packets $\Pi^G(\varphi)$ consisting of a finite set of irreducible Casselman-Wallach (\cite{MR1013462}\cite{MR1170566}) representations of $G$. A local $L$-parameter for $G$, by definition, is a $\widehat{G}$-conjugacy class of \textbf{admissible} homomorphisms
$$
\varphi:{\mathcal {L}}_{\mathbb {R}}\to {}^{L}G
$$
such that the elements in the image $\Im(\varphi)\subset {}^{L}G$ are semi-simple. Here $\widehat{G}$, resp. ${}^{L}G$ is the dual, resp. Langlands dual group of $G$. In particular, $\varphi$ is called \textbf{tempered} if $\Im(\varphi)$ is bounded.
The pure inner forms of $G$ share the same dual, resp. Langlands dual group as $G$. In particular, for any local $L$-parameter $\varphi$ of $G$, it can also be viewed as a local $L$-parameter of any pure inner forms of $G$. Following \cite{MR1216197}, instead of working with a single $L$-packet $\Pi^G(\varphi)$, one should work with the attached Vogan $L$-packet, which is the disjoint union of the $L$-packets for all pure inner forms of $G$,
$$
\bigsqcup_{G^\prime}\Pi^{G^\prime}(\varphi).
$$
Here $G^\prime$ runs over the isomorphism classes of pure inner forms of $G$.
Now let us return to special orthogonal groups situation. For a non-degenerate quadratic space $(V,q)$ over ${\mathbb {R}}$, the isomorphism classes of pure inner forms of ${\mathrm{SO}}(V)$ are classified by the set $H^1({\mathbb {R}},{\mathrm{SO}}(V))$, which in particular classifies quadratic spaces over ${\mathbb {R}}$ of the same dimension and discriminant as $V$ (\cite[\S 8]{MR1295124}). It is known that quadratic spaces $V$ over ${\mathbb {R}}$ are classified by their signature $(p,q)$ with $p,q\in {\mathbb {Z}}_{\geq 0}$, where $p = \mathrm{PI}(V)$ is the positive index of $V$, and $q=\mathrm{NI}(V)$ is the negative index of $V$.
The discriminant of the quadratic space $V$ with signature $(p,q)$ is given by
\begin{equation}\label{eq:discriminantdef}
{\mathrm{disc }} (V)= (-1)^{\floor{\frac{\dim V}{2}}}\cdot (-1)^q\in \{\pm 1\}\simeq {\mathbb {R}}^{\times}/{\mathbb {R}}^{\times 2}.
\end{equation}
Here $\floor{\frac{\dim V}{2}}$ is the maximal integer smaller than or equal to $\frac{\dim V}{2}$. Let the attached special orthogonal group be ${\mathrm{SO}}(p,q)$. By calculation, the pure inner forms of ${\mathrm{SO}}(p,q)$ are given by
\begin{equation}\label{eq:pureinnerofSOpq}
{\mathrm{SO}}(p_{\alpha},q_{\alpha})\qquad \text{with}\qquad \text{$p_{\alpha}+q_{\alpha} = p+q$ and $p\equiv p_{\alpha} \mod 2$}.
\end{equation}
Among all the pure inner forms of ${\mathrm{SO}}(V)$, there is a particular class called \textbf{quasi-split} (resp. \textbf{split}) pure inner forms, which are pure inner forms admitting Borel subgroups (resp. maximal split torus) defined over ${\mathbb {R}}$. The following statements are not hard to justify:
\begin{num}
\item\label{conjec:pureinner:1}
When $\dim V$ is odd, or even and ${\mathrm {P}}{\mathrm {I}}(V)-{\mathrm {N}}{\mathrm {I}}(V)\equiv 0\mod 4$, ${\mathrm{SO}}(V)$ has a unique quasi-split pure inner form that is split over ${\mathbb {R}}$;
\item\label{conjec:pureinner:2} When $\dim V$ is even and
$
{\mathrm {P}}{\mathrm {I}}(V)-{\mathrm {N}}{\mathrm {I}}(V)\equiv 2\mod 4,
$
${\mathrm{SO}}(V)$ has two quasi-split pure inner forms ${\mathrm{SO}}(p+2,p)$ and ${\mathrm{SO}}(p,p+2)$ with $p=\frac{\dim V}{2}-1$. Note that ${\mathrm{SO}}(p+2,p)$ and ${\mathrm{SO}}(p,p+2)$ are isomorphic as inner forms, but not as pure inner forms.
\end{num}
Two admissible pairs $(W,V)$ and $(W^\prime,V^\prime)$ are called \textbf{relevant} if
\begin{equation}\label{eq:defofrelevant}
\dim W=\dim W^\prime,\quad {\mathrm{disc }} (W)={\mathrm{disc }} (W^\prime),\quad \dim V=\dim V^\prime
,
\quad
{\mathrm{disc }}(V)={\mathrm{disc }}(V^\prime).
\end{equation}
Fix an admissible pair $(W,V)$ with attached Gross-Prasad triple $(G,H,\xi)$. For any ${\alpha}\in {\mathrm {H}}^1({\mathbb {R}},{\mathrm{SO}}(W))$ there is a unique relevant admissible pair $(W_{\alpha},V_{\alpha} = W_{\alpha}\oplus^\perp W^\perp)$ with Gross-Prasad triple $(G_{\alpha},H_{\alpha},\xi_{\alpha})$ attached to it. For any local $L$-parameter $\varphi :{\mathcal {L}}_{\mathbb {R}}\to {}^{L}G$, define the \textbf{relevant} Vogan $L$-packet as follows
\begin{equation}\label{eq:relvgpacket}
\Pi^{\mathrm{Vogan}}_\mathrm{rel}(\varphi) =
\bigsqcup_{{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W))}
\Pi^{G_{\alpha}}(\varphi).
\end{equation}
\begin{rmk}\label{rmk:uniqueqsggp}
For an admissible pair $(W,V)$, when $\dim W$ (resp. $\dim V$) is even, as is shown in \eqref{conjec:pureinner:2}, it can happen that ${\mathrm{SO}}(W)$ (resp. ${\mathrm{SO}}(V)$) has two non-isomorphic quasi-split pure inner forms. But since $\dim V$ (resp. $\dim W$) is odd, there is a \textbf{unique} admissible pair $(W_\mathrm{qs},V_\mathrm{qs})$ that is relevant to $(W,V)$, so that ${\mathrm{SO}}(W_\mathrm{qs})\times {\mathrm{SO}}(V_\mathrm{qs})$ is quasi-split.
\end{rmk}
\subsection{The conjecture}\label{subsec:conjecofggp}
In this subsection, we first review the construction of the distinguished character defined in \cite[\S 10]{MR1186476}, then recall the conjecture of Gross and Prasad formulated in \cite{MR1186476}\cite{MR1295124}. We restrict ourselves to ${\mathbb {R}}$.
From \cite[Thm.~6.3]{MR1216197}, the following statement holds:
\begin{num}
\item\label{num:shelstadvogandual}
Given a reductive algebraic group $G$ over ${\mathbb {R}}$ and a local $L$-parameter $\varphi:{\mathcal {L}}_{\mathbb {R}}\to {}^{L}G$, let ${\mathcal {S}}_{\varphi} = \pi_0(S_{\varphi})$ where $S_{\varphi}$ is the centralizer of the image of $\varphi$ in $\widehat{G}$ and ${\mathcal {S}}_{\varphi} = \pi_0(S_{\varphi})$ is the corresponding connected component. After fixing a Whittaker datum for $G$ (i.e. a quasi-split pure inner form of $G$ with a fixed Borel subgroup over ${\mathbb {R}}$, and a generic character of the corresponding unipotent radical), there is a non-degenerate pairing
$$
\Pi^{\mathrm{Vogan}}(\varphi)\times {\mathcal {S}}_\varphi \to \{\pm 1\}.
$$
\end{num}
Returning to special orthogonal groups situation, for a non-degenerate quadratic space $V$ over ${\mathbb {R}}$ with a local $L$-parameter $\varphi_V$ of ${\mathrm{SO}}(V)$, as is suggested in \cite[\S 6,\S7]{MR1186476}, after composing $\varphi_V$ with the standard embedding of ${}^L({\mathrm{SO}}(V))$ into the general linear group ${\mathrm{GL}}({\mathrm {M}}_V)$, one can get a local $L$-parameter $\mathrm{std}\circ \varphi_V:{\mathcal {L}}_{\mathbb {R}}\to {\mathrm{GL}}({\mathrm {M}}_V)$ which determines a non-degenerate bilinear form ${\mathrm {B}}:{\mathrm {M}}_V\times {\mathrm {M}}_V\to {\mathbb {C}}$ preserved under the action of ${\mathcal {L}}_{\mathbb {R}}$ with sign ${\epsilon} = \{\pm 1\}$. When $\dim V$ is odd (resp. even), the bilinear form ${\mathrm {B}}$ is symplectic (resp. symmetric), and hence ${\epsilon}=-1$ (resp. $=1$).
By the semi-simplicity of $\varphi_V$,
$
{\mathrm {M}}_V = \bigoplus_i m_i\cdot {\mathrm {M}}_{i,V}
$
where ${\mathrm {M}}_{i,V}$ are inequivalent irreducible representations of ${\mathcal {L}}_{\mathbb {R}}$ with multiplicity $m_i$.
\begin{num}
\item\label{num:classifycomponentgroup}
Following \cite[Prop.~6.5, Prop.~7.6]{MR1186476} the irreducible representations of ${\mathrm {M}}_{i,V}$ are classified as follows (here ${\mathrm {M}}^\vee_{i,V}$ is the contragredient of ${\mathrm {M}}_{i,V}$):
\begin{enumerate}
\item[(${\mathrm {O}}$-type)] ${\mathrm {M}}_{i,V}\simeq {\mathrm {M}}_{i,V}^\vee$ where the isomorphism is provided by a non-degenerate pairing with sign ${\epsilon}$;
\item[(${\mathrm{Sp}}$-type)] ${\mathrm {M}}_{i,V}\simeq {\mathrm {M}}^\vee_{i,V}$ where the isomorphism is provided by a non-degenerate pairing with sign $-{\epsilon}$. Moreover, in this case, $m_i$ is even;
\item[(${\mathrm{GL}}$-type)] ${\mathrm {M}}_{i,V}\not\simeq {\mathrm {M}}^\vee_{i,V}$, and hence ${\mathrm {M}}_{i,V}\simeq {\mathrm {M}}_{j,V}^\vee$ for some $i\neq j$ and $m_i=m_j$.
\end{enumerate}
\end{num}
Let ${\mathrm {I}}_{\mathrm {O}}$, ${\mathrm {I}}_{\mathrm{Sp}}$ and ${\mathrm {I}}_{\mathrm{GL}}$ be the index set of of the irreducible (resp. direct sum of two irreducible) representations ${\mathrm {M}}_{i,V}$ (resp. ${\mathrm {M}}_{i,V}\oplus {\mathrm {M}}_{i,V}^\vee$) of ${\mathrm {O}}$-type or ${\mathrm{Sp}}$-type (resp. ${\mathrm{GL}}$-type). Following \cite[Prop.~6.6, Prop.~7.7]{MR1186476}, the centralizer of the image of $\varphi_V$ is equal to
\begin{equation}\label{eq:SvphiVdescription}
S_{\varphi_V}
=
\bigg(
\prod_{i\in {\mathrm {I}}_{\mathrm {O}}}{\mathrm {O}}(m_i,{\mathbb {C}})
\bigg)_+
\times
\prod_{i\in {\mathrm {I}}_{\mathrm{Sp}}}{\mathrm{Sp}}(m_i,{\mathbb {C}})
\times
\prod_{i\in {\mathrm {I}}_{\mathrm{GL}}}{\mathrm{GL}}(m_i,{\mathbb {C}})
\end{equation}
where
$
\bigg(
\prod_{i\in {\mathrm {I}}_{\mathrm {O}}}{\mathrm {O}}(m_i,{\mathbb {C}})
\bigg)_+
=
S
\bigg(
\prod_{
\substack{
i\in {\mathrm {I}}_{\mathrm {O}}\\
\dim {\mathrm {M}}_{i,V} \text{ odd}
}
}
{\mathrm {O}}(m_i,{\mathbb {C}})
\bigg)
\times
\prod_{
\substack{
i\in {\mathrm {I}}_{\mathrm {O}}
\\
\dim {\mathrm {M}}_{i,V}\text{ even}
}
}
{\mathrm {O}}(m_i,{\mathbb {C}})
$
and the component group ${\mathcal {S}}_{\varphi_V}$ can be computed as follows
\begin{equation}\label{eq:component}
{\mathcal {S}}_{\varphi_V}
=
\Bigg\{
\begin{matrix}
({\mathbb {Z}}/2{\mathbb {Z}})^r & \text{ if $\dim {\mathrm {M}}_{i,V}$ are even for all ${\mathrm {M}}_{i,V}$ of ${\mathrm {O}}$-type} \\
({\mathbb {Z}}/2{\mathbb {Z}})^{r-1} & \text{ otherwise}
\end{matrix}
\end{equation}
where $r = |{\mathrm {I}}_{\mathrm {O}}|$.
In \cite[\S10]{MR1186476}, Gross and Prasad defined a distinguished character. To make it more precise, let $(G,H,\xi)$ be a Gross-Prasad triple attached to an admissible pair $(W,V)$ over ${\mathbb {R}}$. From \eqref{num:shelstadvogandual}, after fixing a Whittaker datum for $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$, there is a non-degenerate pairing
$
\Pi^{\mathrm{Vogan}}(\varphi)\times {\mathcal {S}}_\varphi\to \{\pm 1\}
$
for any local $L$-parameter $\varphi = \varphi_W\times \varphi_V$ of $G$. Therefore for every $\pi\in \Pi^\mathrm{Vogan}_\varphi$, there is a unique character $\chi_\pi:{\mathcal {S}}_\varphi\to \{\pm 1\}$ corresponding to it. In \cite[\S 10]{MR1186476}, Gross and Prasad defined the following character $\chi_\varphi = \chi^V_{\varphi_W}\times \chi^W_{\varphi_V}$ of ${\mathcal {S}}_\varphi = {\mathcal {S}}_{\varphi_W}\times {\mathcal {S}}_{\varphi_V}$: For every element $s=s_W\times s_V\in {\mathcal {S}}_{\varphi_W}\times {\mathcal {S}}_{\varphi_V}$, set
\begin{align}\label{eq:chivphicharacter:1}
\chi^W_{\varphi_V}(s_V) &=
\det(-{\mathrm{Id}}_{{\mathrm {M}}_V^{s_V = -1}})^{\frac{\dim {\mathrm {M}}_W}{2}}
\cdot
\det(-{\mathrm{Id}}_{{\mathrm {M}}_W})^{\frac{\dim {\mathrm {M}}_V^{s_V=-1}}{2}}
\cdot
{\varepsilon}\big(\frac{1}{2},{\mathrm {M}}_V^{s_V=-1}\otimes {\mathrm {M}}_W,\psi\big)
\\
\chi^V_{\varphi_W}(s_W)&=
\det(-{\mathrm{Id}}_{{\mathrm {M}}_W^{s_W = -1}})^{\frac{\dim {\mathrm {M}}_V}{2}}
\cdot
\det(-{\mathrm{Id}}_{{\mathrm {M}}_V})^{\frac{\dim {\mathrm {M}}_W^{s_W=-1}}{2}}
\cdot
{\varepsilon}\big(\frac{1}{2},{\mathrm {M}}_W^{s_W=-1}\otimes {\mathrm {M}}_V,\psi\big)
\nonumber
\end{align}
Here ${\mathrm {M}}^{s_V=-1}_{V}$ is the $s_V=(-1)$-eigenspace of ${\mathrm {M}}_V$ and ${\varepsilon}(...)$ is the local root number defined by Rankin-Selberg integral (\cite{MR2533003}).
Now let us recall the conjecture of Gross and Prasad. Let $\pi$ be an irreducible Casselman-Wallach representation of $G({\mathbb {R}})$. Set
\begin{equation}\label{eq:multiplicity}
m(\pi) = \dim {\mathrm{Hom}}_{H({\mathbb {R}})}(\pi,\xi).
\end{equation}
Following \cite{szmut1}\cite{MR2720228}, it is known that
$$
m(\pi)\leq 1.
$$
The local Gross-Prasad conjecture studies the refinement behavior of the multiplicity $m(\pi)$ in a relevant Vogan $L$-packet.
\begin{conjec}\label{conjec:gp}
Let $(G,H,\xi)$ be a Gross-Prasad triple attached to an admissible pair $(W,V)$ over ${\mathbb {R}}$. Fix a generic local $L$-parameter $\varphi$ of $G$. Then the following statements hold:
\begin{enumerate}
\item
There exists a unique member $\pi_\varphi\in \Pi^\mathrm{Vogan}_{\mathrm{rel}}(\varphi)$ such that $m(\pi_\varphi) = 1$;
\item
Fix the Whittaker datum for $G$ as \cite[(6.3)]{MR1295124}. Based on \eqref{num:shelstadvogandual}, the character $\chi_{\pi_\varphi}:{\mathcal {S}}_\varphi\to \{\pm 1\}$ attached to the unique member $\pi_\varphi$ is equal to $\chi_\varphi$ defined in \eqref{eq:chivphicharacter:1}.
\end{enumerate}
\end{conjec}
When $\varphi$ is a tempered local $L$-parameter, the second author proved Part (1) of the conjecture in \cite{thesis_zhilin} following the work of Waldspurger (\cite{waldspurger10}\cite{MR3155345}) and Beuzart-Plessis (\cite{beuzart2015local}). Following the work of \cite{MR3155346}, the first author reduced Conjecture \ref{conjec:gp} for generic local $L$-parameters to tempered local $L$-parameters in \cite{chen2021local}. Therefore in order to finish the proof of Conjecture \ref{conjec:gp} over ${\mathbb {R}}$, it remains to establish the following theorem, which is the main result of this paper.
\begin{thm}\label{thm:main}
Over ${\mathbb {R}}$, Part (2) of Conjecture \ref{conjec:gp} holds for tempered local $L$-parameters.
\end{thm}
\section{Some regular semi-simple conjugacy classes}\label{sec:regconj}
In this section, we review the parameterization of certain regular semi-simple conjugacy classes in special orthogonal groups following \cite[\S 1.3]{MR2672539} and \cite[\S 1.3~\S 1.4]{MR3155345}. The Lie algebra analogue is also considered in \cite{wald01nilpotent} (see also \cite[\S 5.1]{thesis_zhilin}). In subsection \ref{subsec:unionoverpure}, we establish Proposition \ref{pro:unionpureinnerconjclassof} which describes the union of the parameterization over pure inner forms of a special orthogonal group with a fixed Kottwitz sign.
\subsection{Parameterization}\label{subsec:parmetrization}
In this subsection, we review the parameterization of certain regular semi-simple conjugacy classes in special orthogonal groups following \cite[\S 1.3]{MR2672539} and \cite[\S 1.3~\S 1.4]{MR3155345}.
\begin{num}
\item\label{num:parameterization:datum}
Consider the following datum:
\begin{itemize}
\item A finite set $I$;
\item For any $i\in I$, fix a finite extension $F_{\pm i}$ of ${\mathbb {R}}$ together with a $2$-dimensional $F_{\pm i}$ commutative algebra $F_i$. Let $\tau_i$ be the unique nontrivial automorphism of $F_i$ over $F_{\pm i}$;
\item For any $i\in I$, fix a constant $u_i\in F_i^\times$ such that $u_i\cdot \tau_i(u_i) = 1$.
\end{itemize}
\end{num}
Let $\underline{\Xi}$ be the set of quadruples $\kappa = (I,(F_{\pm i})_{i\in I}, (F_i)_{i\in I},(u_i)_{i\in I})$ satisfying \eqref{num:parameterization:datum}. For two quadruples in $\underline{\Xi}$
$$
\kappa = (I,(F_{\pm i})_{i\in I}, (F_i)_{i\in I}, (u_i)_{i\in I}),\quad
\kappa^\prime =
(I^\prime, (F^\prime_{\pm i})_{i\in I^\prime}, (F^\prime_i)_{i\in I^\prime}, (u^\prime_i)_{i\in I^\prime}),
$$
$\kappa$ is called isomorphic to $\kappa^\prime$ if there exists a triple $(\iota, (\iota_{\pm i})_{i\in I}, (\iota_i)_{i\in I})$ such that the following properties hold:
\begin{itemize}
\item $\iota:I\to I^\prime$ is a bijective map;
\item For any $i\in I$, $\iota_{\pm i}:F_{\pm i}\to F^\prime_{\pm\iota(i)}$ and $\iota_i:F_i\to F^\prime_{\iota(i)}$ are compatible isomorphisms;
\item $\iota_i(u_i) = u^\prime_{\iota(i)}$.
\end{itemize}
A quadruple $\kappa\in \underline{\Xi}$ is called \textbf{regular} if the identity map is the only automorphism of $\kappa$. Let $\Xi_{\mathrm{reg}}$ be the set of isomorphism classes of regular quadruples.
For any even positive integer $d$, let
$
\Xi_{{\mathrm{reg}},d}
$
be the subset of $\Xi_{\mathrm{reg}}$ consisting of quadruples $\kappa = (I,(F_{\pm i})_{i\in I}, (F_i)_{i\in I}, (u_i)_{i\in I})\in \Xi_{\mathrm{reg}}$ satisfying
$
\sum_{i}[F_i:{\mathbb {R}}] = d.
$
For any $\kappa=(I,(F_{\pm i})_{i\in I}, (F_i)_{i\in I}, (u_i)_{i\in I})\in \Xi_{\mathrm{reg}}$ and $i\in I$, let ${\mathrm{sgn}}_{F_i/F_{\pm i}}$ be the quadratic character of $F_{\pm i}$ associated to $F_i$. Let $I^* = I^*_\kappa$ be the subset of $I$ consisting of $i\in I$ such that $F_i$ is a field, i.e. ${\mathrm{sgn}}_{F_i/F_{\pm i}}$ is nontrivial. Let
$$
C(\kappa) = \prod_{i\in I}F^{\times}_{\pm i}/\mathrm{Norm}_{F_i/F_{\pm i}}(F^\times_i) \simeq
\prod_{i\in I^*}\{\pm 1\}.
$$
Fix $c= (c_i)_{i\in I}\in C(\kappa)$. For a pair $(\kappa,c)\in \Xi_{\mathrm{reg}}\times C(\kappa)$, one may associate to it an even dimensional non-degenerate quadratic space $(W_{\kappa,c},q_{\kappa,c})$ over ${\mathbb {R}}$ with $W_{\kappa,c} = \bigoplus_{i\in I}F_i$ and
\begin{equation}\label{eq:quadkappac:1}
q_{\kappa,c}
\bigg(
\sum_{i\in I}w_i,
\sum_{i\in I}w^\prime_i
\bigg)
=
\sum_{i\in I}{\mathrm{tr}}_{F_i/{\mathbb {R}}}
(\tau_i(w_i)w^\prime_i c_i),\quad w_i,w^\prime_i\in F_i.
\end{equation}
Here we implicitly identify $c_i\in F^{\times}_{\pm i}/\mathrm{Norm}_{F_i/F_{\pm i}}(F^\times_i)$ as one of its representatives in $F_{\pm i}$. It turns out that since $\tau_i(c_i) = c_i$, $q_{\kappa,c}$ is indeed symmetric. Up to isomorphism, the quadratic space $(W_{\kappa,c},q_{\kappa,c})$ is independent of the choice of the representatives of $c$ (\cite[\S 1.3]{MR2672539}).
It turns out that the positive and negative index of $(W_{\kappa,c},q_{\kappa,c})$ can be computed as follows.
\begin{lem}\label{lem:positivenegativeindex}
Let
$$
{\mathrm {I}}^\pm_{{\mathbb {C}}} = \{i\in I\mid F_i\simeq {\mathbb {C}}, c_i = \pm\},
\quad
{\mathrm {I}}_{{\mathbb {R}}\oplus {\mathbb {R}}} = \{i\in I\mid F_i\simeq {\mathbb {R}}\oplus {\mathbb {R}}\},
\quad
{\mathrm {I}}_{{\mathbb {C}}\oplus {\mathbb {C}}} = \{i\in I\mid F_i\simeq {\mathbb {C}}\oplus {\mathbb {C}}\}.
$$
Then the following identities hold:
\begin{align*}
\mathrm{PI}(W_{\kappa,c}) &=
2\cdot |{\mathrm {I}}^{+}_{\mathbb {C}}|
+|{\mathrm {I}}_{{\mathbb {R}}\oplus {\mathbb {R}}}|
+2\cdot |{\mathrm {I}}_{{\mathbb {C}}\oplus {\mathbb {C}}}|
\\
\mathrm{NI}(W_{\kappa,c}) &=
2\cdot |{\mathrm {I}}^{-}_{\mathbb {C}}|
+|{\mathrm {I}}_{{\mathbb {R}}\oplus {\mathbb {R}}}|
+2\cdot |{\mathrm {I}}_{{\mathbb {C}}\oplus {\mathbb {C}}}|.
\end{align*}
\end{lem}
\begin{proof}
For any $i\in I$, from \eqref{num:parameterization:datum}, there are the following situations:
\begin{enumerate}
\item $F_{\pm i} = {\mathbb {R}}$, $F_i = {\mathbb {C}}$. The quadratic form is given as follows
$$
{\mathrm{tr}}_{{\mathbb {C}}/{\mathbb {R}}}(\tau_i(w_i)w^\prime_i c_i),\quad w_i,w^\prime_i\in {\mathbb {C}}
$$
which has positive index $2$ (resp. $0$) and negative index $0$ (resp. $2$) if $c_i = 1 \text{ (resp. $-1$})\in \{\pm 1\} \simeq F_{\pm i}^\times/\mathrm{Norm}(F_i^\times)$;
\item $F_{\pm i} = {\mathbb {R}}$, $F_i = {\mathbb {R}}\oplus {\mathbb {R}}$. The quadratic form is given as follows
$$
{\mathrm{tr}}_{F_i/{\mathbb {R}}}
(\tau_i(w_i^1,w^2_i)\cdot(w^{\prime 1}_{i},w^{\prime 2}_i))
=
w_i^1w^{\prime 2}_i+w_i^2w^{\prime 1}_i,\quad w^{1}_i,w^2_i,w^{\prime 1}_i,w^{\prime 2}_i\in {\mathbb {R}}
$$
which has positive index $1$ and negative index $1$;
\item $F_{\pm i} = {\mathbb {C}}$, $F_i = {\mathbb {C}}\oplus {\mathbb {C}}$. The quadratic form is given by
$$
{\mathrm{tr}}_{F_i/{\mathbb {R}}}
(\tau_i(w_i^1,w^2_i)\cdot(w^{\prime 1}_{i},w^{\prime 2}_i)c_i)
=
c_i\cdot
{\mathrm{tr}}_{{\mathbb {C}}/{\mathbb {R}}}
\big(
w_i^1w^{\prime 2}_i
+
w_i^2w^{\prime 1}_i
\big)
,\quad
w^{1}_i,w^2_i,w^{\prime 1}_i,w^{\prime 2}_i\in {\mathbb {C}}
$$
which has positive index $2$ and negative index $2$.
\end{enumerate}
It follows that the lemma has been proved.
\end{proof}
\begin{rmk}\label{rmk:changeindextosign}
From Lemma \ref{lem:positivenegativeindex}, for $\kappa = (I,(F_{\pm i})_{i\in I}, (F_i)_{i\in I}, (u_i)_{i\in I})\in \Xi_{{\mathrm{reg}},d}$ and $c\in C(\kappa)$, the isomorphism class of the quadratic space $(W_{\kappa,c},q_{\kappa,c})$ is essentially determined by the cardinality of ${\mathrm {I}}^\pm_{\mathbb {C}}$. Note that $
{\mathrm {I}}^+_{\mathbb {C}}\bigsqcup {\mathrm {I}}^-_{\mathbb {C}} = I^*.$ Therefore the cardinality of ${\mathrm {I}}^\pm_{\mathbb {C}}$ is uniquely determined by the difference
$$
|{\mathrm {I}}^+_{\mathbb {C}}|-|{\mathrm {I}}^-_{\mathbb {C}}| =
\sum_{i\in I^*}c_i,\quad c=(c_i)_{i\in I}\in C(\kappa)\simeq
\prod_{i\in I^*}\{\pm 1\}.
$$
For any $\theta\in \big\{-|I^*|, -|I^*|+2,...,|I^*|-2,|I^*|\big\}$, set
$$
C(\kappa)_{\theta} = \big\{
c=(c_i)\in C(\kappa)\mid \sum_{i\in I^*}c_i = \theta
\big\}.
$$
Then for $c,c^\prime\in C(\kappa)$, $(W_{\kappa,c},q_{\kappa,c})\simeq (W_{\kappa,c^\prime},q_{\kappa,c^\prime})$ if and only if there exists $\theta$ such that
$
\sum_{i\in I^*}c_i = \sum_{i\in I^*}c^\prime_i=\theta.
$
When it is the case, by straightforward calculation,
\begin{equation}\label{eq:indexofWkappac}
\mathrm{PI}(W_{\kappa,c}) =
\frac{d}{2}+\theta,
\quad
\mathrm{NI}(W_{\kappa,c}) =
\frac{d}{2}-\theta.
\end{equation}
\end{rmk}
Let $x_{\kappa,c}\in {\mathrm{GL}}(W_{\kappa,c})$ be the element defined by
\begin{equation}\label{eq:quadkappac:2}
x_{\kappa,c}(
\sum_{i\in I}w_i
)
=\sum_{i\in I}u_iw_i,\quad w_i\in F_i.
\end{equation}
Then one can verify that $x_{\kappa,c}\in {\mathrm{SO}}(W_{\kappa,c})$ by the identity $u_i\cdot \tau_i(u_i)=1$.
\begin{defin}\label{defin:parconj}
Let $(V,q)$ be a non-degenerate quadratic space over ${\mathbb {R}}$ and set ${\Delta}_V = \mathrm{PI}(V)-\mathrm{NI}(V)$. Define the following notions:
\begin{enumerate}
\item
When $\dim V$ is even, set
$$
\Xi_{{\mathrm{reg}},V} =
\big\{
(\kappa,c)\mid \kappa\in \Xi_{{\mathrm{reg}},\dim V}, c\in C(\kappa)_{\frac{{\Delta}_V}{2}}
\big\};
$$
\item When $\dim V$ is odd, set
\begin{align*}
\Xi_{{\mathrm{reg}},V} =
\big\{
(\kappa,c)\mid&
\kappa\in \Xi_{{\mathrm{reg}},\dim V-1}, c\in C(\kappa),
\text{ and there exists an anisotropic }
\\
&
\text{
line $(D_{\kappa,V},q_{\kappa,V})$ such that $(W_{\kappa,c},q_{\kappa,c})\oplus^\perp (D_{\kappa,V},q_{\kappa,V})\simeq (V,q)$}
\big\}.
\end{align*}
\end{enumerate}
\end{defin}
\begin{rmk}\label{rmk:afterdefparconj}
When $\dim V$ is odd, through comparing ${\mathrm{disc }}(V)$ and ${\mathrm{disc }}(W_{\kappa,c})$, the signature of $D_{\kappa,c}$ is independent of $c\in C(\kappa)$, and is equal to
\begin{equation}\label{eq:definoffraki}
\mathfrak{i}_{V,\kappa} = (-1)^{
\frac{-{\Delta}_V+1}{2}
+|I^*_\kappa|
}
\quad
\text{ where }
{\Delta}_V = \mathrm{PI}(V)-\mathrm{NI}(V)
.
\end{equation}
Note that ${\mathfrak {i}}_{V,\kappa}$ depends only on $\kappa$ and the pure inner class of $V$.
After comparing ${\Delta}_{W_{\kappa,c}\oplus^\perp D_{\kappa,V}}$ with $
{\Delta}_V$,
$$
\sum_{i\in I^*}c_i = \frac{{\Delta}_V-{\mathfrak {i}}_{V,\kappa}}{2}.
$$
Therefore the following identity holds
\begin{equation}\label{eq:xiregValternative}
\Xi_{{\mathrm{reg}},V}
=
\big\{
(\kappa,c)\mid
\kappa\in \Xi_{{\mathrm{reg}},\dim V-1},c\in C(\kappa)_{\frac{{\Delta}_V-{\mathfrak {i}}_{V,\kappa}}{2}}
\big\}
\end{equation}
\end{rmk}
Following \cite[\S 1.3]{MR2672539} and \cite[\S 1.3,\S1.4]{MR3155345}, or mimicking the proofs from \cite[\S 5.1]{thesis_zhilin}, the following theorem holds.
\begin{thm}\label{thm:parconjso}
Let $(V,q)$ be a non-degenerate quadratic space over ${\mathbb {R}}$. Let ${\mathrm{SO}}(V)^\mathrm{rss}/\sim$ be the set of regular semi-simple conjugacy classes in ${\mathrm{SO}}(V)$. When $\dim V$ is even, let ${\mathrm{SO}}(V)^\mathrm{rss}_{\neq \pm 1}/\sim$ be the subset of ${\mathrm{SO}}(V)^\mathrm{rss}/\sim$ without eigenvalue $\pm 1$.
\begin{enumerate}
\item When $\dim V$ is even, there is a two-to-one map
$$
{\mathrm{SO}}(V)^\mathrm{rss}_{\neq \pm 1}/\sim
\longrightarrow
\Xi_{{\mathrm{reg}},V}
$$
To make it more precise, for $(\kappa,c)\in \Xi_{{\mathrm{reg}}, V}$, let $x_{\kappa,c}\in{\mathrm{SO}}(W_{\kappa,c})$ as in \eqref{eq:quadkappac:2}. Note that $x_{\kappa,c}$ does not have eigenvalue $\pm 1$ since $\kappa$ is regular. Through comparing the signature from \eqref{eq:indexofWkappac},
$
(W_{\kappa,c},q_{\kappa,c})\simeq (V,q).
$
Then $x_{\kappa,c}$ determines an $\mathrm{O}(V)$-conjugacy class in ${\mathrm{SO}}(V)$ which breaks into $2$ different ${\mathrm{SO}}(V)$-conjugacy classes consisting of different elements. Denote them as $x_{\kappa,c}^\pm$;
\item When $\dim V$ is odd, there is a one-to-one map
$$
\Xi_{{\mathrm{reg}},V}
\longleftrightarrow
{\mathrm{SO}}(V)^\mathrm{rss}/\sim
$$
To make it more precise, for $(\kappa,c)\in
\Xi_{{\mathrm{reg}},V}$, with
$
(W_{\kappa,c},q_{\kappa,c})\oplus^\perp (D_{\kappa,V},q_{\kappa,V}) \simeq (V,q),
$
let $x^{D_{\kappa,V}}_{\kappa,c}={\mathrm{Id}}_{D_{\kappa,V}}\oplus x_{\kappa,c}\in {\mathrm{SO}}(D_{\kappa,V}\oplus^\perp W_{\kappa,c})$ where $x_{\kappa,c}$ is constructed in \eqref{eq:quadkappac:2}. Then $x^{D_{\kappa,V}}_{\kappa,c}$ determines a unique ${\mathrm{SO}}(V)$-conjugacy classes in ${\mathrm{SO}}(V)$.
\end{enumerate}
\end{thm}
\subsection{Union over pure inner forms with fixed Kottwitz sign}\label{subsec:unionoverpure}
In this subsection, we discuss the union of the parameterization established in Theorem \ref{thm:parconjso} over pure inner forms of a special orthogonal group with a fixed Kottwitz sign. The main result of this subsection is Proposition \ref{pro:unionpureinnerconjclassof}, which is analogous to the description of the fiber of $p_G$ considered in \cite[\S 1.4]{MR3155345}.
First let us recall the notion of Kottwitz sign defined in \cite{kottsign}.
\begin{defin}\label{defin:kottwitzsign}
Let $G$ be a reductive algebraic group over ${\mathbb {R}}$ with a fixed maximal compact subgroup $K$. Let $G_\mathrm{qs}$ be the unique quasi-split inner form of $G$ with a maximal compact subgroup $K_\mathrm{qs}$. The Kottwitz sign of $G$ is defined as
$$
e(G) = (-1)^{\frac{\dim K_\mathrm{qs}-\dim K}{2}}.
$$
\end{defin}
By straightforward calculation, the Kottwitz sign of ${\mathrm{SO}}(p,q)$ is given as follows.
\begin{lem}\label{lem:kottwitzsignsopq}
$$
e({\mathrm{SO}}(p,q)) =
\Bigg\{
\begin{matrix}
1 & p+q \text{ even}\\
(-1)^{\frac{(p-q)^2-1}{8}} & p+q \text{ odd}
\end{matrix}
$$
\end{lem}
Based on \eqref{eq:pureinnerofSOpq}, the following corollary holds by straightforward calculation.
\begin{cor}\label{cor:pureinnerformsopq}
For any ${\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V) = {\mathrm{SO}}(p,q))$ corresponding to a non-degenerate quadratic space $V_{\alpha}$ with signature $(p_{\alpha},q_{\alpha})$, the following statements hold.
\begin{enumerate}
\item If $p+q$ is odd, then
$$
e({\mathrm{SO}}(V)) = e({\mathrm{SO}}(V_{\alpha}))
\Longleftrightarrow
p \equiv p_{\alpha}\mod 4;
$$
\item If $p+q$ is even, then after fixing an anisotropic line $D$,
$$
e({\mathrm{SO}}(V\oplus^\perp D)) =
e({\mathrm{SO}}(V_{\alpha}\oplus^\perp D))
\Longleftrightarrow p\equiv p_{\alpha} \mod 4.
$$
\end{enumerate}
\end{cor}
Now we are ready to state the main proposition in this subsection.
\begin{pro}\label{pro:unionpureinnerconjclassof}
For $\kappa\in \Xi_{{\mathrm{reg}},d}$, set
$
C(\kappa)^{\pm 1} = \big\{c\in C(\kappa)\mid \prod_{i\in I^*}c_i=\pm 1 \big\}.
$
Then for any non-degenerate quadratic space $V$ over ${\mathbb {R}}$ and $e_0 = \pm 1$, the following identities hold.
\begin{enumerate}
\item
If $\dim V$ is odd, then
$$
\bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))
\\
e({\mathrm{SO}}(V_{\alpha})) = e_0
}
}
\Xi_{{\mathrm{reg}},V_{\alpha}}
=
\Xi_{{\mathrm{reg}},\dim V-1,e_0}.
$$
Here
\begin{align*}
\Xi_{{\mathrm{reg}},\dim V-1,e_0} &=
\big\{
(\kappa,c)\in \Xi_{{\mathrm{reg}},\dim V-1},c\in C(\kappa)^{e_0\cdot {\epsilon}_{V,\kappa}}
\big\},
\\
{\epsilon}_{V,\kappa} &= (-1)^{\frac{
-\mathrm{PI}(V_\mathrm{qs})
+
|I_\kappa^*|+
\big(
\frac{\dim V+{\mathfrak {i}}_{V,\kappa}}{2}
\big)
}{2}}.
\end{align*}
$\mathrm{PI}(V_\mathrm{qs})$ is the positive index of the unique quasi-split pure inner form of ${\mathrm{SO}}(V)$;
\item If $\dim V$ is even, fix an anisotropic line $D$ with signature $\mathrm{sig}(D)\in \{\pm 1\}$. Then
$$
\bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))
\\
e({\mathrm{SO}}(V_{\alpha}\oplus^\perp D)) = e_0
}
}\Xi_{{\mathrm{reg}},V_{\alpha}}
=
\Xi_{{\mathrm{reg}},\dim V,e_0,D}.
$$
Here
\begin{align*}
\Xi_{{\mathrm{reg}},\dim V,e_0,D}
=&\big\{
(\kappa,c)\mid
\kappa\in \Xi_{{\mathrm{reg}},\dim V}, c\in C(\kappa)^{e_0\cdot {\epsilon}_{V,\kappa,D}}
\big\},
\\
{\epsilon}_{V,\kappa,D} = &
(-1)^{\frac{
|I^*_\kappa|
+
\frac{\dim V+1+\mathrm{sig}(D)}{2}
-\mathrm{PI}(V,D)
}{2}}.
\end{align*}
$\mathrm{PI}(V,D)$ is the positive index of the unique quasi-split pure inner form of ${\mathrm{SO}}(V\oplus^\perp D)$.
\end{enumerate}
\end{pro}
\begin{proof}
First let us assume that $\dim V$ is odd. Fix ${\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))$ and $e({\mathrm{SO}}(V_{\alpha})) = e_0$. For any $(\kappa,c)\in \Xi_{{\mathrm{reg}},V_{\alpha}}$, from Remark \ref{rmk:afterdefparconj} there exists an anisotropic line $(D_{\kappa,V_{\alpha}},q_{\kappa,V_{\alpha}})$ with signature ${\mathfrak {i}}_{V_{\alpha},\kappa}$
such that
$
(V_{\alpha},q_{\alpha})\simeq (D_{\kappa,V_{\alpha}},q_{\kappa,V_{\alpha}})
\oplus^\perp
(W_{\kappa,c},q_{\kappa,c}).
$
From \eqref{eq:xiregValternative},
$$
\Xi_{{\mathrm{reg}},V_{\alpha}} =
\big\{
(\kappa,c)\mid
\kappa\in \Xi_{{\mathrm{reg}},\dim V_{\alpha}-1},c\in C(\kappa)_{\frac{{\Delta}_{V_{\alpha}}-{\mathfrak {i}}_{V_{\alpha},\kappa}}{2}}
\big\}.
$$
In general, for $\kappa\in \Xi_{{\mathrm{reg}},\dim V-1}$ and $c\in C(\kappa)$, by direct calculation,
\begin{equation}\label{eq:directcalmod4}
\prod_{i\in I^*}c_i = 1 \text{ (resp. }-1)
\Longleftrightarrow
\bigg(|I_\kappa^*|
-
\sum_{i\in I_\kappa^*}c_i\bigg) \equiv 0 \text{ (resp. }2) \mod 4.
\end{equation}
Equivalently,
\begin{equation}\label{eq:proofunion:1}
\prod_{i\in I_\kappa^*}c_i = (-1)^{\frac{|I_\kappa^*|-\sum_{i\in I^*}c_i}{2}}.
\end{equation}
In particular, when $(\kappa,c)\in \Xi_{{\mathrm{reg}},V_{\alpha}}$,
$
\sum_{i\in I^*_\kappa}c_i = \frac{{\Delta}_{V_{\alpha}}-{\mathfrak {i}}_{V_{\alpha},\kappa}}{2}.
$
Hence
$$
\prod_{i\in I^*_\kappa}c_i =
(-1)^{\frac{|I_\kappa^*|-
\big(
\frac{{\Delta}_{V_{\alpha}}-{\mathfrak {i}}_{V_{\alpha},\kappa}}{2}
\big)
}{2}}
$$
Let $(p_\mathrm{qs},q_\mathrm{qs})$ be the signature of the unique quasi-split pure inner form ${\mathrm{SO}}(V_\mathrm{qs})$ of ${\mathrm{SO}}(V_{\alpha})$. From Corollary \ref{cor:pureinnerformsopq},
$
e_0=e({\mathrm{SO}}(V_{\alpha})) = (-1)^{\frac{p_{\alpha}-p_\mathrm{qs}}{2}}
$
where $(p_{\alpha},q_{\alpha})$ is the signature of $V_{\alpha}$. Therefore $e_0\cdot \prod_{i\in I^*_\kappa}
c_i $ is equal to
\begin{align*}
=
(-1)^{\frac{p_{\alpha}-p_\mathrm{qs}}{2}}
\cdot
(-1)^{\frac{|I_\kappa^*|-
\big(
\frac{{\Delta}_{V_{\alpha}}-{\mathfrak {i}}_{V_{\alpha},\kappa}}{2}
\big)
}{2}}
=
(-1)^{\frac{
p_{\alpha}-p_\mathrm{qs}
+
|I_\kappa^*|-
\big(
\frac{{\Delta}_{V_{\alpha}}-{\mathfrak {i}}_{V_{\alpha},\kappa}}{2}
\big)
}{2}}.
\end{align*}
Since ${\mathrm{SO}}(V_{\alpha})$ is a pure inner form of ${\mathrm{SO}}(V)$,
$
{\Delta}_{V_{\alpha}}\equiv {\Delta}_V\mod 4.
$
Therefore
\begin{align*}
=
(-1)^{\frac{
p_{\alpha}-p_\mathrm{qs}
+
|I_\kappa^*|-
\big(
\frac{{\Delta}_{V_{\alpha}}-{\mathfrak {i}}_{V,\kappa}}{2}
\big)
}{2}}
=
(-1)^{\frac{
-p_\mathrm{qs}
+
|I_\kappa^*|-
\big(
\frac{-\dim V-{\mathfrak {i}}_{V,\kappa}}{2}
\big)
}{2}}=
{\epsilon}_{V,\kappa}.
\end{align*}
It follows that
$$
\bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))
\\
e({\mathrm{SO}}(V_{\alpha})) = e_0
}
}
\Xi_{{\mathrm{reg}},V_{\alpha}}
\subset \Xi_{{\mathrm{reg}},\dim V-1,e_0}.
$$
Conversely, for $(\kappa,c)\in \Xi_{{\mathrm{reg}},\dim V-1,e_0}$ and ${\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))$ with $e({\mathrm{SO}}(V_{\alpha})) = e_0$, by \eqref{eq:proofunion:1}
$$
\sum_{i\in I_\kappa^*}c_i \equiv
\frac{{\Delta}_{V_{\alpha}}-
{\mathfrak {i}}_{V_{\alpha},\kappa}
}{2} \mod 4.
$$
Equivalently,
$$
(-1)^{
\frac{-{\Delta}_{V}+1}{2}
+|I^*_\kappa|}+
2\cdot \sum_{i\in I_\kappa^*}c_i \equiv
{\Delta}_{V_{\alpha}}
\mod 8.
$$
Now
$
\big|(-1)^{
{\mathfrak {i}}_{V,\kappa}}+
2\cdot \sum_{i\in I_\kappa^*}c_i
\big|
\leq 1+2\cdot
\sum_{i\in I^*_\kappa}1
\leq \dim V
$. As ${\alpha}$ runs over $H^1({\mathbb {R}},{\mathrm{SO}}(V))$ with $e({\mathrm{SO}}(V_{\alpha})) = e_0$, ${\Delta}_{V_{\alpha}}$ runs over all the integers between $-\dim V$ and $\dim V$ with fixed congruence class modulo $8$. Hence there exists ${\alpha}_0\in H^1({\mathbb {R}},{\mathrm{SO}}(V))$ such that
$$
{\Delta}_{V_{{\alpha}_0}} = (-1)^{
{\mathfrak {i}}_{V,\kappa}}+
2\cdot \sum_{i\in I_\kappa^*}c_i
\Longleftrightarrow \sum_{i\in I_\kappa^*}c_i =
\frac{{\Delta}_{V_{{\alpha}_0}}-(-1)^{
{\mathfrak {i}}_{V,\kappa}
}}{2}.
$$
It follows that
$$
\Xi_{{\mathrm{reg}},\dim V-1,e_0}\subset \bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))
\\
e({\mathrm{SO}}(V_{\alpha}))=e_0
}
}
\Xi_{{\mathrm{reg}},V_{\alpha}}
$$
and Part (1) of the proposition is established.
It remains to treat the case when $\dim V$ is even. The proof is similar to odd situation. Fix ${\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))$ with $e({\mathrm{SO}}(V_{\alpha}\oplus^\perp D)) = e_0$. For any $(\kappa,c)\in \Xi_{{\mathrm{reg}}, V}$, with the same argument as odd case,
$
\prod_{i\in I^*_\kappa}c_i =
(-1)^{\frac{|I^*_\kappa|-\frac{{\Delta}_{V_{\alpha}}}{2}}{2}}.
$
Let ${\mathrm{SO}}(V_{\alpha}\oplus^\perp D) = {\mathrm{SO}}(p_{\alpha},q_{\alpha})$ with unique quasi-split pure inner form ${\mathrm{SO}}(p_\mathrm{qs},q_\mathrm{qs})$. Then
$e({\mathrm{SO}}(V_{\alpha}\oplus^\perp D)) = (-1)^{\frac{p_{\alpha}-p_\mathrm{qs}}{2}}$. Using the fact that ${\Delta}_{V_{\alpha}} = p_{\alpha}-q_{\alpha}-\mathrm{sig}(D)$, we get
$$
e_0\cdot
\prod_{i\in I^*_\kappa}c_i
=
(-1)^{\frac{|I^*_\kappa|-\frac{{\Delta}_{V_{\alpha}}}{2}+p_{\alpha}-p_\mathrm{qs}}{2}}
=
(-1)^{
\frac{|I^*_\kappa|+\frac{\dim V+1+\mathrm{sig}(D)}{2}-p_\mathrm{qs}}{2}
}
$$
It follows that
$$
\bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))
\\
e({\mathrm{SO}}(V_{\alpha}\oplus^\perp D)) = e_0
}
}\Xi_{{\mathrm{reg}},V}
\subset
\Xi_{{\mathrm{reg}},\dim V,e_0,D}.
$$
For the converse inclusion the same proof as odd case applies verbatim.
\end{proof}
\begin{rmk}\label{rmk:afterpro:unionpureinnerconjclassof}
From Remark \ref{rmk:afterdefparconj}, ${\mathfrak {i}}_{V,\kappa}$ depends only on $\kappa$ and the pure inner class of $V$. It follows that ${\epsilon}_{V,\kappa}$ (resp. ${\epsilon}_{V,\kappa,D}$) defined in Part (1) (resp. (2)) of Proposition \ref{pro:unionpureinnerconjclassof} depends only on $\kappa$ and Kottwitz sign of $V$ (resp. $V\oplus^\perp D$).
\end{rmk}
\section{Geometric multiplicity formula}\label{sec:geomultiplicity}
In this section, for a Gross-Prasad triple $(G,H,\xi)$ attached to an admissible pair $(W,V)$ over ${\mathbb {R}}$, we recall the geometric multiplicity formula for tempered representations of $G({\mathbb {R}})$ established in \cite{thesis_zhilin}. We also establish Lemma \ref{lem:thecalculationofj} and Corollary \ref{cor:germfunaslimitofquasichar}, which are analogous to \cite[Lem.~13.4~(ii)]{waldspurger10} and a formula in \cite[\S 13.6]{waldspurger10}. In subsection \ref{subsec:stablevariant} we recall the stable variant of the geometric multiplicity which was first introduced in \cite{MR3155345}.
\subsection{The formula}\label{subsec:theformula}
In this subsection, we recall the geometric multiplicity formula established in \cite{thesis_zhilin} for tempered representations of $G({\mathbb {R}})$. For details see \cite[\S 7.3]{thesis_zhilin}.
\subsubsection{Geometric support}
For $x\in H_\mathrm{ss}({\mathbb {R}})$ where $H_\mathrm{ss}({\mathbb {R}})$ is the set of semi-simple elements in $H({\mathbb {R}})$, up to conjugation by $H({\mathbb {R}})$ there is no harm to assume that $x\in {\mathrm{SO}}(W)_\mathrm{ss}({\mathbb {R}})$. Let ${\Gamma}(H)$ be the set of semi-simple conjugacy classes in $H({\mathbb {R}})$. Let $W^\prime_x = \ker(1-x|_{W})$, $V^\prime_x = \ker(1-x|_V)$ and $W^{\prime\p}_x = \Im(1-x|_{W})$. Then $W = W^\prime_x\oplus W^{\prime\p}_x$ and $V=V^\prime_x\oplus W^{\prime\p}_x$. Moreover $(V^\prime_x,W^\prime_x)$ is also an admissible pair over ${\mathbb {R}}$. Let $G_x$ be the connected component of the centralizer of $x$ in $G$. Then following \cite[\S7.3.1]{thesis_zhilin},
\begin{equation}\label{eq:formulaGxdecomp}
G_x = G^\prime_x\times G^{\prime\p}_x,\quad \text{ with } \quad G^\prime_x = {\mathrm{SO}}(W^\prime_x)\times {\mathrm{SO}}(V^\prime_x), \quad G^{\prime\p}_x = {\mathrm{SO}}(W^{\prime\p})_x
\times {\mathrm{SO}}(W^{\prime\p})_x.
\end{equation}
Let ${\Gamma}(G,H)$ be the subset of ${\Gamma}(H)$ consisting of $x\in {\Gamma}(G,H)$ satisfying the condition that ${\mathrm{SO}}(W^{\prime\p}_x)_x$ is an anisotropic torus and $G_x$ is quasi-split. Following \cite[(7.3.3)]{thesis_zhilin}, this set ${\Gamma}(G,H)$ is equipped with topology and measure.
\subsubsection{The germ $c_\theta$}
For a reductive algebraic group $G$ over ${\mathbb {R}}$ and $\Theta$ a quasi-character on $G({\mathbb {R}})$ (see \cite[\S 4.4]{beuzart2015local} for the definition), from \cite[Prop.~4.4~(vi)]{beuzart2015local}, the following asymptotic expansion holds for any $x\in G_\mathrm{ss}({\mathbb {R}})$ and $Y\in {\mathfrak {g}}_x({\mathbb {R}}) = ({\mathrm{Lie}} G_x)({\mathbb {R}})$ sufficiently close to $0$,
$$
D^G(xe^Y)^{1/2}\cdot \Theta(xe^Y)
=
D^G(xe^Y)^{1/2}\cdot
\sum_{{\mathcal {O}}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)}c_{\Theta,{\mathcal {O}}}
(x)\cdot \widehat{j}({\mathcal {O}},Y)
+O(|Y|).
$$
Here $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)$ is the set of regular nilpotent orbits in ${\mathfrak {g}}_x({\mathbb {R}})$, $c_{\Theta,{\mathcal {O}}}(x)\in {\mathbb {C}}$, and $\widehat{j}({\mathcal {O}},\cdot)$ is the Fourier transform of the nilpotent orbital integral attached to ${\mathcal {O}}$.
Let $(V,q)$ be a non-degenerate quadratic space over ${\mathbb {R}}$. Recall that $(V,q)$ is called \textbf{quasi-split} (see \cite[\S~6.1.1]{thesis_zhilin} for instance) if
\begin{align}\label{eq:qsquadraticsp}
(V,q)
\simeq
{\mathbb {H}}^{n-1}\oplus^\perp
\Bigg\{
\begin{matrix}
(D,q), & \dim V\equiv 1\mod 2\\
(E=F(\sqrt{b}),c\cdot {\mathrm {N}}_{E/F}), & \dim V\equiv 0\mod 2
\end{matrix}
\end{align}
for some positive integer $n$ and $b,c\in {\mathbb {R}}^\times$. Here ${\mathbb {H}}^{n-1}$ is the unique split quadratic space of dimension $2n-2$, $(D,q)$ is an anisotropic line, and $(E=F(\sqrt{b}),c\cdot {\mathrm {N}}_{E/F})$ is the $2$-dimensional quadratic space sending $m\oplus n\sqrt{b}\mapsto c\cdot (m^2-bn^2)$. Following \cite[\S~6.1.2]{thesis_zhilin}, the set of ${\mathrm{SO}}(V)({\mathbb {R}})$-regular nilpotent orbits in ${\mathfrak {s}}{\mathfrak {o}}(V)({\mathbb {R}})$, denoted as $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))$, has the following description:
\begin{enumerate}
\item $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))\neq \emptyset$ if and only if $(V,q)$ is quasi-split;
\item $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))$ contains a unique element if
\begin{itemize}
\item $\dim V$ is odd;
\item $\dim V\leq 2$;
\item $\dim V\geq 4$ is even and $(V,q)$ is quasi-split but not split;
\end{itemize}
\item When $(V,q)$ is split of even dimension with $\dim V\geq 4$, $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))$ contains two elements ${\mathcal {O}}_{\pm}$ parametrized by $\{\pm 1\}\simeq {\mathbb {R}}^{\times}/{\mathbb {R}}^{\times 2}$. For the explicit parameterization see \cite[\S 6.1.2]{thesis_zhilin}.
\end{enumerate}
Now for a Gross-Prasad triple $(G,H,\xi)$ attached to an admissible pair $(W,V)$ over ${\mathbb {R}}$, and $\Theta$ a quasi-character on $G({\mathbb {R}})$, by \eqref{eq:formulaGxdecomp} and the definition of ${\Gamma}(G,H)$, for any $x\in {\Gamma}(G,H)$, $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x) = \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}^\prime_x)$ with ${\mathfrak {g}}^\prime_x = {\mathfrak {s}}{\mathfrak {o}}(W^\prime_x)\times {\mathfrak {s}}{\mathfrak {o}}(V^\prime_x)$. Following \cite[\S7.3.2]{thesis_zhilin}, for any quasi-character $\Theta$ on $G({\mathbb {R}})$, define the following function $c_{\Theta}(x)$ on ${\Gamma}(G,H)$:
\begin{num}
\item\label{num:definofcgermfun}
\begin{enumerate}
\item If $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)$ is singleton with unique member ${\mathcal {O}}_{\mathrm{reg}}$, set $c_\Theta(x) = c_{\Theta,{\mathcal {O}}_{\mathrm{reg}}}(x)$;
\item If $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)$ has two elements (in particular both $W^\prime_x$ and $V^\prime_x$ are split), write the decomposition $V^\prime_x \simeq W^\prime_x\oplus^\perp D^\prime_x\oplus^\perp Z^\prime_x$ from the definition of admissible pair. Let $\mathrm{sig}(D^\prime_x)\in \{\pm 1\}$ be the signature of the line $D^\prime_x$.
\begin{itemize}
\item If $\dim V^\prime_x$ is even and $\geq 4$, choose ${\mathcal {O}}_{\mathrm{sig}(D^\prime_x)}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)\leftrightarrow \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V^\prime_x))$ and set $c_{\Theta}(x) = c_{\Theta,{\mathcal {O}}_{\mathrm{sig}(D^\prime_x)}}(x)$ ;
\item If $\dim W^\prime_x$ is even and $\geq 4$, choose ${\mathcal {O}}_{-\mathrm{sig}(D^\prime_x)}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)\leftrightarrow \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(W^\prime_x))$ and set $c_{\Theta}(x) = c_{\Theta,{\mathcal {O}}_{-\mathrm{sig}(D^\prime_x)}}(x)$.
\end{itemize}
\end{enumerate}
\end{num}
\subsubsection{Integral formula}
With the above notation, for $x\in {\Gamma}(G,H)$ set
$$
{\Delta}(x) = |\det(1-x)_{|_{W^{\prime\p}_x}}|,\quad
D^G(x) = |\det(1-{\mathrm{Ad}}(x))_{{\mathfrak {g}}/{\mathfrak {g}}_x}|.
$$
For any quasi-character $\Theta$ on $G({\mathbb {R}})$, consider the following integral
\begin{equation}\label{eq:geomultiplicity}
m_\mathrm{geom}(\Theta):=
\int_{{\Gamma}(G,H)}
D^G(x)^{1/2}c_\Theta(x){\Delta}(x)^{-1/2}\,\mathrm{d} x.
\end{equation}
From \cite[Prop.~7.3.3.3]{thesis_zhilin} the integral \eqref{eq:geomultiplicity} is absolutely convergent.
For any tempered representation $\pi$ of $G({\mathbb {R}})$, let $\Theta_\pi$ be the distribution character of $\pi$ (\cite{MR145006}) which in particular is a quasi-character (\cite{MR576644}). Set
$
m_\mathrm{geom}(\pi) :=m_\mathrm{geom}(\Theta_\pi).
$
The following theorem is established in \cite{thesis_zhilin}.
\begin{thm}\label{thm:luothesis}
For any tempered representation $\pi$ of $G({\mathbb {R}})$,
$$
m_\mathrm{geom}(\pi) = m(\pi).
$$
\end{thm}
\subsection{Special values of $\widehat{j}({\mathcal {O}},\cdot)$}\label{subsec:specialvalueregularnilporbitalintegral}
In this subsection, we establish the analogue of \cite[Lem.~13.4~(ii)]{waldspurger10} over ${\mathbb {R}}$. The result will be used to express $c_\Theta(x)$ via the original quasi-character $\Theta$.
Let $(V,q)$ be a split quadratic space of even dimension $\geq 4$ (so that $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))$ has two elements). Consider the following regular semi-simple element in ${\mathfrak {s}}{\mathfrak {o}}(V)({\mathbb {R}})$ (which is also considered in \cite[\S 5.3]{thesis_zhilin} following \cite{wald01nilpotent}):
\begin{num}
\item\label{num:specialvalue:1}
Fix two nonzero purely imaginary numbers $a_1,a_2\in i{\mathbb {R}}^\times$ such that $a_1\neq \pm a_2$. Fix an isomorphism of split quadratic spaces
$$
(V,q)\simeq ({\mathbb {C}},c\cdot \mathrm{Nr})\oplus^\perp ({\mathbb {C}},-c\cdot \mathrm{Nr})\oplus^\perp (\widetilde{Z},q)
$$
where $c=\pm 1$, $\mathrm{Nr} = \mathrm{Norm}_{{\mathbb {C}}/{\mathbb {R}}}$ and $(\widetilde{Z},q)$ is a split quadratic space of dimension $\dim V-4$. Let $\widetilde{T}$ be a fixed maximal split torus of ${\mathrm{SO}}(V)$ with Lie algebra $\widetilde{{\mathfrak {t}}}$, and fix a regular semi-simple element $\widetilde{S}\in \widetilde{{\mathfrak {t}}}({\mathbb {R}})$. Consider the element $X_{a_1,a_2,\widetilde{S}}\in {\mathfrak {s}}{\mathfrak {o}}(V)$ acting on $({\mathbb {C}},c\cdot \mathrm{Nr})$ via $a_1$, $({\mathbb {C}},-c\cdot \mathrm{Nr})$ via $a_2$ and $(\widetilde{Z},q)$ via $\widetilde{S}$. From \cite[Lem.~5.1.0.5]{thesis_zhilin} the conjugacy classes of $X_{a_1,a_2,\widetilde{S}}$ within the stable conjugacy class is determined by $c\in {\mathbb {R}}^{\times}/{\mathbb {R}}^{\times 2}$. Let $X^{\pm}_{a_1,a_2,\widetilde{S}}$ be the element corresponding to $c=c^\pm$ with
${\mathrm{sgn}}_{{\mathbb {C}}/{\mathbb {R}}}(c^\pm) = \pm{\mathrm{sgn}}_{{\mathbb {C}}/{\mathbb {R}}}(\mathrm{Nr}(a_1)-\mathrm{Nr}(a_2))$.
\end{num}
Let $\widehat{j}(X^{\pm}_{a_1,a_2,\widetilde{S}},\cdot)$ be the Fourier transform of the Lie algebra orbital integral at $X^{\pm}_{a_1,a_2,\widetilde{S}}$ (\cite[\S 1.9]{beuzart2015local}). Following \cite[Lem.~4.3.1]{beuzart2015local}, the following limit formula holds
\begin{align}\label{eq:limitforwhjXpm:1}
\lim_{t\in {\mathbb {R}}^{\times 2},t\mapsto 0^+}
D^G(tY)^{1/2}
\cdot
\widehat{j}(X^{\pm}_{a_1,a_2,\widetilde{S}},tY) = &
D^G(Y)^{1/2}
\cdot
\sum_{{\mathcal {O}}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))}
{\Gamma}_{\mathcal {O}}(X^{\pm}_{a_1,a_2,\widetilde{S}})\cdot \widehat{j}({\mathcal {O}},Y).
\end{align}
Here ${\Gamma}_{\mathcal {O}}(X^{\pm}_{a_1,a_2,\widetilde{S}})\in {\mathbb {C}}$ is the regular germs attached to ${\mathcal {O}}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))$ and $X^{\pm}_{a_1,a_2,\widetilde{S}}$, whose explicit formula is established in \cite[Thm.~4.2.0.1]{thesis_zhilin}.
The following lemma can be found in \cite[Lem.~5.3.0.1]{thesis_zhilin}.
\begin{lem}\label{lem:germformula}
\begin{enumerate}
\item
For ${\mathcal {O}}_{\pm}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V))$ with the parameterization in \cite[\S6.1.2]{thesis_zhilin},
$$
{\Gamma}_{{\mathcal {O}}_\pm}(X^+_{a_1,a_2,\widetilde{S}})
-{\Gamma}_{{\mathcal {O}}_\pm}(X^-_{a_1,a_2,\widetilde{S}}) =
\pm 1.
$$
\item
For any $X_\mathrm{qd}\in {\mathfrak {t}}_V$ that is regular semi-simple, where ${\mathfrak {t}}_V$ is the Lie algebra of a maximal split torus in ${\mathrm{SO}}(V)$,
$$
{\Gamma}_{{\mathcal {O}}_\pm}(X_\mathrm{qd}) = 1.
$$
\end{enumerate}
\end{lem}
The main result of this subsection is the following lemma, which is analogues to \cite[Lem.~13.4~(ii)]{waldspurger10} over ${\mathbb {R}}$.
\begin{lem}\label{lem:thecalculationofj}
With the above notation, for $\nu = \pm$,
$$
\widehat{j}({\mathcal {O}}_{\nu},X^+_{a_1,a_2,\widetilde{S}}) =
-\widehat{j}({\mathcal {O}}_{\nu},X^-_{a_1,a_2,\widetilde{S}})
=\nu\cdot \frac{|W_{T_\mathrm{cpt}}|}{2}\cdot D^G(X^+_{a_1,a_2,\widetilde{S}})^{-1/2}
$$
where $T_{\mathrm{cpt}}$ is the centralizer of $X^+_{a_1,a_2,\widetilde{S}}$ in ${\mathrm{SO}}(V)$ and $W_{T_\mathrm{cpt}} = N(T_\mathrm{cpt})/T_\mathrm{cpt}$.
\end{lem}
\begin{proof}
By \eqref{eq:limitforwhjXpm:1},
\begin{align*}
&\lim_{t\in {\mathbb {R}}^{\times 2},t\mapsto 0^+}
D^G(tY)^{1/2}
\cdot
\big(
\widehat{j}(X^+_{a_1,a_2,\widetilde{S}},tY)
-
\widehat{j}(X^-_{a_1,a_2,\widetilde{S}},tY)
\big)
\\
=&
D^G(Y)^{1/2}
\cdot
\sum_{\pm}
\big(
{\Gamma}_{{\mathcal {O}}_\pm}(X^+_{a_1,a_2,\widetilde{S}})
-
{\Gamma}_{{\mathcal {O}}_\pm}(X^-_{a_1,a_2,\widetilde{S}})
\big)
\cdot
\widehat{j}({\mathcal {O}}_\pm,Y).
\end{align*}
By Lemma \ref{lem:germformula}, the above formula is equal to
$$
D^G(Y)^{1/2}
\cdot
\big(
\widehat{j}({\mathcal {O}}_+,Y)
-\widehat{j}({\mathcal {O}}_-,Y)
\big).
$$
From \cite[(3.4.6)]{beuzart2015local}, since $X^\pm_{a_1,a_2,\widetilde{S}}$ are not split,
$$
\widehat{j}({\mathcal {O}}_+,X^\pm_{a_1,a_2,\widetilde{S}})-\widehat{j}({\mathcal {O}}_-,X^{\pm}_{a_1,a_2,\widetilde{S}}) =
2\cdot
\widehat{j}({\mathcal {O}}_+,X^\pm_{a_1,a_2,\widetilde{S}}).
$$
Hence
\begin{align*}
&\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+}}
\frac{D^G(tX^{\pm}_{a_1,a_2,\widetilde{S}})^{1/2}}{
D^G(X^{\pm}_{a_1,a_2,\widetilde{S}})^{1/2}
}
\cdot
\big(
\widehat{j}(X^+_{a_1,a_2,\widetilde{S}},tX^\pm_{a_1,a_2,\widetilde{S}})
-
\widehat{j}(X^-_{a_1,a_2,\widetilde{S}},tX^\pm_{a_1,a_2,\widetilde{S}})
\big)
=2\cdot
\widehat{j}({\mathcal {O}}_+,X^\pm_{a_1,a_2,\widetilde{S}}).
\end{align*}
On the other hand, following the last part of the proof of \cite[Lem.~13.4]{waldspurger10} and \cite[(3.4.4)]{beuzart2015local}, as a distribution, $\widehat{j}(X^{\pm}_{a_1,a_2,\widetilde{S}},\cdot)$ is the parabolic induction from $\widehat{j}(X^{\pm}_{a_1,a_2},\cdot)$ where $\widehat{j}(X^{\pm}_{a_1,a_2},\cdot)$ is the Fourier transform of the Lie algebra orbital integral at $X^{\pm}_{a_1,a_2,\widetilde{S}}$ on the Levi subgroup ${\mathrm{SO}}(2,2)\times \widetilde{T}$ of ${\mathrm{SO}}(V)$ (where ${\mathrm{SO}}(2,2)\simeq {\mathrm{SO}}(({\mathbb {C}},\mathrm{Nr})\oplus^\perp({\mathbb {C}},-\mathrm{Nr}))$. Using the same argument as the last paragraph of \cite[Lem.~13.4]{waldspurger10}, we can reduce the proof to $\dim V=4$ (and hence $\widetilde{Z} =0$) situation. Hence it suffices to evaluate the following limit when $\dim V=4$
\begin{align}\label{eq:lemmaformula:1}
\lim_{t\in {\mathbb {R}}^{\times 2},t\mapsto 0^+}
t^{\frac{\dim {\mathrm{SO}}(V)-\dim T_\mathrm{cpt}}{2}}
\cdot
\big(
\widehat{j}(X^+_{a_1,a_2},tX^\pm_{a_1,a_2})
-
\widehat{j}(X^-_{a_1,a_2},tX^\pm_{a_1,a_2})
\big).
\end{align}
This can be achieved by the formula of Rossmann in \cite[p.217~(15)]{MR508985}. Notice that as a real Lie group, $T_\mathrm{cpt}({\mathbb {R}})\simeq {\mathrm{SO}}(2)({\mathbb {R}})\times {\mathrm{SO}}(2)({\mathbb {R}})$ which is a maximal compact Cartan subgroup in ${\mathrm{SO}}(2,2)({\mathbb {R}})$ containing both $X^{\pm}_{a_1,a_2}$. In particular, since the maximal compact subgroup of ${\mathrm{SO}}(2,2)({\mathbb {R}})$ can be identified with $S({\mathrm {O}}(2)({\mathbb {R}})\times {\mathrm {O}}(2)({\mathbb {R}}))$, the Weyl group
$$
W_{T_\mathrm{cpt}} = N_{{\mathrm{SO}}(2,2)({\mathbb {R}})}(T_\mathrm{cpt}({\mathbb {R}}))/T_\mathrm{cpt}({\mathbb {R}}) \simeq {\mathbb {Z}}/2{\mathbb {Z}}.
$$
Let $W_{T_{\mathrm{cpt},{\mathbb {C}}}}$ be the Weyl group attached to $({\mathfrak {t}}_{\mathrm{cpt},{\mathbb {C}}},{\mathfrak {s}}{\mathfrak {o}}(4,{\mathbb {C}}))$ where ${\mathfrak {t}}_{\mathrm{cpt}}$ is the Lie algebra of $T_{\mathrm{cpt}}$ and ${\mathfrak {t}}_{\mathrm{cpt},{\mathbb {C}}}$ is its complexification. Then $W_{T_\mathrm{cpt}}$ embeds diagonally into the complex Weyl group $W_{T_{\mathrm{cpt},{\mathbb {C}}}}\simeq {\mathbb {Z}}/2{\mathbb {Z}}\times {\mathbb {Z}}/2{\mathbb {Z}}$. Following \cite[p.208]{MR508985}, fix once for all a system of positive roots for $({\mathfrak {t}}_{\mathrm{cpt},{\mathbb {C}}},{\mathfrak {s}}{\mathfrak {o}}(4,{\mathbb {C}}))$ and let $\pi:{\mathfrak {t}}_{\mathrm{cpt},{\mathbb {C}}}\to {\mathbb {C}}$ be the product of these positive roots. Notice that $\pi$ is equal to the square root of Weyl discriminant up to $\frac{\pi(\cdot)}{
|\pi(\cdot)|}$. In particular, one should note that the Lie algebra orbital integral defined in \cite[(1)]{MR508985} differs from ours (following \cite[\S 1.8]{beuzart2015local}) by $\frac{\pi(\cdot)}{
|\pi(\cdot)|}$.
Now by \cite[p.217~(15)]{MR508985}, after straightforward calculation, the following identity holds
$$
\frac{\pi(X^*_{a_1,a_2})}{
|\pi(X^*_{a_1,a_2})|
}
\cdot
\widehat{j}(X^*_{a_1,a_2},tX^{\pm}_{a_1,a_2}) =
\frac{-1}{|W_{T_\mathrm{cpt}}|}
\cdot
\pi(tX^\pm_{a_1,a_2})^{-1}
\cdot
\sum_{w\in W_{T_\mathrm{cpt}}}
(-1)^{\ell(w)}
\cdot
e^{i(w\cdot tX^{\pm}_{a_1,a_2},X^+_{a_1,a_2})}.
$$
where $* = \pm$ and $\ell(w)$ is the length of $w$. Since $W_{T_\mathrm{cpt}}\simeq {\mathbb {Z}}/2{\mathbb {Z}}$ embeds into $W_{T_{\mathrm{cpt},{\mathbb {C}}}}\simeq {\mathbb {Z}}/2{\mathbb {Z}}\times {\mathbb {Z}}/2{\mathbb {Z}}$ diagonally, we get
$$
\frac{\pi(X^*_{a_1,a_2})}{
|\pi(X^*_{a_1,a_2})|
}
\cdot
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}
t^{\frac{\dim {\mathrm{SO}}(2,2)-\dim T_{\mathrm{cpt}}}{2}}
\cdot
\widehat{j}(X^*_{a_1,a_2},tX^{\pm}_{a_1,a_2}) =
\frac{-1}{\pi(X^\pm_{a_1,a_2})}.
$$
Hence the limit \eqref{eq:lemmaformula:1} becomes
$$
=
\bigg(
\frac{\pi(X^+_{a_1,a_2})}{
|\pi(X^+_{a_1,a_2})|}
-
\frac{\pi(X^-_{a_1,a_2})}{
|\pi(X^-_{a_1,a_2})|}
\bigg)^{-1}
\cdot
\frac{-1}{\pi(X^\pm_{a_1,a_2})}
.
$$
By straightforward calculation, $\frac{\pi(X^+_{a_1,a_2})}{
|\pi(X^+_{a_1,a_2})|} = -\frac{\pi(X^-_{a_1,a_2})}{
|\pi(X^-_{a_1,a_2})|}$. Moreover, using the isomorphism ${\mathfrak {s}}{\mathfrak {o}}(2,2)\simeq {\mathfrak {s}}{\mathfrak {l}}_2\times {\mathfrak {s}}{\mathfrak {l}}_2$, the number $\pi(X^{\pm}_{a_1,a_2})$ is purely imaginary. Using the fact that
$
\pi(X^\pm_{a_1,a_2}) =
\frac{\pi(X^\pm_{a_1,a_2})}{
|\pi(X^\pm_{a_1,a_2})|}
\cdot
D^G(X^\pm_{a_1,a_2})^{1/2}$, the limit \eqref{eq:lemmaformula:1} eventually becomes
$$
2\cdot \pm
\cdot
D^G(X^+_{a_1,a_2})^{1/2}.
$$
It follows that
$$
\widehat{j}({\mathcal {O}}_+,X^{\pm}_{a_1,a_2}) =
\pm
\cdot D^G(X^+_{a_1,a_2})^{1/2}.
$$
With the aid of \cite[(3.4.6)]{beuzart2015local} the lemma is established.
\end{proof}
Based on Lemma \ref{lem:thecalculationofj}, the following corollary holds.
\begin{cor}\label{cor:germfunaslimitofquasichar}
Let $(G,H,\xi)$ be a Gross-Prasad triple attached to an admissible pair $(W,V)$ over ${\mathbb {R}}$, and $\Theta$ a quasi-character on $G({\mathbb {R}})$. For any $x\in {\Gamma}(G,H)$, fix a Borel pair $(B_x,T_{\mathrm{qd},x})$ for $G_x$ over ${\mathbb {R}}$. Let ${\mathfrak {t}}_{\mathrm{qd},x} = {\mathrm{Lie}} T_{\mathrm{qd},x}$, and fix a regular semi-simple element $X_{\mathrm{qd},x}\in {\mathfrak {t}}_{\mathrm{qd},x}({\mathbb {R}})$. Set $D^{G_x}(tX_{\mathrm{qd},x}) = |\det \mathrm{ad}(tX_{\mathrm{qd},x})|_{{\mathfrak {g}}_x/{\mathfrak {t}}_{\mathrm{qd},x}}|$ and $W_{T_{\mathrm{qd},x}} = W(G_x,T_{\mathrm{qd},x})$. Then the following limit formulas hold for $c_{\Theta}(x)$ defined in \eqref{num:definofcgermfun}.
\begin{enumerate}
\item If $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}^\prime_x)$ is singleton, then
$$
c_\Theta(x) =
|W_{T_{\mathrm{qd},x}}|^{-1}
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\t\mapsto 0^+
}
}
D^{G_x}(t X_{\mathrm{qd},x})^{1/2}\cdot \Theta(
x\cdot \exp(tX_{\mathrm{qd},x}));
$$
\item
If $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}^\prime_x)$ has two elements, as in Part (3) of \eqref{num:definofcgermfun}, fix an admissible pair decomposition $V^\prime_x\simeq W^\prime_x\oplus^\perp D^\prime_x\oplus^\perp Z^\prime_x$ and let $\mathrm{sig}(D^\prime_x)\in \{\pm 1\}$ be the signature of $D^\prime_x$. Then there exists a pair of regular semi-simple elements $X^{\pm}_x\in {\mathfrak {g}}_x$ with centralizer $T_{X_x}\subset G_x$ that are stably conjugate but not conjugate to each other, such that the following identity holds:
\begin{align*}
c_{\Theta}(x)
=&
\Bigg\{
\frac{
\big\{
\begin{smallmatrix}
\mathrm{sig}(D^\prime_x), & \dim V^\prime_x \text{ even} \\
-\mathrm{sig}(D^\prime_x), &\dim W^\prime_x \text{ even}
\end{smallmatrix}
}{2\cdot
|W_{T_{X_x}}|}
\Bigg\}
\cdot \lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
t\mapsto 0^+
}
}
D^{G_x}(tX_x)
\cdot
\big\{
\Theta(x\exp(tX^+_x))
-
\Theta(x\exp(tX^-_x))
\big\}
\\
&
+\frac{1}{|W_{T_{\mathrm{qd},x}}|}
\cdot
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}
D^{G_x}(tX_{\mathrm{qd},x})^{1/2}
\cdot
\Theta(x\exp(tX_{\mathrm{qd},x}))
\end{align*}
\end{enumerate}
\end{cor}
\begin{proof}
By basic calculus, for any $X\in {\mathfrak {g}}_x$
\begin{equation}\label{eq:calculus}
\lim_{t\mapsto 0}
\frac{D^G(x\cdot \exp(tX))}{D^{G_x}(tX)} = D^G(x).
\end{equation}
From \cite[Prop.~4.4.1~(vi)]{beuzart2015local}, \cite[(1.8.5)]{beuzart2015local} and \eqref{eq:calculus}, for $Y\in {\mathfrak {g}}_{x}({\mathbb {R}})$ that is regular semi-simple
\begin{equation}\label{eq:raphaellimitformula}
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}D^{G_x}(tY)^{1/2}
\cdot
\Theta(x\exp(tY)) =
D^{G_x}(Y)^{1/2}
\cdot \sum_{{\mathcal {O}}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)}
c_{\Theta,{\mathcal {O}}}(x)
\cdot \widehat{j}({\mathcal {O}},Y).
\end{equation}
In particular, by \cite[(3.4.7)]{beuzart2015local}, for $X_{\mathrm{qd},x}\in {\mathfrak {t}}_{\mathrm{qd},x}({\mathbb {R}})$ that is regular semi-simple,
\begin{align}\label{eq:limitgermforaddition:1}
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}D^{G_x}(tX_{\mathrm{qd},x})^{1/2}
\cdot
\Theta(x\exp(tX_{\mathrm{qd},x})) =
\frac{|W_{T_{\mathrm{qd},x}}|}{
|\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)|
}
\cdot \sum_{{\mathcal {O}}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)}
c_{\Theta,{\mathcal {O}}}(x)
\end{align}
Hence Part (1) holds.
For Part (2), by assumption, $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)$ has two elements. Following \eqref{num:definofcgermfun}, let us consider the case when $\dim V^\prime_x$ is even and $\geq 4$. The other case is similar. Fix the unique member ${\mathcal {O}}_{W^\prime_x}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(W^\prime_x))$ and set $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {s}}{\mathfrak {o}}(V^\prime_x)) = \{{\mathcal {O}}_{+,V^\prime_x},{\mathcal {O}}_{-,V^\prime_x}\}$ parametrized as \cite[\S6.1.2]{thesis_zhilin}. Then $\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x) = \{{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{+,V^\prime_x},{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{-,V^\prime_x}\}$. From \eqref{eq:formulaGxdecomp},
$$
G_x=G^\prime_x\times G^{\prime\p}_x,\quad
G^{\prime\p}_x = T^{\prime\p}_x\times T^{\prime\p}_x,
\quad G^\prime_x = {\mathrm{SO}}(W^\prime_x)\times {\mathrm{SO}}(V^\prime_x)
$$
with $T^{\prime\p}_x$ being anisotropic torus from the definition of ${\Gamma}(G,H)$.
From Lemma \ref{lem:thecalculationofj}, we can choose regular semi-simple elements $X^{\pm}_{V^\prime_x}\in {\mathfrak {s}}{\mathfrak {o}}(V^\prime_x)$ with centralizer torus $T_{\mathrm{cpt},V^\prime_x}$ such that for $\nu = \pm$,
$$
\widehat{j}({\mathcal {O}}_{\nu,V^\prime_x},X^+_{V^\prime_x})
=
-
\widehat{j}({\mathcal {O}}_{\nu,V^\prime_x},X^-_{V^\prime_x})
=
\nu
\cdot \frac{|W_{T_{\mathrm{cpt},V^\prime_x}}|}{2}
\cdot D^{{\mathrm{SO}}(V^\prime_x)}(X^+_{V^\prime_x})^{1/2}.
$$
Similarly, from \cite[(3.4.7)]{beuzart2015local}, we may fix a regular semi-simple element $X_{\mathrm{qd},W^\prime_x}\in {\mathfrak {t}}_{\mathrm{qd},W^\prime_x}\subset {\mathfrak {s}}{\mathfrak {o}}(W^\prime_x)$ where ${\mathfrak {t}}_{\mathrm{qd},W^\prime_x}$ is the Lie algebra of a fixed split torus $T_{\mathrm{qd},W^\prime_x}$ in ${\mathrm{SO}}(W^\prime_x)$, such that
$$
\widehat{j}({\mathcal {O}}_{W^\prime_x},X_{\mathrm{qd},W^\prime_x}) = |W_{T_{\mathrm{qd},W^\prime_x}}|
\cdot D^{{\mathrm{SO}}(W^\prime_x)}(X_{\mathrm{qd},W^\prime_x})^{-1/2}.
$$
Fix $X^\pm_x = (X_{\mathrm{qd},W^\prime_x},X^\pm_{V^\prime_x},X_{{\mathfrak {g}}^{\prime\p}_x})\in {\mathfrak {g}}_x$ where $X_{{\mathfrak {g}}^{\prime\p}_x}$ is an arbitrary fixed regular semisimple element in the Lie algebra of $G^{\prime\p}_x$. After plugging $X^{\pm}_x$ into the limit formula \eqref{eq:raphaellimitformula} and using \eqref{eq:calculus},
\begin{align*}
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}D^{G_x}(tX^\pm_x)^{1/2}\cdot \Theta(x\exp(tX^\pm_x))
=&
D^{G_x}(X^\pm_x)^{1/2}
\cdot
\big\{
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{+,V^\prime_x}}(x)
\cdot
\widehat{j}
({\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{+,V^\prime_x},
X^{\pm}_x
)
\\
&+
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{-,V^\prime_x}}(x)
\cdot
\widehat{j}
({\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{-,V^\prime_x},
X^{\pm}_x
)
\big\}
\\
=
&
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{+,V^\prime_x}}(x)
\cdot
\pm\cdot \frac{|W_{T_{\mathrm{cpt},V^\prime_x}}|\cdot
|W_{T_{\mathrm{qd},W^\prime_x}}|
}{2}
\\
&-
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{-,V^\prime_x}}(x)
\cdot
\pm\cdot \frac{|W_{T_{\mathrm{cpt},V^\prime_x}}|\cdot
|W_{T_{\mathrm{qd},W^\prime_x}}|
}{2}.
\end{align*}
For convenience, write $D^{G_x}(tX^{\pm}_x)^{1/2} = D^{G_x}(tX_x)$. Set $T_{X_x}=T_{\mathrm{cpt},V^\prime_x}\times T_{\mathrm{qd},W^\prime_x}\times G^{\prime\p}_x$ so that $W_{T_{X_x}}= W_{T_\mathrm{cpt},V^\prime_x}\times W_{T_\mathrm{qd},W^\prime_x}$. Then we eventually arrive at the following limit formula:
\begin{align}\label{eq:limitgermforsubtraction:1}
&
|W_{T_{X_x}}|^{-1}\cdot \lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
t\mapsto 0^+
}
}
D^{G_x}(tX_x)
\cdot
\big\{
\Theta(x\exp(tX^+_x))
-
\Theta(x\exp(tX^-_x))
\big\} \nonumber
\\
=&
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{+,V^\prime_x}}(x)
-
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{-,V^\prime_x}}(x).
\end{align}
But from \eqref{eq:limitgermforaddition:1},
$$
\frac{2}{|W_{T_{\mathrm{qd},x}}|}
\cdot
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}
D^{G_x}(tX_{\mathrm{qd},x})^{1/2}
\cdot
\Theta(x\exp(tX_{\mathrm{qd},x}))
=
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{+,V^\prime_x}}(x)
+
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{-,V^\prime_x}}(x).
$$
Therefore
\begin{align*}
c_{\Theta,{\mathcal {O}}_{W^\prime_x}\times {\mathcal {O}}_{\pm,V^\prime_x}}(x)
=&
\frac{1}{|W_{T_{\mathrm{qd},x}}|}
\cdot
\lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
\\
t\mapsto 0^+
}
}
D^{G_x}(tX_{\mathrm{qd},x})^{1/2}
\cdot
\Theta(x\exp(tX_{\mathrm{qd},x}))
\\
&\pm
\frac{1}{2\cdot
|W_{T_{X_x}}|}
\cdot \lim_{
\substack{
t\in {\mathbb {R}}^{\times 2}
t\mapsto 0^+
}
}
D^{G_x}(tX_x)
\cdot
\big\{
\Theta(x\exp(tX^+_x))
-
\Theta(x\exp(tX^-_x))
\big\}.
\end{align*}
It follows that Part (2) has been proved.
\end{proof}
\subsection{Stable variant}\label{subsec:stablevariant}
In this subsection, we recall the stable variant of geometric multiplicity introduced in \cite[\S 3.2]{MR3155345}. We follow the notation from subsection \ref{subsec:theformula}.
Recall that $x\in {\Gamma}(G,H)\subset {\Gamma}(H)$ if and only if ${\mathrm{SO}}(W^{\prime\p}_x)_x$ is an anisotropic torus and $G_x$ (equivalently ${\mathrm{SO}}(W^\prime_x)$) is quasi-split. In particular, following the parametrization as Part (1) of Theorem \ref{thm:parconjso} (notice that $W^{\prime\p}_x = \Im(1-x|_W)$ is always of even dimension), $x$ corresponds uniquely to a pair
$
(\kappa^{\prime\p},c^{\prime\p})\in \Xi_{{\mathrm{reg}},W^{\prime\p}_x}.
$ Moreover, following the same argument as \cite[Lem.~5.1.0.4]{thesis_zhilin}, $I_{\kappa^{\prime\p}} = I_{\kappa^{\prime\p}}^*$ if and only if ${\mathrm{SO}}(W^{\prime\p}_x)_x$ is anisotropic.
Let ${\mathcal {C}}(V,W)$ be the set of pairs $(\kappa^{\prime\p},c^{\prime\p})$ satisfying the condition that there exists an orthogonal embedding $(W_{\kappa^{\prime\p},c^{\prime\p}},q_{\kappa^{\prime\p},c^{\prime\p}})
\hookrightarrow (W,q_W)$ whose complement $W^\perp_{\kappa^{\prime\p},c^{\prime\p}}$ is quasi-split, and $I^*_{\kappa^{\prime\p}} = I_{\kappa^{\prime\p}}$. Similarly, set
$
\Xi(d_V,d_W) =
\big\{
\kappa^{\prime\p}\in \Xi_{\mathrm{reg}}\mid
I^*_\kappa = I_\kappa,
2\cdot I^*_\kappa
\leq \min\{ d_W,d_V\}
\big\}
$
where $d_V$ (resp. $d_W$) is the dimension of $V$ (resp. $W$).
By Part (1) of Theorem \ref{thm:parconjso} and the above discussion, any $(\kappa^{\prime\p},c^{\prime\p})\in {\mathcal {C}}(V,W)$ determines uniquely a pair $x_{\kappa^{\prime\p},c^{\prime\p}}^\pm\in {\Gamma}(G,H)$.
Up to a set of measaure zero, there is a two-to-one map
$
{\Gamma}(G,H)\to {\mathcal {C}}(V,W).
$
Similarly, there is a natural projection map ${\mathcal {C}}(V,W)\to \Xi(d_V,d_W)$. Let us equip ${\mathcal {C}}(V,W)$ and $\Xi(d_V,d_W)$ with the push-forward measure. The following lemma, which characterizes the fiber of ${\mathcal {C}}(V,W)\to \Xi(d_V,d_W)$, will be useful in the final proof of Theorem \ref{thm:main}.
\begin{lem}\label{lem:fiberoverkappapp}
The fiber over $\kappa^{\prime\p}\in \Xi(d_V,d_W)$ under the natural map ${\mathcal {C}}(V,W)\to \Xi(d_V,d_W)$ can be identified with
$$
\Bigg\{
\begin{matrix}
C(\kappa^{\prime\p})_{\frac{{\Delta}_W- {\mathfrak {i}}_{W,\kappa^{\prime\p}}}{2}} & \dim W \text{ odd}\\
C(\kappa^{\prime\p})_{\frac{{\Delta}_V- {\mathfrak {i}}_{V,\kappa^{\prime\p}}}{2}} & \dim W \text{ even}
\end{matrix}
$$
\end{lem}
\begin{proof}
By definition, $(\kappa^{\prime\p},c^{\prime\p})\in {\mathcal {C}}(V,W)$ if and only if $(W_{\kappa^{\prime\p},c^{\prime\p}},q_{\kappa^{\prime\p},c^{\prime\p}})\hookrightarrow (W,q_W)$ with $W^\perp_{\kappa^{\prime\p},c^{\prime\p}}$ quasi-split. By \eqref{eq:indexofWkappac}, it happens if and only if
$$
2\cdot \sum_{i\in I^*_{\kappa^{\prime\p}}}c^{\prime\p}_i = {\Delta}_{W_{\kappa^{\prime\p},c^{\prime\p}}}
={\Delta}_W
-{\Delta}_{W^\perp_{\kappa^{\prime\p},c^{\prime\p}}}
.
$$
Let us first consider the case when $\dim W$ (and hence $\dim W^\perp_{\kappa^{\prime\p},c^{\prime\p}}$) is odd. In this case, ${\Delta}_{W^\perp_{\kappa^{\prime\p},c^{\prime\p}}} = \pm 1$ is equal to the discriminant of $W^\perp_{\kappa^{\prime\p},c^{\prime\p}}$ defined in \eqref{eq:discriminantdef}. Since $I^*_{\kappa^{\prime\p}} = I_{\kappa^{\prime\p}}$, based on Lemma \ref{lem:positivenegativeindex}, the discriminant of $W_{\kappa^{\prime\p},c^{\prime\p}}$ is equal to $(-1)^{\frac{\dim W_{\kappa^{\prime\p},c^{\prime\p}}}{2}} = (-1)^{|I^*_{\kappa^{\prime\p}}|}$, and the discriminant of $W$ is equal to $(-1)^{\frac{{\Delta}_W-1}{2}}$. It follows that ${\Delta}_{W^\perp_{\kappa^{\prime\p},c^{\prime\p}}} = {\mathfrak {i}}_{W,\kappa^{\prime\p}}$ and hence
$$
\sum_{i\in I^*_{\kappa^{\prime\p}}}
c^{\prime\p}_i
=\frac{{\Delta}_W-{\mathfrak {i}}_{W,\kappa^{\prime\p}}}{2}.
$$
In particular the fiber is contained in $C(\kappa^{\prime\p})_{\frac{{\Delta}_W-{\mathfrak {i}}_{W,\kappa^{\prime\p}}}{2}}$. For the converse inclusion, it suffices to notice that ${\Delta}_{W^\perp_{\kappa^{\prime\p},c^{\prime\p}}} = \pm 1$ if and only if $W^\perp_{\kappa^{\prime\p},c^{\prime\p}}$ is quasi-split.
Now let us consider the case when $\dim W$ is even. In this case, we can embed $(W_{\kappa^{\prime\p},c^{\prime\p}},q_{\kappa^{\prime\p},c^{\prime\p}})$ into $(V,q_V)$, and by defintion $(\kappa^{\prime\p},c^{\prime\p})\in {\mathcal {C}}(V,W)$ if and only if the orthogonal complement of $W_{\kappa^{\prime\p},c^{\prime\p}}$ in $V$, which is of odd dimension, is quasi-split. Then following the same argument as the odd case,
$$
\sum_{i\in I^*_{\kappa^{\prime\p}}}
c^{\prime\p}_i
=\frac{{\Delta}_V-{\mathfrak {i}}_{V,\kappa^{\prime\p}}}{2}
$$
and the result follows.
\end{proof}
Combining Lemma \ref{lem:fiberoverkappapp} with the notation and the proof from Proposition \ref{pro:unionpureinnerconjclassof}, the following lemma can be established by tedious calculation. We omit the proof.
\begin{lem}\label{lem:fiberoverkappappunionoverpure}
Fix $e_0 = \pm 1$.
\begin{enumerate}
\item
When $\dim W$ is odd, the fiber over $\kappa^{\prime\p}\in \Xi(d_V,d_W)$ under the natural map
$$\bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W))
\\
e({\mathrm{SO}}(W_{\alpha})) = e_0
}
} {\mathcal {C}}(V_{\alpha},W_{\alpha})\to \Xi(d_V,d_W)
$$
is given by
$$
C(\kappa^{\prime\p})^{e_0\cdot {\epsilon}_{W,\kappa^{\prime\p}}}.
$$
\item
When $\dim W$ is even, fix an anisotropic line $D$ with signature $\mathrm{sig}(D)\in \{\pm 1\}$. Then the fiber over $\kappa^{\prime\p}\in \Xi(d_V,d_W)$ under the natural map
$$\bigsqcup_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W\oplus^\perp D))
\\
e({\mathrm{SO}}(W_{\alpha}\oplus^\perp D)) = e_0
}
} {\mathcal {C}}(V_{\alpha},W_{\alpha})\to \Xi(d_V,d_W)
$$
is given by
$$
C(\kappa^{\prime\p})^{e_0\cdot {\epsilon}_{V,W,\kappa^{\prime\p},D}}
$$
where
$$
{\epsilon}_{V,W,\kappa^{\prime\p},D}=
(-1)^{
\frac{|I^*_{\kappa^{\prime\p}}|-
\frac{
{\Delta}_V-
{\mathfrak {i}}_{V,\kappa^{\prime\p}}
}{2}
+
\frac{\dim W+1+{\Delta}_W+\mathrm{sig}(D)}{2}
-\mathrm{PI}(W,D)
}{2}
}.
$$
\end{enumerate}
\end{lem}
Let $\Theta$ be a \textbf{stable} quasi-character on $G({\mathbb {R}})$ in the sense of \cite[\S 12.1]{beuzart2015local}. In particular, either by \cite[(10.1.2)]{thesis_zhilin}, or from Corollary \ref{cor:germfunaslimitofquasichar} (using the fact that $X^\pm_x$ are stably conjugate to each other),
$$
c_\Theta(x) =
|\mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)|^{-1}
\cdot
\sum_{{\mathcal {O}}\in \mathrm{Nil}_{\mathrm{reg}}({\mathfrak {g}}_x)}c_{\Theta,{\mathcal {O}}}(x),\quad x\in G_\mathrm{ss}({\mathbb {R}}).
$$
Following \cite[\S 3.2]{MR3155345}, set
\begin{equation}\label{eq:stablegeomultiplicity}
m_\mathrm{geom}^S(\Theta) =
\int_{
\kappa^{\prime\p} \in \Xi(d_V,d_W)
}
2^{|I^*_{\kappa^{\prime\p}}|}
\cdot
D^G(\kappa^{\prime\p})^{1/2}
c_\Theta(\kappa^{\prime\p})
{\Delta}(\kappa^{\prime\p})^{-1/2}\,\mathrm{d} \kappa^{\prime\p}
\end{equation}
where $D^G(\kappa^{\prime\p}) = D^G(x^\pm_{\kappa^{\prime\p},c^{\prime\p}})$, ${\Delta}(\kappa^{\prime\p}) = {\Delta}(x^\pm_{\kappa^{\prime\p},c^{\prime\p}})$ and
$$
c_\Theta(\kappa^{\prime\p}) =
\Bigg\{
\begin{matrix}
c_\Theta(x^+_{\kappa^{\prime\p},c^{\prime\p}})
=c_\Theta(x^-_{\kappa^{\prime\p},c^{\prime\p}})
& (\kappa^{\prime\p},c^{\prime\p})\in {\mathcal {C}}(V,W) \text{ and $2\cdot |I^*_{\kappa^{\prime\p}}|<\dim W$}
\\
\Theta(x^+_{\kappa^{\prime\p},c^{\prime\p}})+
\Theta(x^-_{\kappa^{\prime\p},c^{\prime\p}}) &
(\kappa^{\prime\p},c^{\prime\p})\in {\mathcal {C}}(V,W) \text{ and $2\cdot |I^*_{\kappa^{\prime\p}}|=\dim W$}
\\
\end{matrix}
$$
\section{Reduction to endoscopic situation and the proof}\label{sec:redtoendoscopypf}
In this section, we establish Theorem \ref{thm:main}, and hence complete the proof of Conjecture \ref{conjec:gp}. In subsection \ref{subsec:classifyLpar}, we establish the structure theorem for tempered local $L$-parameters of special orthogonal groups over ${\mathbb {R}}$. In subsection \ref{subsec:theproof}, based on Theorem \ref{thm:classifyparameters}, we deduce Theorem \ref{thm:main} by mathematical induction and Proposition \ref{pro:endosopicid}, where the latter is analogous to \cite[Prop.~3.3]{MR3155345}.
\subsection{Classification of $L$-parameters}\label{subsec:classifyLpar}
Let $V$ be a non-degenerate quadratic space over ${\mathbb {R}}$.
In this subsection, we classify the tempered local $L$-parameters of ${\mathrm{SO}}(V)$.
\begin{thm}\label{thm:classifyparameters}
For every tempered local $L$-parameter $\varphi_V$ of ${\mathrm{SO}}(V)$ with $\mathrm{std}\circ \varphi_V:{\mathcal {L}}_{\mathbb {R}}\to {\mathrm{GL}}({\mathrm {M}}_V)$, one of the following statements holds:
\begin{enumerate}
\item[(${\mathrm {B}}$)] $\dim V\leq 3$;
\item[(${\mathrm {P}}$)] There exists a decomposition $\varphi_V = (\varphi^{\mathrm{GL}}_V)\oplus \varphi_{V,0}\oplus (\varphi^{\mathrm{GL}}_V)^\vee$ such that $\varphi^{\mathrm{GL}}_V\neq 0$;
\item[(${\mathrm {E}}$)] $\varphi_V$ is not of type $({\mathrm {B}})$ or $({\mathrm {P}})$, and $S_{\varphi_V}\not\subset Z_{{\mathrm{GL}}({\mathrm {M}}_V)}\cdot S^\circ_{\varphi_V}$.
\end{enumerate}
\end{thm}
\begin{rmk}\label{rmk:classifyparameters}
In the above notions, $({\mathrm {B}})$ refers to the base case; $({\mathrm {P}})$ refers to the ``parabolic induction" case, i.e. every representation in the Vogan $L$-packet arises from a non-trivial parabolic induction; $({\mathrm {E}})$ refers to the ``endoscopic induction", i.e. for $s_V\in (S_{\varphi_V}\backslash Z_{{\mathrm{GL}}({\mathrm {M}}_V)}\cdot S^\circ_{\varphi_V})/S^\circ_{\varphi_V}$, the elliptic endoscopic group attached to $s_V$ is ``smaller'' than ${\mathrm{SO}}(V)$.
\end{rmk}
Based on Theorem \ref{thm:classifyparameters}, we can reduce the proof of Theorem \ref{thm:main} via parabolic and endoscopic induction to base case, which was known from \cite[Prop.~7.4]{MR1295124}.
First, from \cite{knapp55local}, the following lemma is known.
\begin{lem}\label{lem:langlandspar:1}
For $\varphi_V$ tempered with the decomposition $\mathrm{std}\circ \varphi_V = \bigoplus_i m_i\cdot {\mathrm {M}}_{i,V}$ as \eqref{num:classifycomponentgroup}, the following statements hold.
\begin{enumerate}
\item ${\mathrm {M}}_{i,V}$ is either of dimension one or two;
\item The only isomorphism classes of one-dimensional representations ${\mathrm {M}}_{i,V}$ are the trivial representation or sign representation.
\end{enumerate}
\end{lem}
\begin{proof}
From \cite[p.403]{knapp55local}, the first statement is clear. When $\dim{\mathrm {M}}_{i,V} = 1$, ${\mathrm {M}}_{i,V}$ cannot be of ${\mathrm{GL}}$-type. Hence ${\mathrm {M}}_{i,V}\simeq {\mathrm {M}}_{i,V}^\vee$ and the second result follows.
\end{proof}
\begin{lem}\label{lem:langlandspar:2}
If $\varphi_V$ is not of type $({\mathrm {P}})$, then all ${\mathrm {M}}_{i,V}$ are of ${\mathrm {O}}$-type and $m_{i,V} = 1$.
\end{lem}
\begin{proof}
Suppose that there exists an ${\mathrm {M}}_{i,V}$ of ${\mathrm {O}}$-type with $m_i\geq 2$, or of ${\mathrm{Sp}}$-type (in this case $m_i$ is even and hence $\geq 2$), or of ${\mathrm{GL}}$-type. Let $\varphi_{i,V}$ be the irreducible component of ${\mathrm {M}}_{i,V}$ in $\varphi_V$, then there exists a decomposition
$$
\varphi_V = (\varphi_{V,i})\oplus \varphi_{V,0}\oplus (\varphi_{V,i})^\vee
$$
for some $\varphi_{V,0}\neq \varphi_V$. But then $\varphi_{V}$ is of type $({\mathrm {P}})$, which is a contradiction.
\end{proof}
Since $\dim {\mathrm {M}}_V$ is even, from Lemma \ref{lem:langlandspar:2}, the following corollaries holds.
\begin{cor}\label{cor:langlandspar:1}
If $\varphi_V$ is not of type $({\mathrm {P}})$, one of the statements hold:
\begin{enumerate}
\item $\dim {\mathrm {M}}_{V,i} = 2$ for any $i\in {\mathrm {I}}_{\mathrm {O}}$;
\item There exists $i_1,i_2\in {\mathrm {I}}_{\mathrm {O}}$ such that ${\mathrm {M}}_{V,i_1}$ is the trivial representation, ${\mathrm {M}}_{V,i_2}$ is the sign representation, and $\dim {\mathrm {M}}_{V,i} = 2$ for any $i\in {\mathrm {I}}_{\mathrm {O}} \backslash \{i_1,i_2\}$.
\end{enumerate}
\end{cor}
\begin{cor}\label{cor:langlandspar:2}
If $\varphi_V$ is not of type $({\mathrm {P}})$, for every $s_V\in {\mathcal {S}}_{\varphi_V}$, the $s_V$-eigenspaces ${\mathrm {M}}^{s_V=1}_V$ and ${\mathrm {M}}^{s_V=-1}_V$ are of even dimension.
\end{cor}
\begin{proof}
Since $\varphi_V$ is not of type $({\mathrm {P}})$, based on Lemma \ref{lem:langlandspar:2}, Corollary \ref{cor:langlandspar:1} and \eqref{eq:SvphiVdescription}, the centralizer of $\varphi_V$ is of the following form
\begin{equation}\label{eq:classifypar:1}
S_{\varphi_V} \simeq
\Bigg\{
\begin{matrix}
{\mathrm {O}}(1,{\mathbb {C}})^{|{\mathrm {I}}_{\mathrm {O}}|}, & \dim {\mathrm {M}}_{V,i} = 2 \text{ for any $i\in {\mathrm {I}}_{\mathrm {O}}$}\\
{\mathrm {O}}(1,{\mathbb {C}})^{|{\mathrm {I}}_{\mathrm {O}}|-2}\times S\big({\mathrm {O}}(1,{\mathbb {C}})\times {\mathrm {O}}(1,{\mathbb {C}})\big), & \text{otherwise}
\end{matrix}
\end{equation}
with each ${\mathrm {O}}(1,{\mathbb {C}}) = \{\pm 1\}$ corresponds to ${\mathrm {M}}_{i,V}$ of dimension $2$, and $S({\mathrm {O}}(1,{\mathbb {C}})\times {\mathrm {O}}(1,{\mathbb {C}})) = \{(1,1),(-1,-1)\}$ corresponds to $\text{trivial}\oplus {\mathrm{sgn}}$. It follows that the corollary holds.
\end{proof}
Now we are ready to prove Theorem \ref{thm:classifyparameters}.
Suppose that $\varphi_V$ is not of type $({\mathrm {P}})$ or $({\mathrm {E}})$, then
$$
S_{\varphi_V}\subset Z_{{\mathrm{GL}}({\mathrm {M}}_V)}\cdot S^\circ_{\varphi_V}.
$$
But following \eqref{eq:classifypar:1}, it happens only when $\dim {\mathrm {M}}_V\leq 2$, from which we immediately deduce that $\dim V\leq 3$. It follows that we have established Theorem \ref{thm:classifyparameters}.
\subsection{The proof}\label{subsec:theproof}
In this subsection, for an admissible pair $(W,V)$ over ${\mathbb {R}}$, through mathematical induction on the dimension of $(W,V)$, we establish Theorem \ref{thm:main}.
In Theorem \ref{thm:classifyparameters}, we have proved that any tempered local $L$-parameters of a special orthogonal group over ${\mathbb {R}}$ are of type $({\mathrm {B}})$, $({\mathrm {P}})$ or $({\mathrm {E}})$. From \cite[Prop.~7.4]{MR1295124}, Theorem \ref{thm:main} is known when $\dim V\leq 3$. Therefore we can make the following induction hypothesis:
\begin{num}
\item[(HYP)]\label{HYP} Theorem \ref{thm:main} holds for any admissible pairs $(W^\prime,V^\prime)$ with the following conditions:
\begin{enumerate}
\item $\dim V^\prime<\dim V$, or
\item $\dim V=\dim V^\prime$ and $\dim W^\prime<\dim W$.
\end{enumerate}
\end{num}
Fix a tempered local $L$-parameter $\varphi = \varphi_W\times \varphi_V$ of $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$. Based on (HYP) and Theorem \ref{thm:classifyparameters}, to prove Theorem \ref{thm:main}, it suffices to treat the following cases:
\begin{enumerate}
\item[(I)] $\varphi_V$ or $\varphi_W$ is of type $({\mathrm {P}})$;
\item[(II)] $\varphi_V$ is of type $({\mathrm {E}})$ and $\varphi_W$ is not of type $({\mathrm {P}})$.
\end{enumerate}
\subsubsection{Case $({\mathrm {I}})$}
Fix a tempered local $L$-parameter $\varphi = \varphi_W\times \varphi_V$ of $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$. Assume that there exists a decomposition $\varphi_V = (\varphi^{\mathrm{GL}}_V)\oplus \varphi_{V,0}\oplus (\varphi^{\mathrm{GL}}_V)^\vee$ and $\varphi_W = (\varphi^{\mathrm{GL}}_W)\oplus \varphi_{W,0}\oplus (\varphi^{\mathrm{GL}}_W)^\vee$ such that at least one of the parameters $\varphi^{\mathrm{GL}}_V$, $\varphi^{\mathrm{GL}}_W$ is non-trivial. In this case, we can apply the induction hypothesis (HYP) to assume that Theorem \ref{thm:main} holds for the admissible pair $(W_0,V_0)$ or $(V_0,W_0)$, depending on $\dim W_0\leq \dim V_0$ or $\dim V_0\leq \dim W_0$.
Set $\varphi_{0} = \varphi_{W_0}\times \varphi_{V_0}$. Let $i$ be the natural embedding from ${\mathrm{GL}}({\mathrm {M}}_{V_0})\times {\mathrm{GL}}({\mathrm {M}}_{W_0})\hookrightarrow {\mathrm{GL}}({\mathrm {M}}_V)\times {\mathrm{GL}}({\mathrm {M}}_W)$. Then $\varphi = i\circ \varphi_0$. Note that the embedding $i$ induces an injective map
$$
i^\mathrm{Vogan}:\Pi^\mathrm{Vogan}(\varphi_0)\to \Pi^{\mathrm{Vogan}}(\varphi).
$$
More precisely, by the local Langlands correspondence for general linear groups over ${\mathbb {R}}$ (\cite{langlandsclassify}), let ${\sigma}$ be the unique tempered representation in the local $L$-packet $\Pi(\varphi^{\mathrm{GL}}_V)\times \Pi(\varphi^{\mathrm{GL}}_W)$. Then for any $\pi_0\in \Pi^\mathrm{Vogan}(\varphi_0)$,
$$
i^\mathrm{Vogan}(\pi_0) = \pi_0\times {\sigma}
$$
where $\pi_0\times {\sigma}$ is the normalized parabolic induction from the parabolic subgroup of ${\mathrm{SO}}(V)\times {\mathrm{SO}}(W)$ with Levi $\big({\mathrm{SO}}(V_0)\times {\mathrm{GL}}_{\dim \varphi^{\mathrm{GL}}_V}\big)\times \big( {\mathrm{SO}}(W_0)\times {\mathrm{GL}}_{\dim \varphi^{\mathrm{GL}}_W}\big)$. Since
$$
|\Pi^\mathrm{Vogan}(\varphi_{0})| = |{\mathcal {S}}_{\varphi_0}| = |{\mathcal {S}}_\varphi| =|\Pi^\mathrm{Vogan}(\varphi)|,
$$
we deduce that $i^\mathrm{Vogan}$ provides a bijection between $\Pi^\mathrm{Vogan}(\varphi_0)$ and $\Pi^\mathrm{Vogan}(\varphi)$. By \eqref{eq:component}, the embedding $i$ also induces an isomorphism of component groups
$$
i_{V_0}:{\mathcal {S}}_{\varphi_{V_0}}\simeq {\mathcal {S}}_{\varphi_V},\quad i_{W_0}:{\mathcal {S}}_{\varphi_{W_0}}\simeq {\mathcal {S}}_{\varphi_{W}}.
$$
By \cite[Cor.~7.3.1]{thesis_zhilin},
$$
m(\pi_{\varphi_0}) = 1\Longrightarrow
m(\pi_{\varphi_0}\times {\sigma}) = 1.
$$
Hence $\pi_\varphi = \pi_{\varphi_0}\times {\sigma} = i^\mathrm{Vogan}(\pi_{\varphi_0})$. Therefore, in order to accomplish the proof of Theorem \ref{thm:main} in Case (I), it suffices to show that
\begin{equation}\label{eq:caseI:1}
\chi_{\varphi}\circ \big(i_{W_0}\times i_{V_0} \big) = \chi_{\varphi_0}.
\end{equation}
But it follows from the same argument as \cite[p.988]{MR1186476}.
\subsubsection{Case $({\mathrm {I}}\RI)$}
Fix a tempered local $L$-parameter $\varphi =\varphi_W\times \varphi_V$ of $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$. Assume that $\varphi_V$ is of type $({\mathrm {E}})$ and $\varphi_W$ is not of type $({\mathrm {P}})$. In this case, $S_{\varphi_V}\not\subset Z_{{\mathrm{GL}}({\mathrm {M}}_V)}\cdot S^\circ_{\varphi_V}$. In particular $S_{Z_V} := \big( Z_{{\mathrm{GL}}({\mathrm {M}}_V)}\cdot S^\circ_{\varphi_V}\cap S_{\varphi_V} \big)/S^\circ_{\varphi_V}$ is a non-trivial subgroup of ${\mathcal {S}}_{\varphi_V}$. Using the fact that both $\chi_\varphi$ and $\chi_{\pi_\varphi}$ are characters of ${\mathcal {S}}_\varphi$, and the fact that $S_{Z_V}$ can be identified as the subgroup $\{(1,1,...,1),(-1,-1,...,-1)\}$ of ${\mathcal {S}}_\varphi$, based on \cite[(10.4)]{MR1186476} and \cite[Prop.~10.5]{MR1186476}, we only need to prove
$$
\chi_{\pi_\varphi} = \chi_\varphi \quad \text{ on \quad $\big({\mathcal {S}}_{\varphi_V}\backslash {\mathcal {S}}_{Z_V} \big)\times {\mathcal {S}}_{\varphi_W}$}.
$$
Since there is a unique representation $\pi_\varphi\in \Pi_\mathrm{rel}^\mathrm{Vogan}(\varphi)$ with $m(\pi) =1$, for any $s\in {\mathcal {S}}_\varphi$, we have the following identity
$$
\chi_{\pi_\varphi}(s) = \sum_{\pi\in \Pi^\mathrm{Vogan}_\mathrm{rel}(\varphi)}\chi_{\pi}(s)\cdot m(\pi).
$$
It follows that to prove Theorem \ref{thm:main} in Case (II), we only need to prove the following identity for any $s\in {\mathcal {S}}_\varphi$,
\begin{equation}\label{eq:endosopic:1}
\sum_{\pi\in \Pi^\mathrm{Vogan}_\mathrm{rel}(\varphi)}\chi_{\pi}(s)
\cdot
m(\pi) = \chi_\varphi(s).
\end{equation}
Let us introduce the following notation.
\begin{defin}\label{defin:virtualrep}
Let $V$ be a non-degenerate quadratic space over ${\mathbb {R}}$. Let $\varphi_V$ be a local $L$-parameter of ${\mathrm{SO}}(V)$. Set the following virtual representation for any $s\in {\mathcal {S}}_{\varphi_V}$,
$$
{\Sigma}^s_{V,\varphi_V} =
\sum_{\pi\in \Pi^{{\mathrm{SO}}(V)}(\varphi_V)}
\chi_\pi(s)\cdot \pi.
$$
When $\varphi_V$ is clear, we abbreviate it as ${\Sigma}^s_{V}$.
\end{defin}
\begin{defin}\label{defin:stablemultiplicity}
Let $(W,V)$ be an admissible pair over ${\mathbb {R}}$ with Gross-Prasad triple $(G,H,\xi)$. Let $\varphi = \varphi_W\times \varphi_V$ be a local $L$-parameter of $G={\mathrm{SO}}(W)\times {\mathrm{SO}}(V)$. Define
$$
m^S_{W,V,\varphi} =
\sum_{{\alpha}\in {\mathrm {H}}^1({\mathbb {R}},{\mathrm{SO}}(W))}m(
{\Sigma}^{-1_{\varphi_W}}_{W_{\alpha}},
{\Sigma}^{1_{\varphi_V}}_{V_{\alpha}}
)
$$
Here
$$
m({\Sigma}^{1_{\varphi_W}}_{W},{\Sigma}^{-1_{\varphi_V}}_{V}) =
\sum_{
\pi\in \Pi^{G}(\varphi)}
\chi_\pi(-1_{\varphi_W},1_{\varphi_V})
\cdot m(\pi).
$$
When $\varphi_*$ is clear, we abbreviate $1_{\varphi_*}$ as $1$ .
\end{defin}
For $s_V\in {\mathcal {S}}_{\varphi_V}\backslash {\mathcal {S}}_{Z_V}$, from Corollary \ref{cor:langlandspar:2}, the eigenspaces
$$
{\mathrm {M}}_{V_\pm} ={\mathrm {M}}_V^{s_V = \pm 1},\quad
{\mathrm {M}}_{W_\pm} = {\mathrm {M}}_W^{s_W = \pm 1}
$$
are both of even dimension. Following \cite[\S 1.8]{MR2672539}, there exists an elliptic endoscopic group ${\mathrm{SO}}(V_+)\times {\mathrm{SO}}(V_-)$ (resp. ${\mathrm{SO}}(W_+)\times {\mathrm{SO}}(W_-)$) of ${\mathrm{SO}}(V)$ (resp. ${\mathrm{SO}}(W)$) attached to the eigenspace decomposition
$$
{\mathrm {M}}_V = {\mathrm {M}}_{V_+}\oplus {\mathrm {M}}_{V_-} (\text{resp. }{\mathrm {M}}_W = {\mathrm {M}}_{W_+}\oplus {\mathrm {M}}_{W_-}),\quad \varphi_V= \varphi_{V_+}\oplus \varphi_{V_-}
(\text{resp. }\varphi_W = \varphi_{W_+}\oplus \varphi_{W_-}).
$$
Moreover, since $s_V\notin {\mathcal {S}}_{Z_V}$, both $\dim {\mathrm {M}}_{V_+}$ and $\dim {\mathrm {M}}_{V_-}$ are strictly smaller than $\dim {\mathrm {M}}_V$. It follows that
$$
\dim V_+<\dim V, \quad \dim V_-<\dim V.
$$
Moreover, by the description of endoscopic groups (\cite[\S 1.7]{MR3155345}), up to permutation, $(W_\pm, V_\pm)$ are still admissible pairs.
It turns out that the following proposition holds.
\begin{pro}\label{pro:endosopicid}
Fix an admissible pair $(W,V)$ over ${\mathbb {R}}$ with Gross-Prasad triple $(G,H,\xi)$. Let $\varphi = \varphi_W\times \varphi_V$ be a tempered local $L$-parameter of $G$, and $s=(s_V,s_W)\in {\mathcal {S}}_\varphi = {\mathcal {S}}_{\varphi_V}\times {\mathcal {S}}_{\varphi_W}$. For $e_0\in \{\pm 1\}$, the following identity holds
\begin{align}\label{eq:endoscopicid}
\sum_{
\substack{
{\alpha}\in {\mathrm {H}}^1({\mathbb {R}},{\mathrm{SO}}(W))
\\
e(G_{\alpha}) = e_0
}
}
m({\Sigma}^{s_W}_{W_{\alpha}},{\Sigma}^{s_V}_{V_{\alpha}})
=
\frac{1}{2}
\big(
e_0\cdot m^S_{W_+,V_+}\cdot m^S_{W_-,V_-}
+
m^S_{W_+,V_-}\cdot m^S_{W_-,V_+}
\big).
\end{align}
Here
$$
m({\Sigma}^{s_W}_{W_{\alpha}},{\Sigma}^{s_V}_{V_{\alpha}}) =
\sum_{
\pi\in \Pi^{G_{\alpha}}(\varphi)}
\chi_\pi(s)
\cdot m(\pi)
,\qquad
m^S_{W_\pm,V_\pm} = m^S_{W_\pm,V_\pm,\varphi_{W_\pm}\oplus \varphi_{V_\pm}}.
$$
\end{pro}
Let us postpone the proof of Proposition \ref{pro:endosopicid} to last subsubsection. Based on Proposition \ref{pro:endosopicid}, we are ready to prove \eqref{eq:endosopic:1}.
Adding \eqref{eq:endoscopicid} for $e_0=\pm 1$,
$$
\sum_{{\alpha}\in {\mathrm {H}}^1({\mathbb {R}},{\mathrm{SO}}(W))}
m({\Sigma}^{s_W}_{W_{\alpha}},{\Sigma}^{s_V}_{V_{\alpha}}) =
m^S_{W_+,V_-}\cdot m^S_{W_-,V_+}.
$$
Now LHS is equal to
$$
\sum_{{\alpha}\in {\mathrm {H}}^1({\mathbb {R}},{\mathrm{SO}}(W))}
\sum_{\pi\in \Pi^{G_{\alpha}}(\varphi)}
\chi_\pi(s)\cdot m(\pi)
=
\sum_{\pi\in \Pi^\mathrm{Vogan}_\mathrm{rel}(\varphi)}
\chi_\pi(s)\cdot m(\pi).
$$
Therefore in order to establish \eqref{eq:endosopic:1}, it suffices to show that
$$
m^S_{W_+,V_-}\cdot m^S_{W_-,V_+} =
\chi_\varphi(s).
$$
By Definition \ref{defin:stablemultiplicity},
$$
m^S_{W_+,V_-}
=
\sum_{\pi\in \Pi^\mathrm{Vogan}_{\mathrm{rel}}(\varphi_{W_+}\times \varphi_{V_-})}
\chi_\pi(-1,1)\cdot m(\pi).
$$
It follows that by (HYP), we have
$$
m^S_{W_+,V_-} =
\chi_{\varphi_{W_+}\times \varphi_{V_-}}(-1,1),
\quad
m^S_{W_-,V_+} =
\chi_{\varphi_{W_-}\times \varphi_{V_+}}(-1,
1
).
$$
Hence we only need to prove the following identity
$$
\chi_{\varphi_{W_+}\times \varphi_{V_-}}(-1,
1
)
\cdot
\chi_{\varphi_{W_-}\times \varphi_{V_+}}(-1,
1
)
=
\chi_\varphi(s).
$$
From \cite[(10.4)]{MR1186476} and \cite[Prop.~10.5]{MR1186476}, it suffices to show the following identity
$$
\chi_{\varphi_{W_+}\times \varphi_{V_-}}(1
,
-1)
\cdot
\chi_{\varphi_{W_-}\times \varphi_{V_+}}( 1
,
-1
)
=
\chi_\varphi(s).
$$
Now from \eqref{eq:chivphicharacter:1},
\begin{align*}
\chi_{\varphi_{W_+}\times \varphi_{V_-}}
(1
,
-1)
= &
\det
(-{\mathrm{Id}}_{{\mathrm {M}}^{s_V=-1}_V})^{\frac{\dim {\mathrm {M}}_{W}^{s_W = 1}}{2}}
\cdot
\det
(-{\mathrm{Id}}_{{\mathrm {M}}_{W}^{s_W=1}})^{\frac{\dim {\mathrm {M}}_V^{s_V=-1}}{2}}
\\
\cdot &
{\varepsilon}
\big(
\frac{1}{2},
{\mathrm {M}}^{s_V=-1}_V\otimes {\mathrm {M}}_{W}^{s_W = 1},\psi
\big)
\end{align*}
and
\begin{align*}
\chi_{\varphi_{W_-}\times \varphi_{V_+}}
(1
,
-1)
=
&
\det
(-{\mathrm{Id}}_{{\mathrm {M}}^{s_W=-1}_W})^{\frac{\dim {\mathrm {M}}_{V}^{s_V = 1}}{2}}
\cdot
\det
(-{\mathrm{Id}}_{{\mathrm {M}}_{V}^{s_V=1}})^{\frac{\dim {\mathrm {M}}_W^{s_W=-1}}{2}}
\\
\cdot &
{\varepsilon}
\big(
\frac{1}{2},
{\mathrm {M}}^{s_W=-1}_W\otimes {\mathrm {M}}_{V}^{s_V = 1},\psi
\big).
\end{align*}
Comparing with the definition of $\chi_\varphi$, we only need to show the following identity
\begin{align*}
&{\varepsilon}
\big(
\frac{1}{2},
{\mathrm {M}}^{s_V=-1}_V
\otimes
{\mathrm {M}}^{s_W=-1}_W,
\psi
\big)^2
\cdot
\det(-{\mathrm{Id}}_{{\mathrm {M}}^{s_V=-1}_V})^{\frac{\dim {\mathrm {M}}_W^{s_W=-1}}{2}}
\cdot
\det(-{\mathrm{Id}}_{{\mathrm {M}}_W^{s_W=-1}})^{\frac{\dim {\mathrm {M}}_V^{s_V=-1}}{2}}
\\
&\cdot
\det(-{\mathrm{Id}}_{{\mathrm {M}}_W^{s_W=-1}})^{
\frac{\dim {\mathrm {M}}^{s_V=-1}_V}{2}
}
\cdot
\det(-{\mathrm{Id}}_{{\mathrm {M}}_V^{s_V=-1}})^{\frac{\dim {\mathrm {M}}_W^{s_W=-1}}{2}}=1.
\end{align*}
But this follows from the fact that ${\varepsilon}(...)\in \{\pm 1\}$. It follows that we completed the proof of Theorem \ref{thm:main} in Case (II).
\subsubsection{Proof of Proposition \ref{pro:endosopicid}}
Let $\Theta_{{\Sigma}^{s_V}_V} = \sum_{\pi\in \Pi^{{\mathrm{SO}}(V)}(\varphi_V)}\chi_\pi(s_V)\cdot \Theta_\pi$ be the virtual distribution character attached to ${\Sigma}^s_V$. Based on the work of Shelstad, the following statements hold.
\begin{num}
\item\label{num:shelstadworkendoscopy}
\begin{enumerate}
\item
From \cite[Lem.~5.2]{shelsteadinner}, the distribution characters $\Theta_{{\Sigma}^{1}_{V_{\pm}}}$ and $\Theta_{{\Sigma}^{1}_{W_\pm}}$ are stable.
\item From \cite{MR532374}\cite{shelsteadinner}\cite{MR627641}\cite{MR2454336}\cite{MR2581952}\cite{MR2448289}, $e({\mathrm{SO}}(V_{\alpha}))\cdot \Theta_{{\Sigma}^{s}_{V_{\alpha}}}$ (resp. $e({\mathrm{SO}}(W_{\alpha}))\cdot \Theta_{{\Sigma}^s_{W_{\alpha}}}$) is the endoscopic transfer (see \cite[\S 1.6]{MR3155345} for instance) of $\Theta_{{\Sigma}^1_{V_+}}\times \Theta_{{\Sigma}^1_{V_-}}$ (resp. $\Theta_{{\Sigma}^1_{W_+}}\times \Theta_{{\Sigma}^1_{W_-}}$) for any ${\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(V))$ (resp. $H^1({\mathbb {R}},{\mathrm{SO}}(W))$).
\end{enumerate}
\end{num}
The following proposition analogous to \cite[Prop.~3.3]{MR3155345} holds.
\begin{pro}\label{pro:analgouewaldspurger3.3}
For $e_0 = \pm 1$,
\begin{align*}
\sum_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W))
\\
e(G_{\alpha}) = e_0
}
}
m_\mathrm{geom}({\Sigma}^{s_W}_{W_{\alpha}},
{\Sigma}^{s_V}_{V_{\alpha}}
) = &
\frac{1}{2}
\big(
e_0\cdot m^S_\mathrm{geom}({\Sigma}^1_{W_+},{\Sigma}^1_{V_+})\cdot m^S_\mathrm{geom}({\Sigma}^1_{W_-},{\Sigma}^1_{V_-})
\\
&+
m^S_\mathrm{geom}({\Sigma}^1_{W_+},{\Sigma}^1_{V_-})
\cdot
m^S_\mathrm{geom}({\Sigma}^1_{W_-},{\Sigma}^1_{V_+})
\big).
\end{align*}
Here
$$
m^S_\mathrm{geom}({\Sigma}^{s_W}_W,{\Sigma}^{s_W}_V) =
m^S_\mathrm{geom}(\Theta_{{\Sigma}^{s_W}_W}\times \Theta_{{\Sigma}^{s_V}_V})
$$
where the latter is defined in \eqref{eq:stablegeomultiplicity}.
\end{pro}
\begin{proof}
Based on \eqref{num:shelstadworkendoscopy}, Lemma \ref{lem:fiberoverkappapp}, Lemma \ref{lem:fiberoverkappappunionoverpure} and Corollary \ref{cor:germfunaslimitofquasichar}, the proof of \cite[Prop.~3.3]{MR3155345} works verbatim. More precisely, let us for convenience assume that $\dim W$ is odd, the other case is similar. By \eqref{eq:geomultiplicity} and Lemma \ref{lem:fiberoverkappappunionoverpure},
$$
\sum_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W))
\\
e(G_{\alpha}) = e_0
}
}
m_\mathrm{geom}({\Sigma}^{s_W}_{W_{\alpha}},
{\Sigma}^{s_V}_{V_{\alpha}}
)
=
\int_{\Xi(d_V,d_W)}
f_{1,e_0}(\kappa^{\prime\p})\,\mathrm{d} \kappa^{\prime\p}
$$
where
$$
f_{1,e_0}(\kappa^{\prime\p}) =
\sum_{
c^{\prime\p}\in C(\kappa^{\prime\p})^{e_0\cdot
{\epsilon}_{W,\kappa^{\prime\p}}}
}
c_{{\Sigma}^{s_W}_W,{\Sigma}^{s_V}_V}(x_{\kappa^{\prime\p},c^{\prime\p}})
D^G(x_{\kappa^{\prime\p},c^{\prime\p}}){\Delta}(x_{\kappa^{\prime\p},c^{\prime\p}})^{-1/2}
$$
and
$
c_{{\Sigma}^{s_W}_W,{\Sigma}^{s_V}_V}=
c_{\Theta_{{\Sigma}^{s_W}_W}\times\Theta_{{\Sigma}^{s_V}_V}}.
$
Notice that from Remark \ref{rmk:afterpro:unionpureinnerconjclassof}, ${\epsilon}_{W,\kappa^{\prime\p}}$ depends only on $\kappa^{\prime\p}$ and $e_0$. Similarly, let $G^{\pm\pm} = {\mathrm{SO}}(W_{\pm})\times {\mathrm{SO}}(V_\pm)$. By \eqref{eq:stablegeomultiplicity} and direct computation,
\begin{align*}
&\frac{1}{2}
\big(
e_0\cdot m^S_\mathrm{geom}({\Sigma}^1_{W_+},{\Sigma}^1_{V_+})\cdot m^S_\mathrm{geom}({\Sigma}^1_{W_-},{\Sigma}^1_{V_-})
+
m^S_\mathrm{geom}({\Sigma}^1_{W_+},{\Sigma}^1_{V_-})
\cdot
m^S_\mathrm{geom}({\Sigma}^1_{W_-},{\Sigma}^1_{V_+})
\big)
\\
=&
\int_{\Xi(d_V,d_W)}
f_{2,e_0}(\kappa^{\prime\p})\,\mathrm{d} \kappa^{\prime\p}
\end{align*}
where $f_{2,e_0}(\kappa^{\prime\p})$ is equal to
\begin{align*}
2^{|I^*_{\kappa^{\prime\p}}|-1}
e_0
\sum_{
(I^{\prime\p}_1,I^{\prime\p}_2)\in {\mathcal {I}}^+(\kappa^{\prime\p})
}
&\big\{c_{{\Sigma}_{W_+}^1,{\Sigma}^1_{V_+}}
(\kappa^{\prime\p}(I^{\prime\p}_1))
D^{G_{++}}(\kappa^{\prime\p}(I_1^{\prime\p}))^{1/2}
{\Delta}(\kappa^{\prime\p}(I_1^{\prime\p}))^{1/2}
\\
\cdot
&c_{{\Sigma}_{W_-}^1,{\Sigma}^1_{V_-}}
(\kappa^{\prime\p}(I^{\prime\p}_2))
D^{G_{--}}(\kappa^{\prime\p}(I_2^{\prime\p}))^{1/2}
{\Delta}(\kappa^{\prime\p}(I_2^{\prime\p}))^{1/2}
\big\}
\end{align*}
\begin{align*}
+2^{|I^*_{\kappa^{\prime\p}}|-1}
\sum_{
(I^{\prime\p}_1,I^{\prime\p}_2)\in {\mathcal {I}}^-(\kappa^{\prime\p})
}
&\big\{c_{{\Sigma}_{W_+}^1,{\Sigma}^1_{V_-}}
(\kappa^{\prime\p}(I^{\prime\p}_1))
D^{G_{+-}}(\kappa^{\prime\p}(I_1^{\prime\p}))^{1/2}
{\Delta}(\kappa^{\prime\p}(I_1^{\prime\p}))^{1/2}
\\
\cdot
&c_{{\Sigma}_{W_-}^1,{\Sigma}^1_{V_+}}
(\kappa^{\prime\p}(I^{\prime\p}_2))
D^{G_{-+}}(\kappa^{\prime\p}(I_2^{\prime\p}))^{1/2}
{\Delta}(\kappa^{\prime\p}(I_2^{\prime\p}))^{1/2}
\big\}
\end{align*}
Here for $\kappa^{\prime\p}\in \Xi(d_V,d_W)$, the pair $(I^{\prime\p}_1,I^{\prime\p}_2)\in {\mathcal {I}}^+ (\kappa^{\prime\p})$ (resp. ${\mathcal {I}}^- (\kappa^{\prime\p})$) if and only if
\begin{itemize}
\item
$I_{\kappa^{\prime\p}} = I^{\prime\p}_1\sqcup I^{\prime\p}_2$,
\item
$\kappa^{\prime\p}(I^{\prime\p}_1) =(I^{\prime\p}_1,(F_{\pm i})_{i\in I^{\prime\p}_1}, (F_i)_{i\in I^{\prime\p}_1}, (u_i)_{i\in I^{\prime\p}_1})\in \Xi(d_{V_+},d_{W_+})$ (resp. $\Xi(d_{V_+},d_{W_-})$) and $\kappa^{\prime\p}(I_2^{\prime\p})\in \Xi(d_{V_-},d_{W_-})$ (resp. $\Xi(d_{V_-},d_{W_+})$).
\end{itemize}
Hence it suffices to show that $f_{1,e_0} = f_{2,e_0}$, which follows from \cite[Prop.~3.3]{MR3155345} verbatim.
\end{proof}
We are ready to deduce Proposition \ref{pro:endosopicid}.
By Corollary \ref{cor:analgouewaldspurger3.3} and Theorem \ref{thm:luothesis}, $m({\Sigma}^{s_W}_{W_{\alpha}},
{\Sigma}^{s_V}_{V_{\alpha}}
) = m_\mathrm{geom}({\Sigma}^{s_W}_{W_{\alpha}},
{\Sigma}^{s_V}_{V_{\alpha}}
)$. In Proposition \ref{pro:analgouewaldspurger3.3}, taking $s_V=1$, $s_W = -1$, we get
$$
V_+ = V_\mathrm{qs}, \quad V_- = 0, \quad W_+ = 0,\quad W_-=W_\mathrm{qs}
$$
where $(W_\mathrm{qs},V_\mathrm{qs})$ is introduced in Remark \ref{rmk:uniqueqsggp}.
Therefore Proposition \ref{pro:analgouewaldspurger3.3} becomes
$$
\sum_{
\substack{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W))
\\
e(G_{\alpha}) = e_0}
}
m_\mathrm{geom}({\Sigma}^{-1}_{W_{\alpha}},{\Sigma}^1_{V_{\alpha}}) =
\frac{1}{2}
\big(
e_0\cdot
m^S_\mathrm{geom}(0,{\Sigma}^1_{V_+})\cdot
m^S_\mathrm{geom}({\Sigma}^1_{W_-},0)
+m^S_\mathrm{geom}({\Sigma}^1_{W_-},{\Sigma}^1_{V_+})
\big).
$$
Taking the sum over $e_0=\pm 1$, and applying Theorem \ref{thm:luothesis}, we get the following corollary.
\begin{cor}\label{cor:analgouewaldspurger3.3}
The following identity holds
$$
m^S_\mathrm{geom}({\Sigma}^1_{W_\mathrm{qs}},{\Sigma}^1_{V_\mathrm{qs}}) =
\sum_{
{\alpha}\in H^1({\mathbb {R}},{\mathrm{SO}}(W))
}
m_\mathrm{geom}({\Sigma}^{-1}_{W_{\alpha}},{\Sigma}^1_{V_{\alpha}}) = m^S_{V_\mathrm{qs},W_\mathrm{qs}}.
$$
\end{cor}
It follows that Proposition \ref{pro:endosopicid} holds.
\bibliographystyle{amsalpha}
|
\section{Introduction}
Diffusion and other transport processes have been widely studied due to
their applications to both academic research and industry; see, for example, \cite{peters,brodin,silva,cussler} and references cited therein.
A simple but yet useful approach to study diffusion are lattice gas models, in which atoms or particles occupy discrete positions in space. They are useful, as a first approximation, to characterize transport phenomena on regular structures, such as solids and surfaces \cite{ala,gomer,antczak,paul,mehrer}.
In these models particles jump to neighboring sites, albeit limited by the interaction with other particles. For instance, for hard-core interaction only one particle is permitted on each lattice site, and thus a particle is only allowed to jump to free-sites or vacancies.
Even with simple models, obtaining closed expressions for the
diffusivity for the complete range of density values has proved to be
challenging. One of the main issues are memory effects, which in
general manifest when the particle concentration becomes
significant. In the case of hard-core interaction, when a tagged
particle jumps to a free site, the empty site that it left behind is
bound to be the target of the next jump of the tagged particle, rather
than moving in other direction. Therefore, there is a spatial
correlation between consecutive jumps which causes the movement of a
tracer to deviate from a standard random walk.
In Ref.\ \cite{dimuro}, a formula for the jump rate in terms of the excess chemical potential was obtained, combining detailed balance with the Widom
insertion formula. The mean field (MF) regime is such that memory effects
can be neglected. Generally, the tracer diffusion coefficient is written in the form $D=D^\text{MF} f$, where $D^\text{MF}$ is the diffusion coefficient in the MF regime, and $f$ is known as the correlation factor, which takes
into account the memory effects discussed previously.
Much work (theoretical, numerical and experimental) has been devoted
to the determination of $f$ in different systems; references
\cite{compaan0,compaan,leclaire,howard,nakazato,kutner,kehr,tahir,chaturvedi,ferrando,hjelt,murch,mantina,bocquet,zhangdeng}
are just a small representative sample. Most applications are related
to diffusion in solids and, more specifically, to the vacancy
mechanism of self-diffusion or substitutional diffusion,
\cite[Ch.\ 7]{mehrer}.
Here we show that it is possible to develop an analytical expression for the correlation factor by taking into account averages of transition rates when a small force is applied to a tagged particle. We analyze a two-dimensional lattice, in which interactions are represented macroscopically by the excess chemical potential $\mu_{ex}$. Apart from hard-core, we focus on soft-core interaction, in which a site can contain up to $\Omega$ particles. Note that hard core is the special case of soft core with $\Omega=1$. In addition, we analyze the case of extended hard core or $k$-NN (nearest-neighbor) hard core, in which the range of a particle occupying a site extends over a disk of a radius proportional to $k$. Extended hardcore is able to mimic, in the asymptotic limit of infinitely large $k$, the continuous dynamics of rigid disks via Monte Carlo simulations.
The paper is organized as follows. In Sec.\ \ref{s.corrfactor} we develop a general expression for the correlation factor using the transition rate formula for two neighboring sites. In Secs.\ \ref{s.softcore} and \ref{s.extended} we present expressions for the correlation factor for soft-core and extended hard-core interactions respectively. In addition, we compare the results predicted by the formulae with Monte Carlo simulations. Finally, in Sec.\ \ref{s.conclusions} we present the conclusions.
\section{Correlation factor}
\label{s.corrfactor}
Let us consider a $d$-dimensional square lattice. Each lattice site, identified with index $i$, is a cell of size $a$ containing an arbitrary number of particles. Interactions are considered only between particles within the same cell. There are $n_i$ particles in cell $i$. Each site is an open system at temperature $T$ and chemical potential $\mu$. We call $\Omega$ the number of microscopic states for one particle in a cell (proportional to its volume), and the (dimensionless) particle density is defined as $\rho_i = n_i/\Omega$.
It has been shown in Ref.\ \cite{dimuro} that the transition rate for one particle to jump from a cell with $n_1$ particles to a neighboring cell with $n_2$ particles is
\begin{equation}\label{e.trans}
W_{n_1,n_2} = \nu \frac{e^{-\beta (\mu_{\text{ex},n_2}-\mu_{\text{ex},n_1})/2}}{\sqrt{\Gamma_{n_1}\Gamma_{n_2}}},
\end{equation}
where $\nu$ is the jump frequency (a quantity that depends on the substratum and that is assumed constant), $\mu_{\text{ex},n_i}$ is the excess chemical potential and $\Gamma_{n_i}$ is the thermodynamic factor, defined as $\beta n_i \frac{\partial \mu}{\partial n_i} = 1+\beta n_i \frac{\partial \mu_{\text{ex},n_i}}{\partial n_i}$. The order of sub-indices in $W_{n_1,n_2}$ indicates the jump direction.
In a mean field approximation, we consider that $n_1=n_2=\bar{n}$ in Eq.\ \eqref{e.trans}, where $\bar{n}$ is the average number of particles. The resulting transition rate is $W_{\bar{n},\bar{n}}=\nu/\Gamma$, where $\Gamma$ without sub-index is evaluated at $\bar{n}$. The mean field tracer diffusivity is calculated using the continuous limit of a random walk with step $a$ \cite[Sect.\ 3.8.2]{gardiner}, giving
\begin{equation}\label{e.DMF}
D^\text{MF} = W_{\bar{n},\bar{n}} a^2 = D_0/\Gamma,
\end{equation}
where $D_0= \nu a^2$ is the diffusivity at small concentration. It is known that this approximation for the tracer diffusivity is not appropriate in most cases \cite{ala}. Memory effects play an important role when considering, for example, hard core interaction; in this case, after a jump, a backward second jump is more likely than a forward jump since the origin site is empty. Memory effects are taken into account by the so-called correlation factor $f$, so that the tracer diffusivity is
\begin{equation}\label{e.Df}
D = D^\text{MF} f,
\end{equation}
or
\begin{equation}\label{e.Df2}
\frac{D}{D_0} = \frac{1}{\Gamma} f.
\end{equation}
The correlation factor is usually written in terms of the average $\langle \cos\theta\rangle$, where $\theta$ is the angle between consecutive jump vectors (see for instance \cite[Ch.\ 7]{mehrer}). In the present approach we follow an alternative path appropriate for diffusion in a lattice that takes advantage from the expression for transition rates.
We consider diffusion of a tagged particle in a system composed by equivalent particles (in this case, tracer diffusivity is equivalent to self-diffusion coefficient). The tracer diffusion coefficient, $D$, can be obtained from the mobility, $B$, using the Einstein relation,
\begin{equation}\label{e.einstein}
D = \beta^{-1} B.
\end{equation}
When a small force, $F$, is applied to the tagged particle, its average velocity is $v = B F$; then, the tracer diffusivity is
\begin{equation}\label{e.DvF}
D = \beta^{-1} v/F.
\end{equation}
Let us assume that the tagged particle is in site number 1, with $n_1$ particles. The velocity in terms of the transition rates is
\begin{equation}\label{e.vel}
v = a \langle W_{n_1,n_2}^F - W_{n_1,n_0}^F\rangle_F
\end{equation}
where
\begin{align}
W_{n_1,n_2}^F &= W_{n_1,n_2} e^{\beta F a/2} = W_{n_1,n_2} (1 + \beta F a/2) \label{e.w12} \\
W_{n_1,n_0}^F &= W_{n_1,n_0} e^{-\beta F a/2} = W_{n_1,n_0} (1 - \beta F a/2) \label{e.w10}
\end{align}
and the average $\langle\rangle_F$ corresponds to a particle distribution with spatial correlations produced by the applied force. In equilibrium (without the external force), correlations are absent since interactions between neighboring sites are neglected. Replacing \eqref{e.w12} and \eqref{e.w10} in \eqref{e.vel}, we get
\begin{equation}\label{e.vel2}
v = a \langle W_{n_1,n_2} - W_{n_1,n_0}\rangle_F + \frac{\beta F a^2}{2} \langle W_{n_1,n_2} + W_{n_1,n_0}\rangle
\end{equation}
where the average $\langle\rangle$ in the last term corresponds to the equilibrium particle distribution with independent number probabilities, $P_{n_i}$, for each site. A small force is assumed in order to have a linear relationship between velocity and force.
Let us call $P^F_{n_0,n_1,n_2}$ the joint probability of having $n_0$, $n_1$ and $n_2$ particles in the corresponding sites when the force $F$ is applied to a tagged particle in site 1. The average of the first term in \eqref{e.vel2} is
\begin{equation}\label{e.averagef}
\langle W_{n_1,n_2} - W_{n_1,n_0}\rangle_F = \sum_{n_0,n_1,n_2} ( W_{n_1,n_2} - W_{n_1,n_0}) P^F_{n_0,n_1,n_2}.
\end{equation}
The joint probability is linearized,
\begin{equation}\label{e.lin}
P^F_{n_0,n_1,n_2} = P_{n_0}P_{n_1}P_{n_2} + \left.\frac{\partial P^F_{n_0,n_1,n_2}}{\partial F}\right|_{F=0} F,
\end{equation}
and the first derivative in $F$ is written in terms of an adimensional function $R_{n_0,n_1,n_2}$, that is defined such that
\begin{equation}\label{e.pr}
\left.\frac{\partial P^F_{n_0,n_1,n_2}}{\partial F}\right|_{F=0} = \beta a P_{n_0}P_{n_1}P_{n_2} R_{n_0,n_1,n_2}.
\end{equation}
Then,
\begin{equation}\label{e.lin2}
P^F_{n_0,n_1,n_2} = (1 + \beta F a R_{n_0,n_1,n_2}) P_{n_0}P_{n_1}P_{n_2}
\end{equation}
and the average in \eqref{e.averagef} is
\begin{equation}\label{e.aver2}
\langle W_{n_1,n_2} - W_{n_1,n_0}\rangle_F = \beta F a \langle (W_{n_1,n_2} - W_{n_1,n_0}) R_{n_0,n_1,n_2}\rangle,
\end{equation}
that is, we have rewritten a non-equilibrium average in terms of an equilibrium average using the function $R_{n_0,n_1,n_2}$ (also, it was used that, at equilibrium, $\langle W_{n_1,n_2}\rangle = \langle W_{n_1,n_0}\rangle$). Using \eqref{e.aver2} in \eqref{e.vel2}, the average velocity is
\begin{equation}\label{e.vel3}
v = \beta F a^2 \langle (W_{n_1,n_2} - W_{n_1,n_0}) R_{n_0,n_1,n_2} + W_{n_1,n_2}\rangle
\end{equation}
and, using \eqref{e.DvF}, the tracer diffusivity is
\begin{equation}\label{e.D}
\frac{D}{D_0} = \frac{1}{\nu} \langle (W_{n_1,n_2} - W_{n_1,n_0}) R_{n_0,n_1,n_2} + W_{n_1,n_2}\rangle.
\end{equation}
The correlation factor, $f$, is obtained combining this equation with \eqref{e.Df2}:
\begin{equation}\label{e.f}
f = \frac{\langle (W_{n_1,n_2} - W_{n_1,n_0}) R_{n_0,n_1,n_2} + W_{n_1,n_2}\rangle}{W_{\bar{n},\bar{n}}},
\end{equation}
where it was used that $W_{\bar{n},\bar{n}}=\nu/\Gamma$. It is shown below, with some examples, that simple approximations for $R_{n_0,n_1,n_2}$ provide appropriate expressions for the correlation factor.
\section{Soft core interaction}
\label{s.softcore}
For soft core interaction, the thermodynamic factor is $\Gamma=1/(1-\rho)$ and, replacing in \eqref{e.DMF}, the mean field tracer diffusivity is \cite{dimuro}
\begin{equation}\label{e.Dmfsoft}
D^\text{MF}/D_0 = 1 - \rho,
\end{equation}
where $\rho=\bar{n}/\Omega$, with $\bar{n}$ the average particle number in one cell and $\Omega$ the maximum number of particles allowed (a measure of the cell's volume). Hard core interaction is obtained for $\Omega=1$. The transition rate is \cite{dimuro}
\begin{equation}\label{e.W}
W_{n_1,n_2} = \nu (1-\rho_2)
\end{equation}
with $\rho_2=n_2/\Omega$. In the present notation, $\rho$ (without sub-index) is the average concentration at any point of the lattice, while $\rho_i=n_i/\Omega$ is the instantaneous concentration at site $i$. The transition rate depends only on the particle number in the destination site. Using this information in Eq.\ \eqref{e.f}, the correlation factor is
\begin{equation}\label{e.fsc}
f = 1 - \frac{1}{1-\rho} \langle (\rho_2-\rho_0) R_{n_0,n_2} \rangle,
\end{equation}
where the dependence of $R$ on $n_1$ can be omitted since the averaged quantity does not depend on $n_1$; more explicitly, $R_{n_0,n_2}=\sum_{n_1} R_{n_0,n_1,n_2} P_{n_1}$.
An expansion in powers of the concentrations in sites 0 and 2 is proposed for $R_{n_0,n_2}$:
\begin{equation}\label{e.expR}
R_{n_0,n_2} = \sum_{i,j=0}^\infty c_{i,j} \rho_0^i \rho_2^j,
\end{equation}
where it is assumed that the coefficients $c_{i,j}$ do not depend on $\Omega$. Keeping terms up to order 2, the average involving $R_{n_0,n_2}$ in \eqref{e.fsc} is
\begin{align}\label{e.averRsc}
\langle (\rho_2-\rho_0)& R_{n_0,n_2} \rangle \nonumber \\
=& \langle (\rho_2-\rho_0) (c_{00} + c_{10}\rho_0 + c_{01}\rho_2 + c_{20}\rho_0^2 + c_{02}\rho_2^2 + c_{11} \rho_0\rho_2)\rangle \nonumber\\
=& \frac{(c_{01}-c_{10})}{\Omega^2}(\langle n^2\rangle - \bar{n}^2) + \frac{(c_{02}-c_{20})}{\Omega^3} (\langle n^3\rangle - \bar{n} \langle n^2\rangle)\nonumber\\
=& \frac{c'}{\Omega^2} \langle\Delta n^2\rangle + \frac{c}{\Omega^3}(\langle\Delta n^3\rangle + 2 \bar{n}\langle\Delta n^2\rangle),
\end{align}
where $c'= c_{01}-c_{10}$, $c=c_{02}-c_{20}$, and $n$ is used indistinctly for sites 0 or 2 since equilibrium averages are the same in both sites. The average particle number for soft-core interaction is given by the Fermi-Dirac distribution:
\begin{equation}\label{e.nmediosoft}
\bar{n} = \frac{\Omega}{1 + e^{-\beta \mu}}
\end{equation}
from which the second and third order moments are obtained: $\langle \Delta n^2\rangle = \frac{1}{\beta} \frac{\partial \bar{n}}{\partial \mu} = \bar{n}(1-\rho)$ and $\langle \Delta n^3\rangle=\frac{1}{\beta^2} \frac{\partial^2 \bar{n}}{\partial \mu^2} = \bar{n}(1-\rho)(1-2\rho)$. Using this information and going back to the expression for the correlation factor, Eq.\ \eqref{e.fsc}, we obtain,
\begin{equation}\label{e.fsc2}
f = 1 + \frac{\rho}{\Omega}[c'+ c/\Omega + c\, 2(1-1/\Omega)\rho].
\end{equation}
Compaan and Haven \cite{compaan} have demonstrated that, for hard core interaction ($\Omega=1$) in a two dimensional lattice and in the limit of $\rho\rightarrow 1$, the correlation factor takes the value $f_{\Omega=1} = 1/(\pi-1)\simeq 0.46694$. This result can be used to set one of the constants in \eqref{e.fsc2}, where $f_{\Omega=1} = 1 + c'+ c$. Then,
\begin{equation}\label{e.fsoft}
f = 1+ \frac{\rho}{\Omega} \left[ \frac{\pi-2}{\pi-1} + c \left(1-\frac{1}{\Omega}\right)(2\rho-1) \right]
\end{equation}
\begin{figure}
\includegraphics[width=\linewidth]{fvsrho1}
\includegraphics[width=\linewidth]{fvsrho2}
\caption{Correlation factor $f$ against density $\rho$ for soft-core interaction and different values of $\Omega$, the maximum number of particles allowed. Points are numerical results and curves correspond to Eq.\ \eqref{e.fsoft} with $c=0.156$. Black dotted and dashed curves correspond to analytical results of Nakazato \cite{nakazato} and Chaturvedi \cite{chaturvedi} respectively, obtained for $\Omega=1$.} \label{f.soft}
\end{figure}
Figure \ref{f.soft} shows the correlation factor against density for different values of $\Omega$; an adjusted value of $c=0.156$ was used. Numerical results (dots) were obtained from diffusion simulations in a two dimensional square lattice. Analytical expressions obtained by Nakazato \cite{nakazato} and Chaturvedi \cite{chaturvedi} for $\Omega=1$ are also shown for comparison, they are slightly larger than the numerical results. The same happens to our linear expression for $\Omega=1$, suggesting that higher order terms have to be included in the expansion of $R_{n_0,n_2}$ to obtain a more accurate approximation. Nevertheless, the expression \eqref{e.fsoft} satisfactorily matches numerical data for different values of $\Omega$ using the same value of $c$ in all cases.
The expressions obtained by Nakazato and Chaturvedi for the two-dimensional square lattice and $\Omega=1$ are
\begin{align}
f_\text{Nakazato} &= \frac{(2-\rho)(1-\alpha)}{2-\rho-\alpha(2-3\rho)} \quad \text{with } \alpha=0.363 \nonumber\\
f_\text{Chaturvedi} &= 1 - \frac{2\rho}{(3-\beta)(2-\rho) + 2\rho} \quad \text{with } \beta=1.1894. \nonumber
\end{align}
\section{Extended hard core}
\label{s.extended}
Diffusion of particles with extended hard core is analyzed in this section. Particles move in a two-dimensional square lattice; see \cite{marques} for a possible order-disorder phase transition in this kind of system. The center of a particle occupies a lattice site and there is an exclusion region around it that the center of other particles cannot occupy; some examples are shown in Fig.\ \ref{f.ext}. A particle's center jumps randomly to neighboring sites, and the jump is allowed if the destination site does not belong to the exclusion region of another particle. Let us call $b$ the number of sites (or the area) of the exclusion region. If $b=1$ we have the situation of the previous section (with $\Omega=1$). The next case is $b=5$, where we have the center plus four nearest neighbors; by including the next-nearest neighbors we have $b=9$, a region of $3\times 3$ sites, etc. Hard disks are obtained in the limit of large $b$. Then, $b$ is not the particle size but the area of the exclusion region. For hard disks, the exclusion region has an area $\pi d^2$, with $d$ the particle diameter, while the particle area is $\pi d^2/4$. In analogy to hard disks, we define the packing fraction as $\xi = \rho b/4$ for $b>1$; and $\xi=\rho$ for $b=1$.
Knowing that the thermodynamic factor is $\Gamma = 1/(1-\rho)$ for $b=1$, we assume that for other values of $b$ it is approximately given by
\begin{equation}\label{e.gammaext}
\Gamma = \frac{1}{1-\rho/\rho_\text{max}},
\end{equation}
where $\rho_\text{max}$ is the maximum possible value of concentration, given by the average number of particles per lattice site. Fig.\ \ref{f.ext} shows different shapes of the exclusion region around one particle for increasing values of $b$; the sequence starts from $b=1$ and a layer of nearest neighbors is added in each step. The crosses represent an example of a configuration with maximum concentration in each case, from which the value of $\rho_\text{max}$ is obtained; the values are given in the following table:
\begin{center}
\begin{tabular}{c|c|c|c|c}
\rule[-1ex]{0pt}{2.5ex} $b$ & 1 & 5 & 9 & 13 \\
\rule[-1ex]{0pt}{2.5ex} $\rho_\text{max}$ & 1 & 1/2 & 1/4 & 1/5 \\
\hline
\rule[-1ex]{0pt}{2.5ex} $b$ & 21 & 25 & 29 & 37 \\
\rule[-1ex]{0pt}{2.5ex} $\rho_\text{max}$ & 1/8 & 1/9 & 1/10 & 1/12 \\
\end{tabular}
\end{center}
\begin{figure}
\includegraphics[width=\linewidth]{fext}
\caption{Exclusion regions, shown in blue around the central particle, for different values of $b$. A configuration with maximum concentration is partially represented in each case with crosses at the centers of the particles.} \label{f.ext}
\end{figure}
The maximum packing fraction is immediately obtained from $\xi_\text{max}=\rho_\text{max} b/4$ for $b>1$, and $\xi_\text{max}=1$ for $b=1$.
Considering transitions between cells of size $m\times m$, where $m$ is approximately equal to the particle diameter, the analysis of the previous section can be applied, resulting a correlation factor that has the concentration dependence given by Eq.\ \eqref{e.fsoft}, that is
\begin{equation}\label{e.fextended}
f = 1 - c_1 \xi + c_2\xi^2,
\end{equation}
where $c_1$ and $c_2$ are adjustable parameters. Using Eq. \eqref{e.Df2}, the tracer diffusivity is
\begin{equation}\label{e.Dext}
D/D_0 = f/\Gamma = (1 - c_1 \xi + c_2\xi^2)(1-\xi/\xi_\text{max}).
\end{equation}
\begin{figure}
\includegraphics[width=\linewidth]{fDext}
\caption{Normalized tracer diffusivity, $D/\nu a^2$, against packing fraction, $\xi$, for different values of $b$. Dots are numerical results and curves represent Eq.\ \eqref{e.Dext} with adjusted values of $c_1$ and $c_2$.}
\label{f.Dext}
\end{figure}
Fig.\ \ref{f.Dext} shows numerical values of the tracer diffusivity, $D$, against packing fraction, $\xi$, for different values of exclusion region size, $b$. The results are well represented by Eq.\ \eqref{e.Dext} adjusting the values of parameters $c_1$ and $c_2$; the resulting parameters are shown in table \ref{tabla}.
\begin{table}
\centering
\begin{tabular}{ccc}
\rule[-1ex]{0pt}{2.5ex} $b$ & $c_1$ & $c_2$ \\
\hline
\rule[-1ex]{0pt}{2.5ex} 1 & 0.593 & 0.0396 \\
\rule[-1ex]{0pt}{2.5ex} 5 & 3.77 & 3.57 \\
\rule[-1ex]{0pt}{2.5ex} 9 & 1.64 & 0.062 \\
\rule[-1ex]{0pt}{2.5ex} 13 & 2.76 & 1.66 \\
\rule[-1ex]{0pt}{2.5ex} 21 & 1.38 & -0.41 \\
\rule[-1ex]{0pt}{2.5ex} 25 & 1.43 & 0.068 \\
\rule[-1ex]{0pt}{2.5ex} 29 & 1.99 & 0.55 \\
\rule[-1ex]{0pt}{2.5ex} 37 & 1.48 & -0.11 \\
\end{tabular}
\caption{Adjusted values of $c_1$ and $c_2$ used in Fig.\ \ref{f.Dext} for different values of $b$.}
\label{tabla}
\end{table}
\section{Summary and conclusions}
\label{s.conclusions}
In this manuscript we address the problem of the spatial correlation
effects when studying the tracer diffusivity on a regular lattice. We
have found a general expression for the correlation factor $f$, which
takes into account the memory effects of consecutive jumps. Interactions are represented macroscopically by the
excess chemical potential. Here we have analysed the cases of
hard-core, soft-core, and extended hard-core interactions. In all cases
the results from the Monte Carlo simulations show a good agreement
with the results predicted by the theoretical expression of $f$. As
expected, we note that for soft core the correlation effects decrease ($f$ tends to 1) when $\Omega$, the number of possible configurations for one particle within a lattice site, increases. This is because when $\Omega$ is large, a jump of a particle represents a minor change in the origin site. Thus, in equilibrium,
the particle will move with almost equal probability to any of its 4
neighboring sites in its next jump.
A minor but evident drawback of the theoretical approach is that it does not
yield a complete expression for $f$, since it contains one
free-parameter that needs to be adjusted for soft-core interactions, or two for extended hard-core. Nevertheless, it has the virtue of being a general expression that can be employed to study tracer diffusion in several systems ruled by different interactions. As we have shown
throughout these lines, memory effects play an important role in the
diffusivity of a tracer particle. A theoretical
understanding of the spatial correlations in a system of interacting
particles is key to gain a full comprehension of the nature of
transport processes.
\section*{Acknowledgments}
This work was partially supported by Consejo Nacional de Investigaciones
Cient\'ificas y T\'ecnicas (CONICET, Argentina, PUE 22920200100016CO).
\bibliographystyle{elsarticle-num}
|
\section{\label{sec1:level1}Introduction}
Electron correlation has long been the key ingredient of modern condensed matter physics. In strongly correlated systems, such as the heavy fermion compound, cuprates and the iron-based superconductors, correlation introduces various unconventional metallic states \cite{RevModPhys.78.17,RevModPhys.83.1589,Powell_2011}. The metallic state with linear-$T$ resistivity is often called strange metal after its discovery in the normal state of high-$T_{c}$ cuprates superconductors. Similar phenomena has also been found in heavy fermion compounds where magnetic quantum critical point is approaching. These unconventional metallic states cannot be understood in the framework of Landau's Fermi liquid (FL) theory and has been empirically classified as non-Fermi liquids (NFLs) due to anomalous thermodynamic and transport properties\cite{PhysRevB.104.235138}. Unfortunately, despite intensive decades of studies, our knowledge on generic NFLs (beyond artificial large-$N$ limits or solvable models) is still very limited because the intrinsic strong correlation effect is beyond the scope of Hartree-Fock mean-field and perturbation theory framework, thus intuitive understanding of these NFLs is largely unknown.
Fortunately, the Luttinger's theorem provides a heuristic way to understand the essence of the NFLs. The Luttinger's theorem suggests the existence of quasiparticles, and has long been taken as a key feature of FL. Recently, several numerical researches have confirmed that the Luttinger's theorem is violated in the doped correlation-driven Mott insulators, e.g., in Fermi-Hubbard model, $t-V$ model and $t-J$ model \cite{landau1957theory,PhysRevB.104.235122,huang2019strange,PhysRevB.54.4336,PhysRevB.78.153103,PhysRevLett.81.2966,PhysRevB.75.045111,PhysRevB.62.4336,Kokalj63,Ortloff58}. However, the accuracy of these studies is affected by the dimension, size of the system, limited temperature region, and crude approximation. Especially for a doped Mott insulator lacking of particle-hole symmetry, due to the notorious minus-sign problem, the most trustworthy determinant quantum Monte Carlo simulation of the Fermi-Hubbard model is limited to a small size and high temperature, whereas the analytical study can rarely predict the nature of a specific system. In our previous work, we have studied the violation of the Luttinger's theorem in the anisotropic limit of Kondo lattice, i.e., the Ising-Kondo lattice model \cite{PhysRevB.104.165146}. Considering the specific feature of Ising-Kondo lattice, the unbiased study of the violation of the Luttinger's theorem is still insufficient and a more generic and transparent system which permits unbiased Monte Carlo simulation on a large size close to the thermodynamic limit is heavily desired..
We note that the Falicov-Kimball (FK) model is such an ideal platform to study the Luttinger's violation of the doped Mott insulator. The FK model is an alternative Hubbard model, which can be exactly solved within the framework of dynamical mean-field theory (DMFT) in infinite dimension \cite{RevModPhys.75.1333}. Hitherto, most research about FKM are restricted to the weak coupling regime, where the long-range correlation and the natural tendency to phase separation conspire to rich patterns in the zero-temperature phase diagram. \cite{PhysRevB.74.035109,RevModPhys.75.1333} These stripelike or inhomogeneous charge orderings can correspond to kinds of strongly correlated electron systems \cite{DAGOTTO20011,Tokura_2006,vojta2009lattice}, such as the doped cuprates \cite{tranquada1995evidence,Berg_2009,parker2010fluctuating,wu2011magnetic,abbamonte2005spatially,PhysRevB.78.174529,Fausti189,Ghiringhelli821,Comin1335,PhysRevLett.120.037003,zhao2019charge}, layered cobalt oxides \cite{boothroyd2011hour,babkevich2016direct} and nickelates \cite{PhysRevLett.104.206403,PhysRevLett.122.247201,Norman_2016,zhang2017large,Cosloviche1600735}. It is used to be thought that the FK model in small interaction (use $U$ as its strength) is more unconventional, while the large $U$ situation can be simply mapped onto effective Ising model and thus its underlying physics has been rarely reported in literature\cite{https://doi.org/10.1002/pssb.200460067}. In this paper we instead focus on the strong coupling regime and reveal that the strong correlation leads to non-trivial quantum states beyond Landau's FL paradigm. At high temperature, the FK system has no long-ranged charge order and thus could be clearly studied in terms of the violation of the Luttinger's theorem. It was reported that the Anderson insulator can be induced by any finite correlation at the thermodynamic limit \cite{PhysRevLett.117.146601}. Here, we find that in the doped Anderson insulator with sufficient correlation ($U>U_c$), single band evolves into a robust two-band structure, with which the Luttinger's theorem is violated. The disappearance of quasiparticle suggests the crossover from FL to NFL occurs in the Anderson insulator with increasing interaction. The evolution of band structure and the Luttinger integral can be exactly covered by a binary disordered system, with which we confirm the two-band structure is underlying the violation of Luttinger's theorem.
Actually, not all the NFLs can violate the Luttinger's theorem.
When Mott gap is absence, where the single-band structure in the NFL sustains under interaction, such as the Sachdev-Ye-Kitaev model, the Luttinger's theorem is satisfied \cite{SYK_Luttinger}.
Combined with previous studies, for the specific NFLs arising from the doped Mott insulator, the NFL nature is attributed to the change of band structure, and this kind of NFLs can be confirmed by the violation of the Luttinger's theorem.
The remainder of this paper is organized as follows: In Sec.~\ref{sec2}, the FK model is introduced and its phase diagram is briefly discussed. Strange metal state is demonstrated by the calculation of resistivity and specific heat. In Sec.~\ref{sec3}, we show that the Luttinger's theorem is violated in the FK model in terms of direct numerical calculation. In Sec.~\ref{sec4}, we provide some analytical result with the Hubbard-I approach and analyses it by the composite fermion. In Sec.~\ref{sec5}, with the assistance of the Luttinger's theorem, the phase diagram in the $U-T$ plane at half filling is elaborated. A binary disordered system is discussed. We compare the violation of the Luttinger's theorem and the spectrum function between the disordered system and the FK model, and finally end the article with a summary.
\section{Model and Method}\label{sec2}
We consider the FK model \cite{RevModPhys.75.1333} on a square lattice, the Hamiltonian is defined as
\begin{equation}
\hat{H}=-t\sum_{i,j}\hat{c}_{i}^{\dag}\hat{c}_{j}+
U\sum_{i}\hat{n}_{i}\hat{w}_{i}-\mu\sum_{i}\hat{n}_{i},
\label{eq:model1}
\end{equation}
where $\hat{c}_{j}^{\dag}(\hat{c}_{j})$ is the itinerant electron's creation (annihilation) operator at site $j$. $\hat{n}_{j}=\hat{c}_{j}^{\dag}\hat{c}_{j}$ denotes the particle number of the itinerant electron, while $\hat{w}_j$ denotes the particle number operator for the electron in the localized state. $t$-term denotes the hopping integral and only nearest-neighbor-hoping is involved. $U$ is the onsite Coulomb interaction between the itinerant electron and the localized electron. The hole doping into the half-filled system ($\mu=\frac{U}{2}$) is realized by tuning the chemical potential $\mu$.
In Eq.~\ref{eq:model1} the number of localized electron at each site is conservative since $[\hat{w}_{j},\hat{H}]=0$.
Therefore, taking the eigenstates of the number operator of localized electron $\hat{w}_{j}$ as bases, the Hamiltonian is automatically reduced to an effective free fermion model
\begin{equation}
\hat{H}(w)=-t\sum_{i,j}\hat{c}_{i}^{\dag}\hat{c}_{j}+
U\sum_{i}{n}_{i}\hat{w}_{i}-\mu\sum_{i} {n}_{i}.
\label{eq:eff_model1}
\end{equation}
Here, $w$ emphasizes its $w$ dependence and $\hat{w}_{i}|w_i>=w_i|w_i>$, $w_i=0,1$.
Now the many-body eigenstate of the original model can be constructed by the single-particle state of the effective Hamiltonian under a given configuration $\{w_i\}$,
and thus the Monte Carlo simulation can be simply carried out.
In the context we consider the square lattice with periodic boundary conditions. The sampling using the truncation algorithm is used, with which we are able to carry on a simulation of the system as large as $N_s=3600$ \cite{Kumar2006}. Since the thermodynamic and transport properties of FK model has been studied in literature, here we focus on the nature of Fermi surface. Therefore, a $60 \times 60$ square lattice is mainly used in the calculation of Luttinger integral, while the others are calculated with a $20 \times 20$ square lattice. Accordingly, the two-dimensional Brillouin zone are sampled by a $60\times 60$ $k$-point grid. Nearest neighbor hopping integral is used as the unit ($t=1$) to measure all energy scales. To attack the NFLs in the doped Mott insulator, we focus on the strong coupling regime ($U=10$).
In this work, we dope the half-filled FK model and consider the hole-doping case by tuning the chemical potential $\mu$. The translation from hole doping to electron doping is straightforward due to the particle-hole symmetry. Compared with Hubbard model, due to the lack of spin degree of freedom, in FK model the itinerant electron on each site contains only one possible state. The particle-hole symmetry also dictates that the particle density is $n_c=0.5$ when $\mu=\frac{U}{2}$.
As indicated in the previous studies in Refs.\cite{PhysRevB.74.035109,doi:10.1080/01411594.2011.604509,https://doi.org/10.1002/pssb.200460067}, the low-temperature ordered states could be identified with structure factor $S_q^{\omega}(Q)$, correlation function $g_n$ and the renormalized correlation function $G_n$, which are respectively defined as :
\begin{equation}
\begin{aligned}
&g_n=\frac{1}{4N}\sum_{i=1}^{N}\sum_{\tau_1,\tau_2=\pm1}w(\textit{\textbf{r}}_i)w(\textit{\textbf{r}}_i+\tau_1 \hat{x}+\tau_2 \hat{y}),\\
&G_n=(-1)^n4(g_n-\rho_i^2),\\
&S_q^w(\mathbf{Q})=\frac{4}{N}\sum_{i,j}e^{i\mathbf{Q}(\mathbf{R}_i-\mathbf{R}_j)}w_iw_j,
\end{aligned}
\label{eq:parameter}
\end{equation}
where $\rho_i=\frac{N_f}{N_s}$ is the concentration of localized electrons.
$N_f$ and $N_s$ are the numbers of lattice sites and localized electrons, respectively.
Above-mentioned quantities reveal a charge density wave (CDW) state at low temperature around the half-filling situation. Away from half filling, inhomogeneous is indicated by the charge distribution in the real space, which has finite density of state around the Fermi energy.
As for high temperature, the phase is detected by the properties of transports, thermodynamics and the spectral function. Accordingly, we elaborate a phase diagram under strong coupling ($U=10$) on the $\mu-T$ plane with rich quantum states. The result is summarized in Fig.~\ref{fig:phase diagram}, where three different kinds of NFLs are uncovered with distinguishable thermodynamics and transports.
At high temperature, around the half filling correlation opens the gap and leads to the Mott insulator.
Doping holes into half-filled FK model drives the Mott insulator-metal transition accompanied with Mott quantum criticality.
With decreasing particle density, the Mott insulator crossovers to the first non-Fermi liquid (NFL-I) state, the strange metal state, the second non-Fermi liquid (NFL-II) state, subsequently.
Actually, the resistivity displays unambiguous bifurcating quantum scaling behavior
\begin{equation}
\rho(T,\delta \mu)=\rho^{\ast}(T)f(T/T_0(\delta \mu)),
\label{eq:quantum_scaling}
\end{equation}
where $T_0(\delta \mu)=c|\delta \mu|^{z\nu}$, $\delta \mu=\mu-\mu^{\ast}(T)$ and $\mu^{\ast}(T)$ is the critical 'zero field' trajectory corresponding to the 'separatrix' line.
$\rho^{\ast}(T)$ is calculated for $\mu=\mu^{\ast}(T)$ and $f(x)$ denotes the unknown scaling function.The resistivity on the separatrix is almost independent with temperature.
The NFL-I and the strange metal together compose the quantum critical region (see Fig.~\ref{fig:Cv}c).
The insulator-like branch and metallic branch satisfy different scaling functions $f$, but have the same critical exponent $z\nu=1.25$.
The NFL-I is near the Mott insulator and has insulator-like feature, where a upturn exists in the resistivity curves at low temperature.
The strange metal state with a lower electron density instead shows the metallic transport properties.
In such a quantum critical region, the FK model transits from insulator to metal.
As shown in Fig.~\ref{fig:Cv}(a,b), the strange metal state is confirmed with linear-$T$ resistivity and logarithm heat capacity coefficient dependence.
In metallic states, resistivity follows $\rho(T)=\rho_{0}+AT^{n}$.
Around the separatrix line of quantum critical region, resistivity shows a quadratic dependence of temperature ($n=2$), and turns to the linear dependence ($n=1$) when crossover to strange metal.
Further decreasing electron density, the scaling behavior is lost but linear-$T$ resistivity sustains, which is referred to NFL-II state.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{phase_diagram5.jpg}
\caption{The phase diagram of the Falicov-Kimball (FK) model in the $\mu-T$ plane at $U=10$. The low temperature regime is divided into a charge density wave (CDW) state and a phase separation (PS) state by a first-order transition. At high temperature, there exists a Mott insulator state around half-filling situation and three distinguishable non-Fermi liquid (NFL) states. A quantum critical region is induced by doping-driven Mott insulator-metal transition. With increasing doped holes, the Mott insulator crossover successively to the first non-Fermi liquid (NFL-I) state, the strange metal (SM) state, and finally to the second non-Fermi liquid (NFL-II) state.
\label{fig:phase diagram}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{sigma_Cv4.jpg}
\caption{The SM behavior in the FK model. (a) Logarithm temperature dependence of the capacity heat coefficient $C(T)/T$. (b) The linear-$T$ resistivity. (c) The quantum critical scaling behavior. In the quantum critical region the resistivity satisfies the quantum critical scaling. The SM is the metal-like part and the NFL-I is referred to as the insulatorlike part.
\label{fig:Cv}
\end{figure}
\section{the violation of Luttinger theorem}\label{sec3}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{schematic3.jpg}
\caption{Schematic picture of the violation of Luttinger's theorem. For a free system with specific filling the volume of Fermi surface can be constructed by the Luttinger's theorem.
The red circle denotes the Fermi surface of a free system.
The white region is enclosed by the Fermi surface of the interacting system.
Adiabatically turning on the interaction in the free system (left) with fixed particle density, the volume enclosed by Fermi surface sustains as the free one until some critical coupling strength (middle).
Under strong enough coupling (right), the interaction changes the volume enclosed by the Fermi surface from the free system.
In fact, The violation of Luttinger's theorem suggests the NFL state has no quasiparticle which can be originated from the free system, we use the magnified grid in the Fermi surface to indicate changes of statistics.
\label{fig:LV3}
\end{figure}
Given the unconventional transports and thermodynamics, it is natural to explore the quasiparticle properties in the high temperature regime of the FK model.
To achieve this aim, we examine the validity of the Luttinger's theorem\cite{PhysRevB.68.085113,PhysRev.118.1417,PhysRev.119.1153,PhysRevLett.84.3370},
which states that if Landau quasiparticle exists, the volume enclosed by Fermi surface is consistent with its density of particles.
Such theorem has been proved originally by Luttinger and Ward in terms of Luttinger-Ward functional in the framework of perturbation theory \cite{PhysRev.118.1417,PhysRev.119.1153}, and later by Oshikawa's non-perturbative topological argument \cite{PhysRevLett.84.3370}. It now has been accepted as a key feature of FL. Mathematically, the Luttinger's theorem defines the Luttinger integral ($\mathrm{LI}$) below
\begin{equation}
\mathrm{LI}=\int_{\theta(\mathrm{Re}G(\textit{\textbf{k}},\omega=0))} \frac{d^dk}{(2\pi)^d},
\label{eq:IL}
\end{equation}
and for general FL it must be equal to density of particles ($n_{c}$)
\begin{equation}
\mathrm{LI}=n_{c}.
\label{eq:LT}
\end{equation}
Here, $G(\textit{\textbf{k}},\omega)$ is the retarded single-electron Green function and $\theta(x)$ is the standard unit-step function with $\theta(x>0)=1$ and $\theta(x<0)=0$.
Intuitively, if we use the momentum space single-particle state $|\textit{\textbf{k}}\rangle$,
the particle density $n_c$ can be obtained by integrating the Fermi-Dirac distribution function with the weight $w(k)$ for each $|\textit{\textbf{k}}\rangle$
\begin{equation}
\mathrm{n_c}=\int w(k) \times f(k)\frac{d^dk}{(2\pi)^d}.
\label{eq:IL}
\end{equation}
At zero temperature in a free system, every $|\textit{\textbf{k}}\rangle$ state has the same spectral weight $w(k)=1$, and the Fermi distribution function $f(k)$ is the unit-step function of energy $f(k)=\theta(-\epsilon_\textit{\textbf{k}})$.
Thus, the particle density is equal to the proportion enclosed by the Fermi surface, which is the same as the Luttinger integral and naturally the Luttinger's theorem is satisfied.
Note that the Luttinger integral is the positive proportion of the real part of retard Green function in the first Brillouin zone, mathematically equal to the proportion enclosed by the Fermi surface.
Actually, for any system the Luttinger integral can be interpreted as the integral of the Fermi-Dirac distribution function with a constant weight $w(k)=1$
\begin{equation}
\mathrm{LI}=\int 1 \times f(k)\frac{d^dk}{(2\pi)^d},
\label{eq:IL2}
\end{equation}
which is the counting of occupied $|\textit{\textbf{k}}\rangle$ state.
According to Eq.~\ref{eq:IL} and Eq.~\ref{eq:IL2}, the working of the Luttinger's theorem suggests $w(k)=1$, i.e., in momentum space every volume element $\frac{d^dk}{(2\pi)^d}$ can contain 1 electron.
Thus the counting of electron number is equal to the counting of occupied $|\textit{\textbf{k}}\rangle$ state.
The many-body state and the excitation can be described clearly by the wave vector $\textit{\textbf{k}}$, which is the characteristic of the quasiparticle picture.
As shown in Fig.~\ref{fig:LV3}, by adiabatically turning on the correlation, the free fermion system evolves into FL, with $|\textit{\textbf{k}}\rangle$ state modified on the perturbative level.
In the FL state, the correspondence between the original state and the final state is valid, although the dispersion relation is modified by interaction.
The correspondence between free system and the FL suggests the presence of quasiparticle, where the weight of each $|\textit{\textbf{k}}\rangle$ state is still invariant as $w(k)=1$.
Actually, although the volume of new Fermi surface in FL is invariant, its shape should depend on the form of self-energy.
With an isotropic interaction, the real part of self-energy in momentum space is also isotropy, leading to an invariant Fermi surface.
Since the interaction in FK model is isotropy, in its FL regime we expect a Fermi surface being the same with free system.
However, with anisotropic interaction, the resultant self-energy is inhomogeneous in momentum space, and thus the Fermi surface can be deformed even within the FL(see the white region in the middle panel of Fig.~\ref{fig:LV3}).
Even though the interaction may change the occupied $|\textit{\textbf{k}}\rangle$ state, if only $w(k)=1$, the volume enclosed by Fermi surface is invariant with fixed particle density. It suggests the key feature of FL is the invariant spectral weight.
We emphasize that the Luttinger's theorem is valuable for study of exotic metallic states.
Since given a specific filling, the Luttinger's theorem provides a method to reconstruct the Fermi surface of a free system with the same particle density.
For a specific correlated system, we can confirm the presence of quasiparticle if its Fermi surface enclose the same volume with the free system.
Figure.~\ref{fig:LV2} displays the imagine and real part of the retard Green function of the FK model at $n_c=0.3$ under different coupling strength, respectively.
The upper panel plots $c$ electron's spectral function $A_{\textit{\textbf{k}}}(\omega=0)$ in the first Brillouin zone and the lower panel plots the real part of Green function $\mathrm{Re}G(\textit{\textbf{k}},\omega=0)$, both of which are plotted around the Fermi energy.
Based on the Luttinger's theorem, we utilize the particle density to plot the Fermi surface of a free system with the white line.
Note that no matter in the FL (left panel), or in the strange metal state (right panel), the location of Fermi surface is always coincident with the zero values of the real part of Green function.
It makes sure that in the strange metal state the Luttinger integral keeps as the proportion enclosed by the Fermi surface.
For a weak coupling situation (left panel, $U=1$), the interacting Fermi surface coincide with the free system, suggesting a FL nature.
As for the strong coupling situation (right panel, $U=10$), the interacting Fermi surface becomes heavily deformed.
The volume of Fermi surface is greater than the free one, which involves more possible occupied $|\textit{\textbf{k}}\rangle$ states.
As shown in the right panel of Fig.~\ref{fig:LV3}, when increasing the interaction, the correspondence between the original state in free system and the final many-body state in NFL disappears at some critical coupling strength.
This suggests that the average spectral weight decreases $\overline{w(k)}<1$
and averagely a $|\textit{\textbf{k}}\rangle$ state can contain electron less than 1.
For conventional fermions, the Pauli exclusion principle permits only one electron in a state,
whereas in the NFL the strong interaction effectively increases the exclusion in statistic.
In Fig.~\ref{fig:LV1}, we demonstrate the overall Luttinger integral with varying coupling strength and varying filling.
As reference, two dotted black line is plotted as $n_c=\mathrm{IL}$ and $n_c=2 \times \mathrm{IL}$. Here, $n_c=\mathrm{IL}$ is the Luttinger integral of a free system, around which the Luttinger's theorem works well.
At exactly half filling, the Luttinger's theorem works well at any coupling strength, due to the existence of particle-hole symmetry.
With increasing coupling strength, the Luttinger's theorem sustains to $U=2$. When $U>2$ the violation emerges and the magnitude increases gradually with increasing interaction.
The degree of deviation corresponds to the degree of deformation of Fermi surface, compared with the free system.
Note that for a specific coupling strength $U$, the particle density changes the degree of deviation, but the deviation/identical of two Fermi surfaces is robust with doping. This robust violation corresponds to the rigid band structure of the FK model.
It suggests a robust NFL-like nature for all paramagnetic phases in the phase diagram (see Fig.~\ref{fig:phase diagram}).
Actually for the lattice model with finite sites, the above momentum integral should be replaced by discrete summation of the $k$-point grid. Frankly speaking, Eq.~\ref{eq:IL} is strictly valid only for zero temperature. Since interesting physics like NFLs and the strange metal behavior often exist at finite-$T$, here we follow the recent quantum determinant Monte Carlo study on the doped Hubbard model in Ref.~\onlinecite{PhysRevB.104.235122} and still use Eq.~\ref{eq:IL} to estimate the validity of the Luttinger's theorem at finite temperature. As indicated in our previous work, at high temperature regime the thermal effect on the Luttinger integral is small \cite{PhysRevB.104.165146}.
As shown in Fig.~\ref{fig:LV1}, at small filling Luttinger integral displays a linear dependence of the particle density $\text{IL}=a \times n_c$ whereas $a \ne 1$ indicates the changes of statistics in NFLs.
Here, the spectral weight is $w(k)=\frac{1}{a}$.
The linear-$n_c$ behavior sustains to some critical particle density, which closely associated with the band structure.
The critical $n_c$ is correspondent to the coincidence of Fermi energy and the pseudo energy gap.
Not only the critical particle density ($n_c \sim 0.25$), but also the critical coupling strength ($U_c=2$) all turn out to be closely related to the band structure.
These properties will be further discussed in the binary disordered system in the final section.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Luttinger_violation2.jpg}
\caption{ Spectral function at Fermi energy $A_{\textit{\textbf{k}}}$$(\omega=0)$ (upper panel) and real part of Green function at Fermi energy $\mathrm{Re}G({\textit{\textbf{k}}},\omega=0)$ (lower panel) in the FL (left panel, $U=1$, $T=0.2$, $n_c=0.3$) and the SM state ($U=10$, $T=0.2$, $n_c=0.3$, left panel). According to the Luttinger's theorem, the Fermi surface of a free system with particle density $n_c=0.3$ is denoted by the white circle. The location of Fermi surface is in accordance with the zero values of the real part of Green function.
\label{fig:LV2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{ReG5.jpg}
\caption{An overall spectacular of the violation of Luttinger's theorem. The Luttinger integral versus density of electron $n_c$ at $T=0.2$. In FL ($U=1$, $U=2$), the electron density and Fermi surface is in accordance with the Luttinger's theorem. At stronger coupling ($U=4$), the proportion of the first Brillouin zone enclosed by Fermi surface deviates from the density of electrons for most doping regime thus confirms their NFL nature. Further increasing coupling strength leads to a stronger deviation ($U=10$).
\label{fig:LV1}
\end{figure}
\section{the Hubbard-I approximation and composite fermions analysis}\label{sec4}
In the last section, with unbiased Monte Carlo simulation we study the violation of Luttinger's theorem in the FK lattice.
In this section, we want to further study the microscopic mechanism of the violation of the Luttinger's theorem analytically.
In the high temperature situation, one can follow the Hubbard-I approximation to give a rough solution for FK model.
At first, we use the equation of motion (EOM) formalism and define the retarded Green's function as \cite{doi:10.1098/rspa.1963.0204}
\begin{equation}
G_{i,j,\sigma}(\omega)=\langle\langle \hat{c}_{i \sigma}|\hat{c}^\dagger_{j \sigma} \rangle\rangle.
\label{eq:Hubbard-I1}
\end{equation}
By using the standard EOM relation,
\begin{equation}
\omega \langle\langle \hat{A}|\hat{B}\rangle\rangle_{\omega}=\langle[\hat{A},\hat{B}]_+ \rangle+\langle \langle[\hat{A},\hat{H}]_-| \hat{B}\rangle\rangle_{\omega}.
\label{eq:Hubbard-I2}
\end{equation}
For the FK model, the Hamiltonian can be rewritten as
\begin{equation}
\hat{H}=-t\sum_{i,j}\hat{c}_{i}^{\dag}\hat{c}_{j}+
\frac{U}{2}\sum_{i}(2\hat{w}_{i}-1)\hat{n}_{i}+(\frac{U}{2}-\mu)\sum_{i}\hat{n}_{i}.
\label{eq:model2}
\end{equation}
At half-filled situation, it follows that
\begin{equation}
\omega \langle\langle \hat{c}_{i}|\hat{c}^\dagger_{j }\rangle\rangle_{\omega}=\delta_{ij}-t \Delta_{im}\langle\langle \hat{c}_{m}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}+\frac{U}{2} \langle\langle (2\hat{w}_{i}-1) \hat{c}_{i}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}.
\label{eq:Hubbard-I3}
\end{equation}
Here, $\Delta_{im}$ denotes m is the nearest-neighbor site of i. For $\langle\langle (2\hat{w}_{i}-1) \hat{c}_{i}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}$, we have
\begin{equation}
\begin{aligned}
\omega \langle\langle (2\hat{w}_{i}-1) \hat{c}_{i }|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}= & \langle (2\hat{w}_{i}-1) \rangle \delta_{ij}
-t \Delta_{il} \langle\langle (2\hat{w}_{i}-1) \hat{c}_{l}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega} \\
&+\frac{U}{2} \langle\langle \hat{c}_{i }|\hat{c}^\dagger_{j }\rangle\rangle_{\omega}.
\end{aligned}
\label{eq:Hubbard-I4}
\end{equation}
If no further EOM is involved, to close EOM we have to decouple $\langle\langle (2\hat{w}_{i}-1) \hat{c}_{l}|\hat{c}^\dagger_{j }\rangle\rangle_{\omega}$ as
\begin{equation}
\langle\langle (2\hat{w}_{i}-1) \hat{c}_{l}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega} \simeq \langle 2\hat{w}_{i}-1\rangle \langle\langle \hat{c}_{l}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}.
\label{eq:Hubbard-I5}
\end{equation}
In paramagnetic strong coupling regime, there is no finite CDW order, thus $\langle(2\hat{w}_{i}-1)\rangle=0$. Meanwhile, the contribution from $\langle\langle (2\hat{w}_{i}-1) \hat{c}_{i}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}$ vanishes due to the decoupling and the above equation has a complete solution:
\begin{equation}
(\omega-\frac{U^2}{4\omega}) \langle\langle \hat{c}_{i}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega}=\delta_{ij}-t \Delta_{im}\langle\langle \hat{c}_{m}|\hat{c}^\dagger_{j}\rangle\rangle_{\omega},
\label{eq:Hubbard-I6}
\end{equation}
which can be written as
\begin{equation}
(\omega-\frac{U^2}{4\omega}) G_{i,j}(\omega)=\delta_{ij}-t \Delta_{im}G_{m,j}(\omega).
\label{eq:Hubbard-I7}
\end{equation}
Now, performing the Fourier transformation
\begin{equation}
G_{i,j}(\omega)=\frac{1}{N_s} \sum_k e^{ik(R_i-R_j)}G(k,\omega),
\label{eq:Hubbard-I8}
\end{equation}
we have
\begin{equation}
\begin{aligned}
&\sum_k (\omega-\frac{J^2}{16\omega}) G(k,\omega) e^{ik(R_i-R_j)}=\sum_k e^{ik(R_i-R_j)} \\
&-t \Delta_{im}G(k,\omega) e^{ik(R_m-R_j+R_i-R_i)}.
\end{aligned}
\label{eq:Hubbard-I9}
\end{equation}
Here, $-t \Delta_{im}e^{ik(R_m-R_i)}=-t\sum_{\delta} e^{-ik\delta}=\varepsilon_k$ and we have $\sum_k[(\omega-\frac{U^2}{4\omega})G(k,\omega)-1]e^{ik(R_i-R_j)}=0 $, which gives the single-particle Green's function as
\begin{equation}
\begin{aligned}
G(k,\omega)=\frac{1}{\omega-\frac{U^2}{4\omega}-\varepsilon_k}=\frac{\alpha_k^2}{\omega-\hat{E}_k^+}+\frac{1-\alpha_k^2}{\omega-\hat{E}_k^-}.
\end{aligned}
\label{eq:Hubbard-I10}
\end{equation}
Here, the coherent factor $\alpha_k^2=\frac{1}{2} \left( 1+\frac{\varepsilon_k}{\sqrt{\varepsilon_k^2+U^2}} \right)$ and
\begin{equation}
\begin{aligned}
\hat{E}_k^{\pm}=\frac{1}{2} \left[ \varepsilon_k\pm \sqrt{\varepsilon_k^2+U^2} \right].
\end{aligned}
\label{eq:Hubbard-I11}
\end{equation}
With above single-particle Green's function, we plot the Luttinger integral at $U=10$ by the Hubbard-I approximation (orange line with circle marker) as a function of particle density and compare it with the Monte Carlo result (blue line with circle marker) in Fig.~\ref{fig:Hubbard-ReG}.
As reference, the Luttinger integral of the free fermion system is plotted in yellow line with circle marker.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{Hubbard-ReG3.jpg}
\caption{Luttinger integral (IL) versus density of electron $n_c$ at $T=0.2$. It displays the violation of the Luttinger's theorem by Monte Carlo simulation (blue, $U=10$) and the Hubbard-I approximation method (orange, $U=10$), respectively. The Luttinger integral of the free fermion system (yellow) is plotted as reference.
\label{fig:Hubbard-ReG}
\end{figure}
It turns out that the Hubbard-I approximation provides an accurate evaluation for the Fermi surface at strong coupling regime. Compared with the Hubbard model, in the FK model Hubbard-I approximation works better, which can almost recover the violation of Luttinger's theorem.
This advantage in the FK model is attributed to the commutation between the interaction and hopping terms
\begin{equation}
[2\hat{w}_{i}-1,\hat{c}_{i}\hat{c}^\dagger_{j }]=0,
\label{eq:commuation}
\end{equation}
where the EOM truncates naturally at the second order only if we assume Eq.~\ref{eq:Hubbard-I5} works.
As mentioned above, the violation of Luttinger's theorem suggests the correspondence between FL and NFL is lost.
We cannot access to an approximated formalism of the quasiparticle by renormalizing parameters, such as the effective mass.
This remind us that an unconventional type of quasiparticle (not Landau quasiparticle) excitation should be expected, if exists.
To access an intuitive understanding of the violation of the Luttinger's theorem in terms of approximated quasiparticle, we provide a composite fermion approach with similar result compared with Hubbard-I approximation.
We note that the interaction term in Eq.~\ref{eq:model1} can be interpreted as the hybridization between the itinerant electrons and some composite fermions $f^{\dagger}_j=(2\hat{w}_j-1)c^{\dagger}_j$, which satisfies the commutation relation of fermions
\begin{equation}
\{f^{\dagger}_i,f_j\}=(2\hat{w}_i-1)(2\hat{w}_j-1)\{c^{\dagger}_i,c_j\}=\delta_{ij}.
\label{eq:commutation}
\end{equation}
However, the commutation relation between the composite fermion and the itinerant electron is nontrivial
\begin{equation}
\{c^{\dagger}_i,f_j\}=(2\hat{w}_j-1)\{c^{\dagger}_i,c_j\}=(2\hat{w}_j-1)\delta_{ij}.
\label{eq:commutation2}
\end{equation}
Now the FK Hamiltonian can be written as
\begin{equation}
\hat{H}=-t\sum_{i,j}\hat{c}_{i}^{\dag}\hat{c}_{j}+
\frac{U}{4}\sum_{i}(\hat{f}^{\dagger}_{i}\hat{c}_{i}+\hat{c}^{\dagger}_{i}\hat{f}_{i})+(\frac{U}{2}-\mu)\sum_{i}\hat{n}_{i}.
\label{eq:model3}
\end{equation}
We assume $\{c^{\dagger}_i,f_j\}$ can be replaced by its average value. At the high temperature paramagnetic regime $<(2\hat{n}_j-1)>=0$ and thus the composite fermion and itinerant electron restore a standard commutation relation $\{c^{\dagger}_i,f_j\}=0$.
It suggests at high temperature we can treat the composite fermion as conventional fermion.
By Fourier transformation, the Hamiltonian can be written as
\begin{equation}
\hat{H}=\sum_{k}
\left(
\begin{array}{cc}
\hat{c}_{k}^{\dag} & \hat{f}_{k}^{\dag}\\
\end{array}
\right)
\left(
\begin{array}{cc}
\epsilon_k+\frac{U}{2}-\mu & \frac{U}{4} \\
\frac{U}{4} & 0 \\
\end{array}
\right)
\left(
\begin{array}{c}
\hat{c}_{k} \\
\hat{f}_{k} \\
\end{array}
\right).
\label{eq:model3_1}
\end{equation}
The dispersion is given by diagonalizing the Hamiltonian
\begin{equation}
\begin{aligned}
\hat{E}_k^{\pm}=\frac{1}{2} \left[ \varepsilon_k\pm \sqrt{\varepsilon_k^2+\frac{U^2}{4}} \right].
\end{aligned}
\label{eq:Hubbard-I11}
\end{equation}
This dispersion is similar with the Hubbard-I results, with a gap $\triangle E=\frac{U}{2}$ smaller than $\triangle E_{Hubbard-I}=U$.
This composite fermion phenomenologically suggests that the elementary excitation of the FK model consists of the contribution of both itinerant electron and also the composite fermion.
Therefore, it is hard to describe the many-body state in terms of a single (Landau) quasiparticle picture.
As particle number increasing by 1, in the momentum space the correspondent many-body state occupied a volume larger than $\frac{d^dk}{(2\pi)^d}$.
\section{Discussion and Conclusion}\label{sec5}
\subsection{Phase diagram at half filling}\label{sec5_1}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{phase2_diagram5.jpg}
\caption{The phase diagram of the FK model in the $U-T$ plane at half-filling situation. The low temperature regime is a CDW state under any finite coupling. At high temperature three different states are uncovered. Under weak coupling system sustains as a FL until $U=2$. At moderate coupling, the Anderson insulator (AI) emerges with a crossover. Further increasing correlation leads to the MI with full opened energy gap.
\label{fig:phase diagram2}
\end{figure}
We notice the robust violation/working of the Luttinger's theorem for a specific coupling strength in the doped FK model, which is invariant under different doping.
Therefore, the Luttinger's theorem can not only help tell NFLs from FLs in the doped system,
but also help distinguish different state under different coupling strength at half filling,
although the working of Luttinger's theorem is protected there by the particle-hole symmetry.
According to the order parameter in Eq.~\ref{eq:parameter}, the density of state, the inverse participation ratio,
and the critical coupling strength where the Luttinger's theorem fails with doping, the half-filling phase diagram of the FK model in the $U-T$ plane can be elaborated as Fig.~\ref{fig:phase diagram2}.
At low temperature the CDW state exists under any finite interaction $U$.
At high temperature, under strong coupling ($U>6$), the FK model is a Mott insulator with fully opened energy gap.
Under weak coupling ($U<2$), the FL nature sustains.
Under a moderate coupling ($2<U<6$), there exists finite density of state around Fermi energy while the two-band structure emerges.
The localization revealed by inverse participation ratio suggests an Anderson insulator.
Note that the phase diagram is different from the previous one in Ref.~\onlinecite{PhysRevLett.117.146601}.
In Ref.~\onlinecite{PhysRevLett.117.146601}, the whole metallic regime is indicated as the localization state, where the Anderson insulator state can be extrapolated to any finite small coupling at thermodynamic limit.
In this work, according to the critical coupling strength where the Luttinger's theorem is violated in the doped FK lattice, we find that the Anderson insulator state can be further divided into a FL state at weak coupling and a NFL state at strong coupling.
What's more, the FL state is not simply the effect of finite size, which can sustain to $U \sim 2$ at thermodynamic limit.
\subsection{A binary disordered system}\label{sec5_2}
Since the Anderson localization in FK model has long been attributed to the disorder effect \cite{PhysRevB.104.045116}, it is valuable to check the pure effect of disorder on the working of Luttinger's theorem.
We wonder whether increasing the strength of disorder will leads to the disappearance of quasiparticle.
Since the doped FK model possesses robust rigid two-band structure at strong coupling, some connection may exist between the robust violation of the Luttinger's theorem and the robust two-band structure, which also corresponds to the singularity of Green's function at high energy regime \cite{Heath_2020}.
However, we wonder whether the origin of band splitting is also important.
To this end, it is necessary to study a disordered system, which could induce a two-band structure at the absence of interaction.
Thus, in this section, we construct a Hamiltonian with interaction replaced by discrete binary disorder to study the Luttinger's theorem in a two-band system, which is written as:
\begin{equation}
\hat{H}=-t\sum_{i,j}\hat{c}_{i}^{\dag}\hat{c}_{j}-(\mu_i-\mu_0)\sum_{i}\hat{n}_{i}.
\label{eq:model4}
\end{equation}
Different from the FK model, where $\mu$ is fixed during a Monte Carlo simulation, in this disordered system chemical potential is composed by two part and the binary disorder is introduced by $\mu_i$.
For each site $i$, $\mu_i$ is randomly taken as $\mu_i=\pm {U}_\text{disorder}$. $\mu_0$ instead is a fixed number in Monte Carlo simulation and it is used to tuning the particle density.
We show the density of state at $n_c=0.3$ for this disordered system (red line) with different disorder strength in Fig.~\ref{fig:dis_dos}(a-c) and compare it with the FK model (blue line).
It turns out that when $|\mu_i|=U_{FK}$, $\mu_0=\mu_{FK}-\frac{U_{FK}}{2}$, the spectral function of the FK model agrees with the disordered system perfectly with varying coupling/disorder strength $U$. Even the location of pseudo gap is in accordance with each other.
\begin{figure}
\centering
\includegraphics[width=1.\columnwidth]{disorder_dos3.jpg}
\caption{(a-c) Density of state under different coupling/disorder strength at $n_c=0.3$ at high temperature $T=0.2$. The density of state of the FK model is denoted by the blue line, and the binary disordered system is denoted by the red line. (d) In FK model the energy gap is opened under small interaction $U=1$ at low temperature $T=0.01$. This opened gap can easily be erased by either thermal fluctuation or doping holes.}
\label{fig:dis_dos}
\end{figure}
At small coupling/disorder strength ($U=1$), only one quasiparticle peak exists and it is the same with the free electron situation.
Under a moderate coupling strength ($2<U<6$), the two-peak structure emerges while the gap is still not fully opened. Finite density of state around the Fermi energy and the two-peak structure leads to the localization of the FK/disordered system. This region is referred to as the Anderson insulator. Increasing coupling/disorder strength makes the gap fully opened, and the scale can be tuned by the magnitude of $U$.
As shown in Fig~.\ref{fig:4}, we further demonstrate the Luttinger integral in the disordered system, as a function of disorder strength and the particle density.
We study the working/violation of the Luttinger's theorem and the evolution of spectral function at different parameter, to confirm the role that band structure plays in the violation of the Luttinger's theorem.
As shown in Fig~.\ref{fig:dis_dos}a, when the one-band structure sustains ($U=1$), the Luttinger's theorem works well where the Fermi surface enclosing a volume consistent with the particle density $n_c$ (see Fig.~\ref{fig:4}).
Increasing the disorder ($U=4$) opens an energy gap $\triangle E \sim U$ in the middle of the spectrum and introduces the deviation of Luttinger integral, suggesting a deformed Fermi surface with larger enclosed volume.
Note that at $U=4$ the FK model is in the Anderson insulator state, where the gap is not fully opened.
It indicates that the violation of the Luttinger's theorem is directly caused by the two-peak structure, i.e, by the singularities of the self-energy at both high and low frequencies, rather than by the
fully opened Mott gap.
Further increasing the disorder ($U=10$) leads to greater deviation and finally accesses to the Mott insulator at half filling.
The Luttinger integral in Fig.~\ref{fig:4} is similar with the FK one shown in Fig.~\ref{fig:LV1}.
In both FK model and the disordered system, the working of Luttinger's theorem depends on the band structure.
We want to emphasize that if the Anderson localization is introduced by the Gaussian-type disorder instead of this discrete disorder,
the Luttinger's theorem will not be violated due to the absence of two-band structure.
Therefore, we conclude the NFL nature is not caused by localization.
As a function of particle density,
the behavior of Luttinger integral is associated closely with the characters of the spectrum function, e.g., another singularity around the pseudo gap.
In both $U=10$ and $U=4$ situations, the Luttinger integral displays a linear dependence of particle density when $n_c < 0.25$, where $n_c = 0.25$ is correspondent to the filling around pseudo gap.
\subsection{Some discussions}\label{sec5_3}
We note that previous analytical studies have demonstrated that two-band structure of the Hubbard and related models guarantees a violation of the Luttinger's theorem,
where the violation stems from the zeros of the single-particle Green function which arises from the separation of scales of the upper and lower Hubbard bands \cite{PhysRevB.75.104503,PhysRevLett.110.090403,Phillips_2020}.
Actually, the proof is also available for the FK model.
In the Mott insulator state, at Fermi energy the self-energy diverges at $\epsilon_k=0$ in the momentum space, leading to the zeros of the single-particle Green function and a fully opened gap.
The divergence of self-energy makes Luttinger-Ward functional ill-defined. Slightly doping away from the half-filling situation, the Luttinger integral counts the zeros instead of poles as the particle density.
As we know, the zeros in momentum space corresponds to the Luttinger surface.
At half filling, the Luttinger surface just locates at the Fermi surface of a half-filling free system.
In fact, with any finite doping, in the FK model the Luttinger surface always exists at a specific energy ($\omega=\mu$).
However, when doping driven the FK model crossover from the Mott insulator to the NFL-I, the energy scale $\omega=\mu$ is too high to dominate the physics.
In the NFL-I, strange metal, and NFL-II, the Fermi surface instead emerges around Fermi energy.
The previous study based on cluster dynamical mean field theory suggests that the FK model is never a strict Landau FL even at weak coupling situation ($U/t \sim 0.5$) \cite{PhysRevB.94.081115}.
However, our work indicates that the Luttinger's theorem is working perfectly at $U/t=1$ as shown in the Fig.~\ref{fig:LV2}, which instead is a clear signal for the presence of quasiparticle.
The disagreement originates from the fact that the NFLs we discussed here are embedded in high temperature paramagnetic regime, where the effect of thermal fluctuation plays an important role.
Actually, as shown in Fig~.\ref{fig:dis_dos}d, CDW ground state of the FK model is robust for any finite $U$, in which the interaction splits the energy band and leads to fully opened gap.
However, different from the real NFLs, this two-band structure is not robust, which can be simply destroyed by either thermal fluctuation or doping holes.
Therefore, only conventional FL is revealed under weak coupling.
Only if the coupling strength is larger than the critical value $U_c$ ($U_c=2$) can the energy gap sustain under thermal fluctuation and doping.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{disorder_ReG3.jpg}
\caption{Luttinger integral (IL) versus density of electron $n_c$ at $T=0.2$. It displays the violation of the Luttinger's theorem in the binary-disordered system. Within a weak disorder (${U}_\text{disorder}=1$, ${U}_\text{disorder}=2$), system behaves as a general FL, where the electron density is in accordance with the Luttinger's theorem. Within stronger disorder (${U}_\text{disorder}=4$), the doped Anderson insulator show the NFL nature, where the proportion enclosed by Fermi surface deviates from the density of electrons for most doping regime. Further increasing disorder strength leads to a stronger deviation (${U}_\text{disorder}=10$).
\label{fig:4}
\end{figure}
\subsection{Conclusion}\label{sec5_4}
In this paper, with unbiased Monte Carlo simulation, we reveal an unambiguously violation of the Luttinger's theorem in the doped FK model.
Under strong coupling, doping holes in the half-filling FK lattice leads to a Mott insulator-metal transition.
Intriguingly, the NFL nature is robust under strong coupling at any finite doping.
As shown in the Fig.~\ref{fig:phase diagram},
at high-temperature regime three different NFLs are revealed.
Empirically, we used to distinguish the NFL features by unconventional transport and thermodynamic properties,
such as the linear-$T$ resistivity and logarithm dependence of heat capacity coefficient in the strange metal state.
Although demonstrating various thermodynamical and transport behaviors, their NFL-like nature do originate from a common ground in spectrum.
Here, with the Luttinger's theorem, we demonstrate the deformation of Fermi surface in the NFLs is associated with the transfer of spectral weight.
After reliable numerical simulation, we try to analytically access to the violation of the Luttinger's theorem.
With the Hubbard-I approximation approach where we have solved the EOM to the second order,
the violation of the Luttinger's theorem is reproduced quantitatively.
To provide a more intuitive picture, we construct the composite fermion. It turns out the excitation in the FK model is
mixed of the itinerant electron and the composite fermion, which cannot simply described by elementary excitation formalism of a specific particle.
Considering the lack of quasiparticle, the breakdown of FL paradigm is predictable.
Finally, a binary disordered model is constructed to discuss the connection between the violation of Luttinger's theorem and the feature of spectrum.
This disordered system can cover the density of states in FK model perfectly.
In such a different model which emphasizes the disorder effect rather than interaction,
the two-band structure can also deform the Fermi surface directly.
As a conclusion, the violation of Luttinger's theorem in the doped Mott insulator is directly connected to the two-band structure, where the change of statistics can be taken as the key feature of NFLs underlying unconventional phenomena.
\begin{acknowledgements}
This research was supported in part by Supercomputing Center of
Lanzhou University and NSFC under Grant No.~$11834005$, No.~$11874188$.
We thank the Supercomputing Center of Lanzhou University for allocation of CPU time.
\end{acknowledgements}
|
\section{introduction}
A geodesic metric space $(H,d)$ is called a $\CAT(0)$ space if every geodesic triangle in $X$ is at least as {\em slim} as its corresponding comparison triangle in the Euclidean plane.
For a detailed study of these spaces,
see, e.g.,~\cite{Alexander2019,Bacak,Brid,Burago}. Throughout this article, we only consider \emph{complete} $\CAT(0)$ spaces, which are often called \emph{Hadamard} spaces.
The notion of \emph{weak convergence} on $\CAT(0)$ spaces has appeared in various contexts over the past years, see, e.g.,~\cite{Bacak, Espi, Jost, Kell, Kirk, Lim, Monod, Sosov}.
Recently, Lytchak and Petrunin~\cite{lytchak2021weak} resolved some open problems concerning the existence of topologies that realize weak convergence for \emph{bounded} sequences and nets in $\CAT(0)$. We are here concerned with similar questions \emph{without} the assumption of boundedness.
\begin{mydef}\label{def:weak-conv}
Let $(H,d)$ be a $\CAT(0)$ space, and let $x,y \in H$. Denote by $[x,y]$ the unique geodesic connecting $x$ and $y$, and let $P_{[x,y]}z$ denote the closest point projection of $z \in H$ onto $[x,y]$ with respect to $d$. A net (in particular a sequence) $(x_\alpha)\subseteq H$ {\em converges weakly} to $x\in H$ if $P_{[x,y]}x_\alpha \to x$ with respect to $d$ for all $y\in H\setminus \{x\}$.
\end{mydef}
For Hilbert spaces the above notion of weak convergence coincides with convergence in the usual weak topology. We are concerned with topologies $\tau$ on general CAT(0) spaces that \emph{realize weak convergence of nets}, by which we mean that a net weakly converges to $x$ if and only if it converges to $x$ with respect to $\tau$. Since any sequence is a net, a topology that realizes weak convergence of nets also realizes weak convergence of sequences.
For \emph{bounded} nets (or sequences), the above definition is identical to the one given by Lytchak and Petrunin in~\cite{lytchak2021weak}, who show the following result:
\begin{thm*}[\cite{lytchak2021weak}]
Let $(H,d)$ be a $\mathrm{CAT(0)}$ space. There exists a unique topology $\tau_{LP}$ on $H$ with the following two properties:
\begin{enumerate}[i.]
\item A sequence $(x_{n})$ converges in $H$ with respect to $\tau_{LP}$ to a point $x$ if and only if the sequence is bounded and converges to $x$ weakly in the sense of Definition~\ref{def:weak-conv}.
\item The topology $\tau_{LP}$ is sequential.
\end{enumerate}
\end{thm*}
With reference to the above theorem, recall that a topology $\tau$ on a set $X$ is called \emph{sequential} if every set $A \subseteq X$ that is $\tau$-sequentially closed is $\tau$-closed. The topology $\tau_{LP}$ can be defined as follows:
A subset $A \subseteq H$ is $\tau_{LP}$-closed if and only if for any bounded sequence $(x_n) \subseteq A$
weakly converging to some point $x \in H$, one has that $x\in A$.
Our interest in unbounded nets (and sequences) is in parts motivated by an attempt to generalize weak topologies on Hilbert spaces to CAT(0) spaces. Since the weak topology on an infinite dimensional Hilbert space is never sequential (see, e.g., Example~\ref{ex:Hilbert}), $\tau_{LP}$ does not agree with the usual weak topology due to Property $(ii)$ of the above theorem.
Another source of motivation for studying the unbounded case is the following example due to Monod~\cite[Example 21]{Monod}:
\begin{example}[infinite spike]\label{ex:spike}
Consider the $\mathrm{CAT(0)}$ space $(H,d)$, which consists of the family of intervals $I_n=[0,n]$, $n\in \mathbb{N}$, glued at the common vertex 0. Consider the unbounded sequence $(x_n)$, where $x_n$ is the nonzero endpoint in $I_n$. Then $(x_n)$ converges weakly to $0$ since every point $y\in H\setminus\{0\}$ belongs to $I_N$ for some unique $N \in \mathbb{N}$, and $P_{[0,y]}x_n=0$ for all $n>N$. Being unbounded, however, this sequence does converge with respect to $\tau_{LP}$.
\end{example}
Below we construct two weak topologies on CAT(0) spaces that agree with the usual weak topology on Hilbert spaces and that also realize the convergence of the sequence in Example~\ref{ex:spike}.
\subsection{Definitions of weak topologies}
Consider a CAT(0) space $(H,d)$. Let $\Gamma(H)$ be the set of all compact geodesics in $H$. For $\gamma\in \Gamma(H)$ and $x\in H$ denote by $P_\gamma x$ the projection of $x$ to $\gamma$. For $x,y \in H$, let $[x,y]$ denote the unique geodesic connecting $x$ to $y$.
We define
\begin{align*}
U_{x}(y):=\{z \in H: P_{[x,y]} z \neq y\} ,
\end{align*}
and we call $U_{x}(y)$ an {\em elementary set} around $x$.
The following definition introduces two topologies related to these elementary sets.
\begin{mydef}
\begin{enumerate}[i.]
\item Let $\tau_w$ be the topology defined by $U\in \tau_w$ if and only if for all $x\in U$ there exist finitely many points $y_1, \ldots y_n\in H\setminus \{x\} $ such that
\[ U_x(y_1)\cap \ldots \cap U_x(y_n) \subseteq U. \]
\item Let $\tau_g$ be the the coarsest topology containing $U_x(y)$ for all $x,y\in H$ with $x\neq y$.
\end{enumerate}
\end{mydef}
The construction of $\tau_{w}$ in this definition was suggested by Ba\v cak~\cite{Bacak3} and was also studied in the thesis of the second author.
Besides these topologies we also consider the so-called \emph{coconvex topology} $\tau_{co}$ introduced by Monod~\cite{Monod}. This topology is defined as the coarsest topology on $H$ such that metrically closed convex sets are $\tau_{co}$-closed. Moreover, we denote by $\tau_{d}$ the metric topology.
\subsection{Main results}
One always has the inclusions
\[\tau_{co}\subseteq \tau_w\subseteq \tau_g\subseteq \tau_d, \]
see Proposition~\ref{prop:relations}. In particular, for convex sets it is equivalent to be closed with respect to $\tau_{w}$, $\tau_{g}$, and $\tau_{d}$. Moreover, we show that one always has that
$\tau_{w}\subseteq \tau_{LP}$
and that $\tau_{w}$ and $\tau_{LP}$ agree on \emph{bounded} CAT(0) spaces, see Corollary~\ref{cor:tau-w-tau-LP} and Lemma~\ref{lem:tau-w-tau-LP}.
Due to Proposition 1.3 in~\cite{lytchak2021weak}, one cannot in general expect the existence of a topology that realizes weak convergence of nets (not even in the bounded case). However, we show the following result:
\begin{thm}\label{thm:tau-w-unique}
If there exists a topology on $H$ that realizes weak convergence of nets, then this topology is equal to $\tau_{w}$.
\end{thm}
This theorem directly follows from Corollary~\ref{thm:uniqueness}. Existence a topology that realizes weak convergence can be characterized in terms of properties of $\tau_{w}$ as follows:
\begin{thm}\label{thm:tau-w-exist}
The following statements are equivalent:
\begin{enumerate}
\item $\tau_w$ realizes weak convergence of nets.
\item For all $x\in H$ and $y\in H\setminus \{x\}$ the point $x$ belongs to the $\tau_w$-interior of $U_x(y)$.
\end{enumerate}
In this case, $\tau_w$ satisfies the Hausdorff property.
\end{thm}
The proof of Theorem~\ref{thm:tau-w-exist} can be found in Section~\ref{sec:weak-top-exist}. For the case of bounded CAT(0) spaces, the Hausdorff property of $\tau_{w}$ is even sufficient for the other conditions in Theorem~\ref{thm:tau-w-exist} to hold; see Corollary~\ref{cor:bounded_space_w}.
Notice that the second condition in Theorem~\ref{thm:tau-w-exist} is weaker than to require that elementary sets be open with respect to $\tau_{w}$. In fact, it is in general unclear whether elementary sets are $\tau_{w}$-open, as, e.g., pointed out by Ba\v cak~\cite{Bacak3}. A necessary and sufficient condition for this to be true is that $\tau_{w}=\tau_{g}$. Moreover, we show that $\tau_{w}=\tau_{g}$ is equivalent to the realization of weak convergence of nets by $\tau_{g}$; see Theorem~\ref{thm:proper}.
Example~\ref{ex:book-triang} showcases a scenario where $\tau_{w}\neq \tau_{g}$.
An interesting instance of the equality $\tau_w=\tau_g$ is provided by locally compact CAT(0) spaces: We show that (similar to Banach spaces), the weak topologies $\tau_w$ and $\tau_g$ agree with the metric topology $\tau_{d}$ in this case:
\begin{thm}\label{thm:loc}
$(H,d)$ is locally compact if and only if $\tau_w$ coincides with the metric topology. In particular, in this case we have that $\tau_w=\tau_g=\tau_d$, and weak and strong convergence of nets agree.
\end{thm}
The proof can be found in Section~\ref{sec:loc-cpct}.
Finally, for separable CAT($0$) spaces we show an analogue of the {Eberlein--\v Smulian Theorem}, namely that $\tau_{g}$-compactness is equivalent to $\tau_{g}$-sequential compactness. Moreover, for separable CAT($0$) spaces we show that $\tau_{g}$ is metrizable on any $\tau_{g}$-compact subset, see Theorem~\ref{thm:comp_implies_seq}.
\section{Relations \& first examples}\label{sec:first}
\subsection{Relations between topologies}
We start by showing inclusions of the topologies introduced above.
\begin{proposition}\label{prop:relations}
One has that $\tau_{co}\subseteq \tau_w\subseteq \tau_g\subseteq \tau_d$.
\end{proposition}
\begin{proof}
To prove the first inclusion we assume that $C\subseteq H$ is convex and metrically closed. We show that $H\setminus C$ is $\tau_w$-open. More precisely, we prove that
\begin{align}\label{eq:convex_complement}
U_x(P_C x)\subseteq H\setminus C
\end{align}
for all $x\in H\setminus C$, where $P_C x$ is the metric projection of $x$ onto $C$. To this end first notice that if $z\in [x,P_C x]$, then
\[ d(z, P_C x)=d(x, P_C x)-d(x,z)\leq d(x, P_C z)-d(x,z) \leq d(z,P_C z),\]
which implies that $P_C x=P_C z$. Now let $c\in C$. Then \[ d(c, P_C x)= d(P_C c, P_C z)\leq d(c,z)\] for all $z\in [x,P_C x]$, since $P_C$ is nonexpansiveness in CAT($0$) spaces. Hence $P_{[x,P_C x]}c= P_C x$. This proves $C\subseteq P_{[x,P_C x]}^{-1}(\{P_C x\})$, which yields \eqref{eq:convex_complement} by taking complements. This proves $\tau_{co}\subseteq \tau_w$. \\
We now turn to the middle inclusion.
By definition of $\tau_w$ every $\tau_w$-open set $U$ is a union $U=\cup_{x\in U}U_x$ where $U_x \subseteq U$ is an intersection of finitely many elementary sets around $x$. Hence $U$ is $\tau_g$-open as $\tau_g$ consists of unions of finite intersections of elementary sets. \\
For the last inclusion is suffices to note that elementary sets are metrically open as they are preimages of metrically open sets under continuous projections.
\end{proof}
\begin{corollary}[closed convex sets]\label{cor:convex}
Let $C\subseteq H$ be convex. Then $C$ is $\tau_w$-closed if and only if $C$ is $\tau_g$-closed if and only if $C$ is metrically closed.
\end{corollary}
\begin{proof}
This is a consequence of Proposition \ref{prop:relations}, since it is straightforward to verify that every topology $\tau$ on $H$ with $\tau_{co}\subseteq \tau\subseteq \tau_d$ satisfies this equivalence.
\end{proof}
\begin{proposition}\label{prop:convex_preimage}
We have $\tau_{co}=\tau_w=\tau_g$ if and only if $U_x(y)$ is $\tau_{co}$-open for all $x,y\in H$ with $x\neq y.$ In particular, if the preimages $P_{[x,y]}^{-1}(\{y\})$ are convex for all $x,y\in H$ with $x\neq y$, then $\tau_{co}=\tau_w=\tau_g.$
\end{proposition}
\begin{proof}
The first statements follows from the fact that $\tau_{co}\subseteq \tau_g$ and the definition of $\tau_g$. For the second statement note that $P_{[x,y]}^{-1}(\{y\})=H\setminus U_x(y)$ is metrically closed. Hence convexity of $P_{[x,y]}^{-1}(\{y\})$ implies that $U_x(y)\in \tau_{co}.$
\end{proof}
\begin{remark}\label{rem:geod-monotone}
"The property of the nice projection onto geodesics"
(property $(N)$ for short) that was introduced in \cite{Espi} says: For all $\gamma\in \Gamma(H)$ and $x,y,m\in H$ we have $P_\gamma m \in [P_\gamma x,P_\gamma y] $ if $m\in [x,y]$. Moreover, we consider the $(Q_4)$ condition
\[ d(x,p)<d(x,q) \text{ and } d(y,p)<d(y,q) \quad \text{implies}\quad d(m,p)\leq d(m,q) \ \forall m\in [x,y].\]
This condition was studied in \cite{Kirk}. We also consider the slightly stronger $(\overline{Q}_4)$ condition
\[ d(x,p)\leq d(x,q) \text{ and } d(y,p)\leq d(y,q) \quad \text{implies}\quad d(m,p)\leq d(m,q) \ \forall m\in [x,y],\]
which was used in \cite{Kakavandi} as a sufficient condition for the realization of weak convergence.
In~\cite[Theorem 5.7]{Espi} it is shown that any $\mathrm{CAT(0)}$ space of constant curvature satisfies the $(Q_4)$ condition.\\
Clearly, the $(\overline{Q}_4)$ condition implies the $(Q_4)$ condition. Furthermore, the $(Q_4)$ condition implies property $(N)$, see~\cite[Lemma 5.15]{Espi}.
If property $(N)$ is satisfied, then $P_{[x,y]}^{-1}(\{y\})$ is convex for all $x,y\in H$ with $x\neq y$. Hence, due to Proposition \ref{prop:convex_preimage}, each of the conditions $(Q_4),$
$(\overline{Q}_4)$ and $(N)$ imply $\tau_{co}=\tau_w=\tau_g$, and we show that these topologies realize weak convergence of nets in this case (see Theorem~\ref{thm:proper}).
\end{remark}
\subsection{Examples}
We provide three simple examples of unbounded CAT($0$) spaces on which the topologies $\tau_g$ and $\tau_w$ agree with the coconvex topology $\tau_{co}$.
\begin{example}[Hilbert spaces]\label{ex:Hilbert}
If $(H,d)$ is a Hilbert space, then $\tau_{co}=\tau_w=\tau_g$ agree with the usual weak topology on $H$: The usual weak topology is finer than $\tau_g$ as for $x,y\in H$ with $x\neq y$ we have
\[ U_x(y)= \left\{ z\in H \colon \, \mathrm{Re}\left( \langle b, z-x \rangle\right)<1 \right\} \]
with $b=\frac{y-x}{\|y-x\|^2}$. Therefore, $U_x(y)$ is open with respect to the usual weak topology. Moreover, the usual weak topology is generated by convex half spaces and hence coarser than $\tau_{co}$ (see also \cite[Ex. 18]{Monod}). Therefore, the claim follows from Proposition \ref{prop:relations}. \\
The following argument shows that the topology $\tau_{LP}$ is strictly finer than the usual weak topology on $H$ if $H$ is infinite dimensional: If $A$ is closed in the usual weak topology, then it is weakly sequentially closed. Therefore, it is $\tau_{LP}$-closed. This shows that $\tau_{LP}$ is finer than the usual weak topology. To see that it is strictly finer consider an orthonormal sequence $(e_n)\subseteq H$. The set $A:=\{ \sqrt{n} e_n \colon n\in \mathbb{N}\}$ is weakly sequentially closed, but 0 is contained in the weak closure of $A$. Hence $A$ is closed in $\tau_{LP}$, but not closed in the usual weak topology.
\end{example}
\begin{example}[infinite spike revisited]
Consider the infinite spike from Example~\ref{ex:spike}. Elementary sets take one of two shapes: (i) half open intervals $(a, n]\subseteq I_{n}$, $a\geq 0$, and (ii) complements of closed intervals $[a, n]\subseteq I_{n}$, $a>0$. Therefore, elementary sets are $\tau_{co}$-open, and Proposition~\ref{prop:relations} implies that $\tau_{co}=\tau_{w}=\tau_{g}$.
Furthermore, the sequence $(x_{n})$ of nonzero endpoints converges to $0$ with respect to the latter topologies, but not with respect to $\tau_{LP}$. The fact that $\tau_{co}\subseteq \tau_{LP}$ (see \cite[Prop. 3.1]{lytchak2021weak}) then implies that $\tau_{LP}$ is strictly finer than $\tau_{co}=\tau_w=\tau_g$ on this space.
Additionally, the infinite spike is compact and sequentially compact with respect to the topology $\tau_{co}=\tau_{w}=\tau_{g}$.
\end{example}
\begin{example}[infinite dimensional hyperbolic space]
Consider the upper half space model of infinite dimensional (separable) real hyperbolic space $\mathbb{H^{\infty}}$. In this model, elementary sets take the form of one of the two connected components that arise when removing a half sphere that touches the boundary of $\mathbb{H^{\infty}}$ orthogonally. Therefore, elementary sets are $\tau_{co}$-open, and Proposition~\ref{prop:relations} implies that $\tau_{co}=\tau_{w}=\tau_{g}$.
\end{example}
\section{Results on \texorpdfstring{$\tau_w$}{tauw}}\label{sec:tau-w}
This section is devoted to the topology $\tau_w$. As a main result we provide the following uniqueness statement: Whenever there exists a topology on $H$ realizing weak convergence of nets, then this topology agrees with $\tau_w$. Moreover, we give a characterization for when $\tau_w$ realizes weak convergence. Before stating these results we require some preliminaries.
\subsection{Realization of weak convergence}\label{sec:weak-top-exist}
First observe that weak convergence always implies convergence with respect to $\tau_w$:
\begin{proposition}\label{prop:w-convergence}
Let $(x_\alpha)\subseteq H$ be a net and $x\in H$. Then
$x_\alpha \to x$ weakly implies $x_\alpha \to x$ with respect to $\tau_w$.
\end{proposition}
\begin{proof}
Let $U$ be $\tau_w$-open with $x\in U$. Then there exists finitely many $y_1, \ldots y_n \in H\setminus \{x\}$ such that
\[ U_x(y_1)\cap \ldots \cap U_x(y_n) \subseteq U. \]
Weak convergence of $x_{\alpha}\to x$ implies that $P_{[x,y_i]} x_\alpha \to x$ with respect to $d$. Hence there exists $\alpha_i$ such that $x_\alpha \in U_x(y_i)$ for all $\alpha \succcurlyeq \alpha_i.$ Let $\alpha_0\succcurlyeq \alpha_i$ for $i=1,\ldots,n$, then $x_\alpha\in U$ for all $\alpha \succcurlyeq \alpha_0.$
\end{proof}
The following lemma concerns properties of elementary sets and will be used frequently in the sequel.
\begin{lemma} \label{prop:elementary_incl}
Let $x,y\in H$ with $x\neq y$, and let $a\in [x,y]$ different from $x$ and $y$. If $z\in U_x(a)$, then $P_{[x,y]}z\in [x,a)$. In particular,
$U_x(a) \subseteq U_x(y)$. Moreover, $U_x(a)\cap U_y(a)= \emptyset$.
\end{lemma}
\begin{proof}
Let $z\in U_x(a)$. Then the convex function $\phi: [x,y] \to \mathbb{R}$, $u\mapsto d(u,z)$ (see \cite[Ex. 2.2.4]{Bacak}), satisfies $\phi(P_{[x,a]}z)={d(P_{[x,a]}z,z)< d(a,z)} =\phi(a)$. Hence $\phi(a)< \phi(u)$ for all $u\in (a,y]$. Therefore, the function $\phi$ attains its minimum in $[x,a).$ Hence, $P_{[x,y]}z\in [x,a)$, which shows the first statement. Since $[x,a)\subseteq [x,y)$, we obtain that $z\in U_x(y).$ Finally, $z\in U_x(a)\cap U_y(a)$ would imply that $\phi(P_{[x,a]})< \phi(a)$ and $\phi(P_{[a,y]}z)< \phi(a)$, contradicting the convexity of $\phi$.
\end{proof}
The following result characterizes $\tau_w$-closed subsets as exactly the ones that are net-closed with respect to weak convergence.
\begin{proposition}[$\tau_w$-closed sets]\label{lem:w-closed-sets}
Let $A\subseteq H$ be a subset. Then the following statements are equivalent:
\begin{enumerate}
\item $A$ is $\tau_w$-closed.
\item Whenever a net in $A$ converges weakly to some $x\in H$, then $x\in A$.
\end{enumerate}
\end{proposition}
\begin{proof}
\item[$(1)\Rightarrow(2)$:] Suppose that $A$ is $\tau_w$-closed, and let $(x_\alpha)$ be a net in $A$ and $x\in H$ with $x_\alpha \to x$ weakly. Proposition \ref{prop:w-convergence} yields $x_\alpha \to x$ with respect to $\tau_w$. Therefore, $x\in A$.
\item[$(2)\Rightarrow(1)$:] To prove the contrapositive we assume that $A$ is not $\tau_w$-closed. Then $H\setminus A$ is not $\tau_w$-open. Hence there exists $x\in H\setminus A$ such that for every finite subset $F\subseteq H\setminus \{x\}$ there exists
\[ x_F \in \bigcap_{z\in F} U_x(z) \quad\text{with} \quad x_F \in A. \]
The family $\mathcal{F}$ of finite subsets of $H\setminus \{x\}$ is a directed set by set inclusion. Together with the above construction we obtain a net $(x_F)_{F \in \mathcal{F}} \subseteq A$. We claim that this net converges weakly to $x$. Let $y\in H$ with $x\neq y$ and $0<\varepsilon < d(x,y)$. Let $y_\epsilon\in [x,y]$ such that $d(x,y_\epsilon)=\varepsilon$, and consider the one point set $F_\varepsilon=\{y_\epsilon\}$. If $F\in \mathcal{F}$ with $F\supseteq F_\varepsilon$, then $x_F\in U_x(y_\epsilon)$. Lemma \ref{prop:elementary_incl} yields $P_{[x,y]} x_F \in [x,y_\varepsilon).$ Therefore, $d(x, P_{[x,y]}x_F)<\varepsilon$, proving weak convergence of $(x_F)_{F\in \mathcal{F}}$ to $x$.
\end{proof}
With Propositions~\ref{prop:w-convergence} and \ref{lem:w-closed-sets} at hand, we deduce a statement that relates weak convergence of nets to the topology $\tau_w$. In particular, Theorem~\ref{thm:tau-w-unique} is a direct consequence.
\begin{corollary}\label{thm:uniqueness}
\begin{enumerate}[1.]
\item Suppose that $\tau$ is a topology on $H$ such that $\tau$-convergence of nets implies weak convergence of nets. Then $\tau_w\subseteq \tau$.
\item Suppose $\tau$ is a topology on $H$ such that weak convergence of nets implies $\tau$-convergence of nets. Then $\tau\subseteq \tau_w$.
\end{enumerate}
\end{corollary}
\begin{proof}
\begin{enumerate}[1.]
\item Due to Proposition \ref{prop:w-convergence} the identity map $(H,\tau)\rightarrow (H,\tau_w)$ is continuous.
\item Suppose $A\subseteq H$ is $\tau$-closed. Let $(x_\alpha)$ be a net in $A$ and $x\in H$ with $x_\alpha \to x$ weakly. Then $x_\alpha \to x$ with respect to $\tau$. Hence $x\in A$. Therefore the statement follows from Proposition~\ref{lem:w-closed-sets}. \qedhere
\end{enumerate}
\end{proof}
We are now in the position to complete the proof of Theorem~\ref{thm:tau-w-exist}.
\begin{proof}[Proof of Theorem~\ref{thm:tau-w-exist}
\item[$(1)\Rightarrow(2)$:] We prove the contrapositive implication. Suppose that there exist $x\in H$ and $y\in H\setminus \{x\}$ such that $x$ does not belong to the $\tau_w$-interior of $U_x(y)$. Then for every $\tau_w$-open set $U$ with $x\in U$ there exists $x_U\in U$ with $x_U\notin U_x(y)$. The family $\mathcal{U}:=\{ U\in \tau_w\colon x\in U \}$ is directed with respect to reverse set inclusion, and the net $(x_U)_{U \in \mathcal{U}}$ converges to $x$ with respect to $\tau_w$. On the other hand $P_{[x,y]} x_U=y$ for all $U\in \mathcal{U}$. Therefore, $(x_U)_{U\in \mathcal{U}}$ does not weakly converge to $x$. Hence $\tau_w$ does not realizes weak convergence.
\item[$(2)\Rightarrow(1)$:]
Let $(x_\alpha)$ be a net in $H$ and $x\in H$. Due to Proposition \ref{prop:w-convergence} it suffices to show that $x_\alpha \rightarrow x$ with respect to $\tau_w$ implies $x_\alpha \rightarrow x$ weakly. To this end let $y\in H$ with $x\neq y$ and $0<\varepsilon< d(x,y)$. Let $y_\epsilon\in [x,y]$ such that $d(x,y_\epsilon)=\varepsilon$. By $(2)$ there exists a $\tau_w$-open set $U\subseteq U_x(y_\varepsilon)$ with $x\in U$. Therefore, we obtain $x_\alpha\in U_x(y_\varepsilon)$ eventually. By Lemma \ref{prop:elementary_incl} this yields $P_{[x,y]}x_\alpha\in [0,y_\varepsilon)$ and therefore $d(x,P_{[x,y]}x_\alpha)<\varepsilon$ eventually. \\
For the last statement in Theorem~\ref{thm:tau-w-exist} one can argue either through uniqueness of weak limits or use the last statement in Lemma \ref{prop:elementary_incl} together with $(2)$.
\end{proof}
We end this subsection with three statements that relate the topology $\tau_w$ to the topology $\tau_{LP}$ introduced by Lytchak and Petrunin.
\begin{corollary}\label{cor:tau-w-tau-LP}
It holds that $\tau_w\subseteq \tau_{LP}$.
\end{corollary}
\begin{proof}
Suppose $A\subseteq H$ is $\tau_w$-closed. Let $(x_n)\subseteq A$ be a bounded sequence weakly converging to some $x$ in $H$. Then Proposition \ref{lem:w-closed-sets} yields $x\in A$. Hence $A$ is $\tau_{LP}$-closed.
\end{proof}
\begin{lemma}\label{lem:tau-w-tau-LP}
Let $A\subseteq H$ be a bounded subset. Then $A$ is $\tau_{LP}$-closed if and only if $A$ is $\tau_w$-closed.
\end{lemma}
\begin{proof}
Suppose $A$ is $\tau_{LP}$-closed. We use the characterization of $\tau_w$-closed sets in Proposition \ref{lem:w-closed-sets} to show that $A$ is $\tau_w$-closed. To this end let $(x_\alpha)\subseteq A$ be a net weakly converging to a point $x\in H.$ Then Lemma $5.2$ in \cite{lytchak2021weak} yields $x_\alpha\to x$ in $\tau_{LP}.$ Hence $x\in A$. The other implication is due to the previous Corollary.
\end{proof}
\begin{corollary}\label{cor:w-equal-LP}
If $(H,d)$ is bounded, then $\tau_w= \tau_{LP}$. In particular, $\tau_w$ is sequential for bounded Hadamard spaces.
\end{corollary}
\subsection{Weak convergence of bounded sequences and nets}
It follows from Corollary~\ref{cor:w-equal-LP} that $\tau_w$ always realizes weak convergence of bounded sequences. Nevertheless, in order to keep our paper self-contained, we provide another proof of this fact that does not invoke results from \cite{lytchak2021weak}. \\
Recall that a topological space is compact if and only if every net has a convergent subnet.
\begin{lemma}[compactness]\label{lem:w-compact}
If $K\subseteq H$ is bounded and $\tau_w$-closed, then $K$ is $\tau_w$-compact. In particular, every metrically closed ball is $\tau_w$-compact.
\end{lemma}
\begin{proof}
First notice that due to \cite[Proposition 3.5.]{Kirk} every net in the bounded set $K$ has a weakly convergent subnet. Together with Proposition \ref{prop:w-convergence} this yields compactness of $K$. The second statement follows from Corollary \ref{cor:convex} since metrically closed balls are convex.
\end{proof}
\begin{proposition}
Let $x\in H$ and $(x_n)$ be a sequence in $H\setminus \{x\}$ . If $x_n\to x$ weakly, then
the set \[ A:=\{ x_n \colon n\in \mathbb{N} \} \cup \{x\} \] is $\tau_w$-closed.
\end{proposition}
\begin{proof}
Let $y$ be an element in the complement of $A$. Denote by $m$ the midpoint between $x$ and $y$. Then there exists $N\in \mathbb{N}$ such that $x_n\in U_x(m)$ for all $n > N$. Moreover, $x\in U_x(m)$. Lemma \ref{prop:elementary_incl} yields $x_n \notin U_y(m)$ for all $n>N$ and $x\notin U_y(m)$. Hence
\[ U_y(m) \cap U_y(x_1) \cap \ldots \cap U_y(x_N) \subseteq H\setminus A.\]
Therefore, $H\setminus A$ is $\tau_w$-open.
\end{proof}
\begin{lemma}[bounded sequences]
Let $(x_n)$ be a bounded sequence in $H$ and $x\in H$. Then $x_n \rightarrow x$ with respect to $\tau_w$ if and only if $x_n\rightarrow x$ weakly.
\end{lemma}
\begin{proof}
Without loss of generality we may assume that $x_n\neq x$ for all $n\in \mathbb{N}$.
Suppose $x_n \rightarrow x$ with respect to $\tau_w$ and let $(x_{m})$ be a subsequence of $(x_n)$. Due to boundedness of $(x_{m})$ there exists a subsequence $(x_k)$ of $(x_{m})$ and $y\in H$, such that $x_k\to y$ weakly \cite[Section 3.]{Kirk}. By the last proposition $A:=\{ x_k \colon k\in \mathbb{N} \} \cup \{y\}$ is $\tau_w$-closed. The convergence of $(x_k)$ to $x$ with respect to $\tau_w$ shows that $x\in A$. Hence $x=y.$ Hence every subsequence of $(x_n)$ has a subsequence that weakly converges to $x$. This shows that $x_n \rightarrow x$ weakly. The converse is due to Proposition \ref{prop:w-convergence}.
\end{proof}
\begin{lemma}[bounded nets]\label{lem:bounded nets}
Suppose that $(H,\tau_w)$ is Hausdorff. Let $(x_\alpha)$ be a bounded net in $H$ and $x\in H$. Then $x_\alpha \rightarrow x$ with respect to $\tau_w$ if and only if $x_\alpha\rightarrow x$ weakly.
\end{lemma}
\begin{proof}
Suppose $x_\alpha \rightarrow x$ with respect to $\tau_w$ and let $(x_\beta)$ be a subnet of $(x_\alpha)$. Boundedness of $(x_\beta)$ yields the existence of a subnet $(x_\delta)$ of $(x_\beta)$ that converges weakly to some $y\in H$. Due to Proposition \ref{prop:w-convergence} we obtain $x_\delta \rightarrow y$ with respect to $\tau_w$. Moreover, $x_\delta\rightarrow x$ with respect to $\tau_w$. The Hausdorff property implies $x=y$. Hence every subnet of $(x_\alpha)$ has a subnet that weakly converges to $x$. This shows that $x_\alpha \rightarrow x$ weakly. The converse is due to Proposition \ref{prop:w-convergence}.
\end{proof}
\begin{corollary}[bounded spaces] \label{cor:bounded_space_w}
Suppose $(H,d)$ is bounded. Then the following statements are equivalent:
\begin{enumerate}
\item $\tau_w$ realizes weak convergence.
\item For all $x\in H$ and $y\in H\setminus \{x\}$ the point $x$ belongs to the $\tau_w$-interior of $U_x(y)$.
\item $\tau_w$ satisfies the Hausdorff property.
\end{enumerate}
\end{corollary}
\begin{proof}
This follows directly from Theorem \ref{thm:tau-w-exist} and Lemma \ref{lem:bounded nets}.
\end{proof}
\begin{remark}
In particular, on a bounded $\mathrm{CAT(0)}$ space the conditions of the previous corollary are satisfied if $\tau_{co}$ is Hausdorff since $\tau_{co}\subseteq \tau_{w}$. Moreover, by Lemma~\ref{lem:w-compact}, $\tau_{w}$ is compact on a bounded space; hence $\tau_{co} = \tau_{w}$ (since two topologies that are both compact and Hausdorff and where one topology contains the other must agree).
\end{remark}
\begin{remark}
It remains open whether the boundedness assumption in Corollary \ref{cor:bounded_space_w} can be dropped or if there exists a space $H$ such that $(H,\tau_w)$ is Hausdorff but $\tau_w$ does not realize weak convergence.
\end{remark}
\section{Results on \texorpdfstring{$\tau_g$}{taug}}
\subsection{Continuous projections}
We show that $\tau_g$ is the initial topology with respect to the projections $P_\gamma\colon (H,\tau_g)\rightarrow (\gamma,d)$ for $\gamma \in \Gamma(H)$. Moreover, we characterize spaces for which
$\tau_g$ realizes weak convergence.
\begin{lemma}[continuous projections] \label{lem:proj}
The projection map $P_\gamma\colon (H,\tau_g)\rightarrow (\gamma,d)$ is continuous for all $\gamma \in \Gamma(H)$. Moreover, $\tau_g$ is the coarsest topology on $H$ with this property.
In other words, $\tau_g$ is the initial topology of the family of maps $P_\gamma\colon (H,\tau_g)\rightarrow (\gamma,d)$ for $\gamma\in \Gamma(H)$.
\end{lemma}
\begin{proof}
Let $ \gamma=[a,b] \in \Gamma(H)$. To show
the first statement it suffices to show that $P_\gamma^{-1} ([x,y])$ is $\tau_g$-closed for all $x,y\in \gamma$ with $d(a,x)\leq d(a,y)$. To this end, we will show the following identity
\begin{align}\label{eq:preimage_of_closed_interval}
P_\gamma^{-1} ([x,y]) = \left(H\setminus U_a(x) \right) \cap \left( H\setminus U_b(y) \right).
\end{align}
First let $z\in P_\gamma^{-1} ([x,y])$. As the function $u\mapsto d(u,z)$ on $(\gamma,d)$ is convex and attains its unique minimum in $[x,y]$ we have that $d(x,z)\leq d(u,z)$ for all $u\in [a,x]$. Hence $P_{[a,x]}z=x$, i.e. $z\notin U_a(x) $. Likewise, $d(y,z)\leq d(u,z)$ for all $u\in [y,b]$ shows $P_{[b,y]}z=y$ and hence $z\notin U_b(y)$.\\
Now suppose $z$ belongs to the right hand side in \eqref{eq:preimage_of_closed_interval}. Then $P_{[a,x]}z=x$ and therefore $d(z, P_\gamma z)\leq d(z,x) < d(z,u)$ for all $u\in [a,x)$. Likewise, $P_{[b,y]}z=y$ yields $d(z, P_\gamma z)\leq d(z,y) < d(z,u)$ for all $u\in (y,b]$. Hence
\[ P_\gamma z\in \gamma\setminus ([a,x) \cup (y,b])= [x,y]. \]
This proves \eqref{eq:preimage_of_closed_interval} and therefore the first statement. \\
To prove the second statement suppose that $\tau$ is a topology on $H$, such that $P_\gamma\colon (H,\tau)\rightarrow (\gamma,d)$ is continuous for all $\gamma\in \Gamma(H)$. Let $x,y\in X$. The subset $[x,y)$ is open in $([x,y],d)$. Hence $U_x(y)=P_{[x,y]}^{-1}([x,y))\in \tau.$ Therefore, $\tau_g\subseteq \tau.$
\end{proof}
\begin{corollary}[Hausdorff property]\label{cor:haus}
$(H,\tau_g)$ satisfies the Hausdorff property.
\end{corollary}
\begin{proof}
Due to Lemma \ref{lem:proj} and since $(\gamma,d)$ is Hausdorff for all $\gamma\in \Gamma(H)$, it suffices to show that the family of projections $P_\gamma\colon H \rightarrow \gamma$ separates points. To this end let $x$ and $y$ be distinct points in $H$. Then $P_{[x,y]} x=x \neq y=P_{[x,y]} y.$
\end{proof}
Another consequence is that we can describe convergence of nets in $\tau_g$ and compare it to weak convergence.
\begin{corollary}[convergence in $\tau_g$]\label{cor:g-convergence}
Let $(x_\alpha)\subseteq H$ be a net and $x\in H$. Then $x_\alpha \to x$ with respect to $\tau_g$ if and only if $P_\gamma x_\alpha \rightarrow P_\gamma x$ with respect to $d$ for all $\gamma\in \Gamma(H)$. In particular, $x_\alpha \to x$ with respect to $\tau_g$ implies $x_\alpha \to x$ weakly.
\end{corollary}
\begin{proof}
The first statement is immediate by Lemma \ref{lem:proj} and convergence in initial topologies. For the second statement let $y\in H\setminus \{x\}$, then the first statement yields $P_{[x,y]}x_\alpha \rightarrow P_{[x,y]} x=x$ with respect to $d$. Hence $x_\alpha \to x$ weakly.
\end{proof}
The following result provides several characterizations for when the two weak topologies $\tau_w$ and $\tau_g$ agree.
\begin{thm}\label{thm:proper}
The following statements are equivalent:
\begin{enumerate}
\item $\tau_w=\tau_g$.
\item For all $x,y\in H$ with $x\neq y$ the elementary set $U_x(y)$ is $\tau_w$-open.
\item $P_\gamma: (H,\tau_w) \rightarrow (\gamma,d)$ is continuous for all $\gamma\in \Gamma(H)$.
\item Weak convergence of a net $(x_{\alpha})$ to a point $x\in H$ implies that $P_{\gamma} x_\alpha \rightarrow P_{\gamma} x$ with respect to $d$ for all $\gamma \in \Gamma(H)$.
\item $\tau_g$ realizes weak convergence of nets.
\end{enumerate}
\end{thm}
\begin{proof}
Due to Proposition \ref{prop:relations} we always have $\tau_w\subseteq \tau_g$. Since $\tau_g$ is generated by elementary sets the equivalence of $(1)$ and $(2)$ is evident. Moreover, we have:
\item[$(1)\Rightarrow(3)$:] See Lemma \ref{lem:proj}.
\item[$(3)\Rightarrow(4)$:] This implication follows from Proposition \ref{prop:w-convergence}.
\item[$(4)\Rightarrow(5)$:] In view of the last statement in Corollary \ref{cor:g-convergence} it suffices to show that weak net-convergence implies net-convergence in $\tau_g$. To this end let $(x_\alpha)\subseteq H$ be a net that converges weakly to $x\in H.$ By $(4)$ we obtain $P_\gamma x_\alpha \to P_\gamma x$ in $d$ for all $\gamma\in \Gamma(H)$. Hence the first statement in Corollary \ref{cor:g-convergence} yields $x_\alpha\to x$ with respect to $\tau_g$.
\item[$(5)\Rightarrow(1)$:] See Theorem~\ref{thm:tau-w-unique}.
\end{proof}
\begin{example}[book of triangles]\label{ex:book-triang}
The following example provides an instance where $\tau_{w}\neq \tau_{g}$.
Consider a Euclidean isosceles right triangle $T$. Label one of its catheti as $a$. Let $A$ denote the vertex at the right angle, let $B$ denote the other vertex on $a$, and let $C$ denote the remaining vertex. Now consider the $\mathrm{CAT(0)}$ space $H$ consisting of countably many copies of $T$ that are isometrically glued along $a$. Enumerate the copies of $T$ in $H$ by $T_{1}, T_{2}, \dots$. For each $n \in \mathbb{N}$, let $C_{n}\in T_{n}$ denote the vertex corresponding to $C\in T$. Due to glueing, the vertices $A$ and $B$ belong to every triangle $T_{n}$.
Let $P$ denote the midpoint of the hypothenuse in $T_{1}$. Then the elementary set $U_{P}(B)$ contains $A$.
We claim that there is no finite intersection of elementary sets around $A$ that is contained in $U_{P}(B)$. In order to see this, choose some point $D\neq A$ in $H$. Then either $D \in T_{n}$ for all $n \in \mathbb{N}$ or $D \in T_{k}$ for exactly one $k\in \mathbb{N}$. Consider the elementary set $U_{A}(D)$. The projection of $C_{n}$ to the geodesic $[A, D]$ is equal to $A$ for for all $n \in \mathbb{N}$ except possibly for $n=k$. Hence, any finite intersection of elementary sets around $A$ contains all but finitely many of the points $C_{1}, C_{2}, \dots$. On the other hand, none of the points $C_{2}, C_{3}, \dots $ belong to $U_{P}(B)$, since their projection onto the geodesic $[P,B]$ is equal to $B$. This proves the claim. In particular, the elementary set $U_{P}(B)$ is not $\tau_{w}$-open; hence $\tau_{w}\neq \tau_{g}$.
\end{example}
\begin{remark}
Consider the compact space that consists of the first three pages of the book of triangles constructed in Example~\ref{ex:book-triang}. On this space, we have that $\tau_{co}=\tau_{w}=\tau_{g}=\tau_{d}$, since the coconvex topology agrees with the metric topology on compact spaces, see~\cite[Lemma 17]{Monod}. However, even for this simple case property $(N)$ from Remark~\ref{rem:geod-monotone} fails to hold. Indeed, the projection of $C_{2}$ and $C_{3}$ to the geodesic $[P,B]$ is equal to $B$, whereas the midpoint of $[C_{2}, C_{3}]$ is $A$, which projects to $P$.
\end{remark}
\begin{remark}
Theorem~\ref{thm:proper} suggests that the coincidence of $\tau_w$ and $\tau_g$ might be a structurally interesting property of $\mathrm{CAT}(0)$ spaces since it might perhaps serve as an analogue of reflexivity in the category of Banach spaces. For example, Theorem~\ref{thm:loc} shows that $\tau_{w}=\tau_{g}$ on all locally compact $\mathrm{CAT}(0)$ spaces.\\
Nonetheless, it is not clear to us whether the coincidence $\tau_{w}=\tau_{g}$ on a $\mathrm{CAT}(0)$ space $(H,d)$ is inherited by closed convex subsets $C\subset H$. While it is not hard to see that this is true if $C$ is bounded, we are not able to prove this permanence property for unbounded $C$. In particular, it is not clear to us whether for an unbounded net $(x_\alpha)\subset C$ and $x\in C$ weak convergence $x_\alpha\rightarrow x$ in the $\mathrm{CAT}(0)$ space $(C,d)$ implies weak convergence in $(H,d)$.
\end{remark}
\subsection{Compactness in \texorpdfstring{$\tau_g$}{taug} }
In this section we study compactness properties of $\tau_g$. We start with observing that $\tau_g$-compactness implies the equality of $\tau_g$ and $\tau_w$.
\begin{proposition}[compactness in $\tau_g$] \label{lem:g-compact}
If $(H,\tau_g)$ is compact, then $\tau_g=\tau_w$.
\end{proposition}
\begin{proof}
Let $x,y\in H$ with $x\neq y$. We prove that the elementary set $U_x(y)$ is $\tau_w$-open, which implies the result due to Theorem \ref{thm:proper}. First notice that
\[A= H\setminus U_x(y)\] is $\tau_g$-closed. Therefore, $A$ is $\tau_g$-compact. Now let $z\in U_x(y)$. For $a\in A$ let $m_a$ be the midpoint between $z$ and $a.$ Clearly, $a\in U_a(m_a)$ and $U_a(m_a)$ is $\tau_g$-open. Moreover, $U_a(m_a)\cap U_z(m_a)=\emptyset$ (see Lemma \ref{prop:elementary_incl}). Since $U_a(m_a)$ for $a\in A$ forms an open cover of $A$ there exists finitely many $a_1,\ldots,a_n\in A$ such that $U_{a_1}(m_{a_1}),\ldots,U_{a_n}(m_{a_n})$ cover $A$. Hence
\[ U_z(m_{a_1}) \cap \ldots \cap U_z(m_{a_n})\subseteq H\setminus A = U_x(y), \] which proves $\tau_w$-openness of $U_x(y)$.
\end{proof}
\begin{corollary}
Suppose that $(H,d)$ is bounded. Then $(H,\tau_g)$ is compact if and only if $\tau_g=\tau_w$.
\end{corollary}
\begin{proof}
See Proposition~\ref{lem:g-compact} and Lemma~\ref{lem:w-compact}.
\end{proof}
Notice that the \emph{book of triangles} constructed in Example~\ref{ex:book-triang} is bounded but cannot be $\tau_{g}$-compact due to the previous corollary.
In the sequel we show an analogue of the {Eberlein--\v Smulian Theorem} and a metrizability result for $\tau_{g}$-compact subsets of separable CAT($0$) spaces.
Aiming for an application of Tychonoff's theorem we consider the map
\[ \Phi \colon (H,\tau_g) \rightarrow \prod_{\gamma\in \Gamma(H)} (\gamma,d) \quad\text{given by} \quad (\Phi(x))_{\gamma}= P_\gamma x. \]
\begin{proposition}[homeomorphism onto image]\label{prop:homeo}
Let $\Phi(H)\subseteq \prod_{\gamma\in \Gamma(H)} (\gamma,d)$ carry the subspace topology $\tau_\textrm{im}$. Then \[ \Phi \colon (H,\tau_g)\rightarrow (\Phi(H),\tau_\textrm{im}) \] is a homeomorphism.
\end{proposition}
\begin{proof}
The same argument as in the proof of Corollary \ref{cor:haus} shows that $\Phi$ is injective. We write $\pi_\gamma \colon \prod_{\gamma\in \Gamma(H)} (\gamma,d)\rightarrow (\gamma,d)$ for the natural projections. Then due to Lemma \ref{lem:proj} the maps $\pi_\gamma \circ \Phi=P_\gamma$ are continuous. The universal property of the product spaces yields that $\Phi$ is continuous. The continuity of the inverse $\Phi^{-1}$ follows from continuity of $P_\gamma\circ \Phi ^{-1}=\pi_\gamma$ together with the universal property of the initial topology $\tau_g$; see Lemma \ref{lem:proj}.
\end{proof}
\begin{thm} \label{thm:comp_implies_seq}
Assume that $(H,d)$ is separable. Let $K\subseteq H$ be a subset. Then $K$ is $\tau_g$-compact if and only if it is $\tau_g$-sequentially compact. In this case, the subspace topology of $\tau_g$ on $K$ is metrizable.
\end{thm}
\begin{proof}
First assume that $K$ is $\tau_g$-compact.
Let $D\subseteq (H,d)$ be countable and dense. Let $\Gamma_D\subseteq \Gamma(H)$ be the set of all $\gamma\in \Gamma(H)$ with both endpoints in $D$. Then $\Gamma_D$ is countable.
Consider the map
\[ \Phi_D \colon (H,\tau_{g}) \rightarrow \prod_{\gamma\in \Gamma_D} (\gamma,d) \quad\text{given by} \quad (\Phi_{D}(x))_{\gamma}= P_\gamma x. \]
The same argument as in the proof of Proposition \ref{prop:homeo} shows that $\Phi_D$ is continuous. We claim that $\Phi_D$ is injective. To see this, let $x,y\in H$ with $x\neq y.$ We set $r=d(x,y)$. Let $\gamma$ be the geodesic connecting $x$ and $y$. There exist \[ x_D, y_D\in D \quad\text{such that}\quad d(x,x_D)< \frac{r}{3} \quad \text{and}\quad d(y,y_D)< \frac{r}{3} .\]
Let $\gamma_D$ be the geodesic connecting $x_D$ and $y_D$.
From $x_D\in \gamma_D$ we obtain
\[ d(P_{\gamma_D} x, x )\leq d(x_D ,x )< \frac{r}{3}.\]
Likewise $d(P_{\gamma_D} y, y )< \frac{r}{3}$. With the triangle inequality we deduce
\[ r= d(x,y)\leq d(x,P_{\gamma_D} x) + d(P_{\gamma_D} x,P_{\gamma_D}y ) + d(P_{\gamma_D} y,y) < \frac{2}{3}r + d(P_{\gamma_D} x,P_{\gamma_D}y ). \]
This implies $0<\frac{1}{3}r < d(P_{\gamma_D} x,P_{\gamma_D}y )$ and we conclude that $\Phi_D$ is injective.\\
Since $K$ is $\tau_g$-compact and $\prod_{\gamma\in \Gamma_D} (\gamma,d)$ satisfies the Hausdorff property $\Phi_D$ restricts to a homeomorphism \[ \phi_D |_K \colon (K,\tau_{g})\rightarrow (\Phi_D(K) , \tau_{\textrm{K}}), \]
where $\tau_{\textrm{K}}$ denotes the subspace topology of $\Phi_D(K)$ in $\prod_{\gamma\in \Gamma_D} (\gamma,d)$. In particular, $\tau_{g}$-compactness of $K$ implies compactness of $(\Phi_D(K) , \tau_{\textrm{K}})$. Being a countable product of metric spaces, the space $\prod_{\gamma\in \Gamma_D} (\gamma,d)$ is metrizable. Therefore, clearly, the subspace $(\Phi_D(K) , \tau_{\textrm{K}})$ is also metrizable. Metrizability and compactness imply that $(\phi_D(K) , \tau_{\textrm{K}})$ is sequentially compact. Therefore, the subspace topology of $\tau_g$ on $K$ is metrizable and sequentially compact.\\
For the converse direction assume that $K$ is $\tau_g$-sequentially compact. Since $(H,d)$ is separable, so is $(K,d)$. Hence $(K,d)$ satisfies the Lindel\"off property. Since $\tau_g$ is coarser than the metric topology also $(K,\tau_g|_K)$ satisfies the Lindel\"off property. Sequential compactness implies that $(K,\tau_g|_K)$ is countably compact. Hence $(K,\tau_g|_K)$ is compact as it is Lindel\"off and countably compact.
\end{proof}
\section{Locally compact spaces}\label{sec:loc-cpct}
We show that $\tau_w$ agrees with the metric topology for locally compact CAT($0$) spaces. We start with the following geometric lemma.
\begin{lemma} \label{lem:cone}
Let $x\in H$, $\varepsilon>0$, $K$ be the closed ball of radius $\varepsilon$ around $x$, $P_K$ the projection to it, $y\in K$, and $m$ the midpoint between $x$ and $y$.
If $z\in H$ satisfies
\[ d(x,z)> \varepsilon \quad \text{and} \quad d(y,P_K z) \leq \frac{\varepsilon}{2} \quad \text{then}\quad P_{[x,m]}z=m.\]
\end{lemma}
\begin{proof}
Notice that $P_{K}z$ is the unique point on $[x,z]$ that satisfies $d(x,P_{K}z)= \varepsilon$. Hence our assumption can only be satisfied if $x\neq y$. Moreover, we have
\[ d(z, P_K z) = d(x,z)-\varepsilon .\]
Let $q\in [x,m)$. Then
\begin{align*}
d(y, z) & \leq d(y,P_K z) + d(P_K z , z) \\
& \leq d(x,z) - \frac{\varepsilon}{2}\\
& \leq d(x, q) + d(q ,z )- \frac{\varepsilon}{2} < d(q,z).
\end{align*}
Since the function $r \mapsto d(r,z)$ is convex on $[x,y]$ and $m\in (q,y)$ the distance $d(m,z)$ is smaller or equal to a convex combination of $d(q,z)$ and $d(y,z)$. Together with the last inequality we obtain $d(m,z)<d(q,z)$. Therefore, $P_{[x,m]}z=m$.
\end{proof}
We are now in the position to prove Theorem~\ref{thm:loc}.
\begin{proof}[Proof of Theorem~\ref{thm:loc}]
If $\tau_w$ and the metric topology coincide then closed balls are metrically compact by Lemma \ref{lem:w-compact}. In particular, $(H,d)$ is locally compact in this case. To prove the converse let $U\subseteq H$ be metrically open and $x\in U$. Then there exists some $\varepsilon>0$ such that the closed ball $K$ of radius $\varepsilon$ around $x$ is compact and contained in $U$. Let $y_1,\ldots,y_n\in K$ be a finite $\frac{\varepsilon}{2}$-net of $K$. With $m_i$ being the midpoint between $x$ and $y_i$ Lemma \ref{lem:cone} implies that
\begin{align}\label{eq:pr_loc}
U_x(m_1) \cap \ldots\cap U_x(m_n) \subseteq K\subseteq U,
\end{align}
since $z\in H\setminus K$ implies $z\notin U_x(m_i)$ for every
$i\in \{ 1,\ldots n\}$ that sasifies $d(y_i,P_K z)\leq \frac{\varepsilon}{2}$. The last statement of the theorem is due to Theorem~\ref{thm:proper}.
\end{proof}
\begin{remark}
Assuming that $(H,d) $ is locally compact Theorem \ref{thm:loc} yields $\tau_w=\tau_g=\tau_d$. Hence every closed ball in $(H,d)$ is metrically compact due to Lemma \ref{lem:w-compact}. This reproves the well known fact, that a Hadamard space $(H,d)$ is locally compact if and only if it is proper. (Here proper means that every closed ball in $(H,d)$ is compact).
\end{remark}
\subsection*{Acknowledgements} We would like to thank Russell Luke and Genaro L\'opez Acedo for valuable discussions on weak convergence in CAT($0$) spaces.
\bibliographystyle{amsplain}
|
\section{Derivations of ELBo}
\label{appd:elbo}
We follow conditional variational autoencoders \cite{NIPS2015_8d55a249} and assume that for given observation $\vec{h}$, $\vec{z}$ is drawn from the
prior distribution $p(\vec{z}\mid \vec{h})$, and the output $\vec{y}$ is generated from the distribution $p(\vec{y}\mid \vec{h}, \vec{z})$.
We learn the variational posterior by minimizing $\text{KL}(q_{\vphi}(\vec{z}\mid\vec{h}, \vec{y}) \Vert p(\vec{z}\mid \vec{h}, \vec{y}))$, which is equivalent to maximizing the evidence lower bound of the marginal log-likelihood (ELBo):
\begin{equation}
\small
\begin{split}
& \text{KL}(q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \Vert p(\vec{z}\mid \vec{h}, \vec{y})) \\
&= \int q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \log \frac{q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})}{p(\vec{z}\mid \vec{h}, \vec{y})} d{\vphi} \\
&= \int q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \log \frac{q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})p(\vec{y}\mid \vec{h})p(\vec{h})}{p(\vec{z}, \vec{h}, \vec{y})} d{\vphi} \\
&= \log p(\vec{y}\mid \vec{h}) \\
&+ \int q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \log \frac{q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})p(\vec{h})}{p(\vec{y}\mid \vec{h}, \vec{z})p(\vec{z}\mid \vec{h})} d{\vphi}
\end{split}
\end{equation}
Since $\text{KL}(q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \Vert p(\vec{z}\mid \vec{h}, \vec{y})) \ge 0$,
we have:
\begin{equation}
\small
\begin{split}
& \log p(\vec{y}\mid \vec{h}) \\
&\ge -\int q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \log \frac{q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})p(\vec{h})}{p(\vec{y}\mid \vec{h}, \vec{z})p(\vec{z}\mid \vec{h})} d{\vphi} \\
&= \mathbb{E}_{q_{\vphi}}[\log p(\vec{y}\mid \vec{z}, \vec{h})+\log p(\vec{z}\mid \vec{x}) \\
&~~~~~~~~~~~~ - q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})]\\
&= \mathbb{E}_{q_{\vphi}}[\log p(\vec{y}\mid \vec{z}, \vec{h})] - \mathbb{E}_{q_{\vphi}}[\frac{q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})}{\log p(\vec{z}\mid \vec{h})}] \\
&= \mathbb{E}_{q_{\vphi}}[\log p(\vec{y}\mid \vec{z}, \vec{h})] \\
&~~~~~~~~~~~~ - \text{KL}[q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y}) \Vert p(\vec{z}\mid \vec{h})]
\end{split}
\end{equation}
where $\vphi$ are parameters for the variational inference networks.
\section{Experiment Setup Details}
\label{appd:exp-setup}
\paragraph{Model Configurations.}
For the implementation details of PG, it is an LSTM-based encoder-decoder model with copying mechanism.
The encoder is a 1-layer Bi-LSTM, and the decoder is a 1-layer uni-directional LSTM.
We set the word embedding size to 300, the hidden dimension for both encoder and decoder to 512.
We let the encoder and decoder shares the same vocabulary list and word embedding, and the vocabulary size is 20000.
For the configuration of the posterior networks, both the mean and covariance network are a single feed-forward neural network, and we set the dimension of the latent variable to 256.
For the implementation details of T5, we use the T5-base implementation from Huggingface \citep{wolf-etal-2020-transformers} \footnote{\url{https://huggingface.co/transformers/model_doc/t5.html}}, and use their default model configuration.
We load the pre-trained weights of T5-base, and fine-tune them on our target task datasets.
For the configuration of the posterior networks, both the mean and covariance network are a single feed-forward neural network, and we set the dimension of the latent variable to 512.
\paragraph{Training Configurations.}
For the training details of PG and T5, we does not apply KL annealing and the coefficient of the KL divergence is always 1.
We use Adam optimizer \cite{duchi2011adaptive} with learning rate of 0.0001, and adopt early stopping if the validation loss does not decrease after 10 epochs.
For the hyper-parameters $\{v, r\}$ of the kernel function in \autoref{eq:gp-ours}, we try a range of values where $v \in [0.01, 100]$ and $r \in [0.0001, 10]$, and do grid search cross validation on the validation set to select the best model.
All experiments are independently conducted on a GPU server (RTX 2090 Ti) with 40cores CPU and 256GB Memory.
\begin{comment}
\section{Dialogue Generation Experiments}
\label{sec:exp-dialog}
\input{tab/dia-sents}
\input{tab/dialog}
We evaluate our model's capability of generating diverse dialogue responses using the multi-turn DailyDialog dataset \cite{li2017dailydialog}.
\paragraph{Dataset.}
The DailyDialog dataset \cite{li2017dailydialog} contains 13,118 multi-turn dialogues, and the average speaker turn per dialogue is 7.9.
We follow the standard data splition, and obtain 76,052 training pairs, 7,069 validation pairs and 6740 testing pairs.
\paragraph{Evaluation Metrics.}
For quality evaluation, we choose to use automatic metrics which are commonly applied in open-domain dialogue generation systems: BLEU with up to bi-grams \cite{papineni2002bleu} and METEOR \cite{banerjee2005meteor}.
For diversity evaluation, we follow Li et al. \shortcite{li2016diversity} to use distinct uni-grams (Dist-1) and bi-grams (Dist-2), which measures the ratio of distinct uni-grams and bi-grams in the generated outputs.
This is a good indicator to show if the dialogue system prefers to generate generic response like "I don't know" or more diverse responses that may be relevant to the contexts, and higher Dist-1 and Dist-2 indicate the model learns to generate more diverse responses.
\paragraph{Competitive Models.}
We choose the 1-layer LSTM Seq2Seq model \citep{cho-etal-2014-learning} with attention mechanism \citep{bahdanau2014neural} as our base Seq2Seq model.
We compare our method with \citep{serban2016building}, where they propose a Hierarchical Recurrent Encoder-Decoder (HRED) that extends the attention-based Seq2Seq model with a higher-level context RNN in order to keep track of past dialogue history.
Another work \citep{serban2017hierarchical} further augments the HRED with a latent variable at the decoder, which introduces random variations on context representations at the utterance-level. We refer this model as Latent Variable Hierarchical Recurrent Encoder-Decoder (VHRED).
\paragraph{Experiment Setups.}
Unlike the twitter dataset and the GYAFC dataset, the source and target texts in the DailyDialog dataset has few word overlaps.
Therefore, we switch from the pointer-generator network to the attention-based Seq2Seq model \cite{bahdanau2014neural} as our base Seq2Seq model, which is a commonly used baseline model in open-domain dialogue generation.
We set the maximum context text length to 150 tokens, maximum response length to 25 tokens, and the vocabulary size to 50K.
The encoder is a 2-layer Bi-GRU, and the decoder is a 2-layer uni-directional GRU.
The word embedding size is 256.
The encoder and decoder hidden states are both 512 dimension.
The dimension of the latent variable is 256.
We set the coefficient of the KL divergence to always be 1, and use Adam optimizer with learning rate of 0.0001.
We set the training epoch to 100, and reload the model with lowest validation loss for evaluation.
For the hyper-parameters $\{v, r\}$ of the kernel function in \autoref{eq:gp-ours}, we try a range of values where $v \in [0.01, 100]$ and $r \in [0.0001, 10]$, and do grid search cross validation on the validation set in order to select the best model.
All experiments are independently conducted on a GPU server (RTX 2090 Ti) with 40cores CPU and 256GB Memory.
\paragraph{Result Analysis.}
In dialogue generation, our method helps the decoder generate diverse responses with high quality.
As shown in \autoref{tab:dialog-acc}, the attention-based Seq2Seq model \citep[Attn-Seq2Seq]{cho-etal-2014-learning} is a strong baseline which outperforms HRED \citep{bahdanau2014neural} and VHRED \citep{serban2016building}.
But Attn-Seq2Seq with the GP prior can achieve comparable performance with the baseline.
\autoref{tab:z-daily} presents an example of the generated outputs from VHRED and Attn-Seq2Seq with the GP prior.
We further generate multiple outputs from by sampling multiple latent representations from $p(\vec{z}\mid \vec{h})$, and select the top-1 output which has the most uni-grams overlap with the reference response.
This oracle ranking method shows the upper-bound of the quality of generated texts by comparing the best prediction with ground-truth reference.
In this case, Attn-Seq2Seq with the GP prior outperforms the baseline.
\input{ethical}
\end{comment}
\section{Conclusion}
\label{sec:con}
In this work, we investigate the problem of generating high quality texts for variational encoder-decoder models.
We propose a novel stochastic function to introduce context-aware variations into encoder hidden states, which provides the decoder with more diverse contextual representations.
To learn this stochastic function, we propose a GP prior to model the dependency between random context variables, and apply an efficient amortized variational inference method to approximate the GP posterior.
Experimental results demonstrate that our method can learn a better contextual representation that leads to higher generation quality compared with deterministic encoder-decoder models and conditional variational autoencoders.
\section{Ethical Considerations}
The goal of this work is to enrich the potential of a language model for generating diverse texts.
Key features of each dataset adopted in this paper are described in the experiment session.
To further guarantee the reproducibility of this work, we provide all the essential details in the main content and the appendices. Codes and models will also open to public after the review period.
Although this work itself does not directly involve any ethical issues, the proposed method faces the same ethical challenges as other natural language generation techniques do.
For example, in dialogue systems, diverse generation systems may produce some offensive or toxic responses. Advanced text generation models could be used in scams.
Additional efforts are needed, both in technology and in society, to alleviate these issues.
Filtering toxic words and deploying spam-detection systems are two of the many examples.
\section{Experiments}
\label{sec:exp}
We evaluate our method on two text generation tasks that require rich contextual representations: paraphrase generation (\autoref{sec:exp-paraphrase}) and text style transfer (\autoref{sec:exp-style}).
We provide some data examples for the two tasks in \autoref{tab:dataset}.
We compared our method with previous works on diverse text generation in terms of quality and diversity.
Empirical experiment results show that our method is able to:
(1) adapt to different encoder-decoder architectures, such as the pointer-generator network \citep[PG]{see-etal-2017-get} and the text-to-text transfer Transformer \citep[T5]{DBLP:journals/corr/abs-1910-10683};
(2) generate higher quality texts compared with deterministic encoder-decoder models \citep{see-etal-2017-get,DBLP:journals/corr/abs-1910-10683} while also enabling diverse generation by conditioning on random context variables.
\subsection{Evaluation Methods}
\paragraph{Text Quality.}
For quality evaluation, we use two commonly used automatic metrics in text generation: \textbf{METEOR} \cite{banerjee2005meteor} and \textbf{BLEU} with up to bi-grams \cite{papineni2002bleu}, which tell us how well the generated outputs match the reference sentences.
\paragraph{Text Diversity.}
For diversity evaluation, we aim at examining how well different latent context variables $\vec{z}_{1:N}$ from $q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})$ can make the decoder generate diverse outputs.
We use self-BLEU with up to bi-grams \cite[\textbf{self-BLEU}]{zhu2018texygen} to measure the mutual bi-gram overlap between the set of outputs per source sentence, lower self-BLEU indicates less bi-gram overlap between generated outputs.
In addition, we use diverse 4-gram \cite[\textbf{Div-4}]{deshpande2019fast} to measure the ratio of distinct 4-grams in the set of outputs per source sentence, higher diverse 4-gram shows more unique 4-grams between generated outputs.
Finally, we use uniqueness \cite[\textbf{Uni.}]{deshpande2019fast} to measure the ratio of unique generated sentences in the set of outputs per source sentence, higher uniqueness suggests different context variables $\vec{z}_{1:N}$ lead to very different output sentences.
\subsection{Competitive Baselines}
We compare our method with competitive deterministic encoder-decoder models and variational encoder-decoder models as follows.
\newparagraph{PG}: \citet{see-etal-2017-get} proposes the pointer-generator network, which is a strong deterministic LSTM-based encoder-decoder baseline. We refer their model as {PG}.
\newparagraph{T5}: \citet{DBLP:journals/corr/abs-1910-10683} propose the text-to-text transfer Transformer, which is a strong deterministic Transformer-based encoder-decoder baseline. We refer their model as {T5}.
\newparagraph{Variation Attention}: \citet{deng2018latent} model the deterministic attention vectors as latent alignment variables to promote diverse text generation. We refer their model as {Variation Attention}.
\newparagraph{Multi-Selectors}: \citet{cho-etal-2019-mixture} use a mixture of experts to sample different binary masks on the source texts for diverse content generation. We refer their model as {Multi-Selectors}.
\newparagraph{T-CVAE}: \citet{ijcai2019-727} model deterministic encoder hidden states as latent context variables with a Transformer-based conditional variational autoencoder. We refer their model as {T-CVAE}.
\newparagraph{PG/T5 + Normal prior}: PG or T5 with a {normal prior} $p(\vec{z})=\mathcal{N}(\vec{0}, \mat{I})$, which follows the conventional variational autoencoders \cite{kingma2013auto,bowman-etal-2016-generating}.
\newparagraph{PG/T5 + GP prior}: PG or T5 with our GP prior defined in \autoref{eq:gp-prior}.
\input{tab/pg}
\subsection{Generation Setups}
For the decoding strategy, we use beam search with beam size of 10.
Note that our method is orthogonal to all diversity-promoting decoding strategies, such as top-k sampling \citep{fan-etal-2018-hierarchical} and nucleus sampling \citep{holtzman2019curious}.
We choose beam search in order to make fair comparison with other works which promotes diversity at the encoder side.
For quality generation, we directly take the mean of $q_{\vphi}(\vec{z}| \vec{h}, \vec{y})$, and generate one $\vec{y}_{1:T}$ based on the sampled context variables $\vec{z}_{1:N}$, since we want to examine how well the posterior network can encode contextual information and make the decoder generate high-quality texts.
For diverse generation, we sample different $\vec{z}_{1:N}$ (instead of directly taking the mean) from $q_{\vphi}(\vec{z}| \vec{h}, \vec{y})$, and generate different $\vec{y}_{1:T}$ based on the sampled context variables $\vec{z}_{1:N}$, since we want to examine how well different latent context variables $\vec{z}_{1:N}$ from $q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})$ can make the decoder generate diverse outputs.
For experiment setups, we sample 10 different $\vec{z}_{1:N}$ and generate 10 different $\vec{y}_{1:T}$ correspondingly.
We compute the diversity scores following the prior work \citep{deshpande2019fast}.
To compute the self-BLEU and Div-4, we randomly sample 5 different $\vec{y}_{1:T}$ out of the 10 generated $\vec{y}_{1:T}$.
To compute the Uni., we compute the unique number of sentences among the 10 generated $\vec{y}_{1:T}$.
In preliminary experiments, we found that sampling from the original variational distribution tends to make the decoder generate same sentences.
We hypothesize that $q_{\vphi}(\vec{z}| \vec{h}, \vec{y})$ is a high-dimensional multivariate Gaussian and sampling from a high-dimensional distribution is a fundamental challenging problem.
Therefore, we applied a simple heuristics to alleviate the sampling issue, where we scale up the covariance matrix of the variational distribution by a numeric scalar.
We find this simple heuristics can help the decoder generate more diverse sentences.
For PG + Normal prior and PG + GP prior, we set the numeric scalar for paraphrase generation task to 25, for style transfer task to 10.
For T5 + Normal prior and T5 + GP prior, we set the numeric scalar for paraphrase generation task to 7, for style transfer task to 4.
\subsection{Paraphrase Generation}
\label{sec:exp-paraphrase}
We first evaluate the model's capability of generating paraphrases using the Twitter URL paraphrasing dataset \cite{lan2017continuously}.
In this task, we aim at comparing the quality of generated texts between our method and other competitive baselines.
We include the experimental setups in \autoref{appd:exp-setup}.
\paragraph{Dataset.}
The Twitter URL paraphrasing dataset \cite{lan2017continuously} contains both positive and negative examples of paraphrases.
We filter out all negative examples from the 1-year 2,869,657 candidate pairs, and divided the remaining paraphrase pairs into 110K training pairs, 3K testing pairs and 1K validation pairs.
\input{tab/pg-diverse}
\paragraph{Result Analysis.}
As shown in \autoref{tab:pg-acc}, for the quality of generated texts, our method is able to well preserve the semantic information from source texts.
For LSTM-based models, PG + GP prior generates better quality texts compared with both its deterministic baseline PG and other variational baselines, e.g. Multi-Selectors and Variation Attention.
Note that PG + Normal prior experiences the posterior collapse problem \citep{bowman-etal-2016-generating,pmlr-v80-kim18e,pmlr-v89-dieng19a} during training, which causes the context variables preserving little semantic information in the source text and the model generating random tokens during inference.
For Transformer-based models, T-CVAE generates the better quality texts than T5, T5 + Normal prior and T5 + GP prior.
But T-CVAE lowercases all input and output tokens while the other models keep both lowercase and capital tokens, this text preprocessing step may bring an unfairly better performance of T-CVAE in quality scores.
Note that the posterior collapse problem does not happen in T5 + Normal prior, and T5 + GP prior still outperforms T5 + Normal prior, which shows the advantage of GP priors in introducing context-aware variations.
\subsection{Text Style Transfer}
\label{sec:exp-style}
We evaluate our model's capability of generating stylistic texts using the Grammarly's Yahoo Answers Formality Corpus (GYAFC) \cite{rao-tetreault-2018-dear}.
In this task, we first compare the quality of generated texts between our method and other competitive baselines, then we test the diversity of generated texts between our GP prior and conditional variational autoencoders.
We include the experimental setups in \autoref{appd:exp-setup}.
\paragraph{Dataset.}
The GYAFC dataset covers two sub-domains: Entertainment \& Music (E\&M), which has 52,593 training pairs, 2,877 validation pairs, 1,416 testing pairs; and Family \& Relationships (F\&R), which has 51,967 training pairs, 2,788 validation pairs, 1,332 testing pairs.
\input{tab/pg-sents}
\paragraph{Result Analysis.}
For the quality of generated texts, GP prior makes the model more robust to generate accurate texts.
As shown in \autoref{tab:pg-acc},
for LSTM-based models, PG + GP prior generates the most accurate texts compared with PG, Multi-Selectors and Variation Attention.
Note that PG + Normal prior also experiences the posterior collapse problem in GYAFC datasets, resulting in very low quality scores on the test set.
For Transformer-based models, T5 + GP prior achieves the best performance than T5, T5 + Normal and T-CVAE, which shows the superiority of GP priors in encoding contextual information.
For the diversity of generated texts, imposing context-aware variations into encoder hidden states is beneficial for generating diverse outputs.
As demonstrated in \autoref{tab:pg-diversity},
for Transformer-based models, T5 + GP prior gives the best diversity performance in both token-level and sentence-level compared with T5 + Normal prior and T-CVAE.
The model performance on the style transfer task verifies the capability of our GP prior in promoting generation diversity.
\autoref{tab:z-em} shows some diverse generation outputs of Transformer-based variational encoder-decoder models.
However, we also notice that increasing diversity will inevitably cause degradation in quality, because $\vec{z}_{1:N}$ are i.i.d. sampled from a high-dimensional multivariate Gaussian $q_\phi(\vec{z}\mid \vec{h}, \vec{y})$.
As discussed in previous work \citep{vono2022high}, multivariate sampling in high-dimensional settings can become computationally demanding.
\paragraph{Computation Complexity Analysis.}
Our GP priors require more computation during training, where the major computation comes from calculating the full co-variance matrix of context variables of the GP prior.
However, during inference, we approximate the GP posterior with a variational posterior $q_\phi(\vec{z}\mid \vec{h}, \vec{y})$ and conducts i.i.d. sampling, which saves the time for multivariate sampling and has the same computation complexity with other conditional variational autoencoder baselines at testing time.
\section{Gaussian Processes as Function Priors}
\label{app:gaussian}
\begin{figure*}[ht]
\subfigure[Sample $g$ from Gaussian process prior.]{
\includegraphics[width=0.45\linewidth]{img/gp_prior.png}
\label{fig:gp-prior}
}
\subfigure[Sample $g$ from Gaussian process posterior.]{
\includegraphics[width=0.5\linewidth]{img/gp_post.png}
\label{fig:gp-post}
}
\caption{Samples of potential functions $g$ from (a) a Gaussian process prior $\mathcal{GP}(\vec{0}, \mat{K})$ which uses a squared exponential kernel $k(h, h')=\exp(-\frac{(h-h')^2}{2})$, (b) a Gaussian process posterior $p(g\mid \mathcal{D})$ conditioning on the training data $\mathcal{D}$.}
\end{figure*}
In this work, we are interested in learning a non-linear mapping function from encoder hidden states to latent context variables.
Gaussian process has the nice property which can represent complex non-linear functions and also allow uncertainty to account for noisy data observations.
Considering a set of observed training data $\mathcal{D} = \{(\vec{h}_i, \vec{z}_i)\}_{i=1}^N$, a Gaussian process defines a probability distribution over possible functions $p(g)$.
Given a Gaussian process prior $\mathcal{GP}(\vec{0}, \mat{K})$ on the function $g(\vec{h})$, we have:
\begin{eqnarray}
\vec{z}_i &=& g(\vec{h}_i) + \epsilon_i \\
g(\vec{h}_i) &\sim& \mathcal{GP}(\vec{0}, \mat{K}) \\
\epsilon_i &\sim& \mathcal{N}(\vec{0}, \sigma^2\mat{I})
\end{eqnarray}
Note that $\epsilon_i$ is the noise of the observed data point $\vec{z}_i$, which is assumed to be an independent identically distributed Gaussian with variance ${\sigma}^2$.
$\mat{K}$ is the covariance matrix which is constructed using a squared exponential covariance function $k(\vec{h}, \vec{h}')=\exp(-\frac{\left\Vert \vec{h}-\vec{h}' \right\Vert^2}{2})$.
Now, we can sample different mapping functions $g(\vec{h})$ from this Gaussian process prior $\mathcal{GP}(\vec{0}, \mat{K})$.
Figure \autoref{fig:gp-prior} illustrates some possible mapping functions $g_1$, $g_2$ and $g_3$.
In Gaussian process, each training and testing data point is treated as random variable which follows Gaussian distribution.
Therefore, we can apply the Bayesian inference to predict a testing data point $\vec{z}_{*}$ conditioning on observed training data points $\mathcal{D}$.
To make concise notation, we let $\vec{h}_{1:N}=\{\vec{h}_i\}_{i=1}^N$, $\vec{z}_{1:N}=\{\vec{z}_i\}_{i=1}^N$, and $\mat{K}^{-1}=[k(\vec{h}_{1:N}, \vec{h}_{1:N})+{\sigma}^2\mat{I}]^{-1}$.
The probability distribution of the testing data point $\vec{z}_{*}$ can be computed by:
\begin{equation}
p(\vec{z}_{*} \mid \vec{h}_*, \mathcal{D}) = \int p(\vec{z}_{*} \mid \vec{h}_*, \vec{g}, \mathcal{D})p(\vec{g} \mid \mathcal{D}) dg
\end{equation}
where
\begin{equation}
\begin{split}
\label{eq:exact-post}
& p(\vec{z}_{*} \mid \vec{h}_*, \mathcal{D}) \sim \mathcal{N}(\vec{\mu}_*, \mat{K}_*) \\
& \vec{\mu}_* = k(\vec{h}_{*}, \vec{h}_{1:N})\mat{K}^{-1}\vec{z}\\
& \mat{K}_* = k(\vec{h}_{*}, \vec{h}_{*}) - k(\vec{h}_{*}, \vec{h}_{1:N})\mat{K}^{-1} k(\vec{h}_{1:N}, \vec{h}_{*})
\end{split}
\end{equation}
Intuitively, training data points $\mathcal{D}$ constrain the set of functions $g$ to pass through them since the covariance becomes smaller when we have training data, as shown in Figure \autoref{fig:gp-post}.
Under our variational encoder-decoder framework, $\vec{h}_{1:N}$ are encoder hidden states and $\vec{z}_{1:N}$ are latent context variables.
Since the Gaussian process induces a distribution over the mapping function $g(\vec{h})$, theoretically we could sample infinite number of mapping functions, where each function gives us a different set of latent context representations $\vec{z}_{1:N}$.
In this way, we managed to obtain diverse context representations in encoder-decoder models.
\begin{comment}
A Gaussian process prior is fully specified by a mean function $m(\cdot)$ and a covariance function $k(\cdot,\cdot)$:
\begin{equation}
\label{eq:gp-prior}
g(\vec{h}) \sim \mathcal{GP}(m(\vec{h}), k(\vec{h}, \vec{h}')),
\end{equation}
where, as a prior, $m(\cdot)$ is usually a zero function and $k(\vec{h}, \vec{h}')=\mathbb{E}{[g(\vec{h})-m(\vec{h})(g(\vec{h}')-m(\vec{h}')]}$.
The choice of $k(\cdot, \cdot)$ in a Gaussian process encodes the correlation of different $\vec{z}$'s in \autoref{eq:z-i}.
Generally, given a sequence of input $\vec{x}_{1:N}$ and its corresponding hidden states $\vec{h}_{1:N}$ from the transition function, the subset of function values $\vec{z}_{1:N}$ computed via \autoref{eq:z-i} are jointly Gaussian distributed:
\begin{equation}
\begin{split}
\label{eq:joint-prior}
p(\vec{z}_{1:N}) = \prod_{i=1}^N \mathcal{N}(m(\vec{h}_i), k(\vec{h}_i,\vec{h}_i'))
\end{split}
\end{equation}
where $m(\vec{h}_i)$ is the value of the prior mean function with input $\vec{h}_i$ and $k(\vec{h}_i,\vec{h}_i')$ gives the covariance between encoder hidden states at time steps $i$ and other encoding time steps.
According to the property of Gaussian processes\cite{rasmussen2006gaussian},
the conditional distribution of a Gaussian process is still a Gaussian process.
If we observed $\vec{z}_{1:N}$, then we can have a good estimation of the random function $g(\cdot)$ conditioned on $\vec{z}_{1:N}$, which is still a Gaussian process
\begin{equation}
\label{eq:post}
g(\cdot)\mid \vec{z}_{1:N} \sim \mathcal{GP}(\vec{\eta}(\cdot), \vec{u}(\cdot,\cdot))
\end{equation}
with
\begin{equation}
\begin{split}
\label{eq:cov}
\vec{\eta}(\cdot) &= {m}(\cdot) + k(\cdot, \mat{H})\mat{K}^{-1}(\vec{z}-\vec{m})\\
\vec{u}(\cdot,\cdot) &= k(\cdot, \cdot) - k(\cdot, \mat{H})\mat{K}^{-1} k^\top(\cdot, \mat{H})
\end{split}
\end{equation}
Theoretically, \autoref{eq:post}
can be used on any given $\vec{h}$.
Therefore, it is also called the predictive distribution in the literature of Gaussian process \cite{rasmussen2006gaussian,deisenroth2011robust,frigola2014variational,eleftheriadis2017identification}.
Given $\vec{z}_i$ as latent state for the encoder hidden state $\vec{h}_i$, and the mapping function $\vec{z}_i=g(\vec{h}_i)+\epsilon_g$ with Gaussian random noise $\epsilon_g \sim \mathcal{N}(\vec{0}, \sigma_g^2 \mat{I})$.
\end{comment}
\section{Efficient Variational Inference}
\label{sec:infer}
With the observed deterministic hidden states $\vec{h}_{1:N}$,
we estimate the GP posterior to make the prediction of context variables $\vec{z}_{1:N}$ more accurate.
Although the posterior estimation of Gaussian processes can be written in a closed form theoretically, the challenge in this work comes from learning with other parts of the model, such as the deterministic encoder producing $\vec{h}_{1:N}$ and the decoder generating $\vec{y}_{1:T}$.
To simplify the inference procedure, we will focus on inferring the samples of the GP posterior regarding the hidden states \emph{only} as $p(g \mid \vec{h}_{1:N})$, which essentially is a Gaussian distribution with non-isotropic covariance.
In this work, we apply variational inference to approximate the GP posterior $p(g \mid \vec{h}_{1:N})$ and learn other model parameters \emph{jointly} with maximum likelihood estimation.
For notation simplicity, we let $\vec{h}=f_{enc}(\vec{x}_{1:N})$, $\vec{z}=\{\vec{z}_{i}\}_{i=1}^N$, $\vec{y}=\{\vec{y}_{i}\}_{i=1}^T$ in this section.
With a {sampled} random function $g(\vec{h})$ from the GP prior as described in line 9 of \autoref{alg:generative}, we will get the joint prior distribution $p(\vec{z}\mid \vec{h})$ according to \autoref{eq:p_z}.
Then we approximate the true posterior $p(\vec{z}\mid \vec{h}, \vec{y})$ with the variational posterior $q_{\vphi}(\vec{z}\mid \vec{h}, \vec{y})$ by maximizing the evidence lower bound of the marginal log-likelihood (ELBo):
\begin{equation}
\label{eq:elbo-final}
\begin{split}
\log p(\vec{y}\mid \vec{h}) \ge
& \mathbb{E}_{q_{\vphi}}[\log p(\vec{y}\mid \vec{z})]\\
& - \text{KL}[q_{\vphi}(\vec{z}\mid \vec{h},\vec{y})\| p(\vec{z}\mid\vec{h})]
\end{split}
\end{equation}
where $\vphi$ denotes the variational parameters.
The derivation of \autoref{eq:elbo-final} is presented in \autoref{appd:elbo}.
During generation, we propose a two-step approximation to simplify $q_{\vphi}(\vec{z}\mid \vec{h},\vec{y})$.
First, to maintain the generative property when using the variational distribution, we propose an approximation of the variational distribution $q_{\vphi}(\vec{z}\mid\vec{h},\vec{y})\approx q_{\vphi}(\vec{z}\mid\vec{h})$.
In this case, random context vector $\vec{z}$ will only depend on $\vec{h}$ during inference.
Second, we apply the mean-field amortized variational approximation \citep{kingma2013auto} to approximating
the parameters of $q_{\vphi}(\vec{z}\mid\vec{h})$:
\begin{equation}
\begin{split}
q_{\vphi}(\vec{z}\mid \vec{h}) &= \prod_{i=1}^N q_{\vphi}(\vec{z}_i\mid \vec{h}_i) \\
&= \prod_{i=1}^N \mathcal{N}(f_\mu(\vec{h}_i), f_{\sigma^2}(\vec{h}_i))
\end{split}
\end{equation}
where
$f_\mu(\cdot)$ and $f_{\sigma^2}(\cdot)$ are the mean and covariance in the amortized variational inference network.
In this work, we use two simple feed-forward neural networks $f_\mu$ and $f_{\sigma^2}$.
The implementation details are included in \autoref{appd:exp-setup}.
\begin{comment}
Furthermore, both of the variational posterior $q_{\vphi}(\vec{z}\mid\vec{h}, \vec{y})$ and the sampled prior $p_{\vth}(\vec{z}\mid\vec{h})$ are multivariate Gaussians.
Let
$q_{\vphi}(\vec{z}\mid\vec{h}, \vec{y})=\mathcal{N}(\vec{\mu}_q, \mat{K}_q)$, where $\vec{\mu}_q \in \mathbb{R}^{N\times 1}$, $\mat{K}_q \in \mathbb{R}^{N\times N}$ and
$p_{\vth}(\vec{z}\mid\vec{h})=\mathcal{N}(\vec{\mu}_p, \mat{K}_p)$, where $\vec{\mu}_p \in \mathbb{R}^{N\times 1}$, $\mat{K}_p \in \mathbb{R}^{N\times N}$,
the KL-divergence
$\text{KL}[q_{\vphi}(\vec{z} \mid \vec{h}, \vec{y})\Vert p_{\vth}(\vec{z}\mid\vec{h})]$
in \autoref{eq:elbo-final} can be computed in closed-form,
\begin{equation}
\begin{split}
& \text{KL}[q_{\vphi}(\vec{z} \mid \vec{h}, \vec{y})\Vert p_{\vth}(\vec{z}\mid\vec{h})] \\
& = \frac{1}{2}[\log \frac{|\mat{K}_{p}|}{|\mat{K}_{q}|}
- N + tr(\mat{K}_p^{-1}\mat{K}_q) \\
& ~~~~~ +(\vec{m}_p-\vec{m}_q)^\top \mat{K}_p^{-1}(\vec{m}_p-\vec{m}_q)]
\end{split}
\end{equation}
While the computation complexity of $\mat{K}_p^{-1}$ is $\mathcal{O}(N^3)$, $N$ is the length of the input sentence in our application, the computation of $\mat{K}_p^{-1}$ is still tractable.
\end{comment}
\section{Introduction}
\label{sec:intro}
Generating high quality texts with high diversity is an important requirement for many text generation applications, such as paraphrase generation \citep{prakash2016neural,li-etal-2018-paraphrase}, style transfer \citep{jhamtani2017shakespearizing,rao-tetreault-2018-dear}, dialog generation \citep{sordoni2015hierarchical,serban2016building}, etc.
The encoder-decoder framework \citep{cho-etal-2014-learning,sutskever2014sequence} is widely adopted \citep{bahdanau2014neural,luong2015effective,gu-etal-2016-incorporating, see-etal-2017-get,DBLP:journals/corr/abs-1910-10683} to generate high-quality texts, where an encoder is applied to learn contextual information from source texts and a decoder is used to generate texts for target tasks.
To improve the diversity of generated texts, prior works propose to introduce variations into either the encoder \citep{bahuleyan2018variational,deng2018latent,ijcai2019-727,cho-etal-2019-mixture,qian-cheung-2019-enhancing,wu-etal-2020-encoder,duan-etal-2020-pre,sun-etal-2021-generating} or the decoder \citep{vijayakumar2016diverse,holtzman2019curious,he2018sequence,shen2019mixture}.
However, it is difficult to incorporate meaningful variations into encoder-decoder models without hurting the quality of generated texts.
\begin{figure}[t]
\centering
\subfigure[With a normal Gaussian prior]{
\scalebox{0.55}{\input{img/fig-ind-prior}}
\label{fig:vae-normal}
}
\quad
\subfigure[With a Gaussian process prior]{
\scalebox{0.55}{\input{img/fig-gp-prior}}
\label{fig:vae-gp}
}
\caption{A simple illustration on a variational encoder-decoder model with (a) an normal Gaussian prior and (b) a Gaussian process prior.
With a normal Gaussian prior, each hidden state $\vec{h}_i$ will be mapped into a random vector $\vec{z}_i$ independently; while a Guassian process prior imposes dependency constraints among $\{\vec{z}_i\}$. Double circles on $\{\vec{h}_i\}$ indicate they are deterministic variables.}
\end{figure}
Promoting diversity at the encoder side mainly focuses on modelling the probabilistic distribution of contextual representations.
Some prior works \citep{deng2018latent,bahuleyan2018variational} propose to model attention alignments between encoder and decoder hidden states as latent variables, and generate diverse texts by sampling from the latent attention variables.
Other existing works \citep{ijcai2019-727,liu2019transformer,shinoda-etal-2021-improving,sun-etal-2021-generating} directly apply conditional variational autoencoders to model encoder hidden states as latent variables, and generate high-diversity texts by sampling from the latent context variables.
However, when modelling the latent variables, they treat each latent variable as independent to each other, which inevitably causes the loss of some contextual information during learning, as shown in Figure \autoref{fig:vae-normal}.
Other works turns towards designing diversity-promoting decoding strategies at the decoder side, such as diverse beam search \citep{vijayakumar2016diverse}, top-k sampling \citep{fan-etal-2018-hierarchical}, and nucleus sampling \citep{holtzman2019curious}.
But for those decoding strategies, there is often a trade-off between quality and diversity, and the generation models have to sacrifice quality for a higher diversity.
Another line of works suggest to learn a mixture of expert encoders \citep{cho-etal-2019-mixture} or decoders \citep{he2018sequence,shen2019mixture}, and generate diverse texts by sampling from different encoders or decoders.
While different expert encoders or decoders can introduce some diversity, the model capacities are limited within the pre-defined set of experts.
In this work, we propose a novel approach to introduce context-aware variations into the encoder in order to generate high-quality and high-diversity texts.
For an encoder-decoder model, we introduce a stochastic function to map deterministic encoder hidden states $\{\vec{h}_i\}$ into a set of {random context variables} $\{\vec{z}_i\}$.
The advantage of this stochastic function is that it explicitly models the dependency between each context variable, as shown in Figure \autoref{fig:vae-gp}, which can help preserve more semantic information from source texts.
During generation, the decoder generates diverse outputs conditioning on sampled different context variables.
In other words, by learning a stochastic function on top of \emph{one} deterministic encoder, the proposed approach offers \emph{many} versions of random context variables for a decoder to generate diverse texts.
To learn the stochastic function over hidden states, we propose a Gaussian process prior \citep[GP]{rasmussen2006gaussian} to model the joint distribution of all encoder hidden states.
The major differences between GP priors and other priors used in previous works \citep{bahuleyan2018variational,deng2018latent,ijcai2019-727,cho-etal-2019-mixture,wu-etal-2020-encoder,duan-etal-2020-pre,shinoda-etal-2021-improving,sun-etal-2021-generating} have two-folds:
(1) GP priors explicitly model the dependency between latent variables of varying sizes as illustrated in Figure \autoref{fig:vae-gp}, while previous works consider latent variables as independent with each other as shown in Figure \autoref{fig:vae-normal};
(2) GP priors provide \emph{infinite} number of joint Gaussian distributions of latent variables as shown in Figure \autoref{fig:seq2seq-gp}, while previous works have to pre-define a fixed set of Gaussian distributions (e.g. a standard normal distribution, or a mixture of Gaussian distributions) with the risk of experiencing the posterior collapse problem \citep{bowman-etal-2016-generating,pmlr-v80-kim18e,pmlr-v89-dieng19a}.
Besides, the proposed random function only introduces variations into the encoder, and is orthogonal to diversity-promoting decoding strategies at the decoder side.
Users can freely adopt different decoding strategies to further encourage diverse generation outputs.
The major contributions of this work are three-fold:
\begin{enumerate}
\item We propose a novel method to introduce context-aware variations into encoder-decoder models, which can help the model learn rich contextual representations and also promote diversity in generation.
\item We propose an efficient variational inference method to approximate the joint distribution of fully-connected random context variables.
\item We test our proposed method in both LSTM-based \citep{{see-etal-2017-get}} and Transformers-based \citep{DBLP:journals/corr/abs-1910-10683} encoder-decoder models on paraphrase generation and style transfer tasks. Empirical experimental results show that, on one hand, the proposed method can generate higher quality texts than deterministic encoder-decoder models and conditional variational auto-encoders; on the other hand, it also supports diverse generation by conditioning on different sets of sampled random context variables.
\end{enumerate}
\begin{figure*}[t]
\centering
\subfigure[Variational encoder-decoder model with a normal Gaussian prior. Note that we simplify the latent variable $\vec{z}_i$ from a vector to a scalar in order to plot out the Gaussian distribution for better illustration.]{
\includegraphics[width=0.9\linewidth]{img/vae.png}
\label{fig:seq2seq-normal}
}
\quad
\subfigure[Variational encoder-decoder model with a GP prior. Note that we simplify the latent variable $\vec{z}_i$ from a vector to a scalar in order to plot out the joint Gaussian distribution for better illustration.]{
\includegraphics[width=0.9\linewidth]{img/gpvae.png}
\label{fig:seq2seq-gp}
}
\caption{A simple illustration for comparison between our GP priors and the priors in conditional variational autoencoders \cite{NIPS2015_8d55a249} under the variational encoder-decoder framework. }
\end{figure*}
\section{Model Description}
This section discusses our novel latent structured variable model on learning rich context representations by transforming the hidden states from a deterministic encoder into random hidden states via stochastic functions.
\subsection{Encoding with Stochastic Functions}
\label{subsec:generative}
Let $\vec{x}_{1:N}=\{\vec{x}_i\}_{i=1}^N$ be the source sentence of length $N$, and $\vec{y}_{1:T}=\{\vec{y}_t\}_{t=1}^T$ be the target sentence of length $T$.
In encoder-decoder models, an encoder is used to obtain deterministic context representations of the source sentence, i.e. the encoder hidden states:
$\vec{h}_{1:N} = f_{enc}(\vec{x}_{1:N})$, where $f_{enc}(\cdot)$ is a nonlinear transition function implemented by LSTM \citep{sutskever2014sequence} or Transformer \citep{vaswani2017attention}.
To introduce context-aware variations
into the encoder, we propose to learn a stochastic function that maps the deterministic hidden states to variables.
Specifically, after computing the hidden states $\vec{h}_{1:N}$ from the transition function $f_{enc}(\cdot)$, the proposed method employs a stochastic mapping function $g(\cdot)$ to model the deterministic context representations as a series of random context variables:
\begin{equation}
\label{eq:p_z}
p(\vec{z}_{1:N} \mid \vec{h}_{1:N}) = g(\vec{h}_{1:N}) + \vec{\epsilon}
\end{equation}
where $\vec{\epsilon}\sim \mathcal{N}(\vec{0}, \sigma^2\mat{I})$ is a Gaussian noise.
Then, the decoder can generate diverse texts conditioning on different sets of context variables sampled from $p(\vec{z}_{1:N} \mid \vec{h}_{1:N})$, as shown in Figure \autoref{fig:seq2seq-gp}.
Considering that natural language texts are always context-dependent, we expect the random context variables $\vec{z}_{1:N}$ to encode the context dependency to some extent.
In other words, the distribution of $\vec{z}_i$ {will} not only depend on $\vec{h}_i$, but also depend on other $\{\vec{z}_{j}\}_{j\not= i}$, as shown in Figure \autoref{fig:vae-gp}.
Under this framework, variational encoder-decoder models \citep{bahuleyan2018variational,deng2018latent,ijcai2019-727} can be viewed as a special case, as illustrated in Figure \autoref{fig:seq2seq-normal}, where random context variables $\{\vec{z}_i\}_{i=1}^N$ are \emph{independent} from each other.
In this work, we consider this special case as generation with {normal priors}.
Empirical comparison between normal priors and GP priors will be given in \autoref{sec:exp}.
\subsection{Gaussian Process Priors for Stochastic Functions}
\label{subsec:prior}
The learning of stochastic function $g(\vec{h})$ is the key for the proposed method to be successful.
Intuitively, we design $g(\vec{h})$ to satisfy two constraints simultaneously: (1) it can introduce some variation to the deterministic encoder hidden states; (2) it should preserve the contextual information in the deterministic encoder hidden states to be a faithful representation.
In this work, we propose to learn $g(\vec{h})$ with a functional prior defined by Gaussian processes.
As shown in Figure \autoref{fig:gp-prior} in \autoref{app:gaussian}, we can sample very different functions $g(\vec{h})$ from the same GP prior, which ensures randomness when sampling $\vec{z}_{1:N}$.\footnote{Please refer to \autoref{app:gaussian} and \citep{rasmussen2006gaussian} for detailed introduction of Gaussian processes.}
We define the stochastic function $g(\vec{h})$ following a GP prior:
\begin{equation}
\label{eq:gp-prior}
g(\vec{h}) \sim \mathcal{GP}(m(\vec{h}), k(\vec{h}, \vec{h}'))
\end{equation}
with the mean function $m(\vec{h})$ and covariance function $k(\vec{h},\vec{h}')$ as
\begin{equation}
\begin{split}
\label{eq:gp-ours}
m(\vec{h}) &= \vec{h} \\
k(\vec{h}, \vec{h}') &=
v^2 \exp\{-\frac{\|\vec{h}-\vec{h}'\|_2^2}{2r^2}\}
\end{split}
\end{equation}
where $\vec{h}$ indicates the current observed encoder hidden state, and $\vec{h}'$ indicates the other contextual encoder hidden states;
$v$ controls the average distance between a sampled function $g(\vec{h})$ and the mean function $m(\vec{h})$, and $r$ controls the covariance between random variables, increasing $r$ will make $\vec{z}$ and $\vec{z}'$ become more correlated.
In this work, $v$ and $r$ are chosen based on the text generation performance on development sets.
By setting $m(\vec{h})=\vec{h}$, we actually define a semi-parametric GP prior \citep[Sec. 15.2.6]{murphy2012machine} instead of a fully non-parameteric prior, since $\vec{h}$ as a hidden state is computed from the deterministic encoder with learnable parameters.
The intuition behind this definition is that, although we want to introduce some variations, taking the expectation of the sampled random states $\vec{z}$ should still be $\vec{h}$.
The {main advantage}
of applying GP priors is that we can sample infinite number of random functions $g(\vec{h})$ thus obtaining infinite sets of random context variables $\vec{z}_{1:N}$, as illustrated in Figure \autoref{fig:seq2seq-gp}.
In contrast, standard variational encoder-decoder models can only learn a fixed set of $C$ joint distributions $p(\vec{z}_{1:N}|\vec{h}_{1:N})$, where $1\le C \ll \infty$.\footnote{When $C=1$, it represents a conventional variational autoencoder \citep{bowman-etal-2016-generating}; when $C=5$, it represents a variational autoencoder with a mixture of Gaussians prior (component number = 5); when $C\to \infty$, it represents a variational autoencoder with a GP prior.}
\subsection{Generation with Random Context Variables}
\label{subsec:generation}
In this section, we demonstrate how to incorporate $\vec{z}_{1:N}$ into two typical encoder-decoder models for text generation: a LSTM-based encoder-decoder model \cite{see-etal-2017-get} and a Transformer-based encoder-decoder model \citep{DBLP:journals/corr/abs-1910-10683}.
The performance of these two variational encoder-decoder models with will be evaluated in \autoref{sec:exp}.
Given the deterministic encoder hidden states $\vec{h}_{1:N}$, we first sample a function $g(\vec{h})$ from the GP prior in \autoref{eq:gp-prior}; then sample a set of random context variables $\vec{z}_{1:N}$ from $g(\vec{h})$; finally generate a output sentence $\vec{y}_{1:T}$ based on the sampled $\vec{z}_{1:N}$.
The generative story with random context variables $\vec{z}_{1:N}$ is detailed in \autoref{alg:generative}.
For a LSTM-based encoder-decoder model \cite{see-etal-2017-get}, we apply the attention mechanism \citep{bahdanau2014neural} over the random context variables $\{\vec{z}_i\}_{i=1}^N$ to construct $\vec{c}_{t}$ for the decoder.
At each decoding time step $t$, the decoder computes the attention vector $\vec{c}_t$ and decoder hidden state $\vec{s}_t$ as follows:
\begin{eqnarray}
\alpha_{ti} &=& \frac{\exp{(a(\vec{s}_{t-1}, \vec{z}_i))}}{\sum_{j=1}^N \exp{(a(\vec{s}_{t-1}, \vec{z}_j))}} \\
\vec{c}_t &=& \sum_{i=1}^N \alpha_{ti}\cdot \vec{z}_i \\
\vec{s}_t &=& f_{dec}(\vec{s}_{t-1}, \vec{y}_{t-1}, \vec{c}_t)
\label{eq:py-yz}
\end{eqnarray}
where $a(\vec{s}_{t-1}, \vec{z}_i) = \vec{v}_a^\top \text{tanh}(\mat{W}_a \vec{s}_{t-1} + \mat{U}_a \vec{z}_i)$, $\mat{W}_a$, $\mat{U}_a$ and $\vec{v}_a^\top$ are parameter matrices.
Finally, the decoder outputs a word distribution based on the context representations and previous decoded words at each decoding time step $t$:
\begin{equation}
p(\vec{y}_t\mid\vec{y}_{t-1}, \vec{z}_{1:N}) = \text{softmax}(\mat{W}_b \cdot \vec{s}_t)
\label{eq:p_vocab}
\end{equation}
where $\mat{W}_b$ is a parameter matrix.
\input{alg}
For a Transformer-based encoder-decoder model \citep{DBLP:journals/corr/abs-1910-10683}, we take the output of the last layer in the encoder as $\vec{h}_{1:N}$, and feed them into $g(\vec{h})$ to get random context variables $\vec{z}_{1:N}$.
For the decoder, the inputs $\mat{K}$ and $\mat{V}$ are the combination of $\vec{h}_{1:N}$ and $\vec{z}_{1:N}$:
\begin{eqnarray}
\mat{K}^{l} = \mat{V}^{l} &=& \mat{W}_z[\vec{z}_{1:N}; \vec{h}_{1:N}]\\
\mat{A} &=& \text{MultiHead}(\mat{S}^{l-1}, \mat{K}^{l}, \mat{V}^{l}) \\
\mat{B} &=& \text{LayerNorm}(\mat{A} + \mat{S}^{l-1}) \\
\mat{S}^{l} &=& \text{LayerNorm}(\text{FFN}(\mat{B})+\mat{B})
\end{eqnarray}
where $\mat{S}^{l}=\{\vec{s}_t\}_{t=1}^T$ is the last layer of decoder hidden states, and $\text{MultiHead}(\cdot)$, $\text{LayerNorm}(\cdot)$ and $\text{FFN}(\cdot)$ follow the standard implementation in \citep{vaswani2017attention}.
The word distribution $p(\vec{y}_t\mid\vec{y}_{t-1}, \vec{z}_{1:N})$ at each decoding time step $t$ is computed the same way as in \autoref{eq:p_vocab}.
\section{Learning with Gaussian Process Priors}
\label{sec:model}
Intuitively, the random function $g(\vec{h}_i)$ needs to satisfy two constraints simultaneously: (1) it can introduce some randomness that allows some variations; (2) it should maintain the majority of the information in $\vec{h}_i$ in order to be a faithful representation.
Therefore, the definition of the random function $g(\vec{h}_i)$ is the key for the proposed method to be successful.
In this section, we start with standard attention-based Seq2Seq models, then we introduce a random function which imposes randomness in the encoder hidden states.
Finally, we propose our Gaussian process prior to learn a better random function, which not only introduces randomness into context representations, but also well preserve the semantic information.
In this work, instead of giving an explicit form of $g(\vec{h}_i)$, we assign $g(\vec{h}_i)$ with a probability distribution using Gaussian process.
As shown in Figure\autoref{fig:gp-prior}, we can sample very different functions $g(\vec{h}_i)$ from the same Gaussian process prior, which ensures randomness when generating $\vec{z}_i$.
\begin{comment}
A Gaussian process prior is fully specified by a mean function $m(\cdot)$ and a covariance function $k(\cdot,\cdot)$:
\begin{equation}
\label{eq:gp-prior}
g(\vec{h}) \sim \mathcal{GP}(m(\vec{h}), k(\vec{h}, \vec{h}')),
\end{equation}
where, as a prior, $m(\cdot)$ is usually a zero function and $k(\vec{h}, \vec{h}')=\mathbb{E}{[g(\vec{h})-m(\vec{h})(g(\vec{h}')-m(\vec{h}')]}$.
The choice of $k(\cdot, \cdot)$ in a Gaussian process encodes the correlation of different $\vec{z}$'s in \autoref{eq:ssm-obs}.
Generally, given a sequence of input $\vec{x}_{1:N}$ and its corresponding hidden states $\vec{h}_{1:N}$ from the transition function, the subset of function values $\vec{z}_{1:N}$ computed via \autoref{eq:ssm-obs} are jointly Gaussian distributed:
\begin{equation}
\begin{split}
\label{eq:joint-prior}
p(\vec{z}_{1:N}) = \prod_{i=1}^N \mathcal{N}(m(\vec{h}_i), k(\vec{h}_i,\vec{h}_i'))
\end{split}
\end{equation}
where $m(\vec{h}_i)$ is the value of the prior mean function with input $\vec{h}_i$ and $k(\vec{h}_i,\vec{h}_i')$ gives the covariance between encoder hidden states at time steps $i$ and other encoding time steps.
According to the property of Gaussian processes\cite{rasmussen2006gaussian},
the conditional distribution of a Gaussian process is still a Gaussian process.
If we observed $\vec{z}_{1:N}$, then we can have a good estimation of the random function $g(\cdot)$ conditioned on $\vec{z}_{1:N}$, which is still a Gaussian process
\begin{equation}
\label{eq:post}
g(\cdot)\mid \vec{z}_{1:N} \sim \mathcal{GP}(\vec{\eta}(\cdot), \vec{u}(\cdot,\cdot))
\end{equation}
with
\begin{equation}
\begin{split}
\label{eq:cov}
\vec{\eta}(\cdot) &= {m}(\cdot) + k(\cdot, \mat{H})\mat{K}^{-1}(\vec{z}-\vec{m})\\
\vec{u}(\cdot,\cdot) &= k(\cdot, \cdot) - k(\cdot, \mat{H})\mat{K}^{-1} k^\top(\cdot, \mat{H})
\end{split}
\end{equation}
Theoretically, \autoref{eq:post}
can be used on any given $\vec{h}$.
Therefore, it is also called the predictive distribution in the literature of Gaussian process \cite{rasmussen2006gaussian,deisenroth2011robust,frigola2014variational,eleftheriadis2017identification}.
Given $\vec{z}_i$ as latent state for the encoder hidden state $\vec{h}_i$, and the mapping function $\vec{z}_i=g(\vec{h}_i)+\epsilon_g$ with Gaussian random noise $\epsilon_g \sim \mathcal{N}(\vec{0}, \sigma_g^2 \mat{I})$.
\end{comment}
Specifically, in this work, we use a Gaussian process defined with a
In order to preserve the majority of the information in $\vec{h}_{1:N}$, we introduce a specific Gaussian process prior for the mapping function $g(\vec{h}_i)$:
\begin{equation}
\begin{split}
\label{eq:gp-ours}
g(\vec{h}) &\sim \mathcal{GP}(m(\vec{h}), k(\vec{h}, \vec{h}'))\\
m(\vec{h}_i) &= f(\vec{h}_{i-1}, \vec{x}_i) \\
k(\vec{h}_i, \vec{h}_j) &=
v^2 \exp\{-\frac{1}{2}\sum_{d=1}^D(\frac{|\vec{h}^{(d)}_{i} - \vec{h}^{(d)}_j|}{r})^2\}
\end{split}
\end{equation}
We use a mean function $m(\vec{h}_i)=\vec{h}_i$ for our Gaussian process prior.
The interpretation for this mean function is that when we have not observed any training data $\mathcal{D}$, we assume all possible function values $g(\vec{h}_i)$ are centered around $\vec{h}_i$, which enables us to introduce random perturbation on the encoder hidden state $\vec{h}_i$.
The covariance function $k$ is a squared exponential function, where $\{v, r\}$ is a set of hyper-parameters used to control the signal variance and length scale of the covariance function, and $d$ is the dimension index of the encoder hidden state. The squared exponential function is widely used as a linear smoother to reconstruct the underlying signal $g(\vec{h}_i)$ by removing the noise $\epsilon_i$.
When predicting the output sentence $\vec{y}_{1:T}$, we first sample a function $g$ from our Gaussian process prior, then sample the latent context variables $\vec{z}_{1:N}$ conditioning on function values computed from that sampled function $g$, finally predict the output sentence $\vec{y}_{1:T}$ based on the sampled latent context variables $\vec{z}_{1:N}$.
The overall generative process for our variational Seq2Seq framework is demonstrated in \autoref{alg:generative}.
Note that the standard variational Seq2Seq model \cite{bowman2016generating} with a standard Gaussian prior $p(\vec{z})=\mathcal{N}(\vec{0}, \mat{I})$ is a special case under our variational Seq2Seq framework.
In this case, the mapping function $g$ always maps the encoder hidden state $\vec{h}_i$ into zero mean with variance equals to one.
\input{alg}
\section{Related Works}
\label{sec:related}
\paragraph{Diverse text generation.}
Related works on diverse text generation mainly focus on changing decoding strategies at the decoder side or introducing randomness at the encoder side.
At the decoder side, recent works apply various decoding algorithms to promote diversity, such as diverse beam search \cite{vijayakumar2016diverse
, top-k sampling \cite{fan-etal-2018-hierarchical}
and nucleus sampling \cite{holtzman2019curious}
Our model is orthogonal to these diverse decoding algorithms since we focus on the encoder side.
Another group of works \cite{he2018sequence,shen2019mixture} propose to use a mixture of decoders to generate multiple outputs, where the context encodings are shared across multiple decoders.
At the encoder side, \citet{cho-etal-2019-mixture} propose to leverage a mixture of selectors to identify key contents from the source text, where each selector samples a sequential binary latent variables as a hard attention mask on every source token.
\citet{xu2018d} train different pattern embeddings, and generate diverse paraphrases conditioning on different pattern embeddings.
\paragraph{Conditional variational autoencoders.}
Variational encoder-decoder models \cite{deng2018latent,bahuleyan2018variational,ijcai2019-727,sun-etal-2021-generating} are related to our method.
\citet{deng2018latent} formulate the attention vector as latent alignment variables, and use the latent variables as hard attention for the decoder to select which source words to focus on during generation.
\citet{ijcai2019-727} present a conditional variational autoencoder based on Transformer, and learn a latent variable for generating diverse texts for the story completion task.
\citet{sun-etal-2021-generating} propose a self-separated conditional variational autoencoder that introduces group information to regularize the latent variables, which alleviates the posterior collapse problem and improves the model performance in the dialogue generation task.
|
\section{Introducton}
\subsection{Background}
In this paper we study the homogenization of homogeneous incompressible Navier-Stokes equations in a perforated domain in $\mathbb{R}^{3}$ under Dirichlet boundary condition. Our goal is to describe the limit behavior of the (weak) solutions as the number of holes goes to infinity and the size of holes goes to zero simultaneously.
Let $\Omega \subset {\mathbb{R}^3}$ be a bounded domain of class $C^{2,\beta}~(0<\beta<1)$. The holes in $\Omega$ are denoted by $T_{\varepsilon,k}$ which are assumed to satisfy
\begin{equation}\label{1-hole}
B(\varepsilon {x_k},{\delta _0}{a_\varepsilon}) \subset T_{\varepsilon ,k} = \varepsilon {x_k} + a_\varepsilon T_{0} \subset \subset B(\varepsilon {x_k},{\delta _1}{a_\varepsilon}) \subset \subset B(\varepsilon {x_k},{\delta _2}{a_\varepsilon}) \subset B(\varepsilon {x_k}, \delta_3\varepsilon) \subset \varepsilon {Q_k},
\end{equation}
where the cube $Q_k: = (-\frac{1}{2},\frac{1}{2})^3+k$ and $x_k=x_0+k$ with $x_0\in T_{0}$, for each $k\in \mathbb{Z}^3$; $T_{0}$ is a model hole which is assumed to be a closed bounded and simply connected $C^{2,\beta}$ domain; $\delta _i, \ i = 0,1,2,3$ are fixed positive numbers. The perforation parameters $\varepsilon$ and $a_\varepsilon$ are used to measure the mutual distance of holes and the size of holes, and $\varepsilon x_k=\varepsilon x_0+\varepsilon k$ present the locations of holes. Without loss of generality, we assume that
$x_0=0$ and $0<a_\varepsilon \leqslant \varepsilon \leqslant 1$.
The perforated domain $\Omega_{\varepsilon}$ under consideration is described as follows:
\begin{equation}\label{1-domain}
\Omega _\varepsilon : = \Omega \backslash \bigcup\limits_{k \in K_\varepsilon} {T_{\varepsilon ,k}},~~~~{K_\varepsilon}:= \{ k \in \mathbb{Z}^3: \varepsilon \overline {Q_k} \subset \Omega \}.
\end{equation}
The study of homogenization problems in fluid mechanics have gained a lot interest. In particular, the homogenization of Stokes system in perforated domains has been systematically studied. In 1980s, Tartar \cite{ref7} considered the case where the size of holes is proportional to the mutual distance of holes and derived Darcy's law. In 1990s, Allaire \cite{ref3,ref4} considered general size of holes and obtained complete results in periodic setting. By introducing a local problem and employing an abstract framework the idea of which goes back to \cite{CM82}, Allaire found that the homogenized limit equations are determined by the ratio $\sigma_\varepsilon$ given as
\begin{equation}\label{1-sigma}
\sigma _\varepsilon: = \Big(\frac{\varepsilon^d}{a_\varepsilon^{d-2}}\Big)^{\frac{1}{2}}, \ d \geqslant 3;\quad{\sigma _\varepsilon}: = \varepsilon \left| \log \frac{{a_\varepsilon }}{\varepsilon} \right|^{\frac{1}{2}}, \ d = 2,
\end{equation}
where $d$ is the spatial dimension. More precisely, if ${\lim _{\varepsilon \to 0}}{\sigma _\varepsilon }= 0$ corresponding to the case of large holes, the homogenized system is the
Darcy's law; if ${\lim _{\varepsilon \to 0}}{\sigma _\varepsilon}= \infty$ corresponding to the case of small holes, the motion of the fluid does not change much in the homogenization process and in the limit there arise the same Stokes equations; if ${\lim _{\varepsilon \to 0}}{\sigma _\varepsilon }= \sigma _* \in(0,\infty) $ corresponding to the case of critical size of holes, the homogenized system is governed by the Brinkman's law—a combination of the Darcy's law and the original Stokes equations.
The homogenization study is extended to more complicated models decribing fluid flows: Mikeli\'c \cite{ref2} has studied the incompressible Navier-Stokes equations in a porous medium; Masmoudi \cite{ref12} studied the compressible Navier-Stokes equations; Feireisl, Novotn\'y and Takahashi \cite{ref13} studied the Navier-Stokes-Fourier equations. In all these previous studies, only the case where the size of holes is proportional to the mutual distance of holes is considered and the Darcy's law is recovered in the limit. Recently, Feireisl, Namlyeyeva and Ne\v{c}asov\'a \cite{ref14} studied the case with critical size of holes for the incompressible Navier-Stokes equations and they derived Brinkman's law; Feireisl et al. also considered the case of small holes for the compressible Navier-Stokes equations \cite{ref15,ref10,ref9}.
\medskip
In \cite{ref3,ref4}, Allaire also gave a rather complete description concerning the homogenization of stationary incompressible Navier-Stokes equations and the results coincide with the Stokes equations: for the case of small holes, the equations remain unchanged; for the case of large holes, Darcy's law is derived; for the case of critical size of holes, Brinkmann type equations are recovered. While, for the evolutionary incompressible Navier-Stokes equations, the study is not compete with respect to the size of holes, even in periodic setting: in \cite{ref2}, Mikeli\'c considered the case when the size of holes is proportional to the mutual distance of holes, and in \cite{ref14} the critical size of holes is considered.
In this paper we shall consider the homogenization of evolutionary incompressible Navier-Stokes equations in perforated domain $\Omega _\varepsilon$ with Dirichlet boundary condition and our goal is to give a complete description for the homogenization process related to small and large sizes of holes in periodic setting. Let $T>0$, the initial boundary problem in space-time cylinder $\Omega_{\varepsilon} \times (0,T)$ under consideration is the following:
\begin{equation}\label{1-equation}
\begin{cases}
\partial _t {\mathbf{u}}_\varepsilon + {\mathrm{div}\,} ({\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon)
-\mu \Delta {\mathbf{u}}_\varepsilon + \nabla {p_\varepsilon} = {\mathbf{f}}_{\varepsilon}, &\mbox{in}\; \Omega _\varepsilon \times (0,T),\\
{\mathrm{div}\,} {\mathbf{u}}_\varepsilon = 0, &\mbox{in}\;\Omega _\varepsilon \times (0,T),\\
{\mathbf{u}}_\varepsilon = 0, &\mbox{on}\;\partial\Omega _\varepsilon \times (0,T),\\
{\mathbf{u}}_\varepsilon |_{t = 0} = {\mathbf{u}}_{\varepsilon}^0 \in L^2(\Omega_{\varepsilon};\mathbb{R}^3).
\end{cases}
\end{equation}
Here ${\mathbf{u}}_{\varepsilon}$ is the fluid velocity field in ${\mathbb{R}}^{3}$ and $p_{\varepsilon}$ is the fluid pressure. The external force ${\mathbf{f}}_{\varepsilon}$ is assume to be in $L^{2}(\Omega_{\varepsilon}\times (0,T);{\mathbb{R}}^{3})$.
We recall some notations. Let $W^{1,q}_{0}(\Omega)$ be the collection of Sobolev functions in $W^{1,q}(\Omega)$ with zero trace, and let $W^{-1,q}$ be the dual space of $W^{1,q}_{0}(\Omega)$. We set $V^{1,q}(\Omega):=\{v\in W_0^{1,q}(\Omega),\ {\mathrm{div}\,} v =0\}$ with $W^{1,q}$ norm and ${V^{-1,q}}(\Omega)$ be the dual space of ${V^{1,q}}(\Omega)$. Let $L^{q}_{0}(\Omega)$ be the collection of $L^{q}(\Omega)$ integrable functions that are of zero average. We sometimes use $L^r L^s$ to denote the Bochner space $L^r(0,T;L^s(\Omega))$ or $L^r(0,T;L^s(\Omega);{\mathbb{R}}^{3})$ for short. For a function $g$ in $\Omega_{\varepsilon}$, we use the notation $\widetilde g$ to represent its zero extension in $\Omega$:
$$
\widetilde g = g \ \mbox{in $\Omega_{\varepsilon}$}, \quad \widetilde g = 0 \ \mbox{in $\Omega\setminus \Omega_{\varepsilon} = \bigcup\limits_{k \in {K_\varepsilon }} {T_{\varepsilon ,k}}$}.
$$
We now recall the definition of (finite energy) weak solutions:
\begin{defn}\label{def-weak}
We call ${\mathbf{u}}_\varepsilon$ a weak solution of \eqref{1-equation} in $\Omega_{\varepsilon}\times (0,T)$ provided:
\begin{itemize}
\item There holds:
\ba\label{def-weak-1}
& {\mathbf{u}}_\varepsilon \in L^2(0,T;V^{1,2}(\Omega _\varepsilon))\cap C_{\rm weak}([0,T], L^{2}(\Omega_{\varepsilon}))\cap C([0,T], L^{q}(\Omega_{\varepsilon})), \\
& \partial _t{\mathbf{u}}_\varepsilon \in L^\frac{4}{3}}(0,T;{V^{ - 1,2}}({\Omega _\varepsilon) ),
\ea
for any $1\leq q <2$.
\item For any $\varphi\in C^\infty_c(\Omega_{\varepsilon} \times [0,T) ;\mathbb{R}^3)$ with ${\mathrm{div}\,} \varphi =0$, there holds
\ba\label{5}
& \int_0^T \int_{\Omega _\varepsilon} - {\mathbf{u}}_\varepsilon \cdot \partial_t\varphi - {\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon :\nabla \varphi
+ \mu \nabla {\mathbf{u}}_\varepsilon:\nabla \varphi {\mathrm{d}} x{\mathrm{d}} t \\
& = \int_0^T \int_{\Omega_\varepsilon} {\mathbf{f}}_{\varepsilon} \cdot \varphi {\mathrm{d}} x{\mathrm{d}} t + \int_{\Omega _\varepsilon} {\mathbf{u}}_\varepsilon^0 \cdot \varphi (x,0) \,{\mathrm{d}} x{\mathrm{d}} t.
\ea
The pressure $p_{\varepsilon}$ is determined by
\ba\nn
\l \nabla p_{\varepsilon}, \psi \r = - \l \partial _t {\mathbf{u}}_\varepsilon + {\mathrm{div}\,} ({\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon)
-\mu \Delta {\mathbf{u}}_\varepsilon - {\mathbf{f}}_{\varepsilon}, \psi \r, \quad \forall \,\psi \in C_{c}^{\infty}(\Omega_{\varepsilon} \times (0,T);{\mathbb{R}}^{3}).
\ea
We further call ${\mathbf{u}}_{\varepsilon}$ a finite energy weak solution of \eqref{1-equation} provided there holds in addition the energy inequality: for a.a. $t\in (0,T),$
\begin{equation}
\frac{1}{2}\int_{\Omega _\varepsilon} \left| {\mathbf{u}}_\varepsilon (x,t) \right|^2 {\mathrm{d}} x
+\int_0^t \int_{\Omega _\varepsilon} \left| \nabla {\mathbf{u}}_\varepsilon(x,s) \right|^2 {\mathrm{d}} x{\mathrm{d}} s
\leqslant \int_0^t \int_{\Omega _\varepsilon} {\mathbf{f}}_{\varepsilon} \cdot {\mathbf{u}}_\varepsilon {\mathrm{d}} x{\mathrm{d}} s
+ \frac{1}{2}\int_{\Omega _\varepsilon} \left| {\mathbf{u}}^0_{\varepsilon}(x) \right|^2{\mathrm{d}} x.\label{6}
\end{equation}
\end{itemize}
\end{defn}
For each fixed $\varepsilon$, the existence of a global finite energy weak solution ${\mathbf{u}}_\varepsilon$ is known, see for example Leray's pioneer work \cite{ref18} or the classical books \cite{ref6,ref5}. We need to investigate the behavior of the solutions as $\varepsilon \to 0.$
\subsection{Main Results}
From \eqref{1-sigma}, we see that the size of holes $a_\varepsilon$ is typically chosen to be $\varepsilon^\alpha~ (\alpha\geqslant 1)$ in three or higher dimensional spaces, while $a_\varepsilon$ is typically chosen to be $e^{-\varepsilon^{-\alpha}}~(\alpha>0)$ in two dimensional case. Here for the study of three dimensional case, we shall take
\be\label{ae-se}
a_{\varepsilon} = \varepsilon^{\alpha} \ \mbox{with} \ \alpha\geqslant 1, \ \mbox{which implies} \ \sigma_{\varepsilon} = \varepsilon^{\frac{3-\alpha}{2}}.
\ee
Throughout the paper, we will assume the zero extension of the initial datum and the external force satisfy
\be\label{ini-force}
\widetilde {\mathbf{u}}_{\varepsilon}^{0} \to {\mathbf{u}}^{0} \ \mbox{strongly in} \ L^{2}(\Omega),\quad\widetilde {\mathbf{f}}_{\varepsilon} \to {\mathbf{f}} \ \mbox{strongly in} \ L^{2}(\Omega\times (0,T)).
\ee
Let ${\mathbf{u}}_\varepsilon$ be a finite energy weak solution of \eqref{1-equation}. We employ the idea of Mikeli\'c \cite{ref2} and Temam \cite{ref5} and introduce for any $t\in(0,T)$,
\begin{equation}\label{1-new-def}
{\mathbf{U}}_\varepsilon(\cdot, t): =\int_0^t{\mathbf{u}}_\varepsilon (\cdot, s) {\mathrm{d}} s,\quad \Psi_{\varepsilon} (\cdot, t) := \int_0^t {\mathbf{u}}_\varepsilon (\cdot, s) \otimes {\mathbf{u}}_\varepsilon (\cdot, s) {\mathrm{d}} s, {\mathbf{F}}_\varepsilon(\cdot, t) : =\int_0^t{\mathbf{f}}_\varepsilon (\cdot, s) {\mathrm{d}} s.
\end{equation}
Then $ {\mathbf{U}}_{\varepsilon} \in C([0, T];W^{1,2}_{0}(\Omega_\varepsilon), {\div {\mathbf{U}}}_\varepsilon=0, \ {\mathbf{F}}_\varepsilon\in C([0, T];L^2(\Omega_\varepsilon))$ and $\div \Psi _\varepsilon\in C([0,T]; L^{\frac{3}{2}}(\Omega_\varepsilon))$. As shown in the proof of our theorems, instead of showing the limit behavior of ${\mathbf{u}}_{\varepsilon}$, we turn to study the limit behavior of ${\mathbf{U}}_{\varepsilon}$ which has better regularity in time variable. Clearly
\be\label{Fe-F-st}
\widetilde {\mathbf{F}}_{\varepsilon} \to {\mathbf{F}} = \int_0^t{\mathbf{f}} (s) {\mathrm{d}} s \ \mbox{strongly in} \ L^{2}(\Omega \times (0,T)).
\ee
The classical theory on Stokes equations implies that there exists $P_\varepsilon\in C([0,T];L^2_{0}(\Omega_\varepsilon))$ (see Chapter 3 in \cite{ref5}), such that for any $t\in (0,T)$,
\begin{equation}\label{2-equa}
{\mathbf{u}}_\varepsilon(t) - {\mathbf{u}}^0_\varepsilon + {\mathrm{div}\,}\Psi _\varepsilon(t) - \mu \Delta {\mathbf{U}}_\varepsilon (t)+\nabla P_\varepsilon(t) = {\mathbf{F}}_\varepsilon(t),\quad\text{in}~W^{ - 1,2}(\Omega _\varepsilon).
\end{equation}
Now we state our results corresponding to different sizes of holes. The case of critical size of holes is considered by Feireisl-Namlyeyeva-Ne\v{c}asov\'a in \cite{ref14}, so we are focusing only on the case of small holes and large holes. Note that the limits are taken up to possible extractions of subsequences. The first result corresponds to the case of small holes:
\begin{theorem}\label{thm-1}
Let ${\mathbf{u}}_\varepsilon$ be a finite energy weak solution of the Navier-Stokes system \eqref{1-equation} in the sense of Definition \ref{def-weak} with initial datum and external force satisfying \eqref{ini-force}. Let $\widetilde p_\varepsilon$ be the extension of $p_\varepsilon$ defined by $\widetilde p_\varepsilon = \partial _t\widetilde P_\varepsilon$ where $\widetilde P_\varepsilon$ is the extension of $ P_{\varepsilon}$ defined in \eqref{def-F-P} and \eqref{small-F-P}. If $\alpha>3$, i.e. $\mathop {\lim }\limits_{\varepsilon \to\infty} \sigma _\varepsilon = \infty$, then
\ba\label{thm1-conv}
{\widetilde {{\mathbf{u}}}_\varepsilon } \to {{\mathbf{u}}} \ \mbox{weakly(*) in} \ L^{\infty}(0,T;L^{2}(\Omega)) \cap L^2(0,T;W_0^{1,2}(\Omega)),
\ea
and
\ba\label{thm1-conv-p}
\widetilde p_\varepsilon \to p\ \mbox{weakly\ in} \ W^{-1,2} (0,T;L_0^2(\Omega)).
\ea
Moreover, $({\mathbf{u}},p)$ is a weak solution of the Navier-Stokes equations in homogeneous domain $\Omega$:
\begin{equation}
\begin{cases} \label{equa-u-p-small}
\partial_t {\mathbf{u}} + {\mathrm{div}\,} ({\mathbf{u}} \otimes {\mathbf{u}}) - \mu \Delta {\mathbf{u}} + \nabla p = {\mathbf{f}},&\ \mbox{in} \;\Omega \times (0,T) ,\\
{\mathrm{div}\,} {\mathbf{u}} = 0, &\ \mbox{in} \;{\Omega} \times (0,T),\\
{\mathbf{u}}= 0, &\ \mbox{on} \; \partial\Omega \times (0,T),\\
{\mathbf{u}}{|_{t=0}} = {\mathbf{u}}^0.
\end{cases}
\end{equation}
\end{theorem}
\medskip
For the case of large holes, we consider the time-scaled Navier-Stokes system:
\begin{equation}\label{1-time-equa}
\begin{cases}
\sigma_\varepsilon^2 \partial _t {\mathbf{u}}_\varepsilon + {\mathrm{div}\,} ({\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon)
-\mu \Delta {\mathbf{u}}_\varepsilon + \nabla p_\varepsilon = {\mathbf{f}}, &\ \mbox {in} \;\Omega_\varepsilon \times (0,T),\\
{\mathrm{div}\,}{\mathbf{u}}_\varepsilon = 0, &\ \mbox{in} \; \Omega _\varepsilon \times (0,T),\\
{\mathbf{u}}_\varepsilon = 0, &\ \mbox{on} \; \partial\Omega _\varepsilon \times (0,T),\\
{\mathbf{u}}_\varepsilon |_{t = 0} = {\mathbf{u}}^0_{\varepsilon}\in L^2(\Omega_\varepsilon;\mathbb{R}^3).
\end{cases}
\end{equation}
For a solution ${\mathbf{u}}_{\varepsilon}$ to \eqref{1-time-equa}, we similarly introduce ${\mathbf{U}}_{\varepsilon}, \Psi_{\varepsilon}, {\mathbf{F}}_{\varepsilon}$ as in \eqref{1-new-def}. There exists $P_\varepsilon\in C([0,T];L^2_{0}(\Omega_\varepsilon))$, such that for any $t\in (0,T)$,
\begin{equation}\label{2-equa-large-holes}
\sigma_{\varepsilon}^{2} {\mathbf{u}}_\varepsilon(t) - \sigma_{\varepsilon}^{2} {\mathbf{u}}^0_\varepsilon+\Phi _\varepsilon(t) - \mu \Delta {\mathbf{U}}_\varepsilon (t)+\nabla P_\varepsilon(t) = {\mathbf{F}}_\varepsilon(t),\quad\text{in}~W^{ - 1,2}(\Omega _\varepsilon).
\end{equation}
We have the following theorem concerning the case of large holes:
\begin{theorem}\label{thm-3}
Let ${\mathbf{u}}_\varepsilon$ be a finite energy weak solution of the time-scaled Navier-Stokes system \eqref{1-time-equa} in the sense of Definition \ref{def-weak} with initial datum and external force satisfying \eqref{ini-force}. Let $\widetilde p_\varepsilon$ be the extension of $p_\varepsilon$ defined by $\widetilde p_\varepsilon = \partial _t\widetilde P_\varepsilon$ where $\widetilde P_\varepsilon$ is the extension of $ P_{\varepsilon}$ defined in \eqref{large-F-P-1} and \eqref{large-F-P-2}. If $1<\alpha<3$, i.e. $\lim _{\varepsilon \to 0}{\sigma _\varepsilon}=0$, then
\ba\label{thm2-u}
\sigma_{\varepsilon}^{-2} \widetilde \uu_{\e} \to {{\mathbf{u}}} \ \mbox{weakly in} \ {L^2}(0,T;L^{2}(\Omega ;{\mathbb{R}^3})) ,
\ea
and
\ba\label{thm2-p-1}
\widetilde p_\varepsilon=\widetilde p_\varepsilon^{(1)}+\sigma_\varepsilon^{\frac12}\widetilde p_\varepsilon^{(2)}+\sigma_\varepsilon^2\widetilde p_\varepsilon^{(3)},
\ea
with
\ba\label{thm2-p-2}
&\widetilde p_\varepsilon^{(1)} \to p \ \mbox{weakly in} \ L^2 (0,T;W^{1,2}(\Omega)),\\
&\widetilde p_\varepsilon^{(2)} \ \mbox{bounded in} \ L^\frac{4}{3}(0,T;L_0^2(\Omega)),\\
&\widetilde p_\varepsilon^{(3)} \ \mbox{bounded in} \ W^{-1,2} (0,T;L_0^2(\Omega)).
\ea
Moreover $({{\mathbf{u}}},p)$ is a weak solution of the Darcy's law:
\begin{equation}\label{Darcy}
\begin{cases}
\mu {\mathbf{u}} = A({{\mathbf{f}}} - \nabla p),&\ \mbox{in} \;\Omega \times (0,T),\\
{\mathrm{div}\,}{\mathbf{u}} = 0, &\ \mbox{in} \;\Omega \times (0,T) ,\\
{\mathbf{u}}\cdot \mathbf{n} = 0, &\ \mbox{on} \;\partial\Omega \times (0,T),
\end{cases}
\end{equation}
where $\mathbf{n}$ is the unit normal vector on the boundary of $\Omega$.
\end{theorem}
Here in \eqref{Darcy}, the permeability tensor $A$ is a constant positive definite matrix (see \cite{Allaire91}) given by:
\[A_{i,j}: = \mathop {\lim}\limits_{\eta \to 0} c_\eta ^{- 2}\int_{Q_\eta} \nabla w_\eta ^i:\nabla w_\eta ^j{\mathrm{d}} x
= \mathop {\lim }\limits_{\eta \to 0} \int_{Q_\eta} {(w_\eta ^j)}_i {\mathrm{d}} x: = ({\overline w}^j)_i,\]
where $w_\eta ^i$ satisfies the generalized cell problem introduced in Section \ref{sec:cell}.
\begin{remark}
The case $\alpha=1$ is considered by Mikeli\'c \cite{ref2}. In this case the permeability tensor $A_{i,j}$ is defined by the classical cell problem where $\eta = 1$ (see also \cite{ref7}) which is slightly different from the definition above.
\end{remark}
Sections \ref{sec-thm1} and \ref{sec-thm2} are devoted to the proofs of Theorem \ref{thm-1} and Theorem \ref{thm-3} respectively. In the sequel, $C$ denotes a constant independent of $\varepsilon$, while its value may differ from line to line.
\subsection{Generalized cell problem}\label{sec:cell}
We introduce the idea of \textit{generalized cell problem} \cite{ref19, ref1,ref7} which is used to study the homogenization process. Near each single hole, after a scaling of size $\varepsilon^{-1}$ such that the controlling cube becomes the $O(1)$ size, one can consider the following \textit{modified cell problem}:
\[\begin{cases}
-\Delta w_\eta ^i + \nabla q_\eta ^i = c_\eta ^2{e^i},&\text{in}\;Q_\eta : = {Q_0}\backslash (\eta T_{0}),\\
{\mathrm{div}\,} w_\eta ^i = 0, & \mbox{in}\;Q_\eta,\\
w_\eta ^i = 0, & \mbox{on}\;\eta T_{0},\\
(w_\eta ^i,q_\eta ^i) \ \mbox{is $Q_0$-periodic}.
\end{cases}\]
Here $Q_{0} = \big(-\frac 12, \frac 12\big)^{d}, \ \eta : = \frac{a_\varepsilon}{\varepsilon},\ c_\eta := \frac{\varepsilon }{\sigma _\varepsilon}$, and $\{ e^i\}_{i = 1,...,d}$ is the standard Euclidean coordinate of ${\mathbb{R}^d}$. Clearly ${c_\eta } \to 0$ when $\eta \to 0$. When $a_\varepsilon$ is proportional to $\varepsilon$, $\eta$ becomes a positive constant independent of $\varepsilon$ and $Q_\eta$ becomes a fixed domain of type ${Q_0}\backslash T_{0}$; this is the case considered by Tartar \cite{ref7}.
For each fixed $\eta>0$, the generalized cell problem admits a unique regular solution. We now recall two lemmas concerning the estimates of the cell problems. The proofs can be found in \cite{ref1}.
\begin{lemma}
The solution $(w^i_\eta, q_\eta^i)$ of the generalized cell problem has the estimates:
\ba\label{lem-cell-1}
\| \nabla w_{\eta}^i\| _{L^2(Q_\eta)}\leqslant C c_\eta,\quad\| w_\eta ^i\| _{L^2(Q_\eta)} \leqslant C,\quad
\| q_{\eta}^i\| _{L^2(Q_\eta)}\leqslant C c_\eta.
\ea
\end{lemma}
Define the scaled cell solutions
\[w_{\eta ,\varepsilon }^i( \cdot ): = w_\eta ^i(\frac{ \cdot }{\varepsilon }),\quad q_{\eta ,\varepsilon }^i( \cdot ): = q_\eta ^i(\frac{ \cdot }{\varepsilon })\]
which solves
\ba\label{cell-general}
\begin{cases}
-\varepsilon ^2\Delta w_{\eta,\varepsilon}^i + \varepsilon\nabla q_{\eta,\varepsilon }^i = c_\eta ^2{e^i},&{\text{in}}\;\varepsilon {Q_0}\backslash ({a_\varepsilon}T_{0}),\\
{\mathrm{div}\,} w_{\eta,\varepsilon}^i= 0, & \mbox{in} \;{\varepsilon Q_\eta},\\
w_{\eta, \varepsilon}^i= 0, & \mbox{on} \;a_\varepsilon T_{0},\\
(w_{\eta, \varepsilon}^i,q_{\eta ,\varepsilon }^i) \ \mbox{is $\varepsilon Q_0$-periodic.}
\end{cases}
\ea
Employing the estimates of $(w_\eta ^i,q_\eta ^i)$ in \eqref{lem-cell-1} gives
\begin{lemma}
The scaled cell soluiton $(w_{\eta, \varepsilon}^i, q_{\eta, \varepsilon}^i)$ has the estimates:
\ba\label{lem-cell-2}
&\| w_{\eta, \varepsilon}^i\| _{L^2(\Omega )} \leqslant C\| {w_\eta ^i}\| _{L^2(Q_0)} \leqslant C,\\
&\| q_{\eta, \varepsilon}^i\| _{L^2(\Omega )} \leqslant C\| q_{\eta, \varepsilon}^i\| _{L^2(Q_0}
\leqslant C {c_\eta},\\
&\| \nabla w_{\eta, \varepsilon}^i\| _{L^2(\Omega)}
\leqslant C \varepsilon^{-1}\| \nabla w_\eta ^i\| _{L^2(Q_0)}
\leqslant C \varepsilon^{-1}{c_\eta} \leqslant C\sigma_\varepsilon^{-1}.
\ea
\end{lemma}
From these estimates we have
\begin{equation}\label{cell-3}
w_{\eta,\varepsilon}^i \to \overline w^i\ \mbox{weakly} \ \mbox{in} \ L^2 (\Omega),\quad
c_\eta^{-1}q_{\eta, \varepsilon}^i \to \overline q^i\ \mbox{weakly} \ \mbox{in}\ L^2(\Omega).
\end{equation}
\section{Proof of Theorem \ref{thm-1}}\label{sec-thm1}
This section is devoted to proving Theorem \ref{thm-1} concerning the case of small holes where $a_{\varepsilon} = \varepsilon^{\alpha}$ with $\alpha >3$ and $\sigma_{\varepsilon} = \varepsilon^{\frac{3-\alpha}{2}} \to \infty$ \eqref{ae-se}.
\subsection{Estimates of velocity}
In this case, the uniform estimates of ${\mathbf{u}}_{\varepsilon}$ follow directly from the energy inequality \eqref{6}. Indeed, using H\"older's inequality and Poincar\'e inequality gives
\ba\nn
&\frac{1}{2}\int_{\Omega_\varepsilon} \left| {{{{\mathbf{u}}}_\varepsilon}(x,t)} \right|^2 {\mathrm{d}} x
+ \int_0^t \int_{\Omega _\varepsilon} {\left| {\nabla {{{\mathbf{u}}}_\varepsilon }(x,s)} \right|}^2 {\mathrm{d}} x{\mathrm{d}} s\\
&\leqslant \int_0^t \int_{{\Omega _\varepsilon }} {\mathbf{f}}_{\varepsilon} \cdot {\mathbf{u}}_\varepsilon {\mathrm{d}} x{\mathrm{d}} s
+ \frac{1}{2}\int_{{\Omega _\varepsilon }} \left| {{{\mathbf{u}}}^0_{\varepsilon}(x)} \right|^2 {\mathrm{d}} x\\
& \leqslant C \sup_{0<\varepsilon\leqslant 1}\| {\mathbf{f}}_{\varepsilon}\| _{L^{2}(0,T; {L^2}({\Omega _\varepsilon }))}^{2} + \frac{1}{2}\| {{{\nabla {\mathbf{u}}}_\varepsilon }}\| _{L^{2}(0,t; {L^2}({\Omega _\varepsilon }))}^{2} + \frac{1}{2}\sup_{0<\varepsilon\leqslant1} \|{\mathbf{u}}_{\varepsilon}^{0}\|_{L^{2}(\Omega_{\varepsilon})}^{2}.
\ea
Together with the assumption on the initial datum and the external force in \eqref{ini-force}, we deduce
\begin{equation}\label{est-u-0}
\| {\mathbf{u}}_\varepsilon \| _{{L^\infty}(0,T;{L^2}({\Omega _\varepsilon}))} \leqslant C, \quad\| {\nabla {{{\mathbf{u}}}_\varepsilon }}\| _{{L^2}(0,T;{L^2}({\Omega _\varepsilon }))} \leqslant C.
\end{equation}
Since ${{\mathbf{u}}}_\varepsilon \in L^{2}(0,T; W^{1,2}_{0}(\Omega_{\varepsilon}))$ has zero trace on the boundary, its zero extension
$\widetilde {\mathbf{u}}_{\varepsilon} \in L^{\infty}(0,T; \\ L^{2}(\Omega)) \cap L^{2}(0,T; W^{1,2}_{0}(\Omega))$ has the estimates:
\begin{equation}\label{est-u}
\| \widetilde {\mathbf{u}}_\varepsilon \| _{L^\infty(0,T;L^2(\Omega))} \leqslant C,
\quad\| \widetilde {\mathbf{u}}_\varepsilon \| _{L^2 (0,T;W_0^{1,2}(\Omega))} \leqslant C.
\end{equation}
Thus, up to a subsequence, there holds the convergence
\ba\label{thm1-conv-0}
{\widetilde {{\mathbf{u}}}_\varepsilon } \to {{\mathbf{u}}} \ \mbox{weakly(*) in} \ L^{\infty}(0,T;L^{2}(\Omega)) \cap L^2(0,T;W_0^{1,2}(\Omega)),
\ea
which is exactly \eqref{thm1-conv} in Theorem \ref{thm-1}. Moreover, by the definition of ${\mathbf{U}}_\varepsilon$ in \eqref{1-new-def} we have
\ba\label{est-small-U}
\| \widetilde {\mathbf{U}}_\varepsilon \|_{W^{1,\infty}((0,T); L^2(\Omega))}\leqslant C,
\quad \|\widetilde {\mathbf{U}}_\varepsilon \|_{W^{1,2}((0,T); W^{1,2}_{0}(\Omega))}\leqslant C, \quad \| \widetilde {\mathbf{U}}_\varepsilon \|_{C([0,T]; W^{1,2}_{0}(\Omega))}\leqslant C.
\ea
\subsection{Extension of pressure}\label{sec:extP-small}
The extension of the pressure is given by using the dual formula and employing the so-called {\em restriction operator} due to Allaire \cite{ref3, ref4} for general sizes of holes, and due to Tartar \cite{ref7} for the case where the size of the holes is proportional to their mutual distance. A restriction operator $R_{\varepsilon} $ is a linear operator $R_{\varepsilon} : W^{1,2}_{0}(\Omega;{\mathbb{R}}^{d}) \to W^{1,2}_{0}(\Omega_{\varepsilon};{\mathbb{R}}^{d}) $ such that:
\ba\label{pt-res}
&{\mathbf{u}} \in W_0^{1,2}(\Omega_\varepsilon;{\mathbb{R}}^d) \Longrightarrow R_\varepsilon (\widetilde {\mathbf{u}})={\mathbf{u}} \ \mbox{in}\ \Omega_\varepsilon,\ \mbox{where} \ \widetilde
{\mathbf{u}}:=\begin{cases}{\mathbf{u}} \ &\mbox{in}\ \Omega_\varepsilon,\\ 0 \ &\mbox{on}\ \Omega\setminus \Omega_\varepsilon, \end{cases}\\
&{\mathbf{u}} \in W_0^{1,2}(\Omega;{\mathbb{R}}^d),\ {\mathrm{div}\,}{\mathbf{u}} =0 \ \mbox{in} \ \Omega \Longrightarrow {\mathrm{div}\,} R_\varepsilon ({\mathbf{u}}) =0 \ \mbox{in} \ \Omega_\varepsilon,\\
&{\mathbf{u}} \in W_0^{1,2}(\Omega;{\mathbb{R}}^d)\Longrightarrow \|\nabla R_\varepsilon({\mathbf{u}})\|_{L^2(\Omega_\varepsilon)} \leqslant C \, \big( \|\nabla {\mathbf{u}}\|_{L^2(\Omega)} + (1+\sigma_{\varepsilon}^{-1}) \|{\mathbf{u}}\|_{L^2(\Omega)}\big).
\ea
For each $\varphi\in L^{q}(0,T; W^{1,2}_0(\Omega))$ with $1<q<\infty$, the restriction $R_{\varepsilon}(\varphi)$ is taken only on spatial variable:
\[R_\varepsilon(\varphi)(\cdot,t)=R_\varepsilon(\varphi(\cdot,t))(\cdot) \quad \mbox{for each $t\in (0,T)$}. \]
Clearly $R_{\varepsilon}$ maps $L^{q}(0,T; W^{1,2}_0(\Omega))$ onto $L^{q}(0,T; W^{1,2}_0(\Omega_{\varepsilon}))$ with the estimate:
\ba\label{pt-res-2}
\|\nabla R_\varepsilon(\varphi)\|_{L^{q}(0,T; L^{2}(\Omega_\varepsilon))} \leqslant C \, \big( \|\nabla \varphi\|_{L^{q}(0,T; L^{2}(\Omega))} + (1+\sigma_{\varepsilon}^{-1}) \|\varphi\|_{L^{q}(0,T; L^{2}(\Omega))}\big).
\ea
\begin{lemma}\label{lemma-large}
Let $1<q<\infty$. Assume $H \in L^{q}(0,T; W^{-1,2}(\Omega;{\mathbb{R}}^{3}))$ satisfying
\ba
\l H, \varphi \r_{\Omega\times(0,T)}=0,\ \forall \varphi\in C_{c}^{\infty}(\Omega \times (0,T) ;{\mathbb{R}}^{3}), \ {\mathrm{div}\,} \varphi=0.
\nn\ea
Then there exists a scalar function $P\in L^{q}(0,T; L_0^2(\Omega))$ such that:
\ba\label{H-nablaP}
H = \nabla P, \ \mbox{with} \ \| P\| _{L^{q}(0,T;L_0^2(\Omega))} \leqslant C\| H\| _{L^{q}(0,T;W^{-1,2}(\Omega))}.
\ea
\end{lemma}
\begin{proof}
For each $\psi \in C_c^\infty(\Omega;{\mathbb{R}}^{3})$ with ${\mathrm{div}\,} \psi\equiv 0$ and $ \phi\in C_c^\infty(0,T)$, there holds
\ba\nonumber
0 = \l H, \phi(t)\psi(x) \r_{\Omega\times(0,T)} = \int_0^T \l H(\cdot,t), \psi(\cdot) \r_{\Omega} \phi(t){\mathrm{d}} t.
\ea
Thus we have for a.a. $t\in (0,T)$ that
\ba\nonumber
\l H(\cdot,t), \psi(\cdot) \r_{\Omega}=0, \quad \forall \psi \in C_c^\infty(\Omega;{\mathbb{R}}^{3}), \ {\mathrm{div}\,} \psi\equiv 0.
\ea
Therefore, for a.a. $t\in(0,T)$, there exists $P(\cdot,t)\in L_0^2(\Omega)$ (see \cite{ref8}), such that
\ba\nonumber
H (\cdot,t)=\nabla P(\cdot, t),\ \mbox{with}\ \| P(\cdot, t)\| _{L_0^2(\Omega)} \leqslant\| H (\cdot,t)\| _{W^{-1,2}(\Omega)}.
\ea
This implies immediately \eqref{H-nablaP}.
\end{proof}
Now we define a functional $\widetilde H_\varepsilon$ in $\mathcal{D}'(\Omega\times(0,T))$ by the following dual formulation:
\ba\nn
\l\widetilde H_\varepsilon,\varphi\r_{\Omega\times(0,T)}=\l\nabla P_\varepsilon,R(\varphi)\r_{\Omega_\varepsilon\times(0,T)},\ \forall \varphi \in C_c^\infty(\Omega\times(0,T);{\mathbb{R}}^{3}),
\ea
where $P_{\varepsilon}\in C([0,T],L_{0}^{2}(\Omega_{\varepsilon}))$ is given in \eqref{2-equa}. Then for any $\varphi \in C_c^\infty (\Omega\times(0,T); \mathbb{R}^3)$,
\ba\label{def-F-P}
\l \widetilde H_\varepsilon,\varphi \r_{\Omega\times(0,T)}&
=\l\nabla {P_\varepsilon}(t), R_\varepsilon (\varphi) \r_{\Omega_\varepsilon\times(0,T)}\\
&=\l{\mathbf{F}}_\varepsilon(t)-{\mathbf{u}}_\varepsilon(t)+{\mathbf{u}}^0_\varepsilon + \mu \Delta {\mathbf{U}}_\varepsilon(t)
-{\Phi_\varepsilon}(t),{R_\varepsilon}(\varphi) \r_{\Omega_\varepsilon\times(0,T)}.
\ea
Using the uniform estimates in \eqref{est-u-0}--\eqref{est-small-U} implies
\ba\label{est-P-small-1}
\left| \l \Delta {\mathbf{U}}_\varepsilon(t), R_\varepsilon (\varphi) \r_{\Omega_\varepsilon\times(0,T)} \right|
&\leqslant\| \nabla {\mathbf{U}}_\varepsilon(t)\| _{L^2 L^2}
\| \nabla R_\varepsilon (\varphi)\| _{L^2 L^2} \leqslant C\| \varphi \| _{L^2 W^{1,2}_0},\\
\left| \l\Phi_\varepsilon(t), R_\varepsilon(\varphi)\r_{\Omega_\varepsilon\times(0,T)} \right|
&= \left| \l \Psi_\varepsilon(t), \nabla R_\varepsilon(\varphi)\r_{\Omega_\varepsilon\times(0,T)} \right| \leqslant\| \Psi_{\varepsilon}\| _{L^\infty L^2}\| \nabla R_{\varepsilon} (\varphi )\| _{L^2 L^{2}}\\
&\leqslant C\|{\mathbf{u}}_{\varepsilon}\otimes {\mathbf{u}}_{\varepsilon}\|_{L^{1}L^{2}}\| \varphi\| _{L^2 W^{1,2}_0}\leqslant C\| \varphi\| _{L^2 W^{1,2}_0},\\
\left| \l {{\mathbf{u}}}_\varepsilon, R_\varepsilon (\varphi) \r_{\Omega_\varepsilon\times(0,T)} \right|
&\leqslant C\| {\mathbf{u}}_\varepsilon\| _{L^\infty(0,T; L^2(\Omega _\varepsilon))} \| R_\varepsilon (\varphi) \| _{L^2 L^2}\leqslant C\| \varphi \| _{L^2 W^{1,2}_0},\\
\left| \l {{\mathbf{u}}}^0_\varepsilon, R_\varepsilon (\varphi) \r_{\Omega _\varepsilon\times(0,T)} \right|
&\leqslant C \| {\mathbf{u}}^0_\varepsilon\| _{L^2} \| R_\varepsilon (\varphi)\| _{L^2 L^2} \leqslant C\| \varphi\| _{L^2 W^{1,2}_0},\\
\left| \l {\mathbf{F}}_\varepsilon, R_\varepsilon (\varphi) \r_{\Omega _\varepsilon\times(0,T)} \right|
&\leqslant C \| {\mathbf{F}}_\varepsilon\| _{L^2 L^2}\| R_\varepsilon (\varphi) \| _{L^2 L^2}\leqslant C\| \varphi\| _{L^2 W^{1,2}_0}.
\ea
Estimates in \eqref{est-P-small-1} imply
\ba\label{est-small-F}
\l \widetilde H_\varepsilon,\varphi \r_{\Omega\times(0,T)}
&\leqslant C \| \varphi \| _{L^2 W^{1,2}_0}.
\ea
Thus $\widetilde H_\varepsilon$ is bounded in $L^{2}(0,T; W^{-1,2}(\Omega;{\mathbb{R}}^{3}))$. Moreover, the second property of the restriction operator in \eqref{pt-res} implies that
\ba
\l \widetilde H_\varepsilon, \varphi \r_{\Omega\times(0,T)}=0,\ \forall \varphi\in C_{c}^{\infty}(\Omega \times (0,T) ;{\mathbb{R}}^{3}), \ {\mathrm{div}\,} \varphi=0.
\nn\ea
Thus, by \eqref{est-small-F}, we can apply Lemma \ref{lemma-large} and deduce that there exists $\widetilde P_\varepsilon \in L^2(0,T; L^2_0(\Omega)) $ such that
\ba\label{small-F-P}
\widetilde H_\varepsilon=\nabla \widetilde P_\varepsilon,
\ea
and
\ba\label{est-small-P}
\| \widetilde P_\varepsilon \|_{L^2(0,T; L^2_0(\Omega))} \leqslant C \| \widetilde H_\varepsilon \|_{L^2(0,T; W^{-1,2}(\Omega))} \leqslant C.
\ea
\subsection{Momentum equations in the homogeneous domain}
In the case of small holes, we find that the extension $\widetilde {\mathbf{u}}_\varepsilon$ satisfies the Navier-Stokes equations in $\Omega$ up to a small remainder:
\begin{proposition}\label{moment-equa}
Under the assumptions in Theorem \ref{thm-1}, the extension $\widetilde {\mathbf{u}}_\varepsilon$ satisfies the following equations in the sense of distribution:
\[\partial _t \widetilde {\mathbf{u}}_\varepsilon + {\mathrm{div}\,} \left(\widetilde {\mathbf{u}}_\varepsilon \otimes \widetilde {\mathbf{u}}_\varepsilon \right)- \mu \Delta \widetilde {\mathbf{u}}_\varepsilon + \nabla \hat p_\varepsilon = \widetilde {\mathbf{f}}_\varepsilon +\mathbf{G}_\varepsilon, \ {\mathrm{div}\,} \widetilde {\mathbf{u}}_{\varepsilon} = 0, \]
where ${\mathbf{G}}_\varepsilon \in \mathcal{D}' (\Omega \times (0,T))$ satisfying
\begin{equation}\label{est-F-3D}
\left| \l {\mathbf{G}}_\varepsilon, \varphi \r \right| \leqslant C\varepsilon ^\sigma \big(\| \partial _t \varphi \| _{L^{\frac 43} L^{2}}
+\| \nabla \varphi\| _{L^{4} L^{r_1}} \big), \ \forall \, \varphi\in C_{c}^{\infty}(\Omega\times (0,T);\mathbb{R}^3), \ {\mathrm{div}\,} \varphi =0.
\end{equation}
Here $\sigma:=((3 - q)\alpha - 3)/q >0$ for some $q>2$ close to $2$, and $ 2<r_1<3$ given in \eqref{r1}.
\end{proposition}
\begin{proof}
Let $\varphi \in C_c^\infty (\Omega \times (0,T);{\mathbb{R}^3})$ with ${\mathrm{div}\,}\varphi = 0$. To extend the Navier-Stokes equations from $\Omega$ to $\Omega_\varepsilon$, an idea is to find a family of functions ${\{ {g_\varepsilon}\} _{\varepsilon>0}}$ vanishing on the holes and converges to $1$ in some Sobolev space $W^{1,q}(\Omega)$ and decompose $\varphi$ as
\be\label{dec-g}
\varphi = {g_\varepsilon }\varphi + (1-{g_\varepsilon})\varphi.
\ee
Then $g_\varepsilon\varphi$ can be treated as a test function for the momentum equations in $\Omega_\varepsilon$. While for the terms related the other part $(1-g_\varepsilon)\varphi$, we show that they are small and converge to zero. However, such a decomposition destroyed the divergence free property of $\varphi$: ${\mathrm{div}\,} (g_{\varepsilon} \varphi) \neq 0$. To overcome this trouble, we introduce the following Bogovskii type operator in perforated domain $\Omega_{\varepsilon}$ (see Proposition 2.2 in \cite{ref9} and Theorem 2.3 in \cite{ref15}):
\begin{lemma}\label{lem-Bog}
Let $\Omega_\varepsilon$ defined as in \eqref{1-hole} and \eqref{1-domain} with $\alpha\geq 1$. Then for any $1<q<\infty$, there exists a linear operator $\mathcal{B}_\varepsilon \,: \, L^q_{0}(\Omega_\varepsilon)\to W_{0}^{1,q}(\Omega_\varepsilon;{\mathbb{R}}^3)$ such that for any $f\in L^q_{0}(\Omega_\varepsilon)$, there holds
\be\label{pro-div1}
{\mathrm{div}\,} \mathcal{B}_\varepsilon(f)=f\ \mbox{in} \ \Omega_\varepsilon, \quad \|\mathcal{B}_\varepsilon(f)\|_{W_{0}^{1,q}(\Omega_\varepsilon;{\mathbb{R}}^3)}\leq C\big(1+\varepsilon^{\frac{(3-q)\alpha-3}{q}}\big)\|f\|_{L^q(\Omega_\varepsilon)}
\ee
for some constant $C$ independent of $\varepsilon$.
For any $r >3/2$, the linear operator $\mathcal B_\varepsilon$ can be extended as a linear operator from $\{{\mathrm{div}\,} {\mathbf{g}} : {\mathbf{g}}\in L^r(\Omega_\varepsilon;{\mathbb{R}}^3),
{\mathbf{g}} \cdot {\bf n}=0 \mbox{ on } \partial \Omega_\varepsilon\}$ to $L^r(\Omega_\varepsilon;{\mathbb{R}}^3)$ satisfying
\be\label{pro-div2}
\|\mathcal{B}_\varepsilon({\mathrm{div}\,} {\bf g})\|_{L^{r}(\Omega_\varepsilon;{\mathbb{R}}^3)}\leq C \|{\bf g}\|_{L^r(\Omega_\varepsilon;{\mathbb{R}}^3)},
\ee
for some constant $C$ independent of $\varepsilon$.
\end{lemma}
By the description of the holes in \eqref{1-hole}, there exists cut-off functions $\{ g_\varepsilon \}_{\varepsilon>0} \subset C^\infty({\mathbb{R}}^{3})$ such that $0\leqslant g_{\varepsilon} \leqslant 1$ and
\begin{equation}\nn
g_\varepsilon= 0 \ \mbox{on} \ \bigcup_{k \in K_\varepsilon} B(\varepsilon {x_k},\delta_1 \varepsilon^\alpha),\quad g_\varepsilon = 1 \ \mbox{on} \ (\bigcup_{k \in K_\varepsilon} B(\varepsilon {x_k},\delta_2 \varepsilon^\alpha))^c,\quad |\nabla g_{\varepsilon}| \leqslant C \varepsilon^{-\alpha}.
\end{equation}
Then for each $1\leqslant q \leqslant \infty$ there holds
\begin{equation}\label{est-g}
\| g_\varepsilon- 1\| _{L^q({\mathbb{R}}^{3})} \leqslant C \varepsilon^{\frac{3\alpha-3}{q}},\quad
\| \nabla g_\varepsilon\| _{L^q({\mathbb{R}}^{3})} \leqslant C\varepsilon ^{\frac{3\alpha-3}{q}-\alpha}.
\end{equation}
Now we estimate
\ba
I^\varepsilon &:= \int_0^T \int_\Omega \widetilde {\mathbf{u}}_\varepsilon \partial _t \varphi
+\widetilde {\mathbf{u}}_\varepsilon \otimes \widetilde {\mathbf{u}}_\varepsilon :\nabla \varphi
-\nabla \widetilde {\mathbf{u}}_\varepsilon :\nabla \varphi + \widetilde {\mathbf{f}}_\varepsilon \varphi \, {\mathrm{d}} x {\mathrm{d}} t.
\nonumber\ea
Using the decomposition \eqref{dec-g} we write
\ba
I^\varepsilon &=\int_0^T \int_{\Omega_\varepsilon}{\mathbf{u}}_\varepsilon \partial_t (g_\varepsilon \varphi)
+ {\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon:\nabla (g_\varepsilon\varphi)
-\nabla {\mathbf{u}}_\varepsilon:\nabla (g_\varepsilon\varphi ) + {\mathbf{f}}_\varepsilon (g_\varepsilon \varphi ){\mathrm{d}} x {\mathrm{d}} t +\sum_{j=1}^{4} {I_j}
\nonumber\ea
with
\ba\nn
I_1 & = \int_0^T \int_\Omega \widetilde {\mathbf{u}}_\varepsilon(1-g_\varepsilon) \partial _t\varphi {\mathrm{d}} x {\mathrm{d}} t, \\
I_2 & = \int_0^T \int_\Omega \widetilde {\mathbf{u}}_\varepsilon \otimes \widetilde {\mathbf{u}}_\varepsilon :(1-g_\varepsilon)\nabla \varphi
-\widetilde {\mathbf{u}}_\varepsilon \otimes \widetilde {\mathbf{u}}_\varepsilon :\nabla g_\varepsilon \otimes \varphi {\mathrm{d}} x {\mathrm{d}} t ,\\
I_3 &= \int_0^T \int_\Omega \nabla \widetilde {\mathbf{u}}_\varepsilon :(1-g_\varepsilon)\nabla \varphi
- \nabla \widetilde {\mathbf{u}}_\varepsilon:\nabla g_\varepsilon \otimes \varphi {\mathrm{d}} x {\mathrm{d}} t ,\\
I_4 &= \int_0^T \int_\Omega \widetilde {\mathbf{f}}_\varepsilon (g_\varepsilon - 1)\varphi {\mathrm{d}} x{\mathrm{d}} t.
\ea
Observe that
$$
\int_{\Omega_{\varepsilon}} \varphi \cdot \nabla g_{\varepsilon}{\mathrm{d}} x = \int_{\Omega_{\varepsilon}} {\mathrm{div}\,}(\varphi g_{\varepsilon}) {\mathrm{d}} x = 0.
$$
Thus, we can apply Lemma \ref{lem-Bog} and introduce
\ba\nn
\varphi_{1} := \varphi - \varphi_{2},\quad \varphi_{2} := \mathcal{B}_{\varepsilon} ({\mathrm{div}\,}(\varphi g_{\varepsilon})) = \mathcal{B}_{\varepsilon} (\varphi \cdot \nabla g_{\varepsilon}).
\ea
Then $\varphi_{1} \in C_{c}^{\infty}( (0,T); W^{1,2}_{0}(\Omega_{\varepsilon}))$ satisfying ${\mathrm{div}\,} \varphi_{1} = 0$. Using the weak formulation \eqref{5} gives
\ba
I^\varepsilon & = \int_0^T \int_{\Omega_\varepsilon}{\mathbf{u}}_\varepsilon \partial_t \varphi_{1}
+ {\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon:\nabla \varphi_{1}
-\nabla {\mathbf{u}}_\varepsilon:\nabla \varphi_{1} + {\mathbf{f}}_\varepsilon \cdot \varphi_{1} {\mathrm{d}} x {\mathrm{d}} t + \sum_{ j=1}^{4} {I_j} + \sum_{ j=5}^{8} {I_j} \\
&= \sum_{ j=1}^{4} {I_j} + \sum_{ j=5}^{8} {I_j} ,
\nonumber\ea
where
\ba
I_5 &=\int_0^T \int_{\Omega_\varepsilon}{\mathbf{u}}_\varepsilon \partial_t \varphi_{2} {\mathrm{d}} x {\mathrm{d}} t,\quad &&I_6 = \int_0^T \int_{\Omega_\varepsilon} {\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon : \nabla \varphi_{2} {\mathrm{d}} x {\mathrm{d}} t , \\
I_7 &= \int_0^T \int_{\Omega_\varepsilon} -\nabla {\mathbf{u}}_\varepsilon:\nabla \varphi_{2} {\mathrm{d}} x {\mathrm{d}} t ,\quad &&I_8 = \int_0^T \int_{\Omega_{\varepsilon}} {\mathbf{f}}_\varepsilon \cdot \varphi_{2}{\mathrm{d}} x{\mathrm{d}} t.
\nonumber \ea
Now we estimate $I_{j}$ term by term. Since $\alpha>3$, there exists $q\in(2, 3)$ close to $2$ such that
\be\nn
\sigma:=\frac{(3 - q)\alpha - 3}{q} >0.
\ee
Let $q^*$ be the Sobolev conjugate component to $q$ with $\frac{1}{q^{*}} = \frac{1}{q} - \frac{1}{3}$. Clearly $6<q^{*}<\infty$.
\medskip
By the uniform estimates \eqref{est-u} and \eqref{est-g}, together with interpolation and Sobolev embedding, we obtain
\[|I_1| \leqslant\| \widetilde {\mathbf{u}}_\varepsilon \| _{L^4 L^3}\| 1- g_\varepsilon \| _{L^6}\| \partial_t\varphi\| _{L^{\frac 43} L^2} \leqslant \| \widetilde {\mathbf{u}}_\varepsilon \| _{L^\infty L^2 \cap L^{2} L^{6}} \| 1- g_\varepsilon \| _{L^6}\| \partial_t\varphi\| _{L^{\frac 43} L^2} \leqslant C \varepsilon ^\sigma\| \partial _t\varphi \| _{L^{\frac 43} L^2} .\]
Again by interpolation and using Sobolev embedding, we have
\ba
|I_2| \leqslant \| \widetilde {\mathbf{u}}_\varepsilon \otimes \widetilde {\mathbf{u}}_\varepsilon \| _{L^{\frac{4}{3}} L^2}\big(
\| g_\varepsilon - 1\| _{L^{q^{*}}}\| \nabla \varphi \| _{L^4L^{r_{1}}} + \| \nabla g_\varepsilon \| _{L^q}\| \varphi \| _{L^{4} L^{r_2}}\big) \leqslant C \varepsilon ^\sigma\| \nabla \varphi \| _{L^{4} L^{r_1}} ,
\nonumber\ea
where
\ba\label{r1}
\frac{1}{r_{1}} = \frac{1}{2} - \frac{1}{q_{*}} = \frac{5}{6} - \frac{1}{q}>\frac{1}{3}, \quad \frac{1}{r_{2}} = \frac{1}{2} - \frac{1}{q} = \frac{1}{r_{1}^{*}}.
\ea
Similarly,
\ba
|I_3| \leqslant\| \nabla \widetilde {\mathbf{u}}_\varepsilon \| _{L^2 L^2} \big(\| g_\varepsilon - 1\| _{L^{q^*}}
\| \nabla \varphi\| _{L^2 L^{r_1}} +\| \nabla g_\varepsilon\| _{L^q}\| \varphi\| _{L^2 L^{r_2}}\big)\leqslant C\varepsilon ^\sigma\| \nabla \varphi\| _{L^2 L^{r_1}}.
\nonumber\ea
For $I_{4}$,
\[|I_4| \leqslant\| {\mathbf{f}}_\varepsilon\| _{L^2 L^2}\| g_\varepsilon - 1\| _{L^{q^*}}\| \varphi \| _{L^2 L^{r_1}}
\leqslant C \varepsilon ^\sigma\| \varphi \| _{L^2 L^{r_1}}.\]
\medskip
Next we estimate $I_{j}, j=5,6,7,8$ for which the estimates of the Bogovskii operator $\mathcal{B}_{\varepsilon}$ in \eqref{pro-div1} and \eqref{pro-div2} will be repeatedly used. Since the Bogovskii operator $\mathcal{B}_{\varepsilon}$ only applies on spatial variable, then $$\partial_{t} \varphi_{2} = \partial_{t} \mathcal{B}_{\varepsilon}(\varphi \cdot \nabla g_{\varepsilon}) = \mathcal{B}_{\varepsilon}(\partial_{t}\varphi \cdot \nabla g_{\varepsilon}) = \mathcal{B}_{\varepsilon} \big({\mathrm{div}\,}((\partial_{t} \varphi) g_{\varepsilon})\big) = \mathcal{B}_{\varepsilon} \big({\mathrm{div}\,}((\partial_{t} \varphi) (g_{\varepsilon}-1))\big).$$ Thus, taking $r_{5}>3/2$ and using \eqref{pro-div2} gives
\ba
|I_5|& \leqslant\| {\mathbf{u}}_\varepsilon \| _{L^4 L^3} \|\partial_{t} \varphi_{2} \|_{L^{\frac{4}{3}} L^{r_{5}}} \leqslant C \| \partial_{t}\varphi ( g_\varepsilon -1 )\| _{L^{\frac 43} L^{r_{5}}} \\
& \leqslant C\| g_\varepsilon - 1\| _{L^{q^*}}\| \partial_{t} \varphi \| _{L^{\frac 43} L^{r_{6}}} \leqslant C \varepsilon^\sigma\| \partial_{t} \varphi \| _{L^{\frac 43} L^{r_{6}}},
\nn\ea
where
$$
\frac{1}{r_{6}} = \frac{1}{r_{5}} - \frac{1}{q^{*}} = \frac{1}{r_{5}} - \frac{1}{q} + \frac{1}{3}.
$$
Since $q>2$, we can choose $r_{5}>3/2$ close to $3/2$ such that
$$
\frac{1}{r_{5}} - \frac{1}{q} = \left(\frac{1}{r_{5}} - \frac{2}{3} \right) + \left( \frac{1}{2} - \frac{1}{q} \right)+ \frac{1}{6} = \frac{1}{6}.
$$
With such a choice of $r_{5}$ we have $ r_{6} = 2$ and
\ba
|I_5| \leqslant C \varepsilon^\sigma\| \partial_{t} \varphi \| _{L^{\frac 43} L^{2}}.
\nn\ea
Using \eqref{pro-div1}, by similar argument one has
\ba
|I_6| + |I_{7}| + |I_{8}| \leqslant C \varepsilon ^\sigma\| \nabla \varphi \| _{L^{4} L^{r_1}}.
\nonumber\ea
Summing up the above estimates for $I_{j}$ gives
$$
|I^\varepsilon| \leqslant C\varepsilon ^\sigma \big(\| \partial _t \varphi \| _{L^{\frac 43} L^{2}}
+\| \nabla \varphi\| _{L^{4} L^{r_1}} \big).
$$
This implies our desired result in \eqref{est-F-3D}.
\end{proof}
\subsection{Convergence of the nonlinear convective term}
Here we will show $\widetilde {\mathbf{u}}_{\varepsilon}$ has certain compactness such that the convergence of the nonlinear convective term $\widetilde {\mathbf{u}}_{\varepsilon}\otimes \widetilde {\mathbf{u}}_{\varepsilon}$ can be obtained. A key observation is that some uniform estimates related to time derivative can be deduced from Proposition \ref{moment-equa}. Indeed, By Proposition \ref{moment-equa}, for any $\varphi \in C_c^\infty ( \Omega \times (0,T) ;{\mathbb{R}^3})$ with ${\mathrm{div}\,}\varphi=0$, we have
\ba\nn
\left| \l \partial _t \widetilde {\mathbf{u}}_\varepsilon, \varphi \r \right| &
\leqslant \int_0^T \int_\Omega \left| \widetilde {\mathbf{u}}_\varepsilon
\otimes \widetilde {\mathbf{u}}_\varepsilon :\nabla \varphi \right| + \left| \nabla \widetilde {\mathbf{u}}_\varepsilon:\nabla \varphi \right| + \left| {\mathbf{f}}_\varepsilon \cdot \varphi \right| {\mathrm{d}} x{\mathrm{d}} t + \left| \l {\mathbf{G}}_\varepsilon,\varphi\r \right|\\
&\leqslant C\| \nabla \varphi\| _{L^4 L^2} + C\varepsilon^\sigma\big(\| \partial _t \varphi \| _{L^{\frac 43} L^{2}}
+\| \nabla \varphi\| _{L^{4} L^{r_1}} \big) \\
& \leqslant C\| \nabla \varphi\| _{L^{4} L^{r_1}} + C \varepsilon^\sigma\| \partial _t \varphi \| _{L^{\frac 43} L^{2}}.
\ea
Recall that $\sigma>0$ and $r_{1} \in (2,3)$. Thus we have the following decomposition
\[\widetilde {\mathbf{u}}_\varepsilon = \widetilde {\mathbf{u}}_\varepsilon ^{(1)} + \varepsilon ^\sigma \widetilde {\mathbf{u}}_\varepsilon ^{(2)},\]
where $\partial _t\widetilde {\mathbf{u}}_\varepsilon ^{(1)}$ is uniformly bounded in $L^{\frac{4}{3}}(0,T; V^{-1,r_{1}'}(\Omega)) $ and $ \widetilde \uu_{\e}^{(2)}$ is uniformly bounded in $L^{4}(0,T; L^{2}(\Omega))$. Since $\widetilde \uu_{\e}$ is uniformly bounded in $L^{\infty}(0,T; L^{2}(\Omega))\cap L^{2}(0,T;L^{6}(\Omega))$, then $\widetilde \uu_{\e}^{(1)} = \widetilde \uu_{\e} - \varepsilon^\sigma\widetilde \uu_{\e}^{(2)}$ is uniformly bounded in $L^{4}(0,T; L^{2}(\Omega))$. Thus, up to a subsequence,
$$
\widetilde \uu_{\e}^{(1)} \to {\mathbf{u}} \ \mbox{weakly in} \ L^{4}(0,T; L^{2}(\Omega)).
$$
Recall $\widetilde \uu_{\e}$ is uniformly bounded in $L^{2}(0,T; W^{1,2}_{0}(\Omega))$. Then apply Aubin-Lions type argument (see Lemma 5.1 in \cite{ref6.1}) gives
\[\widetilde {\mathbf{u}}_\varepsilon ^{(1)} \otimes \widetilde {\mathbf{u}}_\varepsilon \to {\mathbf{u}} \otimes {\mathbf{u}} \ \mbox{in} \ \mathcal{D}'( \Omega \times (0,T) ).\]
Clearly
$$
\varepsilon^{\sigma} \widetilde \uu_{\e}^{(2)} \otimes \widetilde \uu_{\e} \to 0 \ \mbox{strongly in} \ L^{\frac{4}{3}}(0,T; L^{\frac{3}{2}}(\Omega)) \cap L^{4}(0,T; L^{1}(\Omega)).
$$
Thus,
\ba\label{convective-conv-0}
\widetilde \uu_{\e}\otimes \widetilde \uu_{\e} \to {\mathbf{u}}\otimes {\mathbf{u}} \ \mbox{in} \ \mathcal{D}'( \Omega \times (0,T) ).
\ea
Together with the uniform estimates of $\widetilde \uu_{\e}$ in \eqref{est-u}, we deduce from \eqref{convective-conv-0} that
\ba\label{convective-conv}
\widetilde \uu_{\e}\otimes \widetilde \uu_{\e} \to {\mathbf{u}}\otimes {\mathbf{u}} \ \mbox{weakly in} \ L^{\frac{4}{3}}(0,T; L^{2}(\Omega) ).
\ea
\subsection{Passing to the limit}
Now we are ready to pass $\varepsilon \to 0$ and prove the following result in ${\mathbf{U}}_{\varepsilon}$ from which we can deduce our desired limit equations in ${\mathbf{u}}_{\varepsilon}$.
\begin{proposition}\label{thm-UP-crit}
Let $({\mathbf{U}}_\varepsilon, P_\varepsilon)$ be the solution of the equation \eqref{2-equa} and $\widetilde {P}_\varepsilon$ be the extension in $\Omega$ defined by \eqref{small-F-P}. If $\alpha>3$, i.e. $\lim _{\varepsilon \to 0}\sigma _\varepsilon=\infty$, then
\ba\label{U-P-conv}
&\widetilde {\mathbf{U}}_\varepsilon \to {\mathbf{U}} \ \mbox{weakly({*}) \ in}\ W^{1,\infty}((0,T); L^2(\Omega))\cap W^{1,2}((0,T); W^{1,2}_{0}(\Omega)),\\
&\widetilde P_\varepsilon \to P\ \mbox{weakly\ in}\ L^2(0,T; L_0^2(\Omega)).
\ea
Moreover, $({\mathbf{U}},P)$ solves
\begin{equation}\label{equa-U-P-small}
{\mathbf{u}} - {\mathbf{u}}^0 + {\mathrm{div}\,} \Psi - \mu \Delta {\mathbf{U}}+\nabla P= {\mathbf{F}},\ {\mathrm{div}\,} {\mathbf{U}}=0 \ \mbox{in}\ \mathcal{D}'(\Omega\times(0,T)),
\end{equation}
with
\ba\label{U-Psi-F}
{\mathbf{U}}(\cdot ,t) = \int_0^t {\mathbf{u}} (\cdot,s){\mathrm{d}} s, \quad \Psi (\cdot ,t): = \int_0^t {\mathbf{u}}(\cdot,s) \otimes {\mathbf{u}}(\cdot,s){\mathrm{d}} s, \quad {\mathbf{F}}(\cdot ,t) = \int_0^t {\mathbf{f}} (\cdot,s){\mathrm{d}} s,
\ea
where ${\mathbf{u}}$ is the limit of $\widetilde \uu_{\e}$ given in \eqref{thm1-conv-0}, ${\mathbf{u}}^{0}$ and ${\mathbf{f}}$ are given in \eqref{ini-force}.
\end{proposition}
\begin{proof}
The weak convergences in \eqref{U-P-conv} follow immediately from the uniform estimates \eqref{est-small-U} and \eqref{est-small-P}. The divergence free condition ${\mathrm{div}\,} {\mathbf{U}} = 0$ follows from ${\mathrm{div}\,} \widetilde {\mathbf{U}}_{\varepsilon}=0$. Moreover, the weak convergence of $\widetilde \uu_{\e}$ in \eqref{thm1-conv-0} implies for each $\varphi \in C_{c}^{\infty}(\Omega\times (0,T);{\mathbb{R}}^{3})$ that
\ba\label{U-u-small}
\int_{0}^{T} \int_{\Omega} \widetilde {\mathbf{U}}_{\varepsilon} \cdot \varphi(t,x) {\mathrm{d}} x {\mathrm{d}} t & = \int_{0}^{T} \int_{\Omega} \int_{0}^{t} \widetilde \uu_{\e}(s,x) {\mathrm{d}} s \cdot \varphi(t,x) {\mathrm{d}} x {\mathrm{d}} t \\
& = \int_{0}^{T} \int_{0}^{t} \int_{\Omega} \widetilde \uu_{\e}(s,x) \cdot \varphi(t,x) {\mathrm{d}} x {\mathrm{d}} s {\mathrm{d}} t \\
& \to \int_{0}^{T} \int_{0}^{t} \int_{\Omega} {\mathbf{u}}(s,x) \cdot \varphi(t,x) {\mathrm{d}} x {\mathrm{d}} s {\mathrm{d}} t \\
& = \int_{0}^{T} \int_{\Omega} \int_{0}^{t} {\mathbf{u}}(s,x) {\mathrm{d}} s \cdot \varphi(t,x) {\mathrm{d}} x {\mathrm{d}} t.
\ea
Thus, the uniqueness of weak limits implies
$$
{\mathbf{U}}(\cdot ,t) = \int_0^t {\mathbf{u}} (\cdot,s){\mathrm{d}} s,
$$
which is the first equality in \eqref{U-Psi-F}.
\medskip
Given any scalar test function $\phi \in C_{c}^{\infty}(\Omega \times (0,T) )$, we choose $w ^i_{\eta,\varepsilon}\phi$ as a test function to \eqref{2-equa} to obtain the weak formulation:
\ba\label{weak-equa-small}
&\int_0^T \int_\Omega \widetilde {\mathbf{u}}_\varepsilon \cdot (w_{\eta, \varepsilon}^i\phi)
-\widetilde \Psi _\varepsilon:\nabla (w_{\eta, \varepsilon}^i\phi)
+\mu \nabla {\mathbf{U}}_\varepsilon :\nabla (w_{\eta ,\varepsilon }^i\phi)
-\widetilde P_\varepsilon{\mathrm{div}\,}(w_{\eta, \varepsilon}^i\phi ){\mathrm{d}} x{\mathrm{d}} t \\
&=\int_0^T \int_\Omega {\mathbf{F}}_\varepsilon \cdot (w_{\eta,\varepsilon}^i\phi ){\mathrm{d}} x{\mathrm{d}} t+ \int_{0}^{T}\int_\Omega \widetilde {\mathbf{u}}_\varepsilon^0 \cdot (w_{\eta,\varepsilon}^i\phi ) \,{\mathrm{d}} x{\mathrm{d}} t.
\ea
By \eqref{lem-cell-2}, we know that $\| \nabla w_{\eta, \varepsilon}^i\| _{L^2(\Omega)} \leqslant C\sigma_\varepsilon ^{-1} = C \varepsilon^{\frac{\alpha-3}{2}} \to 0$. Then $w_{\eta, \varepsilon}^i \to {\overline w ^i}$ strongly in $L^2(\Omega)$ by Sobolev compact embedding. This implies
\ba\label{conv-small-1}
&\int_0^T \int_\Omega \widetilde{\mathbf{u}}_\varepsilon \cdot w_{\eta, \varepsilon}^i \phi {\mathrm{d}} x{\mathrm{d}} t \to
\int_0^T \int_\Omega {\mathbf{u}} \cdot {\overline w}^i \phi {\mathrm{d}} x{\mathrm{d}} t,\\
&\int_{0}^{T}\int_\Omega \widetilde {\mathbf{u}}_\varepsilon^0 \cdot (w_{\eta,\varepsilon}^i\phi ) \,{\mathrm{d}} x{\mathrm{d}} t \to
\int_{0}^{T} \int_\Omega {\mathbf{u}}^0 \cdot {\overline w}^i \phi {\mathrm{d}} x{\mathrm{d}} t.
\ea
For the nonlinear convective term one has
\[\int_0^T \int_\Omega \widetilde \Psi_\varepsilon :\nabla (w_{\eta, \varepsilon }^i\phi ){\mathrm{d}} x{\mathrm{d}} t
= \int_0^T \int_\Omega \widetilde \Psi_\varepsilon:\nabla w_{\eta, \varepsilon}^i\phi
+ \widetilde \Psi_\varepsilon: \nabla \phi \otimes w_{\eta, \varepsilon}^i {\mathrm{d}} x{\mathrm{d}} t,\]
where the first term on the right-hand side satisfies
\[\left| \int_0^T \int_\Omega \widetilde \Psi_\varepsilon :\nabla w_{\eta ,\varepsilon }^i\phi {\mathrm{d}} x{\mathrm{d}} t \right| \leqslant C\| \widetilde {\mathbf{u}}_\varepsilon \| _{L^2 L^6}^2
\| \nabla w_{\eta, \varepsilon}^i\| _{L^2}\| \phi \| _{L^\infty L^6}
\leqslant C \varepsilon ^{\frac{\alpha-3}{2}} \to 0.\]
For the second term, by the convergence of the convective term in \eqref{convective-conv} and the strong convergence of ${w_{\eta ,\varepsilon }^i}$, we have
\ba\nonumber
\int_0^T \int_\Omega \widetilde \Psi_\varepsilon :\nabla \phi \otimes w_{\eta ,\varepsilon }^i{\mathrm{d}} x{\mathrm{d}} t \to \int_0^T \int_\Omega \Psi :\nabla \phi \otimes \overline w ^i{\mathrm{d}} x{\mathrm{d}} t.
\ea
So we obtain
\ba\label{conv-small-3}
\int_0^T \int_\Omega \widetilde \Psi_\varepsilon:\nabla (w_{\eta, \varepsilon}^i\phi){\mathrm{d}} x{\mathrm{d}} t
\to \int_0^T \int_\Omega \Psi:\nabla ({\overline w}^i\varphi ){\mathrm{d}} x{\mathrm{d}} t.
\ea
Using $\nabla w_{\eta ,\varepsilon }^i \to 0$ and the strong convergence of ${w_{\eta ,\varepsilon }^i}$ again implies
\ba\label{conv-small-4}
\int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon:\nabla (w_{\eta, \varepsilon}^i\varphi){\mathrm{d}} x{\mathrm{d}} t &
=\int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla w_{\eta, \varepsilon}^i\varphi {\mathrm{d}} x{\mathrm{d}} t
+\int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon : \nabla \varphi \otimes w_{\eta, \varepsilon}^i{\mathrm{d}} x{\mathrm{d}} t\\
&\to \int_0^T \int_\Omega \nabla {\mathbf{U}} :\nabla \varphi \otimes {\overline w}^i{\mathrm{d}} x{\mathrm{d}} t
=\int_0^T \int_\Omega \nabla {\mathbf{U}} :\nabla ({\overline w}^i\varphi){\mathrm{d}} x{\mathrm{d}} t.
\ea
For the term related to the pressure, again by the strong convergence of ${w_{\eta ,\varepsilon }^i}$, we have
\ba\label{conv-small-5}
\int_0^T \int_\Omega \widetilde P_\varepsilon {\mathrm{div}\,} (w_{\eta,\varepsilon}^i \phi) {\mathrm{d}} x{\mathrm{d}} t
&=\int_0^T \int_\Omega \widetilde P_\varepsilon ( w_{\eta,\varepsilon}^i \cdot \nabla \phi) {\mathrm{d}} x{\mathrm{d}} t\\
&\to\int_0^T \int_\Omega P(\overline w^i\cdot\phi) {\mathrm{d}} x{\mathrm{d}} t
=\int_0^T \int_\Omega P {\mathrm{div}\,}(\overline w^i \phi) {\mathrm{d}} x{\mathrm{d}} t.
\ea
For the external force term, by the strong convergence of $\widetilde {\mathbf{F}}_{\varepsilon}$ in \eqref{Fe-F-st}, we have
\ba\label{conv-small-6}
\int_0^T \int_\Omega \widetilde {\mathbf{F}}_\varepsilon ( w_{\eta,\varepsilon}^i \phi) {\mathrm{d}} x{\mathrm{d}} t
\to\int_0^T \int_\Omega {\mathbf{F}} (\overline w^i\phi) {\mathrm{d}} x{\mathrm{d}} t.
\ea
Summarizing the convergences in \eqref{conv-small-1}--\eqref{conv-small-6} and passing $\varepsilon\to 0$ in \eqref{weak-equa-small} implies
\ba\label{equa-U-P-small-weak}
&\int_0^T \int_\Omega {\mathbf{u}} \cdot (\overline w^i\phi)-\Psi :\nabla (\overline w^i \phi)+\mu \nabla {\mathbf{U}}:\nabla (\overline w^i \phi)- P {\mathrm{div}\,}(\overline w^i \phi ){\mathrm{d}} x{\mathrm{d}} t \\&=\int_0^T \int_\Omega {\mathbf{F}}\cdot (\overline w^i \phi ){\mathrm{d}} x{\mathrm{d}} t+ \int_0^T \int_\Omega {\mathbf{u}}^0 \cdot (\overline w^i\phi ) \,{\mathrm{d}} x{\mathrm{d}} t.
\ea
Equation \eqref{equa-U-P-small-weak} is equivalent to
\ba\nonumber
&\int_0^T \int_\Omega {\mathbf{u}} \cdot (A\varphi)-\Psi :\nabla (A\varphi)+\mu \nabla {\mathbf{U}}:\nabla (A\varphi)- P {\mathrm{div}\,}(A\varphi ){\mathrm{d}} x{\mathrm{d}} t \\&=\int_0^T \int_\Omega {\mathbf{F}}\cdot (A\varphi ){\mathrm{d}} x{\mathrm{d}} t+\int_0^T \int_\Omega {\mathbf{u}}^0 \cdot (A\varphi ) \,{\mathrm{d}} x{\mathrm{d}} t
\ea
for any $\varphi\in C_c^\infty( \Omega \times (0,T) ;\mathbb{R}^3)$ with $A=(A_{i,j}) = (\overline w^{j})_{i} = (\overline w^{i})_{j}.$ As shown in \cite{Allaire91}, $A$ is a positive definite matrix. This implies
\[{\mathbf{u}} - {\mathbf{u}}^0+{\mathrm{div}\,}\Psi - \mu \Delta {\mathbf{U}}+\nabla P-{\mathbf{F}} = 0, \quad \mbox{in} \ \mathcal{D}'(\Omega\times (0,T)).\]
The proof of Proposition \ref{thm-UP-crit} is completed.
\end{proof}
\subsection{End of the proof}
From Proposition \ref{thm-UP-crit}, we can deduce the limit equations in ${\mathbf{u}}$. Indeed, differentiating in $t$ to equation \eqref{equa-U-P-small} implies
\ba\label{eq-uu-pf-0}
\partial_t {\mathbf{u}}+ {\mathrm{div}\,}({\mathbf{u}} \otimes {\mathbf{u}}) - \mu \Delta {\mathbf{u}}+ \nabla {p} = {\mathbf{f}}, \ \mbox{in} \ \mathcal{D}'(\Omega \times (0,T)),
\ea
with $p = \partial_{t} P$. By $\eqref{U-P-conv}_{2}$, we know that
\ba\nn
\widetilde p_{\varepsilon} : = \partial_{t} \widetilde P_{\varepsilon} \to p \ \mbox{weakly in} \ W^{-1,2}(0,T;L_{0}^{2}(\Omega)),
\ea
which is exactly \eqref{thm1-conv-p}.
Clearly ${\mathrm{div}\,} {\mathbf{u}} = 0$ which follows from ${\mathrm{div}\,} \widetilde {\mathbf{u}}_{\varepsilon} = 0$. Since ${\mathbf{u}} \in L^{\infty}(0,T; L^{2}(\Omega)) \cap L^{2}(0,T;W^{1,2}_{0}(\Omega))$, using equation \eqref{eq-uu-pf-0} implies
\ba\nn
\partial_t {\mathbf{u}} \in L^{\frac{4}{3}}(0,T;V^{-1,2}(\Omega)).
\ea
Thus
\ba\label{u-contin-t}
{\mathbf{u}} \in C_{weak}([0,T), L^{2}(\Omega))\cap C([0,T), L^{q}(\Omega)), \ \mbox{for any $1\leqslant q<2$}.
\ea
We shall further show the attainment of initial datum for ${\mathbf{u}}$. With such continuity of ${\mathbf{u}}$ in time variable in \eqref{u-contin-t}, by density argument, we deduce from \eqref{eq-uu-pf-0} that for each $\varphi \in C_{c}^{1}(\Omega \times [0,T);{\mathbb{R}}^{3}), \ {\mathrm{div}\,} \varphi =0$,
\ba\label{u-weak-1}
\int_{0}^{T} \int_\Omega - {\mathbf{u}} \partial_{t} \varphi - {\mathbf{u}}\otimes {\mathbf{u}} : \nabla \varphi + \nabla {\mathbf{u}} :\nabla \varphi {\mathrm{d}} x {\mathrm{d}} t = \int_{0}^{T} \int_\Omega {\mathbf{f}} \cdot \varphi + \int_{\Omega} {\mathbf{u}}(x, 0) \varphi(x,0) {\mathrm{d}} x.
\ea
Following the proof of Proposition \ref{moment-equa}, we can deduce for each $\varphi \in C_{c}^{1}(\Omega \times [0,T);{\mathbb{R}}^{3}), \ {\mathrm{div}\,} \varphi =0$ that
\ba\label{u-weak-2}
\int_{0}^{T} \int_\Omega - \widetilde \uu_{\e} \partial_{t} \varphi - \widetilde \uu_{\e} \otimes \widetilde \uu_{\e} : \nabla \varphi + \nabla \widetilde \uu_{\e} :\nabla \varphi {\mathrm{d}} x {\mathrm{d}} t = \int_{0}^{T} \int_\Omega \widetilde {\mathbf{f}}_{\varepsilon} \cdot \varphi + \int_{\Omega} {\mathbf{u}}_\varepsilon^{0}(x) \varphi(x,0) {\mathrm{d}} x + {\mathbf H}_{\varepsilon}(\varphi),
\ea
where the remainder term ${\mathbf H}_{\varepsilon}(\varphi)$ satisfies
$$
|{\mathbf{H}}_{\varepsilon}(\varphi)| \leq C \varepsilon^{\sigma} \big(\| \partial _t \varphi \| _{L^{\frac 43} L^{2}}
+\| \nabla \varphi\| _{L^{4} L^{r_1}} + \| \varphi(\cdot,0)\|_{L^{r_{1}^{*}}(\Omega)}\big),
$$
with $\sigma>0$ and $2<r_{1}<3$ given in Proposition \ref{moment-equa} and $r_{1}^{*}$ the Sobolev conjugate number of $r_{1}$ such that $\frac{1}{r_{1}^{*}} = \frac{1}{r_{1}} - \frac{1}{3}$.
By the convergence of the convective term in \eqref{convective-conv}, passing $\varepsilon\to 0$ in \eqref{u-weak-2} implies
\ba\label{u-weak-3}
\int_{0}^{T} \int_\Omega - {\mathbf{u}} \partial_{t} \varphi - {\mathbf{u}}\otimes {\mathbf{u}} : \nabla \varphi + \nabla {\mathbf{u}} :\nabla \varphi {\mathrm{d}} x {\mathrm{d}} t = \int_{0}^{T} \int_\Omega {\mathbf{f}} \cdot \varphi + \int_{\Omega} {\mathbf{u}}^{0}(x) \varphi(x,0) {\mathrm{d}} x.
\ea
Comparing \eqref{u-weak-1} and \eqref{u-weak-3} implies the attainment of initial datum:
$$
{\mathbf{u}}|_{t=0} = {\mathbf{u}}^{0}.
$$
\medskip
We obtain system \eqref{equa-u-p-small} and thus complete the proof of Theorem \ref{thm-1}.
\section{Proof of Theorem \ref{thm-3}}\label{sec-thm3}\label{sec-thm2}
In this section we shall prove Theorem \ref{thm-3} where $a_{\varepsilon} = \varepsilon^{\alpha}$ \eqref{ae-se} with $1<\alpha <3$ and $\sigma_{\varepsilon} = \varepsilon^{\frac{3-\alpha}{2}}$ concerning the case of large holes. In this case we consider the time-scaled Navier-Stokes system \eqref{1-time-equa}.
In the case of large holes, one can benefit from the zero boundary condition on the holes and obtain the following perforation version of Poincar\'e inequality (see Lemma 3.4.1 in \cite{ref4}):
\begin{lemma}\label{lem-Poincare}
Let $\Omega_{\varepsilon}$ be the perforated domain defined by \eqref{1-hole} and \eqref{1-domain} with $a_{\varepsilon} = \varepsilon^{\alpha}, \ 1<\alpha <3$. There exists a constant $C$ independent of $\varepsilon$ such that
\begin{equation}
\|u\|_{{L^2}({\Omega _\varepsilon })} \leqslant C\sigma_\varepsilon \|\nabla u\|_{{L^2}({\Omega _\varepsilon })}, \quad \mbox{for any $u \in W_0^{1,2}({\Omega _\varepsilon })$.}
\label{11}
\end{equation}
\end{lemma}
\subsection{Estimates of velocity}
From the energy inequality \eqref{6}, using Poincar\'e inequality \eqref{11} and H\"older's inequality gives
\ba\nn
&\frac{1}{2}\sigma _\varepsilon^2\int_{\Omega_\varepsilon} \left| {{\mathbf{u}}}_\varepsilon(x,t) \right|^2 {\mathrm{d}} x
+ \int_0^t \int_{\Omega _\varepsilon} \left| \nabla {\mathbf{u}}_\varepsilon (x,s) \right|^2 {\mathrm{d}} x{\mathrm{d}} s\\
&\leqslant \int_0^t \int_{\Omega _\varepsilon} {\mathbf{f}}_{\varepsilon} \cdot {\mathbf{u}}_\varepsilon {\mathrm{d}} x{\mathrm{d}} s
+ \frac{1}{2}\sigma _\varepsilon^2 \int_{\Omega_\varepsilon} \left| {{\mathbf{u}}}^0_{\varepsilon}(x) \right|^2 {\mathrm{d}} x\\
&\leqslant\| {\mathbf{f}}_{\varepsilon}\| _{L^{2}(0,t; L^2(\Omega_\varepsilon))}
\| {{\mathbf{u}}}_\varepsilon \| _{L^{2}(0,t; L^2(\Omega_\varepsilon ))}
+ \frac{1}{2}\sigma_\varepsilon^2\sup_{0<\varepsilon\leqslant 1} \|{\mathbf{u}}_{\varepsilon}^{0}\|_{L^{2}(\Omega_{\varepsilon})}^{2} \\
&\leqslant C \sigma _\varepsilon\| {\mathbf{f}}_{\varepsilon}\| _{L^{2}(0,t; L^2(\Omega _\varepsilon))}
\| \nabla {\mathbf{u}}_\varepsilon\| _{L^{2}(0,t; L^2(\Omega_\varepsilon))}
+\frac{1}{2}\sigma _\varepsilon^2 \sup_{0<\varepsilon\leqslant 1} \|{\mathbf{u}}_{\varepsilon}^{0}\|_{L^{2}(\Omega_{\varepsilon})}^{2}\\
& \leqslant C \sigma _\varepsilon^2 \sup_{0<\varepsilon\leqslant 1}\| {\mathbf{f}}_{\varepsilon}\| _{L^{2}(0,T; {L^2}({\Omega _\varepsilon }))}^{2} + \frac{1}{2}\| {{{\nabla {\mathbf{u}}}_\varepsilon }}\| _{L^{2}(0,t; {L^2}({\Omega _\varepsilon }))}^{2} + \frac{1}{2}\sigma _\varepsilon^2\sup_{0<\varepsilon\leqslant1} \|{\mathbf{u}}_{\varepsilon}^{0}\|_{L^{2}(\Omega_{\varepsilon})}^{2}.
\ea
This implies
\begin{equation}\label{est-large-ue}
\| {{\mathbf{u}}}_\varepsilon \| _{L^{\infty}(0,T;L^2(\Omega _\varepsilon))} \leqslant C, \quad\| \nabla {\mathbf{u}}_\varepsilon \| _{L^2 (0,T;L^2(\Omega _\varepsilon))}
\leqslant C \sigma_\varepsilon,\quad\| {\mathbf{u}}_\varepsilon \| _{L^2 (0,T;L^2(\Omega_\varepsilon))}
\leqslant C\sigma_\varepsilon^2,
\end{equation}
where we used again \eqref{11}. Therefore the extension satisfies
\begin{equation}\label{3-est-u}
\| \widetilde \uu_{\e}\| _{L^{\infty}(0,T;L^2(\Omega))} \leqslant C, \quad\| \nabla \widetilde \uu_{\e} \| _{L^2 (0,T;L^2(\Omega))}
\leqslant C \sigma_\varepsilon,\quad\| \widetilde \uu_{\e} \| _{L^2 (0,T;L^2(\Omega))}
\leqslant C \sigma_\varepsilon^2.
\end{equation}
Then, up to a subsequence, we have the convergence
\ba\label{UUe-conv-large}
\sigma_\varepsilon^{-2} \widetilde \uu_{\e}\to {\mathbf{u}} \ \mbox{weakly in} \ L^{2}(0,T; L^2(\Omega)),
\ea
which is exactly \eqref{thm2-u}.
By \eqref{1-new-def}, it follows from \eqref{3-est-u} that
\begin{equation}\label{3-estimate-U}
\|\widetilde {\mathbf{U}}_\varepsilon \|_{W^{1,\infty}(0,T; L^2(\Omega))}\leqslant C,\quad \| \nabla \widetilde {\mathbf{U}}_\varepsilon \|_{W^{1,2}(0,T; L^{2} (\Omega))}\leqslant C \sigma_\varepsilon,\quad
\| \widetilde {\mathbf{U}}_\varepsilon \|_{W^{1,2}(0,T; L^{2} (\Omega))}\leqslant C \sigma_\varepsilon^{2}.
\end{equation}
Then, up to a subsequence,
\ba\label{UUe-conv-large}
\sigma_\varepsilon^{-2} \widetilde {\mathbf{U}}_\varepsilon \to {\mathbf{U}} \ \mbox{weakly in} \ W^{1,2}(0,T; L^2(\Omega)).
\ea
As the argument in \eqref{U-u-small}, the uniqueness of the weak limits implies that
\ba\label{U-u-large}
{\mathbf{U}}(\cdot ,t) = \int_0^t {\mathbf{u}} (\cdot,s){\mathrm{d}} s.
\ea
\subsection{Extension of pressure}
Let $P_{\varepsilon}$ be given in the Stokes equations \eqref{2-equa-large-holes}. Recall the estimate of $R_\varepsilon$ in \eqref{pt-res-2}:
\[
\|\nabla R_\varepsilon(\varphi)\|_{L^q (0,T; L^{2}(\Omega_\varepsilon))} \leqslant C \, \big( \|\nabla \varphi\|_{L^q (0,T; L^{2}(\Omega_\varepsilon))} + \sigma_{\varepsilon}^{-1} \|\varphi\|_{L^q (0,T; L^{2}(\Omega_\varepsilon))}\big),
\]
which will be repeatedly used in this section. Here $\sigma_{\varepsilon} \to 0$ for the case of large holes. As in Section \ref{sec:extP-small}, we define a functional $\widetilde H_\varepsilon$ in $\mathcal{D}'(\Omega\times(0,T);{\mathbb{R}}^{3})$ by the following dual formulation: for each $\varphi \in C_c^\infty(\Omega\times(0,T);{\mathbb{R}}^{3}),$
\ba\label{def-large-F}
\l \widetilde H_\varepsilon ,\varphi \r_{\Omega\times(0,T)}&
=\l\nabla {P_\varepsilon}(t), R_\varepsilon (\varphi) \r_{\Omega_{\varepsilon}\times(0,T)}\\
&=\l{\mathbf{F}}_\varepsilon(t)-\sigma^2_\varepsilon{\mathbf{u}}_\varepsilon(t)+\sigma^2_\varepsilon{\mathbf{u}}^0_\varepsilon + \mu \Delta {\mathbf{U}}_\varepsilon(t)
- {\mathrm{div}\,} \Psi_\varepsilon (t),{R_\varepsilon}(\varphi) \r_{\Omega_{\varepsilon}\times(0,T)}.
\ea
Using the uniform estimates in \eqref{est-large-ue}--\eqref{3-estimate-U} implies
\ba\label{est-P-large-1}
\left| \l {\mathrm{div}\,} \Psi_\varepsilon(t), R_\varepsilon(\varphi)\r_{\Omega_\varepsilon\times(0,T)} \right| &
= \left| \l \Psi_\varepsilon(t), \nabla R_\varepsilon(\varphi)\r_{\Omega_\varepsilon\times(0,T)} \right| \\
&\leqslant\| \Psi_{\varepsilon}\| _{L^\infty L^2}
\| \nabla R_\varepsilon (\varphi)\| _{L^2 L^2}\\
&\leqslant\| {\mathbf{u}}_\varepsilon \otimes {\mathbf{u}}_\varepsilon\| _{L^1 L^2}
\| \nabla R_\varepsilon (\varphi)\| _{L^2 L^2}\\
&\leqslant C\| {\mathbf{u}}_\varepsilon \| _{L^2 L^6}^2(\| \nabla \varphi \| _{L^2 L^2}
+ \sigma_\varepsilon^{-1} \| \varphi \| _{L^2 L^2})\\
&=C\sigma_\varepsilon^2\| \nabla \varphi \| _{L^2 L^2}
+ C\sigma_\varepsilon\| \varphi \| _{L^2 L^2}.
\ea
and
\ba\label{est-P-large-2}
\left| \l \Delta {\mathbf{U}}_\varepsilon(t), R_\varepsilon (\varphi) \r_{\Omega_\varepsilon\times(0,T)} \right|&
\leqslant\| \nabla {\mathbf{U}}_\varepsilon(t)\| _{L^2 L^2}
\| \nabla R_\varepsilon (\varphi)\| _{L^2 L^2}\\
&\leqslant C{\sigma _\varepsilon }(\| \nabla \varphi \| _{L^2 L^2}
+ \sigma_\varepsilon^{-1} \| \varphi \| _{L^2 L^2})\\
&=C\sigma _\varepsilon\| \nabla \varphi \| _{L^2 L^2}
+ C\| \varphi \| _{L^2 L^2}.
\ea
By Poinca\'re inequality in Lemma \ref{lem-Poincare}, we obtain
\ba\label{est-P-large-3}
\left| \l \sigma^2_\varepsilon{{\mathbf{u}}}_\varepsilon, R_\varepsilon (\varphi) \r_{\Omega _\varepsilon\times(0,T)} \right|
&\leqslant C\sigma^2_\varepsilon \| {\mathbf{u}}_\varepsilon\| _{L^\infty(0,T; L^2(\Omega _\varepsilon))} \sigma_\varepsilon
\| \nabla R_\varepsilon (\varphi) \| _{L^2 L^2} \\
&\leqslant C(\sigma_\varepsilon^3\| \nabla \varphi\| _{L^2 L^2}+\sigma_\varepsilon^2\| \varphi\| _{L^2 L^2}),\\
\left| \l \sigma^2_\varepsilon{{\mathbf{u}}}^0_\varepsilon, R_\varepsilon (\varphi) \r_{\Omega _\varepsilon\times(0,T)} \right|
&\leqslant C\sigma^2_\varepsilon \| {\mathbf{u}}^0_\varepsilon\| _{L^2} \sigma_\varepsilon
\| \nabla R_\varepsilon (\varphi) \| _{L^2 L^2}\\
&\leqslant C(\sigma_\varepsilon^3\| \nabla \varphi\| _{L^2 L^2}+\sigma_\varepsilon^2\| \varphi\| _{L^2 L^2}),\\
\left| \l {\mathbf{F}}_\varepsilon (t), R_\varepsilon (\varphi) \r_{\Omega _\varepsilon\times(0,T)} \right|
&\leqslant C \sigma_\varepsilon\| {\mathbf{F}}_\varepsilon(t)\| _{L^2 L^2}\| \nabla R_\varepsilon (\varphi) \| _{L^2 L^2} \\
&\leqslant C(\sigma_\varepsilon\| \nabla \varphi\| _{L^2 L^2}+\| \varphi\| _{L^2 L^2}).
\ea
From \eqref{est-P-large-1}--\eqref{est-P-large-3} we deduce that
\ba\label{est-large-F}
\l \widetilde H_\varepsilon,\varphi \r_{\Omega\times(0,T)}
&\leqslant C (\| \varphi \| _{L^2 L^2}+ \sigma_\varepsilon\| \nabla \varphi\| _{L^2 L^2}).
\ea
This indicates $\widetilde H_{\varepsilon}$ is bounded in $L^{2}(0,T, L^{2}(\Omega)) + \sigma_{\varepsilon} L^{2}(0,T, W^{-1,2}(\Omega)) $.
On the other hand,
\ba
\l \partial_t \widetilde H_\varepsilon, \varphi \r_{\Omega\times(0,T)}&
=-\l \widetilde H_\varepsilon, \partial_t\varphi \r_{\Omega\times(0,T)}
=-\l \nabla P_\varepsilon, R_\varepsilon (\partial_t \varphi) \r_{\Omega_\varepsilon\times(0,T)}\\
&=-\l{\mathbf{F}}_\varepsilon(t)-\sigma^2_\varepsilon{\mathbf{u}}_\varepsilon(t)+\sigma^2_\varepsilon{\mathbf{u}}^0_\varepsilon + \mu \Delta {\mathbf{U}}_\varepsilon(t)
-{\Phi_\varepsilon}(t), R_\varepsilon (\partial_t \varphi) \r_{\Omega_\varepsilon\times(0,T)}\\
&=\l {\mathbf{f}}_\varepsilon, R_\varepsilon (\varphi) \r + \l\sigma^2_\varepsilon {\mathbf{u}}_\varepsilon, R_\varepsilon (\partial_t \varphi) \r
+\l \mu \Delta {\mathbf{u}}_\varepsilon, R_\varepsilon (\varphi) \r-\l {\mathrm{div}\,} ({\mathbf{u}}_\varepsilon\otimes{\mathbf{u}}_\varepsilon), R_\varepsilon (\varphi)\r.
\nn\ea
By the uniform estimates in \eqref{est-large-ue}--\eqref{3-estimate-U} and Lemma \ref{lem-Poincare}, we have
\ba
\left|\l {\mathbf{f}}_\varepsilon ,R_\varepsilon (\varphi) \r \right| & \leqslant C \| {\mathbf{f}}_\varepsilon\| _{L^2 L^2}\| R_\varepsilon (\varphi)\| _{L^2 L^2}
\leqslant C (\sigma_\varepsilon\| \nabla \varphi\| _{L^2 L^2}
+\| \varphi\| _{L^2 L^2}), \\
\left|\l\sigma^2_\varepsilon {\mathbf{u}}_\varepsilon, R_\varepsilon (\partial_t \varphi)\r\right|
& \leqslant C \sigma_\varepsilon^2 \sigma_\varepsilon\| \nabla R_\varepsilon (\partial_t\varphi)\| _{L^2 L^2}
\leqslant C\sigma_\varepsilon^2 (\sigma_\varepsilon\| \nabla \partial_t\varphi\| _{L^2 L^2}
+\| \partial_t\varphi\| _{L^2 L^2}),\\
| \l \mu \Delta {\mathbf{u}}_\varepsilon, R_\varepsilon (\varphi) \r| & \leqslant C\| \nabla {\mathbf{u}}_\varepsilon\| _{L^2 L^2}
\| \nabla R_\varepsilon (\varphi)\| _{L^2 L^2} \leqslant C (\sigma_\varepsilon\| \nabla \varphi\| _{L^2 L^2}
+\| \varphi\| _{L^2 L^2}),\\
| \l {\mathrm{div}\,} ({\mathbf{u}}_\varepsilon\otimes{\mathbf{u}}_\varepsilon), R_\varepsilon (\varphi)\r | &\leqslant C
\| \widetilde {\mathbf{u}}_\varepsilon \otimes\widetilde {\mathbf{u}}_\varepsilon\| _{L^{\frac43} L^2}
\| \nabla R_\varepsilon (\varphi)\| _{L^4 L^2} \leqslant C\sigma_\varepsilon^{\frac32}
(\| \nabla \varphi\| _{L^4 L^2}
+ \sigma_\varepsilon^{-1}\| \varphi\| _{L^4 L^2}).
\nn\ea
This shows that
\ba\nn
\left| \l \partial_t \widetilde H_\varepsilon, \varphi \r_{\Omega\times(0,T)}\right|
&\leqslant C (\| \varphi\| _{L^2 L^2}+\sigma_\varepsilon\| \nabla \varphi\| _{L^2 L^2}
+\sigma_\varepsilon^{\frac12}\| \varphi\| _{L^4 L^2}+\sigma_\varepsilon^{\frac32}\| \nabla \varphi\| _{L^4 L^2}\\
&\quad\quad\quad\quad\quad\quad\sigma_\varepsilon^2\| \partial_t\varphi\| _{L^2 L^2}+\sigma_\varepsilon^3\| \nabla \partial_t\varphi\| _{L^2 L^2})\\
&\leqslant C\| \varphi\| _{L^2 L^2}+C\sigma_\varepsilon^{\frac12}\| \nabla \varphi\| _{L^4 L^2}
+C\sigma_\varepsilon^2\| \nabla \partial_t\varphi\| _{L^2 L^2}.
\ea
Therefore, $\widetilde H_\varepsilon$ is uniformly bounded in $W^{1,2} L^2 +\sigma_\varepsilon^{\frac12} W^{1,\frac{4}{3}} W^{-1,2} + \sigma_\varepsilon^2 L^2 W^{-1,2}$. Moreover, the second property of the restriction operator in \eqref{pt-res} implies that
\ba
\l \widetilde H_\varepsilon, \varphi \r_{\Omega\times(0,T)}=0,\ \forall \varphi\in C_{c}^{\infty}(\Omega \times (0,T) ;{\mathbb{R}}^{3}), \ {\mathrm{div}\,} \varphi=0.
\nn\ea
Thus, we can apply Lemma \ref{lemma-large} and deduce that there exists $\widetilde P_\varepsilon \in W^{1,2} (W^{1,2}\cap L_{0}^{2}) + \sigma_\varepsilon^{\frac12} W^{1,\frac43} L^2_0+\sigma_\varepsilon^2 L^2 L^2_0 $ such that
\ba\label{large-F-P-1}
\widetilde H_\varepsilon =\widetilde H_\varepsilon^{(1)} + \sigma_\varepsilon^{\frac12} \widetilde H_\varepsilon^{(2)}+\sigma_\varepsilon^2 \widetilde H_\varepsilon^{(3)},\quad
\widetilde P_\varepsilon =\widetilde P_\varepsilon^{(1)} + \sigma_\varepsilon^{\frac12} \widetilde P_\varepsilon^{(2)}+\sigma_\varepsilon^2 \widetilde P_\varepsilon^{(3)},
\ea
with
\ba\label{large-F-P-2}
\widetilde H_\varepsilon^{(1)} =\nabla \widetilde P_\varepsilon^{(1)},\quad
\widetilde H_\varepsilon^{(2)} =\nabla \widetilde P_\varepsilon^{(2)},\quad
\widetilde H_\varepsilon^{(3)} =\nabla \widetilde P_\varepsilon^{(3)},
\ea
and
\ba\label{est-large-P-2}
& \| \widetilde P_\varepsilon^{(1)} \|_{W^{1,2} W^{1,2}}\leqslant \| \widetilde H_\varepsilon^{(1)} \|_{W^{1,2} L^2} \leqslant C,\\
& \| \widetilde P_\varepsilon^{(2)} \|_{W^{1,\frac43} L^2_0}\leqslant \| \widetilde H_\varepsilon^{(2)} \|_{W^{1,\frac43} W^{-1,2}} \leqslant C,\\
& \| \widetilde P_\varepsilon^{(3)} \|_{L^2 L^2_0}\leqslant \| \widetilde H_\varepsilon^{(3)} \|_{L^2 W^{-1,2}}\leqslant C.
\ea
\subsection{Passing to the limit}
Instead of showing the limit equations in ${\mathbf{u}}$ directly, we prove the following results in ${\mathbf{U}}$:
\begin{proposition}\label{prop-UP-large}
Let $({\mathbf{U}}_\varepsilon, P_\varepsilon)$ be the weak solution of equation \eqref{2-equa-large-holes} and $(\widetilde {\mathbf{U}}_\varepsilon, \widetilde P_\varepsilon)$ be their extension in $\Omega$ defined in \eqref{1-new-def}, \eqref{def-large-F}, \eqref{large-F-P-1} and \eqref{large-F-P-2}. If ${\lim _{\varepsilon \to 0}}{\sigma _\varepsilon } = 0~i.e. ~1<\alpha<3$, then
\ba\label{conv-large-U}
\sigma_\varepsilon^{-2}\widetilde {\mathbf{U}}_\varepsilon \to {\mathbf{U}}\ \mbox{weakly\ in} \ L^2 (0,T;L^2(\Omega;\mathbb{R}^3)),
\ea
and
\ba\label{conv-large-P-1}
\widetilde P_\varepsilon=\widetilde P_\varepsilon^{(1)}+\sigma_\varepsilon^{\frac{1}{2}}\widetilde P_\varepsilon^{(2)}+\sigma_\varepsilon^2 \widetilde P_\varepsilon^{(3)},
\ea
which satisfies the estimates in \eqref{est-large-P-2} and
\ba\label{conv-large-P-2}
\widetilde P_\varepsilon^{(1)} \to P\ \mbox{weakly\ in} \ W^{1,2}(0,T;W^{1,2}(\Omega)).
\ea
Moreover, $({\mathbf{U}},P)$ solves
\begin{equation}\label{equa-U-P-large}
\mu {\mathbf{U}} = A({\mathbf{F}}- \nabla P),\ \mbox{in}\ \mathcal{D}'(\Omega\times (0,T)),
\end{equation}
with
\ba\label{U-F}
{\mathbf{U}}(\cdot ,t) = \int_0^t {\mathbf{u}} (\cdot,s){\mathrm{d}} s, \quad {\mathbf{F}}(\cdot ,t) = \int_0^t {\mathbf{f}} (\cdot,s){\mathrm{d}} s.
\ea
\end{proposition}
\begin{proof}
The convergence of $\widetilde {\mathbf{U}}_{\varepsilon}$ in \eqref{conv-large-U} follows immediately from the uniform estimates \eqref{3-estimate-U} and was given already in \eqref{UUe-conv-large}. The weak convergence of $\widetilde P_{\varepsilon}$ in \eqref{conv-large-P-2} follows from the uniform estimates \eqref{est-large-P-2}. The relations in \eqref{U-F} follows from the uniqueness fo weak limits and were given already in \eqref{Fe-F-st} and \eqref{U-u-large}.
Given any scalar test function $\phi \in C_{c}^{\infty}(\Omega \times (0,T) )$, we choose $w ^i_{\eta,\varepsilon}\phi$ as a test function to \eqref{2-equa-large-holes} and obtain
\ba\label{weak-equa-large}
&\int_0^T \int_\Omega \sigma_\varepsilon^2\widetilde {\mathbf{u}}_\varepsilon \cdot (w_{\eta, \varepsilon}^i\phi)
-\widetilde \Psi _\varepsilon:\nabla (w_{\eta, \varepsilon}^i\phi)
+\mu \nabla {\mathbf{U}}_\varepsilon :\nabla (w_{\eta ,\varepsilon }^i\phi)
-\widetilde P_\varepsilon{\mathrm{div}\,} (w_{\eta, \varepsilon}^i\phi ){\mathrm{d}} x{\mathrm{d}} t \\
&=\int_0^T \int_\Omega {\mathbf{F}}_\varepsilon \cdot (w_{\eta,\varepsilon}^i\phi ){\mathrm{d}} x{\mathrm{d}} t+\sigma_\varepsilon^2 \int_0^T \int_\Omega \widetilde {\mathbf{u}}_\varepsilon^0 \cdot (w_{\eta,\varepsilon}^i\phi ) \,{\mathrm{d}} x{\mathrm{d}} t.
\ea
By the uniform estimates in \eqref{3-est-u} and \eqref{lem-cell-2}, and the convergences in \eqref{cell-3} and \eqref{Fe-F-st}, we have
\ba\label{conv-large-1}
\left| \sigma_\varepsilon^2\int_0^T \int_\Omega \widetilde {\mathbf{u}}_\varepsilon \cdot (w_{\eta,\varepsilon}^i\phi){\mathrm{d}} x{\mathrm{d}} t \right|
&\leqslant \sigma_\varepsilon^2\| \widetilde {\mathbf{u}}_\varepsilon \| _{L^2 L^6}\| w_{\eta, \varepsilon}^i\| _{L^2(\Omega)}
\| \phi \| _{L^2 L^3} \leqslant C\sigma _\varepsilon^3 \to 0,\\
\left| \sigma_\varepsilon^2\int_0^T \int_\Omega {\mathbf{u}}^0_\varepsilon \cdot (w_{\eta,\varepsilon}^i\phi ){\mathrm{d}} x{\mathrm{d}} t \right|
& \leqslant C \sigma_\varepsilon^2 ||{\mathbf{u}}^0||_{L^2(\Omega)} ||w_{\eta,\varepsilon}^i||_{L^2(\Omega)}\| \phi \| _{L^1 L^\infty} \leqslant C\sigma _\varepsilon^2 \to 0 \to 0,\\
\int_0^T \int_\Omega \widetilde {\mathbf{F}}_\varepsilon \cdot (w_{\eta,\varepsilon}^i\phi) {\mathrm{d}} x{\mathrm{d}} t
& \to \int_0^T \int_\Omega {\mathbf{F}} \cdot ({\overline w^i}\phi) {\mathrm{d}} x{\mathrm{d}} t.
\ea
For the convective term, by \eqref{lem-cell-2} and \eqref{3-est-u}, we have
\ba\label{conv-large-4}
&\left| \int_0^T \int_\Omega \widetilde \Psi _\varepsilon:\nabla (w_{\eta, \varepsilon}^i\phi) {\mathrm{d}} x{\mathrm{d}} t \right| \leqslant \| \widetilde \Psi_{\varepsilon} \|_{L^{\infty} L^{2}}\| w_{\eta, \varepsilon}^i\| _{W^{1,2}}\| \phi \| _{W^{1,\infty}}\\
& \leqslant \| \widetilde \uu_{\e}\otimes \widetilde \uu_{\e}\| _{L^1 L^2}\| w_{\eta, \varepsilon}^i\| _{W^{1,2}}\| \phi \| _{W^{1,\infty}}\leqslant C\| \widetilde \uu_{\e}\| _{L^2 L^6}^{2}\| w_{\eta, \varepsilon}^i\| _{W^{1,2}} \leqslant C \sigma_{\varepsilon} \to 0.
\ea
Using equation \eqref{cell-general} in $(w_{\eta,\varepsilon}^i,q_{\eta,\varepsilon}^i)$ in Section \ref{sec:cell} gives us
\ba\label{conv-large-2}
&\int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla (w_{\eta, \varepsilon}^i\phi ){\mathrm{d}} x{\mathrm{d}} t
= \int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla \phi \otimes w_{\eta,\varepsilon}^i
+ \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla w_{\eta,\varepsilon}^i\phi {\mathrm{d}} x{\mathrm{d}} t\\
&= \int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla \phi \otimes w_{\eta,\varepsilon}^i
+ \nabla(\phi \widetilde {\mathbf{U}}_\varepsilon ):\nabla w_{\eta,\varepsilon}^i
- \nabla \phi \otimes \widetilde {\mathbf{U}}_\varepsilon :\nabla w_{\eta,\varepsilon}^i {\mathrm{d}} x{\mathrm{d}} t\\
&= \int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla \phi \otimes w_{\eta,\varepsilon}^i
+ \varepsilon^{-1} {\mathrm{div}\,} (\phi \widetilde {\mathbf{U}}_\varepsilon) q_{\eta,\varepsilon}^i
+\sigma_\varepsilon^{-2}(\phi \widetilde {\mathbf{U}}_\varepsilon) \cdot e^i
- \nabla \phi \otimes \widetilde {\mathbf{U}}_\varepsilon :\nabla w_{\eta,\varepsilon}^i{\mathrm{d}} x{\mathrm{d}} t.
\ea
For the first term and the last term on the right-hand side of \eqref{conv-large-2}, we have
\ba\nonumber
\left| \int_0^T \int_\Omega \nabla \widetilde {\mathbf{U}}_\varepsilon :\nabla \phi \otimes w_{\eta,\varepsilon}^i{\mathrm{d}} x{\mathrm{d}} t \right|
& \leqslant \| \nabla \widetilde {\mathbf{U}}_\varepsilon \|_{L^2 L^2}
\| w_{\eta,\varepsilon}^i\| _{L^2 (\Omega)}
\| \nabla \phi \|_{L^2 L^\infty(\Omega)} \leqslant C{\sigma_\varepsilon} \to 0, \\
\left| \int_0^T \int_\Omega \nabla \phi \otimes \widetilde {\mathbf{U}}_\varepsilon :\nabla w_{\eta,\varepsilon}^i {\mathrm{d}} x{\mathrm{d}} t \right|&
\leqslant\| \widetilde {\mathbf{U}}_\varepsilon \| _{L^2 L^2}
\| \nabla w_{\eta,\varepsilon}^i\| _{L^2(\Omega)}\| \nabla \phi \| _{L^2 L^\infty(\Omega)}\leqslant C{\sigma_\varepsilon} \to 0.
\ea
Using ${\mathrm{div}\,} \widetilde {\mathbf{U}}_\varepsilon = 0$ and $\varepsilon^{-1} c_\eta= \sigma_\varepsilon^{-1}$ implies
\ba\label{conv-large-6}
\left| \int_0^T \int_\Omega \varepsilon ^{-1}{\mathrm{div}\,} (\phi \widetilde {\mathbf{U}}_\varepsilon )q_{\eta,\varepsilon}^i
{\mathrm{d}} x{\mathrm{d}} t \right| &\leqslant \varepsilon ^{-1}\| \nabla \phi \| _{L^2 L^\infty (\Omega)}\| \widetilde {\mathbf{U}}_\varepsilon \| _{L^2 L^2}\| q_{\eta,\varepsilon}^i\| _{L^2(\Omega)}\\
&\leqslant C \varepsilon ^{-1}\sigma _\varepsilon ^2 c_\eta = C\sigma _\varepsilon \to 0.
\ea
By the weak convergence of $\sigma_\varepsilon^{-2} \widetilde {\mathbf{U}}_\varepsilon$ in \eqref{UUe-conv-large}, we have
\ba\label{conv-large-7}
\int_0^T \int_\Omega \sigma_\varepsilon^{-2} \phi \widetilde {\mathbf{U}}_\varepsilon\cdot {e^i} {\mathrm{d}} x{\mathrm{d}} t \to \int_0^T \int_\Omega \phi {\mathbf{U}} \cdot {e^i} {\mathrm{d}} x{\mathrm{d}} t.
\ea
Now we deal with the term related to the pressure. By the decomposition in \eqref{large-F-P-1}--\eqref{est-large-P-2} we have
\ba\label{large-P-1}
& \int_0^T \int_\Omega \widetilde P_\varepsilon {\mathrm{div}\,} (w_{\eta, \varepsilon}^i \phi ) {\mathrm{d}} x{\mathrm{d}} t
=\int_0^T \int_\Omega \widetilde P_\varepsilon^{(1)} w_{\eta, \varepsilon}^i \cdot \nabla \phi {\mathrm{d}} x{\mathrm{d}} t \\
&\qquad +\sigma_\varepsilon^{\frac{1}{2}}\int_0^T \int_\Omega \widetilde P_\varepsilon^{(2)} w_{\eta, \varepsilon}^i \cdot \nabla \phi {\mathrm{d}} x{\mathrm{d}} t +\sigma_\varepsilon^2 \int_0^T \int_\Omega \widetilde P_\varepsilon^{(3)} w_{\eta, \varepsilon}^i \cdot \nabla \phi {\mathrm{d}} x{\mathrm{d}} t.
\ea
By the uniform estimates in \eqref{est-large-P-2}, there holds
\ba\label{large-P-2}
\sigma_\varepsilon^{\frac{1}{2}}\int_0^T \int_\Omega \widetilde P_\varepsilon^{(2)} w_{\eta, \varepsilon}^i \cdot \nabla \phi {\mathrm{d}} x{\mathrm{d}} t
\leqslant \sigma_\varepsilon^{\frac{1}{2}}\| \widetilde P_\varepsilon^{(2)}\| _{L^{\frac43} L^2}
\| w_{\eta, \varepsilon}^i\| _{L^2}\| \nabla \phi\| _{L^{4} L^\infty}
\leqslant C\sigma_\varepsilon^{\frac{1}{2}} \to 0, \\
\sigma_\varepsilon^2 \int_0^T \int_\Omega \widetilde P_\varepsilon^{(3)} w_{\eta, \varepsilon}^i \cdot \nabla \phi {\mathrm{d}} x{\mathrm{d}} t
\leqslant \sigma_\varepsilon^2\| \widetilde P_\varepsilon^{(3)}\| _{L^2 L^2}
\| w_{\eta, \varepsilon}^i\| _{L^2}\| \nabla \phi\| _{L^2 L^\infty}
\leqslant C\sigma_\varepsilon^2\to 0.
\ea
Since $\widetilde P_\varepsilon^{(1)}\to P$ weakly in $W^{1,2}(0,T;W^{1,2}(\Omega))$, by Sobolev compact embedding, one has $\widetilde P_\varepsilon^{(1)}\to P$ strongly in $L^2(0,T;L^2_0(\Omega))$. Thus,
\ba\label{large-P-3}
\int_0^T \int_\Omega \widetilde P_\varepsilon^{(1)} w_{\eta, \varepsilon}^i \cdot \nabla \phi {\mathrm{d}} x{\mathrm{d}} t
\to \int_0^T \int_\Omega P (\overline w^i \cdot \nabla \phi) {\mathrm{d}} x{\mathrm{d}} t
=\int_0^T \int_\Omega P {\mathrm{div}\,} (\overline w^i \phi ) {\mathrm{d}} x{\mathrm{d}} t.
\ea
From \eqref{large-P-1}--\eqref{large-P-3} we deduce that
\ba\label{conv-large-8}
\int_0^T \int_\Omega \widetilde P_\varepsilon {\mathrm{div}\,} (w_{\eta,\varepsilon}^i\phi ) {\mathrm{d}} x{\mathrm{d}} t
\to \int_0^T \int_\Omega P{{\mathrm{div}\,} }(\overline w ^i\phi ) {\mathrm{d}} x{\mathrm{d}} t.
\ea
By \eqref{conv-large-1}--\eqref{conv-large-8}, passing $\varepsilon\to 0$ in \eqref{weak-equa-large} implies our desired equation \eqref{equa-U-P-large}. The proof of Proposition \ref{prop-UP-large} is completed.
\end{proof}
\subsection{End of the proof}
By Proposition \ref{prop-UP-large}, differentiating in $t$ to equation \eqref{equa-U-P-large} implies
\ba\nn
\mu {\mathbf{u}} = A({{\mathbf{f}}} - \nabla p), \ \mbox{in} \ \mathcal{D}'(\Omega \times (0,T)),
\ea
with $p = \partial_{t} P$. By \eqref{est-large-P-2}, \eqref{conv-large-P-1} and \eqref{conv-large-P-2}, we know that
\ba\nn
\widetilde p_\varepsilon=\widetilde p_\varepsilon^{(1)}+\sigma_\varepsilon^{\frac12}\widetilde p_\varepsilon^{(2)}+\sigma_\varepsilon^2\widetilde p_\varepsilon^{(3)},
\ea
with
\ba\nn
&\widetilde p_\varepsilon^{(1)} \to p \ \mbox{weakly in} \ L^2 (0,T;W^{1,2}(\Omega)),\\
&\widetilde p_\varepsilon^{(2)} \ \mbox{bounded in} \ L^\frac{4}{3}(0,T;L_0^2(\Omega)),\\
&\widetilde p_\varepsilon^{(3)} \ \mbox{bounded in} \ W^{-1,2} (0,T;L_0^2(\Omega)),
\ea
which are exactly \eqref{thm2-p-1} and \eqref{thm2-p-2}.
Since $\widetilde {\mathbf{u}}_\varepsilon \in L^2(0,T;W_0^{1,2}(\Omega ))$, ${\mathrm{div}\,} \widetilde {\mathbf{u}}_\varepsilon=0$ and $ \sigma_\varepsilon^{-2} \widetilde {\mathbf{u}}_\varepsilon \to {\mathbf{u}}$ weakly in $L^2(0,T;L^2(\Omega))$, then ${\mathrm{div}\,} {\mathbf{u}}=0$ in $\Omega\times (0,T) $ and ${\mathbf{u}}\cdot \mathbf{n}=0$ on $\partial \Omega \times (0,T)$ (see \cite{ref20}). We thus complete the proof of Theorem \ref{thm-3}.
\section*{Acknowledgments}
Yong Lu has been supported by the Recruitment Program of Global Experts of China. Both authors are partially supported by the NSF of China under Grant 12171235. On behalf of all authors, the corresponding author states that there is no conflict of interest.
|
\section{Introduction}
In this paper we consider the numerical discretization of the two-dimensional
stationary Stokes problem by Crouzeix-Raviart elements. They were introduced
in the seminal paper \cite{CrouzeixRaviart} in 1973 by Crouzeix and Raviart
with the goal to obtain a stable and economic discretization of the Stokes
equation. They can be considered as an non-conforming enrichment of conforming
finite elements of polynomial degree $k$ for the velocity and discontinuous
pressures of degree $k-1$. It is well known that the conforming $\left(
k,k-1\right) $ pair of finite elements can be unstable; for two-dimensions
the proof of the inf-sup stability of Crouzeix-Raviart discretizations of
general order $k$ has been evolved over the last 50 years, the inf-sup
stability for $k=1$ has been proved in \cite{CrouzeixRaviart} and only
recently the last open case $k=3$, has been proved in \cite{CCSS_CR_2}. We
mention the papers \cite{Fortin_Soulie}, \cite{ScottVogelius},
\cite{Crouzeix_Falk}, \cite{Baran_Stoyan}, \cite{GuzmanScott2019},
\cite{CCSS_CR_1} which contain essential milestones in this development.
Since higher order methods are becoming increasingly popular a natural
question arises how the inf-sup constant depends on the polynomial degree $k$.
It is the goal of this paper to investigate this dependency.
The paper is organized as follows. In Section \ref{NuDiscrete} we introduce
the Stokes problem and the Crouzeix-Raviart discretization of polynomial order
$k$. We state our main theorem that the discrete inf-sup constant can be
estimated from below by $ck^{-\alpha}$, where the positive constant $c$
depends only on the shape-regularity of the mesh. The explicit value of
$\alpha$ depends on the mesh topology. The simplest case is that each triangle
in triangulation contains at least one inner vertex and then $\alpha=2.25$
holds. We will give the value of $\alpha$ also for more general triangulations.
The proof is given in Section \ref{SecProof}. The key result is that for any
discrete pressure $q$, there exists a velocity $\mathbf{v}$ from the
Crouzeix-Raviart space such that $q-\operatorname{div}_{\mathcal{T}
\mathbf{v}$ belongs to a certain subspace (see (\ref{defMSV})) of the
\textit{Scott-Vogelius pressure space} \cite[R.1, R.2]{vogelius1983right},
\cite[R.1, R.2]{ScottVogelius} and $\mathbf{v}$ depends continuously on $q$.
The subspace property allows us to conclude from \cite[R.1, R.2
{ScottVogelius}, \cite{GuzmanScott2019} that the divergence operator has a
right-inverse in the space of continuous velocity approximations of order $k$
while it is chosen in such a way that, in contrast to the setting in
\cite[Thm. 5.1]{ScottVogelius}, \cite[Thm. 1]{GuzmanScott2019}, the norm of
the right-inverse does not deteriorate if a triangle vertex is a
\textquotedblleft nearly-critical\textquotedblright\ point -- a notion which
will be introduced in Definition \ref{Defetacriticalpoint}. This key result is
proved for odd polynomial degree in Section \ref{PTpge5} and for even
polynomial degree in Section \ref{SecEvenk}.
In the conclusions (Sec. \ref{SecConcl}) we summarize our main findings and
compare our results with existing results in the literature for other pairs of
finite elements for the Stokes equation.
In the appendix, we prove a technical result for a Gram matrix related to the
bilinear form $\left( \operatorname*{div}_{\mathcal{T}}\cdot,\cdot\right)
_{L^{2}\left( \Omega\right) }$ applied to the Crouzeix-Raviart element and
discontinuous pressure space.
\subsection{The Stokes problem and its numerical
discretization\label{NuDiscrete}}
Let $\Omega\subset\mathbb{R}^{2}$ denote a bounded polygonal Lipschitz domain
with boundary $\partial\Omega$. We consider the Stokes equatio
\
\begin{array}
[c]{llll
-\Delta\mathbf{u} & -\nabla p & =\mathbf{f} & \text{in }\Omega,\\
\operatorname*{div}\mathbf{u} & & =0 & \text{in }\Omega
\end{array}
\]
with Dirichlet boundary conditions for the velocity and a normalization
condition for the pressur
\[
\mathbf{u}=\mathbf{0}\quad\text{on }\partial\Omega\quad\text{and\quad
\int_{\Omega}p=0.
\]
To formulate this equation in a variational form we first introduce the
relevant function spaces. Throughout the paper we restrict to vector spaces
over the field of real numbers.
For $s\geq0$, $1\leq p\leq\infty$, $W^{s,p}\left( \Omega\right) $ denotes
the classical Sobolev spaces of functions with norm $\left\Vert \cdot
\right\Vert _{W^{s,p}\left( \Omega\right) }$. As usual we write
$L^{p}\left( \Omega\right) $ instead of $W^{0,p}\left( \Omega\right) $ and
$H^{s}\left( \Omega\right) $ for $W^{s,2}\left( \Omega\right) $. For
$s\geq0$, we denote by $H_{0}^{s}\left( \Omega\right) $ the closure of the
space of infinitely smooth functions with compact support in $\Omega$ with
respect to the $H^{s}\left( \Omega\right) $ norm. Its dual space is denoted
by $H^{-s}\left( \Omega\right) $.
The scalar product and norm in $L^{2}\left( \Omega\right) $ are denoted
respectively b
\
\begin{array}
[c]{llll
\left( u,v\right) _{L^{2}\left( \Omega\right) }:=\int_{\Omega}uv &
\text{and} & \left\Vert u\right\Vert _{L^{2}\left( \Omega\right) }:=\left(
u,u\right) _{L^{2}\left( \Omega\right) }^{1/2} & \text{in }L^{2}\left(
\Omega\right) .
\end{array}
\]
Vector-valued and $2\times2$ tensor-valued analogues of the function spaces
are denoted by bold and blackboard bold letters, e.g., $\mathbf{H}^{s}\left(
\Omega\right) =\left( H^{s}\left( \Omega\right) \right) ^{2}$ and
$\mathbb{H}^{s}=\left( H^{s}\left( \Omega\right) \right) ^{2\times2}$ and
analogously for other quantities.
The $\mathbf{L}^{2}\left( \Omega\right) $ scalar product and norm for vector
valued functions are given b
\[
\left( \mathbf{u},\mathbf{v}\right) _{\mathbf{L}^{2}\left( \Omega\right)
}:=\int_{\Omega}\left\langle \mathbf{u},\mathbf{v}\right\rangle \quad
\text{and\quad}\left\Vert \mathbf{u}\right\Vert _{\mathbf{L}^{2}\left(
\Omega\right) }:=\left( \mathbf{u},\mathbf{u}\right) _{\mathbf{L
^{2}\left( \Omega\right) }^{1/2},
\]
where $\left\langle \mathbf{u},\mathbf{v}\right\rangle $ denotes the Euclidean
scalar product in $\mathbb{R}^{3}$. In a similar fashion, we define for
$\mathbf{G},\mathbf{H}\in\mathbb{L}^{2}\left( \Omega\right) $ the scalar
product and norm b
\[
\left( \mathbf{G},\mathbf{H}\right) _{\mathbb{L}^{2}\left( \Omega\right)
}:=\int_{\Omega}\left\langle \mathbf{G},\mathbf{H}\right\rangle \quad
\text{and\quad}\left\Vert \mathbf{G}\right\Vert _{\mathbb{L}^{2}\left(
\Omega\right) }:=\left( \mathbf{G},\mathbf{G}\right) _{\mathbb{L
^{2}\left( \Omega\right) }^{1/2},
\]
where $\left\langle \mathbf{G},\mathbf{H}\right\rangle =\sum_{i,j=1
^{3}G_{i,j}H_{i,j}$. Finally, let $L_{0}^{2}\left( \Omega\right) :=\left\{
u\in L^{2}\left( \Omega\right) :\int_{\Omega}u=0\right\} $.
We introduce the bilinear form $a:\mathbf{H}^{1}\left( \Omega\right)
\times\mathbf{H}^{1}\left( \Omega\right) \rightarrow\mathbb{R}$ b
\begin{equation}
a\left( \mathbf{u},\mathbf{v}\right) :=\left( \nabla\mathbf{u
,\nabla\mathbf{v}\right) _{\mathbb{L}^{2}\left( \Omega\right) },
\label{defabili
\end{equation}
where $\nabla\mathbf{u}$ and $\nabla\mathbf{v}$ denote the derivatives of
$\mathbf{u}$ and $\mathbf{v}$. The variational form of the Stokes problem is
given by: For given $\mathbf{F}\in\mathbf{H}^{-1}\left( \Omega\right) ,
\begin{equation}
\text{find }\left( \mathbf{u},p\right) \in\mathbf{H}_{0}^{1}\left(
\Omega\right) \times L_{0}^{2}\left( \Omega\right) \;\text{s.t.\ }\left\{
\begin{array}
[c]{lll
a\left( \mathbf{u},\mathbf{v}\right) +\left( p,\operatorname*{div
\mathbf{v}\right) _{L^{2}\left( \Omega\right) } & =\mathbf{F}\left(
\mathbf{v}\right) & \forall\mathbf{v}\in\mathbf{H}_{0}^{1}\left(
\Omega\right) ,\\
\left( \operatorname*{div}\mathbf{u},q\right) _{L^{2}\left( \Omega\right)
} & =0 & \forall q\in L_{0}^{2}\left( \Omega\right) .
\end{array}
\right. \label{varproblemstokes
\end{equation}
It is well-known (see, e.g., \cite{Girault86}) that (\ref{varproblemstokes})
is well posed. Since we consider non-conforming discretizations we restrict
the space $\mathbf{H}^{-1}\left( \Omega\right) $ for the right-hand side to
a smaller space and assume from now on for simplicity that $\mathbf{F}\left(
\mathbf{v}\right) =\left( \mathbf{f},\mathbf{v}\right) _{\mathbf{L
^{2}\left( \Omega\right) }$ for some $\mathbf{f}\in\mathbf{L}^{2}\left(
\Omega\right) $; for a more general setting we refer to
\cite{Veeser_qo_part1}, \cite{Veeser_qo}).\bigskip
In the following a discretization for problem (\ref{varproblemstokes}) is
introduced. Let $\mathcal{T}=\left\{ K_{i}:1\leq i\leq n\right\} $ denote a
triangulation of $\Omega$ consisting of closed triangles which are
\textit{conforming}: the intersection of two different triangles is either
empty, a common edge, or a common point. We also assume $\Omega
=\operatorname*{dom}\mathcal{T}$, where
\begin{equation}
\operatorname*{dom}\mathcal{T}:=\operatorname*{int}\left(
{\displaystyle\bigcup\limits_{K\in\mathcal{T}}}
K\right) \label{defdomT
\end{equation}
and $\operatorname*{int}\left( M\right) $ denotes the interior of a set
$M\subset\mathbb{R}^{2}$. An important measure for the quality of a finite
element triangulation is the shape-regularity constant, which we define b
\begin{equation}
\gamma_{\mathcal{T}}:=\max_{K\in\mathcal{T}}\frac{h_{K}}{\rho_{K}}
\label{defgammat
\end{equation}
with the local mesh width $h_{K}:=\operatorname*{diam}K$ and $\rho_{K}$
denoting the diameter of the largest inscribed ball in $K$. The global mesh
width is given by $h_{\mathcal{T}}:=\max\left\{ h_{K}:K\in\mathcal{T
\right\} $.
Piecewise versions of differential operators\ such as $\nabla$ and
$\operatorname*{div}$ are defined for functions $u$ and vector fields
$\mathbf{w}$ which are sufficiently smooth in the interior of the triangles
$K\subset\mathcal{T}$ b
\[
\left. \left( \nabla_{\mathcal{T}}u\right) \right\vert _{\overset{\circ
}{K}}:=\nabla\left( \left. u\right\vert _{\overset{\circ}{K}}\right)
\quad\text{and\quad}\left. \left( \operatorname*{div}\nolimits_{\mathcal{T
}\mathbf{w}\right) \right\vert _{\overset{\circ}{K}}:=\operatorname*{div
\left( \left. \mathbf{w}\right\vert _{\overset{\circ}{K}}\right) .
\]
The values on $\partial K$ are arbitrary since $\partial K$ has measure zero.
\begin{remark}
\label{Remangle}It is well known that the shape-regularity constant implies
that there exists some $\phi_{\mathcal{T}}>0$ depending only on $\gamma
_{\mathcal{T}}$ such that every triangle angle in $\mathcal{T}$ is bounded
from below by $\phi_{\mathcal{T}}$. In turn, every triangle angle in
$\mathcal{T}$ is bounded from above by $\pi-2\phi_{\mathcal{T}}$.
\end{remark}
The set of edges in $\mathcal{T}$ is denoted by $\mathcal{E}\left(
\mathcal{T}\right) $, while the subset of boundary edges is $\mathcal{E
_{\partial\Omega}\left( \mathcal{T}\right) :=\left\{ E\in\mathcal{E}\left(
\mathcal{T}\right) :E\subset\partial\left( \operatorname*{dom
\mathcal{T}\right) \right\} $; the subset of inner edges is given by
$\mathcal{E}_{\Omega}\left( \mathcal{T}\right) :=\mathcal{E}\left(
\mathcal{T}\right) \backslash\mathcal{E}_{\partial\Omega}\left(
\mathcal{T}\right) $. For each edge $E\in\mathcal{E}$ we fix a unit vector
$\mathbf{n}_{E}$ orthogonal to $E$ with the convention that $\mathbf{n}_{E}$
is the outer normal vector for boundary edges $E\in\mathcal{E}_{\partial
\Omega}$.
The set of triangle vertices in $\mathcal{T}$ is denoted by $\mathcal{V
\left( \mathcal{T}\right) $, while the subset of inner vertices is
$\mathcal{V}_{\Omega}\left( \mathcal{T}\right) :=\left\{ \mathbf{V
\in\mathcal{V}\left( \mathcal{T}\right) :\mathbf{V}\notin\partial\left(
\operatorname*{dom}\mathcal{T}\right) \right\} $ and $\mathcal{V
_{\partial\Omega}\left( \mathcal{T}\right) :=\mathcal{V}\left(
\mathcal{T}\right) \backslash\mathcal{V}_{\Omega}\left( \mathcal{T}\right)
$. For $K\in\mathcal{T}$, the set of its vertices is denoted by $\mathcal{V
\left( K\right) $. For $E\in\mathcal{E}\left( \mathcal{T}\right) $, we
define the edge patch b
\[
\mathcal{T}_{E}:=\left\{ K\in\mathcal{T}:E\subset K\right\} \quad
\text{and\quad}\omega_{E}:
{\displaystyle\bigcup\limits_{K\in\mathcal{T}_{E}}}
K.
\]
For $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) $, the nodal patch is
defined b
\begin{equation}
\mathcal{T}_{\mathbf{z}}:=\left\{ K\in\mathcal{T}:\mathbf{z}\in K\right\}
\quad\text{and\quad}\omega_{\mathbf{z}}:
{\displaystyle\bigcup\limits_{K\in\mathcal{T}_{\mathbf{z}}}}
K \label{nodalpatch
\end{equation}
with local mesh width $h_{\mathbf{z}}:=\max\left\{ h_{K}:K\in\mathcal{T
_{\mathbf{z}}\right\} $. For $K\in\mathcal{T}$, we set
\begin{equation}
\mathcal{T}_{K}:=\left\{ K^{\prime}\in\mathcal{T}\mid K\cap K^{\prime
\neq\emptyset\right\} \quad\text{and\quad}\omega_{K}:
{\displaystyle\bigcup\limits_{K^{\prime}\in\mathcal{T}_{K}}}
K^{\prime}. \label{trianglepatch
\end{equation}
For a subset $M\subset\mathbb{R}^{2}$, we denote by $\left[ M\right] $ its
convex hull; in particular, any edge $E$ with endpoints $\mathbf{a
,\mathbf{b}$ can be written as $E=\left[ \mathbf{a},\mathbf{b}\right]
=\left[ \mathbf{b},\mathbf{a}\right] $.
Let $\mathbb{N}=\left\{ 1,2,\ldots\right\} $ and $\mathbb{N}_{0
:=\mathbb{N\cup}\left\{ 0\right\} $. For $m\in\mathbb{N}$, we employ the
usual multiindex notation for
\mbox{\boldmath$ \mu$
=\left( \mu_{i}\right) _{i=1}^{m}\in\mathbb{N}_{0}^{m}$ and points
$\mathbf{x}=\left( x_{i}\right) _{i=1}^{m}\in\mathbb{R}^{m}
\[
\left\vert
\mbox{\boldmath$ \mu$
\right\vert :=\mu_{1}+\ldots+\mu_{m},\quad\mathbf{x}^
\mbox{\boldmath$ \mu$
}:
{\displaystyle\prod\limits_{j=1}^{m}}
x_{j}^{\mu_{j}}.
\]
Let $\mathbb{P}_{m,k}$ denote the space of $m$-variate polynomials of maximal
degree $k$, consisting of functions of the for
\[
\sum_{\substack
\mbox{\boldmath$ \mu$
\in\mathbb{N}_{0}^{m}\\\left\vert
\mbox{\boldmath$ \mu$
\right\vert \leq k}}a_
\mbox{\boldmath$ \mu$
}\mathbf{x}^
\mbox{\boldmath$ \mu$
\]
for real coefficients $a_
\mbox{\boldmath$ \mu$
}$. Formally, we set $\mathbb{P}_{m,-1}:=\left\{ 0\right\} $. To indicate
the domain explicitly in notation we write sometimes $\mathbb{P}_{k}\left(
D\right) $ for $D\subset\mathbb{R}^{m}$ and skip the index $m$ since it is
then clear from the argument $D$.
We introduce the following finite element space
\begin{equation
\begin{array}
[c]{ll}
& \mathbb{P}_{k}\left( \mathcal{T}\right) :=\left\{ q\in L^{2}\left(
\Omega\right) \mid\forall K\in\mathcal{T}:\left. q\right\vert
_{\overset{\circ}{K}}\in\mathbb{P}_{k}\left( \overset{\circ}{K}\right)
\right\} ,\\
\text{and (cf. (\ref{defdomT}))} & \mathbb{P}_{k,0}\left( \mathcal{T}\right)
:=\left\{ q\in\mathbb{P}_{k}\left( \mathcal{T}\right) :\in
_{\operatorname*{dom}\mathcal{T}}q=0\right\} .
\end{array}
\label{Pkdefs
\end{equation}
Furthermore, le
\
\begin{array}
[c]{ll}
& S_{k}\left( \mathcal{T}\right) :=\mathbb{P}_{k}\left( \mathcal{T}\right)
\cap H^{1}\left( \operatorname*{dom}\mathcal{T}\right) ,\\
\text{and} & S_{k,0}\left( \mathcal{T}\right) :=S_{k}\left( \mathcal{T
\right) \cap H_{0}^{1}\left( \operatorname*{dom}\mathcal{T}\right) .
\end{array}
\]
The vector-valued versions are denoted by $\mathbf{S}_{k}\left(
\mathcal{T}\right) :=S_{k}\left( \mathcal{T}\right) ^{2}$ and
$\mathbf{S}_{k,0}\left( \mathcal{T}\right) :=S_{k,0}\left( \mathcal{T
\right) ^{2}$. Finally, we define the Crouzeix-Raviart space b
\begin{subequations}
\label{CRdeffull}
\end{subequations
\begin{align}
\operatorname*{CR}\nolimits_{k}\left( \mathcal{T}\right) & :=\left\{
v\in\mathbb{P}_{k}\left( \mathcal{T}\right) \mid\forall q\in\mathbb{P
_{k-1}\left( E\right) \quad\forall E\in\mathcal{E}_{\Omega}\left(
\mathcal{T}\right) \quad\int_{E}\left[ v\right] _{E}q=0\right\} ,\tag
\ref{CRdeffull
a}\label{CRdeffulla}\\
\operatorname*{CR}\nolimits_{k,0}\left( \mathcal{T}\right) & :=\left\{
v\in\operatorname*{CR}\nolimits_{k}\left( \mathcal{T}\right) \mid\forall
q\in\mathbb{P}_{k-1}\left( E\right) \quad\forall E\in\mathcal{E
_{\partial\Omega}\left( \mathcal{T}\right) \quad\int_{E}vq=0\right\} .
\tag
\ref{CRdeffull
b}\label{PCR0RB
\end{align}
Here, $\left[ v\right] _{E}$ denotes the jump of $v\in\mathbb{P}_{k}\left(
\mathcal{T}\right) $ across an edge $E\in\mathcal{E}_{\Omega}\left(
\mathcal{T}\right) $ and $\mathbb{P}_{k-1}\left( E\right) $ is the space of
polynomials of maximal degree $k-1$ with respect to the local variable in $E$.
We have collected all ingredients for defining the Crouzeix-Raviart
discretization for the Stokes equation. For $k\in\mathbb{N}$, let the discrete
velocity space and pressure space be defined by
\[
\mathbf{CR}_{k,0}\left( \mathcal{T}\right) :=\left( \operatorname*{CR
\nolimits_{k,0}\left( \mathcal{T}\right) \right) ^{2}\quad\text{and\quad
}M_{k-1}\left( \mathcal{T}\right) :=\mathbb{P}_{k-1,0}\left( \mathcal{T
\right) .
\]
Then, the discretization is given by: find $\left( \mathbf{u
_{\operatorname*{CR}},p_{\operatorname*{disc}}\right) \in\mathbf{CR
_{k,0}\left( \mathcal{T}\right) \times M_{k-1}\left( \mathcal{T}\right)
\;$such tha
\begin{equation}
\left\{
\begin{array}
[c]{lll
a_{\mathcal{T}}\left( \mathbf{u}_{\operatorname*{CR}},\mathbf{v}\right)
-b_{\mathcal{T}}\left( \mathbf{v},p_{\operatorname*{disc}}\right) & =\left(
\mathbf{f},\mathbf{v}\right) _{\mathbf{L}^{2}\left( \Omega\right) } &
\forall\mathbf{v}\in\mathbf{CR}_{k,0}\left( \mathcal{T}\right) ,\\
b_{\mathcal{T}}\left( \mathbf{u}_{\operatorname*{CR}},q\right) & =0 &
\forall q\in M_{k-1}\left( \mathcal{T}\right) ,
\end{array}
\right. \label{discrStokes
\end{equation}
where the bilinear forms $a_{\mathcal{T}}:\mathbf{CR}_{k,0}\left(
\mathcal{T}\right) \times\mathbf{CR}_{k,0}\left( \mathcal{T}\right)
\rightarrow\mathbb{R}$ and $b_{\mathcal{T}}:\mathbf{CR}_{k,0}\left(
\mathcal{T}\right) \times M_{k-1}\left( \mathcal{T}\right) \rightarrow
\mathbb{R}$ are given b
\[
a_{\mathcal{T}}\left( \mathbf{u},\mathbf{v}\right) :=\left( \nabla
_{\mathcal{T}}\mathbf{u},\nabla_{\mathcal{T}}\mathbf{v}\right) _{\mathbb{L
^{2}\left( \Omega\right) }\quad\text{and\quad}b_{\mathcal{T}}\left(
\mathbf{v},q\right) :=\left( \operatorname*{div}\nolimits_{\mathcal{T
}\mathbf{v},q\right) _{L^{2}\left( \Omega\right) }.
\]
It is well known that problem (\ref{discrStokes}) is well-posed if the
bilinear form $a_{\mathcal{T}}\left( \cdot,\cdot\right) $ is coercive and
$b_{\mathcal{T}}\left( \cdot,\cdot\right) $ satisfies the inf-sup condition.
To verify the first condition we introduce, for a conforming triangulation
$\mathcal{T}$ of the domain $\Omega$, the \textit{broken Sobolev space
\[
H^{1}\left( \mathcal{T}\right) :=\left\{ u\in L^{2}\left( \Omega\right)
\mid\forall K\in\mathcal{T}:\left. u\right\vert _{K}\in H^{1}\left(
K\right) \right\}
\]
and define, for $u\in H^{1}\left( \mathcal{T}\right) $, the \textit{broken
}$H^{1}$\textit{--seminorm} b
\[
\left\Vert u\right\Vert _{H^{1}\left( \mathcal{T}\right) }:=\left\Vert
\nabla_{\mathcal{T}}u\right\Vert _{\mathbf{L}^{2}\left( \Omega\right)
}=\left( \sum_{K\in\mathcal{T}}\left\Vert \nabla u\right\Vert _{\mathbf{L
^{2}\left( K\right) }^{2}\right) ^{1/2}.
\]
In \cite[Lem. 2]{CrouzeixRaviart}) it is proved that $\left\Vert
\cdot\right\Vert _{\mathbf{H}^{1}\left( \mathcal{T}\right) }$ defines a norm
in $\mathbf{CR}_{k,0}\left( \mathcal{T}\right) +\mathbf{H}_{0}^{1}\left(
\Omega\right) $ which is equivalent to the norm $\left( \sum_{K\in
\mathcal{T}}\left\Vert \mathbf{u}\right\Vert _{\mathbf{H}^{1}\left( K\right)
}^{2}\right) ^{1/2}$ with equivalence constants independent of the polynomial
degree and the mesh width (see Theorem \ref{TheoDiscFried}). This directly
implies the coercivity of $a_{\mathcal{T}}\left( \cdot,\cdot\right) $
\[
a_{\mathcal{T}}\left( \mathbf{u},\mathbf{u}\right) \geq\left\Vert
\mathbf{u}\right\Vert _{\mathbf{H}^{1}\left( \mathcal{T}\right) }^{2
\quad\forall\mathbf{u}\in\mathbf{CR}_{k,0}\left( \mathcal{T}\right) .
\]
Hence, well-posedness of (\ref{discrStokes}) follows from the inf-sup
condition for $b_{\mathcal{T}}\left( \cdot,\cdot\right) $.
\begin{definition}
Let $\mathcal{T}$ denote a conforming triangulation for $\Omega$. The pair
$\mathbf{CR}_{k,0}\left( \mathcal{T}\right) \times M_{k-1}\left(
\mathcal{T}\right) $ is \emph{inf-sup stable} if there exists a constant
$c_{\mathcal{T},k}$ such tha
\begin{equation}
\inf_{p\in M_{k-1}\left( \mathcal{T}\right) \backslash\left\{ 0\right\}
}\sup_{\mathbf{v}\in\mathbf{CR}_{k,0}\left( \mathcal{T}\right)
\backslash\left\{ 0\right\} }\frac{\left( p,\operatorname*{div
_{\mathcal{T}}\mathbf{v}\right) _{L^{2}\left( \Omega\right) }}{\left\Vert
\mathbf{v}\right\Vert _{\mathbf{H}^{1}\left( \mathcal{T}\right) }\left\Vert
p\right\Vert _{L^{2}\left( \Omega\right) }}\geq c_{\mathcal{T},k}>0.
\label{infsupcond
\end{equation}
\end{definition}
We are now in the position to formulate our main theorem.
\begin{theorem}
\label{Theomain}Let $\Omega\subset\mathbb{R}^{2}$ be a bounded polygonal
Lipschitz domain and let $\mathcal{T}$ denote a conforming triangulation of
$\Omega$ consisting of more than a single triangle. Let $k\in\mathbb{N}$. If
$k\geq3$ is odd we assume that $\mathcal{T}$ contains at least one inner
vertex. Then, the inf-sup condition (\ref{infsupcond}) hold
\begin{equation}
c_{\mathcal{T},k}\geq c_{\mathcal{T}}k^{-\alpha} \label{infsupp_dep
\end{equation}
for a constant $c_{\mathcal{T}}>0$ depending only on the shape-regularity of
the mesh. In particular $c_{\mathcal{T}}$ is independent of the mesh width
$h_{\mathcal{T}}$ and the polynomial degree $k$. The value of $\alpha\geq0$
depends on the mesh topology
\[
\alpha=\left\{
\begin{array}
[c]{ll
1/4+\kappa & \left\{
\begin{array}
[c]{l
\text{if }k\text{ is even,}\\
\text{or }k\geq3\text{ is odd and all triangles in }\mathcal{T}\text{ have at
least one inner vertex,
\end{array}
\right. \\
\left( 1/4+\kappa\right) \left( 1+L\right) & \text{otherwise,
\end{array}
\right.
\]
for $\kappa=2$ and $L$ denotes the number of steps involved in the
step-by-step construction introduced in (\ref{stepbystepextension}).
\end{theorem}
\proof
The estimate $c_{\mathcal{T},k}>0$ follows, for $k=1$ from
\cite{CrouzeixRaviart}, for $k=2$ from \cite[Thm. 3.1]{ChaLeeLee}, for even
$k\geq4$ from \cite{Baran_Stoyan}, for odd $k\geq5$ from \cite{CCSS_CR_1}, and
for $k=3$ from \cite{CCSS_CR_2}. We se
\[
c_{\mathcal{T},\operatorname*{low}}:=\min\left\{ c_{\mathcal{T},j}:1\leq
j\leq3\right\} .
\]
Estimate (\ref{infsupp_dep}) for some $c_{\mathcal{T}}:=c_{\mathcal{T
,\operatorname*{odd},\operatorname*{high}}>0$ for odd $k\geq5$ is proved in
Section \ref{PTpge5}, Lem. \ref{LemmaFinOdd} while this estimate for some
$c_{\mathcal{T}}:=c_{\mathcal{T},\operatorname*{even},\operatorname*{high}}>0$
for even $k\geq4$ is proved in Section \ref{SecEvenk}. Both constants
$c_{\mathcal{T},\operatorname*{odd},\operatorname*{high}}$, $c_{\mathcal{T
,\operatorname*{even},\operatorname*{high}}$ depend only on the
shape-regularity of the mesh. Hence, $c_{\mathcal{T}}\geq\min\left\{
c_{\mathcal{T},\operatorname*{low}},c_{\mathcal{T},\operatorname*{odd
,\operatorname*{high}},c_{\mathcal{T},\operatorname*{even
,\operatorname*{high}}\right\} $
\endproof
We emphasize that the original definition in \cite{CrouzeixRaviart} allows for
slightly more general finite element spaces, more precisely, the spaces
$\operatorname*{CR}_{k}\left( \mathcal{T}\right) $ can be enriched by
locally supported functions. From this point of view, the definition
(\ref{CRdeffull}) describes the \textit{natural} Crouzeix-Raviart space. The
possibility for enrichment has been used frequently in the literature to prove
inf-sup stability for the arising finite element spaces (see, e.g.,
\cite{CrouzeixRaviart}, \cite{Guzman_divfree}, \cite{Matthies_nonconf_2005}).
In contrast, we will prove the $k$-explicit estimate for the inf-sup constant
for the natural Crouzeix-Raviart family.
\section{Proof of Theorem \ref{Theomain}\label{SecProof}}
In this section, we will analyse the $k$-dependence of the inf-sup constant in
the form (\ref{infsupp_dep}), first for odd polynomial degree $k\geq5$ and
then for even degree $k\geq4$.
\subsection{Barycentric coordinates and basis functions for the
velocity\label{VelBasis}}
In this section, we introduce basis functions for the finite element spaces in
Section \ref{NuDiscrete}. We begin with introducing some general notation.
\begin{notation}
\label{Notation}For vectors $\mathbf{a}_{i}\in\mathbb{R}^{n}$, $1\leq i\leq
m$, we write $\left[ \mathbf{a}_{1}\mid\mathbf{a}_{2}\mid\ldots\mid
\mathbf{a}_{m}\right] $ for the $n\times m$ matrix with column vectors
$\mathbf{a}_{i}$. For $\mathbf{v}=\binom{v_{1}}{v_{2}}\in\mathbb{R}^{2}$ we
set $\mathbf{v}^{\perp}:=\left( v_{2},-v_{1}\right) ^{T}$. Let\ $\mathbf{e
_{k,i}\in\mathbb{R}^{k}$\ be the $i$-th canonical unit vector in
$\mathbb{R}^{k}$.
For $\mathbf{v}\in\mathbb{R}^{n}$, $\left\Vert \mathbf{v}\right\Vert $ is the
Euclidean vector norm and for a matrix $\mathbf{B}\in\mathbb{R}^{n\times n}$,
$\left\Vert \mathbf{B}\right\Vert :=\sup\left\{ \left\Vert \mathbf{Bx
\right\Vert /\left\Vert \mathbf{x}\right\Vert :\mathbf{x}\in\mathbb{R
^{n}\backslash\left\{ \mathbf{0}\right\} \right\} $ is the induced matrix norm.
Vertices in a triangle are always numbered counterclockwise. In a triangle $K$
with vertices $\mathbf{A}_{1}$, $\mathbf{A}_{2}$, $\mathbf{A}_{3}$ the angle
at $\mathbf{A}_{i}$ is called $\alpha_{i}$. If a triangle is numbered by an
index (e.g., $K_{\ell}$), the angle at $A_{\ell,i}$ is called $\alpha_{\ell
,i}$. For quantities in a triangle $K$ as, e.g., angles $\alpha_{j}$, $1\leq
j\leq3$, we use the cyclic numbering convention $\alpha_{3+1}:=\alpha_{1}$ and
$\alpha_{1-1}:=\alpha_{3}$.
For a $d$-dimensional measurable set $D,$ we write $\left\vert D\right\vert $
for its measure; for a discrete set, say $\mathcal{J}$, we denote by
$\left\vert \mathcal{J}\right\vert $ its cardinality.
In the proofs, we consider frequently nodal patches $\mathcal{T}_{\mathbf{z}}$
for inner vertices $\mathbf{z}\in\mathcal{V}_{\Omega}\left( \mathcal{T
\right) $. The number $m$ denotes the number of triangles in $\mathcal{T
_{\mathbf{z}}$. Various quantities in this patch such as, e.g., the triangles
in $\mathcal{T}_{\mathbf{z}}$, have an index which runs from $1$ to $m$. Here,
we use the cyclic numbering convention $K_{m+1}:=K_{1}$ and $K_{1-1}:=K_{m}$
and apply this analogously for other quantities in the nodal patch.
\end{notation}
Let the closed reference triangle $\widehat{K}$ be the triangle with vertices
$\mathbf{\hat{A}}_{1}:=\left( 0,0\right) $, $\mathbf{\hat{A}}_{2}:=\left(
1,0\right) $, $\mathbf{\hat{A}}_{3}:=\left( 0,1\right) $. The nodal points
on the reference element of order $k\in\mathbb{N}_{0}$ are given b
\[
\widehat{\mathcal{N}}_{k}:=\left\{
\begin{array}
[c]{ll
\left\{ \dfrac{1}{k
\mbox{\boldmath$ \mu$
\mi
\mbox{\boldmath$ \mu$
\in\mathbb{N}_{0}^{2}:\mathbb{\quad}\left\vert
\mbox{\boldmath$ \mu$
\right\vert \leq k\right\} & k\geq1,\\
& \\
\left\{ \left( \dfrac{1}{3},\dfrac{1}{3}\right) \right\} & k=0.
\end{array}
\right.
\]
For a triangle $K\subset\mathbb{R}^{2}$, we denote by $\chi_{K}:\widehat{K
\rightarrow K$ an affine bijection. The mapped nodal points of order
$k\in\mathbb{N}_{0}$ on $K$ are given b
\[
\mathcal{N}_{k}\left( K\right) :=\left\{ \chi_{K}\left( \mathbf{z}\right)
:\mathbf{z}\in\widehat{\mathcal{N}}_{k}\right\} .
\]
The nodal points of order $k$ on $\mathcal{T}$ are defined b
\[
\mathcal{N}_{k}\left( \mathcal{T}\right) :
{\displaystyle\bigcup\limits_{K\in\mathcal{T}}}
\mathcal{N}_{k}\left( K\right) \quad\text{and\quad}\mathcal{N
_{\partial\Omega}^{k}\left( \mathcal{T}\right) :=\mathcal{N}_{k}\left(
\mathcal{T}\right) \cap\partial\Omega,\quad\mathcal{N}_{k,\Omega}\left(
\mathcal{T}\right) :=\mathcal{N}_{k}\left( \mathcal{T}\right) \cap\Omega.
\]
We introduce the Lagrange basis for the space $S_{k}\left( \mathcal{T
\right) $, which is indexed by the nodal points $\mathbf{z}\in\mathcal{N
_{k}\left( \mathcal{T}\right) $ and characterized by
\begin{equation}
B_{k,\mathbf{z}}\in S_{k}\left( \mathcal{T}\right) \quad\text{and\quad
}\forall\mathbf{z}^{\prime}\in\mathcal{N}_{k}\left( \mathcal{T}\right)
\qquad B_{k,\mathbf{z}}\left( \mathbf{z}^{\prime}\right) =\delta
_{\mathbf{z},\mathbf{z}^{\prime}}, \label{basisfunctions
\end{equation}
where $\delta_{\mathbf{z},\mathbf{z}^{\prime}}$ is the Kronecker delta. A
basis for the space $S_{k,0}\left( \mathcal{T}\right) $ is given by
$B_{k,\mathbf{z}}$, $\mathbf{z}\in\mathcal{N}_{k,\Omega}\left( \mathcal{T
\right) $.\bigskip
Let $K$ denote a triangle with vertices $\mathbf{A}_{i}$, $1\leq i\leq3$, and
let $\lambda_{K,\mathbf{A}_{i}}\in\mathbb{P}_{1}\left( K\right) $ be the
\textit{barycentric coordinate} for the node $\mathbf{A}_{i}$ defined b
\begin{equation}
\lambda_{K,\mathbf{A}_{i}}\left( \mathbf{A}_{j}\right) =\delta_{i,j
\quad1\leq i,j\leq3. \label{lambdaintro1
\end{equation}
If the numbering of the vertices in $K$ is fixed, we write $\lambda_{K,i}$
short for $\lambda_{K,\mathbf{A}_{i}}$. For the barycentric coordinate on the
reference element $\widehat{K}$ for the vertex $\mathbf{\hat{A}}_{j}$ we write
$\widehat{\lambda}_{j}$, $j=1,2,3$. Elementary calculation yield (see, e.g.,
\cite[Appendix A]{CCSS_CR_1}
\begin{equation}
\partial_{\mathbf{n}_{k}}\lambda_{K,\mathbf{A}_{i}}=\frac{\left\vert
E_{i}\right\vert }{2\left\vert K\right\vert }\times\left\{
\begin{array}
[c]{ll
-1 & i=k,\\
\cos\alpha_{\ell} & \ell\text{ s.t. }\left\{ \ell,i,k\right\} =\left\{
1,2,3\right\} ,
\end{array}
\right. \label{normcomp
\end{equation}
where $E_{i}$ is the edge of $K$ opposite to $\mathbf{A}_{i}$, $\mathbf{n
_{k}$ the outward unit normal at $E_{k}$, and $\alpha_{\ell}$ the angle in $K$
at $\mathbf{A}_{\ell}$.
\begin{definition}
Let $L_{k}$ denote the usual univariate Legendre polynomial of degree $k$ (see
\cite[Table 18.3.1]{NIST:DLMF}). Let $k\in\mathbb{N}$ be even and
$K\in\mathcal{T}$. Then, the \emph{non-conforming triangle bubble }is given b
\[
B_{k,K}^{\operatorname*{CR}}:=\left\{
\begin{array}
[c]{ll
\dfrac{1}{2}\left( -1
{\displaystyle\sum\limits_{i=1}^{3}}
L_{k}\left( 1-2\lambda_{K,i}\right) \right) & \text{on }K,\\
0 & \text{on }\Omega\backslash K.
\end{array}
\right.
\]
For $k$ odd and $E\in\mathcal{E}\left( \mathcal{T}\right) $, the
\emph{non-conforming edge bubble} is given b
\begin{equation}
B_{k,E}^{\operatorname*{CR}}:=\left\{
\begin{array}
[c]{ll
L_{k}\left( 1-2\lambda_{K,\mathbf{A}_{K,E}}\right) & \text{on }K\text{ for
}K\in\mathcal{T}_{E},\\
0 & \text{on }\Omega\backslash\omega_{E},
\end{array}
\right. \label{BkEdef
\end{equation}
where $\mathbf{A}_{K,E}$ denotes the vertex in $K$ opposite to $E$.
\end{definition}
Different representations of the functions $B_{k,E}^{\operatorname*{CR}}$,
$B_{k,K}^{\operatorname*{CR}}$ exist in the literature, see \cite{BaranCVD},
\cite{Ainsworth_Rankin}, \cite[for $p=4,6.$]{ChaLeeLee}, \cite{ccss_2012}
while the formula for $B_{k,K}^{\operatorname*{CR}}$ has been introduced in
\cite{Baran_Stoyan} and the one for $B_{k,E}^{\operatorname*{CR}}$ in
\cite{CCSS_CR_1}.
\begin{proposition}
\label{ThmBasisCRscalar}A basis for the space $\operatorname*{CR}_{k,0}\left(
\mathcal{T}\right) $ is given
\begin{enumerate}
\item for even $k$ b
\[
\left\{ B_{k,\mathbf{z}}\mid\mathbf{z}\in\mathcal{N}_{k,\Omega}\left(
\mathcal{T}\right) \right\} \cup\left\{ B_{k,K}^{\operatorname*{CR}}\mid
K\in\mathcal{T}\right\} ,
\]
\item for odd $k$ b
\[
\left\{ B_{k,\mathbf{z}}\mid\mathbf{z}\in\mathcal{N}_{k,\Omega}\left(
\mathcal{T}\right) \backslash\mathcal{V}_{\Omega}\left( \mathcal{T}\right)
\right\} \cup\left\{ B_{k,E}^{\operatorname*{CR}}\mid E\in\mathcal{E
_{\Omega}\left( \mathcal{T}\right) \right\} .
\]
\end{enumerate}
\end{proposition}
The proof of this proposition and the following corollary can be found, e.g.,
in \cite[Rem. 3]{BaranCVD}, \cite[Thm. 22]{ccss_2012}, \cite[Cor.
3.4]{CCSS_CR_1}.
\begin{corollary}
\label{CorBasis}A basis for the space $\mathbf{CR}_{k,0}\left( \mathcal{T
\right) $ is given
\begin{enumerate}
\item for even $k$ b
\begin{equation
\begin{array}
[c]{l
\left\{ B_{k,\mathbf{z}}\mathbf{v}_{\mathbf{z}}\mid\mathbf{z}\in
\mathcal{N}_{k,\Omega}\left( \mathcal{T}\right) \right\} \cup\left\{
B_{k,\mathbf{z}}\mathbf{w}_{\mathbf{z}}\mid\mathbf{z}\in\mathcal{N}_{k,\Omega
}\left( \mathcal{T}\right) \right\} \\
\quad\\
\quad\cup\left\{ B_{k,K}^{\operatorname*{CR}}\mathbf{v}_{K}\mid
K\in\mathcal{T}\right\} \cup\left\{ B_{k,K}^{\operatorname*{CR}
\mathbf{w}_{K}\mid K\in\mathcal{T}\right\} ,
\end{array}
\label{basisvel
\end{equation}
\item for odd $k$ b
\begin{equation
\begin{array}
[c]{l
\left\{ B_{k,\mathbf{z}}\mathbf{v}_{\mathbf{z}}\mid\mathbf{z}\in
\mathcal{N}_{k,\Omega}\left( \mathcal{T}\right) \backslash\mathcal{V
_{\Omega}\left( \mathcal{T}\right) \right\} \cup\left\{ B_{k,\mathbf{z
}\mathbf{w}_{\mathbf{z}}\mid\mathbf{z}\in\mathcal{N}_{k,\Omega}\left(
\mathcal{T}\right) \backslash\mathcal{V}_{\Omega}\left( \mathcal{T}\right)
\right\} \\
\quad\\
\quad\cup\left\{ B_{k,E}^{\operatorname*{CR}}\mathbf{v}_{E}\mid
E\in\mathcal{E}_{\Omega}\left( \mathcal{T}\right) \right\} \cup\left\{
B_{k,E}^{\operatorname*{CR}}\mathbf{w}_{E}\mid E\in\mathcal{E}_{\Omega}\left(
\mathcal{T}\right) \right\} .
\end{array}
\label{basisvelodd
\end{equation}
\end{enumerate}
Here, for any nodal point $\mathbf{z}$, the linearly independent vectors
$\mathbf{v}_{\mathbf{z}},\mathbf{w}_{\mathbf{z}}\in\mathbb{R}^{2}$ can be
chosen arbitrarily. The same holds for any triangle $K$ for the vectors
$\mathbf{v}_{K},\mathbf{w}_{K}\in\mathbb{R}^{2}$ in (\ref{basisvel}) and for
any $E\in\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ for the vectors
$\mathbf{v}_{E},\mathbf{w}_{E}\in\mathbb{R}^{2}$ in (\ref{basisvelodd}).
\end{corollary}
\begin{remark}
The original definition by \cite{CrouzeixRaviart} is implicit and given for
conforming simplicial finite element meshes in $\mathbb{R}^{d}$, $d=2,3$. For
their practical implementation, a basis is needed and Corollary \ref{CorBasis}
provides a simple definition. A basis for Crouzeix-Raviart finite elements in
$\mathbb{R}^{3}$ has been introduced in \cite{Fortin_d3} for $k=2$ and a
general construction is given in \cite{CDS}.
\end{remark}
\subsection{The case of odd $k\geq5$\label{PTpge5}}
In this section, we assume for the followin
\begin{equation
\begin{array}
[c]{ll
\text{a)} & k\geq5\text{ is odd and}\\
\text{b)} & \mathcal{T}\text{ is a conforming triangulation and has at least
one inner vertex.
\end{array}
\label{Assumptab
\end{equation}
This section is structured as follows. In \S \ref{GeoPrel} we generalize the
concept of critical points (see \cite{vogelius1983right}, \cite{ScottVogelius
) to $\eta$-critical points which turn out to be essential for estimates with
constants depending on the mesh only via the shape-regularity constant. We
split these $\eta$-critical points into a set of \textquotedblleft
obtuse\textquotedblright\ $\eta-$critical points and \textquotedblleft
acute\textquotedblright\ $\eta-$critical points. In \S \ref{Obtuse}, we
provide the proof of Theorem \ref{Theomain} for a maximal partial
triangulation that does not contain acute $\eta-$critical points and satisfies
(\ref{Assumptab}). Finally, in \S \ref{SecAcute} we present the argument to
allow for acute $\eta-$critical points.
\subsubsection{Geometric preliminaries\label{GeoPrel}}
For the analysis of the inf-sup constant we start with the definition of
\textit{critical points} (see \cite{vogelius1983right}, \cite{ScottVogelius}).
\begin{definition}
\label{DefCritpoint}Let $\mathcal{T}$ denote a triangulation as in
\S \ref{NuDiscrete}. For $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right)
$, let
\[
\mathcal{E}_{\mathbf{z}}:=\left\{ E\in\mathcal{E}\left( \mathcal{T}\right)
:\mathbf{z}\text{ is an endpoint of }E\right\} .
\]
The point $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) $ is a
\emph{critical point} for $\mathcal{T}$ if there exist two straight infinite
lines $L_{1}$, $L_{2}$ in $\mathbb{R}^{2}$ such that all edges $E\in
\mathcal{E}_{\mathbf{z}}$ satisfy $E\subset L_{1}\cup L_{2}$. The set of all
critical points in $\mathcal{T}$ is $\mathcal{C}_{\mathcal{T}}$.
\end{definition}
\begin{figure}[ptb
\centering
\includegraphics[
height=2.9386in,
width=5.5737in
{critcases.eps
\caption{Illustration of the four critical cases as in Remark
\ref{RemCritGeom}. Left top: inner critical point, right top: acute critical
point, left bottom: flat critical point, right bottom: obtuse critical point.
\label{FigCritCases
\end{figure}
\begin{remark}
\label{RemCritGeom}Geometric configurations where critical points occur are
well studied in the literature (see, e.g., \cite{ScottVogelius}). Any critical
point $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}$ belongs to one of the following
cases (see Fig. \ref{FigCritCases}):
\begin{enumerate}
\item $\mathbf{z}\in\mathcal{V}_{\Omega}\left( \mathcal{T}\right) $ and
$\mathcal{T}_{\mathbf{z}}$ consists of four triangles and $\mathbf{z}$ is the
intersections of the two diagonals in $\omega_{\mathbf{z}}$.
\item $\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left( \mathcal{T}\right) $
and $\operatorname*{card}\mathcal{E}_{\mathbf{z}}=2$, i.e., both edges
$E\in\mathcal{E}_{\mathbf{z}}$ are boundary edges which joint $\mathbf{z}$.
\item $\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left( \mathcal{T}\right) $
and $\operatorname*{card}\mathcal{E}_{\mathbf{z}}=3$ and two edges
$E\in\mathcal{E}_{\mathbf{z}}$ are boundary edges which lie on a straight
boundary piece.
\item $\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left( \mathcal{T}\right) $
and $\operatorname*{card}\mathcal{E}_{\mathbf{z}}=4$ and each of the two
boundary edges is aligned with one edge $\mathcal{E}_{\mathbf{z}
\cap\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $.
\end{enumerate}
\end{remark}
\begin{definition}
Let $\mathcal{T}$ denote a triangulation as in \S \ref{NuDiscrete}. Let
$\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) $ and the nodal patch
$\mathcal{T}_{\mathbf{z}}$ as in (\ref{nodalpatch}). Let the triangles
$K_{\ell}$, $1\leq\ell\leq m$, in $\mathcal{T}_{\mathbf{z}}$ be numbered
counterclockwise and denote the angle in $K_{\ell}$ at $\mathbf{z}$ by
$\omega_{\ell}$. Then,
\[
\Theta\left( \mathbf{z}\right) :=\left\{
\begin{array}
[c]{ll
\max\left\{ \left\vert \sin\left( \omega_{1}+\omega_{2}\right) \right\vert
,\left\vert \sin\left( \omega_{2}+\omega_{3}\right) \right\vert
,\ldots,\left\vert \sin\left( \omega_{m}+\omega_{1}\right) \right\vert
\right\} & \text{if }\mathbf{z}\in\mathcal{V}_{\Omega}\left( \mathcal{T
\right) ,\\
\max\left\{ \left\vert \sin\left( \omega_{1}+\omega_{2}\right) \right\vert
,\left\vert \sin\left( \omega_{2}+\omega_{3}\right) \right\vert
,\ldots,\left\vert \sin\left( \omega_{m-1}+\omega_{m}\right) \right\vert
\right\} & \text{if }\mathbf{z}\in\Gamma\wedge m>1,\\
0 & \text{if }\mathbf{z\in}\Gamma\wedge m=1.
\end{array}
\right.
\]
\end{definition}
\begin{remark}
It is easy to see that $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}$ if and only if
$\Theta\left( \mathbf{z}\right) =0$.
\end{remark}
\begin{lemma}
\label{Lemangle}Let $\phi_{\mathcal{T}}$ be as in Remark \ref{Remangle}. Se
\[
\eta_{0}:=\min\left\{ \frac{1}{2},c_{1},\frac{3\phi_{\mathcal{T}}}{\pi
,\sin\phi_{\mathcal{T}}\right\}
\]
wit
\[
c_{1}:=\left\{
\begin{array}
[c]{ll
\min\left\{ \sin2\phi_{\mathcal{T}},\left\vert \sin\left( 2\pi
-4\phi_{\mathcal{T}}\right) \right\vert \right\} & \phi_{\mathcal{T}}\leq
\pi/8,\\
\sin2\phi_{\mathcal{T}} & \pi/8<\phi_{\mathcal{T}}\leq\pi/4,\\
1 & \phi_{\mathcal{T}}>\pi/4.
\end{array}
\right.
\]
Let $0\leq\eta<\eta_{0}$ be fixed. If, for $\mathbf{z}\in\mathcal{V}\left(
\mathcal{T}\right) $, it holds $\Theta\left( \mathbf{z}\right) \leq\eta$,
then, for any edge $E=\left[ \mathbf{z},\mathbf{z}^{\prime}\right]
\in\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ it hold
\[
\Theta\left( \mathbf{z}^{\prime}\right) \geq\eta_{0}.
\]
\end{lemma}
\proof
Let $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) $ and consider an
edge $E=\left[ \mathbf{z},\mathbf{z}^{\prime}\right] \in\mathcal{E}_{\Omega
}\left( \mathcal{T}\right) $. Then, there are two triangles $K,K^{\prime
\in\mathcal{T}$ which are adjacent to $E$. The angle in $K$ resp. $K^{\prime}$
at $\mathbf{z}$ is denoted by $\omega$ resp. $\omega^{\prime}$.
\textbf{1st case. }Let $\omega+\omega^{\prime}\leq\pi/2$ or $\omega
+\omega^{\prime}\geq\frac{3}{2}\pi$. Then, we conclude from Remark
\ref{Remangle} tha
\[
2\phi_{\mathcal{T}}\leq\omega+\omega^{\prime}\leq\frac{\pi}{2}\quad
\text{or\quad}\frac{3}{2}\pi\leq\omega+\omega^{\prime}\leq2\pi-4\phi
_{\mathcal{T}}.
\]
For the left inequality to hold, the minimal angle must satisfy $\phi
_{\mathcal{T}}\leq\pi/4$ while for the right inequality, it must hold
$\phi_{\mathcal{T}}\leq\pi/8$. For $\phi_{\mathcal{T}}>\pi/4$, the \textbf{1st
case} is empty. We hav
\[
\Theta\left( \mathbf{z}\right) \geq\left\{
\begin{array}
[c]{ll
\min\left\{ \sin2\phi_{\mathcal{T}},\left\vert \sin\left( 2\pi
-4\phi_{\mathcal{T}}\right) \right\vert \right\} & \phi_{\mathcal{T}}\leq
\pi/8\\
\sin2\phi_{\mathcal{T}} & \pi/8<\phi_{\mathcal{T}}\leq\pi/4
\end{array}
\right\} \geq c_{1}\geq\eta_{0}.
\]
Since $\eta<\eta_{0}\leq c_{1}$ this case cannot appear.
\textbf{2nd case. }Let $\pi/2\leq\omega+\omega^{\prime}\leq3\pi/2$. The
condition $\left\vert \sin\left( \omega+\omega^{\prime}\right) \right\vert
\leq\eta$ then implies that $\omega+\omega^{\prime}=\pi+\delta$ with
\begin{equation}
\left\vert \delta\right\vert \leq\arcsin\eta=\eta\sum_{\ell=0}^{\infty
\eta^{2\ell}\frac{\left( 2\ell\right) !}{\left( \ell!\right) ^{2}4^{\ell
}\left( 2\ell+1\right) }\overset{\eta_{0}\leq1/2}{\leq}\eta\sum_{\ell
=0}^{\infty}2^{-2\ell}\frac{\left( 2\ell\right) !}{\left( \ell!\right)
^{2}4^{\ell}\left( 2\ell+1\right) }=\frac{\pi\eta}{3}. \label{deltaest
\end{equation}
Consequently the two angles $\alpha$ in $K$ and $\alpha^{\prime}$ in
$K^{\prime}$ at $\mathbf{z}^{\prime}$ satisf
\[
\alpha+\alpha^{\prime}=2\pi-\omega-\omega^{\prime}-\beta-\beta^{\prime
=\pi-\delta-\beta-\beta^{\prime}\leq\pi+\frac{\pi\eta}{3}-2\phi_{\mathcal{T
}\overset{\pi\eta/3\leq\phi_{\mathcal{T}}}{\leq}\pi-\phi_{\mathcal{T}},
\]
where $\beta$ (resp. $\beta^{\prime}$) denotes the third angle in $K$ (resp.
$K^{\prime}$). Hence, in this cas
\[
\Theta\left( \mathbf{z}^{\prime}\right) \geq\left\vert \sin\left( \pi
-\phi_{\mathcal{T}}\right) \right\vert =\sin\phi_{\mathcal{T}}\geq\eta_{0}.
\
\endproof
\begin{definition}
\label{Defetacriticalpoint}Let $\eta_{0}$ be as in Lemma \ref{Lemangle}. For
$0\leq\eta<\eta_{0}$, the set of $\eta$\emph{-critical points }$\mathcal{C
_{\mathcal{T}}\left( \eta\right) $ is given b
\[
\mathcal{C}_{\mathcal{T}}\left( \eta\right) :=\left\{ \mathbf{z
\in\mathcal{V}\left( \mathcal{T}\right) \mid\Theta\left( \mathbf{z}\right)
\leq\eta\right\} .
\]
A point $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left( \eta\right)
\backslash\mathcal{C}_{\mathcal{T}}\left( 0\right) $ is called a
\emph{nearly critical point}. An $\eta$-critical point $\mathbf{z
\in\mathcal{C}_{\mathcal{T}}\left( \eta\right) $ is \emph{isolated} if all
edge $\left[ \mathbf{z,z}^{\prime}\right] \in\mathcal{E}\left(
\mathcal{T}\right) $ satisfy: $\mathbf{z}^{\prime}$ is not an $\eta$-critical point.
\end{definition}
By perturbing the geometric configurations in Remark \ref{RemCritGeom} we
obtain the following subcases (see Fig. \ref{RemCritGeom2}).
\begin{definition}
\label{Remetacritcases}Let $\eta_{0}$ be as in Lemma \ref{Lemangle} and
$0\leq\eta<\eta_{0}$. If $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right)
$ satisfies
\begin{enumerate}
\item $\mathbf{z}\in\mathcal{V}_{\Omega}\left( \mathcal{T}\right) $ and
$\operatorname*{card}\mathcal{T}_{\mathbf{z}}=4$ and $\Theta\left(
\mathbf{z}\right) \leq\eta$. Then $\mathbf{z}$ is an \emph{inner} $\eta
$\emph{-critical point}. Le
\[
\mathcal{C}_{\mathcal{T}}^{\operatorname*{inner}}\left( \eta\right)
:=\left\{ \mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left( \eta\right)
:\mathbf{z}\text{ is an inner }\eta\text{-critical point}\right\} .
\]
\item $\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left( \mathcal{T}\right) $
and $\operatorname*{card}\mathcal{E}_{\mathbf{z}}=2$. Then $\mathbf{z}$ is an
\emph{acute critical point}. Let\footnote{Note that the set of acute critical
boundary points is independent of $\eta$.
\[
\mathcal{C}_{\mathcal{T}}^{\operatorname{acute}}:=\left\{ \mathbf{z
\in\mathcal{C}_{\mathcal{T}}:\mathbf{z}\text{ is an acute critical boundary
point}\right\} .
\]
\item $\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left( \mathcal{T}\right) $
and $\operatorname*{card}\mathcal{E}_{\mathbf{z}}=3$ and $\Theta\left(
\mathbf{z}\right) \leq\eta$. Then $\mathbf{z}$ is \emph{flat }$\eta
$\emph{-critical point}. Le
\[
\mathcal{C}_{\mathcal{T}}^{\operatorname*{flat}}\left( \eta\right)
:=\left\{ \mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left( \eta\right)
:\mathbf{z}\text{ is a flat }\eta\text{-critical boundary point}\right\} .
\]
\item $\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left( \mathcal{T}\right) $
and $\operatorname*{card}\mathcal{E}_{\mathbf{z}}=4$ and $\Theta\left(
\mathbf{z}\right) \leq\eta$. Then $\mathbf{z}$ is a (locally) \emph{concave
}$\eta$-\emph{critical point.} Le
\[
\mathcal{C}_{\mathcal{T}}^{\operatorname*{concave}}\left( \eta\right)
:=\left\{ \mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left( \eta\right)
:\mathbf{z}\text{ is a concave }\eta\text{-critical point}\right\} .
\]
\end{enumerate}
\end{definition}
\begin{figure}[ptb
\centering
\includegraphics[
height=2.9386in,
width=5.5737in
{etacritcases.eps
\caption{Illustration of the four $\eta$-critical cases as in Remark
\ref{Remetacritcases}. Left top: inner $\eta-$critical point, right top: acute
critical point, left bottom: flat $\eta$-critical point, right bottom: concave
$\eta$-critical point.
\label{RemCritGeom2
\end{figure}
The acute critical points require some special treatment and we denote the
union of the others by
\[
\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right)
:=\mathcal{C}_{\mathcal{T}}^{\operatorname*{inner}}\left( \eta\right)
\cup\mathcal{C}_{\mathcal{T}}^{\operatorname*{flat}}\left( \eta\right)
\cup\mathcal{C}_{\mathcal{T}}^{\operatorname*{concave}}\left( \eta\right) .
\]
The following lemma states that for a possibly adjusted $\eta_{0}$, still
depending only on the shape-regularity of the mesh, the $\eta$-critical points
belong to one of the four categories described in Definition
\ref{Remetacritcases}\emph{.}
\begin{lemma}
There exists some $\eta_{0}^{\prime}\in\left] 0,\eta_{0}\right] $ depending
only on the shape-regularity of the mesh such that for $0\leq\eta<\eta
_{0}^{\prime}$ any $\eta$-critical point belongs to one of the four categories
described in Definition \ref{Remetacritcases}.
\end{lemma}
\proof
We only consider the case $\mathbf{z}\in\mathcal{V}_{\Omega}\left(
\mathcal{T}\right) $ while the other cases follow by similar arguments. Let
$m:=\operatorname*{card}\mathcal{T}_{\mathbf{z}}$ and choose a
counterclockwise numbering for the triangles in $\mathcal{T}_{\mathbf{z}}$,
i.e., $K_{i}$, $1\leq i\leq m$. The shape-regularity of the mesh implies that
there is some $m_{\max}$ depending only on $\phi_{\mathcal{T}}$ such that
$m\leq m_{\max}$. Denote by $\omega_{i}$ the angle in $K_{i}$ at $\mathbf{z}$.
Let $\eta_{0}^{\prime}\in\left] 0,\eta_{0}\right] $ which will be fixed
later and assume $0\leq\eta<\eta_{0}^{\prime}$. Since $\mathbf{z}$ is an
$\eta$-critical point we have
\[
\left\vert \sin\left( \omega_{i}+\omega_{i+1}\right) \right\vert \leq
\eta<\eta_{0}^{\prime}\quad\forall1\leq i\leq m.
\]
The shape-regularity implies $\phi_{\mathcal{T}}\leq\omega_{i}\leq\pi
-2\phi_{\mathcal{T}}$ and, for $\delta=\arcsin\eta_{0}^{\prime}$, we ge
\begin{equation}
\omega_{i}+\omega_{i+1}\in\left[ 2\phi_{\mathcal{T}},\delta\right]
\cup\left[ \pi-\delta,\pi+\delta\right] \cup\left[ 2\pi-\delta,2\pi
-4\phi_{\mathcal{T}}\right] \quad\text{for all }1\leq i\leq m.
\label{intervals
\end{equation}
Since $\arcsin:\left[ 0,1\right[ \rightarrow\mathbb{R}_{\geq0}$ is
monotonously increasing with $\arcsin0=0$ and $\lim_{x\rightarrow1}\arcsin
x=+\infty$, we can select $\eta_{0}^{\prime}$ such $0<\delta<2\phi
_{\mathcal{T}}$. In turn, the first and last interval in (\ref{intervals}) are
empty an
\[
\omega_{i}+\omega_{i+1}\in\left[ \pi-\delta,\pi+\delta\right] \quad\text{for
all }1\leq i\leq m.
\]
We set $\omega_{i}+\omega_{i+1}=:\pi+\delta_{i}$ for some $\left\vert
\delta_{i}\right\vert \leq\delta$ and obtai
\begin{equation}
4\pi=\sum_{i=1}^{m}\left( \omega_{i}+\omega_{i+1}\right) =m\pi+\sum
_{i=1}^{m}\delta_{i}. \label{4pieq
\end{equation}
By adjusting $\eta_{0}^{\prime}$ such that $m_{\max}\delta<\pi$ we conclude
that $m=4$ and $\sum_{i=1}^{m}\delta_{i}=0$. Hence, $\mathbf{z}$ is an inner
$\eta$-critical point
\endproof
Next, we collect the $\eta-$critical points in pairwise disjoint,
edge-connected sets which we will define in the following.
\begin{figure}[ptb
\centering
\includegraphics[
height=1.7089in,
width=6.6193in
{fan4.eps
\caption{Three types of obtuse $\eta$-critical points $\mathbf{z
\in\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right) $
with associated inner edge $\mathfrak{E}\left( \mathbf{z}\right) $, normal
vector $\mathfrak{N}\left( \mathbf{z}\right) $ and opposite endpoint
$\mathfrak{V}\left( \mathbf{z}\right) $ of $\mathfrak{E}\left(
\mathbf{z}\right) $; left: inner $\eta$-critical point, middle: flat $\eta
$-critical point, right: concave $\eta$-critical point.
\label{Figobtuse
\end{figure}
We say two points $\mathbf{y},\mathbf{y}^{\prime}\in\mathcal{V}\left(
\mathcal{T}\right) $ are \textit{edge-connected} if there is an edge
$E\in\mathcal{E}\left( \mathcal{T}\right) $ with endpoints $\mathbf{y,y
^{\prime}$. A subset $\mathcal{V}^{\prime}\subset\mathcal{V}\left(
\mathcal{T}\right) $ is edge-connected if there is a numbering of the points
in $\mathcal{V}^{\prime}=\left\{ \mathbf{y}_{j}:1\leq j\leq n\right\} $ such
that $\mathbf{y}_{j-1}$, $\mathbf{y}_{j}$ are edge-connected for all $2\leq
j\leq n$. A point $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) $ is
edge-connected to $\mathcal{V}^{\prime}$ if $\mathbf{z}\in\mathcal{V}^{\prime
}$ or there is $\mathbf{y\in}\mathcal{V}^{\prime}$ such that $\mathbf{z}$,
$\mathbf{y}$ are edge-connected.
From Lemma \ref{Lemangle} we know that two edge-connected points
$\mathbf{z},\mathbf{z}^{\prime}\in\mathcal{V}\left( \mathcal{T}\right) $ can
be both critical only if the connecting edge $E$ belongs to $\mathcal{E
_{\partial\Omega}\left( \mathcal{T}\right) $; as a consequence
$\mathbf{z},\mathbf{z}^{\prime}\in\mathcal{V}_{\partial\Omega}\left(
\mathcal{T}\right) $. Next, we will group the points in $\mathcal{C
_{\mathcal{T}}\left( \eta\right) $ into subsets called \textit{fans}.
From Lemma \ref{Lemangle} it follows that the points in $\mathcal{C
_{\mathcal{T}}^{\operatorname*{inner}}\left( \eta\right) $ are isolated. All
other $\eta$-critical points lie on the boundary. Next, we define mappings
$\mathfrak{E}:\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left(
\eta\right) \rightarrow\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $,
$\mathfrak{N}:\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left(
\eta\right) \rightarrow\mathbb{S}_{2}$, and $\mathfrak{V}:\mathcal{C
_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right) \rightarrow
\mathcal{V}\left( \mathcal{T}\right) \backslash\mathcal{C}_{\mathcal{T
}\left( \eta\right) $. The construction is illustrated in Figure
\ref{Figobtuse}.
For $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left(
\eta\right) $, Definition \ref{Remetacritcases} implies that $\left\vert
\mathcal{E}_{\mathbf{z}}\right\vert \geq3$ and hence $\mathcal{E}_{\mathbf{z
}\cap\mathcal{E}_{\Omega}\left( \mathcal{T}\right) \neq0$. We fix one edge
$E\in\mathcal{E}_{\mathbf{z}}\cap\mathcal{E}_{\Omega}\left( \mathcal{T
\right) $ and set $\mathfrak{E}\left( \mathbf{z}\right) :=E$. Note that the
choice of $E$ is unique for $\mathbf{z}\in\mathcal{C}_{\mathcal{T
}^{\operatorname*{flat}}\left( \eta\right) $. For $\mathbf{z}\in
\mathcal{C}_{\mathcal{T}}^{\operatorname*{inner}}\left( \eta\right) $ the
choice is arbitrary. For $\mathbf{z}\in\mathcal{C}_{\mathcal{T}
^{\operatorname*{concave}}\left( \eta\right) $, the set $\mathcal{E
_{\mathbf{z}}\cap\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ consists of
two edges, say $E_{1}$, $E_{2}$. We fix one of them and set $\mathfrak{E
\left( \mathbf{z}\right) :=E_{2}$. Let $\mathbf{z}^{\prime}\in
\mathcal{V}\left( \mathcal{T}\right) $ be such that $\mathfrak{E}\left(
\mathbf{z}\right) =\left[ \mathbf{z},\mathbf{z}^{\prime}\right] $. Then
$\mathfrak{V}\left( \mathbf{z}\right) :=\mathbf{z}^{\prime}$. Lemma
\ref{Lemangle} implies that $\mathbf{z}^{\prime}$ is not an $\eta$-critical
point. A unit vector $\mathfrak{N}\left( \mathbf{z}\right) $ orthogonal to
$\mathfrak{E}\left( \mathbf{z}\right) $ is defined by the condition that
$\mathbf{z}-\mathbf{z}^{\prime}$ and $\mathfrak{N}\left( \mathbf{z}\right) $
form a right-handed system.
\begin{definition}
\label{DefFan}We decompose $C_{\mathcal{T}}^{\operatorname*{obtuse}}\left(
\eta\right) $ into disjoint \emph{fans} $\mathcal{C}_{\mathcal{T},\ell
}\left( \eta\right) $, $\ell\in\mathcal{J}$, such that the following
conditions are satisfies
\begin{enumerate}
\item $\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right)
{\displaystyle\bigcup\limits_{\ell\in\mathcal{J}}}
\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) ,$
\item for any $\ell\in\mathcal{J}$, the set $\mathcal{C}_{\mathcal{T},\ell
}\left( \eta\right) $ is edge-connected,
\item for any $\ell\in\mathcal{J}$, there is $\mathbf{z}_{\ell}\in
\mathcal{V}\left( \mathcal{T}\right) \backslash\mathcal{C}_{\mathcal{T
}\left( \eta\right) $ such that for all $\mathbf{z}\in\mathcal{C
_{\mathcal{T},\ell}\left( \eta\right) $ it holds $\mathfrak{V}\left(
\mathbf{z}\right) =\mathbf{z}_{\ell}$ and, vice versa:
\item any $\mathbf{z}^{\prime}\in\mathcal{C}_{\mathcal{T}
^{\operatorname*{obtuse}}\left( \eta\right) $ which is edge-connected to
some $\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) $ and satisfies
$\mathfrak{V}\left( \mathbf{z}^{\prime}\right) =\mathbf{z}_{\ell}$ belongs
to $\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) $.
\end{enumerate}
\end{definition}
The following lemma will allow us to construct the right-inverse of the
divergence separately for each fan.
\begin{lemma}
\label{Lemomegaell}Let $\eta_{0}$ be as in in Lemma \ref{Lemangle} and let
$0\leq\eta<\eta_{0}$ be fixed.
\begin{enumerate}
\item[a.] Then, the mapping $\mathfrak{E}:\mathcal{C}_{\mathcal{T
}^{\operatorname*{obtuse}}\left( \eta\right) \rightarrow\mathcal{E}_{\Omega
}\left( \mathcal{T}\right) $ is injective.
\item[b.] For $\ell\in\mathcal{J}$, let $\omega_{\ell}:
{\displaystyle\bigcup\limits_{\mathbf{z}\in\mathcal{C}_{\mathcal{T},\ell}}}
\mathcal{\omega}_{\mathfrak{E}\left( \mathbf{z}\right) }$. The domains
$\omega_{\ell}$ have pairwise disjoint interior.
\end{enumerate}
\end{lemma}
\proof
\textbf{Part} \textbf{a. }The injectivity of the mapping $\mathfrak{E
:\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right)
\rightarrow\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ follows from
Lemma \ref{Lemangle}: if $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left(
\eta\right) $ and $\mathbf{z}^{\prime}\in\mathcal{V}\left( \mathcal{T
\right) $ is such that $E:=\left[ \mathbf{z},\mathbf{z}^{\prime}\right]
\in\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ then $\mathbf{z}^{\prime
}\notin\mathcal{C}_{\mathcal{T}}\left( \eta\right) $.
\textbf{Part b. }The following construction is illustrated in Figure
\ref{Fig_nod_patch}
\begin{figure}[ptb
\centering
\includegraphics[
height=3.0761in,
width=5.7432in
{fan5.eps
\caption{Nodal patch, illustrating edge-connected obtuse $\eta$-critical
points of a fan $\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) $. The
left-most $\eta$-critical point is $\mathbf{z}_{\ell,n_{\ell}}$ and of type
\textquotedblleft flat\textquotedblright, the right-most is $\mathbf{z
_{\ell,1}$ of type \textquotedblleft concave\textquotedblright\ with
$\mathfrak{E}\left( \mathbf{z}_{\ell,1}\right) =\left[ \mathbf{z}_{\ell
,1},\mathbf{z}\right] $. The extremal points $\mathbf{z}_{\ell,0}$ and
$\mathbf{z}_{\ell,n_{\ell}+1}$ do not belong to $\mathcal{C}_{\mathcal{T
,\ell}\left( \eta\right) $. The edge connecting $\mathbf{z}_{\ell}$ with
$\mathbf{z}_{\ell,j}$ is denoted by $E_{\ell,j}$.
\label{Fig_nod_patch
\end{figure}
For fixed $\ell\in\mathcal{J}$, we number the points in $\mathcal{C
_{\mathcal{T},\ell}$ by $\mathbf{z}_{\ell,j}$, $1\leq j\leq n_{\ell}$, such
that $\mathbf{z}_{\ell,j-1}$ and $\mathbf{z}_{\ell,j}$ are edge-connected for
all $2\leq j\leq n_{\ell}$ and $\mathbf{z}_{\ell,1}$ and $\mathbf{z
_{\ell,n_{\ell}}$ are the first and last point in the polygonal line through
these points. Let $K_{\ell,j}$ be the triangle with vertices $\mathbf{z
_{\ell,j-1}$, $\mathbf{z}_{\ell}$, $\mathbf{z}_{\ell,j}$, $2\leq j\leq
n_{\ell}\ $and let $\mathcal{T}_{\ell}:=\left\{ K_{\ell,j}:2\leq j\leq
n_{\ell}\right\} $. Note that this set is empty if $\left\vert \mathcal{C
_{\mathcal{T},\ell}\right\vert =1$. Let $K_{\ell,1},K_{\ell,n_{\ell}+1
\in\mathcal{T}\backslash\mathcal{T}_{\ell}$ be two different triangles such
that $E_{\ell,1}:=\left[ \mathbf{z}_{\ell,1},\mathbf{z}_{\ell}\right]
\subset\partial K_{\ell,1}$ and $E_{\ell,n_{\ell}}:=\left[ \mathbf{z
_{\ell,n_{\ell}},\mathbf{z}_{\ell}\right] \subset\partial K_{\ell,n_{\ell
+1}$. Let $\mathbf{z}_{\ell,0}$ be the third vertex in $K_{\ell,1}$ and
observe that it does not belong to $\mathcal{C}_{\mathcal{T},\ell}$. Since
$E_{\ell,1}$ is an inner edge and $\mathbf{z}_{\ell,1}$ is an $\eta$-critical
point, Lemma \ref{Lemangle} implies that $\mathbf{z}_{\ell}$ is not an $\eta
$-critical point. Next, we show that $E_{\ell,0}:=\left[ \mathbf{z}_{\ell
,0},\mathbf{z}_{\ell}\right] $ does not belong to $\mathcal{E}_{\ell
}:=\left\{ \mathfrak{E}\left( \mathbf{z}\right) :\mathbf{z}\in
\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right)
\right\} $; from this the assertion follows. We assume $E_{\ell,0
\in\mathcal{E}_{\ell}$ and derive a contradiction. Since $\mathbf{z}_{\ell
}\notin\mathcal{C}_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right)
$, this assumption implies that $\mathbf{z}_{\ell,0}\in\mathcal{C
_{\mathcal{T}}^{\operatorname*{obtuse}}\left( \eta\right) $. If
$\mathfrak{V}\left( \mathbf{z}_{\ell,0}\right) =\mathbf{z}_{\ell}$ then
Definition \ref{DefFan}(4) implies that $\mathbf{z}_{\ell,0}\in\mathcal{C
_{\mathcal{T},\ell}$ and this is a contradiction. If $\mathfrak{V}\left(
\mathbf{z}_{\ell,0}\right) \neq\mathbf{z}_{\ell}$, then $E_{\ell,0
\notin\mathcal{E}_{\ell}$.
\endproof
Since the acute critical boundary points need some special treatment we define
a sequence of triangulations $\mathcal{T}_{i}$, $1\leq i\leq L$, with the properties
\begin{enumerate}
\item
\begin{equation}
\mathcal{T}_{1}\subset\mathcal{T}_{2}\subset\ldots\subset\mathcal{T
_{L}=\mathcal{T}, \label{stepbystepextension
\end{equation}
\item $\mathcal{T}_{1}$ is a maximal subset of $\mathcal{T}$ such that
$\mathcal{C}_{\mathcal{T}_{1}}^{\operatorname{acute}}=\emptyset$,
\item for $j=1,2,\ldots,L$
\[
\mathcal{T}_{j}=\left\{ K\in\mathcal{T}\mid\exists E\subset\partial K\text{
s.t. }E\in\mathcal{E}\left( \mathcal{T}_{j-1}\right) \right\} .
\]
\end{enumerate}
By this step-by-step procedure, triangles are attached to a previous
triangulation $\mathcal{T}_{j-1}$ which have an edge in common with the set of
edges in $\mathcal{T}_{j-1}$. The proof of Theorem \ref{Theomain} under
assumption (\ref{Assumptab}) then consists of first proving the inf-sup
stability for $\mathcal{T}_{1}$ and then to investigate the effect of
attaching a triangle to an inf-sup stable triangulation. A sufficient
condition for $L=0$ is that every triangle in $\mathcal{T}$ has an interior point.
\subsubsection{The case $\mathcal{C}_{\mathcal{T}}^{\operatorname{acute
}=\emptyset$\label{Obtuse}}
In this section, we prove the inf-sup stability for the triangulation
$\mathcal{T}_{1}$ in (\ref{stepbystepextension}) where $\mathcal{C
_{\mathcal{T}_{1}}^{\operatorname{acute}}=\emptyset$. For simplicity we skip
the index $1$ and write $\mathcal{T}$, $\mathcal{C}_{\mathcal{T}}\left(
\eta\right) $, etc.
Next, we define some fundamental non-conforming Crouzeix-Raviart vector field
which will be used to eliminate the critical pressures in the Stokes pair
$\left( \mathbf{S}_{k,0}\left( \mathcal{T}\right) ,\mathbb{P
_{k-1,0}\left( \mathcal{T}\right) \right) $.
The essential properties of the Crouzeix-Raviart function $B_{k,E
^{\operatorname*{CR}}$ are: it is a polynomial of degree $k$ and a Legendre
polynomial on each edge $E^{\prime}\subset\partial\omega_{E}$ so that the jump
relations in (\ref{PCR0RB}) are satisfied. Furthermore, $\left[
B_{k,E}^{\operatorname*{CR}}\right] _{E}=0$ and $\operatorname*{supp
B_{k,E}^{\operatorname*{CR}}\subset\omega_{E}$.
In the first step, we modify the function $B_{k,E}^{\operatorname*{CR}}$ by
adding a continuous edge bubble in $S_{k,0}\left( \mathcal{T}\right) $ such
that the $H^{1}\left( \Omega\right) $ norm of the modified function has an
improved behaviour with respect to $k$.
Let $E\in\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ with endpoints
$\mathbf{V}_{1}$, $\mathbf{V}_{2}$. Set $\mathbf{t}_{E}=\left( \mathbf{V
_{2}-\mathbf{V}_{1}\right) /\left\Vert \mathbf{V}_{2}-\mathbf{V
_{1}\right\Vert $ and consider a function $w_{E}\in\mathbb{P}_{k}\left(
\mathcal{T}\right) $ with $\operatorname*{supp}w_{E}=\omega_{E}$ an
\begin{subequations}
\label{Defweall}
\end{subequations
\begin{align}
\left. \left. w_{E}\right\vert _{K}\right\vert _{E^{\prime}} & =\left.
\left. B_{k,E}^{\operatorname*{CR}}\right\vert _{K}\right\vert _{E^{\prime
}\quad\forall K\in\mathcal{T}_{E}\text{ and }E^{\prime}\subset\partial
K\cap\partial\omega_{E},\tag
\ref{Defweall
a}\label{Defwe}\\
\left[ w_{E}\right] _{E} & =0\quad\text{and\quad}\partial_{\mathbf{t}_{E
}w_{E}\left( \mathbf{V}_{1}\right) =\partial_{\mathbf{t}_{E}}w_{E}\left(
\mathbf{V}_{2}\right) =0. \tag
\ref{Defweall
b}\label{Defweallb
\end{align}
Then, $w_{E}$ also belongs to the space $\operatorname*{CR}_{k,0}\left(
\mathcal{T}\right) $ an
\[
\left. \operatorname*{div}\nolimits_{\mathcal{T}}\left( w_{E}\mathbf{n
_{E}\right) \right\vert _{K}\left( \mathbf{z}\right) =\left.
\operatorname*{div}\nolimits_{\mathcal{T}}\left( B_{k,E}^{\operatorname*{CR
}\mathbf{n}_{E}\right) \right\vert _{K}\left( \mathbf{z}\right)
\quad\forall K\in\mathcal{T\quad\forall}\mathbf{z}\in\mathcal{V}\left(
K\right) .
\]
The last relation can be seen from the following reasoning. For $K\in
\mathcal{T}$ and $\mathbf{z}\in\mathcal{V}\left( K\right) $, set
$\mathbf{t}_{\mathbf{y}}:=\left( \mathbf{y}-\mathbf{z}\right) /\left\Vert
\mathbf{y-z}\right\Vert $ for all $\mathbf{y}\in\mathcal{V}\left( K\right)
\backslash\left\{ \mathbf{z}\right\} $. Clearly, $\mathbf{n}_{E
=\sum_{\mathbf{y}\in\mathcal{V}\left( K\right) \backslash\left\{
\mathbf{z}\right\} }\alpha_{\mathbf{y}}\mathbf{t}_{\mathbf{y}}$ for some
$\alpha_{\mathbf{y}}\in\mathbb{R}$. The conditions (\ref{Defweall}) implies
that for $\mathbf{y}\in\mathcal{V}\left( K\right) \backslash\left\{
\mathbf{z}\right\} $ it hold
\[
\frac{\partial\left. w_{E}\right\vert _{K}}{\partial\mathbf{t}_{\mathbf{y}
}=\frac{\partial\left. B_{k,E}^{\operatorname*{CR}}\right\vert _{K}
{\partial\mathbf{t}_{\mathbf{y}}}.
\]
Hence
\begin{align}
\left. \operatorname*{div}\nolimits_{\mathcal{T}}\left( w_{E}\mathbf{n
_{E}\right) \right\vert _{K}\left( \mathbf{z}\right) & =\frac
{\partial\left. w_{E}\right\vert _{K}}{\partial\mathbf{n}_{E}}\left(
\mathbf{z}\right) =\sum_{\mathbf{y}\in\mathcal{V}\left( K\right)
\backslash\left\{ \mathbf{z}\right\} }\alpha_{\mathbf{y}}\frac
{\partial\left. w_{E}\right\vert _{K}}{\partial\mathbf{t}_{\mathbf{y}
}\left( \mathbf{z}\right) \label{reasoningz}\\
& =\sum_{\mathbf{y}\in\mathcal{V}\left( K\right) \backslash\left\{
\mathbf{z}\right\} }\alpha_{\mathbf{y}}\frac{\partial\left. B_{k,E
^{\operatorname*{CR}}\right\vert _{K}}{\partial\mathbf{t}_{\mathbf{y}}}\left(
\mathbf{z}\right) =\frac{\partial\left. B_{k,E}^{\operatorname*{CR
}\right\vert _{K}}{\partial\mathbf{n}_{E}}\left( \mathbf{z}\right) =\left.
\operatorname*{div}\nolimits_{\mathcal{T}}\left( B_{k,E}^{\operatorname*{CR
}\mathbf{n}_{E}\right) \right\vert _{K}\left( \mathbf{z}\right) .\nonumber
\end{align}
\begin{lemma}
\label{Lemtrianglemean}Let $k\geq5$ be odd and for $E\in\mathcal{E}_{\Omega
}\left( \mathcal{T}\right) $, let $B_{k,E}^{\operatorname*{CR}}$ be as in
(\ref{BkEdef}). Then, there exists a function $\tilde{B}_{k,E
^{\operatorname*{CR}}\in\operatorname*{CR}_{k,0}\left( \mathcal{T}\right) $ with
\begin{enumerate}
\item $\operatorname*{supp}\tilde{B}_{k,E}^{\operatorname*{CR}}=\omega_{E},$
\item for all $K\in\mathcal{T}$, for all $\mathbf{z}\in\mathcal{V}\left(
K\right) $
\begin{equation}
\operatorname{div}\left( \left. \tilde{B}_{k,E}^{\operatorname*{CR
}\mathbf{n}_{E}\right\vert _{K}\right) \left( \mathbf{z}\right)
=\operatorname{div}\left( \left. B_{k,E}^{\operatorname*{CR}}\mathbf{n
_{E}\right\vert _{K}\right) \left( \mathbf{z}\right) , \label{neinher
\end{equation}
\item for all $K\in\mathcal{T}_{E}$, for all $E^{\prime}\subset\partial
K\cap\mathcal{\partial\omega}_{E}$
\[
\left. \left. \tilde{B}_{k,E}^{\operatorname*{CR}}\right\vert _{K
\right\vert _{E^{\prime}}=\left. \left. B_{k,E}^{\operatorname*{CR
}\right\vert _{K}\right\vert _{E^{\prime}}\quad\text{and\quad}\left[
\tilde{B}_{k,E}^{\operatorname*{CR}}\right] _{E}=0,
\]
\item for all $K\in\mathcal{T}
\begin{equation}
\int_{K}\operatorname{div}_{\mathcal{T}}\left( \tilde{B}_{k,E
^{\operatorname*{CR}}\mathbf{n}_{E}\right) =0. \label{localintzero
\end{equation}
\item The piecewise gradient is bounded b
\begin{equation}
\left\Vert \nabla_{\mathcal{T}}\tilde{B}_{k,E}^{\operatorname*{CR}}\right\Vert
_{\mathbf{L}^{2}\left( \Omega\right) }\leq Ck^{1/4}. \label{pwgrad
\end{equation}
\end{enumerate}
\end{lemma}
\proof
We employ the reference triangle as in \cite{BabuskaMandel} in order to apply
the polynomial extension theorem therein. Let $\tilde{K}$ be the equilateral
triangle with vertices $\mathbf{\tilde{A}}_{1}:=\left( -1,0\right) $,
$\mathbf{\tilde{A}}_{2}:=\left( 1,0\right) $, $\mathbf{\tilde{A}
_{3}:=\left( 0,\sqrt{3}\right) $ and let $\tilde{E}_{j}$ denote the edge in
$\tilde{K}$ opposite to $\mathbf{\tilde{A}}_{j}$, $1\leq j\leq3$.
Let $E\in\mathcal{E}_{\Omega}\left( \mathcal{T}\right) $ with endpoints
$\mathbf{V}_{1}$, $\mathbf{V}_{2}$, and let $K\in\mathcal{T}_{E}$. Choose an
affine pullback $\phi_{K}:\tilde{K}\rightarrow K$ such that $\phi_{K}\left(
\tilde{E}_{3}\right) =E$.
We employ the function $\tilde{\psi}_{k}^{-}\in\mathbb{P}_{k}\left( \left[
-1,1\right] \right) $ which is analysed in the proof of Lemma A.1
\cite[denoted by $F_{k}$]{Ainsworth_parker_I} and set $\tilde{\psi}_{k
^{+}\left( x\right) :=-\tilde{\psi}_{k}^{-}\left( -x\right) $. These
functions satisfy $\tilde{\psi}_{k}^{\pm}\left( \pm1\right) =\left(
\tilde{\psi}_{k}^{-}\right) ^{\prime}\left( +1\right) =\left( \tilde{\psi
}_{k}^{+}\right) ^{\prime}\left( -1\right) =0$ and $\left( \tilde{\psi
}_{k}^{-}\right) ^{\prime}\left( -1\right) =\left( \tilde{\psi}_{k
^{+}\right) ^{\prime}\left( +1\right) =1$. Their norms can be estimated b
\[
\left\Vert \tilde{\psi}_{k}^{\pm}\right\Vert _{L^{2}\left( \left[
-1,1\right] \right) }\leq Ck^{-3}\quad\text{and\quad}\left\Vert \left(
\tilde{\psi}_{k}^{\pm}\right) ^{\prime}\right\Vert _{L^{2}\left( \left[
-1,1\right] \right) }\leq Ck^{-1}.
\]
We set
\[
\tilde{\varphi}_{k}\left( x\right) =L_{k-1}\left( x\right) -L_{k-1
^{\prime}\left( -1\right) \tilde{\psi}_{k}^{-}\left( x\right)
-L_{k-1}^{\prime}\left( 1\right) \tilde{\psi}_{k}^{+}\left( x\right) .
\]
Clearly, it hold
\[
\tilde{\varphi}_{k}\left( \pm1\right) =1,\quad\tilde{\varphi}_{k}^{\prime
}\left( \pm1\right) =0.
\]
By using Lemma \ref{LemExInt}, we ge
\[
\left\Vert \tilde{\varphi}_{k}\right\Vert _{L^{2}\left( \left[ -1,1\right]
\right) }\leq Ck^{-1/2},\quad\left\Vert \tilde{\varphi}_{k}\right\Vert
_{H^{1}\left( \left[ -1,1\right] \right) }\leq Ck.
\]
From \cite[Thm. 7.4]{BabuskaMandel} we conclude that there is $\tilde{w
_{E}\in\mathbb{P}_{k}\left( \tilde{K}\right) $ with
\[
\left. \tilde{w}_{E}\right\vert _{\tilde{E}_{j}}=\left. \left. \tilde
{B}_{k,E}^{\operatorname*{CR}}\right\vert _{K}\circ\phi_{K}\right\vert
_{\tilde{E}_{j}},\quad j=1,2\quad\text{and\quad}\left. \tilde{w
_{E}\right\vert _{\tilde{E}_{3}}=\tilde{\varphi}_{k
\]
which satisfie
\[
\left\Vert \tilde{w}_{E}\right\Vert _{H^{1}\left( \tilde{K}\right) }\leq
C\left\Vert \tilde{w}_{E}\right\Vert _{H^{1/2}\left( \partial\tilde
{K}\right)
\]
for a constant $C$ independent of $k$. Since $\partial\tilde{K}$ is closed we
can interpolate between the spaces $L^{2}\left( \partial\tilde{K}\right) $
and $H^{1}\left( \partial\tilde{K}\right) $ and obtai
\begin{equation}
\left\Vert \tilde{w}_{E}\right\Vert _{H^{1}\left( \tilde{K}\right) }\leq
C\left\Vert \tilde{w}_{E}\right\Vert _{L^{2}\left( \partial\tilde{K}\right)
}^{1/2}\left\Vert \tilde{w}_{E}\right\Vert _{H^{1}\left( \partial\tilde
{K}\right) }^{1/2}\leq Ck^{1/4}.\label{wtildeest
\end{equation}
By pullback to the triangle $K$ via $\left. w_{E}\right\vert _{K}:=\tilde
{w}_{E}\circ\phi_{K}^{-1}$, $K\in\mathcal{T}_{E}$, we see that this function
is continuous across $E$ (with value $\tilde{\varphi}_{k}\circ\left. \phi
_{K}^{-1}\right\vert _{E}$) and, on $E^{\prime}\subset\partial\omega_{E}$, it
is a lifted Legendre polynomial. This implies property (3) for $w_{E}$.
Outside $\omega_{E}$ we extend $w_{E}$ by zero so that (1) holds for $w_{E}$.
Since the construction implies that the derivative of $w_{E}$ in the direction
of $E$, evaluated at the endpoints $\mathbf{V}_{1}$, $\mathbf{V}_{2}$ of $E$,
is zero, we may apply the reasoning in (\ref{reasoningz}) to obtain property
(2) for $w_{E}$.
From (\ref{wtildeest}) we obtain in a standard wa
\[
\left\Vert \nabla w_{E}\right\Vert _{\mathbf{L}^{2}\left( K\right) }\leq
Ck^{1/4}.
\]
Next, we modify $w_{E}$ such that property (4) holds without affecting the
other properties. Let $\psi_{E}\in S_{4,0}\left( \mathcal{T}\right) $ with
$\operatorname*{supp}\psi_{E}=\omega_{E}$ and
\begin{equation}
\left. \psi_{E}\right\vert _{K}:=\alpha_{K}\lambda_{K,\mathbf{V}_{1}
^{2}\lambda_{K,\mathbf{V}_{2}}^{2}\quad\text{with\quad}\alpha_{K}:=\left(
\int_{K}\partial_{\mathbf{n}_{E}}w_{E}\right) /\left( \int_{K
\partial_{\mathbf{n}_{E}}\left( \lambda_{K,\mathbf{V}_{1}}^{2}\lambda
_{K,\mathbf{V}_{2}}^{2}\right) \right) \quad\forall K\in\mathcal{T
_{E}.\label{DefPsieta
\end{equation}
The modified function $\tilde{B}_{k,E}^{\operatorname*{CR}}$ finally is
defined b
\begin{equation}
\tilde{B}_{k,E}^{\operatorname*{CR}}=w_{E}-\psi_{E}.\label{defboldpsi
\end{equation}
Since $\left. \operatorname*{div}_{\mathcal{T}}\left( \tilde{B
_{k,E}^{\operatorname*{CR}}\mathbf{n}_{E}\right) \right\vert _{K
=\partial_{\mathbf{n}_{E}}\left( \tilde{B}\left. _{k,E}^{\operatorname*{CR
}\right\vert _{K}\right) $ property (4) follows by construction. The gradient
$\nabla_{\mathcal{T}}\psi_{E}$ vanishes in the vertices of $K$ so that
$\left( \partial_{\mathbf{n}_{E}}\psi_{E}\right) \left( \mathbf{z}\right)
=0$ for all $\mathbf{z}\in\mathcal{V}\left( K\right) $ and (2) is inherited
from $w_{E}$. Properties (1), (3) are obvious. Next, we verify (5). Let
$\mathbf{V}_{3}$ denote the vertex in $K$ opposite to $E$. We first comput
\begin{align*}
\int_{K}\partial_{\mathbf{n}_{E}}\left( \lambda_{K,\mathbf{V}_{1}}^{2
\lambda_{K,\mathbf{V}_{2}}^{2}\right) & =\sum_{j=1}^{2}2\partial
_{\mathbf{n}_{E}}\lambda_{K,\mathbf{V}_{j}}\int_{K}\lambda_{K,\mathbf{V}_{1
}\lambda_{K,\mathbf{V}_{2}}\lambda_{K,\mathbf{V}_{3-j}}=\int_{K
\lambda_{K,\mathbf{V}_{1}}^{2}\lambda_{K,\mathbf{V}_{2}}\sum_{j=1
^{2}2\partial_{\mathbf{n}_{E}}\lambda_{K,\mathbf{V}_{j}}\\
& =-\frac{1}{15}\partial_{\mathbf{n}_{E}}\lambda_{K,\mathbf{V}_{3}}\left\vert
K\right\vert \overset{\text{(\ref{normcomp})}}{=}\frac{\left\vert E\right\vert
}{30},\\
\left\vert \int_{K}\partial_{\mathbf{n}_{E}}w_{E}\right\vert &
\leq\left\vert K\right\vert ^{1/2}\left\Vert \nabla w_{E}\right\Vert
_{\mathbf{L}^{2}\left( K\right) }\leq C\left\vert K\right\vert ^{1/2
k^{1/4}.
\end{align*}
In this way, $\left\vert \alpha_{K}\right\vert \leq Ck^{1/4}$ and an inverse
inequality for quartic polynomials gives u
\[
\left\Vert \nabla\psi_{E}\right\Vert _{\mathbf{L}^{2}\left( K\right) }\leq
Ch_{K}^{-1}\left\Vert \psi_{E}\right\Vert _{L^{2}\left( K\right) }\leq
Ch_{K}^{-1}k^{1/4}\left\Vert 1\right\Vert _{L^{2}\left( K\right) }\leq
Ck^{1/4}.
\]
Hence, property (5) follows
\endproof
Next we recall a result which goes back to Vogelius \cite{vogelius1983right}
and Scott-Vogelius \cite{ScottVogelius}, see also \cite[Proof of Thm.
1]{GuzmanScott2019}.
\begin{definition}
\label{DefAz}Let $\eta_{0}$ be as in Lemma \ref{Lemangle}. For $0\leq\eta
<\eta_{0}$, the subspace $M_{\eta,k-1}^{\operatorname*{SV}}\left(
\mathcal{T}\right) $ of the pressure space $M_{k-1}\left( \mathcal{T
\right) $ is given b
\begin{equation}
M_{\eta,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) :=\left\{ q\in
M_{k-1}\left( \mathcal{T}\right) \mid\forall\mathbf{z}\in\mathcal{C
_{\mathcal{T}}\left( \eta\right) :A_{\mathcal{T},\mathbf{z}}\left(
q\right) =0\right\} , \label{defMSV
\end{equation}
where, for $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left( \eta\right) $, the
functional $A_{\mathcal{T},\mathbf{z}}\left( q\right) $ is as follows:
number the triangles in the patch $\mathcal{T}_{\mathbf{z}}$ counterclockwise
$K_{\ell}$, $1\leq\ell\leq m$, and set\footnote{Note that for $\left\vert
\mathcal{T}_{\mathbf{z}}\right\vert \geq2$ the definition of $A_{\mathcal{T
,\mathbf{z}}$ is unique only up to a multiplication by a factor $\left(
-1\right) $. This can be fixed by a numbering convention: $\mathfrak{E
\left( \mathbf{z}\right) =K_{1}\cap K_{2}$.
\begin{equation}
A_{\mathcal{T},\mathbf{z}}\left( q\right) =\sum_{\ell=1}^{m}\left(
-1\right) ^{\ell}\left( \left. q\right\vert _{K_{\ell}}\right) \left(
\mathbf{z}\right) . \label{DefATz
\end{equation}
\end{definition}
Note that $M_{0,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) $ is the
pressure space introduced by Vogelius \cite{vogelius1983right} and
Scott-Vogelius \cite{ScottVogelius} and the following inclusions hold: for
$0\leq\eta\leq\eta^{\prime}\leq\eta_{0}
\[
M_{\eta^{\prime},k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) \subset
M_{\eta,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) \subset
M_{0,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) =Q_{h}^{k-1
\]
with the pressure space $Q_{h}^{k-1}$ in \cite[p. 517]{GuzmanScott2019}.
For the Scott-Vogelius pressure space $M_{0,k-1}^{\operatorname*{SV}}\left(
\mathcal{T}\right) $, the existence of a continuous right-inverse of the
divergence operator into $\mathbf{S}_{k,0}\left( \mathcal{T}\right) $ was
proved in Vogelius \cite{vogelius1983right} and Scott-Vogelius
\cite{ScottVogelius}.
\begin{proposition}
[Scott-Vogelius]For any $p\in M_{0,k-1}^{\operatorname*{SV}}\left(
\mathcal{T}\right) $ there exists some $\mathbf{v}\in\mathbf{S}_{k,0}\left(
\mathcal{T}\right) $ such tha
\[
\operatorname*{div}\mathbf{v}=q\mathbf{\quad}\text{and\quad}\left\Vert
\mathbf{v}\right\Vert _{\mathbf{H}^{1}\left( \Omega\right) }\leq C\left\Vert
q\right\Vert _{L^{2}\left( \Omega\right) },
\]
for a constant which only depends on the shape-regularity of the mesh, the
polynomial degree $k$, and on $\Theta_{\min}^{-1}$, wher
\begin{equation}
\Theta_{\min}:=\min_{\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right)
\backslash\mathcal{C}_{\mathcal{T}}}\Theta\left( \mathbf{z}\right) .
\label{thetamin
\end{equation}
In particular, the constant $C$ is independent of $h$.
\end{proposition}
In Lemma \ref{LemProject}, we show that, by subtracting the divergence of a
suitable Crouzeix-Raviart velocity from a given pressure in $M_{k-1}\left(
\mathcal{T}\right) $, the resulting modified pressure belongs to the reduced
pressure space $M_{\eta,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right)
$. As a preliminary, we need a bound of the functional $A_{\mathcal{T
,\mathbf{z}}$ in (\ref{DefATz}) which is explicit with respect to the local
mesh size and polynomial degree.
\begin{lemma}
\label{LemFunctional}There exists a constant $C$ which only depends on the
shape-regularity of the mesh such tha
\[
\left\vert A_{\mathcal{T},\mathbf{z}}\left( q\right) \right\vert \leq
C\frac{k^{2}}{h_{\mathbf{z}}}\left\Vert q\right\Vert _{L^{2}\left(
\omega_{\mathbf{z}}\right) }\quad\forall q\in\mathbb{P}_{k-1}\left(
\mathcal{T}\right)
\]
for any $k\in\mathbb{N}$.
\end{lemma}
\proof
Let $\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) $ and $K\in
\mathcal{T}_{\mathbf{z}}$. The affine pullback to the reference triangle is
denoted by $\chi_{K}:\widehat{K}\rightarrow K$. For $q\in\mathbb{P}_{k}\left(
K\right) $, let $\widehat{q}:=q\circ\chi_{K}$ and $\mathbf{\hat{z}}:=\chi
_{K}^{-1}\left( \mathbf{z}\right) $. The
\begin{equation}
\left\vert q\left( \mathbf{z}\right) \right\vert =\left\vert \widehat{q
\left( \widehat{\mathbf{z}}\right) \right\vert
\overset{\text{\cite{Hesthaven_trace_2003}, \cite[Lem. 6.1
{Ainsworth_Jiang_hp}}}{\leq}\frac{\left( k+1\right) \left( k+2\right)
}{\sqrt{2}}\left\Vert \widehat{q}\right\Vert _{L^{2}\left( \widehat{K
\right) }=\binom{k+2}{2}\left\vert K\right\vert ^{-1/2}\left\Vert
q\right\Vert _{L^{2}\left( K\right) }. \label{estfunctform
\end{equation}
A summation over all $K\in\mathcal{T}_{\mathbf{z}}$ leads t
\[
\left\vert A_{\mathcal{T},\mathbf{z}}\left( q\right) \right\vert \leq
\binom{k+2}{2}\sum_{\ell=1}^{m}\left\vert K_{\ell}\right\vert ^{-1/2
\left\Vert q\right\Vert _{L^{2}\left( K_{\ell}\right) }\leq C\binom{k+2
{2}h_{\mathbf{z}}^{-1}\left\Vert q\right\Vert _{L^{2}\left( \omega
_{\mathbf{z}}\right) },
\]
where $C$ only depends on the shape-regularity of the mesh
\endproof
The following lemma shows that the non-conforming Crouzeix-Raviart elements
allow us to modify a general pressure $q\in\mathbb{P}_{k-1,0}\left(
\mathcal{T}\right) $ in such a way that the result belongs to $M_{\eta
,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) $ \textit{provided
}$\mathcal{C}_{\mathcal{T}}^{\operatorname{acute}}=\emptyset$.
\begin{lemma}
\label{LemProject}Let assumption (\ref{Assumptab}) be satisfied and
$\mathcal{C}_{\mathcal{T}}^{\operatorname{acute}}=\emptyset$. There exists a
constant $\eta_{2}>0$ which only depends on the shape-regularity of the mesh
such that for any fixed $0\leq\eta<\eta_{2}$ and any $q\in\mathbb{P
_{k-1,0}\left( \mathcal{T}\right) $, there exists some $\mathbf{v}_{q
\in\mathbf{CR}_{k,0}\left( \mathcal{T}\right) $ such tha
\begin{align}
\int_{K}\operatorname{div}\mathbf{v}_{q} & =0\quad\forall K\in
\mathcal{T},\label{trianglemean}\\
q-\operatorname*{div}\nolimits_{\mathcal{T}}\mathbf{v}_{q} & \in
M_{\eta,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right)
\label{inclusionMV
\end{align}
an
\begin{equation}
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{q}\right\Vert _{\mathbb{L
^{2}\left( \Omega\right) }\leq C_{\operatorname*{CR}}k^{1/4}\left\Vert
q\right\Vert _{L^{2}\left( \Omega\right) }. \label{CRlifting1
\end{equation}
The constant $C_{\operatorname*{CR}}$ depends only on the shape-regularity of
the mesh.
\end{lemma}
\proof
Let $q\in\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) $. Let the fans
$\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) $, $\ell\in\mathcal{J}$,
be as in Definition \ref{DefFan}. For each fan $\mathcal{C}_{\mathcal{T},\ell
}\left( \eta\right) $ we employ an ansat
\begin{equation}
\mathbf{v}_{\ell}:=\sum_{\mathbf{z}\in\mathcal{C}_{\mathcal{T},\ell}\left(
\eta\right) }\alpha_{\ell,\mathbf{z}}\tilde{B}_{k,\mathfrak{E}\left(
\mathbf{z}\right) }^{\operatorname*{CR}}\mathfrak{N}\left( \mathbf{z
\right) \label{defvl
\end{equation}
for $\tilde{B}_{k,\mathfrak{E}\left( \mathbf{z}\right) }^{\operatorname*{CR
}$ as in (\ref{defboldpsi}), where the coefficients $\alpha_{\ell,\mathbf{z
}\in\mathbb{R}$ are defined next. The global function $\mathbf{v}_{q}$ is then
given b
\[
\mathbf{v}_{q}=\sum_{\ell=1}^{N}\mathbf{v}_{\ell}.
\]
Property (\ref{trianglemean}) follows from this ansatz by using Lemma
\ref{Lemtrianglemean}. Next, we define the coefficients $\alpha_{\ell
,\mathbf{z}}$ in (\ref{defvl}) such that (\ref{inclusionMV}) holds and prove
the norm estimates for $\mathbf{v}_{q}$. Our construction of the fans imply
that open interiors of the supports of $\mathbf{v}_{\ell}$ are pairwise
disjoint (see Lem. \ref{Lemomegaell}); as a consequence the definition of
$\left( \alpha_{\ell,\mathbf{z}}\right) _{\mathbf{z}\in\mathcal{C
_{\mathcal{T}}\left( \eta\right) }$ and the estimate of $\nabla
_{\mathcal{T}}\mathbf{v}_{q}$ can be performed for each fan separately.
For $\mathbf{z}\in\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) $, let
$E:=\mathfrak{E}\left( \mathbf{z}\right) $, $\mathbf{n}_{E}:=\mathfrak{N
\left( \mathbf{z}\right) $. Let $K_{\mathbf{z}}^{-}$, $K_{\mathbf{z}}^{+}$,
denote the triangles in $\mathcal{T}_{E}$ with the convention that
$\mathbf{n}_{E}$ points into $K_{\mathbf{z}}^{+}$. We us
\[
\operatorname*{div}\left( \left. \tilde{B}_{k,E}^{\operatorname*{CR
}\mathbf{n}_{E}\right\vert _{K}\right) \left( \mathbf{V}\right)
\overset{\text{(\ref{neinher})}}{=}\operatorname*{div}\left( \left.
B_{k,E}^{\operatorname*{CR}}\mathbf{n}_{E}\right\vert _{K}\right) \left(
\mathbf{V}\right) \quad\forall\mathbf{V}\in\mathcal{V}\left( K\right)
\]
and compute the divergence of $B_{k,E}^{\operatorname*{CR}}\mathbf{n}_{E}
\begin{equation}
\operatorname*{div}\left( \left. B_{k,E}^{\operatorname*{CR}}\mathbf{n
_{E}\right\vert _{K}\right) =\left( -1\right) ^{i+1}\left\{
\begin{array}
[c]{ll
\frac{\left\vert E\right\vert }{\left\vert K\right\vert }L_{k}^{\prime}\left(
1-2\lambda_{K,E}\right) & \text{on }K=K_{i},\quad i=1,2,\\
0 & \text{otherwise.
\end{array}
\right. \label{divrep
\end{equation}
Well-known properties of Legendre polynomials applied to (\ref{divrep}) imply
that for any vertex $\mathbf{V}$ of $K$ and odd polynomial degree $k
\begin{equation}
\operatorname*{div}\left( \left. \tilde{B}_{k,E}^{\operatorname*{CR
}\mathbf{n}_{E}\right\vert _{K}\right) \left( \mathbf{V}\right)
=-\binom{k+1}{2}\times\left\{
\begin{array}
[c]{ll
\sigma\frac{\left\vert E\right\vert }{\left\vert K\right\vert } &
\forall\mathbf{V}\in\mathcal{V}\left( K\right) ,\quad\text{if
K=K_{\mathbf{z}}^{\sigma}\quad\sigma\in\left\{ +,-\right\} \\
0 & \text{otherwise.
\end{array}
\right. \label{divcomp
\end{equation}
Hence the condition $A_{\mathcal{T},\mathbf{y}}\left( q-\operatorname*{div
_{\mathcal{T}}\mathbf{v}_{\ell}\right) =0$ for all $\mathbf{y}\in
\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) $ is equivalent to the
system of linear equatio
\begin{equation}
\mathbf{M}_{\ell
\mbox{\boldmath$ \alpha$
_{\ell}=\mathbf{r}_{\ell} \label{defalphaell
\end{equation}
wit
\begin{equation}
\mathbf{M}_{\ell}:=\left( A_{\mathcal{T},\mathbf{y}}\left(
\operatorname*{div}\nolimits_{\mathcal{T}}\left( B_{k,\mathfrak{E}\left(
\mathbf{z}\right) }^{\operatorname*{CR}}\mathfrak{N}\left( \mathbf{z
\right) \right) \right) \right) _{\substack{\mathbf{y}\in\mathcal{C
_{\mathcal{T},\ell}\left( \eta\right) \\\mathbf{z}\in\mathcal{C
_{\mathcal{T},\ell}\left( \eta\right) }},\qua
\mbox{\boldmath$ \alpha$
_{\ell}:=\left( \alpha_{\ell,\mathbf{z}}\right) _{\mathbf{z}\in
\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) },\quad\mathbf{r}_{\ell
}:=\left( A_{\mathcal{T},\mathbf{y}}\left( q\right) \right) _{\mathbf{y
\in\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) }. \label{DefMl
\end{equation}
The matrix $\mathbf{M}_{\ell}$ is explicitly given b
\[
\mathbf{M}_{\ell}:=-\binom{k+1}{2}\left[
\begin{array}
[c]{ccccc
\frac{\left\vert E_{\ell,1}\right\vert }{\left\vert K_{\ell,1}\right\vert
}+\frac{\left\vert E_{\ell,1}\right\vert }{\left\vert K_{\ell,2}\right\vert }
& \frac{\left\vert E_{\ell,2}\right\vert }{\left\vert K_{\ell,2}\right\vert }
& 0 & \ldots & 0\\
\frac{\left\vert E_{\ell,1}\right\vert }{\left\vert K_{\ell,2}\right\vert } &
\frac{\left\vert E_{\ell,2}\right\vert }{\left\vert K_{\ell,2}\right\vert
}+\frac{\left\vert E_{\ell,2}\right\vert }{\left\vert K_{\ell,3}\right\vert }
& \ddots & \ddots & \vdots\\
0 & \ddots & \ddots & & 0\\
\vdots & \ddots & & & \frac{\left\vert E_{\ell,n_{\ell}}\right\vert
}{\left\vert K_{\ell,n_{\ell}}\right\vert }\\
0 & \ldots & 0 & \frac{\left\vert E_{\ell,n_{\ell}-1}\right\vert }{\left\vert
K_{\ell,n_{\ell}}\right\vert } & \frac{\left\vert E_{\ell,n_{\ell}}\right\vert
}{\left\vert K_{\ell,n_{\ell}}\right\vert }+\frac{\left\vert E_{\ell,n_{\ell
}\right\vert }{\left\vert K_{\ell,n_{\ell+1}}\right\vert
\end{array}
\right] .
\]
We use (cf. Fig. \ref{Figfan2}
\begin{figure}[ptb
\centering
\includegraphics[
height=2.2805in,
width=4.2004in
{fan2.eps
\caption{Local numbering convention of the angles in $K_{\ell,j}$ and
$K_{\ell,j+1}$. The angle in $K_{\ell,j}$ at $\mathbf{z}_{\ell}$ is denoted by
$\alpha_{\ell,j,1}$, at $\mathbf{z}_{\ell,j-1}$ by $\alpha_{\ell,j,2}$, at
$\mathbf{z}_{\ell,j}$ by $\alpha_{\ell,j,3}$ and in $K_{\ell,j+1}$
accordingly.
\label{Figfan2
\end{figure}
\[
\frac{\left\vert E_{\ell,j}\right\vert }{\left\vert K_{\ell,j}\right\vert
}=\frac{2\sin\left( \alpha_{\ell,j,1}+\alpha_{\ell,j,2}\right) }{\left\vert
E_{\ell,j}\right\vert \sin\alpha_{\ell,j,1}\sin\alpha_{\ell,j,2}
\quad\text{and\quad}\frac{\left\vert E_{\ell,j}\right\vert }{\left\vert
K_{\ell,j+1}\right\vert }=\frac{2\sin\left( \alpha_{\ell,j+1,1}+\alpha
_{\ell,j+1,3}\right) }{\left\vert E_{\ell,j}\right\vert \sin\alpha
_{\ell,j+1,1}\sin\alpha_{\ell,j+1,3}
\]
and obtain
\begin{equation}
\mathbf{M}_{\ell}=\mathbf{D}_{\ell}\left( \mathbf{T}_{\ell}
\mbox{\boldmath$ \Delta$
_{\ell}\right) \label{DefMlTlDl
\end{equation}
with $\mathbf{D}_{\ell}=-k\left( k+1\right) \operatorname*{diag}\left[
\left\vert E_{\ell,j}\right\vert ^{-1}:1\leq j\leq n_{\ell}\right] $ an
\begin{align}
\mathbf{T}_{\ell} & :=\left[
\begin{array}
[c]{ccccc
\frac{\sin\left( \alpha_{\ell,1,1}+\alpha_{\ell,2,1}\right) }{\sin
\alpha_{\ell,1,1}\sin\alpha_{\ell,2,1}} & \frac{1}{\sin\alpha_{\ell,2,1}} &
0 & \ldots & 0\\
\frac{1}{\sin\alpha_{\ell,2,1}} & \frac{\sin\left( \alpha_{\ell,2,1
+\alpha_{\ell,3,1}\right) }{\sin\alpha_{\ell,2,1}\sin\alpha_{\ell,3,1}} &
\ddots & \ddots & \vdots\\
0 & \ddots & \ddots & & 0\\
\vdots & \ddots & & & \frac{1}{\sin\alpha_{\ell,n_{\ell},1}}\\
0 & \ldots & 0 & \frac{1}{\sin\alpha_{\ell,n_{\ell},1}} & \frac{\sin\left(
\alpha_{\ell,n_{\ell},1}+\alpha_{\ell,n_{\ell}+1,1}\right) }{\sin\alpha
_{\ell,n_{\ell},1}\sin\alpha_{\ell,n_{\ell}+1,1}
\end{array}
\right] ,\label{defTl}\\%
\mbox{\boldmath$ \Delta$
_{\ell} & :=\operatorname*{diag}\left[ \frac{\sin\left( \alpha_{\ell
,j,2}+\alpha_{\ell,j+1,3}\right) }{\sin\alpha_{\ell,j,2}\sin\alpha
_{\ell,j+1,3}}:1\leq j\leq n_{\ell}\right] .\nonumber
\end{align}
In Lemma \ref{Tlest}, we prove that the matrix $\mathbf{T}_{\ell}
\mbox{\boldmath$ \Delta$
_{\ell}$ is invertible and the inverse is bounded by a constant independent of
$h_{\mathcal{T}}$ and $k$. Hence
\begin{equation}
\left\Vert
\mbox{\boldmath$ \alpha$
_{\ell}\right\Vert \leq\tilde{C}\frac{h_{\mathbf{z}_{\ell}}}{k\left(
k+1\right) }\left\Vert \mathbf{r}_{\ell}\right\Vert . \label{estboldalpha
\end{equation}
Let $\mathcal{T}_{\ell}:=\left\{ K_{\ell,j}:1\leq j\leq n_{\ell}+1\right\} $
and $D_{\ell}:=\operatorname*{dom}\left( \mathcal{T}_{\ell}\right) $. We
estimate the function $\mathbf{v}_{\ell}$ in (\ref{defvl}) b
\begin{align*}
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{\ell}\right\Vert _{\mathbb{L
^{2}\left( D_{\ell}\right) } & \leq\left( \sum_{\mathbf{z}\in
\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) }\left\vert \alpha
_{\ell,\mathbf{z}}\right\vert ^{2}\left\Vert \nabla_{\mathcal{T}}\tilde
{B}_{k,\mathfrak{E}\left( \mathbf{z}\right) }^{\operatorname*{CR
}\mathfrak{N}\left( \mathbf{z}\right) \right\Vert _{\mathbb{L}^{2}\left(
D_{\ell}\right) }^{2}\right) ^{1/2}\\
& \leq\max_{\mathbf{z}\in\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right)
}\left\Vert \nabla_{\mathcal{T}}\tilde{B}_{k,\mathfrak{E}\left(
\mathbf{z}\right) }^{\operatorname*{CR}}\mathfrak{N}\left( \mathbf{z
\right) \right\Vert _{\mathbb{L}^{2}\left( D_{\ell}\right) }\left\Vert
\mbox{\boldmath$ \alpha$
_{\ell}\right\Vert \\
& \overset{\text{(\ref{pwgrad})}}{\leq}\tilde{C}k^{1/4}\frac{h_{\mathbf{z
_{\ell}}}{k\left( k+1\right) }\left\Vert \mathbf{r}_{\ell}\right\Vert \leq
Ch_{\mathbf{z}_{\ell}}k^{-7/4}\left\Vert \mathbf{r}_{\ell}\right\Vert .
\end{align*}
In both estimates the constant $C$ only depends on the shape-regularity of the
mesh. We use Lemma \ref{LemFunctional} and conclude tha
\[
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{\ell}\right\Vert _{\mathbb{L
^{2}\left( D_{\ell}\right) }\leq Ck^{1/4}\left\Vert q\right\Vert
_{L^{2}\left( D_{\ell}\right) }\text{.
\]
Since the interiors of the supports $D_{\ell}$ have pairwise empty
intersection the estimat
\[
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{q}\right\Vert _{\mathbb{L
^{2}\left( \Omega\right) }\leq Ck^{1/4}\left\Vert q\right\Vert
_{L^{2}\left( \Omega\right)
\]
follows
\endproof
\begin{definition}
\label{DefPiCR}Let assumption (\ref{Assumptab}) be satisfied and $\eta_{2}>0$
as in Lemma \ref{LemProject}. Fix $\eta\in\left] 0,\eta_{2}\right[ $. For
$q\in\mathbb{P}_{k,0}\left( \mathcal{T}\right) $, the linear map $\Pi
_{k}^{\operatorname*{CR}}:\mathbb{P}_{k-1,0}\left( \mathcal{T}\right)
\rightarrow\mathbf{CR}_{k,0}\left( \mathcal{T}\right) $ is given b
\[
\Pi_{k}^{\operatorname*{CR}}q:=\sum_{\ell\in\mathcal{J}}\sum_{\mathbf{z
\in\mathcal{C}_{\mathcal{T},\ell}\left( \eta\right) }\alpha_{\ell
,\mathbf{z}}\tilde{B}_{k,\mathfrak{E}\left( \mathbf{z}\right)
^{\operatorname*{CR}}\mathfrak{N}\left( \mathbf{z}\right)
\]
with
\mbox{\boldmath$ \alpha$
_{\ell}$ as in (\ref{defalphaell}).
\end{definition}
For the proof of the following lemma we recall the definitions of some linear
maps from the literature which are related to the right inverse of the
divergence acting on some polynomial spaces.
Bernardi and Raugel introduced in \cite[Lem. II.4]{Bernardi_Raugel_1985} a
linear mapping $\Pi^{\operatorname*{BR}}:M_{k-1}\left( \mathcal{T}\right)
\rightarrow\mathbf{S}_{2,0}\left( \mathcal{T}\right) $ with the property:
for any $q\in M_{k-1}\left( \mathcal{T}\right) $, the function
$\Pi^{\operatorname*{BR}}q\in\mathbf{S}_{2,0}\left( \mathcal{T}\right) $
satisfie
\begin{equation}
\int_{K}q=\int_{K}\operatorname*{div}\left( \Pi^{\operatorname*{BR}}q\right)
\mathbf{\quad}\forall K\in\mathcal{T} \label{defPiBR
\end{equation}
and
\begin{equation}
\left\Vert \Pi^{\operatorname*{BR}}q\right\Vert _{\mathbf{H}^{1}\left(
\Omega\right) }\leq C_{\operatorname*{BR}}\left\Vert q\right\Vert
_{L^{2}\left( \Omega\right) } \label{CBR
\end{equation}
for a constant $C_{\operatorname*{BR}}$ which is independent of the mesh width
and the polynomial degree.
Next we consider some right inverse of the divergence operator on the spac
\[
M_{k-1}^{\operatorname*{V}}\left( \mathcal{T}\right) :=\left\{ q\in
M_{k-1}\left( \mathcal{T}\right) \mid\left(
\begin{array}
[c]{ll
\int_{K}q=0 & \forall K\in\mathcal{T}\\
\left. q\right\vert _{K}\left( \mathbf{A}\right) =0 & \forall
\mathbf{A}\text{ vertex of }K
\end{array}
\right) \right\} .
\]
Le
\[
\mathbf{S}_{k}^{\operatorname*{V}}\left( \mathcal{T}\right) :=\left\{
\mathbf{u}\in\mathbf{S}_{k}\left( \mathcal{T}\right) \mid\forall
K\in\mathcal{T}:\quad\left. \mathbf{u}\right\vert _{\partial K
=\mathbf{0}\right\} .
\]
There exists a linear operator $\Pi^{\operatorname*{V}}:M_{k-1
^{\operatorname*{V}}\left( \mathcal{T}\right) \rightarrow\mathbf{S
_{k}^{\operatorname*{V}}\left( \mathcal{T}\right) $ such that for all $q\in
M_{k-1}^{\operatorname*{V}}\left( \mathcal{T}\right)
\begin{align}
q & =\operatorname*{div}\Pi^{\operatorname*{V}}q,\nonumber\\
\left\Vert \Pi^{\operatorname*{V}}q\right\Vert _{\mathbf{H}^{1}\left(
\Omega\right) } & \leq C_{\operatorname*{V}}\left\Vert q\right\Vert
_{L^{2}\left( \Omega\right) }, \label{optests
\end{align}
where the constant $C_{\operatorname*{V}}$ is independent of the mesh width
and the polynomial degree. Note that in the original paper \cite[Lem.
2.5]{vogelius1983right} by M. Vogelius, the right-hand side in the estimate
(\ref{optests}) contains an additional factor $k^{\beta_{\operatorname*{V}}}$
for some positive $\beta_{\operatorname*{V}}$ (independent of the mesh width).
In \cite[Thm. 3.4]{Ainsworth_parker_I}, the operator in \cite[Lem.
2.5]{vogelius1983right} is modified and the estimate in the form
(\ref{optests}) is proved for the modified operator.
Finally, we employ the linear operator $\Pi^{\operatorname*{GS}}:M_{\eta
,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) \rightarrow
\mathbf{S}_{4,0}\left( \mathcal{T}\right) $ introduced by Guzm\'{a}n and
Scott in \cite[Proof of Lem. 6 and Lem. 7]{GuzmanScott2019} with the property
that, for any $q\in M_{\eta,k-1}^{\operatorname*{SV}}\left( \mathcal{T
\right) $, it hold
\begin{subequations}
\label{Thetamin_eta}
\end{subequations
\begin{align}
\left( I-\operatorname*{div}\Pi^{\operatorname*{GS}}\right) q & \in
M_{k-1}^{\operatorname*{V}}\left( \mathcal{T}\right) ,\tag
\ref{Thetamin_eta
a}\label{Thetamin_etaa}\\
\left\Vert \nabla\Pi^{\operatorname*{GS}}q\right\Vert _{\mathbb{L}^{2}\left(
\Omega\right) } & \leq C_{\operatorname*{GS}}k^{\kappa}\eta^{-1}\left\Vert
q\right\Vert _{L^{2}\left( \Omega\right) }\quad\text{for }\kappa=2. \tag
\ref{Thetamin_eta
b}\label{Thetamin_etac
\end{align}
We emphasize that in \cite[Lemma 7]{GuzmanScott2019} the constant
$\Theta_{\min}^{-1}$ (cf. (\ref{thetamin})) instead of $\eta^{-1}$ appears in
(\ref{Thetamin_etac}) so that the estimate of $\left\Vert \nabla
\Pi^{\operatorname*{GS}}q\right\Vert _{\mathbb{L}^{2}\left( \Omega\right) }$
for $q\in M_{0,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right) $
deteriorates in cases where the $\mathbf{z}$ is a \textit{nearly singular}
point, i.e., very close to the geometric situations described in Remark
\ref{RemCritGeom}. The proof of \cite[Lemma 7]{GuzmanScott2019} is split into
an estimate related to points $\mathbf{z}$ with $A_{\mathcal{T},\mathbf{z
}\left( q\right) =0$ (cf. (\ref{DefATz})) and an estimate for the remaining
points $\mathbf{z}$ with $A_{\mathcal{T},\mathbf{z}}\left( q\right) \neq0$.
Only in this second part, the constant $\Theta_{\min}^{-1}$ is involved. From
Lemma \ref{LemProject}, we conclude that $q-\operatorname*{div}_{\mathcal{T
}\left( \Pi_{k}^{\operatorname*{CR}}q\right) \in M_{\eta,k-1
^{\operatorname*{SV}}\left( \mathcal{T}\right) $ and the second part of the
proof in \cite[Lemma 7]{GuzmanScott2019} is applied only to points wit
\[
\min_{\mathbf{z}\in\mathcal{V}\left( \mathcal{T}\right) \backslash
\mathcal{C}_{\mathcal{T}}\left( \eta\right) }\Theta\left( \mathbf{z
\right) \geq\eta.
\]
Hence, (\ref{Thetamin_etac}) follows for $\eta$ depending only on the
shape-regularity of the mesh.
\begin{lemma}
\label{LemmaFinOdd}Let assumption (\ref{Assumptab}) be satisfied and let
$\mathcal{C}_{\mathcal{T}}^{\operatorname{acute}}=\emptyset$. There exists a
constant $\eta_{2}>0$ which only depends on the shape-regularity of the mesh
such that for any fixed $0\leq\eta<\eta_{2}$ and any $q\in\mathbb{P
_{k-1,0}\left( \mathcal{T}\right) $, there exists some $\mathbf{w}_{q
\in\mathbf{CR}_{k,0}\left( \mathcal{T}\right) $ such tha
\[
q=\operatorname*{div}\nolimits_{\mathcal{T}}\mathbf{w}_{q
\]
an
\[
\left\Vert \mathbf{w}_{q}\right\Vert _{\mathbf{H}^{1}\left( \mathcal{T
\right) }\leq Ck^{1/4+\kappa}\eta^{-1}\left\Vert q\right\Vert _{L^{2}\left(
\Omega\right) }.
\]
The constant $C$ only depends on the shape-regularity of the mesh and is
independent of the mesh width and the polynomial degree $k$.
\end{lemma}
\proof
For the construction of $\mathbf{w}_{q}$ we follow and modify the lines of
proof in \cite[Thm. 1]{GuzmanScott2019} by a) involving the operator $\Pi
_{k}^{\operatorname*{CR}}$ and b) employing the concept of $\eta$-critical points.
For given $q\in\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) $, we employ the
operators $\Pi_{k}^{\operatorname*{CR}}$, $\Pi^{\operatorname*{BR}}$,
$\Pi^{\operatorname*{GS}}$, $\Pi^{\operatorname*{V}}$ in the definition of the
function $\mathbf{w}_{q}$
\begin{align}
\mathbf{w}_{q} & =\mathbf{T}_{1}+\mathbf{T}_{2}+\mathbf{T}_{3
+\mathbf{T}_{3},\label{defwq}\\
\mathbf{T}_{1} & :=\Pi^{\operatorname*{BR}}q,\nonumber\\
\mathbf{T}_{2} & :=\Pi_{k}^{\operatorname*{CR}}\left( I-\operatorname*{div
\Pi^{\operatorname*{BR}}\right) q,\nonumber\\
\mathbf{T}_{3} & :=\Pi^{\operatorname*{GS}}\left( I-\operatorname*{div
\nolimits_{\mathcal{T}}\Pi_{k}^{\operatorname*{CR}}\right) \left(
I-\operatorname*{div}\Pi^{\operatorname*{BR}}\right) q,\nonumber\\
\mathbf{T}_{4} & :=\Pi^{\operatorname*{V}}\left( I-\operatorname*{div
\Pi^{\operatorname*{GS}}\right) \left( I-\operatorname*{div
\nolimits_{\mathcal{T}}\Pi_{k}^{\operatorname*{CR}}\right) \left(
I-\operatorname*{div}\Pi^{\operatorname*{BR}}\right) q.\nonumber
\end{align}
By construction we hav
\[
\operatorname{div}_{\mathcal{T}}\mathbf{w}_{q}=q.
\]
The first two summands in (\ref{defwq}) satisf
\begin{align}
\left\Vert \mathbf{T}_{1}\right\Vert _{\mathbf{H}^{1}\left( \Omega\right) }
& \overset{\text{(\ref{CBR})}}{\leq}C_{\operatorname*{BR}}\left\Vert
q\right\Vert _{L^{2}\left( \Omega\right) },\label{estPi1}\\
\left\Vert \mathbf{T}_{2}\right\Vert _{\mathbf{H}^{1}\left( \mathcal{T
\right) } & \overset{\text{(\ref{CRlifting1})}}{\leq}C_{\operatorname*{CR
}k\left( \left\Vert q\right\Vert _{L^{2}\left( \Omega\right) }+\left\Vert
\Pi^{\operatorname*{BR}}q\right\Vert _{\mathbf{H}^{1}\left( \Omega\right)
}\right) \overset{\text{(\ref{estPi1})}}{\leq}C_{\operatorname*{CR}}\left(
1+C_{\operatorname*{BR}}\right) k\left\Vert q\right\Vert _{L^{2}\left(
\Omega\right) }. \label{estPi2
\end{align}
For the third term in (\ref{defwq}) we ge
\begin{align}
\left\Vert \nabla\mathbf{T}_{3}\right\Vert _{\mathbb{L}^{2}\left(
\Omega\right) } & \leq C_{\operatorname*{GS}}k^{\kappa}\eta^{-1}\left\Vert
\left( I-\operatorname*{div}\nolimits_{\mathcal{T}}\Pi_{k
^{\operatorname*{CR}}\right) \left( I-\operatorname*{div}\Pi
^{\operatorname*{BR}}\right) q\right\Vert _{L^{2}\left( \Omega\right)
}\nonumber\\
& \leq C_{\operatorname*{GS}}k^{\kappa}\eta^{-1}\left( \left\Vert \left(
I-\operatorname*{div}\Pi^{\operatorname*{BR}}\right) q\right\Vert
_{L^{2}\left( \Omega\right) }+\left\Vert \operatorname*{div
\nolimits_{\mathcal{T}}\Pi_{k}^{\operatorname*{CR}}\left(
I-\operatorname*{div}\Pi^{\operatorname*{BR}}\right) q\right\Vert
_{L^{2}\left( \Omega\right) }\right) . \label{CGSest2
\end{align}
The combination with (\ref{CBR}), (\ref{CRlifting1}) leads t
\begin{equation}
\left\Vert \nabla\mathbf{T}_{3}\right\Vert _{\mathbb{L}^{2}\left(
\Omega\right) }\leq CC_{\operatorname*{GS}}\left( 1+C_{\operatorname*{BR
}\right) \left( 1+C_{\operatorname*{CR}}\right) k^{1/4+\kappa}\eta
^{-1}\left\Vert q\right\Vert _{L^{2}\left( \Omega\right) }. \label{estPi3
\end{equation}
For the fourth term we get in a similar wa
\begin{align}
\left\Vert \nabla\mathbf{T}_{4}\right\Vert _{\mathbb{L}^{2}\left(
\Omega\right) } & \overset{\text{(\ref{optests})}}{\leq
C_{\operatorname*{V}}\left( \left\Vert \left( I-\operatorname*{div
\nolimits_{\mathcal{T}}\Pi_{k}^{\operatorname*{CR}}\right) \left(
I-\operatorname*{div}\Pi^{\operatorname*{BR}}\right) q\right\Vert
_{L^{2}\left( \Omega\right) }+\left\Vert \operatorname*{div}\mathbf{T
_{3}\right\Vert _{L^{2}\left( \Omega\right) }\right) \nonumber\\
& \leq C_{\operatorname*{V}}\left( \left\Vert q\right\Vert _{L^{2}\left(
\Omega\right) }+\left\Vert \operatorname{div}\mathbf{T}_{1}\right\Vert
_{L^{2}\left( \Omega\right) }+\left\Vert \operatorname{div}_{\mathcal{T
}\mathbf{T}_{2}\right\Vert _{L^{2}\left( \Omega\right) }+\left\Vert
\operatorname*{div}\mathbf{T}_{3}\right\Vert _{L^{2}\left( \Omega\right)
}\right) \nonumber\\
& \overset{\text{(\ref{estPi1}), (\ref{estPi2}), (\ref{estPi3})}}{\leq
}C_{\operatorname*{V}}\left( 1+C_{\operatorname*{BR}}\right) \left(
1+C_{\operatorname*{CR}}k^{1/4}+CC_{\operatorname*{GS}}\left(
1+C_{\operatorname*{CR}}\right) k^{1/4+\kappa}\eta^{-1}\right) \left\Vert
q\right\Vert _{L^{2}\left( \Omega\right) }. \label{estPi4
\end{align}
The combination of (\ref{estPi1}), (\ref{estPi2}), (\ref{estPi3}),
(\ref{estPi4}) with (\ref{defwq}) leads to the assertion.
\endproof
Lemma \ref{LemmaFinOdd} implies that for conforming triangulations
$\mathcal{T}$ which satisfy (\ref{Assumptab}) and $\mathcal{C}_{\mathcal{T
}^{\operatorname{acute}}=\emptyset$, there exists a bounded linear operato
\[
\Pi_{\mathcal{T},k}^{\operatorname*{inv}}:\mathbb{P}_{k-1,0}\left(
\mathcal{T}\right) \rightarrow\mathbf{CR}_{k,0}\left( \mathcal{T}\right)
\]
such that $\operatorname*{div}_{\mathcal{T}}\circ\Pi_{\mathcal{T
,k}^{\operatorname*{inv}}$ is the identity on $\mathbb{P}_{k-1,0}\left(
\mathcal{T}\right) $ an
\[
\left\Vert \Pi_{\mathcal{T},k}^{\operatorname*{inv}}q\right\Vert
_{\mathbf{H}^{1}\left( \mathcal{T}\right) }\leq C_{\operatorname*{inv
}k^{1/4+\kappa}\left\Vert q\right\Vert _{L^{2}\left( \Omega\right)
\]
for a constant $C_{\operatorname*{inv}}$ which only depends on the
shape-regularity of the mesh.
\subsubsection{The inf-sup stability for $\mathcal{T}$\label{SecAcute}}
In this section, we remove the condition $\mathcal{C}_{\mathcal{T
}^{\operatorname{acute}}=\emptyset$ and construct a bounded right-inverse of
the piecewise divergence operator for odd $k\geq5$ and conforming
triangulation which contains at least one inner point. The construction is
based on the step-by-step procedure (cf. (\ref{stepbystepextension})) of the
triangulation $\mathcal{T}_{1}$ to $\mathcal{T}$. Inductively, we assume that
there is a triangulation $\mathcal{T}_{j}$ along a bounded right-inverse
$\Pi_{j,k}^{\operatorname*{inv}}:\mathbb{P}_{k-1,0}\left( \mathcal{T
_{j}\right) \rightarrow\mathbf{CR}_{k,0}\left( \mathcal{T}_{j}\right) $ of
the piecewise divergence operator. A single extension step is analysed by the
following lemma.
\begin{lemma}
Let $\mathcal{T}$ denote a conforming triangulation for the domain
$\Omega:=\operatorname*{dom}\mathcal{T}$ and let $\mathcal{T}^{\prime
\subset\mathcal{T}$ be a subset such that every triangle $K\in\mathcal{T
\backslash\mathcal{T}^{\prime}$ has one edge, say $E$, which belongs to
$\mathcal{E}\left( \mathcal{T}^{\prime}\right) $. We assume that
$\mathcal{T}^{\prime}$ has at least one inner vertex and set $\Omega^{\prime
}:=\operatorname*{dom}\mathcal{T}^{\prime}$. Assume that there exists a
bounded linear operator $\Pi_{\mathcal{T}^{\prime},k}^{\operatorname*{inv
}:\mathbb{P}_{k-1,0}\left( \mathcal{T}^{\prime}\right) \rightarrow
\mathbf{CR}_{k,0}\left( \mathcal{T}^{\prime}\right) $ with
$\operatorname*{div}_{\mathcal{T}^{\prime}}\circ\Pi_{\mathcal{T}^{\prime
,k}^{\operatorname*{inv}}=\operatorname*{Id}$ on $\mathbb{P}_{k-1,0}\left(
\mathcal{T}^{\prime}\right) $ an
\[
\left\Vert \Pi_{\mathcal{T}^{\prime},k}^{\operatorname*{inv}}q\right\Vert
_{\mathbf{H}^{1}\left( \Omega^{\prime}\right) }\leq C_{\mathcal{T}^{\prime
}\left\Vert q\right\Vert _{L^{2}\left( \Omega^{\prime}\right) }.
\]
Then, there exists a linear operator $\Pi_{\mathcal{T},k}^{\operatorname*{inv
}:\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) \rightarrow\mathbf{CR
_{k,0}\left( \mathcal{T}\right) $ with $\operatorname*{div}_{\mathcal{T
}\circ\Pi_{\mathcal{T},k}^{\operatorname*{inv}}=\operatorname*{Id}$ on
$\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) $ an
\[
\left\Vert \Pi_{\mathcal{T},k}^{\operatorname*{inv}}q\right\Vert
_{\mathbf{H}^{1}\left( \Omega\right) }\leq C_{\mathcal{T}}\left\Vert
q\right\Vert _{L^{2}\left( \Omega\right) }\quad\text{with\quad
}C_{\mathcal{T}}:=C_{3}k^{1/4+\kappa}C_{\mathcal{T}^{\prime}
\]
for a constant $C_{3}$ which depends only on the shape-regularity of the mesh.
\end{lemma}
\proof
Let $q\in\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) $. We set
$\mathbf{v}_{0}:=\Pi^{\operatorname*{BR}}q$ where the operator $\Pi
^{\operatorname*{BR}}$ is as in (\ref{defPiBR}) and satisfie
\[
\left\Vert \mathbf{v}_{0}\right\Vert _{\mathbf{H}^{1}\left( \Omega\right)
}\leq C_{\operatorname*{BR}}\left\Vert q\right\Vert _{L^{2}\left(
\Omega\right) }.
\]
Hence, $q_{1}:=q-\operatorname*{div}\mathbf{v}_{0}$ belongs to $\mathbb{P
_{k-1,0}\left( \mathcal{T}\right) $ and has trianglewise integral mean zero.
The following construction is illustrated in Figure \ref{Fig_glue1}
\begin{figure}[ptb
\centering
\includegraphics[
height=2.84in,
width=5.7276in
{fan_glue1.eps
\caption{The black triangles form the triangulation $\mathcal{T}^{\prime}$.
Left: One triangle $K$ is attached to $\mathcal{T}^{\prime}$ having a common
side $E$ with $K^{\prime}\in\mathcal{T}^{\prime}$ and $\mathcal{T}\left(
K^{\prime}\right) =\left\{ K\right\} $. Right: Here two triangles $K_{1}$,
$K_{2}\protect\notin\mathcal{T}^{\prime}$ are attached to a triangle
$K^{\prime}\in\mathcal{T}^{\prime}$ and $\mathcal{T}\left( K^{\prime}\right)
=\left\{ K_{1},K_{2}\right\} $.
\label{Fig_glue1
\end{figure}
Let $K^{\prime}\in\mathcal{T}^{\prime}$ such there exists a non-empty subset
$\mathcal{T}\left( K^{\prime}\right) \subset\mathcal{T}\backslash
\mathcal{T}^{\prime}$ and any $K\in\mathcal{T}\left( K^{\prime}\right) $
shares an edge with $K^{\prime}$. We have $\left\vert \mathcal{T}\left(
K^{\prime}\right) \right\vert \leq2$; indeed, if $\left\vert \mathcal{T
\left( K^{\prime}\right) \right\vert =3$, then all three edges of
$K^{\prime}$ are boundary edges which implies $\mathcal{T}^{\prime}=\left\{
K^{\prime}\right\} $ and violates the condition that $\mathcal{T}^{\prime}$
must contain an inner vertex.
For $K\in\mathcal{T}\left( K^{\prime}\right) $, let $\mathbf{z}$ denote the
vertex in $K$ opposite to $E$ and set $\omega_{E}=K\cup K^{\prime}$. The
endpoints of $E$ are denoted by $\mathbf{y}_{1}$, $\mathbf{y}_{2}$. We employ
the ansatz (cf. (\ref{defboldpsi})
\begin{equation}
\mathbf{v}_{1}:=\alpha\tilde{B}_{k,E}^{\operatorname*{CR}}\mathbf{n}_{E}
\label{ansatzv1
\end{equation}
with the convention that $\mathbf{n}_{E}$ is the unit vector orthogonal to $E$
and directed into $K^{\prime}$. By definition $\mathbf{v}_{1}\in
\mathbf{CR}_{k,0}\left( \mathcal{T}\right) $ and $\operatorname*{supp
\mathbf{v}_{1}\subset\omega_{E}$. We determine $\alpha$ in (\ref{ansatzv1})
such that $\operatorname*{div}\left( \left. \mathbf{v}_{1}\right\vert
_{K}\right) \left( \mathbf{z}\right) =q_{1}\left( \mathbf{z}\right) $ and
employ (\ref{divcomp}) to ge
\[
\operatorname{div}\left( \tilde{B}_{k,E}^{\operatorname*{CR}}\mathbf{n
_{E}\right) \left( \mathbf{z}\right) =\binom{k+1}{2}\frac{\left\vert
E\right\vert }{\left\vert K\right\vert }.
\]
Hence $\alpha=q_{1}\left( \mathbf{z}\right) \frac{\left\vert K\right\vert
}{\left\vert E\right\vert }/\binom{k+1}{2}$ and we conclude as in the proof of
Lemma \ref{LemProject} tha
\[
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{1}\right\Vert _{\mathbb{L
^{2}\left( \omega_{E}\right) }\leq C_{\operatorname*{CR}}k^{1/4}\left\Vert
q_{1}\right\Vert _{L^{2}\left( K\right) }.
\]
We se
\[
q_{1}=\operatorname*{div}\mathbf{v}_{1}+q_{2}\quad\text{for }q_{2
:=q_{1}-\operatorname*{div}\mathbf{v}_{1
\]
and note tha
\begin{equation}
\left\Vert q_{2}\right\Vert _{L^{2}\left( K\right) }\leq\left(
1+C_{\operatorname*{CR}}k^{1/4}\right) \left\Vert q_{1}\right\Vert
_{L^{2}\left( K\right) }. \label{q2est
\end{equation}
The construction implies $q_{2}\in\mathbb{P}_{k-1,0}\left( \mathcal{T
\right) $, $q_{2}$ has trianglewise integral mean zero, and $q_{2}\left(
\mathbf{z}\right) =0$. Next, we employ the function defined in \cite[(3.5)
{GuzmanScott2019}. For $\tilde{K}\in\left\{ K,K^{\prime}\right\} $, let
$\left. \eta_{E,1}\right\vert _{\tilde{K}}:=\lambda_{\tilde{K},\mathbf{y
_{1}}^{2}\lambda_{\tilde{K},\mathbf{y}_{2}}$ and $\left. \eta_{E,2
\right\vert _{\tilde{K}}:=\lambda_{\tilde{K},\mathbf{y}_{1}}\lambda_{\tilde
{K},\mathbf{y}_{2}}^{2}$. For $j=1,2$ se
\[
\mathbf{v}_{E,j}=\left\vert E\right\vert \mathbf{t}_{E,j}\eta_{E,j
\quad\text{for }\mathbf{t}_{E,1}=\frac{\mathbf{y}_{2}-\mathbf{y}_{1
}{\left\vert E\right\vert }\quad\text{and\quad}\mathbf{t}_{E,2}=-\mathbf{t
_{E,1}.
\]
The properties of $\mathbf{v}_{E,j}$ are listed in \cite[(3.6)
{GuzmanScott2019}: for $j=1,2$ it hold
\begin{equation
\begin{array}
[c]{ll
\mathbf{v}_{E,j}\in\mathbf{S}_{3,0}\left( \mathcal{T}\right) , &
\operatorname*{supp}\mathbf{v}_{E,j}\subset\omega_{E},\\
\int_{K}\operatorname{div}\mathbf{v}_{E,j}=0 & \forall K\in\mathcal{T},\\
\left( \operatorname{div}\left. \mathbf{v}_{E,j}\right\vert _{K}\right)
\left( \mathbf{v}\right) =\delta_{\mathbf{v},\mathbf{y}_{j}} & \forall
K\in\mathcal{T},\quad\forall\mathbf{v}\in\mathcal{V}\left( K\right) ,\\
\left\Vert \nabla\mathbf{v}_{E,j}\right\Vert _{\mathbb{L}^{2}\left(
\omega_{E}\right) }\leq C_{0}\left\vert E\right\vert & \text{for
C_{0}\text{ depending only on the shape-regularity.
\end{array}
\label{vEj
\end{equation}
We set
\[
\mathbf{v}_{2}:=\sum_{j=1}^{2}\left. q_{2}\right\vert _{K}\left(
\mathbf{y}_{j}\right) \mathbf{v}_{E,j
\]
and defin
\begin{equation}
q_{2}=\operatorname{div}\mathbf{v}_{2}+q_{3}\quad\text{for }q_{3
:=q_{2}-\operatorname{div}\mathbf{v}_{2}. \label{defq3
\end{equation}
The function $q_{3}\in\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) $ has
trianglewise integral mean zero and $\left. q_{3}\right\vert _{K}$ vanishes
in all vertices of $K$. The norm $\left\Vert \nabla\mathbf{v}_{2}\right\Vert
_{\mathbb{L}^{2}\left( K^{\prime}\cup K\right) }$ can be estimated in the
same way as the function $\mathbf{T}_{3}$ in (\ref{estPi3}); however the
factor $\eta^{-1}$ does not appear as in (\ref{estPi3}) since the last
estimate in (\ref{vEj}) does not depend on $\eta$. In this way, we ge
\begin{align}
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{2}\right\Vert _{\mathbb{L
^{2}\left( \omega_{E}\right) } & \leq C_{1}k^{\kappa}\left\Vert
q_{2}\right\Vert _{L^{2}\left( K\right) }\label{v2est1}\\
& \overset{\text{(\ref{q2est})}}{\leq}C_{1}\left( 1+C_{\operatorname*{CR
}k^{1/4}\right) k^{\kappa}\left\Vert q_{1}\right\Vert _{L^{2}\left(
K\right) }. \label{v2est2
\end{align}
Hence, from \cite[Thm. 3.4]{Ainsworth_parker_I} we deduce that there exists
$\mathbf{v}_{3}\in\mathbf{S}_{k,0}\left( \mathcal{T}\right) $ with
$\operatorname*{supp}\mathbf{v}_{3}=K$ such that $\operatorname*{div
\mathbf{v}_{3}=q_{3}$ on $K$ an
\begin{align*}
h_{K}^{-1}\left\Vert \mathbf{v}_{3}\right\Vert _{\mathbf{L}^{2}\left(
K\right) }+\left\Vert \nabla\mathbf{v}_{3}\right\Vert _{\mathbb{L}^{2}\left(
K\right) }\leq & C_{\operatorname*{V}}\left\Vert q_{3}\right\Vert
_{L^{2}\left( K\right) }\overset{\text{(\ref{defq3}), (\ref{v2est1})}}{\leq
}C_{\operatorname*{V}}\left( 1+C_{1}k^{\kappa}\right) \left\Vert
q_{2}\right\Vert _{L^{2}\left( K\right) }\\
& \overset{\text{(\ref{q2est})}}{\leq}C_{\operatorname*{V}}\left(
1+C_{1}k^{\kappa}\right) \left( 1+C_{\operatorname*{CR}}k^{1/4}\right)
\left\Vert q_{1}\right\Vert _{L^{2}\left( K\right) }.
\end{align*}
In this way we have constructed the function $\mathbf{v}_{K}\in\mathbf{CR
_{k,0}\left( \mathcal{T}\right) $ b
\[
\mathbf{v}_{K}=\mathbf{v}_{1}+\mathbf{v}_{2}+\mathbf{v}_{3
\]
such that $\operatorname*{div}\mathbf{v}_{K}=q_{1}$ on $K$,
$\operatorname*{supp}\mathbf{v}_{K}\subset\omega_{E}$ an
\begin{align*}
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{K}\right\Vert _{\mathbb{L
^{2}\left( \Omega\right) } & =\left\Vert \nabla_{\mathcal{T}}\mathbf{v
_{K}\right\Vert _{\mathbb{L}^{2}\left( \omega_{E}\right) }\leq\sum_{\ell
=1}^{3}\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{\ell}\right\Vert
_{\mathbb{L}^{2}\left( \omega_{E}\right) }\\
& \leq C_{2}k^{1/4+\kappa}\left\Vert q_{1}\right\Vert _{L^{2}\left(
K\right) }\leq C_{2}k^{1/4+\kappa}\left( \left\Vert q\right\Vert
_{L^{2}\left( K\right) }+\left\Vert \nabla\mathbf{v}_{0}\right\Vert
_{\mathbb{L}^{2}\left( K\right) }\right) ,
\end{align*}
where $C_{2}$ only depends on the shape-regularity of the mesh through the
constants $C_{\operatorname*{V}}$, $C_{\operatorname*{CR}}$, $C_{1}$. Let
$\mathbf{v}_{q}:=\mathbf{v}_{0}+\sum_{K\in\mathcal{T}\backslash\mathcal{T
^{\prime}}\mathbf{v}_{K}$ and note that by constructio
\[
\operatorname*{div}\nolimits_{\mathcal{T}}\mathbf{v}_{q}=q\quad\text{on
}\Omega\backslash\overline{\Omega^{\prime}
\]
an
\begin{align*}
\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{q}\right\Vert _{\mathbb{L
^{2}\left( \Omega\right) } & \leq\left\Vert \nabla\mathbf{v}_{0}\right\Vert
_{\mathbb{L}^{2}\left( \Omega\right) }+\sum_{K\in\mathcal{T}\backslash
\mathcal{T}^{\prime}}\left\Vert \nabla_{\mathcal{T}}\mathbf{v}_{K}\right\Vert
_{\mathbb{L}^{2}\left( \omega_{E}\right) }\\
& \leq\left\Vert \nabla\mathbf{v}_{0}\right\Vert _{\mathbb{L}^{2}\left(
\Omega\right) }+2C_{2}\sum_{K\in\mathcal{T}\backslash\mathcal{T}^{\prime
}k^{1/4+\kappa}\left( \left\Vert q\right\Vert _{L^{2}\left( K\right)
}+\left\Vert \nabla\mathbf{v}_{0}\right\Vert _{\mathbb{L}^{2}\left( K\right)
}\right) \\
& \leq Ck^{1/4+\kappa}\left\Vert q\right\Vert _{L^{2}\left( \Omega\right)
}.
\end{align*}
Finally, the linear map $\Pi_{\mathcal{T},k}^{\operatorname*{inv}
:\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) \rightarrow\mathbf{CR
_{k,0}\left( \mathcal{T}\right) $ is defined b
\[
\Pi_{\mathcal{T},k}^{\operatorname*{inv}}q=\mathbf{v}_{q}+\Pi_{\mathcal{T
^{\prime},k}^{\operatorname*{inv}}\left( \left. \left(
q-\operatorname*{div}\mathbf{v}_{q}\right) \right\vert _{\Omega^{\prime
}\right)
\]
and satisfies $\operatorname*{div}\circ\Pi_{\mathcal{T},k
^{\operatorname*{inv}}=\operatorname*{Id}$ on $\mathbb{P}_{k-1,0}\left(
\mathcal{T}\right) $ an
\[
\left\Vert \nabla_{\mathcal{T}}\Pi_{\mathcal{T},k}^{\operatorname*{inv
}q\right\Vert _{\mathbb{L}^{2}\left( \Omega\right) }\leq C_{3}k^{1/4+\kappa
}C_{\mathcal{T}^{\prime}}\left\Vert q\right\Vert _{L^{2}\left( \Omega\right)
\]
for some $C_{3}$ which only depends on the shape-regularity of the mesh.
\endproof
By iterating this argument we can prove Theorem \ref{Theomain} for the case
(\ref{Assumptab}).
\begin{theorem}
Let $\mathcal{T}$ be a conforming triangulation which contains at least one
interior vertex. Let $k\geq5$ be odd and let $L\in\mathbb{N}_{0}$ the number
of steps in the construction (\ref{stepbystepextension}). Then, the inf-sup
constant for the corresponding Crouzeix-Raviart discretization satisfie
\[
c_{\mathcal{T},k}\geq c_{\mathcal{T}}k^{-\left( 1/4+\kappa\right) \left(
L+1\right)
\]
for some constant $c_{\mathcal{T}}$ which depends only on the shape-regularity
of the mesh. If every triangle in $\mathcal{T}$ has an interior vertex than
$L=0$.
\end{theorem}
\subsection{The case of even $k$\label{SecEvenk}}
This case is slightly simpler that the case of odd $k$ since the
non-conforming Crouzeix-Raviart functions for even $k$ have smaller support
(i.e., one triangle) compared to two triangles (which share an edge) for odd
$k$.
In this section, we assum
\begin{equation
\begin{array}
[c]{ll
\text{a)} & k\geq4\text{ is even and}\\
\text{b)} & \mathcal{T}\text{ is a conforming triangulation and contains more
than a single triangle.
\end{array}
\label{asseven
\end{equation}
\begin{remark}
\label{RemTle1}It is easy to verify that $\left\vert \mathcal{T}\right\vert
>1$ implies that there exists a mapping $\mathfrak{K}:\mathcal{C
_{\mathcal{T}}\left( \eta\right) $ $\rightarrow\mathcal{T}$ with
$\mathfrak{K}\left( \mathbf{z}\right) \in\mathcal{T}_{\mathbf{z}}$ and not
all vertices of $\mathfrak{K}\left( \mathbf{z}\right) $ are $\eta$-critical.
\end{remark}
For an $\eta$-critical point $\mathbf{z}\in\mathcal{C}_{\mathcal{T}}\left(
\eta\right) $, let $n_{\mathbf{z}}:=\left\vert \mathcal{T}_{\mathbf{z
}\right\vert $ and fix a counterclockwise numbering of the triangles in
\begin{equation}
\mathcal{T}_{\mathbf{z}}=\left\{ K_{j}^{\mathbf{z}}:1\leq j\leq
n_{\mathbf{z}}\right\} \label{tznumbconv
\end{equation}
such that $K_{j}^{\mathbf{z}}$ and $K_{j+1}^{\mathbf{z}}$ share an edge for
all $1\leq j\leq n_{\mathbf{z}}-1$. For the case of even $k$, we define the
functional $A_{\mathcal{T},\mathbf{z}}$ for $q\in\mathbb{P}_{k-1,0}\left(
\mathcal{T}\right) $ b
\[
A_{\mathcal{T},\mathbf{z}}q:=\sum_{j=1}^{n_{\mathbf{z}}}\left( -1\right)
^{j}\left. q\right\vert _{K_{j}^{\mathbf{z}}}\left( \mathbf{z}\right) .
\]
\begin{lemma}
Let assumption (\ref{asseven}) be satisfied. There exists a constant $\eta
_{2}>0$ which only depends on the shape-regularity of the mesh such that for
any fixed $0\leq\eta<\eta_{2}$ and any $q\in\mathbb{P}_{k-1,0}\left(
\mathcal{T}\right) $, there exists some $\mathbf{v}_{q}\in\mathbf{CR
_{k,0}\left( \mathcal{T}\right) $ such tha
\begin{align}
\int_{K}\operatorname{div}\mathbf{v}_{q} & =0\quad\forall K\in
\mathcal{T},\label{intcondeven}\\
q-\operatorname*{div}\nolimits_{\mathcal{T}}\mathbf{v}_{q} & \in
M_{\eta,k-1}^{\operatorname*{SV}}\left( \mathcal{T}\right)
\label{qcorecteven
\end{align}
an
\begin{equation}
\left\Vert \mathbf{v}_{q}\right\Vert _{\mathbf{H}^{1}\left( \mathcal{T
\right) }\leq C_{\operatorname*{CR}}k^{1/4}\left\Vert q\right\Vert
_{L^{2}\left( \Omega\right) }. \label{vq1
\end{equation}
The constant $C_{\operatorname*{CR}}$ depends only on the shape-regularity of
the mesh.
\end{lemma}
\proof
For a triangle $K\in\mathcal{T}$, we set $\mathcal{C}_{K}\left( \eta\right)
:=\left\{ \mathbf{z}\in\mathcal{V}\left( K\right) \cap\mathcal{C
_{\mathcal{T}}\left( \eta\right) :K=\mathfrak{K}\left( \mathbf{z}\right)
\right\} $ and $n_{K}:=\left\vert \mathcal{C}_{K}\left( \eta\right)
\right\vert $. Note that $0\leq n_{K}\leq2$ (cf. Rem. \ref{RemTle1}). We
number the vertices $\mathbf{V}_{j}$ in $K$ counterclockwise with the
convention $\left\{ \mathbf{V}_{j}:1\leq j\leq n_{K}\right\} =\mathcal{C
_{K}\left( \eta\right) $. The angle in $K$ at $\mathbf{V}_{j}$ is denoted by
$\alpha_{j}$. Let $E_{j}$ be the edge in $K$ opposite to $\mathbf{V}_{j}$ and
let $\mathbf{n}_{j}$ denote the outward unit normal vector at $E_{j}$.
In a similar way as for the construction of $\tilde{B}_{k,E
^{\operatorname*{CR}}$ in Lemma \ref{Lemtrianglemean} there exists for any
$\mathbf{c}\in\mathbb{R}^{2}\backslash\left\{ \mathbf{0}\right\} $ a
function $\tilde{B}_{k,K}^{\operatorname*{CR}}\in\operatorname*{CR
_{k,0}\left( \mathcal{T}\right) $ (depending on $\mathbf{c}$) with
\begin{enumerate}
\item $\operatorname*{supp}\tilde{B}_{k,K}^{\operatorname*{CR}}=K,$
\item for all $K\in\mathcal{T}$, for all $\mathbf{z}\in\mathcal{V}\left(
K\right)
\begin{equation}
\operatorname{div}\left( \left. \tilde{B}_{k,K}^{\operatorname*{CR
}\mathbf{c}\right\vert _{K}\right) \left( \mathbf{z}\right)
=\operatorname{div}\left( \left. B_{k,K}^{\operatorname*{CR}}\mathbf{c
\right\vert _{K}\right) \left( \mathbf{z}\right) , \label{divtrianglerel
\end{equation}
\item for all $K\in\mathcal{T}$, for all $E^{\prime}\subset\partial K$
\[
\left. \left. \tilde{B}_{k,K}^{\operatorname*{CR}}\right\vert _{K
\right\vert _{E^{\prime}}=\left. \left. B_{k,K}^{\operatorname*{CR
}\right\vert _{K}\right\vert _{E^{\prime}},
\]
\item for all $K\in\mathcal{T}
\begin{equation}
\int_{K}\operatorname{div}_{\mathcal{T}}\left( \tilde{B}_{k,K
^{\operatorname*{CR}}\mathbf{c}\right) =0. \label{divKeven=0
\end{equation}
\item The piecewise gradient is bounded b
\begin{equation}
\left\Vert \tilde{B}_{k,K}^{\operatorname*{CR}}\right\Vert _{H^{1}\left(
\mathcal{T}\right) }\leq Ck^{1/4}.\label{Btildeest
\end{equation}
\end{enumerate}
For $j=1,2$, we define $\tilde{B}_{k,K}^{\operatorname*{CR}}$ by choosing
$\mathbf{c}=\mathbf{n}_{j}$ and se
\begin{equation
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}:=\left\{
\begin{array}
[c]{ll
\tilde{B}_{k,K}^{\operatorname*{CR}}\mathbf{n}_{j} & \text{in }K,\\
0 & \text{in }\Omega\backslash K,
\end{array}
\right. \label{defboldpsieven
\end{equation}
i.e., we fix $\mathbf{v}_{K}:=\mathbf{n}_{1}$ and $\mathbf{w}_{K
=\mathbf{n}_{2}$ in (\ref{basisvel}). The divergence of
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}$ evaluated at a vertex $\mathbf{V}_{s}$,
$s=1,2$, is given b
\begin{align*}
\operatorname{div}\left( \left.
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}\right\vert _{K}\right) \left( \mathbf{V
_{s}\right) & \overset{\text{(\ref{divtrianglerel})}}{=}\operatorname{div
\left( \left. B_{k,K}^{\operatorname*{CR}}\mathbf{n}_{j}\right\vert
_{K}\right) \left( \mathbf{V}_{s}\right) =
{\displaystyle\sum\limits_{i=1}^{3}}
L_{k}^{\prime}\left( 1-2\lambda_{K,i}\left( \mathbf{V}_{s}\right) \right)
\partial_{\mathbf{n}_{j}}\lambda_{K,i}\\
& =k\left( k+1\right) \partial_{\mathbf{n}_{j}}\lambda_{K,s
\overset{\text{(\ref{normcomp})}}{=}\binom{k+1}{2}\frac{\left\vert
E_{s}\right\vert }{\left\vert K\right\vert }\times\left\{
\begin{array}
[c]{ll
-1 & j=s,\\
\cos\alpha_{3} & j\neq s.
\end{array}
\right.
\end{align*}
Let $q\in\mathbb{P}_{k-1,0}\left( \mathcal{T}\right) $. We choose
\mbox{\boldmath$ \delta$
_{K}:=\left( \delta_{K,j}\right) _{j=1}^{2}$ by the conditio
\begin{equation}
A_{\mathcal{T},\mathbf{V}_{s}}\left( \operatorname*{div
\nolimits_{\mathcal{T}}\left( \sum_{j=1}^{2}\delta_{K,j
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}\right) \right) \overset{!}{=}\left\{
\begin{array}
[c]{ll
A_{\mathcal{T},\mathbf{V}_{s}}\left( q\right) & \text{if }\mathbf{V}_{s
\in\mathcal{C}_{K}\left( \eta\right) ,\\
0 & \text{otherwise.
\end{array}
\right. \label{condkeven
\end{equation}
For $s=1,2$, let $\ell_{s}$ be defined by $K_{\ell_{s}}^{\mathbf{V}_{s}}=K$
(cf. (\ref{tznumbconv})). Then
\begin{align*}
A_{\mathcal{T},\mathbf{V}_{s}}\left( \operatorname*{div
\nolimits_{\mathcal{T}}\left( \sum_{j=1}^{2}\delta_{K,j
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}\right) \right) & =\left( -1\right)
^{\ell_{s}}\sum_{j=1}^{2}\delta_{K,j}\left( \operatorname*{div}\left.
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}\right\vert _{K}\right) \left( \mathbf{V
_{s}\right) \\
& =\mathbf{M}_{K
\mbox{\boldmath$ \delta$
_{K
\end{align*}
fo
\[
\mathbf{M}_{K}=\left( -1\right) ^{\ell_{s}+1}\binom{k+1}{2}\left\vert
K\right\vert ^{-1}\left[
\begin{array}
[c]{cc
\left\vert E_{1}\right\vert & -\left\vert E_{1}\right\vert \cos\alpha_{3}\\
-\left\vert E_{2}\right\vert \cos\alpha_{3} & \left\vert E_{2}\right\vert
\end{array}
\right] .
\]
We define $\mathbf{r}_{K}=\left( r_{K,s}\right) _{s=1}^{2}$ b
\[
r_{K,s}:=\left\{
\begin{array}
[c]{ll
A_{\mathcal{T},\mathbf{V}_{s}}\left( q\right) & \text{if }\mathbf{V}_{s
\in\mathcal{C}_{K}\left( \eta\right) ,\\
0 & \text{otherwise
\end{array}
\right.
\]
so that
\mbox{\boldmath$ \delta$
_{K}$ is the solution o
\[
\mathbf{M}_{K
\mbox{\boldmath$ \delta$
_{K}=\mathbf{r}_{K}.
\]
Observe tha
\[
\det\left[
\begin{array}
[c]{cc
\left\vert E_{1}\right\vert & -\left\vert E_{1}\right\vert \cos\alpha_{3}\\
-\left\vert E_{2}\right\vert \cos\alpha_{3} & \left\vert E_{2}\right\vert
\end{array}
\right] =\left\vert E_{1}\right\vert \left\vert E_{2}\right\vert \sin
^{2}\alpha_{3}=2\left\vert K\right\vert \sin\alpha_{3
\]
and $\sin\alpha_{3}\geq\sin\phi_{\mathcal{T}}>0$ due to the shape-regularity
of the mesh. For the coefficient
\mbox{\boldmath$ \delta$
_{K}$ we get explicitl
\begin{equation
\mbox{\boldmath$ \delta$
_{K}=\frac{\left( -1\right) ^{\ell_{s}+1}}{\left( k+1\right) k\sin
\alpha_{3}}\left[
\begin{array}
[c]{cc
\left\vert E_{2}\right\vert & \left\vert E_{1}\right\vert \cos\alpha_{3}\\
\left\vert E_{2}\right\vert \cos\alpha_{3} & \left\vert E_{1}\right\vert
\end{array}
\right] \mathbf{r}_{K}\label{Defalphazeven
\end{equation}
with an estimat
\[
\left\Vert
\mbox{\boldmath$ \delta$
_{K}\right\Vert \leq C\frac{h_{K}}{k\left( k+1\right) }\left\Vert
\mathbf{r}_{K}\right\Vert \overset{\text{Lem. \ref{LemFunctional}}}{\leq
}C\left\Vert q\right\Vert _{L^{2}\left( \omega_{K}\right) }\quad\text{if
}\mathcal{C}_{K}\left( \eta\right) \neq\emptyset,
\]
where $C$ only depends on the shape-regularity of the mesh. Note that this is
the analogue for even $k$ to (\ref{estboldalpha}). If $\mathcal{C}_{K}\left(
\eta\right) =\emptyset$, it holds
\mbox{\boldmath$ \delta$
_{K}=\mathbf{0}$.
We define the global functio
\begin{equation}
\mathbf{v}_{q}:=\sum_{\substack{K\in\mathcal{T}\\\mathcal{C}_{K}\left(
\eta\right) \neq\emptyset}}\sum_{j=1}^{2}\delta_{K,j
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}. \label{v1globaleven
\end{equation}
From (\ref{divKeven=0}) we conclude that $\mathbf{v}_{q}$ satisfies
(\ref{intcondeven}).
Next, we verify (\ref{qcorecteven}). Let $\mathbf{y}\in\mathcal{C
_{\mathcal{T}}\left( \eta\right) $ and recall the notation and convention as
in (\ref{tznumbconv}). Let $K_{\ell}^{\mathbf{y}}=\mathfrak{K}\left(
\mathbf{y}\right) $. Then (\ref{qcorecteven}) follows fro
\begin{align*}
A_{\mathcal{T},\mathbf{y}}\left( \operatorname*{div}\nolimits_{\mathcal{T
}\mathbf{v}_{q}\right) & =\left( -1\right) ^{\ell}\left(
\operatorname{div}\left. \mathbf{v}_{q}\right\vert _{K_{\ell}^{\mathbf{y}
}\right) \left( \mathbf{y}\right) =\left( -1\right) ^{\ell}\left(
\operatorname{div}\left. \sum_{j=1}^{2}\delta_{K_{\ell}^{\mathbf{y}},j
\mbox{\boldmath$ \psi$
_{k,K_{\ell}^{\mathbf{y}}}^{\operatorname*{CR},j}\right\vert _{K_{\ell
}^{\mathbf{y}}}\right) \left( \mathbf{y}\right) \\
& =A_{\mathcal{T},\mathbf{y}}\left( q\right) .
\end{align*}
The estimat
\[
\left\Vert \nabl
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR},j}\right\Vert _{\mathbb{L}^{2}\left( K\right)
}\leq Ck^{1/4
\]
for a constant $C$ which only depends on the shape-regularity of the mesh
follows directly from (\ref{Btildeest}) and the final estimate (\ref{vq1}) is
derived by repeating the arguments as in the proof of Lemma \ref{LemProject}
\endproof
This lemma allows us to extend Definition \ref{DefPiCR} to the case of even
$k$ by defining the coefficients
\mbox{\boldmath$ \delta$
_{K}$ by (\ref{Defalphazeven}) and the functions
\mbox{\boldmath$ \psi$
_{k,K}^{\operatorname*{CR}}$ by (\ref{defboldpsieven}) and set (cf.
(\ref{v1globaleven})) $\Pi_{k}^{\operatorname*{CR}}q:=\mathbf{v}_{q}$. Since
$\left( I-\Pi_{k}^{\operatorname*{CR}}\right) q\in M_{\eta,k-1
^{\operatorname*{SV}}\left( \mathcal{T}\right) $ we may apply the further
steps in the proof of Lemma \ref{LemmaFinOdd} to obtain the inf-sup stability
for even $k$.
\begin{theorem}
Let $\mathcal{T}$ be a conforming triangulation satisfies (\ref{asseven}).
Then, the inf-sup constant for the corresponding Crouzeix-Raviart
discretization satisfie
\[
c_{\mathcal{T},k}\geq c_{\mathcal{T}}k^{-\left( 1/4+\kappa\right)
\]
for some constant $c_{\mathcal{T}}$ which depends only on the shape-regularity
of the mesh.
\end{theorem}
\section{Conclusion\label{SecConcl}}
In this paper, we have derived lower bounds for the inf-sup constant for
Crouzeix-Raviart elements for the Stokes equation which are explicit with
respect to the polynomial degree $k$ and are independent of the mesh size.
Under very mild assumptions on the topology of the triangulation, more precisely,
\begin{enumerate}
\item for odd $k\geq3$
\begin{enumerate}
\item we assume that $\mathcal{T}$ has at least one interior point and
\item for $k\geq5$ we assume that $\mathcal{T}$ has no acute critical point,
\end{enumerate}
\item for even $k\geq4$ we assume that $\mathcal{T}$ contains more than one triangle,
\end{enumerate}
the inf-sup constant can be bounded from below by $c_{\mathcal{T},k}\geq
c_{\mathcal{T}}k^{-\left( 1/4+\kappa\right) }$ with $\kappa=2$. If for odd
$k$, condition 1.b is not satisfied but a step-by-step construction
(\ref{stepbystepextension}) for some $L\geq1$ is possible, then,
$c_{\mathcal{T},k}\geq c_{\mathcal{T}}k^{-\left( 1/4+\kappa\right) \left(
1+L\right) }$.
Finally, we compare these findings with some other stable pairs of Stokes
elements on triangulations in the literature. The element $\left(
\mathbf{S}_{k,0}\left( \mathcal{T}\right) ,\mathbb{P}_{k-2,0}\left(
\mathcal{T}\right) \right) $ has a discrete inf-sup constant which can be
estimated from below by $Ck^{-3}$ (see \cite{Schwab_Suri_hp_infsup},
\cite{Stenberg_Suri_hp}). The discrete inf-sup constant for the Scott-Vogelius
element $\left( \mathbf{S}_{k,0}\left( \mathcal{T}\right) ,\mathbb{P
_{k-1,0}\left( \mathcal{T}\right) \right) $ can be estimated from below by
$c/\left( \Theta_{\min}k^{m}\right) $ for some integer $m$ sufficiently
large (see \cite{ScottVogelius}, \cite{vogelius1983right}). In
\cite{Ainsworth_parker_I}, a conforming stable pair $\left( \mathbf{X
_{k}\left( \mathcal{T}\right) ,M_{k-1}\left( \mathcal{T}\right) \right) $
of Stokes elements on triangulations is introduced and it is proved that the
discrete inf-sup constant can be estimated from below by $c/\tilde{\Theta
}_{\min}$ for a constant $c$ independent of $h$ and $k$ and $\tilde{\Theta
}_{\min}:=\min_{\mathbf{z}\in\mathcal{V}_{\partial\Omega}\left(
\mathcal{T}\right) \backslash\mathcal{C}_{\mathcal{T}}}\Theta\left(
\mathbf{z}\right) $. However, the implementation requires finite elements for
the velocity with $C^{1}$ continuity at the triangle vertices and pressures
which are continuous in the triangle vertices.
|
\section*{Acknowledgements}
\label{sec:acks}
SH is grateful for support from the NASA Ames Research Center, from NASA Academic Mission Services (NAMS) under Contract No. NNA16BD14C, and from the DARPA ONISQ program under interagency agreement IAA 8839, Annex 114.
\bibliographystyle{abbrvnat}
\section{ZX-Calculus for Parametrized Quantum Circuits} \label{sec:ZXforPQC}
\label{sec:zx_for_pqc}
In this section we extend the ZX-calculus to
accommodate linear combinations of (conventional) ZX-diagrams.
Then, toward its application to parametrized quantum circuits we derive a collection of general rules and useful identities within the new framework.
We will apply these rules to several concrete quantum circuit examples in Section~\ref{sec:example_applications}.
\subsection{Diagrammatic Rules for Linear Combinations}
Here we define linear combinations of diagrams, in which case diagram constants give the relative weights of the sum. For example, for computing the expectation value of an observable $H=\sum_{j=1}^m a_j H_j$ for some quantum circuit state $\ket{\psi}=U\ket{\psi_0}$ we have $\langle H \rangle_\psi = \sum_{j=1}^m a_j\langle H_j \rangle_\psi$, which hence corresponds to a single ZX-diagram or equivalently to a sum of $m$ weighted diagrams. This idea generalizes in the natural way to sums of linear maps and more general ZX objects.
We also show new ZX-diagram rules which relate single (sub)diagrams to sums or products of (sub)diagrams, such that the resulting diagram reductions involve differing numbers of diagrams.
As mentioned, we do not use the common convention of considering diagrams equivalent up to scalars or phases; hence we include complex scalar multipliers explicitly in our diagrams and rules to follow, i.e.,
\begin{equation}
a \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$m$}{$A$}{$n$}[2.5em]}}
\, \neq \, \,\,
b \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$m$}{$A$}{$n$}[2.5em]}} \label{eqn:zxsumrules_scalar}
\qquad
\text{unless } a=b \,,
\end{equation}
where $a, b$ are complex scalar multipliers and the diagram is a placeholder for an arbitrary ZX-diagram with $m$ inputs and $n$ outputs.
In particular, care must be taken in applying the usual rules of ZX-calculus to account for any implicit constant factors.
We note that scalar factors are also retained in the distinct application of~\cite{zhao2021analyzing}; see~\cite[Fig. 4]{zhao2021analyzing} for an example list of some ZX-diagram rewrite rules with explicit scalars.
\begin{defn}[Sum notation]
We define novel diagram notation
for describing arbitrary linear combinations.
The linear combination of two ZX-diagrams with $m$ inputs and $n$ outputs is written as
\begin{equation}
a \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$m$}{$A$}{$n$}[2.5em]}}
+
\,\, b \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$m$}{$B$}{$n$}[2.5em]}} \\
=: \vcenter{\hbox{
\externalize{zxsumrules_sum}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_sums_of_diagrams}
\end{tikzpicture}
}
}} \label{eqn:zxsumrules_sum} \, .
\end{equation}
\end{defn}
Each summand is written inside a \textit{bubble} and the scalar factors are written on the line combining the new summation symbol and the bubbles.
This definition naturally extends to an arbitrary number of summands.
Summand diagrams are required have the same numbers of input $(m)$ and output $(n)$ lines as each other and
as the those of the sum object.
Note that $m$ or $n$ are zero for diagrams representing states ($m=0$), effects ($n=0$), or constants ($m=n=0$).
Sums of diagrams also arise in~\cite{yeung2020diagrammatic,toumi2021diagrammatic,zhao2021analyzing} in the context of differentiating diagram components (where sums arise, for example, from the product rule of calculus).
Our consideration of arbitrary complex linear combinations of diagrams is more general.
\subsubsection{Rules}
Now we state the two rules needed for the extension of ZX-calculus to linear combinations.
\begin{enumerate}
\item {Diagram Pull Rule}
The first rule
applies if diagrams in a linear combination are equal up to a certain subdiagram ($A$ and $B$ below).
Then we can write a single diagram containing the linear combination of the beforementioned subdiagrams
\combineddiagplaceholdconfig
\begin{align}
&
a \, \cdot \, \;
\vcenter{\hbox{
\scalebox{1.0}{
\combineddiagplacehold{$A$}
}
}}
+
b \, \cdot \, \;
\vcenter{\hbox{
\scalebox{1.0}{
\combineddiagplacehold{$B$}
}
}} \\[2em]
\stackrel{\eqref{eqn:zxsumrules_sum}}{=}
&
\vcenter{\hbox{
\scalebox{1.0}{
\externalize{zxsumrules_partial_sum_all_combined}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_partial_sum_notation_all_combined}
\end{tikzpicture}
}
}
}} \\[2em]
= &
\vcenter{\hbox{
\scalebox{1.0}{
\externalize{zxsumrules_partial_sum_combined}{%
\begin{tikzpicture}[node distance=4ex and 1.5em]
\input{./images/tikz/rules_partial_sum_notation_combined}
\end{tikzpicture}
}
}
}} \label{eqn:zxsumrules_pull} \, . \tag{i}
\end{align}
The last equality we call the \textit{diagram pull rule}.
This also hold if any of the $\ell$, $p$, $k$, $m$, $n$ vanish.
Thus describing how to pull scalars, effects and states in and out of the bubbles.
If $p=0$, the rule describes how to pull in and out diagrams only from the left or only from the right.
We will make heavy use of this in Section~\ref{sec:example_applications}.
\item {Product (Composition) Rule}
The second rule describes how to combine products
(i.e., compositions)
of linear combinations of diagrams.
We state the rule for a product of two linear combinations comprised of two summands each
\begin{align}
&
\biggl(
a \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$m$}{$A$}{$n$}[2.5em]}}
+ \, b \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$m$}{$B$}{$n$}[2.5em]}}
\biggr)
\, \circ \,
\biggl(
c \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$n$}{$C$}{$\ell$}[2.5em]}}
+ \, d \, \cdot \, \;
\vcenter{\hbox{\diagplacehold{$n$}{$D$}{$\ell$}[2.5em]}}
\biggr)
\\[2em]
= &
\vcenter{\hbox{
\externalize{zxsumrules_product_part_comb1}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_products_of_sums_partially_combined_1}
\end{tikzpicture}
}
}}
\, \circ \,
\vcenter{\hbox{
\externalize{zxsumrules_product_part_comb2}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_products_of_sums_partially_combined_2}
\end{tikzpicture}
}
}}
\\[2em]
= &
\vcenter{\hbox{
\externalize{zxsumrules_product_full_comb}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_products_of_sums_fully_combined}
\end{tikzpicture}
}
}} \label{eqn:zxsumrules_product_of_sums} \tag{ii} \,.
\end{align}
The product rule
extends in the obvious way
to the case of more than two factors or summands.
\end{enumerate}
\subsubsection{Additional Rules}%
We introduce several additional rules which
turn out to be useful in the remainder of this work.
\paragraph{Scalar-pull rule}
First, scalars can be pulled through the bubble.
I.e.\ it does not matter if we write them to the left or right of the bubbles.
\begin{equation}
\vcenter{\hbox{
\externalize{zxsumrules_scalar_pull_lhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_scalar_pull_lhs}
\end{tikzpicture}
}
}}
=
\vcenter{\hbox{
\externalize{zxsumrules_scalar_pull_rhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_scalar_pull_rhs}
\end{tikzpicture}
}
}} \label{eqn:zxsumrules_scalarpull} \,.
\end{equation}
\paragraph{Linear combinations for states and effects}
Since we can put the scalar factor left or right of the bubbles, we can simplify linear combinations in the case of states or effects.
For states (no inputs), we can cut the left half of the diagram
\begin{equation}
\vcenter{\hbox{
\externalize{zxsumrules_cut_left_summator_lhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_cut_left_summator_lhs}
\end{tikzpicture}
}
}}
=:
\vcenter{\hbox{
\externalize{zxsumrules_cut_left_summator_rhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_cut_left_summator_rhs}
\end{tikzpicture}
}
}} \label{eqn:zxsumrules_cutleftsummator}\, .
\end{equation}
For effects (no outputs), we can cut the right half of the diagram
\begin{equation}
\vcenter{\hbox{
\externalize{zxsumrules_cut_right_summator_lhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_cut_right_summator_lhs}
\end{tikzpicture}
}
}}
=:
\vcenter{\hbox{
\externalize{zxsumrules_cut_right_summator_rhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_cut_right_summator_rhs}
\end{tikzpicture}
}
}} \label{eqn:zxsumrules_cutrightsummator} \, .
\end{equation}
\paragraph{Direct connection of diagrams (no bubbles)}
We can also completely drop the bubbles and continue the input and output wires through the sum symbols
\begin{equation}
\vcenter{\hbox{
\externalize{zxsumrules_direct_notation_lhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_direct_notation_lhs}
\end{tikzpicture}
}
}}
=:
\vcenter{\hbox{
\externalize{zxsumrules_direct_notation_rhs}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/rules_direct_notation_rhs}
\end{tikzpicture}
}
}} \label{eqn:zxsumrules_direct_notation}\,.
\end{equation}
However, we will not
require this notation
in the examples considered
in this paper.
\subsection{ZX-Calculus for Expectation Values of Quantum Circuits}
In this section, we will present various identities within the extended ZX-calculus framework, that are useful for the analysis of parametrized quantum circuits.
While we primarily consider Pauli operators here, similar results may derived in different
basis or gate sets.
See \cite{cowtan2020generic} for some additional useful rules regarding Pauli operator exponentials.
\subsubsection{Rotations}
First, we can write rotation operators in terms of linear combinations of Clifford gates
\begin{align}
\mathrm{e}^{\mathrm{i} \gamma Z}
&=
\mathrm{e}^{\mathrm{i} \gamma}
\begin{ZX}[ampersand replacement=\&]
\zxN{} \rar \& \zxZ{-2\gamma} \rar \& \zxN{}
\end{ZX}
\\
&=
\vcenter{\hbox{
\externalize{z_spider}{%
\tikz{%
\input{./images/tikz/identities/z_spider.tex}
}
}
}} \label{eqn:zspider_as_lincomb} \, ,
\end{align}
\begin{align}
\mathrm{e}^{\mathrm{i} \beta X}
&=
\mathrm{e}^{\mathrm{i} \beta}
\begin{ZX}[ampersand replacement=\&]
\zxN{} \rar \& \zxX{-2\beta} \rar \& \zxN{}
\end{ZX}
\\
&=
\vcenter{\hbox{
\externalize{x_spider}{%
\tikz{%
\input{./images/tikz/identities/x_spider.tex}
}
}
}} \label{eqn:xspider_as_lincomb} \, .
\end{align}
In both cases the proof easily follows from the identity $\mathrm{e}^{\mathrm{i}\alpha A}=\cos\alpha I +\mathrm{i} \sin\alpha A$ for operators satisfying $A^2=I$.
\subsubsection{Rotation commutation}
In Equation~\eqref{eqn:zx_commute} we saw how to \say{commute} X- and Z-spiders.
As a special case for this, we can rederive the $\pi$-copy rule \ensuremath{(\hyperref[fig:zx_rules]{\bm \pi})}\xspace in order to demonstrate how the extended ZX-calculus works.
For this, we use the diagram pull rule~\eqref{eqn:zxsumrules_pull} from above, as well as~\eqref{eqn:zspider_as_lincomb} and $XZ=-ZX$:
\begin{align}
\vcenter{\hbox{
\externalize{rederive_pi_copy_rule_00}{%
\begin{tikzpicture}[node distance=2ex and 1.5em]
\input{./images/tikz/rederive_pi_copy_rule/00.tex}
\end{tikzpicture}
}
}}
& \stackrel{\eqref{eqn:zspider_as_lincomb}}{=}
\vcenter{\hbox{
\externalize{rederive_pi_copy_rule_01}{%
\begin{tikzpicture}[node distance=2ex and 1.5em]
\input{./images/tikz/rederive_pi_copy_rule/01.tex}
\end{tikzpicture}
}
}} \\[1ex]
& \stackrel{\eqref{eqn:zxsumrules_pull}}{=}
\vcenter{\hbox{
\externalize{rederive_pi_copy_rule_02}{%
\begin{tikzpicture}[node distance=2ex and 1.5em]
\input{./images/tikz/rederive_pi_copy_rule/02.tex}
\end{tikzpicture}
}
}}
\\[1ex]
& \oset[2ex]{\mathmakebox[\widthof{=}]{XZ=-ZX}}{=}
\vcenter{\hbox{
\externalize{rederive_pi_copy_rule_03}{%
\begin{tikzpicture}[node distance=2ex and 1.5em]
\input{./images/tikz/rederive_pi_copy_rule/03.tex}
\end{tikzpicture}
}
}}
\\[1ex]
& \stackrel{\eqref{eqn:zxsumrules_pull}}{=}
\vcenter{\hbox{
\externalize{rederive_pi_copy_rule_04}{%
\begin{tikzpicture}[node distance=2ex and 1.5em]
\input{./images/tikz/rederive_pi_copy_rule/04.tex}
\end{tikzpicture}
}
}}
\\[1ex]
& \stackrel{\eqref{eqn:zspider_as_lincomb}}{=}
\vcenter{\hbox{
\externalize{rederive_pi_copy_rule_05}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/rederive_pi_copy_rule/05.tex}
\end{tikzpicture}
}
}} \, . %
\label{eqn:rederive_pi_copy_rule}
\end{align}
\subsubsection{Phase-Gadgets}
Important for parameterized quantum circuits are multi-qubit rotations, so-called \textit{phase-gadgets} (cf.~\cite{cowtan2020generic}),
for example
\begin{align}
\mathrm{e}^{\mathrm{i} \gamma Z_u Z_v}
&=
\sqrt{2} \mathrm{e}^{\mathrm{i} \gamma}
\begin{ZX}[ampersand replacement=\&]
\zxN{} \rar \& \zxZ{} \rar \dar \& \zxN{} \& \zxN{} \\
\& \zxX{} \rar \& \zxZ{-2 \gamma} \& \\
\zxN{} \rar \& \zxZ{} \rar \uar \& \zxN{} \& \zxN{}
\end{ZX}
\\
&=
\vcenter{\hbox{
\externalize{phase_gadget}{%
\tikz{%
\input{./images/tikz/identities/phase_gadget.tex}
}
}
}} \label{eqn:phase_gadget_as_lincomb} \, .
\end{align}
A proof of the first equality is given in \cite[Corollary 3.4]{cowtan2020generic}.
The second equality is derived similarly to \eqref{eqn:zspider_as_lincomb}, \eqref{eqn:xspider_as_lincomb}.
In particular for the analysis of QAOA expectation values, we will encounter conjugates of phase-gadgets in conjunction with $\pi$-X-spiders.
We will make heavy use of the following identity which is proven in~Appendix~\ref{sec:app_proof_conjugates_cancel_phase_gadgets}.
\begin{equation}\label{eqn:conjugates_cancel_phase_gadgets}
\vcenter{\hbox{
\externalizezx{conjugates_cancel_phase_gadgets_lhs}{%
\input{./images/tikz/phase_gadgets/lhs.tex}
}
}}
=
\vcenter{\hbox{
\externalizezx{conjugates_cancel_phase_gadgets_rhs}{%
\input{./images/tikz/phase_gadgets/rhs.tex}
}
}} \, .
\end{equation}
Phase-gadgets can be combined to implement so-called \textit{phase polynomials},
i.e., parameterized exponentials of diagonal Hamiltonians such as utilized in QAOA circuits~\cite{de2020architecture,cowtan2020generic,hadfield2021representation}.
\subsubsection{Lightcones}
For quantum circuits of limited depth or connectivity,
it is often the case when computing a particular quantity that
a significant fraction of the gates and qubits can be ignored or discarded due to having no effect,
in analogy with spacelike-separated events in relativity.
Naturally, the same principle may be fruitfully applied to diagrammatic analysis.
Given an observable $C=\sum_j C_j$, typically each $C_j$ acts nontrivially on a subset of $\ell<n$ qubits.
Hence, depending on the structure of the problem and given quantum circuit ansatz $U\ket{\psi_0}$, the $n$-qubit expectation
values $\langle C_j \rangle$ may be equivalently reduced to ones over $L$ qubits,
$\ell \leq L \leq n$, by in each case restricting the quantum circuit in the natural way.
This phenomena is known, generally, as the \textit{lightcone} or \textit{causal cone} rule~\cite{evenbly2009algorithms,farhi2014quantum, streif2019training,hadfield2021analytical}, and is clearly exhibited with the ZX-calculus.
For example, if $\ket{\psi_0}$ is a product state and $U$ consists of only $1$-local gates,
then $L=\ell$ independently of the circuit depth (cf. the example of Section~\ref{sec:example_applications_single_qubit}).
For QAOA applied to MaxCut, $\ell=2$ and it is easily shown that the lightcone after each $q$th QAOA layer consists of the restriction to the subgraph within distance $q$ of the given edge~\cite{farhi2014quantum,hadfield2021analytical}, i.e., its size $L$ depends on the vertex degrees in the graph neighborhood.
Hence, importantly, for QAOA or similar layered ansatz we may apply the lightcone rule layer-by-layer.
Applying this restriction, the inner operator for a MaxCut QAOA expectation value reads
\begin{align}
\mathcal{O}^p_{uv}
&:=
\prod_{\ell=1}^p \,
\mathrm{e}^{\mathrm{i}\gamma_\ell C} \mathrm{e}^{\mathrm{i} \tilde \beta_\ell B} \,
Z_u Z_v \,
\prod_{k=p}^1 \,
\mathrm{e}^{-\mathrm{i}\tilde \beta_k B} \mathrm{e}^{-\mathrm{i}\gamma_k C}
\\[2ex]
&=
\vcenter{\hbox{
\externalize{light_cone_rule}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/light_cone_rule.tex}
\end{tikzpicture}
}
}}\label{eqn:light_cone_rule} \, ,
\end{align}
where we used placeholder diagrams for the \textit{reduced} phase-separation layer
\begin{align}
\vcenter{\hbox{
\externalize{ps_layer_definition}{%
\input{./images/tikz/ps_layer_definition/equation.tex}
}
}}\label{eqn:ps_layer_definition} \, ,
\end{align}
and the \textit{reduced} mixing layer
\begin{align}
\vcenter{\hbox{
\externalize{mixing_layer_definition}{%
\input{./images/tikz/mixing_layer_definition/equation.tex}
}
}}\label{eqn:mixing_layer_definition}
\end{align}
where $\beta:=-2\tilde \beta$ for convenience.
For the reduced phase-separation layer we have used the MaxCut cost function Hamiltonian
$
C = \frac{1}{2} \sum_{uv \in E} \left(1 - Z_u Z_v\right)
$
and
\begin{align}
\mathrm{e}^{\mathrm{i} \gamma C}
&=
\prod_{(u,v) \in E} \mathrm{e}^{\frac{\mathrm{i} \gamma}{2}} \mathrm{e}^{\frac{-\mathrm{i} \gamma}{2} Z_u Z_v}
\\
&\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
\sqrt{2}^{|E|}
\prod_{(u,v) \in E}
\begin{ZX}[ampersand replacement=\&]
\zxN{} \rar \&[1mm] \zxZ{} \rar
\& \zxN{} \rar \& \zxN{}
\& u
\\
\& \zxX{} \uar \dar \rar
\& \zxZ{\gamma} \&
\&
\\
\zxN{} \rar \& \zxZ{} \rar
\& \zxN{} \rar \& \zxN{}
\& v
\end{ZX}
\, .
\end{align}
This leads to the factor $\sqrt{2}^{n_L}$ in~\eqref{eqn:ps_layer_definition}, where $n_L$ is the number of phase-gadgets in the right diagram of~\eqref{eqn:ps_layer_definition}.
Also, we used the neighborhood of a set of nodes $L$
\begin{equation}
N_{L} := \bigcup_{\ell \in L} \text{nbhd}(\ell) \, ,
\end{equation}
as well as the exclusive $p$-th neighborhood of $\{u, v\}$, recursively defined by
\begin{align}
\mathcal{N}_{uv}^p &:= \bigcup_{i, j \in \mathcal{N}_{uv}^{p-1} \times \mathcal{N}_{uv}^{p-1} \cap E}
N_{\{i, j\}} \setminus \cup_{k=0}^{p-1} \mathcal{N}_{uv}^k \, , \\
\mathcal{N}_{uv}^0 &:= \{u, v\} \, ,
\end{align}
and the complement of the same
\begin{equation}
\mathcal{M}_{uv}^p = \mathcal{N}_{uv}^p \setminus E \, .
\end{equation}
While here we have considered QAOA circuits as a demonstrative example, the same principle may be applied to or formalized for more general ans\"atze.
\section{Introduction}
\label{sec:introduction}
Diagrammatic approaches to quantum mechanics~\cite{coecke2008interacting,coecke2017picturing,coecke2021kindergarden}
have gained much attention in recent years as an advantageous alternative approach to analyzing and understanding quantum systems,
providing simpler intuition and in some cases improved algorithmic approaches.
These methods provide straightforward rules for representing, manipulating,
and simplifying quantum objects, while at the same time are underpinned by sophisticated mathematical ideas
(in particular, category theory~\cite{abramsky2004categorical,van2020zx}).
An important example is the ZX-calculus~\cite{coecke2008interacting,coecke2011interacting,van2020zx} and its closely related variants~\cite{ranchin2014depicting,wang2014qutrit,hadzihasanovic2015diagrammatic,jeandel2017calculus,hadzihasanovic2018diagrammatic,backens2018zh,east2022aklt} which have seen a number of successful applications in quantum computing,
ranging from circuit optimization~\cite{duncan2020graph,kissinger2020reducing,backens2021there,gorard2021zx}
and synthesis~\cite{cowtan2019phase,de2020architecture},
to algorithm analysis~\cite{carette2021quantum,townsend2021classifying},
natural language processing~\cite{coecke2020foundations} and
machine learning~\cite{yeung2020diagrammatic,toumi2021diagrammatic,zhao2021analyzing}, among others.
In this paper we show how the ZX-calculus is also useful for analyzing
algorithms based on
parameterized quantum circuits (PQCs),
such as variational quantum algorithms,
in particular for
calculating important derived quantities such as expectation values of quantum observables, or their gradients. Such quantities may be computed as functions of the circuit parameters, in which case the parameters are symbolically carried through subsequent ZX-diagrams, or as numbers for the case of fixed parameters of interest.
To enable this, we present several new ZX-rules generalizing the standard ones appearing in the literature; in particular,
we present rules and notation for explicitly handling linear combinations of ZX-diagrams which naturally arise,
for example, when incorporating commutation rules for unitary operators which are used for instance in computing
expectation values.
For linear combination of diagrams, clearly, it is critical to keep track of the scalar multiplier of each diagram,
whereas in previous single-diagram applications such global phases or normalization constants can typically be ignored.
In our application these multipliers will typically be complex-valued functions of the quantum circuit parameters.
Furthermore, our formalism then allows direct importation of a number of useful relations from the operator analysis to ZX calculus setting, such as causal cone and operator commutation rules, among others.
After stating the new rules we demonstrate their efficacy with
several prototypical examples of parameterized quantum circuits in the context of combinatorial optimization,
including straightforward derivation of some new and existing results concerning
example circuits drawn from the literature.
While for computing expectation values of relatively shallow circuits we are able to show most of the key diagram reduction steps explicitly,
for deeper circuits our approach can be aided by integration with software implementations of the ZX-calculus (e.g., \cite{kissinger2015quantomatic,pyzx2019}).
Though we focus on the common task of analyzing quantum circuit expectation values, important in particular for assessing algorithm performance, our proposed rules are general and may find much broader application in future work.
For instance, toward analyzing phenomena related to parameter setting, expectation value gradients may be obtained either by
differentiating directly, or by
reducing the calculation to that of computing further circuit expectation values as in
parameter shift rules~\cite{crooks2019gradients,Wierichs2022generalparameter}.
We emphasize that our approach may be applied to a wide variety of application problems and related quantum circuits beyond those explicitly considered in our examples,
and further
ZX results and
generalizations from the literature may be leveraged,
including extensions to qudits~\cite{wang2014qutrit} or fermions~\cite{hadzihasanovic2018diagrammatic,cowtan2020generic}, among others.
The paper is structured as follows. We next elaborate on our contribution and the connection to
relevant prior work.
In Section~\ref{sec:prelimiaries}, for the benefit of the reader we briefly overview the ZX-calculus and parameterized quantum circuits. In Section~\ref{sec:zx_for_pqc} we show a number of ZX rules, both new and existing, useful for PQCs. Then in Section~\ref{sec:example_applications} we apply these rules to analysis of four different circuit examples taken from the literature. We conclude in Section~\ref{sec:summary_and_outlook} with an outlook on future promising directions.
\subsection{Related work}
Several
recent papers
provide distinct but complementary results towards applying the ZX calculus in the PQC setting.
A recent paper~\cite{zhao2021analyzing} considers using the ZX-calculus for computing and analyzing expectation values of derivatives of the cost expectation for particular classes of random parameterized quantum circuits built from particular gate sets (see in particular~\cite[Assumption 1]{zhao2021analyzing}),
in the context of detecting possible barren plateaus~\cite{mcclean2018barren}.
Our work differs in that we consider expectation values of the cost function themselves, and make no similar assumption of randomly selected parameters.
A particular similarity with~\cite{zhao2021analyzing} is both their application and ours require explicit accounting of scalar factors associated to ZX-diagrams (see Section~\ref{sec:ZXforPQC}).
However, while it is observed in~\cite[Eq. 7]{zhao2021analyzing} that quantum expectation values may represented with the ZX-calculus in~\cite[Eq. 7]{zhao2021analyzing}, the authors do not apply the decomposition $C=\sum_jC_j$,
which we exploit to derive novel ZX rules and analysis.
Our approach and results are complementary to those of~\cite{zhao2021analyzing}.
Two recent papers \cite{wang2022differentiating,jeandel2022addition} have appeared during finalization of this work that consider addition and differentiation of ZX-diagrams.
In particular these papers introduce diagrammatic extensions related to ours.
We believe our work nicely complements these recent results.
The papers \cite{wang2022differentiating,jeandel2022addition} do not deal with expectation values explicitly which is the focus of this work.
A natural next research direction is to combine our and other recent results \cite{zhao2021analyzing,wang2022differentiating,jeandel2022addition} to explore application of ZX-calculus to optimizing parameterized quantum circuits.
In terms of previous applications to variational quantum algorithms, a different diagrammatic approach to constructing parameterized quantum circuits was consider in~\cite{herasymenko2021diagrammatic}.
Another work~\cite{fontana2020optimizing} applied ZX-calculus in analysis of symmetries in the parameter landscape of the cost function expectation.
To our knowledge, no prior work explicitly uses the cost Hamiltonian decomposition $C=\sum_j C_j$ to represent $\langle C\rangle$ as a sum of simpler diagrams, to which the new rules we show particularly apply.
\subsection{Our contributions: A Motivating Example}%
\label{sec:intro_first_example}
Consider the subsequent application of a Z-rotation followed by an X-rotation gate and its equivalent representations as a formula, quantum circuit, or ZX-diagram (we use the convention to read the diagrams and circuits from left to right)
\begin{equation}
\mathrm{e}^{\mathrm{i} \beta X} \mathrm{e}^{\mathrm{i} \gamma Z}
\,
=
\,
\begin{quantikz}[thin lines, row sep=1ex, column sep=0.5em]
\ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_Z(-2\gamma)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_X(-2\beta)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} \\[0.3ex]
\end{quantikz}
\,
\propto
\begin{ZX}
\zxN{} \rar & \zxZ{-2\gamma} \rar & \zxX{-2\beta} \rar & \zxN{}
\end{ZX}
\,. %
\label{eqn:two_single_qubit_rotations}
\end{equation}
Note we use $X,Y,Z$ to denote the single-qubit Pauli matrices throughout.
In order to \say{commute} the X-rotation through the Z-rotation, we need to expand both rotations and apply the commutation rules for Pauli operators
$XZ=-XZ$
\begin{align}
\mathrm{e}^{\mathrm{i} \beta X} \mathrm{e}^{\mathrm{i} \gamma Z}
&=
\left( c_\beta I + \mathrm{i} s_\beta X \right)
\left( c_\gamma I + \mathrm{i} s_\gamma Z \right)
\\
&=
c_\beta c_\gamma I
+ \mathrm{i} c_\beta s_\gamma Z
+ \mathrm{i} s_\beta c_\gamma X
- s_\beta s_\gamma XZ
\\
&=
c_{\gamma} c_\beta I
+ \mathrm{i} c_{\gamma} s_\beta X
+ \mathrm{i} s_{\gamma} c_\beta Z
+ s_{\gamma} s_\beta ZX
\end{align}
Where $s_\alpha := \sin \alpha$ and $c_\alpha := \cos \alpha$.
In most use cases it is preferable to avoid the intricate bookkeeping which comes with the expansion of single qubit rotations as linear combinations of primitive Pauli operators.
However there are occasions where this is unavoidable.
One example is the analytical calculation of expectation values for parametrized quantum circuits.
During such calculations (as we will see below), we repeatedly encounter situations were we need to pull an X-spider from the right through a Z-spider on the left (cf.~\eqref{eqn:two_single_qubit_rotations}), or vice versa, in order to fully contract a diagram.
Only if the angle in one of the spiders is $0$ or $\pi$ does the
conventional ZX-calculus give you the means to perform such an operation within the framework itself, using the $\pi$-copy rule, e.g.\
\begin{equation}
\begin{ZX}
\zxN{} \rar & \zxZ{-2\gamma} \rar & \zxX{\pi} \rar & \zxN{}
\end{ZX}
\stackrel{\footnotesize\PiRule}{=}
\mathrm{e}^{- 2 \mathrm{i} \gamma}
\begin{ZX}
\zxN{} \rar & \zxX{\pi} \rar & \zxZ{2\gamma} \rar & \zxN{}
\end{ZX}
\, .
\end{equation}
However, in order to accommodate the general case, we
introduce a novel diagrammatic extension to the ZX-calculus for the description of linear combinations of diagrams.
\begin{align}
\mathrm{e}^{\mathrm{i} \gamma Z} \mathrm{e}^{\mathrm{i} \beta X}
& \stackrel{\eqref{eqn:zspider_as_lincomb},\eqref{eqn:xspider_as_lincomb}}{=}
\mathrm{e}^{\mathrm{i} \gamma} \mathrm{e}^{\mathrm{i} \beta}
\begin{ZX}
\zxN{} \rar & \zxZ{-2\gamma} \rar & \zxX{-2\beta} \rar & \zxN{}
\end{ZX}
\\[2ex]
& \stackrel{\eqref{eqn:zspider_as_lincomb},\eqref{eqn:xspider_as_lincomb}}{=}
\vcenter{\hbox{
\externalize{zx_commutation_02}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/zx_commutation/02.tex}
\end{tikzpicture}
}
}}
\\[2ex]
& \stackrel{\eqref{eqn:zxsumrules_product_of_sums}}{=}
\vcenter{\hbox{
\externalize{zx_commutation_03}{%
\begin{tikzpicture}[baseline=-0.3em, node distance=5ex and 2.5em]
\input{./images/tikz/zx_commutation/03.tex}
\end{tikzpicture}
}
}} %
\stackrel{\footnotesize\PiRule}{=}
\vcenter{\hbox{
\externalize{zx_commutation_04}{%
\begin{tikzpicture}[node distance=5ex and 2.5em]
\input{./images/tikz/zx_commutation/04.tex}
\end{tikzpicture}
}
}}
\\[-6ex]
\label{eqn:zx_commute}
\\[4ex]
\end{align}
The details of this calculation will be explained below.
With our extension, we are able to pull all grey X-spiders from the right to the left.
Later on, we will see that this is necessary in order to fully contract diagrams representing expectation values of parametrized quantum circuits.
\section{Summary and Outlook}
\label{sec:summary_and_outlook}
We introduced an extension of the ZX-calculus to
conveniently incorporate linear combinations of ZX-diagrams.
Moreover we demonstrated how this generalized diagrammatic framework can be applied to the analysis of parametrized quantum circuits,
in particular to the calculation of observable expectation values.
Further quantities of interest such as gradients may be similarly derived,
as well as more complicated PQC phenomenon such as barren plateaus studied,
by combining our framework with several distinct but complementary recent ZX-calculus advances~\cite{zhao2021analyzing,jeandel2022addition,wang2022differentiating}.
Future research could incorporate our approach with further variants of ZX-calculus,
like the ZH-calculus~\cite{backens2018zh} or the ZX-framework for qudits~\cite{ranchin2014depicting,wang2014qutrit}.
In particular the latter could facilitate novel insights into performance analysis of quantum alternating operator ans\"atze~\cite{hadfield2019quantum} for problems like graph-coloring~\cite{wang2020xymixers} and beyond~\cite{stollenwerk2020toward}.
Similarly, our approach could be likewise applied to applications beyond combinatorial optimization, like variational quantum eigensolvers for quantum
chemistry applications~\cite{cowtan2020generic}.
Generally, it is of interest to explore to what extent diagrammatic approaches may ultimately aid in the design and analysis of better performing parameterized quantum circuit ans\"atze,
as well as help with important related challenges such as alleviating the cost of parameter setting, avoiding undesirable features such as barren plateaus, or tailoring ansatz design to a given set of hardware constraints.
\section{Example Applications to Combinatorial Optimization}%
\label{sec:example_applications}
Expectation values of quantum circuit observables -- i.e., constants--may be represented with ZX-diagrams,
as has been previously observed in \cite[Eq. 7]{zhao2021analyzing}.
In doing so, in some cases the structure of the original problem
may be directly reflected in the structure of the corresponding ZX-diagrams.
This is demonstrated by the examples of this section, in which
apply our ZX-calculus extension to
calculate cost expectation values for several different ans\"atze for combinatorial optimization.
The purpose of this section is twofold.
First, we want to demonstrate that calculations with parametrized quantum circuits, like the finding an analytical expression for expectation values, can sometimes become
more intuitive and simplified by using ZX-calculus in conjunction with our extension to linear combinations.
Second, we show that our extension is indeed necessary to achieve the aforementioned task diagrammatically by, for instance, providing means to \say{commute} X- and Z-spiders (cf.~\eqref{eqn:zx_commute} or~\eqref{eqn:rederive_pi_copy_rule}), while explicitly keeping track of all resulting terms.
For each of the examples to follow we show how the cost function expectation value $\langle C\rangle$ may be computed and analyzed using our extended ZX-calculus. Recall that given a decomposition of the cost Hamiltonian $C=\sum C_\ell$ it suffices to compute the $\langle C_\ell \rangle$ values independently, which typically correspond to similar diagrams. In particular (sub)graph symmetry can be
exploited to reduced the number of unique diagrams required \cite{farhi2014quantum,shaydulin2021classical,shaydulin2021exploiting}.
Generally the quantity $\langle C\rangle$ is important in parameter setting, as well as bounding algorithm performance in terms of important quantities of interest such as the approximation ratio achieved~\cite{hadfield2018quantum}.
\subsection{Independent Single-Qubit Rotations Ansatz}%
\label{sec:example_applications_single_qubit}
We begin with
a simple but important example.
Consider an arbitrary cost function and corresponding cost (diagonal) Hamiltonian $C$ on $n$ qubits we seek to extremize, together with the simple depth-1 ansatz consisting of a free single-qubit Pauli-$Y$ rotation on each qubit, applied to the initial state $\ket{00\dots 0}=\ket{0}^{\otimes n}$,
\begin{equation}
\vcenter{\hbox{
\begin{quantikz}[thin lines, row sep={1.0ex}]
\lstick{\ket{0}} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_Y(\alpha_1)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} \\
\lstick{\ket{0}} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_Y(\alpha_2)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} \\
\vdots & & \vdots & \\
\lstick{\ket{0}} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_Y(\alpha_n)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
\end{quantikz}
}}
\,\,=\,\,
\frac{1}{\sqrt{2^n}}
\vcenter{\hbox{
\begin{ZX}[row sep=1.9ex, column sep=1em]
\zxX{} \rar & \zxFracZ-{\pi}{2} \rar & \zxX{\alpha} \rar & \zxFracZ{\pi}{2} \rar & \zxN{} \\
\zxX{} \rar & \zxFracZ-{\pi}{2} \rar & \zxX{\alpha} \rar & \zxFracZ{\pi}{2} \rar & \zxN{} \\
\vdots & & \vdots & & \\
\zxX{} \rar & \zxFracZ-{\pi}{2} \rar & \zxX{\alpha} \rar & \zxFracZ{\pi}{2} \rar & \zxN{}
\end{ZX}
}}\label{eqn:single_qubit_rotation_ansatz} \,.
\end{equation}
Further consider, for example, an arbitrary instance of MaxCut; though the same argument we show here applies
similarly to many other problems. For a graph with edge set $E$ the cost Hamiltonian is $C=\tfrac{|E|}2-\tfrac12\sum_{(ij)\in E}Z_iZ_j$.
As shown below in Equation~\eqref{eqn:single_qubit_rotation_exp_value} the derivation of
$\langle Z_u Z_v \rangle$
becomes very simple with ZX-calculus.
\def0.5em{0.5em}
\tikzset{dots/.style={%
scale=0.6,
ultra thick}
}
\begin{equation}
\langle Z_u Z_v \rangle
\;
=
\begin{tikzpicture}[baseline=8ex, node distance=4ex and 2.0em]
\input{./images/tikz/single_qubit_rotation/derivation.tex}
\end{tikzpicture}
\label{eqn:single_qubit_rotation_exp_value}\,.
\end{equation}
Here, again, $s_\alpha=\sin(\alpha)$ and $c_\alpha=\cos(\alpha)$.
Note that after the second step the usage of linear combinations to handle the X-spider with phase $(-2\alpha_u)$ provides a way to continue the calculation, which would not be possible within the conventional ZX-framework. From the permutation symmetry of the ansatz, the expectation value $\langle Z_i Z_j\rangle$ of each edge is the same~\cite{shaydulin2021classical}.
Hence we have
\begin{equation} \label{eq:npansatz}
\langle C\rangle= \tfrac{|E|}2- \tfrac12 \sum_{(u, v)\in E} \cos(\alpha_u) \cos(\alpha_v),
\end{equation}
which implies
\begin{equation}
\max_{\alpha} \langle C\rangle = \max_{\alpha\in\{0,\pi\}^n} \langle C\rangle = \max_x c(x)= c(y^*),
\end{equation}
where angles $\alpha^* \in \{0,\pi\}^n$ encode a string $y^*$ via $y^*_i=\tfrac12-\tfrac12 \cos(\alpha^*_i)$.
Hence, as any optimal angles directly encode an optimal solution to the MaxCut instance, the expectation value $\langle C\rangle$ is NP-hard to optimize.
Indeed, for MaxCut, \eqref{eq:npansatz} reproduces the quantity of Equation~1 of~\cite{bittel2021training} (up to an affine shift).
This result is used throughout~\cite{bittel2021training} via further reductions to show that optimizing a number of other classes of PQCs is NP-hard.
We have similarly demonstrated that the single-qubit rotations ansatz is NP-hard to optimize for problems such as MaxCut, but via a compact derivation using ZX-diagrams.
\subsection{Hardware Efficient Ansatz}%
\label{sec:example_applications_hweff}
The second example we consider is a
variant of a hardware efficient SU-2 2-local ansatz from Qiskit~\cite{qiskit2019}.
This ansatz was also studied in~\cite{funcke2021dimensional}.
For simplicity here we consider a 3-qubit realization,
\begin{align}
&
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_example_ansatz_00}{%
\input{./images/tikz/hw_eff_ansatz_example/ansatz_00.tex}
}
}}
\\[4ex]
&=
\frac{1}{\sqrt{2^3}}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_example_ansatz_01}{%
\input{./images/tikz/hw_eff_ansatz_example/ansatz_01.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\SpiderRule,\footnotesize\CopyRule}{=}
\frac{1}{\sqrt{2^3}}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_example_ansatz_02}{%
\input{./images/tikz/hw_eff_ansatz_example/ansatz_02.tex}
}
}}\label{eqn:hw_eff_ansatz} \, ,
\end{align}
where we conveniently set
$\gamma_{ij} := \tilde \gamma_{ij} + \frac{\pi}{2}$
and
$\beta_{ij} := \frac{-\tilde \beta{ij}}{2}$.
To compute the expectation value of
a given cost Hamiltonian, we
we again requite expectation values of products of
Pauli-Z operators.
We demonstrate how such calculations can be performed diagramatically by considering again MaxCut as an example. For the expectation value corresponding to a given edge (23) we have
\begin{align}
&\langle Z_2 Z_3 \rangle
=
\\[4ex]
&
\frac{1}{{2^3}}
\vcenter{\hbox{
\externalize{hw_eff_ansatz_example_01}{%
\begin{tikzpicture}[node distance=8ex and 2.0em]
\input{./images/tikz/hw_eff_ansatz_example/01.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\footnotesize\SpiderRule, \footnotesize\PiRule}{=}
&
\frac{1}{{2^3}}
\vcenter{\hbox{
\externalize{hw_eff_ansatz_example_02}{%
\begin{tikzpicture}[node distance=8ex and 2.0em]
\input{./images/tikz/hw_eff_ansatz_example/02.tex}
\end{tikzpicture}\label{eq:hwexpeczz}
}
}}
\\[4ex]
\stackrel{\eqref{eqn:xspider_as_lincomb}}{=}
&
\frac{1}{{2^3}}
\vcenter{\hbox{
\externalize{hw_eff_ansatz_example_03}{%
\begin{tikzpicture}[node distance=12ex and 2.0em]
\input{./images/tikz/hw_eff_ansatz_example/03.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\footnotesize\SpiderRule, \eqref{eqn:zxsumrules_pull}}{=}
&
\frac{1}{{2^3}}
\vcenter{\hbox{
\externalize{hw_eff_ansatz_example_04}{%
\begin{tikzpicture}[node distance=12ex and 1.0em]
\input{./images/tikz/hw_eff_ansatz_example/04.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\substack{\footnotesize\CopyRule\\ \footnotesize\PiRule \\ \footnotesize\SpiderRule}}{=}
&
\frac{1}{{2^3}}
\vcenter{\hbox{
\externalize{hw_eff_ansatz_example_05}{%
\begin{tikzpicture}[node distance=14ex and 1.0em]
\input{./images/tikz/hw_eff_ansatz_example/05.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\eqref{eqn:zxidentity_hweffansatz}}{=}
\input{./images/tikz/hw_eff_ansatz_example_result.tex} \, ,
\end{align}
where in the last step we have used the identity
\begin{align}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_lhs}{%
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_00.tex}
}
}}
& =
\frac{f_{m_2 m_3}^{r_1 r_2 r_3}}{2^3}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_rhs}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_09.tex}
}
}}
\\
&=
\begin{cases}
0 & \text{if } \quad
\begin{aligned}
& \ell_1 + m_2 + \ell_3 + m_3 + r_3 \text{ odd} \\
\vee \, & r_1 + m_2 + \ell_3 + m_3 + r_3 \text{ odd} \\
\vee \, & \ell_2 + m_2 + r_2 \text{ odd}
\end{aligned} \\
f_{m_2 m_3}^{r_1 r_2 r_3}
& \text{else}
\end{cases}%
\label{eqn:zxidentity_hweffansatz} \, ,
\end{align}
with
$
\ell_1,
\ell_2,
\ell_3,
m_2,
m_3,
r_1,
r_2,
r_3
\in \{0, 1\}^{\times 8}
$
and
$
f_{m_2 m_3}^{r_1 r_2 r_3}
:=
(-1)^{m_2 r_1 + (m_2 \oplus m_3) r_2 + m_3 r_3}
$,
which is the proven in Appendix~\ref{sec:app_proof_zxidentity_hweffansatz}.
Observe that in the second step above any dependency on the parameters $\gamma_{12},\gamma_{22}$, and $\gamma_{32}$ was immediately
shown to cancel out (due to commuting with the diagonal cost Hamiltonian),
and likewise for $\beta_{12}$ (due to the locality of $Z_2Z_3$).
Similar simplifications are often easily obtained from the diagrammatic perspective.
The formula \eqref{eq:hwexpeczzformula} exemplifies the significant difficulty faced in obtaining analytical results for
PQCs, even for relatively small ansatz.
Nevertheless, in our analysis the complexity remained manageable with the diagrammatic approach up until the very last step, were we applied a simple numerical procedure to collect all the surviving terms (according to~\eqref{eqn:zxidentity_hweffansatz}) of the contraction.
Different hardware-efficient ans\"atze may be similarly considered, including ones tailored to specific hardware topology.
As mentioned, for deeper or more complicated ans\"atze, analysis may
be aided or automated through implementation in software.
Here \eqref{eq:hwexpeczz} demonstrates how diagrammatic approaches can yield more compact representations of expectation values (as compared to \eqref{eq:hwexpeczzformula}).
\subsection{QAOA$_1$ for MaxCut on Simple Graph}%
\label{sec:example_applications_qaoa_example_graph}
Next we turn to QAOA~\cite{farhi2014quantum,hadfield2019quantum},
for which we continue our use of MaxCut as a running example.
For simplicity we consider QAOA$_1$, the lowest depth realization,
which is indicative of the $p>1$ case due to the alternating structure of the ansatz.
Recall that for a QAOA state the MaxCut expectation value reads
$\langle C \rangle
= \tfrac{|E|}{2} - \tfrac12\sum_{i, j\in E} \langle Z_i Z_j \rangle \, .$
We begin with the specific graph $G$
of Figure~\ref{fig:qaoa_exp_value_simple_example}, before we consider ring graphs in Section~\ref{sec:example_applications_qaoa_ring}, and arbitrary graphs in Appendix~\ref{sec:app_qaoa}.
\begin{figure}[htpb]
\centering
\externalize{qaoa_max_cut_example}{%
\begin{tikzpicture}
\node (lhs) {$G$ \; \; = };
\node[right=2pt of lhs] {%
\begin{tikzpicture}[scale=1.0]
\node[draw, circle, label={left:\small 1}, inner sep=2pt] at (0, 0) (1) {};
\node[draw, circle, label={right:\small 2}, inner sep=2pt] at (1, -1) (2) {};
\node[draw, circle, label={left:\small 3}, inner sep=2pt] at (-1, -2) (3) {};
\node[draw, circle, label={right:\small 4}, inner sep=2pt] at (0, -3) (4) {};
\draw (1) -- (2);
\draw (2) -- (3);
\draw (3) -- (4);
\draw (2) -- (4);
\end{tikzpicture}
};
\end{tikzpicture}
}
\caption{Simple example graph to consider for MaxCut with QAOA}%
\label{fig:qaoa_exp_value_simple_example}
\end{figure}
Observe how the structure of the graph directly reappears in the diagrams below,
which reflects the fact that the QAOA phase operator is derived from the cost Hamiltonian.
For deeper QAOA circuits, the graph structure will again appear at each layer in the diagrammatic representation.
Hence ZX-calculus provides a toolkit toward directly incorporating or better understanding the relationship between the cost function and a given parameterized quantum algorithm.
Here we demonstrate the edge expectation value calculation for QAOA$_{1}$,
\begin{align}
\langle Z_2 Z_3 \rangle_{\text{QAOA}_1}
= &
\frac{2^4}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_01}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_01.tex}
\end{tikzpicture}
}
}}\\
\stackrel{\footnotesize\SpiderRule \footnotesize\PiRule}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_02}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_02.tex}
\end{tikzpicture}
}
}}\\
\stackrel{\eqref{eqn:xspider_as_lincomb}}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_03}{%
\begin{tikzpicture}[node distance=9ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_03.tex}
\end{tikzpicture}
}
}}\\
\stackrel{\eqref{eqn:zxsumrules_pull}~\eqref{eqn:zxsumrules_product_of_sums}}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_04}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_04.tex}
\end{tikzpicture}
}
}}\\
\stackrel{\eqref{eqn:zxsumrules_pull}}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_05}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_05.tex}
\end{tikzpicture}
}
}}\label{eqn:qaoa_max_cut_example_terms}\\
\stackrel{%
\eqref{eqn:qaoa_max_cut_example_summand_ii}%
\eqref{eqn:qaoa_max_cut_example_summand_ix}%
\eqref{eqn:qaoa_max_cut_example_summand_xi}%
\eqref{eqn:qaoa_max_cut_example_summand_xx}%
}{=}
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_06}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_06.tex}
\end{tikzpicture}
}
}}\\
=
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_07}{
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/qaoa_max_cut_example_07.tex}
\end{tikzpicture}
}
}} \, .
\end{align}
The remaining expectation values can be similarly computed for each of the edges in $E$ to give $\langle C\rangle$.
In the third step above, we could not have easily continued within the conventional ZX-calculus framework.
Whenever one needs to pull parametrized X-spiders through parametrized Z-spiders or vice versa, our extension is utilized.
The detailed calculation of the four contributions used in the last step
is given in Appendix~\ref{sec:app_qaoa_example_graph}.
Note that calculation of the general $n$-qubit case (cf.~Appendix~\ref{sec:app_qaoa}) is surprisingly concise compared to the special case of 4-qubits considered here.
\subsection{QAOA for MaxCut on Ring graphs}%
\label{sec:example_applications_qaoa_ring}
Here we consider the simple example of the one-dimensional \say{ring-of-disagrees}, i.e., 2-regular connected graphs, and rederive the
QAOA$_1$ expectation value as previously shown in~\cite{farhi2014quantum,wang2018quantum}.
First consider the case of QAOA with arbitrary number of layers~$p$, with $n\gg p$.
From the problem symmetry, it suffices to consider the expectation value of a single edge term $\langle Z_iZ_{i+1} \rangle_{QAOA_p}$.
Applying the lightcone rule~\eqref{eqn:light_cone_rule},
the outermost
reduced phase-separation layer%
~\eqref{eqn:ps_layer_definition}
reads
\begin{equation}
\vcenter{\hbox{
\externalize{qaoa_ring_ps_layer}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 3.0em]
\def0.5em{0.5em}
\def2.0em{2.0em}
\def-30{-30}
\def0.2em{0.2em}
\tikzset{dots/.style={%
scale=1.0,
ultra thick}
}
\input{./images/tikz/qaoa_ring/ps_layer_lhs.tex}
\node[name=anchor, right=2em of 1-r, invisible];
\node[anchor=center] at ($(7-r.center)!0.5!(anchor.center)$) {$=$};
\input{./images/tikz/qaoa_ring/ps_layer_rhs.tex}
\end{tikzpicture}
}
}} \label{eqn:qaoa_ring_ps_layer} \, .
\end{equation}
Hence for
the QAOA$_p$ expectation value we obtain
\begin{align}
&\langle Z_i Z_{i+1} \rangle_{\text{QAOA}_p}
\\[4ex]
&=
\frac{1}{2^{2(p+1)}}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_general_01}{%
\begin{tikzpicture}[node distance=6ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_general/00.tex}
\end{tikzpicture}
}
}} \,.
\end{align}
Observe how the problem and
structure
again appears in the above diagram (i.e., $p$-neighborhoods of the edge $(i, i+1)$ are line graphs).
Furthermore, the utility of the lightcone rule is clearly demonstrated here.
Continuing for the $p=1$ case, we get
\begin{align}
\langle Z_i Z_{i+1} \rangle_{\text{QAOA}_1}
&=
\frac{2^3}{2^4}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_01}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/01.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\footnotesize\PiRule, \footnotesize\SpiderRule}{=}
\frac{\mathrm{e}^{-2\mathrm{i} \beta }}{2}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_02}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/02.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\eqref{eqn:xspider_as_lincomb}}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_03}{%
\begin{tikzpicture}[node distance=10ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/03.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\eqref{eqn:zxsumrules_pull}, \eqref{eqn:zxsumrules_product_of_sums}}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_04}{%
\begin{tikzpicture}[node distance=10ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/04.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\eqref{eqn:zxsumrules_pull}}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_05}{%
\begin{tikzpicture}[node distance=10ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/05.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_06}{%
\begin{tikzpicture}[node distance=10ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/06.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\substack{
\footnotesize\PiRule\\
\eqref{eqn:conjugates_cancel_phase_gadgets}
}}{=}
\frac{\mathrm{e}^{2\mathrm{i}\gamma}}{2^3}
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_07}{%
\begin{tikzpicture}[node distance=10ex and 2.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/07.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
-2 \mathrm{i} c_{\beta} s_{\beta}
\frac{\mathrm{e}^{2\mathrm{i}\gamma}}{2^3}
\;
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_08}{%
\begin{tikzpicture}[node distance=5ex and 1.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/summand_ix/02.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
\frac{- 2 \mathrm{i} c_{\beta} s_{\beta}}{2^4}
\;
\vcenter{\hbox{
\externalize{qaoa_ring_exp_val_p1_09}{%
\begin{tikzpicture}[node distance=5ex and 1.5em]
\input{./images/tikz/qaoa_ring/exp_val_p1/summand_ix/03.tex}
\end{tikzpicture}
}
}}
\\[2ex]
&
=
2 c_{\beta} s_{\beta}
s_{\gamma} c_{\gamma}
\end{align}
The result is consistent with that of~\cite[Thm. 1]{wang2018quantum}.
The expression obtained for $\langle C\rangle$ is easily optimized to reproduce the performance result obtained numerically for the ring of disagrees in~\cite{farhi2014quantum}.
Similar to the previous examples, here we saw the necessity of our extension for handling X-Z commutations in the third and seventh steps of the derivation above.
In Appendix~\ref{sec:app_qaoa} we show the same calculation for MaxCut on general graphs,
as obtained for QAOA$_1$ in \cite[Thm. 1]{wang2018quantum}. Similar techniques may be applied and results
obtained for a wide variety of important problems, for instance quadratic binary optimization problems of which MaxCut is a special case.
\section{Preliminaries}%
\label{sec:prelimiaries}
\subsection{ZX-Calculus}
Here we very briefly
overview some
primary aspects of the ZX-calculus; we refer the reader to~\cite{vilmart2018near,van2020zx} and the references therein for comprehensive introductions, including complete sets of graphical rewrite rules
as well as their mathematical details.
Note that throughout we
employ grayscale notation for ZX diagrams as used in~\cite{van2020zx} instead of the original red and green diagram colorings.
The ZX-calculus allows translation between
linear maps,
such as quantum circuits and expectation values,
and \textit{string diagrams} that equivalently represent or encode the same object.
After translation, ZX-diagrams may be
conveniently rearranged or simplified using a sets of standard,
mathematically rigorous
diagram manipulation rules.
While in the usual circuit diagram representation
quantum circuits
correspond to directed acyclic graphs with nodes labeled by a unitary quantum gate, string diagrams are more general
and hence facilitate a
potentially more powerful approach.
In particular, string diagrams correspond to \textit{undirected} graphs, i.e., ZX-diagrams and sub-objects do not inherit the temporal ordering implicit in
quantum circuit diagrams.
Hence, while a quantum circuit can always be efficiently translated to an equivalent ZX-diagram, the converse is not true in general~\cite{duncan2020graph}.
We note that while ZX-diagrams can incorporate the density matrix picture of quantum mechanics, for our purposes it will suffice to consider pure states.
The basic building blocks of a ZX-diagram are so-called \textit{Z-spiders} and \textit{X-spiders}, denoted as
\begin{align}
\begin{ZX}[ampersand replacement=\&]
\leftManyDots{m} \zxZ{\theta} \rightManyDots{n}
\end{ZX}
& :=
\underbrace{\ket{0\dots 0}}_{n}\underbrace{\bra{0\dots 0}}_m + e^{\mathrm{i}\theta}\underbrace{\ket{1\dots 1}}_{n}\underbrace{\bra{1\dots 1}}_m %
\label{eqn:def_z_spider}
\\
\begin{ZX}[ampersand replacement=\&]
\leftManyDots{m} \zxX{\theta} \rightManyDots{n}
\end{ZX}
& :=
\underbrace{\ket{+\dots +}}_{n}\underbrace{\bra{+\dots +}}_m + e^{\mathrm{i}\theta}\underbrace{\ket{-\dots -}}_{n}\underbrace{\bra{-\dots -}}_m %
\label{eqn:def_x_spider}
\end{align}
Here we employ the common approach of defining various ZX-diagram objects in terms of vectors and matrices in the usual bra-ket notation; later it will suffice to consider and manipulate ZX-diagrams alone.
Here the numbers of input and output lines $n,m\geq 0$ can be the same or different. Roughly speaking, each line corresponds to a qubit,
though this association may be broken by subsequent diagram manipulations. Note that spiders are symmetric tensors and as such are invariant under same-side permutations of wires.
As special cases of~\eqref{eqn:def_z_spider} and~\eqref{eqn:def_x_spider} for a single qubit, $Z$ and $X$ Pauli operators and their (unnormalized) eigenstates can be denoted as
\begin{align}
\begin{ZX}
\zxN{} \rar &[1mm] \zxZ{\pi} \rar & [1mm] \zxN{}
\end{ZX}
& =
\ket{0}\bra{0} - \ket{1}\bra{1}
=
Z
&
\begin{ZX}
\zxN{} \rar &[1mm] \zxX{\pi} \rar & [1mm] \zxN{}
\end{ZX}
& =
\ket{0}\bra{1} + \ket{1}\bra{0}
=
X
\\
\begin{ZX}
\zxZ{} \rar & [1mm,1mm] \zxN{}
\end{ZX}
& =
\ket{0} + \ket{1}
=
\sqrt{2} \ket{+}
&
\begin{ZX}
\zxX{} \rar & [1mm,1mm] \zxN{}
\end{ZX}
& =
\ket{+} + \ket{-}
=
\sqrt{2} \ket{0}
\\
\begin{ZX}
\zxZ{\pi} \rar & [1mm] \zxN{}
\end{ZX}
& =
\ket{0} - \ket{1}
=
\sqrt{2} \ket{-}
&
\begin{ZX}
\zxX{\pi} \rar & [1mm] \zxN{}
\end{ZX}
& =
\ket{+} - \ket{-}
=
\sqrt{2} \ket{1}
\end{align}
The Hadamard gate is given a special symbol and denoted
$H =
\begin{ZX}
\zxN{} \rar & [1mm] \zxH{} \rar & [1mm] \zxN{}
\end{ZX}
=
\mathrm{e}^{-\mathrm{i} \frac{\pi}{4}}
\begin{ZX}
\zxN{} \rar & \zxFracZ{\pi}{2} \rar & \zxFracX{\pi}{2} \rar & \zxFracZ{\pi}{2} \rar \rar & \zxN{}
\end{ZX}
$
.
Other single-qubit gates may be derived similarly by combining these primitives.
For multi-qubit gates, like the CNOT gate, we have
\begin{equation}
\text{CNOT}
=
\begin{quantikz}[baseline=-2pt, thin lines, row sep=3ex, column sep=1em]
\ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ctrl{1} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} \\
\ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \targ{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
\end{quantikz}
=
\sqrt{2}
\begin{ZX}
\zxN{} \rar &[1mm] \zxZ{} \dar \rar & [1mm] \zxN{}
\\[2ex]
\zxN{} \rar &[1mm] \zxX{} \rar & [1mm] \zxN{}
\end{ZX}
\end{equation}
A number of the most important ZX-diagram rewrite rules are displayed in Figure~\ref{fig:zx_rules}. We use the label attached to each equation to reference these rules when we apply them in the examples we consider below.
\begin{figure}[htpb]
\centering
\begin{tikzpicture}
\input{./images/tikz/zx_rules.tex}
\end{tikzpicture}
\caption{The ZX-diagram rewrite rules (cf.\ for example~\cite{van2020zx} or~\cite{zhao2021analyzing}). Note that we explicitly keep the phase factors.}%
\label{fig:zx_rules}
\end{figure}
\subsection{Parameterized Quantum Circuits}
Parameterized quantum circuits (PQC) have gained much attention in recent years, in particular as heuristic approaches suitable for NISQ~\cite{preskill2018quantum} era devices that are classically optimized (often variationally) as part of a hybrid
protocol, though we emphasize they are by no means restricted to this setting; see \cite{cerezo2021variational,bharti2022noisy} for reviews of recent developments.
Two particular approaches of interest are the QAOA (quantum alternating operator ansatz~\cite{hadfield2019quantum}, which generalizes the quantum approximate optimization algorithm~\cite{farhi2014quantum}) and VQE (variational quantum eigensolver~\cite{peruzzo2014variational,mcclean2016theory}) paradigms,
as well as a number of more recent variants of these approaches.
Here we briefly review the original QAOA paradigm and its application to combinatorial optimization, though our results to follow may be applied more generally to a variety of problems and algorithms.
In QAOA we are given a cost function $c(x)$ and corresponding classical Hamiltonian $C$ (i.e., diagonal in the computational basis, $C\ket{x}=c(x)\ket{x}$) we seek to optimize over bit strings $x\in\{0,1\}^n$.
A QAOA$_p$ circuit has
consists of $2p$ alternating layers
specified by $2p$ angles $\gamma_i,\beta_i$ in some domain (e.g.~$[-\pi,\pi]$) to create the state
\begin{equation}
\ket{\boldsymbol{\gamma \beta}}= U_M(\beta_p)U_P(\gamma_p)\dots U_M(\beta_1)U_P(\gamma_1)\ket{s} \, ,
\end{equation}
for phase operator $U_P(\gamma)=\exp(-\mathrm{i} \gamma C)$, (transverse-field) mixing operator $U_M(\beta)=\exp(-\mathrm{i} \beta B)$ where $B=\sum_{i=1}^n X_i$, and standard initial product state $\ket{s}=\ket{+}^{\otimes n}$. The state is then measured in the computational basis
which returns some $y\in\{0,1\}^n$ achieving cost $c(y)$.
Figure~\ref{fig:example_pqc} shows a simple example of a QAOA circuit.
\begin{figure}[htpb]
\centering
\begin{quantikz}[thin lines, row sep=1ex, column sep=1em]
\ket{0} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{H} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ctrl{1} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
& \ctrl{1} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
& \ \ldots\ \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_X(\beta)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ \ldots\ \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} \\
\ket{0} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{H} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \targ{}
& \gate{R_Z(\gamma)} & \targ{} & \ctrl{1} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ctrl{1} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
& \ \ldots\ \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_X(\beta)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ \ldots\ \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} \\
\ket{0} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{H} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
& \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \targ{} & \gate{R_Z(\gamma)} & \targ{} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
& \ \ldots\ \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \gate{R_X(\beta)} & \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{} & \ \ldots\ \ifthenelse{\the\pgfmatrixcurrentcolumn>1}{\arrow[arrows]{l}}{}
\end{quantikz}
\caption{Example of a parametrized quantum circuit: QAOA on 3 qubits. Here the phase and mixing operators as well as initial state preparation have been compiled to basic quantum gates.}%
\label{fig:example_pqc}
\end{figure}
Repeated state preparation and measurement gives further samples which may be used to estimate the cost expectation $\langle C\rangle_p$ or other important quantities.
These quantities may be used to update or search for better circuit parameters if desired; we emphasize that in different cases parameters may be found through analytic~\cite{wang2018quantum}, numeric~\cite{farhi2014quantum}, or average-case~\cite{streif2019training} techniques, or, distinctly, searched for empirically
(e.g., variationally).
After a set number of runs overall, or when other suitable termination criteria has been reached, the best solution found is returned.
A fundamentally important quantity for QAOA as well as related approaches is the cost expectation value $\langle C\rangle$, which can be used to bound the expected approximation ratio achieved~\cite{farhi2014quantum,hadfield2018quantum,hadfield2019quantum,hadfield2021analytical}.
Importantly, we are often given a decomposition of the cost Hamiltonian such as $C=\sum_j C_j$ which we may exploit in computing $\langle C\rangle = \sum_j \langle C_j\rangle$ as a sum of terms (typically, a linear combination of Pauli $Z$ operators~\cite{hadfield2021representation}), which directly motivates the
rules we introduce for accommodating linear combinations of ZX-diagrams.
For combinatorial optimization the $C_j$ terms mutually commute which leads to further simplifications, whereas this may not be true for more general problems and applications such as quantum chemistry (though linearity of expectation still applies).
In general, many quantities of interest for PQCs can be expressed as expectation values and are hence amenable to similar analysis via diagrammatic techniques as we explore below.
\section{Appendix}
Here we diagrammatically
derive the QAOA$_1$ expectation value for MaxCut on general graphs, and
provide proofs of several results given in the main text.
\subsection{QAOA$_1$ for MaxCut on General Graphs}%
\label{sec:app_qaoa}
For the QAOA expectation value for MaxCut $
\langle C \rangle
= \tfrac{|E|}{2} - \tfrac12\sum_{u, v\in E} \langle Z_u Z_v \rangle
$ we need to calculate the contributions $\langle Z_u Z_v \rangle$.
In this section, we perform the calculation for general graphs in the QAOA, $p=1$ case, reproducing results obtained in \cite{wang2018quantum}.
Following the lightcone rule from Equation~\eqref{eqn:light_cone_rule} we obtain for $Z$-$Z$ terms in the MaxCut QAOA$_1$ expectation value on a general graph $G=(V, E)$
\begin{align}
\langle Z_u Z_{v} \rangle_{\text{QAOA}_1}
&
\stackrel{\eqref{eqn:light_cone_rule}}{=}
\frac{1}{2^{|V|}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_00}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/00.tex}
\end{tikzpicture}
}
}}
\\[4ex]
&
\stackrel{\substack{%
\eqref{eqn:mixing_layer_definition}\\
\footnotesize\PiRule%
}}{=}
\frac{\mathrm{e}^{-2\mathrm{i}\beta}}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_01}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/01.tex}
\end{tikzpicture}
}
}}
\\[4ex]
&
\stackrel{\substack{%
\eqref{eqn:xspider_as_lincomb}\\
\eqref{eqn:zxsumrules_product_of_sums}%
}}{=}
\frac{1}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_02}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/02.tex}
\end{tikzpicture}
}
}}
\\[4ex]
&
\stackrel{\eqref{eqn:zxsumrules_pull}}{=}
\frac{1}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_03}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/03.tex}
\end{tikzpicture}
}
}} \, .
\end{align}
The first summand vanishes and the second and third are linked by symmetry.
We continue with the second summand (the \textit{I-X}-term)
\begin{align}
&
\frac{1}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_00}{
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/00.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\eqref{eqn:ps_layer_definition}}{=}
&
\frac{2^{n_u + 2 n_{uv} + n_v + 1}}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_01}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 1.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/01.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\substack{%
\footnotesize\PiRule\\
\footnotesize\CopyRule\\
\eqref{eqn:conjugates_cancel_phase_gadgets}%
}}{=}
&
\frac{2^{n_u + n_{uv} + 1}}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_02}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 1.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/02.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{%
\footnotesize\SpiderRule%
}{=}
&
\frac{2^{n_u + n_{uv} + n_v + 1}}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_03}{
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/03.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{%
\substack{%
\footnotesize\SpiderRule\\
\eqref{eqn:qaoa_max_cut_summand_ix_side_calculation}
}
}{=}
&
\frac{c_{\gamma}^{n_u + n_{uv}}}{2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_04}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/04.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{%
\substack{%
\eqref{eqn:conjugates_cancel_phase_gadgets}\\
\eqref{eqn:phase_gadget_as_lincomb}
}
}{=}
&
\frac{c_{\gamma}^{n_u + n_{uv}}}{2^2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_05}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/05.tex}
\end{tikzpicture}
}
}}
\\[4ex]
=
&
\mathrm{i} s_\gamma c_{\gamma}^{n_u + n_{uv}} \, ,
\end{align}
where we have used the size of the exclusive neighborhoods
$n_u:=|N_{u} \setminus \{N_v \cup u\}|$,
$n_v:=|N_{v} \setminus \{N_u \cup v\}|$,
and the joined neighborhood $n_{uv}:=|N_u \cap N_v|$,
the relation
$|\mathcal{N}^1_{uv}| = n_u + n_{uv} + n_v$,
as well as
\begin{align}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_side_calculation_00}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/side_calculation_00.tex}
\end{tikzpicture}
}
}}
&
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\frac{\mathrm{e}^{\mathrm{i} \gamma}}{\sqrt{2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_side_calculation_01}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/side_calculation_01.tex}
\end{tikzpicture}
}
}}
\\[4ex]
&
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
\frac{1}{2}
\;
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_side_calculation_02}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/side_calculation_02.tex}
\end{tikzpicture}
}
}}
\\[4ex]
&
=
c_\gamma
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_ix_side_calculation_03}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_ix/side_calculation_03.tex}
\end{tikzpicture}
}
}}\label{eqn:qaoa_max_cut_summand_ix_side_calculation} \,.
\end{align}
Analogously the third summand (the \textit{X-I}-term) can be obtained as
\begin{align}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xi_00}{%
\begin{tikzpicture}[every node/.style={font=\small}, node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_xi/00.tex}
\end{tikzpicture}
}
}}
=
\mathrm{i} s_\gamma c_{\gamma}^{n_v + n_{uv}} \, .
\end{align}
The fourth summand (the \textit{X-X}-term) reads
\begin{align}
&
\frac{1}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_00}{
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/00.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\eqref{eqn:ps_layer_definition}}{=}
&
\frac{2^{n_u + 2 n_{uv} + n_v + 1}}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_01}{%
\begin{tikzpicture}[node distance=4ex and 1.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/01.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\substack{%
\footnotesize\PiRule\\
\eqref{eqn:conjugates_cancel_phase_gadgets}%
}}{=}
&
\frac{2^{n_u + 2 n_{uv} + n_v}}{2^{|\mathcal{N}^1_{uv}| + 2}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_02}{%
\begin{tikzpicture}[node distance=4ex and 1.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/02.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\substack{%
\footnotesize\SpiderRule\\
\eqref{eqn:phase_gadget_as_lincomb}%
}}{=}
&
\frac{2^{n_{uv}}}{2^2}
\left(\frac{\mathrm{e}^{\mathrm{i} \gamma}}{\sqrt{2}}\right)^{n_u + 2 n_{uv} + n_v}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_03}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/03.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\substack{%
\eqref{eqn:qaoa_max_cut_summand_ix_side_calculation}\\
\eqref{eqn:qaoa_max_cut_summand_xx_side_calculation}%
}}{=}
&
\frac{c_\gamma^{n_u + n_v}}{2^2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_04}{%
\begin{tikzpicture}[node distance=2ex and 1.0em]
\input{./images/tikz/qaoa_max_cut/summand_xx/04.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\eqref{eqn:zxsumrules_product_of_sums}}{=}
&
\frac{-c_\gamma^{n_u + n_v}}{2^2}
\biggl\{
\biggr.
\binom{n_{uv}}{1}
s_\gamma^2
c_\gamma^{n_{uv}-2}
\;\;\;
\begin{ZX}[ampersand replacement=\&]
\zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\& \zxZ{}
\\
\zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\& \zxZ{}
\end{ZX}
\\[3ex]
&
\qquad
\qquad
+
\binom{n_{uv}}{3}
s_\gamma^6
c_\gamma^{n_{uv}-6}
\;\;\;
\begin{ZX}[ampersand replacement=\&]
\zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\& \zxZ{}
\\
\zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\&[1.2em] \zxN{} \rar
\&[1.2em] \zxN{} \rar
\& \zxZ{}
\end{ZX}
\\[3ex]
&
\qquad
\qquad
+
\binom{n_{uv}}{5}
s_\gamma^{10}
c_\gamma^{n_{uv}-10}
\begin{ZX}[ampersand replacement=\&]
\zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{}
\\
\zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{\pi} \rar
\& \zxZ{}
\end{ZX}
\\[3ex]
&
\qquad
\qquad
+
\;
\dots
\qquad \qquad \qquad \qquad
\qquad \qquad \qquad \qquad
\qquad \qquad
\biggl.
\biggr\}
\\[3ex]
& =
-c_\gamma^{n_u + n_v}
\sum_{i=1, 3, \dots} \binom{n_{uv}}{i} (s_\gamma^2)^i (c_\gamma^2)^{n_{uv} - i } \, ,
\end{align}
where we have used
\begin{align}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_side_calculation_00}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/side_calculation_00.tex}
\end{tikzpicture}
}
}}
\stackrel{\footnotesize\SpiderRule}{=}
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_side_calculation_01}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/side_calculation_01.tex}
\end{tikzpicture}
}
}}
\\[4ex]
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
&
\left(\frac{\mathrm{e}^{-\mathrm{i} \gamma}}{\sqrt{2}}\right)^2
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_side_calculation_02}{%
\begin{tikzpicture}[node distance=4ex and 1.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/side_calculation_02.tex}
\end{tikzpicture}
}
}}
\\[4ex]
=
&
2\, \left(\frac{\mathrm{e}^{-\mathrm{i} \gamma}}{\sqrt{2}}\right)^2
\vcenter{\hbox{
\externalize{qaoa_max_cut_summand_xx_side_calculation_03}{%
\begin{tikzpicture}[node distance=2ex and 1.5em]
\input{./images/tikz/qaoa_max_cut/summand_xx/side_calculation_03.tex}
\end{tikzpicture}
}
}}\label{eqn:qaoa_max_cut_summand_xx_side_calculation} \, .
\end{align}
Hence, the total $Z$-$Z$-expectation value reads
\begin{align}
\langle Z_u Z_{v} \rangle_{\text{QAOA}_1}
& =
c_\beta s_\beta
s_\gamma
\left(%
c_{\gamma}^{n_u + n_{uv}}
+ c_{\gamma}^{n_v + n_{uv}}
\right)
+
c_\gamma^{n_u + n_v}
s_\beta^2
\sum_{i=1, 3, \dots} \binom{n_{uv}}{i} (s_\gamma^2)^i (c_\gamma^2)^{n_{uv} - i }.
\end{align}
This result is consistent with the corresponding QAOA$_1$ performance analysis of \cite{wang2018quantum}; applying the binomial theorem to write the sum above in closed form then leads directly to the result of~\cite[Thm. 1]{wang2018quantum}.
\subsection{Details on QAOA$_1$ for MaxCut on Simple Graph}%
\label{sec:app_qaoa_example_graph}
We calculate each of the four summands in~\eqref{eqn:qaoa_max_cut_example_terms}.
The first summand (the \textit{I-I}-term) reads
\begin{align}
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ii_00}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ii/00.tex}
\end{tikzpicture}
}
}}\\[1em]
\stackrel{\footnotesize\SpiderRule}{=}
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ii_01}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ii/01.tex}
\end{tikzpicture}
}
}}
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\; \frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ii_02}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ii/02.tex}
\end{tikzpicture}
}
}}
=
\;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ii_03}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ii/03.tex}
\end{tikzpicture}
}
}} \label{eqn:qaoa_max_cut_example_summand_ii} \, .
\end{align}
The second summand (the \textit{I-X}-term) reads
\begin{align}
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_00}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/00.tex}
\end{tikzpicture}
}
}}
\stackrel{\footnotesize\PiRule}{=} \;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_01}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/01.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\footnotesize\SpiderRule}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_02}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/02.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
&
\frac{1}{2^2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_03}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/03.tex}
\end{tikzpicture}
}
}}
=
\frac{1}{2^2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_04}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/04.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
&
\frac{1}{2^3}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_05}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/05.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_06}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/06.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:zxsumrules_pull}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_07}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/07.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:zxsumrules_product_of_sums}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_08}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/08.tex}
\end{tikzpicture}
}
}}
= \;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_ix_09}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_ix/09.tex}
\end{tikzpicture}
}
}}
\label{eqn:qaoa_max_cut_example_summand_ix}
\end{align} \,.
The third summand (the \textit{X-I}-term) reads
\begin{align}
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_00}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/00.tex}
\end{tikzpicture}
}
}}
\stackrel{\footnotesize\PiRule}{=}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_01}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/01.tex}
\end{tikzpicture}
}
}}\\[1em]
\stackrel{\footnotesize\SpiderRule}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_02}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/02.tex}
\end{tikzpicture}
}
}}
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_03}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/03.tex}
\end{tikzpicture}
}
}}\\[1em]
= &
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_04}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/04.tex}
\end{tikzpicture}
}
}}
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\frac{1}{\sqrt{2^5}}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_05}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/05.tex}
\end{tikzpicture}
}
}}\\[1em]
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_06}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/06.tex}
\end{tikzpicture}
}
}}\\[1em]
\stackrel{\eqref{eqn:zxsumrules_pull}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_07}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/07.tex}
\end{tikzpicture}
}
}}\\[1em]
\stackrel{\eqref{eqn:zxsumrules_product_of_sums}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_08}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/08.tex}
\end{tikzpicture}
}
}}
= \;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xi_09}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xi/09.tex}
\end{tikzpicture}
}
}}
\label{eqn:qaoa_max_cut_example_summand_xi}
\end{align} \, .
The fourth summand (the \textit{X-X}-term) reads
\begin{align}
& \vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_00}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/00.tex}
\end{tikzpicture}
}
}}
\stackrel{\footnotesize\PiRule}{=}
\; \vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_01}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/01.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\footnotesize\SpiderRule}{=}
&
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_02}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/02.tex}
\end{tikzpicture}
}
}}
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\frac{1}{2}
\;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_03}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/03.tex}
\end{tikzpicture}
}
}}
\\[3ex]
=
&
\frac{1}{2}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_04}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/04.tex}
\end{tikzpicture}
}
}}
\stackrel{\eqref{eqn:conjugates_cancel_phase_gadgets}}{=}
\frac{1}{\sqrt{2^5}}
\;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_05}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/05.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:phase_gadget_as_lincomb}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_06}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/06.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:zxsumrules_pull}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_07}{%
\begin{tikzpicture}[node distance=8ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/07.tex}
\end{tikzpicture}
}
}}
\\[3ex]
\stackrel{\eqref{eqn:zxsumrules_product_of_sums}}{=}
&
\frac{1}{2^4}
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_08}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/08.tex}
\end{tikzpicture}
}
}}
= \;
\vcenter{\hbox{
\externalize{qaoa_max_cut_example_summand_xx_09}{%
\begin{tikzpicture}[node distance=4ex and 2.5em]
\input{./images/tikz/qaoa_max_cut_example/summand_xx/09.tex}
\end{tikzpicture}
}
}}\label{eqn:qaoa_max_cut_example_summand_xx} \, .
\end{align}
\subsection{Proofs of useful ZX-diagram Identities}
\subsubsection{Phase-gadget identity}%
\label{sec:app_proof_conjugates_cancel_phase_gadgets}
\begin{proof}[Proof of~\eqref{eqn:conjugates_cancel_phase_gadgets}]
First, we can us the spider fusion rule to write
\begin{align}
\vcenter{\hbox{
\externalizezx{phase_gadgets_00}{%
\input{./images/tikz/phase_gadgets/00.tex}
}
}}
&\stackrel{\footnotesize\SpiderRule}{=}
\vcenter{\hbox{
\externalizezx{phase_gadgets_01}{%
\input{./images/tikz/phase_gadgets/01.tex}
}
}} \, .
\end{align}
Then, we just consider the inner part
\begin{align}
\vcenter{\hbox{
\externalizezx{phase_gadgets_02}{%
\input{./images/tikz/phase_gadgets/02.tex}
}
}}
&\stackrel{\footnotesize\PiRule}{=}
\vcenter{\hbox{
\externalizezx{phase_gadgets_03}{%
\input{./images/tikz/phase_gadgets/03.tex}
}
}}
\\[4ex]
&\stackrel{\footnotesize\SpiderRule}{=}
\vcenter{\hbox{
\externalizezx{phase_gadgets_04}{%
\input{./images/tikz/phase_gadgets/04.tex}
}
}}
\\[4ex]
&=
\vcenter{\hbox{
\externalizezx{phase_gadgets_05}{%
\input{./images/tikz/phase_gadgets/05.tex}
}
}}
\\[4ex]
&\stackrel{\footnotesize\SpiderRule}{=}
\vcenter{\hbox{
\externalizezx{phase_gadgets_06}{%
\input{./images/tikz/phase_gadgets/06.tex}
}
}}
\\[4ex]
&\stackrel{{\footnotesize\BialgRule}}{=}
\frac{1}{\sqrt{2}}
\vcenter{\hbox{
\externalizezx{phase_gadgets_07}{%
\input{./images/tikz/phase_gadgets/07.tex}
}
}}
\\[4ex]
&\stackrel{\footnotesize\PiRule \footnotesize\SpiderRule}{=}
\frac{1}{\sqrt{2}}
\vcenter{\hbox{
\externalizezx{phase_gadgets_08}{%
\input{./images/tikz/phase_gadgets/08.tex}
}
}} \, .
\end{align}
If $t+l+b+r$ even, we have
\begin{align}
\frac{1}{\sqrt{2}}
\vcenter{\hbox{
\externalizezx{phase_gadgets_09even}{%
\input{./images/tikz/phase_gadgets/09even.tex}
}
}}
&\stackrel{{\footnotesize\IdRule}, \footnotesize\SpiderRule}{=}
\frac{1}{\sqrt{2}}
\vcenter{\hbox{
\externalizezx{phase_gadgets_10even}{%
\input{./images/tikz/phase_gadgets/10even.tex}
}
}}
\\[4ex]
&\stackrel{\footnotesize\CopyRule}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalizezx{phase_gadgets_11even}{%
\input{./images/tikz/phase_gadgets/11even.tex}
}
}}
\\[4ex]
&\stackrel{{\footnotesize\IdRule},\footnotesize\SpiderRule}{=}
\frac{1}{2}
\vcenter{\hbox{
\externalizezx{phase_gadgets_12even}{%
\input{./images/tikz/phase_gadgets/12even.tex}
}
}} \,,
\end{align}
else, if $t+l+b+r$ odd, we have
\begin{align}
\frac{1}{\sqrt{2}}
\vcenter{\hbox{
\externalizezx{phase_gadgets_09odd}{%
\input{./images/tikz/phase_gadgets/09odd.tex}
}
}}
&\stackrel{\footnotesize\SpiderRule,\footnotesize\PiRule}{=}
\frac{1}{\sqrt{2}} \mathrm{e}^{\mathrm{i} \gamma}
\vcenter{\hbox{
\externalizezx{phase_gadgets_10odd}{%
\input{./images/tikz/phase_gadgets/10odd.tex}
}
}} \, ,
\end{align}
which proves \eqref{eqn:conjugates_cancel_phase_gadgets}.
\end{proof}
\subsubsection{Hardware Efficient Ansatz}%
\label{sec:app_proof_zxidentity_hweffansatz}
\begin{proof}[Proof of \eqref{eqn:zxidentity_hweffansatz}]
\begin{align}
&
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_00}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_00.tex}
}
}}
\stackrel{\footnotesize\PiRule, \footnotesize\SpiderRule}{=}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_01}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_01.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\CopyRule, \footnotesize\PiRule, \footnotesize\SpiderRule}{=}
\underbrace{(-1)^{m_2 r_1 + (m_2 \oplus m_3) r_2 + m_3 r_3}}_{=:f_{m_2 m_3}^{r_1 r_2 r_3}}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_02}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_02.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\SpiderRule}{=}
f_{m_2 m_3}^{r_1 r_2 r_3}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_03}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_03.tex}
}
}}
\stackrel{\footnotesize\SpiderRule}{=}
f_{m_2 m_3}^{r_1 r_2 r_3}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_04}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_04.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\PiRule, \footnotesize\SpiderRule}{=}
f_{m_2 m_3}^{r_1 r_2 r_3}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_05}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_05.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\SpiderRule}{=}
f_{m_2 m_3}^{r_1 r_2 r_3}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_06}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_06.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\HopfRule, \footnotesize\SpiderRule}{=}
\frac{f_{m_2 m_3}^{r_1 r_2 r_3}}{2}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_07}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_07.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\HopfRule, \footnotesize\SpiderRule}{=}
\frac{f_{m_2 m_3}^{r_1 r_2 r_3}}{2^2}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_08}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_08.tex}
}
}}
\\[4ex]
&
\stackrel{\footnotesize\CopyRule, \footnotesize\SpiderRule}{=}
\frac{f_{m_2 m_3}^{r_1 r_2 r_3}}{2^3}
\vcenter{\hbox{
\externalizezx{hw_eff_ansatz_side_calculation_09}{
\input{./images/tikz/hw_eff_ansatz_example/side_calculation_09.tex}
}
}}
\end{align}
\end{proof}
|
\section{Introduction}
\label{sec:introduction}
Massive stars are the significant sources of ultraviolet (UV) radiation in galaxies with energies high enough ($>$\,13.6\,eV) to ionize the neutral gas in the interstellar medium (ISM). A part of this high-energy radiation is absorbed by the neutral gas and heats the surrounding medium creating the ionized H\,{\sc ii} regions. A large fraction of this ionizing radiation escapes into the diffuse medium, penetrating the molecular gas, if some part of the H\,{\sc ii} region is optically thin. This creates an ionization zone (H$^+$), ionization front (H$^0$), and a photodissociation region (PDR). The ionization front is at the outer boundary of an H\,{\sc ii} region that lies inside the PDR. There have been several studies of Galactic and extragalactic H\,{\sc ii} regions and PDRs, e.g., Orion Nebula \citep{pogge92, garcia07}, 30\,Doradus \citep{pellegrini10_30dor}, NGC\,364 \citep{peimbert00, Relano_2002}, dense H\,{\sc ii} regions in IC\,10 \citep{polles19} and NGC\,595 \citep{relano10}. The impact of ionizing radiation on the surrounding medium and the physical properties of H\,{\sc ii} regions are normally obtained by strong emission lines in the optical spectrum, which is mainly populated by hydrogen recombination lines and forbidden lines of other common elements. These gas emission lines are sensitive to physical conditions such as density and temperature; hence their relative intensities can probe the physical mechanism involved in the ionization processes.
The Large Magellanic Cloud (LMC) is an ideal laboratory to study the properties of H\,{\sc ii} regions and the massive star feedback in a low-metallicity galaxy due to its sub-solar metallicity ( Z\,=\,0.5Z$_\odot$, \citealt{westerlund97}); face-on viewing angle \citep{van01}; reduced extinction along the line of sight; a distance of 50\,kpc \citep{pietrzynki19} allowing the spatially resolved observations of the ISM structures on sub-parsec scales. \citet{Naslim15} have reported on ten PDRs in the LMC using H$_2$ pure rotational transition emission obtained with {\it Spitzer}. These regions include intense H\,{\sc ii} regions, diffuse ISM clouds, and dense molecular clouds. We study the individual clouds in detail using various observations to investigate the high-mass star formation (see \citealt{naslim18, nayana20}) and its impact on the ISM. To explore the impact of ionizing radiation from massive stars on the surrounding medium, we revisit two H\,{\sc ii} regions in a well-studied star-forming complex of the LMC, N44. We examine the physical conditions and ionization structure of N44D and N44C by comparing the observations with the predictions of the photoionization model, CLOUDY \citep{ferland2017}.
The N44 superbubble is one of the brightest star-forming regions in the LMC, which can be clearly traced by its compact H\,{\sc ii} regions along the main shell rim in an H$\alpha$ map (Fig.\,\ref{N44_OB}). The region is powered by nearly 35$-$38 hot stars \citep{Oye_Massey1995, McLeod19} of stellar associations LH47, LH48 and LH49 \citep{lucke70}. \citet{McLeod19} report the analysis of radiative and mechanical feedback from massive stars in the H\,{\sc ii} regions of N44 (N44\,A, N44\,B, N44\,C and N44\,D) using the optical integral field data from Multi-Unit Spectroscopic Explorer (MUSE). They used He\,{\sc ii}\,5412\AA\, line to identify the feedback driving massive stars, and estimated the spectral types and luminosity classes of these stars for determining the stellar radiative output. Using the nebular emission line maps of H$\alpha$, H$\beta$, [S\,{\sc ii}] 6717, 6732\AA, [N\,{\sc ii}] 6584\AA, and [O\,{\sc iii}] 5007\AA, they derived the electron density from [S\,{\sc ii}] 6717/6732 ratio assuming a nebular electron temperature of 10\,000\,K. They also derived the degree of ionization using [O\,{\sc ii}]/[O\,{\sc iii}] ratio, kinematics, and the oxygen abundances. In addition, they explored the role of different stellar feedback mechanisms by estimating various pressures and found that the H\,{\sc ii} region expansion is mainly driven by stellar radiation pressure and ionized gas.
In this paper, we further explore the rich nebular emission lines in MUSE data set \citep{McLeod19} of N44 for a detailed understanding of their spatial distributions, ionization structures and physical conditions. We compare these results with the photoionization models to interpret the observations of this star-forming complex. The MUSE archival data of N44 provides many iconic emission lines, such as H$\alpha$, H$\beta$, [S\,{\sc iii}] 9069\AA, [S\,{\sc ii}] 6717, 6732\AA, [O\,{\sc iii}] 4959, 5007\AA, [O\,{\sc ii}] 7318, 7329\AA, [O\,{\sc i}] 6300\AA, and [N\,{\sc ii}] 6584\AA, that have only been partially utilized in \citet{McLeod19}. We exploit various emission line ratios to study the behavior of different ionization zones and compare the observed line ratios with the best-fit photoionization models to further understand the physical process.
The N44 main shell is surrounded by several H$\alpha$ bright regions (Fig.\,\ref{N44_OB}). The compact H\,{\sc ii} region on the southwest rim of the shell, N44\,D, is the most luminous one in N44 \citep{McLeod19}. We choose two brightest H\,{\sc ii} regions N44\,D and N44\,C for our study, which show higher degree of ionization than N44\,A and N44\,B, and nearly spherical ionization structures with different feedback characteristics based on studies by \citet{McLeod19}. N44\,D encloses three hot stars of spectral types O5\,V, O9.5\,V and O5.5\,V \citep{McLeod19}. The second brightest H\,{\sc ii} region, N44\,C, is adjacent to N44D on the western rim of the shell, and the region harbors three hot stars O5\,III, O9.5\,V and O5\,III \citep{McLeod19}. One reason for selecting N44\,D is that MUSE observations show an edge on ionization front, allowing a detailed study of ionization as a function of depth into the cloud. At first glance, N44\,C does not show such a well-defined ionization front at the H\,{\sc ii} region boundary, while it shows a higher photon escape fraction relative to N44\,D that reported by \citet{McLeod19}. These two nebulae show different ionization structures, hence physical process of two different types of H\,{\sc ii} regions can be compared. Moreover, the hot star content of N44 H\,{\sc ii} regions are extensively studied and high-spatial-resolution observations are readily available. This study allows us to directly apply the observed stellar parameters, gas densities, and emission line intensities to constrain the photoionization model without arbitrary assumption, and test their influence on geometry of the H\,{\sc ii} regions.
\begin{figure*}
\centering
\includegraphics[width= 150 mm, height= 120 mm]{N44_latest2.eps}
\caption{The H$\alpha$ map of whole N44 superbubble obtained from Magellanic Cloud Emission Line Survey (MCELS; \citealt{smith98}). Spectral type of hot stars \citep{McLeod19} are labeled on the H$\alpha$ emission map and the magenta coloured boxes indicate the N44\,D and N44\,C H\,{\sc ii} regions. }
\label{N44_OB}
\end{figure*}
\section{Observations}
We used the MUSE archival data of N44\,C and N44\,D (program ID: 096.C-0137(A), PI: A. F. McLeod). MUSE is a large field-of-view (FOV) integral field unit (IFU) panchromatic optical instrument on the European Southern Observatory's (ESO's) Very Large Telescope (VLT) in Paranal, Chile. This instrument provides high-spatial-resolution observations at a pixel scale of 0.2$^{\prime\prime}$ with a resolving power ranging from 1770 to 3590. The observations of N44\,C and N44\,D have been taken on 21 October 2015 and 25 February 2016, with the MUSE$\_$wfm-noao$\_$obs$\_$genericoffset observing template, in a wide-field observing mode covering a wavelength range 475$-$935\,nm. The reduced MUSE data is retrieved from the ESO science archive\footnote{http://archive.eso.org/cms.html}. The data were reduced using the MUSE-1.6.1 pipeline. The MUSE pipeline process automatically removes most of the instrumental signatures. The raw data were pre-processed, and bias subtraction, flat fielding, sky-subtraction, wavelength calibration and flux calibration were applied. These data were not taken with the Adaptive Optics System of MUSE and the seeing-limited angular resolutions 0.98$^{\prime\prime}$ and 1.30$^{\prime\prime}$ are achieved for N44D and N44C respectively. We note that, no point spread function (PSF) matching was applied for subsequent analysis and all the analysed regions are resolved regardless of the achieved seeing. Our analysis is based on integrated line flux maps, hence no PSF information is retained.
\section{Emission line maps of N44\,D and N44\,C}
\begin{figure*}
\hspace*{0mm}\includegraphics[width=180mm, height=45mm]{n44d_gnuplot2.eps}
\hspace*{0mm}\includegraphics[width=180mm, height=45mm]{n44c_gnuplot2.eps}
\caption{The MUSE H$\alpha$ maps of N44\,D (top) and N44\,C (bottom). The emission line fluxes are extracted from the polygon regions (white) of N44\,D1, N44\,D2 and N44\,C for the analysis of photoionized gas in this work.}
\label{fig:Halpha}
\end{figure*}
Fig.\,\ref{fig:Halpha} shows the extinction corrected H$\alpha$ line flux maps of N44\,D and N44\,C obtained with MUSE. The ionized gas traced by the H$\alpha$ emission shows two H\,{\sc ii} regions in N44\,D, those we label as N44\,D1 and N44\,D2, and one in N44\,C. Even though the regions are nearly spherical, their boundaries cannot be directly specified in a circular aperture. Hence for determining the boundaries of these H$\alpha$ bright regions based on their surface brightness, we use the Python package ASTRODENDRO \citep{Rosolowsky08}. This algorithm identifies and characterizes the hierarchical structures in the emission line map as a structure tree, where each entity is represented as an isosurface. The local maxima represent the top level of the dendrogram and are identified from the emission line map with the flux $>$ 3$\sigma$. The isosurfaces (two-dimensional contours) that surround the local maxima are leaves, branches, and trunks. The trunks represent parent structures that enclose the branches connecting two leaves. Further description and methods of using ASTRODENDRO can be found in \citet{naslim18}. \citet{bruna20} have recently used ASTRODENRO to identify H$\alpha$ bright regions in MUSE maps of NGC\,7793 by applying a similar method. We define the boundary of H\,{\sc ii} regions in H$\alpha$ map that are identified as trunks with ASTRODENDRO. The lower contour levels of these trunks are taken as the boundaries of H\,{\sc ii} regions within the chosen observation field. These regions appear as polygons in Fig.\,\ref{fig:Halpha}.
\begin{figure}
\centering
\hspace{0mm}\includegraphics[scale=0.35]{N44D_lines_grace.eps}
\hspace{0.5 cm}
\hspace{0mm}\includegraphics[scale=0.35]{N44C_lines_grace.eps}
\caption{The identified emission lines are labelled in the MUSE spectra extracted from a 1.0$^{\prime\prime}$ radius circular region close to the cross-cut indicated as red line in Fig.\,\ref{fig:spatialmap}. This is to show the rich emission lines available for analysis in the MUSE observation of N44\,D1 and N44\,C. The continua of the spectra are normalized to 1 and the peak of certain strong lines are cut out of the scale for the weaker lines to be visible in the plot properly.}
\label{N44_lines}
\end{figure}
In addition to H$\alpha$\,\,6562.8\AA\, emission, the MUSE spectra of N44\,D1 and N44\,C show emission due to H$\beta$\,4861\AA, [O\,{\sc iii}] 5007\AA, 4959\AA, [O\,{\sc ii}] 7318\AA, 7329\AA, [O\,{\sc i}] 6300\AA, [N\,{\sc ii}] 6584\AA, [S\,{\sc ii}] 6717\AA, 6732\AA, [S\,{\sc iii}] 9069\AA, [Ar\,{\sc iii}] 7135\AA, 7751\AA\, and many He\,{\sc i} and Paschen hydrogen lines ( Fig \ref{N44_lines}). Fig.\,\ref{fig:spatialmap} shows integrated line flux maps of H$\beta$, [S\,{\sc ii}]\,6717\AA, [O\,{\sc iii}]\,5007\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc i}]\,6300\AA, and [N\,{\sc ii}]\,6584\AA\, in N44\,D1, N44\,D2, and N44\,C. H$\beta$ and [O\,{\sc iii}]\,5007\AA\, emission show similar spatial distribution as the H$\alpha$ emission in both N44\,D and N44\,C, whilst [S\,{\sc ii}], [O\,{\sc ii}], [N\,{\sc ii}] and [O\,{\sc i}] emission in N44\,D1 show a shell structure. The morphology of N44\,D2 is irregular with pillars or filamentary structures.
The continuum subtraction is applied by creating continuum maps from the user defined line-free portions of the spectrum around each emission line. The line fluxes (erg s$^{-1}$ cm$^{-2}$) of N44\,D1, N44\,D2 and N44\,C regions are then extracted from the integrated line flux maps by applying aperture photometry within the specified regions as polygon structures obtained from ASTRODENDRO. Only the pixel values with S/N\,$>$\,10 are considered within all the polygons. These fluxes are then point source flux removed by subtracting the point source fluxes, which are also extracted within the user defined apertures. For uncertainties, we added in quadrature the error in aperture photometry, and an expected 20$\%$ calibration error in every line flux measurements. The error in photometry is the quadratically added uncertainty in measurements over all pixels within a region.
To check the data reduction and flux calibration quality of data cube retrieved from the MUSE archive, we compared the H$\alpha$ line luminosities (erg s$^{-1}$) of N44\,C and N44\,D1 from the MUSE archival data to the H$\alpha$ luminosities presented in \citet{McLeod19}. In Table \ref{comparison} we show the comparison of H$\alpha$ luminosities obtained from the data presented in this work and \citet{McLeod19}. We note that the line luminosities obtained from the two data sets agree within the estimated uncertainties, hence we are confident to proceed with the analysis of pipeline reduced MUSE archival data. The observed line luminosities (erg s$^{-1}$) are given in Table \ref{lines}. We choose two bright regions, N44\,D1 and N44\,C, for further analysis with photoionization models.
\begin{table*}
\centering
\caption{Comparison of H$\alpha$ luminosities obtained from the MUSE archival pipeline data (this work) with the data obtained from \citet{McLeod19}}.
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{lcccccccc}
\hline
\multicolumn{2}{c}{N44\,D1}&
\multicolumn{2}{c}{N44\,C}&\\
$^{a}L_{\textrm{obs}}$(ergs$^{-1}$)&$^{b}L_{\textrm{obs}}$(erg s$^{-1}$)&$^{a}L_{\textrm{obs}}$(erg s$^{-1}$)&$^{b}L_{\textrm{obs}}$(erg s$^{-1}$)\\
$(\times$10$^{37}$)&($\times$10$^{37}$)&($\times$10$^{37}$)& ($\times$10$^{37}$) \\
\hline
1.47$\pm$0.16&1.47&1.39$\pm$0.15&1.51 \\
\hline
\label{comparison}
\end{tabular}
\parbox{100mm}{
a: MUSE archival data used in this work.
b: \citet{McLeod19}.
}
\end{table*}
\section{Emission line ratios: [S\,{\scriptsize II}]/H$\alpha$, [N\,{\scriptsize II}]/H$\alpha$, [O\,{\scriptsize III}]/H$\alpha$ and [O\,{\scriptsize III}]/H$\beta$}
We present the extinction corrected [S\,{\sc ii}]\,6717/H$\alpha$, [N\,{\sc ii}]\,6584/H$\alpha$, [O\,{\sc iii}]\,5007/H$\alpha$, and [O\,{\sc iii}]\,5007/H$\beta$ ratio maps of N44\,D and N44\,C in Figs.\,\ref{lineratioN44D} and \ref{lineratioN44C}, respectively. These ratios allow us to study the ionization structure of the region. [S\,{\sc ii}]\,6717/H$\alpha$ and [N\,{\sc ii}]\,6584/H$\alpha$ are lower at the central regions closer to the ionizing stars implying a higher ionization zone, while at the periphery the values of these ratios are higher indicating low ionization zone. We find that both [S\,{\sc ii}]\,6717/H$\alpha$ and [N\,{\sc ii}]\,6584/H$\alpha$ maps of N44\,D1 and N44\,C show a shell morphology. The central part of the N44\,D1 has a lower [S\,{\sc ii}]\,6717/H$\alpha$ ratio ($\sim$\,0.02), and at the periphery, its value is $\sim$\,0.3. N44\,C shows a [S\,{\sc ii}]\,6717/H$\alpha$ value of $\sim$\,0.04 at the center and a value of $\sim$\,0.20 at the periphery. Similarly, the value of [N\,{\sc ii}]\,6584/H$\alpha$ ratio ranges from 0.03 to 0.20 in N44\,D1, and 0.04 to 0.15 in N44\,C. The model calculations by \citet{Allen08} have shown that [S\,{\sc ii}]/H$\alpha$ and [N\,{\sc ii}]/H$\alpha$ ratios greater than 0.39 and 0.79 would be a result of strong contributions from shocks. The [S\,{\sc ii}]/H$\alpha$ and [N\,{\sc ii}]/H$\alpha$ ratios of both N44\,D1 and N44\,C are well below the values 0.39 and 0.79, respectively, indicating a substantial contribution from photoionization. However, in the regions outside the boundary of these H\,{\sc ii} regions, we find the enhanced [S\,{\sc ii}]/H$\alpha$ and [N\,{\sc ii}]/H$\alpha$ ratios; hence the contribution from shocks cannot be totally ignored.
A similar effect is found in the [O\,{\sc iii}]\,5007/H$\beta$ and [O\,{\sc iii}]\,5007/H$\alpha$ ratios. The values are higher in the regions closer to the central ionizing stars showing a high degree of ionization, while the ratios are lower at the outer regions indicating a low degree of ionization.
We note that, the integrated [O\,{\sc iii}]\,5007/H$\alpha$ ratio of N44\,D1 (2.52) is much higher than N44\,D2 (1.67) and N44\,C (0.55), which indicates the hardness of radiation field in N44\,D1. The [O\,{\sc iii}]\,5007/H$\beta$ ratio is also higher in N44\,D1, representing the effect of high effective temperature of the ionizing star of spectral type O5\,V. The N44\,D1, N44\,D2, and N44\,C show an [O\,{\sc iii}]\,5007/H$\beta$ ratio of 7.89, 5.18 and 2.07 respectively. In N44\,C, the high values of [O\,{\sc iii}]\,5007/H$\alpha$ ($\sim$\,1.0 -- 0.3) and [O\,{\sc iii}]\,5007/H$\beta$ ($\sim$\,3.0 -- 1.0) are near the O5\,III star and the values decrease toward the edge of the bubble.
\begin{figure*}
\centering
\includegraphics[scale=0.35]{N44D1_SII_Ha_ratiomap_old.eps}
\includegraphics[scale=0.35]{N44D1_NII_Ha_ratiomap_old.eps}\\
\includegraphics[scale=0.35]{N44D1_OIII_Ha_ratiomap_old.eps}
\includegraphics[scale=0.35]{N44D1_OIII_Hb_ratiomap_old.eps}
\caption{
[S\,{\sc ii}]6717/H$\alpha$, [N\,{\sc ii}]6584/H$\alpha$, [O\,{\sc iii}]\,5007/H$\alpha$ and [O\,{\sc iii}]\,5007/H$\beta$ ratio maps of N44\,D1 and N44\,D2 regions. The white polygons represent the regions taken for analysis in this work and blue cross labels are the locations of hot stars.}
\label{lineratioN44D}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=0.35]{N44C_SII_Ha_ratiomap_old.eps}
\includegraphics[scale=0.35]{N44C_NII_Ha_ratiomap_old.eps}\\
\includegraphics[scale=0.35]{N44C_OIII_Ha_ratiomap_old.eps}
\includegraphics[scale=0.35]{N44C_OIII_Hb_ratiomap_old.eps}
\caption{
[S\,{\sc ii}]\,6717/H$\alpha$, [N\,{\sc ii}]\,6584/H$\alpha$, [O\,{\sc iii}]\,5007/H$\alpha$ and [O\,{\sc iii}]\,5007/H$\beta$ ratio maps of N44\,C region. The white polygons represent the regions taken for analysis in this work and blue cross labels are the locations of hot stars.}
\label{lineratioN44C}
\end{figure*}
\section{H$\alpha$ and H$\beta$ emission}
\subsection{Extinction correction}
The line luminosities are corrected for extinction using the intensity ratios (H$\alpha$/H$\beta$)$_{\textrm{obs}}$. Since, H$\alpha$/H$\beta$ ratio is relatively sensitive to temperature, it can be used as a reliable reddening indicator. This ratio is compared with the theoretically expected value of the Balmer decrement (H$\alpha$/H$\beta)_{\textrm{exp}}$ for Case B recombination \citep{Osterbrock2006}.
Any deviation from expected value of H$\alpha$/H$\beta$ ratio for a particular electron temperature can be associated with extinction. We estimate the nebular emission line color excess $E(B-V)$ from the H$\alpha$/H$\beta$ ratio using the equation from \citet{Dom_nguez_2013},
\begin{equation}
E(B-V) = \frac{2.5}{k(\lambda_{\textrm{H}\beta})-k(\lambda_{\textrm{H}\alpha})} log_{10}\Bigg[\frac{({\textrm{H}\alpha}/{\textrm{H}\beta})_{\textrm{obs}}}{({\textrm{H}\alpha}/{\textrm{H}\beta})_{\textrm{exp}}}\Bigg]
\end{equation}
The expected value of (H$\alpha$/H$\beta)_{\textrm{exp}}$ flux ratio is $\sim$\,2.86 \citep{Osterbrock2006}. This value is obtained by assuming Case B recombination at an electron temperature 10\,000\,K and density 100\,cm$^{-3}$. Then, following the extinction curve estimated by \citet{calzetti2000},
\begin{equation}
k(\lambda)=2.659(-1.857+1.040/\lambda)+ R_{\textrm{V}}
\end{equation}
for $\lambda$ = 0.63 $\mu m$ to 2.2 $\mu m$, and
\begin{equation}
k(\lambda)=2.659(-2.156+1.509/\lambda - 0.198/\lambda^2 +0.011/\lambda^3)+ R_{\textrm{V}}
\end{equation}
for $\lambda$ = 0.12 $\mu m$ to 0.63 $\mu m$.\\
Here, $k(\lambda_{\textrm{H}\alpha})$ and $k(\lambda_{\textrm{H}\beta})$ are the extinction curves at H$\alpha$ and H$\beta$ wavelengths respectively.
Assuming the ratio of total to selective extinction $R_\textrm{V} (=A_{\textrm{V}}/E(B-V))=3.1$, which is valid at optical wavelength for the LMC \citep{Gordon03}, we get $k(\lambda_{\textrm{H}\alpha})$ = 2.38 , and $k(\lambda_{\textrm{H}\beta})$ = 3.65 .
Using the color excess $E(B-V)$, the extinction in magnitude for H$\alpha$ line is obtained by
\begin{equation}
A_{\textrm{H}\alpha} = k(\lambda_{\textrm{H}\alpha}) E(B-V)
\end{equation}
and for H$\beta$ line,
\begin{equation}
A_{\textrm{H}\beta} = k(\lambda_{\textrm{H}\beta}) E(B-V)
\end{equation}
Then extinction corrected H$\alpha$ luminosity, $L$(H$\alpha)$ is,
\begin{equation}
L(\textrm{H}\alpha) = L(\textrm{H}\alpha)_{obs} 10^{0.4 A_{\textrm{H}\alpha}}
\end{equation}
and the extinction corrected H$\beta$ luminosity, $L$(H$\beta)$ is,
\begin{equation}
L(\textrm{H}\beta) = L(\textrm{H}\beta)_{obs} 10^{0.4 A_{\textrm{H}\beta}}
\end{equation}
Here $L$(H$\alpha)_{obs}$ and $L$(H$\beta)_{obs}$ are the observed luminosities of H$\alpha$ and H$\beta$ emission respectively.
The values of the color excess $E(B-V)$ are 0.08 for N44\,D1 and N44\,D2, and 0.20 for N44\,C. The extinction toward N44\,D1 is $A_{\textrm{V}} = 0.25$\,mag. Our value agrees with the calculations by \citet{Garnett_2000} and \citet{Lopez_2014} for N44. The N44\,C has a significantly higher extinction, $A_{\textrm{V}} = 0.62$\,mag.
We applied the same method of extinction correction to other line emission, and the extinction corrected luminosities (erg s$^{-1}$) are given in Table\,\ref{lines}.
\subsection{Lyman Continuum Photon Flux}
The O-type stars in H\,{\sc ii} regions are the prominent sources of Lyman continuum photons. These stars deposit a bulk of their high energy photons into the surrounding H\,{\sc ii} region within the Stromgren radii. If the gas is optically thick in the Lyman continuum, we expect all the ionizing photons emitted by the star to be absorbed. However, a significant fraction of these photons can escape on a larger scale outside of the H\,{\sc ii} region into the ISM. This fraction of photon leakage from H\,{\sc ii} regions needs to be measured to understand the overall energy budget, and to probe whether the dominant hot stars in the region are responsible for the photoionization in the surrounding medium. We calculate the number of L[OIII]/H$\beta$yman continuum photons ($Q$) absorbed in the region surrounding a hot star corresponding to H$\alpha$ luminosities by assuming the Case B recombination for electron temprature $T_{\textrm{e}}$ = 10\,000\,K, and density $n_{\textrm{e}}$ = 100\,cm$^{-3}$. The number of Lyman continuum photon related to H$\alpha$ luminosity is obtained by $Q(\textrm{H}\alpha)$ = 7.31$\times$10$^{11}L$(H$\alpha$)\,ph\,s$^{-1}$ \citep{kennicutt1998, Osterbrock2006}.
The number of ionizing photons derived from the H$\alpha$ luminosity ($Q$) for N44\,D1, N44\,D2, and N44\,C are tabulated in Table\,\ref{properties}.
To calculate the photon escape fraction, we also need to know the number of total Lyman continuum photons emitted by the ionizing stars. We adopt the model calculations ($Q_0$) for hot stars of appropriate spectral types from \citet{martins05}. The $Q_0$ values of the only O5\,V star in N44\,D1, a combination of three ionizing stars of spectral types O5\,III, O8\,V and O9.5\,V in N44\,C, and O5.5\,V star in N44\,D2 are given in Table\,\ref{properties}.
Using the $Q$ and $Q_0$ values we calculate the photon escape fraction,
\begin{equation}
f_{\textrm{esc}} = \frac{Q_0-Q}{Q_0}
\end{equation}
\citet{McLeod19} reported $f_{\textrm{esc}} \backsim$ 0.37 and 0.68 for N44\,D1 and N44\,C respectively, using this method. We verify their determination and find that $f_{\textrm{esc}}$ for N44\,D1, N44\,D2, and N44\,C are 0.36, 0.71, and 0.70 respectively. These values imply that about 36$\%$ of the ionizing photons escape from N44\,D1, 71$\%$ of ionizing photons escape from N44\,D2 and 70$\%$ from N44\,C. N44\,C shows a relatively larger amount of photon leakage than N44\,D1, and is more optically thin to the ionizing photon. This observation is consistent with the study of H\,{\sc ii} regions in the LMC by \citet{pellegrini12} that boundaries of optically thick regions are generally characterized by stratification in the ionization structures. The ionization structure of N44\,D1 shows a well-defined nebular boundary where [O\,{\sc i}] and [S\,{\sc ii}] emission peak at the outer boundary of the H$\alpha$ and [O\,{\sc iii}] emission. The [S\,{\sc ii}]/H$\alpha$ ratio map clearly shows the transition region in the ionization structure. N44\,C does not show such an ionization stratification, while a slightly extended shell structure is found in the [S\,{\sc ii}]/H$\alpha$ map.
Using the number of ionizing photons $Q$, we can also estimate the average electron density of emitting gas in an H\,{\sc ii} regions, $\langle n_{\textrm{e}}\rangle$. Assuming the spherical nebula where H is fully ionized, the recombination balance equation is,
\begin{equation}
Q=\frac{4\pi}{3} \alpha_{\textrm{B}} n_{\textrm{e}}^2 R^3_{\textsc{Hii}}
\end{equation}
Here $\alpha_{\textrm{B}}$ is the Case B recombination coefficient $\sim$ 2.59$\times$10$^{-13}$ cm$^{3}$ s$^{-1}$ for gas at $T$\,=\,10\,000\,K. $R_{\textrm{H{\sc ii}}}$ is the mean radius of t[OIII]/H$\beta$he H\,{\sc ii} region. Then the average electron density from H$\alpha$ emission is obtained by,
\begin{equation}
\langle n_{\textrm{e}}\rangle\, = 177 \sqrt{\frac{Q_{48}}{R^3_{\textsc{Hii}}}}
\end{equation}
Here $Q_{48}$(=Q/10$^{48}$) is the number of Lyman continuum photons derived from the H$\alpha$ luminosity and $R_{\textrm{H{\sc ii}}}$ is the radius in parsec.
The $\langle n_{\textrm{e}}\rangle$ derived from H$\alpha$ emission of N44\,D1, N44\,D2, and N44\,C are 31, 26, and 38 cm$^{-3}$ respectively.
\subsection{Electron density}
Electron density ($n_{\textrm{e}}$) and electron temperature ($T_{\textrm{e}}$) are two important physical parameters for characterizing an H\,{\sc ii} region. The $n_{\textrm{e}}$ can be determined from the observed line intensities of two different energy levels with nearly equal excitation energy of the same ion. Their line ratios are generally not sensitive to $T_{\textrm{e}}$. The forbidden lines ratio, [S\,{\sc ii}]\,6717/6732 is usually used to determine $n_{\textrm{e}}$, where [S\,{\sc ii}]\,6717\AA\, and 6732\AA\, emission are relatively strong in the ionized nebula. Their corresponding critical density is $\sim$\,10$^{3}$\,cm$^{-3}$, hence probing the low-density regimes. \citet{McLeod19} have reported $n_{\textrm{e}}$\,$\sim\,152\pm42$\,cm$^{-3}$ for N44\,C and $\sim\,143\pm42$\,cm$^{-3}$ for N44\,D1, applying the analytical solution given in \citet{McCall84} and assuming a $T_{\textrm{e}}$ of 10\,000\,K. \citet{Toribio17} have derived $n_{\textrm{e}}\,\sim\,200\pm150$\,cm$^{-3}$ for a 3.0$\times$9.4 arcsec$^2$ region closer to the ionizing star in N44\,D1 using [S\,{\sc ii}]\,6717/6732 ratio obtained from VLT-UVES spectrum. \citet{Lopez_2014} have reported a relatively low value of $n_{\textrm{e}}$\,$\sim\,60$\,cm$^{-3}$ for the entire N44 using the flux density of the free-free emission at 3.5 cm. \citet{Garnett_2000} derived $n_{\textrm{e}}$\,$ <\,160\,cm^{-3}$ for N44D1 using [S\,{\sc ii}]\,6717/6732 ratio obtained with 0.9 m telescope at Cerro Tololo Inter-American Observatory. \citet{McLeod19} emphasize that, densities derived from radio emission by \citet{Lopez_2014} are expected to be smaller than those derived from the ratio of collisionally excited lines \citep{peimbert2017}. These studies show a discrepancy in derived values of $n_{\textrm{e}}$ for N44\,D1 and N44\,C, hence we calculate the electron densities using [S\,{\sc ii}]\,6717/6732 ratios of N44\,D1 and N44\,C derived in our analysis of MUSE spectra. Electron temperature can be obtained by the forbidden line ratios [SIII]6312/9069 and [NII]5755/6384, however the MUSE observations of N44\,D1 and N44\,C show very weak [SIII]\,6312\AA\, and [NII]\,5755\AA\, emission, which cannot be extracted from the data cube within a 5$\sigma$ detection threshold in most of the pixels inside the defined polygons. We calculate the electron density as in \citet{McLeod2015} by applying the analytical solution in \citet{McCall84}.
\begin{equation}
n_e = \frac{1.49-R_{\textsc{sii}}}{12.8\times R_{\textsc{sii}}- 5.6713}\times10^4 \;\;\mbox{cm}^{-3}
\end{equation}
Here, R$_{\textsc{sii}}$ is the [S\,{\sc ii}]\,6717/6732 ratio, and electron temperature is assumed to be 10\,000\,K as in \citet{McLeod19}. The derived $n_{\textrm{e}}$ for N44D1, N44D2 and N44C using this method are 141$\pm$43, 121$\pm$37 and 92$\pm$35 \,cm$^{-3}$ respectively. We also estimate $n_e$ using the publicly available Python based package, PYNEB \citep{Luridiana2013} for a comparison. This algorithm includes FIVEL \citep{robertis1987} and NEBULAR \citep{shaw1995} packages for analyzing nebular emission lines. The package calculates the physical conditions ($T_{\textrm{e}}$ and $n_{\textrm{e}}$) for a given set of emission line intensities, and returns the diagnostic plots. We use the density sensitive [S\,{\sc ii}]\,6732/6717 line ratio to determine $n_{\textrm{e}}$ using the diags.getTemDen task in PYNEB for a given $T_{\textrm{e}}$ of 10\,000\,K. The estimated electron densities from PYNEB for N44D1, N44D2 and N44C are 132$\pm$50, 115$\pm$45 and 66$\pm$40 \,cm$^{-3}$ respectively (Table\,\ref{properties}). These density values are comparable with the density derived from the equation 11 within the estimated uncertainties.
\section{Structure of ionized gas}
\begin{figure*}
\centering
\includegraphics[scale=0.35]{OIII_OII_OI.eps}\\
\includegraphics[scale=0.35]{SII_NII_Hbeta.eps}\\
\includegraphics[scale=0.35]{N44C_OIII_OII_OI.eps}\\
\includegraphics[scale=0.35]{N44C_Hbeta_SII_NII_2.eps}
\caption{(a) Spatial distribution of N44\,D1 and N44\,D2 are shown in [O\,{\sc iii}]\,5007\AA\, (blue), [O\,{\sc ii}]\,7318\AA\, (red) and [O\,{\sc i}]\,6300\AA\, emission (green). (b) N44\,D1 and N44\,D2 are shown in H$\beta$ (blue), [S\,{\sc ii}]\,6717\AA\, (red) and [N\,{\sc ii}]\,6584\AA\, emission (green). (c) N44\,C is shown in [O\,{\sc iii}]\,5007\AA\, (blue), [O\,{\sc ii}]\,7318\AA\, (red) and [O\,{\sc i}]\,6300\AA\, emission (green). (d) N44\,C is shown in H$\beta$ (blue), [S\,{\sc ii}]\,6717\AA\, (red), and [N\,{\sc ii}]\,6584\AA\, (green). The ionization zone is traced by [O\,{\sc iii}] and H$\beta$, the partially ionized zone is traced by [O\,{\sc ii}] and [N\,{\sc ii}] and the ionization front is traced by [O\,{\sc i}]\,6300\AA\, and [S\,{\sc ii}]\,6717\AA\,. The observed spatial profiles in Fig. \ref {fig:spatialcut1} and \ref{lineprof_mod_N44D} are taken along the cross-cuts indicated as red lines in these maps.}
\label{fig:spatialmap}
\end{figure*}
To investigate the structure of ionization zones in N44\,D and N44\,C we compare the spatial distribution of H$\beta$, [O\,{\sc iii}], [O\,{\sc ii}], [O\,{\sc i}], [N\,{\sc ii}] and [S\,{\sc ii}] emission line maps (Fig.\,\ref{fig:spatialmap}). The spatial distribution of [O\,{\sc iii}], [O\,{\sc ii}], and [O\,{\sc i}] of N44\,D1 in Fig.\,\ref{fig:spatialmap}(a) shows a clear stratification from ionization zones where [O\,{\sc iii}]\,5007\AA\, emission peaks around the O5\,V star, to the [O\,{\sc i}]\,6300\AA\, emission at the ionization front. \citet{odell1992} have reported that [O\,{\sc i}]\,6300\AA\, emission arises at the ionization front behind the PDR, between the ionization zone and molecular cloud. The majority of the [O\,{\sc i}]\,6300\AA\, emission in H\,{\sc ii} regions is due to the collisional excitation by thermal electrons and atomic hydrogen via the charge exchange, hence the intensity of [O\,{\sc i}]\,6300\AA\, emission can be a measure of neutral hydrogen content. [O\,{\sc i}]\,6300\AA\, in PDRs can also be collisionally excited by electrons ejected by dust grains and Polycyclic Aromatic Hydrocarbon (PAHs) that absorb far-UV radiation emitted by the massive stars. In Fig.\,\ref{fig:spatialmap}(b), we compare the spatial distribution of H$\beta$ (blue) and [N\,{\sc ii}]\,6584\AA (red) with [S\,{\sc ii}]\,6717\AA (green). [S\,{\sc ii}] emission is expected to peak at ionization front. [N\,{\sc ii}] emission is found to be co-spatial with [O\,{\sc ii}] emission that peaks at partial-ionization zone, and concentrated in the outer boundary of the H$\beta$ and [O\,{\sc iii}] emission. The structure of N44\,D1 is nearly spherical with only one important source of ionizing photons.
In Fig.\,\ref{fig:spatialcut1} we show the spatial profiles of various nebular emission lines by taking a cross-cut along the north-western edge of N44\,D1. [O\,{\sc ii}] and [N\,{\sc ii}] emission are in a thin layer at the partial ionization zone outside [O\,{\sc iii}], but slightly interior to the [O\,{\sc i}] emission. The [O\,{\sc i}] emission is concentrated in a thin zone $\sim$\,4 -- 7\,pc located in the outer boundary of the H\,{\sc ii} region. H$\alpha$, H$\beta$, [O\,{\sc iii}] and He\,{\sc i} are found to be co-spatial in the ionization zone. In the outer layer of the ionization front in this H\,{\sc ii} region, we expect a well-defined PDR with a layer of C\,{\sc ii}, then H$_2$ emission, and a molecular cloud. High-spatial resolution spectroscopic observations in infrared and submillimeter wavelengths are required for further studies of PDR properties. A similar ionization structure is reported in the Orion Nebula H\,{\sc ii} region. The [O\,{\sc i}]\,6300\AA\, and [S\,{\sc ii}]\,6717\AA\, emission appear to peak along a bright bar at the ionization front in the outer boundary, that forms a thin transition layer of thickness $\sim$\,10$^{15}$--10$^{16}$ cm between the ionization zone and PDR.The emission from higher ionization species [O\,{\sc iii}] arises away from the ionization front close to the ionizing star $\theta_1$Ori C \citep{Odell2017, Odell01, Hester1991, odell1992}.
In N44\,C, the [O\,{\sc iii}]\,5007\AA\, and H$\beta$ emission appear to peak in the interior of the bubble near the ionizing star O5\,III (Figs.\,\ref{fig:spatialmap} c, d). We show the spatial profiles of [O\,{\sc iii}], [O\,{\sc ii}], [S\,{\sc ii}], [N\,{\sc ii}], and H$\alpha$ emission in N44\,C taking a cross-cut from the position centroid of three ionizing stars to the north-east of N44\,C (Fig.\,\ref{fig:spatialcut1}). The spatial distribution of [O\,{\sc iii}]\,5007\AA\, and H$\beta$, are similar and co-spatial in the fully ionized zone. [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [N\,{\sc ii}]\,6584\AA\, and [S\,{\sc ii}]\,6717\AA\, emission do not form a well-defined outer boundary, but appear as patches of emission bars within [O\,{\sc iii}]\,5007\AA\, and H$\beta$ emission region.
\section{Comparison with photoionization model}
The MUSE observations of N44 in the LMC show that the two bright H\,{\sc ii} regions N44\,D1 and N44\,C have different ionization geometries and physical characteristics. N44\,D1 has H$\alpha$ and H$\beta$ surface brightness values 0.03\,dex and 0.1\,dex higher than N44\,C.
[O\,{\sc iii}]/H$\beta$ and [O {\sc iii}]/H$\alpha$ ratios in N44\,D1 are considerably larger than in N44\,C, indicating a higher degree of ionization. N44\,D1 shows a higher [O\,{\sc i}]\,6300/H$\beta$ ratio with a well-defined [O\,{\sc i}]\,6300\AA\,emission at the H\,{\sc ii} region outer boundary, indicating an ionization front.
In the ideal case, the [O\,{\sc i}]\,6300\AA\, emission dominates at the outer boundary of the H\,{\sc ii} region where the neutral hydrogen density dominates; hence we expect most of the Lyman continuum photons are absorbed by the nebula is ionization-bounded. This is an optically thick case, where we find a shell of [O\,{\sc i}]\,6300\AA\, emission at the ionization front, indicating the border of the H\,{\sc ii} region. In the density-bounded H\,{\sc ii} regions, there is no well-defined ionization front surrounding highly ionized gas, and most of the Lyman continuum photons leak from the cloud contributing to the diffuse ionized gas outside of the cloud. In such cases, there is a weak [O\,{\sc i}]\,6300\AA\, emission condensation within the H\,{\sc ii} region or no shell structure at the boundary. In a Blister H\,{\sc ii} region, there is a partial ionization front at the boundary, which does not cover the nebula completely; hence the photons escape in certain directions. Our study of the [O\,{\sc i}]\,6300\AA\, emission in N44\,D1 and N44\,C reveals these different observational properties of H\,{\sc ii} regions. Moreover, the measurement of the photon leakage using the H$\alpha$ emission indicates that N44\,D1 and N44\,C have photon escape fractions 36$\%$ and 70$\%$, respectively. These calculations are in agreement with the study of the photon leakage by \citet{McLeod19}. The remaining ionizing photons are trapped within the H\,{\sc ii} region itself and affect the overall ionization balance. \citet{pellegrini11,pellegrini12} reported that the ionization-bounded H\,{\sc ii} regions are constrained to an escape fraction $<$\,0.6 and those with density-bounded are $>$\,0.6.
To further interpret these observations, we compute various photo-ionization models for comparing the emission line ratios and the geometry of ionization structure.
We model the ionization structure of N44\,D1 and N44\,C using the photoionization code CLOUDY \citep{ferland2017}. We derive the emissivities of prominent ionic species across the H\,{\sc ii} region, from the illuminated face of the H$^+$ region through the partially ionized zone to the neutral ionization front, where ionizing radiation has been attenuated and becoming neutral to the molecular zone. We develop various photoionization models for N44\,D1 and N44\,C, and test which model can better match the observed ionization geometry of the cloud and emission line ratios. To compare the ionization geometry, we use the spatial profiles of the line emissions along the H\,{\sc ii} regions (Figs.\,\ref{fig:spatialcut1} and \ref{lineprof_mod_N44D}). The emission line ratios we use for tests are given in Tables \ref{linesratio_N44D1} and \ref{linesratio_N44C}. The [S\,{\sc ii}]\,6717/6732 line ratio is sensitive to electron density, and [O\,{\sc ii}]\,(7318+7329)/[O\,{\sc iii}]\,5007 to the ionization parameter. [O\,{\sc iii}]\,5007/H$\alpha$, [O\,{\sc iii}]\,5007/H$\beta$, [O\,{\sc ii}]/H$\beta$, [N\,{\sc ii}]/H$\beta$, and [O\,{\sc i}]/H$\beta$ give the behavior of different ionization zones, and are also sensitive to the metallicity. These line ratios are very sensitive to the adopted input parameters. A built-in optimization program based on the PHYMIR algorithm \citep{vanhoof97, vanhoof13, Ferland13} is used to obtain the best fit model, which applies a $\chi^2$ minimization to determine the goodness-of-fit by varying the input parameters. We vary the total hydrogen density ($hden$), $\Phi$(H), $brems$ and filling factor to find the best agreement between the model with the observed line ratios and geometry. Sometimes, for a given set of constraints some observables are optimized very well compared to the other sets and the best-fit model is obtained by the overall $\chi^2$. Therefore, we manually fine tuned certain parameters until the best matching line ratios with the observations are obtained. Finally, the model line ratios are compared with the observed line ratios and the best-fit model was determined by calculating the $\chi ^2$ as \citep{mondal2017, pavana_2019},
\begin{equation}
\chi^2 = \sum_{i=1} ^{n}(M_i - O_i)^2/\sigma_i ^2
\end{equation}
Here, the number of observed lines is n, M$_i$ is the model line ratio, O$_i$ is the observed line ratio and the $\sigma_i$ is the error in the observed flux ratio. The best optimized model line ratios, observed line ratios and the $\chi^2$ values are given in Tables \ref{linesratio_N44D1} and \ref{linesratio_N44C}.
The basic input parameters to CLOUDY require the geometry of the cloud, intensity of incident ionizing photon flux, elemental abundances and gas density.
\subsection{Geometry}
Observations show a clear ionization stratification in N44\,D1. The ionized gas is traced by [O\,{\sc iii}]\,5007\AA\, and H$\beta$ emission and the ionization front is traced by [O\,{\sc i}]\,6300\AA\, and [S\,{\sc ii}]\,6717\AA\, emission. The line ratio maps show a nearly radial symmetry around the ionizing star. This structure is quite simple to model as observation shows there is only a prominent source of the ionizing photon at the cloud's interior. For N44\,D1, we calculate an optically thick spherical model with a covering factor 0.64. We choose an outer radius, $R$ = 7.4\,pc based on the observed geometry (Fig.\,\ref{fig:spatialcut1}). \citet{McLeod19} report a radius containing 90 per\,cent of the H$\alpha$ emission, $R_{90}$ = 7.4\,pc for N44\,D1 which agrees with the adopted radius in our model. The model constitutes the exciting star at the center of a spherical cloud, surrounded by the layers of H\,{\sc ii} region and PDR.
MUSE observation of N44\,C does not show a clear ionization stratification as N44\,D1. However, observations indicate that the majority of photons escape from N44\,C and the region harbors three ionizing stars (See Fig.\,\ref{N44_OB}). Therefore, we adopt an optically thin open geometry for N44\,C with a covering factor 0.3 and a radius of $R$ = 6.93\,pc.
\subsection{Stellar Continuum}
We use OSTAR TLUSTY models in CLOUDY for defining the stellar continuum.
We choose a model with the effective temperature ($T_{\textrm{eff}}$) of 41\,540\,K, gravity (log$g$) of 3.92, and a metallicity of 0.5Z$_\odot$, that is appropriate for a spectral class O5\,V star \citep{McLeod19} in N44\,D1. The radiation field from this O5\,V star corresponds to an incident flux of ionizing photons, log\,$\Phi(H) = 9.98\,$ph$\,s^{-1}$\,cm$^{-2}$ at the ionization front of N44\,D1 traced by the peak of [O\,{\sc i}]\,6300\AA\, emission. Here, $\Phi(H) = Q(\textrm{H})/4\pi r^2$, where the number of ionizing photons per second $Q(\textrm{H})$ is 1.82$\times10^{49}$\,ph\,s$^{-1}$ and $r$ is the distance from ionizing star to the cloud illumination face. We note that the flux of ionizing photons $\Phi$(H), plays a significant role in shape of spatial profiles and lines strengths; hence we test the models with varying $\Phi$(H) until the best-fit model is obtained. For N44\,D1 we vary the $\Phi$(H) in a range 9.98$-$10.60\,ph\,s$^{-1}$\,cm$^{-2}$.
Observations show that, N44\,C encloses three hot stars of spectral types O5\,III, O8\,V, and O9.5\,V, hence we chose three hot star model atmospheres from TLUSTY. We adopt models with $T_{\textrm{eff}}$ = 39\,500\,K and log$g$ = 3.69 for spectral type O5\,III, $T_{\textrm{eff}}$ = 33\,400\,K and log$g$ = 3.92 for O8\,V star, and $T_{\textrm{eff}}$ = 30\,500\,K and log$g$ = 3.92 for O9.5\,V star with a metallicity of 0.5\,Z$_{\odot}$.
We expect the X-rays to significantly affect the ionization balance of the gas, since the observed total X-ray luminosity is in comparable range as that of the observed H$\alpha$ luminosity in both N44\,D1 and N44\,C. \citet{chu1993} have presented the global X-ray emission in N44 using observations with the ROSAT satellite. They reported a diffuse X-ray luminosity of (0.29$-$3.5)$\times 10^{37}$ erg\,s$^{-1}$ at a characteristic temperature (1.6$-$2.5)$\times 10^{6}$\,K. We vary the Bremsstrahlung temperature ($brems$) in a range (1.6$-$2.5)$\times 10^{6}$ K.
\subsection{Abundances}
We follow the abundances provided by \citet{Toribio17} for C and O, and \citet{Garnett_2000} for He, N, Ne, S, and Ar. For the remaining species, we use the standard values included in CLOUDY for H\,{\sc ii} regions \citep{Baldwin91} and adopt an overall gaseous metallicity of 0.5\,Z$_{\odot}$. Since the dust contributes significantly to heating and the overall equilibrium of the photoionized gas in the cloud, we also include the dust grains with a metallicity scaled to half solar.
\subsection{Density}
For gas density at the illuminated face of the cloud, we use the range of electron density values obtained from the H$\alpha$ luminosity and [S\,{\sc ii}]\,6717/6732 line ratio (30 $\le n_{\textrm{e}} \le$ 180)\,cm$^{-3}$ for N44\,D1. For N44\,C, we varied the density between 40$-$100\,cm$^{-3}$, and finally obtained a value consistent with the observed electron density obtained from [S\,{\sc ii}]\,6717/6732 ratio (66\,cm$^{-3}$). We compute models with constant density as well as constant pressure in a time-steady hydrostatic cloud. When we use a filling factor of 1, we find most of the predicted line ratios differ from the observed values, and the ratios considerably change when we use a filling factor $<$\,1. We estimate an approximate range of filling factors using the relation $N^{2}_{\textrm{e}}$(rms) = $\epsilon$ $N^{2}_{\textrm{e}}$(local) \citep{Relano_2002}. Here, $N_{\textrm{e}}$(rms) is taken as the average electron density derived from H$\alpha$ flux and $N_{\textrm{e}}$(local) is the electron density obtained from the [S\,{\sc ii}]\,6717/6732 ratio.
We note that the total gas and radiation pressure vary with the position and width of the ionization front. This can also be a result of varying density at the ionization front as pressure changes. If density increases, the ionization front pushes the interior to the cloud, increasing the line emissivities. We therefore, tested two models: one with constant pressure and the other with constant density distribution.
\subsection{Constant pressure and constant density distribution}
\begin{figure}
\centering
\includegraphics[scale=0.35]{OIII_OII_OI_profile4_grace_pc_latest.eps}
\hspace{0.2 cm}
\includegraphics[scale=0.35]{N44D_constantdensity_spatialprof_mod_grace.eps}\\
\vspace{0.5 cm}
\includegraphics[scale=0.35]{N44D_constantpressure_spatialprof_grace_mod.eps}
\caption{ The spatial profiles of various emission lines in MUSE observations of N44\,D1 are shown for comparison with the constant density model and constant pressure model. a) The observed spatial profiles of [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc iii}]\,5007\AA, [N\,{\sc ii}]\,6584\AA, and [S\,{\sc ii}]\,6717\AA \,emission of N44\,D1 obtained by taking a cross-cut (red line) shown in Fig.\,\ref{fig:spatialmap}. b) The emission line profiles of [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc iii}]\,5007\AA, [N\,{\sc ii}]\,6584\AA, and [S\,{\sc ii}]\,6717\AA \, emission from constant density model of N44\,D1 are shown for comparison. c) The spatial profiles of [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc iii}]\,5007\AA, [N\,{\sc ii}]\,6584\AA, [S\,{\sc ii}]\,6717\AA \,emission from constant pressure model of N44\,D1 are shown for comparison. The best-fit model line ratios for these models are given in Table \ref{linesratio_N44D1}. All the line emissivities are normalized to 1.0.}
\label{fig:spatialcut1}
\end{figure}
\begin{figure*}
\includegraphics[width=8.5cm]{N44C_2-spatial_prof_grace_latest.eps}
\hspace{0.2 cm}
\includegraphics[width=8.5cm]{N44C_constantdensity_spatial_prof_grace.eps}
\caption{ a) The observed spatial profiles of [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc iii}]\,5007\AA, [N\,{\sc ii}]\,6584\AA, and [S\,{\sc ii}]\,6717\AA \,emission of N44\,C obtained by taking a cross-cut (red line) shown in Fig.\,\ref{fig:spatialmap}. b) The spatial profiles of [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc iii}]\,5007\AA, [N\,{\sc ii}]\,6584\AA, and [S\,{\sc ii}]\,6717\AA \, emission from constant density model obtained for N
4\,C are shown for comparison. The best-fit model line ratios for this model are given in Table \ref{linesratio_N44C}.}
\label{lineprof_mod_N44D}
\end{figure*}
In constant pressure model, the total pressure is kept constant throughout the cloud. The sum of gas pressure, line radiation pressure, turbulent pressure and the outward pressure of star light remain constant. At any particular region on the cloud, the resulting forces are due to the various contributions to the pressure balance. Therefore the cloud remains in hydrostatic equilibrium.
The hydrostatic equilibrium model would balance the pressure gradient due to the kinetic energy and momentum carried by stellar photon flux, with the thermal gas pressure exerted by ionized gas. However, this model significantly changes the gas density and width of the ionization front. The constant density represents the total hydrogen density constant throughout the nebula, but electron and molecular fractions vary with depth. In Fig.\,\ref{lineprof_mod_N44D}, we compare the spatial profiles of various emission lines, [O\,{\sc i}]\,6300\AA, [O\,{\sc ii}]\,7318\AA, [O\,{\sc iii}]\,5007\AA, [N\,{\sc ii}]\,6584\AA, [S\,{\sc ii}]\,6717\AA \,obtained for N44\,D1 with constant density as well as constant pressure models. We note that constant pressure models provide a relatively narrower ionization fronts than constant density models. \citet{pellegrini07,pellegrini09} applied constant pressure models to Orion bar and M17 PDRs for a self-consistent simulation of H$^+$, H$^0$ and H$_2$ regions, including additional turbulent pressure and magnetic field. In that model, the physical depth, the separation of H$^0$, and H$_2$, and overall geometry depend on the gas density. In our constant pressure model, we include additional turbulence of 10\,km\,s$^{-1}$. This provides turbulent line broadening in addition to the thermal line broadening, and slightly increases the widths of the spatial profiles in the ionization front. \citet{pellegrini07,pellegrini09} found that the presence of a magnetic field significantly increases the physical width of the PDR, as a result of reduced gas density and increased photon path length. We calculate the constant pressure models consisting of gas pressure, and turbulent pressure. We did not include the pressure due to the magnetic field, as currently there is no observational evidence of a magnetic field in N44. In Table\,\ref{linesratio_N44D1} we show the diagnostic line ratios obtained for both constant pressure and constant density model along with the observed line ratios of N44\,D1. We note that the constant pressure model describes most of the line ratios, however, we obtain reduced $\chi^2$ for most of the line ratios in the constant density model. We also note that geometry of O\,{\sc iii}]\,5007\AA\, spatial profile matches reasonably well with observation in the constant density model. The constant pressure model predicts a relatively larger electron density than observed. Our these studies show that the depth of the H$^+$ region, the width of ionization front, and overall geometry of the emission line profiles with depth are largely dependent on the thermal gas pressure and stellar radiation pressure. Within H$^+$ regions, the density can be constant with the depth; hence the predicted line ratios have a minimal effect on the chosen equation of state.
\subsection{Photoionization model of N44\,D1}
Even though observations give a photon escape fraction of 0.36 for N44\,D1, it does not mean that the H\,{\sc ii} region is completely ionization-bounded. We expect the photon leakage from some part of the cloud; hence that direction can be density-bounded, and the remaining part of the cloud can be ionization-bounded. This model can be a Blister type H\,{\sc ii} region as suggested by \citet{pellegrini11}.
It is also interesting to note that N44\,D1 shows relatively large [O\,{\sc iii}]/H$\beta$ ratios, implying a large ionization parameter which can be a result of the high effective temperature of the only ionizing, O5\,V, star in N44\,D1. However, to obtain a high [O\,{\sc iii}]/H$\beta$ ratio, we include an additional contribution from a Bremsstrahlung component with the temperature in a range 10$^{6.2}$--10$^{6.4}$\,K. This plasma temperature is inferred from the studies of the X-ray emission by \citep{chu1993}. These authors have reported an excess of X-ray emission in the N44\,D region. This excess X-ray emission can be shock-driven due to massive stellar winds or the off-center supernovae remnant. We find that partial ionization-bounded geometry with constant density models can reproduce reasonably well the observed geometry and line ratios of N44\,D1 for a log\,$\Phi$(H)$=10.14$\,ph\,s$^{-1}$\,cm$^{-2}$, electron density 136\,cm$^{-3}$, and Bremsstrahlung temperature 1.67$\times 10^{6}$\,K.
\subsection{Photoionization model of N44\,C}
The observed geometry of N44\,C does not give a well-defined ionization front at the H\,{\sc ii} region boundary, and the patches of [S\,{\sc ii}] and [O\,{\sc ii}] concentrations are found closer to the cloud centroid. We obtain an optimal model for N44\,C by applying an optically-thin constant density distribution. We note that the shape of the incident radiation mainly depends on three fixed stellar atmosphere models, which are appropriate for three enclosed hot stars O5\,III, O8.5\,V, and O9.5\,V. The [O\,{\sc iii}]/H$\beta$ ratio in N44\,C is relatively low compared to N44\,D1, indicating low ionization parameter and to obtain this gas ionization we did not include any additional Bremsstrahlung component as in N44\,D1. Since the observations show that 70$\%$ of photons escape from N44\,C, we use an optically thin open geometry where the Lyman continuum optical depth ($\tau_{912}$) is found to be very low ($<$\,1), hence the majority of the cloud is open to the Lyman photons. We note, this model can reasonably well describe the observed geometry and the line ratios of N44\,C. Fig.\,\ref{lineprof_mod_N44D} shows the model emission line profiles as a function of depth in the nebula, and the line ratios are given in Table\,\ref{linesratio_N44C}. Comparison of our models with the observed geometry and line ratios indicate that N44\,C has an optically thin geometry and the region is being energized mainly by the outward momentum carried by the radiation pressure from three ionizing stars.
\section{Conclusions}
We carry out a detailed analysis of two H\,{\sc ii} regions in N44 using the integral field optical spectroscopic observations obtained with MUSE. Comparing these observations with the photoionization models computed with CLOUDY, we describe the spatial distribution of emission line geometry and the physical conditions.
Our results are summarized as follow:
\begin{enumerate}
\item Our analysis reveals that the spatial distribution of various spectral lines in N44\,D1 provides a stratified ionization geometry. The central ionizing star is covered by a fully ionized hydrogen gas and at the periphery, there is a well-defined transition zone from O$^{++}$ through O$^+$ to the neutral zone O$^0$. H$\alpha$, H$\beta$ and [O\,{\sc iii}] emission are co-spatial and peak at the fully ionized zone, while [O\,{\sc ii}], [N\,{\sc ii}], [S\,{\sc ii}] and [O\,{\sc i}] emission peak at the outer boundary. This region provides an excellent site for modeling an ideal H\,{\sc ii} region with a stratified ionization geometry. The [O\,{\sc i}]\,6300\AA\, emission of the N44 D1 region reveals a clear boundary/transition zone in the outer boundary of [O\,{\sc iii}]\,5007\AA\, and [O\,{\sc ii}]\,7329\AA\, emission, which does not cover the entire nebula completely, indicating a partial ionization front. Comparing these studies with \citet{pellegrini11, pellegrini12}, we suggest that the N44\,D1 is a Blister H\,{\sc ii} region.
\item The spatial distributions of various spectral lines in N44\,C do not show a stratified ionization front at the boundary. However, it shows the condensations of [S\,{\sc ii}] and [N\,{\sc ii}] emission within the H\,{\sc ii} region. The [O\,{\sc i}]\,6300\AA\, emission is relatively weak in N44\,C and does not show a well-defined outer boundary as in N44\,D1. These observations support relatively higher photon escape fraction reported by \citet{McLeod19}, suggesting that N44\,C is a density-bounded optically thin H\,{\sc ii} region.
\item Our studies reveal that the [O\,{\sc iii}]/H$\alpha$ and [O\,{\sc iii}]/H$\beta$ line ratios give a clear indication of a higher degree of ionization in the regions closer to the ionizing stars and the ratios are lower toward the boundary of N44\,D1 and N44\,C. [S\,{\sc ii}]/H$\alpha$ and [N\,{\sc ii}]/H$\beta$ line ratio maps show a shell structure in both N44\,D1 and N44\,C. [O\,{\sc iii}]/H$\alpha$ and [O\,{\sc iii}]/H$\beta$ of N44\,D1 are much higher than N44\,C, indicating a harder radiation field. The effective temperature of the hot star plays a key role here because N44\,D1 has a hotter ionizing star (O5\,V) than N44\,C (O5\,III).
\item We use our results of spatially resolved MUSE spectra to explore the photoionization models with CLOUDY that can well describe the observed geometry and emission line ratios. We find that the constant density model gives better geometry and line ratios than the constant pressure model in the N44\,D1. An ionization-bounded geometry with a partial covering factor can well reproduce the observed geometry and line ratios, indicating that N44\,D1 is a Blister H\,{\sc ii} region. The spatial profile of [O\,{\sc iii}]\,5007\AA\, matches very nicely with the observation. Model calculations reveal that a significant amount of X-ray emission takes part in shaping the geometry of the emission line profiles, in addition to the ionizing radiation from the O5\,V star. The electron density and temperature values from this model are also consistent with our measurements from the observed values.
\item An optically thin and open geometry model has been applied to reproduce the observed geometry and line ratios in N44\,C. The modeling results show that N44\,C region is mainly energized by the radiation from three ionizing stars. Our studies indicate that the ionization structure and physical conditions in N44\,D1 and N44\,C are set by the stellar radiation pressure and gas thermal pressure.
\end{enumerate}
\section*{Acknowledgments}
This research has been supported by the United Arab Emirates University (UAEU) through UAEU Program for Advanced Research (UPAR) grant G00003479 and start-up grant G00002964. This paper makes use of the following MUSE (VLT) data: program ID: 096.C--0137(A). F. Kemper acknowledges the Ministry of Science and Technology of Taiwan for the grant MOST107-2119-M-001-031-MY3 and Academia Sinica Investigator Award, AS-IA-106-M03. M. Sewilo acknowledges the NASA award, 80GSFC21M0002 (M. S.).
\begin{table*}
\centering
\caption{Luminosities of observed emission lines }
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{lcccccccc}
\hline
\multicolumn{1}{c}{Emission lines}&
\multicolumn{2}{c}{N44\,D1}&
\multicolumn{2}{c}{N44\,D2}&
\multicolumn{2}{c}{N44\,C}&\\
&$L_{\textrm{obs}}$(erg s$^{-1}$)$^{\text{a}}$&$L_{\textrm{int}}$(erg s$^{-1}$)$^{\text{b}}$&$L_{\textrm{obs}}$(erg s$^{-1}$)&$L_{\textrm{int}}$(erg s$^{-1}$)&$L_{\textrm{obs}}$(erg s$^{-1}$)&$L_{\textrm{int}}$(erg s$^{-1}$)\\
&$\times$10$^{36}$&$\times$10$^{36}$&$\times$10$^{36}$&$\times$10$^{36}$&$\times$10$^{36}$&$\times$10$^{36}$&\\
\hline
H$\alpha$ &14.70\,$\pm$\,1.62 &17.5\,$\pm$\,1.93 &5.36\,$\pm$\,0.58 &5.45\,$\pm$\,0.69 &13.9\,$\pm$\,1.53 &21.5\,$\pm$\,1.86\\
H$\beta$ &4.68\,$\pm$\,0.47 &6.12\,$\pm$\,0.61 &1.71\,$\pm$\,0.24 &1.76\,$\pm$\,0.31 &3.69\,$\pm$\,0.55 &7.24\,$\pm$\,0.75\\
\big[N {\sc ii}\big]6584\AA &0.80\,$\pm$\,0.13 &0.96\,$\pm$\,0.16 &0.29\,$\pm$\,0.07 &0.30\,$\pm$\,0.08 &0.97\,$\pm$\,0.16 &1.50\,$\pm$\,0.20\\
\big[O {\sc i}\big]6300\AA &0.32\,$\pm$\,0.07 &0.39\,$\pm$\,0.08 &0.07\,$\pm$\,0.03 &0.07\,$\pm$\,0.03 &0.13\,$\pm$\,0.04 &0.21\,$\pm$\,0.05\\
\big[O {\sc ii}\big]7318\AA &0.16\,$\pm$\,0.05 &0.19\,$\pm$\,0.06 &0.06\,$\pm$\,0.03 &0.06\,$\pm$\,0.03 &0.17\,$\pm$\,0.05 &0.24\,$\pm$\,0.06\\
\big[O {\sc ii}\big]7329\AA &0.15\,$\pm$\,0.04 &0.17\,$\pm$\,0.04 &0.05\,$\pm$\,0.02 &0.05\,$\pm$\,0.02 &0.14\,$\pm$\,0.04 &0.20\,$\pm$\,0.04\\
\big[O {\sc iii}\big]4959\AA &12.40\,$\pm$\,0.64 &16.1\,$\pm$\,2.13 &2.98\,$\pm$\,0.38 &3.06\,$\pm$\,0.50 &2.61\,$\pm$\,0.04 &5.02\,$\pm$\,0.56\\
\big[O {\sc iii}\big]5007\AA &36.90\,$\pm$\,3.85 &48.5\,$\pm$\,5.06 &8.87\,$\pm$\,0.99 &9.11\,$\pm$\,1.30 &7.66\,$\pm$\,1.09 &15.1\,$\pm$\,1.48\\
\big[S {\sc ii}\big]6717\AA &1.11\,$\pm$\,0.16 &1.31\,$\pm$\,0.18 &0.34\,$\pm$\,0.07 &0.35\,$\pm$\,0.08 & 0.94\,$\pm$\,0.15&1.43\,$\pm$\,0.18\\
\big[S {\sc ii}\big]6732\AA &0.83\,$\pm$\,0.13 &0.98\,$\pm$\,0.15 &0.26\,$\pm$\,0.06 &0.26\,$\pm$\,0.07 &0.68\,$\pm$\,0.11 &1.04\,$\pm$\,0.14\\
\big[S {\sc iii}\big]9069\AA &1.48\,$\pm$\,0.21 &1.62\,$\pm$\,0.23 &0.60\,$\pm$\,0.10 &0.61\,$\pm$\,0.11 &1.45\,$\pm$\,0.18 &1.82\,$\pm$\,0.20\\
\\
\hline
\label{lines}
\end{tabular}
\parbox{160 mm}{a: $L_{\textrm{obs}}$ is the observed luminosity.\\ b: $L_{\textrm{int}}$ is the reddening corrected luminosity.}
\end{table*}
\begin{table*}
\centering
\caption{Emission line properties}
\setlength{\tabcolsep}{11pt}
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{@{}lccccccc}
\hline
\multicolumn{1}{c}{Regions}&
\multicolumn{1}{c}{$Q$}&
\multicolumn{1}{c}{$Q_{\textrm{o}}$}&
\multicolumn{1}{c}{$f_{\textrm{esc}}$}&
\multicolumn{1}{c}{$\langle n_{\textrm{e}}\rangle ^{\text{a}}$}&
\multicolumn{1}{c}{$n_{\textrm{e}}$ [S {\sc ii}] $^{\text{b}}$}&
\multicolumn{1}{c}{$n_{\textrm{e}}$ [S {\sc ii}] $^{\text{c}}$}\\
&(s$^{-1}$)&(s$^{-1}$)&&(cm$^{-3}$)&(cm$^{-3}$)&(cm$^{-3}$)&\\
\hline
N44\,D1&1.07$\times 10^{49}$&1.66 $\times 10^{49}$&0.36&31&132$\pm$50& 141$\pm$43\\
N44\,D2&0.36$\times 10^{49}$&1.26$\times 10^{49}$&0.71&26&115$\pm$45& 121$\pm$37\\
N44\,C&1.02$\times 10^{49}$&3.37$\times 10^{49}$&0.70 &38 &66$\pm$40 & 92$\pm$35\\
\hline
\label{properties}
\end{tabular}
\parbox{150 mm}{a: $\langle n_{\textrm{e}}\rangle$ is the average electron density from H$\alpha$ emission.\\ b: $n_{\textrm{e}}$[S {\sc ii}] is the electron density derived using PYNEB.\\ c: $n_{\textrm{e}}$[S {\sc ii}] is the electron density derived from the equation no. 11.}
\end{table*}
\begin{table*}
\centering
\caption{Model line ratios compared with observations for N44\,D1}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{lcccccccccc}
\hline
\multicolumn{1}{c}{Line ratios}
&
\multicolumn{1}{c}{Observed}&
\multicolumn{1}{c}{Pressure Model}&
\multicolumn{1}{c}{$\chi^{2}$}&
\multicolumn{1}{c}{Density Model}
\multicolumn{1}{c}{$\chi^{2}$}&\\
&ratios &$\Phi$(H)$_{10.15}$& & $\Phi$(H)$_{10.14}$& \\
\hline
\big[S {\sc ii}\big]\,6717/6732 &1.32$\pm$0.40 &1.28& 0.01 &1.32& 0.00\\
\big[O {\sc ii}\big]\,(7318+7329)/[O {\sc iii}]5007 &0.008$\pm$0.002&0.004 & 4.0 &0.006& 1.00\\
\big[O {\sc iii}\big]\,4959/5007 &0.33$\pm$0.07 &0.33& 0.00 &0.33& 0.00\\
\big[O {\sc iii}\big]\,5007/H$_\beta$ &7.89$\pm$1.60 &6.76& 0.50 &7.10& 0.24\\
\big[S {\sc iii}\big]\,9069/H$_\beta$ &0.31$\pm$0.08 &0.36& 0.40 &0.37& 0.56\\
\big[O {\sc ii}\big]\,7329/H$_\beta$ &0.03$\pm$0.01 &0.02& 1.0 &0.02& 1.00\\
\big[S {\sc ii}\big]\,6717/H$_\beta$ &0.24$\pm$0.10 &0.36& 1.44 &0.34& 1.00\\
\big[N {\sc ii}\big]\,6584/H$_\beta$ &0.17$\pm$0.05 &0.14& 0.36 &0.14& 0.36\\
\big[O {\sc i}\big]\,6300/H$_\beta$ &0.07$\pm$0.05 &0.15& 2.56 &0.14& 1.96\\
\big[O {\sc iii}\big]\,5007/H$_\alpha$ &2.52$\pm$0.50 &2.40& 0.06 &2.51& 0.00\\
\big[S {\sc ii}\big]\,6717/H$_\alpha$ &0.08$\pm$0.02 &0.13& 6.30 &0.12& 4.00\\
\big[N {\sc ii}\big]\,6584/H$_\alpha$ &0.05$\pm$0.02 &0.05& 0.00 &0.05& 0.00\\
$T_{\textrm{e}}$ over radius (K) & &12\,491& &12\,410& \\
$n_{\textrm{e}}$ over radius (cm$^{-3}$) & &175& &136&
\\
\hline
\label{linesratio_N44D1}
\end{tabular}
\vspace{-8mm}
\end{table*}
\begin{table*}
\centering
\caption{Model line ratios compared with observations for N44\,C}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{lcccc}
\hline
\multicolumn{1}{c}{Line ratios}&
\multicolumn{1}{c}{Observed ratios}&
\multicolumn{1}{c}{Model ratios}&
\multicolumn{1}{c}{$\chi^{2}$}&\\
\hline
\big[S {\sc ii}\big]\,6717/6732 &1.37$\pm$0.4 &1.35 & 0.002 \\
\big[O {\sc ii}\big]\,(7318+7329)/[O {\sc iii}]5007 &0.04$\pm$0.02 &0.02 & 1.00\\
\big[O {\sc iii}\big]\,4959/5007 &0.34$\pm$0.01 &0.33 & 1.00\\
\big[O {\sc iii}\big]\,5007/H$_\beta$ &2.07$\pm$0.6 &1.75& 0.28 \\
\big[S {\sc iii}\big]\,9069/H$_\beta$ &0.39$\pm$0.1 & 0.28 & 1.20\\
\big[O {\sc ii}\big]\,7329/H$_\beta$ &0.03$\pm$0.02 &0.02 & 0.25\\
\big[S {\sc ii}\big]\,6717/H$_\beta$ &0.25$\pm$0.08 &0.30 & 0.39\\
\big[N {\sc ii}\big]\,6584/H$_\beta$ &0.26$\pm$0.08 &0.19 & 0.76\\
\big[O {\sc i}\big]\,6300/H$_\beta$ &0.03$\pm$0.01 &0.03 & 0.00\\
\big[O {\sc iii}\big]\,5007/H$_\alpha$ &0.55$\pm$0.14 &0.63 & 0.32 \\
\big[S {\sc ii}\big]\,6717/H$_\alpha$ &0.07$\pm$0.02 &0.10 & 2.25\\
\big[N {\sc ii}\big]\,6584/H$_\alpha$ &0.07$\pm$0.02 &0.07 & 0.00 \\
$T_{\textrm{e}}$ over radius (K) &&10\,709 & \\
$n_{\textrm{e}}$ over radius (cm$^{-3}$) &&84 & \\
\hline
\label{linesratio_N44C}
\end{tabular}
\vspace{-8mm}
\end{table*}
|
\section{Introduction}
In this article we prove convergence results of orthogonal
series of certain tensor products of splines in the spirit of
the known results for martingales.
We begin by discussing the
situation for martingales and, subsequently, for univariate splines.
For martingales, we use \cite{Neveu1975} and \cite{Pisier2016} as
references.
Let $(\Omega, (\mathscr F_n), \mathbb P)$ be a filtered probability space. A
sequence of integrable functions $(X_n)_{n\geq 1}$ is a \emph{martingale} if $\mathbb
E(X_{n+1} | \mathscr F_n) = X_n$ for every $n$, where we denote by $\mathbb
E(\cdot | \mathscr F_n)$ the conditional expectation operator with respect to
the $\sigma$-algebra $\mathscr F_n$. This operator is the orthoprojector onto
the space of $\mathscr F_n$-measurable $L^2$-functions and it
can be extended to $L^1$.
Observe that if $X\in L^1$, the sequence $(\mathbb E(X|\mathscr F_n))$
is a martingale.
In this case, we have that $\mathbb E(X|\mathscr F_n)$ converges
almost surely to $\mathbb E(X|\mathscr F)$ with $\mathscr F = \sigma(
\cup_n\mathscr F_n)$.
For general scalar-valued martingales, we have the following convergence
theorem:
any martingale $(X_n)$ that is bounded in $L^1$ has
an almost sure limit function contained in $L^1$.
Additionally we know that martingale differences $dX_n = X_n - X_{n-1}$ converge unconditionally in
$L^p$ for $1<p<\infty$, i.e., we have the inequality
\begin{equation}\label{eq:strong_p_mart}
\Big\| \sum_n \varepsilon_n dX_n \Big\|_p \leq C_p \Big\| \sum_n dX_n
\Big\|_p
\end{equation}
for all sequences of signs $(\varepsilon_n)$ and some
constant $C_p$ depending only on $p$.
More precisely, we have the following inequality of weak type:
\begin{equation}\label{eq:wt_mart}
\sup_{\lambda >0} \lambda \cdot \mathbb P\Big( \sup_{n} \Big| \sum_{\ell\leq n}
\varepsilon_\ell dX_\ell\Big| > \lambda\Big) \leq C \sup_{n} \| X_n \|_1,
\end{equation}
where $C$ some absolute constant.
Equation \eqref{eq:strong_p_mart} is a consequence of \eqref{eq:wt_mart} since by
orthogonality of martingale differences $dX_n$ we have
\eqref{eq:strong_p_mart} for $p=2$ and the Marcinkiewicz interpolation theorem
then implies \eqref{eq:strong_p_mart} for every $p$ in the range $1<p<\infty$.
Consider now the special case where each $\sigma$-algebra $\mathscr F_n$ is generated by
a partition of a bounded interval $I\subset \mathbb R$ into finitely many
intervals $(I_{n,i})_i$ of positive length as atoms of $\mathscr F_n$. In this case, $(\mathscr F_n)$ is
called an \emph{interval filtration} on $I$.
Then, the
characteristic functions $(\ensuremath{\mathbbm 1}_{I_{n,i}})$ of those atoms are a sharply localized
orthogonal basis of $L^2(\mathscr F_n)$ with respect to Lebesgue measure
$|\cdot|$.
If we want to preserve the localization property of the basis functions, but
at the same time consider spaces of functions with higher smoothness, a natural
candidate are spaces of piecewise polynomial functions of order $k$, given by
\begin{equation}\label{eq:defSk}
\begin{aligned}
S_{k}(\mathscr F_n) = \{f : I\to \mathbb R\ |\ &\text{$f$ is $k-2$ times continuously
differentiable and} \\
&\qquad\text{a polynomial of order $k$ on
each atom of $\mathscr F_n$} \},
\end{aligned}
\end{equation}
where $k$ is an arbitrary positive integer.
One reason for this is that $S_k(\mathscr F_n)$ admits a special basis, the so
called
B-spline basis $(N_{n,i})_i$, that consists of non-negative and localized
functions $N_{n,i}$.
Here, the term ``localized'' means that the support of each function $N_{n,i}$
consists of at most $k$ neighbouring atoms of $\mathscr F_n$.
A second reason is that if $(\mathscr F_n)$ is an increasing sequence of
interval $\sigma$-algebras, then the sequence of corresponding spline spaces $S_k(\mathscr
F_n)$ is increasing as well.
Note that the aforementioned properties of the B-spline functions $(N_{n,i})$ imply that they
do not form an orthogonal basis of $S_k(\mathscr F_n)$ for $k\geq 2$.
For more information on spline functions,
see e.g. \cite{Schumaker2007}.
Let $P_n^{k}$ be the orthogonal projector onto
$S_{k}(\mathscr F_n)$ with respect to the
$L^2$ inner product on $I$ equipped with the Lebesgue measure.
Since the space $S_{1}(\mathscr F_n)$ consists of
piecewise constant functions,
$P_n^{1}$ is the conditional expectation
operator with respect to the $\sigma$-algebra $\mathscr F_n$ and the Lebesgue
measure.
In general, the operator $P_n^{k}$ can be written in terms of the B-spline basis $(N_{n,i})$ as
\[
P_n^{k} f = \sum_i \int_I f(x) N_{n,i}(x)\,\mathrm{d} x \cdot N_{n,i}^*,
\]
where the functions $(N_{n,i}^*)$,
contained in the spline space $S_{k}(\mathscr F_n)$, are the biorthogonal (or
dual) system to the
B-spline basis $(N_{n,i})$.
Due to the uniform boundedness of the B-spline functions
$N_{n,i}$, we are able to insert functions $f$ in formula
\eqref{eq:Pn} that are contained not only in $L^2$, but in
$L^1$, thereby extending the operators
$P_n^{k}$ to $L^1$.
Similarly to the definition of martingales, we adopt the following
notion introduced in \cite{Passenbrunner2020}:
let $(X_n)_{n\geq 1}$ be a sequence of functions in the space $L^1$. We call this
sequence a \emph{$k$-martingale spline sequence} (adapted to $(\mathscr F_n))$
if
\[
P_n^{k} X_{n+1} = X_n,\qquad n\geq 1.
\]
The local nature of the B-splines and the nestedness of the spaces
$(S_k(\mathscr F_n))_n$ ultimately allow us to
transfer the classical martingale theorems discussed above
to
$k$-martingale spline sequences adapted to \emph{arbitrary}
interval filtrations ($\mathscr F_n$) and for any positive integer $k$, just by replacing
conditional expectation operators with the spline projection operators
$P_n^{k}$.
Assume that, for all $n$,
$\mathscr F_n$ arises from $\mathscr F_{n-1}$ by the subdivision of exactly
one atom of $\mathscr F_{n-1}$ into two intervals $L_n, R_n$ as atoms of
$\mathscr F_n$.
In the case $k=1$, spline differences $dX_n$ are the same as martingale
differences and then given by a constant multiple
of the generalized Haar function $|R_{n}| \ensuremath{\mathbbm 1}_{L_n} - |L_n|
\ensuremath{\mathbbm 1}_{R_n}$. Similarly, for $k>1$, there exists a system of (unlocalized)
orthogonal spline functions $(f_n)$ so that $dX_n$ is a constant multiple of $f_n$.
The following statements are true:
\begin{enumerate}
\item $L^1$-bounded $k$-martingale spline sequences $(X_n)$
converge almost everywhere to some $L^1$-function.
\cite{PassenbrunnerShadrin2014,
MuellerPassenbrunner2020}
\item
Inequality \eqref{eq:strong_p_mart} holds for $k$-martingale
spline sequences $(X_n)$ with a constant $C_{p,k}$
depending only on $p$ and $k$ but \emph{not} on the interval filtration
$(\mathscr F_n)$ (see \cite{GevorkyanKamont2004} for $k=2$ and
\cite{Passenbrunner2014} for general $k$).
\item \label{it:cz} By using Calderon-Zygmund operator techniques, A. Kamont and K. Keryan
\cite{KamontKeryan2021}
showed under certain geometric conditions ($(k-1)$-regularity, cf.
Definition~\ref{def:regularity} for $d=1$) on the filtration
$(\mathscr F_n)$ that \eqref{eq:wt_mart} also holds for $k$-martingale
spline sequences $(X_n)$.
\end{enumerate}
In this article we are concerned with similar results pertaining to
tensor product spline projections.
Let $d$ be a positive integer and, for $j=1,\ldots,d$, let $(\mathscr
F_n^{j})$ be an interval filtration on the bounded interval $I^j\subset \mathbb R$. Filtrations $(\mathscr
F_n)$ of the form $\mathscr F_n = \mathscr F_n^1\otimes \cdots \otimes
\mathscr F_n^d$ will be called an \emph{interval filtration} on
the $d$-dimensional rectangle $I^1 \times \cdots \times I^d$.
Then, the atoms of $\mathscr F_n$ are of the form
$A_1\times \cdots\times A_d$ with atoms $A_j$ in $\mathscr F_n^j$.
For a tuple $k=(k_1,\ldots,k_d)$ consisting of $d$ positive integers,
denote by $P_n^{k}$
the orthogonal projector with respect to
$d$-dimensional Lebesgue measure $|\cdot|$ onto the tensor product
spline space $S_k(\mathscr F_n) = S_{k_1}(\mathscr F_n^{1}) \otimes \cdots \otimes
S_{k_d}(\mathscr F_n^{d})$.
In \cite{Passenbrunner2021} we show that an $L^1$-bounded sequence of functions $(X_n)$
with $P_n^k X_{n+1} = X_n$ converges almost everywhere to some
$L^1$-function.
Now we assume that, for $n\geq 1$, $\mathscr F_{n}$ is of the form that $\mathscr
F_{n}=\mathscr F_{n}^{1} \otimes \cdots \otimes \mathscr F_{n}^d$ arises
from $\mathscr F_{n-1} = \mathscr F_{n-1}^{1} \otimes \cdots \otimes \mathscr
F_{n-1}^d$
in the way that there exists a coordinate $\delta_0\in \{1,\ldots,d\}$ so that
$\mathscr F_{n}^\delta = \mathscr F_{n-1}^\delta$ for $\delta\neq \delta_0$ and
$\mathscr F_{n}^{\delta_0}$
arises from $\mathscr F_{n-1}^{\delta_0}$ by splitting exactly one atom of $\mathscr
F_{n-1}^{\delta_0}$ into two atoms of $\mathscr F_{n}^{\delta_0}$.
In Section~\ref{sec:tensor_def} we describe an
orthonormal system $(f_\ell)$ consisting of tensor products of spline
functions so that there exists an increasing sequence of integers
$(M_n)$ satisfying
\[
S_k(\mathscr F_n) = \operatorname{span}\{ f_\ell : \ell \leq M_n \}\qquad
\text{for all $n$}.
\]
In the special case $k=(1,\ldots,1)$, if $M_{n-1} < \ell \leq M_n$,
those functions $f_\ell$ are given by (a constant multiple of) the tensor product of
one generalized Haar function in direction $\delta_0$ with
$(d-1)$ characteristic
functions of atoms in $(\mathscr F_n^\delta)$ in the directions
$\delta\neq \delta_0$. Therefore, in this case, $(f_\ell)$ is a
martingale difference sequence.
Here, we extend the result concerning a.e. convergence from
\cite{Passenbrunner2021} and show that partial sums of the form
$X_n=\sum_{\ell\leq n} a_\ell f_\ell$, that are uniformly
$L^1$-bounded, converge almost
everywhere. Moreover, we give sufficient geometric conditions on the
filtration $(\mathscr F_n)$ (cf. Theorem~\ref{thm:main}) so that we
have the following weak type inequality,
similar to \eqref{eq:wt_mart}:
\begin{equation}\label{eq:wt_intro}
\sup_{\lambda > 0} \lambda \cdot \Big| \Big\{ \sup_n \Big|
\sum_{\ell\leq n} \varepsilon_\ell a_\ell f_\ell \Big|
> \lambda\Big\} \Big| \leq C \sup_n \| X_n \|_1
\end{equation}
for some constant $C$,
all sequences $(\varepsilon_n)$ of signs and all sequences of
coefficients $(a_n)$.
We note two things.
Firstly, by specializing Theorem~\ref{thm:main} to $d=1$,
our sufficient conditions on $(\mathscr F_n)$ for inequality
\eqref{eq:wt_intro} are less restrictive than the sufficient condition in
item \eqref{it:cz} ($(k-1)$-regularity implies $k$-regularity,
cf. Definition~\ref{def:regularity} and the succeeding remark).
Secondly, for $d\geq 2$, our sufficient conditions allow for an arbitrary
ratio of sidelengths of atoms of $\mathscr F_n$, meaning that the
rectangles that are atoms of $\mathscr F_n$ can be very long in one
direction and very short in another direction.
The organization of the article is as follows. In
Section~\ref{sec:prelim} we collect known results about polynomials and
spline functions that are used in the
sequel. In Section~\ref{sec:tensor_orth} we construct multivariate
orthonormal spline functions $(f_n)$. Section~\ref{sec:main} contains
the formulation of our main result (Theorem~\ref{thm:main}) that
inequality \eqref{eq:wt_intro} is valid under
certain geometric conditions on the filtration $(\mathscr F_n)$ that
are also defined and analyzed here. Finally,
Section~\ref{sec:proof_main} contains the proof of the main result.
\section{Preliminaries}\label{sec:prelim}
\subsection{Polynomials}
We will need the following multi-dimensional version of Remez' theorem
(see \cite{Ga2001,BrudnyiGanzburg1973}). If
$p(x)=\sum_{\alpha\in \Lambda} a_\alpha x^\alpha$ is a $d$-variate polynomial
where $\Lambda$ is a finite set containing $d$-dimensional multiindices, the
degree of $p$ is defined as $\max\{\sum_{i=1}^d \alpha_i :
\alpha\in \Lambda\}$.
Recall that a convex body in $\mathbb R^d$ is a compact, convex set with non-empty interior.
\begin{thm}[Remez, Brudnyi, Ganzburg]
Let $d\in\mathbb N$, $V\subset \mathbb R^d$ a convex body and $E\subset V$ a measurable
subset. Then, for all polynomials $p$ of degree $r$ on $V$,
\begin{equation*}
\| p \|_{L^\infty(V)} \leq \bigg( 4d \frac{|V|}{|E|}\bigg)^r \| p
\|_{L^\infty(E)}.
\end{equation*}
\end{thm}
We have the following corollary:
\begin{cor} \label{cor:remez}
Let $p$ be a polynomial of degree $r$ on a convex body $V\subset \mathbb R^d$. Then
\begin{equation*}
\big|\big\{ x \in V : |p(x)| \geq (8d)^{-r} \|p\|_{L^\infty(V)} \big\}\big| \geq |V|/2.
\end{equation*}
\end{cor}
\begin{proof}
This follows from an application of the above theorem to the set $E = \{x\in
V : |p(x)| \leq (8d)^{-r} \|p\|_{L^\infty(V)}\}$.
\end{proof}
\subsection{Spline spaces}
Consider an \emph{interval $\sigma$-algebra $\mathscr F$}, i.e. a
$\sigma$-algebra that is generated by
a partition of a bounded interval $I\subset \mathbb R$ into finitely many
intervals of positive length as atoms of $\mathscr F$. Let $k$ be an arbitrary
positive integer.
Let $S_k(\mathscr F)$ be the spline space of order $k$ corresponding to
the $\sigma$-algebra $\mathscr F$ defined in~\eqref{eq:defSk} and let $(N_i)$ be
the B-spline basis of $S_k(\mathscr F)$ that forms a partition of unity.
In the next result and in what follows,
we use the notation $A(t)\lesssim_x B(t)$ if there exists a
constant $c$ depending only on the order parameter $k$ and on $x$ so that $A(t)\leq c B(t)$
for all $t$, where $t$ denotes all implicit or explicit dependencies that the
symbols $A$ and $B$ might have. Similarly we use the symbols $\gtrsim_x$ and $\simeq_x$.
\begin{prop}[B-spline stability]\label{prop:lpstab}
Let $1\leq p< \infty$ and $g=\sum_{j} a_j N_{j}$ be a linear
combination of B-splines. Then,
\begin{equation}\label{eq:lpstab}
|a_j|\lesssim |K_j|^{-1/p}\|g\|_{L^p(K_j)},\qquad \text{for all $j$},
\end{equation}
where $K_j\subseteq \operatorname{supp} N_{j}$ is an atom of $\mathscr F$ having maximal length. Additionally,
\begin{equation}\label{eq:deboorlpstab}
\|g\|_p\simeq \Big(\sum_{j} |a_j|^p \cdot |\operatorname{supp} N_j|\Big)^{1/p}.
\end{equation}
\end{prop}
The two inequalites \eqref{eq:lpstab} and \eqref{eq:deboorlpstab} are Lemma 4.1
and Lemma 4.2 in \cite[Chapter 5]{DeVoreLorentz1993}, respectively.
For more information on spline functions,
see e.g. \cite{Schumaker2007}.
Let $P$ be the orthogonal projector onto
$S_k(\mathscr F)$ with respect to the
$L^2$ inner product $\langle\cdot,\cdot\rangle$ on $I$ equipped with the Lebesgue measure.
Since the space $S_{1}(\mathscr F)$ consists of
piecewise constant functions,
for the choice $k=1$, the operator $P$ is the conditional expectation
operator with respect to the $\sigma$-algebra $\mathscr F$ and the Lebesgue
measure.
This orthogonal projector is uniformly bounded on $L^\infty$ by a
constant depending only on the spline order $k$, which is content of the
following celebrated theorem by A. Shadrin \cite{Shadrin2001}:
\begin{thm}\label{thm:shadrin}
Let $P$ be the orthogonal projector onto {$S_k({\mathscr F})$}
with respect to the canonical inner product in $L^2(I)$.
Then,
\begin{equation*}
\| P : L^\infty(I) \to L^\infty(I) \| \lesssim 1.
\end{equation*}
\end{thm}
Since orthogonal projectors are self-adjoint, this also implies that the
operators $P$ are uniformly bounded on $L^1(I)$ by the same constant.
The operator $P$ can be written in terms of the
B-spline basis $(N_{i})$ and its dual basis $(N_i^*)$ as
\begin{equation}\label{eq:Pn}
P f = \sum_i \int_I f(x) N_{i}(x)\,\mathrm{d} x \cdot N_{i}^*,
\end{equation}
where the functions $N_i^*\in S_k(\mathscr F)$ are given by the conditions
$\langle N_i^*, N_j\rangle = \delta_{ij}$ for all $i,j$ and $\delta_{ij} = 1$ if
$i=j$ and $\delta_{ij}=0$ otherwise, where we denote $\langle f,g\rangle = \int_I f(x)g(x)\,\mathrm{d} x$.
The dual B-spline functions $N_i^*$ can be written in terms of the B-spline
basis $N_{i}^* = \sum_j a_{ij}
N_{j}$ for some coefficients $(a_{ij})$.
Those coefficients $(a_{ij})$ admit some fast decay away from the diagonal:
\begin{thm}[\cite{PassenbrunnerShadrin2014}]\label{thm:maintool}
There exists a constant $q\in(0,1)$ depending only on the spline order
$k$ so that
\[
|a_{ij}|\lesssim \frac{q^{|i-j|}}{|\conv(\operatorname{supp} N_i , \operatorname{supp} N_j)|},\qquad
\text{for all } i,j,
\]
where $\conv(A,B)$ denotes the smallest interval containing both sets $A,B$.
\end{thm}
\subsection{Totally positive matrices}\label{sec:tp}
We say that a matrix $B=(b_{ij})_{i,j=1}^n$ is \emph{totally positive} if for
any choice of ${\bf m_1, m_2}\subset \{1,\ldots,n\}$
with the same cardinality,
the determinant of the matrix $B({\bf m_1}, {\bf m_2})$, resulting from
$B$ by taking the rows with indices in ${\bf m_1}$ and columns with indices in
${\bf m_2}$, is non-negative.
For totally positive matrices, we have the following well known lemma, which can be found
in \cite{deBoorJiaPinkus}.
\begin{lem} \label{lem:tp0}
If $B\in \mathbb R^{n\times n}$ is invertible and totally positive,
then, for any integer interval $\bf m$ $\subset \{1,\ldots,n\}$, so is
the principal submatrix $C := B({\bf m},{\bf m})$ of $B$
and
\[
0 \leq (-1)^{i+j} C^{-1}(i,j) \leq (-1)^{i+j} B^{-1}(i,j),
\qquad i,j\in \bf m.
\]
\end{lem}
A proof of this result is contained in \cite{deBoor2012}.
We also have the following straightforward extension, whose proof is similar to
the proof of Lemma~\ref{lem:tp0}, but we include it here for
completeness.
\begin{lem}\label{lem:tp}
If $B\in \mathbb R^{n\times n}$ is invertible and totally positive,
then, for every ${\bf m} \subset \{1,\ldots,n\}$, so is
the principal submatrix $C := B({\bf m},{\bf m})$ of $B$
and
\begin{equation}\label{eq:inverse_relative}
0 \leq (-1)^{i+j} C^{-1}(i,j) \leq (-1)^{i+j} B^{-1}(i,j),
\qquad i,j\in\bf m.
\end{equation}
\end{lem}
\begin{proof}
Let ${\bf m_1} \subseteq \{1,\ldots,n\}$ and ${\bf m} ={\bf m_1}
\setminus \{\ell\}$ for some $\ell\in {\bf m_1}$.
Denote by $C$ the matrix $B({\bf m}, {\bf m})$ and by $D$ the matrix
$B({\bf m_1}, {\bf m_1})$, which are both totally positive.
Then we show that if $D$ is invertible, so is $C$ and the matrix $E$, given by
\begin{equation}\label{eq:E}
E(i,j) = D^{-1}(i,j) - \frac{ D^{-1}(i,\ell)
D^{-1}(\ell,j)}{D^{-1}(\ell,\ell)},\qquad i,j\in {\bf m},
\end{equation}
is the inverse of $C$.
First we note that by the Hadamard inequality for totally positive
matrices (see e.g. \cite[Theorem 1.21]{Pinkus2010}), we have
$\det D \leq D(\ell,\ell)\det C$ and we obtain by the invertibility of
$D$ that $\det D>0$ and therefore also $\det C>0$. Now we invoke the
formula $D^{-1}(\ell,\ell)= \det C/\det D$ to see that
$D^{-1}(\ell,\ell)>0$ and the formula in \eqref{eq:E} makes sense.
Next, we calculate for $r,s\in {\bf m}$
\begin{align*}
(CE)(r,s) &= \sum_{u\in {\bf m}} C(r,u)E(u,s) = \sum_{u\in {\bf
m}} D(r,u) \Big( D^{-1}(u,s) - \frac{ D^{-1}(u,\ell)
D^{-1}(\ell,s)}{D^{-1}(\ell,\ell)}\Big) \\
&= \delta_{rs} - D(r,\ell)D^{-1}(\ell,s) - \big(\delta_{r\ell} -
D(r,\ell)
D^{-1}(\ell,\ell)\big) \frac{D^{-1}(\ell,s)}{D^{-1}(\ell,\ell)}.
\end{align*}
Since $r\neq \ell$ we have $\delta_{r\ell} = 0$, which gives that $(CE)(r,s) =
\delta_{rs}$ for $r,s\in{\bf m}$. A similar calculation yields
$(EC)(r,s) = \delta_{rs}$ for $r,s\in {\bf m}$ which implies that the
matrix $E$, given by formula \eqref{eq:E}, is
indeed the inverse of $C$.
Since $D$ is totally positive, for any choice of $i,j\in{\bf m}$
we know that $(-1)^{i+j} D^{-1}(i,j)\geq 0$
and $ (-1)^{i+j}D^{-1}(i,\ell) D^{-1}(\ell,j) /
D^{-1}(\ell,\ell)\geq 0$. By equation \eqref{eq:E}, this implies the inequality
\[
0\leq (-1)^{i+j} E(i,j) = (-1)^{i+j} C^{-1}(i,j) \leq
(-1)^{i+j}D^{-1}(i,j),\qquad i,j\in {\bf m}.
\]
Therefore, by induction on the cardinality of ${\bf m}$, this implies
inequality \eqref{eq:inverse_relative}.
\end{proof}
\subsection{Application of Lemma~\ref{lem:tp} to B-spline matrices}\label{sec:discuss}
Observe that the matrix $(a_{ij})$ -- satisfying $N_i^* = \sum_j a_{ij}N_j$ for
all $i$ -- is given by the inverse of the B-spline Gram
matrix $B=(\langle N_i, N_j\rangle)_{i,j=1}^n$.
This matrix is totally positive, which is
a consequence of
the fact that the kernel $N_{i}(x)$, depending on the variables $i$ and $x$,
is totally positive \cite[Theorem 4.1, Chapter 10]{Karlin1968} and
the so called basic composition formula \cite[Chapter 1, Equation
(2.5)]{Karlin1968}.
This means that we can apply Lemma~\ref{lem:tp} to
submatrices of B-spline Gram matrices.
Let $(I_{i})_i$ be an enumeration of the atoms of $\mathscr F$ for consecutive integers $i$ in the way that
if $i<j$ then $I_{i}$ is to the left of $I_{j}$.
Let $A,B$ be two atoms of $\mathscr F$. If $A=I_{i}$ and $B=I_{j}$ for two
integers $i,j$, we set $d_{\mathscr F}(A,B) = j-i$.
Denote $F_i = \operatorname{supp} N_i$,
and by $A(x)$ the atom of $\mathscr F$ containing the point
$x\in I$.
Denote
by $K_i\subset F_i$ an atom of $\mathscr F$ contained in $F_i$ having maximal
length.
Then, Theorem~\ref{thm:maintool} implies the following pointwise estimate for the dual B-spline
functions:
\begin{equation}\label{eq:dualspline}
| N_i^*(x) | \lesssim \frac{q^{ |d_{\mathscr F}(K_i, A(x))| }}{
|\conv(F_i, A(x))| },\qquad \text{$1\leq i\leq n,$ $x\in I$.}
\end{equation}
Choose an arbitrary subset ${\bf m} \subset \{1,\ldots,n\}$. Let $N_i^{{\bf m}*}$,
$i\in \bf m$ be the dual functions to $\{ N_i : i\in {\bf m} \}$.
Then, we can write
\[
N_i^{{\bf m}*} = \sum_{j\in \bf m} a_{ij}^{\bf m} N_j,
\]
where the coefficients $(a_{ij}^{\bf m})_{i,j\in \bf m}$ are given as the inverse to the matrix $(\langle
N_i,N_j\rangle)_{i,j\in\bf m}$.
Since the matrix $B = (\langle N_i, N_j\rangle)_{i,j=1}^n$ is totally
positive and invertible,
we can invoke Lemma \ref{lem:tp} to deduce from Theorem~\ref{thm:maintool} that
\begin{equation}\label{eq:geom_aijm}
|a_{ij}^{\bf m}|\lesssim \frac{ q^{|i-j|} }{|\conv (F_i,
F_j)|},\qquad i,j\in{\bf m}.
\end{equation}
Therefore, we have the following estimate for the functions $N_i^{ {\bf m}*}$ similar to \eqref{eq:dualspline}:
\begin{equation}\label{eq:dual_estimate}
|N_i^{ {\bf m}*}(x)| \lesssim \frac{ q^{ |d_{\mathscr F}( K_i, A(x) )|} }{|\conv(F_i,
A(x))|},\qquad i\in{\bf m}, x\in I.
\end{equation}
This estimate does not depend on the subset ${\bf m}$ of $\{1,\ldots,n\}$.
If ${\bf m_1} \subset {\bf m}\subset \{1,\ldots,n\}$ with $ {\bf m} \setminus {\bf m_1} =
\{i\}$, we have
$ N_{i}^{ {\bf m}*} \in \operatorname{span} \{ N_j : j\in {\bf m} \}$ and
\[
\langle N_{i}^{ {\bf m}*}, N_\ell \rangle = 0,\qquad \ell\in {\bf m_1}.
\]
This means that $N_{i}^{ {\bf m}*}$ is
orthogonal to the span of $ \{N_\ell : \ell\in \bf m_1 \}$.
We know by Lemma~\ref{lem:tp} applied to the subset $\{i\}$ of ${\bf m}$ that
\begin{equation}\label{eq:lowerestimate}
|a_{ii}^{\bf m}| \geq \frac{1}{\langle N_i, N_i\rangle} \geq
\frac{1}{|F_i|}.
\end{equation}
Using \eqref{eq:lowerestimate}
and the local stability \eqref{eq:lpstab} of B-splines, we
obtain (for $1\leq p<\infty$)
\begin{equation}\label{eq:J1}
\| N_{i}^{ {\bf m}*} \|_{L^p(K_{i})} \gtrsim |K_{i}|^{1/p} |a_{ii}^{\bf
m}| \gtrsim
|K_{i}|^{1/p
- 1}.
\end{equation}
On the other hand, by \eqref{eq:deboorlpstab} and \eqref{eq:geom_aijm},
\begin{equation}\label{eq:J2}
\int |N_{i}^{ {\bf m}*}|^p \simeq \sum_{j\in {\bf m}} |a_{ij}^{\bf m}|^p
|F_j| \lesssim
\sum_{j\in {\bf m}} \frac{ q^{ p |i-j|} }{ |\conv(F_{i},
F_{j})|^p} |F_j| \lesssim |F_{i}|^{1-p}\lesssim |K_{i}|^{1-p}.
\end{equation}
Inequalities \eqref{eq:J1} and \eqref{eq:J2} together
imply that $\|N_{i}^{ {\bf m} *}\|_p \simeq |K_{i}|^{1/p - 1}$.
\subsection{Orthogonal spline functions}\label{sec:franklin}
Let $(\mathscr F_n)_{n\geq 0}$ be an \emph{interval filtration} on an interval $I$,
which means that $(\mathscr F_n)$ is an increasing sequence of interval $\sigma$-algebras on
the interval~$I$.
Additionally, we assume that $(\mathscr F_n)$ satisfies $\mathscr F_0 = \{\emptyset,I\}$ and,
is in \emph{standard form},
meaning that for all $n\geq 1$,
$\mathscr F_n$ arises from $\mathscr F_{n-1}$ by the subdivision of exactly
one atom of $\mathscr F_{n-1}$ into two intervals $L_n$ and $R_n$ as atoms of
$\mathscr F_n$.
Then, the codimension
of $S_k(\mathscr F_{n-1})$ in $S_k(\mathscr F_n)$ is one and thus there exists a
unique (up to sign) function $f_n\in S_k(\mathscr F_n)$ that is orthonormal to $S_k(\mathscr
F_{n-1})$.
In the case $k=1$, the function $f_n$ is a constant multiple of the generalized
Haar function $|R_{n}| \ensuremath{\mathbbm 1}_{L_n} - |L_n|
\ensuremath{\mathbbm 1}_{R_n}$.
We denote $d_n(A,B) = d_{\mathscr F_n}(A,B)$ for two atoms $A,B$ of $\mathscr
F_n$ and we let $A_n(x)$ be the atom of
$\mathscr F_n$ containing the point $x\in I$.
The functions $(f_n)$ satisfy that for every $n$, there exists an atom $J_n$ of $\mathscr F_n$ with the
following properties (\cite{Passenbrunner2014}).
\begin{enumerate}
\item $|d_n(J_n, L_n)| \leq k$.
\item There exists a support $F$ of a B-spline
function in $S_k(\mathscr F_n)$ with $F\supset J_n$ and
$|F|\lesssim |J_n|$.
\item Pointwise estimate for $f_n$:
\begin{equation}\label{eq:estfranklin}
|f_n(x)| \lesssim \frac{q^{| d_n(A_n(x),J_n) | }
|J_n|^{1/2} } { |\conv(J_n, A_n(x))| },\qquad x\in I.
\end{equation}
\item $\| f_n \|_p \simeq |J_n|^{1/p - 1/2}$ for $1\leq p\leq \infty$.
\end{enumerate}
We say that $J_n$ is the \emph{characteristic interval} of the function
$f_n$.
Additionally we have the following lemma, which is also contained in
\cite{Passenbrunner2014}.
\begin{lem}\label{lem:Jint}
Let $V\subset I$ be an interval. Then, the cardinality of the set
\[
\{ n : J_n \subset V, |J_n|\geq |V|/2\}
\]
is bounded by some constant depending only on $k$.
\end{lem}
\subsection{Tensor product splines}
\label{sec:tensor}
Let $d$ be a positive integer and let for any coordinate $\delta \in
\{1,\ldots,d\}$ the sequence $(\mathscr F_n^{\delta})$ be an interval filtration
on the bounded interval $I^{\delta}$, generated by
the intervals $(I_{n,i}^{\delta})_i$. Put $I = I^1\times \cdots\times I^d$.
If $\mathscr F_n = \mathscr F_n^1\otimes \cdots\otimes \mathscr F_n^d$, then
the sequence $(\mathscr F_n)$ is called an \emph{interval filtration} on the
$d$-dimensional rectangle $I$.
Every $\sigma$-algebra $\mathscr F_n$ is generated by the finite, mutually
disjoint family
$\{ I_{n,i} : i\in \Lambda \}$, $\Lambda\subset \mathbb Z^d$, of
$d$-dimensional rectangles given by $I_{n,i} = \prod_{\ell=1}^d
I_{n,i_\delta}^\delta$ for $i\in \Lambda$.
We assume that
$\Lambda$ is of the form $\Lambda^1\times \cdots\times \Lambda^d$ where for each
$\ell=1,\ldots,d$, $\Lambda^\ell$ is a finite set of consecutive integers and the
rectangles $I_{n,i}$ have the property that they are ordered in the same way as
$\mathbb R^d$, i.e., if $i,j\in\Lambda$ with $i_\ell < j_\ell$ then the
projection of $I_{n,i}$ onto the $\ell$th coordinate axis lies to the left of the
projection of $I_{n,j}$ onto the $\ell$th coordinate axis.
For $x\in I$, let $A_n(x)$ be the uniquely determined atom (rectangle) $A\in\mathscr
F_n$ so that $x\in A$. For two atoms $A,B\in \mathscr F_n$, define
$d_n(A,B) := d_{\mathscr F_n}(A,B):= j-i\in\mathbb Z^d$ if $A=I_{n,i}$ and $B=I_{n,j}$. For
$s\in\mathbb Z^d$, we put $|s|_1 = \sum_{j=1}^d |s_j|$.
For each $\ell=1,\ldots,d$, let $k_\ell$ be a positive integer.
Define the tensor product spline space of order $k= (k_1,\ldots,k_d)$ associated to $\mathscr
F_n$ as
\[
S_k(\mathscr F_n) := S_{k_1}(\mathscr F_n^{1}) \otimes \cdots \otimes
S_{k_d}(\mathscr F_n^{d}).
\]
The space $S_k(\mathscr F_n)$ admits the tensor product B-spline basis
$(N_{n,i})_{i}$ defined by
\[
N_{n,i} = N_{n,i_1}^1\otimes \cdots \otimes N_{n,i_d}^d,
\]
where $(N_{n,i_\ell}^\ell)_{i_\ell}$ denotes the B-spline basis of
$S_{k_\ell}(\mathscr F_n^{\ell})$ that forms a partition of unity.
The support $F_{n,i}$ of
$N_{n,i}$ is composed of at most $k_1\cdots k_d$ neighouring atoms of $\mathscr
F_n$.
Consider the orthogonal projection operator $P_n$ onto $S_k(\mathscr F_n)$ with respect to
the $d$-dimensional Lebesgue measure.
A direct consequence of Shadrin's theorem \ref{thm:shadrin} and using its tensor
product structure is that $P_n$ is uniformly bounded on $L^\infty(I)$ (and
therefore also on $L^1(I)$).
Using the B-spline basis and its biorthogonal system $(N_{n,i}^*)$,
the orthogonal projector $P_n$ is given by
\begin{equation}\label{eq:rep_Pn}
P_n f =\sum_i \int_I f(x)N_{n,i}(x)\,\mathrm{d} x \cdot N_{n,i}^{*},\qquad
f\in L^1(I).
\end{equation}
In the following, the symbols $\lesssim,\gtrsim,\simeq$ are used with
the same meaning as before, but note that the dependence of the constants on the
parameter $k=(k_1,\ldots,k_d)$ also includes an implicit dependence on the
dimension $d$.
The dual B-spline functions $N_{n,i}^*$ admit the following crucial geometric decay estimate
\begin{equation}
\label{eq:mainestimate}
| N_{n,i}^*(x) | \lesssim \frac{ q^{|d_n(K_{n,i}, A_n(x))|_1} }{ |\conv
(F_{n,i},
A_n(x))|}, \qquad x\in I
\end{equation}
for some constant $q\in [0,1)$ that depends only on $k$, where $\conv(A,B)$
denotes the smallest, axis-parallel rectangle containing both sets $A,B$ and
$K_{n,i}\subset F_{n,i}$ is an atom of $\mathscr F_n$ having maximal volume.
This inequality is a consequence of Theorem~\ref{thm:maintool} and
the fact that $N_{n,i}^*$ is the
tensor product of one-dimensional dual B-spline functions.
Inserting this estimate in formula \eqref{eq:rep_Pn} for $P_n f$ and as $F_{n,i}$
consists of
at most $k_1\cdots k_d$ neighbouring atoms of $\mathscr F_n$, setting $C_k :=
C(k_1\cdots k_d) q^{-|k|_1}$, we get the
pointwise estimate
\begin{equation}
\label{eq:estPn}
|P_n f(x)| \lesssim \sum_{A \text{ atom of }\mathscr F_n}
\frac{q^{|d_n(A,A_n(x))|_1}}{|\conv(A , A_n(x))|}\int_A
|f(t)|\,\mathrm{d} t,\qquad
f\in L^1(I).
\end{equation}
Introducing the maximal function
\begin{equation}\label{eq:max_fct}
\mathscr M f(x) = \sup_{n} \sum_{A\text{ atom of }\mathscr F_n}
\frac{\rho^{|d_n(A,A_n(x))|_1}}{|\conv(A , A_n(x))|}\int_A
|f(t)|\,\mathrm{d} t,
\qquad x\in I, f\in L^1(I)
\end{equation}
for some fixed parameter $\rho\in[0,1)$,
we have the following Theorem \cite{Passenbrunner2021}.
\begin{thm}\label{prop:maximal}
The maximal function $\mathscr M$ is of weak type
(1,1), i.e. there exists a constant $C$ depending only on the dimension
$d$ and on the parameter $\rho<1$, so that we
have the inequality
\[
| \{ \mathscr M f > \lambda \}| \leq \frac{C}{\lambda}\|
f\|_{L^1},\qquad \lambda>0,\ f\in L^1(I).
\]
\end{thm}
This theorem and estimate \eqref{eq:estPn} imply that for any $f\in L^1$, the
sequence of orthogonal projections $P_n f$ on the spline spaces $S_k(\mathscr
F_n)$ of the function $f$ converges almost everywhere with respect to
$d$-dimensional Lebesgue measure (see also \cite{Passenbrunner2021}).
\section{Orthonormal tensor spline functions}\label{sec:tensor_orth}
\label{sec:tensor_def}
In this section, we construct orthonormal spline functions based on a given
interval filtration $(\mathscr F_n)_{n\geq 0}$ on a $d$-dimensional rectangle
$I$ and based
on the parameter $k = (k_1,\ldots,k_d)$ containing the orders $k_\delta$ of the
polynomials in direction $\delta$.
For $n\geq
1$ assume that $\mathscr F_{n}$ is of the form that $\mathscr
F_{n}=\mathscr F_{n}^{1} \otimes \cdots \otimes \mathscr F_{n}^d$ arises
from $\mathscr F_{n-1} = \mathscr F_{n-1}^{1} \otimes \cdots \otimes \mathscr
F_{n-1}^d$
in the way that there exists a coordinate $\delta_0\in \{1,\ldots,d\}$ so that
$\mathscr F_{n}^\delta = \mathscr F_{n-1}^\delta$ for $\delta\neq \delta_0$ and
$\mathscr F_{n}^{\delta_0}$
arises from $\mathscr F_{n-1}^{\delta_0}$ by splitting exactly one atom of $\mathscr
F_{n-1}^{\delta_0}$ into two atoms of $\mathscr F_{n}^{\delta_0}$.
If the sequence $(\mathscr F_n)$ of interval $\sigma$-algebras on a
$d$-dimensional rectangle $I$ has these properties, we say that $(\mathscr F_n)$ is
of \emph{standard form}.
Additionally, assume that $\mathscr F_0=\{\emptyset,I\}$.
Let $(f_{0,m})_{m=1}^{M_0}$ be an orthonormal basis of the space of polynomials $S_k(\mathscr
F_0)$ on $I$ that are of order $k_\delta$ in direction $1\leq \delta\leq d$.
Fix the index $n\geq 1$.
In the following construction of orthonormal spline functions
$(f_{n,m})_{m=1}^{M_n}$
corresponding to the $\sigma$-algebra $\mathscr F_{n}$, we assume
without restriction that $\delta_0=1$. For other values of $\delta_0$, the
construction proceeds similarly with obvious modifications.
Let $f$ be the $L^2$-normalized function that is
contained in the space $S_{k_1}(\mathscr F_{n}^1)$ and
orthogonal to $S_{k_1}(\mathscr F_{n-1}^1)$ and let $J$ be its corresponding
characteristic interval (cf. Section \ref{sec:franklin}).
For $j\geq 2$, let $(N_{n,i}^j)_{i\in\Lambda_n^j}$ be the B-spline basis of $S_{k_j}(\mathscr
F_n^j)$.
We successively define index sets $\Omega_m^j\subset \Lambda_n^j$, functions $f_{n,m} = f\otimes
D_m^{2}\otimes\cdots\otimes D_m^{d}$ and corresponding characteristic intervals
$J_{n,m}^j$ for $j\geq 2$ and $1\leq m\leq \prod_{j=2}^d
|\Lambda_n^j|=:M_n$, which will be given by the following
inductive procedure on $m$.
Let $\pi$ be an arbitrary permutation of the set
$\{2,\ldots,d\}$, which is allowed to depend on the value of $n$.
If $m=1$, and for $j\geq 2$, we choose $\nu_j \in \Lambda_n^j$ arbitrarily and
set $\Omega_1^j = \{ \nu_j\}$.
Define
\[
D_1^j = \frac{N_{n,\nu_j}^j }{\| N_{n,\nu_j}^j \|_2 },\qquad j\geq 2,
\]
and let $J_{n,1}^j$ be an atom of $\mathscr F_n^j$ contained in the
support of $N_{n,\nu_j}^j$ having maximal length.
Assume that $m\in \{2,\ldots, M_n\}$ and that $\Omega_\ell^j, D_\ell^j,
J_{n,\ell}^j$ are
defined for $\ell < m$ and $j\geq 2$.
Let $j_0=\max\{ \pi(j) : \Omega_{m-1}^j \subsetneq \Lambda_n^j \}$.
Then we distinguish three cases for the parameter $j$ according to the value of $\pi(j)$.
\begin{enumerate}
\item $\pi(j)= j_0$:
Choose $\mu_{j} \in \Lambda_n^{j}\setminus \Omega_{m-1}^{j}$ arbitrarily and
set $\Omega_{m}^{j} = \Omega_{m-1}^{j} \cup \{ \mu_{j} \}$. Let
$D_{m}^{j}$ be the $L^2$-normalized function contained in $\operatorname{span} \{
N_{n,\ell}^{j} :
\ell\in \Omega_{m}^{j} \}$ that is orthogonal to $\operatorname{span} \{
N_{n,\ell}^{j} : \ell\in\Omega_{m-1}^{j}\}$. This function
is uniquely given up to sign.
Let $J_{n,m}^j$ be an atom of $\mathscr F_n^j$
contained in the support of $N_{n,\mu_j}^j$ with maximal
length.
\item $\pi(j)>j_0$: we choose $\mu_j \in \Lambda_n^j$
arbitrarily, set
$\Omega_{m}^j = \{\mu_j\}$ and $D_{m}^j = N_{n,\mu_j}^j /
\| N_{n,\mu_j}^j \|_2$.
Let $J_{n,m}^j$ be a largest atom of $\mathscr F_{n}^j$
contained in the support of $N_{n,\mu_j}^j$.
\item $\pi(j)<j_0$: we put $\Omega_{m}^j = \Omega_{m-1}^j$, $D_{m}^j =
D_{m-1}^j$ and $J_{n,m}^j = J_{n,m-1}^j$.
\end{enumerate}
Then, for any $m\in \{1,\ldots,M_n\}$, we define the \emph{characteristic
interval} $J_{n,m}$ of $f_{n,m}=f\otimes
D_m^{2}\otimes\cdots\otimes D_m^{d}$ by
\[
J_{n,m} = J \times J_{n,m}^{2} \cdots \times J_{n,m}^d,
\]
which is an atom of $\mathscr F_n$. By construction, each atom of $\mathscr F_n$
appears at most $k_2\cdots k_d$ times among the sets $J_{n,m}$ for $1\leq m\leq
M_n$.
By the discussion in Section \ref{sec:discuss}, we know that for any $m$ and
$\pi(j)=j_0$, the function $D_{m}^j$ is a renormalization of the
function $N_{\nu_j}^{{\bf m}*}$ with ${\bf m} = \Omega_{m}^j$ and $\nu_j$ being the
only element in the set $\Omega_{m}^j \setminus \Omega_{m-1}^j$ (with the
understanding that $\Omega_0^j = \emptyset$).
Combining the estimates \eqref{eq:dual_estimate}, \eqref{eq:J1}, \eqref{eq:J2},
and \eqref{eq:estfranklin}, we obtain the following pointwise estimate for the
functions $f_{n,m}$:
\begin{equation}\label{eq:pw}
| f_{n,m}(x) | \lesssim \frac{ q^{|d_n(A_n(x), J_{n,m})|_1}
|J_{n,m}|^{1/2}}{|\conv(J_{n,m}, A_n(x)) |}, \qquad x\in I.
\end{equation}
The same estimates imply $\|f_{n,m}\|_p \simeq |J_{n,m}|^{1/p - 1/2}$ for $1\leq
p\leq\infty$.
By construction, the functions $(f_{n,m})$ are orthonormal in $L^2(I)$.
For arbitrary $n\geq 0$ and $0\leq m\leq M_n$, we denote by $P_{n,m}$ the orthogonal projection operator onto the span of
\[
S_k(\mathscr F_{n-1}) \cup \{ f_{n,\mu} : 1\leq \mu\leq m \},
\]
which we decompose as
\begin{equation}\label{eq:Pnm}
P_{n,m} f(x) = P_{n,0}f(x) + \sum_{\mu = 1}^m \langle f,
f_{n,\mu}\rangle f_{n,\mu}.
\end{equation}
We now show that $P_{n,m}$ is uniformly bounded on $L^1$ by a constant
depending only on $k$ (and therefore also on $d$). Indeed, the
operator $P_{n,0}$ equals the operator $P_{n-1}$ in
Section~\ref{sec:tensor}, for which we already know the uniform
$L^1$-boundedness. This means, in order to estimate $\|P_{n,m} : L^1\to
L^1\|$, we only have to estimate
\[
\Big\| \sum_{\mu\leq m} \langle f,f_{n,\mu}\rangle f_{n,\mu}\Big\|_1
\lesssim \sum_{\text{$A$ atom of $\mathscr F_n$}} \Big(\sum_{\mu\leq m}
\frac{ |J_{n,\mu}| q^{|d_n(J_{n,\mu},A)|_1}} {|\conv(J_{n,\mu}, A)|}\Big)
\int_A |f(t)|\,\mathrm{d} t,
\]
which follows from \eqref{eq:pw}.
Since each atom of $\mathscr F_n$ occurs at most $k_1\cdots k_d$ times among the
sets $J_{n,\mu}$, $1\leq \mu\leq m$, the latter sum over $\mu$ is bounded by a
constant depending only on $k$, which already gives the claimed
uniform boundedness of $\|P_{n,m} : L^1 \to L^1\|$.
Estimate \eqref{eq:pw} also yields the following pointwise bound for $P_{n,m}f$
by the maximal function $\mathscr Mf$ introduced in \eqref{eq:max_fct}.
\begin{prop}\label{prop:proj_max}
For any $n\geq 0$ and any $m\in \{1,\ldots,M_n\}$, we have the
inequality
\[
|P_{n,m}f(x)| \lesssim \mathscr Mf(x),\qquad x\in I,
\]
for $\rho = q^{1/2}$ (with $q$ as in \eqref{eq:pw}) in the definition \eqref{eq:max_fct} of $\mathscr
M$.
\end{prop}
\begin{proof}
Since we already know the desired bound for $P_{n,0}f(x)$ by
\eqref{eq:estPn}, it suffices
to consider the second term in equation \eqref{eq:Pnm}. Using estimate \eqref{eq:pw}, we obtain
\begin{equation}
\label{eq:mf}
\begin{aligned}
\Big|\sum_{\mu\leq m} \langle f,f_{n,\mu}\rangle
f_{n,\mu}(x)\Big| \lesssim
\sum_{\text{ $A$ atom of $\mathscr F_n$ }}
\Big(\sum_{\mu\leq m}\frac{q^{|d_n(J_{n,\mu}, A_n(x))|_1 +
|d_n(J_{n,\mu},A)|_1
}|J_{n,\mu}|}{|\conv(J_{n,\mu},A) |\cdot |\conv(J_{n,\mu}, A_n(x)) |} \Big)
\int_A |f(t)|\,\mathrm{d} t.
\end{aligned}
\end{equation}
Define $\rho = q^{1/2}$. For any atom $A$ of $\mathscr F_n$ and any
index $\mu\leq m$ we
have the inequalities $|d_n(A_n(x), A)|_1 \leq |d_n(J_{n,\mu}, A_n(x))|_1 +
|d_n(J_{n,\mu},A)|_1$. Moreover, since for any coordinate $\delta$, we
have $|\conv(A_n^\delta(x), A^\delta)| \leq |\conv(J_{n,\mu}^\delta,
A_n^{\delta}(x))| + |\conv(J_{n,\mu}^\delta, A^\delta)|$, we also have
the inequality
\begin{align*}
\frac{|J_{n,\mu}|}{|\conv(J_{n,\mu},A) |\cdot |\conv(J_{n,\mu},
A_n(x)) |}
\leq 2^d\frac{1}{|\conv(A_n(x), A)|}.
\end{align*}
Inserting this in \eqref{eq:mf}, we obtain
\begin{align*}
\Big|\sum_{\mu\leq m} \langle f,f_{n,\mu}\rangle
f_{n,\mu}(x)\Big| &\lesssim \sum_{\text{$A$ atom of $\mathscr F_n$}}
\frac{\rho^{|d_n(A_n(x), A)|_1}}{|\conv(A_n(x), A)|} \Big(
\sum_{\mu\leq m} \rho^{|d_n (J_{n,\mu}, A)|_1}\Big)
\int_A |f(t)|\,\mathrm{d} t \\
&\lesssim \mathscr Mf(x),
\end{align*}
since each atom of $\mathscr F_n$ only occurs at most $k_1\cdots k_d$
times among the sets $J_{n,\mu}$ for $1\leq \mu\leq m$.
\end{proof}
Combining this pointwise inequality for $P_{n,m}f$ and the weak type estimate
for the maximal function $\mathscr M f$ contained in Theorem \ref{prop:maximal}
yields -- as in \cite{Passenbrunner2021} -- that for every $f\in L^1(I)$, the series $\sum_{\ell} \langle
f,f_\ell\rangle f_\ell$
converges almost everywhere with respect to $d$-dimensional Lebesgue measure, if we use the
rearrangement $(f_\ell)$ of the functions $(f_{n,m})$ described in the following.
\subsection{Rearrangement of the functions $f_{n,m}$}\label{sec:rearr}
To each function $f_{n,m}$ we associate the $\sigma$-algebra $\mathscr F_{n}$
(so that $J_{n,m}$ is an atom of $\mathscr F_n$).
Now we enumerate the functions $(f_{n,m})$ as $(f_\ell)_{\ell\geq 0}$ according
to the lexicographic ordering on the pairs $(n,m)$. If $f_{n,m} = f_{\ell}$ for
some indices $n,m,\ell$, then we define the associated $\sigma$-algebra
$\mathscr A_\ell$ to the function $f_\ell$ by $\mathscr A_\ell = \mathscr
F_{n}$ and also we define the characteristic interval $J_\ell = J_{n,m}$
corresponding to the function $f_\ell$.
Observe that -- as opposed to the situation for $(\mathscr F_n)$ in standard
form -- two different values $\ell_1,\ell_2$ can give
$\mathscr A_{\ell_1} = \mathscr A_{\ell_2}$ by this definition.
\subsection{(Quasi-)Dyadic extension of interval
$\sigma$-algebras}\label{sec:dyadic}
Let $\mathscr F$ be an interval $\sigma$-algebra on a one-dimensional interval and let $(A_j)_{j=1}^m$ be an
enumeration of the intervals that are atoms of $\mathscr F$. For each $j=1,\ldots,m$, let $A_j =
L_j\cup R_j$ be a decomposition of $A_j$ into two disjoint intervals.
Define
\begin{equation}\label{eq:dyadic}
\mathscr D_{1,\ell}(\mathscr F) = \sigma(\mathscr F, \{ L_j, R_j :
1\leq j\leq \ell\}),\qquad \ell=1,\ldots,m
\end{equation}
to be the $\sigma$-algebra generated by $\mathscr F$ and by the splitting of the
first $\ell$ atoms $A_j$ into the two intervals $L_j, R_j$.
For an integer $\nu\geq 1$, we define
inductively
\[
\mathscr D_{\nu+1,\ell}(\mathscr F) = \mathscr D_{1,\ell} \big( \mathscr
D_{\nu, 2^{\nu-1}m}(\mathscr F)\big),\qquad \ell = 1,\ldots,2^\nu m.
\]
Let $\mathscr A = \mathscr A^{1}\otimes \cdots \otimes \mathscr A^{d}$ be an
interval $\sigma$-algebra on a $d$-dimensional rectangle.
A \emph{quasi-dyadic extension} of $\mathscr A$ consists of an
interval filtration $(\mathscr A_n)_{n\geq 0}$ in standard form with $\mathscr
A_0 = \mathscr A$ so that each $\mathscr
A_n$ equals
\[
\mathscr D_{\nu_1,\ell_1} (\mathscr A^{1}) \otimes \cdots \otimes
\mathscr D_{\nu_d,\ell_d} (\mathscr A^{d})
\]
for some $(\nu_i,\ell_i)$
satisfying $|\nu_i - \nu_j| \leq 1$ for all $i,j=1,\ldots,d$.
If the intervals $L_j$ and $R_j$ in \eqref{eq:dyadic} are chosen so that $|L_j|
= |R_j| = |A_j|/2$, a quasi-dyadic extension $(\mathscr A_n)$ of $\mathscr A$
will be called a \emph{dyadic extension} of $\mathscr A$.
\section{Unconditional convergence of multivariate orthogonal spline series}\label{sec:main}
Let $\mathscr F$ be an interval $\sigma$-algebra on a bounded interval $U$. Let
$(N_i)$ be the B-spline basis of $S_r(\mathscr F)$ for some positive integer
$r$. We say that a subset $A\subset U$ is a \emph{B-spline support of order $r$
in $\mathscr F$}, if $A$ is the support of one of the B-spline
functions $N_i$. Observe that $A$ is a union of at most $r$ neighbouring atoms of $\mathscr
F$.
\begin{defin}(Regularity)\label{def:regularity}
Let $(\mathscr F_n)$ be an interval filtration on $I$ with
$\mathscr F_n = \mathscr F_n^1 \otimes \cdots \otimes \mathscr F_n^d$ and let
$r=(r_1,\ldots,r_d)$ be a $d$-tuple of positive integers.
\begin{enumerate}
\item
We say that $(\mathscr F_n)$ is
\emph{$r$-regular
with parameter $\gamma$}, if for all $\delta\in \{1,\ldots,d\}$, for all $n$ and
for any two B-spline supports $A,B$ of order $r_\delta$ in $\mathscr F_n^\delta$
with vanishing Euclidean distance, we have
\[
\max \Big(\frac{|A|}{|B|}, \frac{|B|}{|A|} \Big) \leq \gamma.
\]
\item
We say that $(\mathscr F_n)$
is \emph{direction
$r$-regular with parameter $\beta$} if,
for all $\delta\in
\{1,\ldots,d\}$,
for all strictly decreasing sequences $(A_j)_{j=1}^\beta$ of atoms in
some $\mathscr F_{n_j}$, respectively,
and for all B-spline supports
$B$ of order $r_\delta$ in $\mathscr F_{n_1}^{\delta}$ with
$A_{1}^{\delta}\subset B$, the set $B$ is not a B-spline support of order
$r_\delta$ in $\mathscr F_{n_\beta}^{\delta}$.
\end{enumerate}
\end{defin}
\begin{remark}We make a few comments on the definition of regularity and
direction regularity.
\begin{enumerate}
\item If $d=1$, for any interval filtration $(\mathscr
F_n)$ and any choice of positive integer $r$, the filtration
$(\mathscr F_n)$ is direction $r$-regular with parameter $r+1$.
\item It is easily seen that if $(\mathscr F_n)$ is $r$-regular with
parameter $\gamma$, then, for every $m$ with $m_i\geq r_i$ for
every $i\in \{1,\ldots,d\}$, the filtration $(\mathscr F_n)$ is
$m$-regular for some parameter $\gamma'$. The same statement
holds for direction regularity instead of regularity.
\item Observe that regularity and direction regularity do not
assume any condition on the relative sidelengths of atoms of
$\mathscr F_n$, we only have regularity in every fixed
direction and direction regularity which basically says that we are not
allowed to refine too often while neglecting a particular direction.
\item If $(\mathscr F_n)$ is a quasidyadic interval filtration (meaning
that $(\mathscr F_n)$ is a quasidyadic extension of
the trivial $\sigma$-algebra on a $d$-dimensional rectangle), the sequence $(\mathscr
F_n)$ is direction $(1,\ldots,1)$-regular for some parameter
$\beta$.
This observation also allows us to give examples of interval
filtrations that are direction $(1,\ldots,1)$-regular, but not $r$-regular
for any choice of $d$-tuples of integers~$r$.
\item The notions of regularity and direction regularity are invariant
under the rearrangement $(\mathscr A_n)$ of $(\mathscr F_n)$ described
in Section~\ref{sec:rearr}.
\end{enumerate}
\end{remark}
Given an interval filtration $(\mathscr F_n)_{n\geq 0}$ in standard form on a $d$-dimensional rectangle $I$ with $\mathscr F_0 = \{\emptyset, I\}$
and a parameter
$k=(k_1,\ldots,k_d)$, we let $(f_n)$ be the sequence of orthonormal spline
functions constructed in Section~\ref{sec:tensor_def} in the order described in
Section~\ref{sec:rearr}.
\begin{thm}\label{thm:main}
Let $k=(k_1,\ldots,k_d)$ be a tuple of positive integers.
Let $(\mathscr F_n)$ be an interval filtration in standard form on a $d$-dimensional
rectangle $I$ (with $\mathscr F_0 = \{\emptyset, I\}$) that is $k$-regular with parameter $\gamma$ and direction $k$-regular with parameter
$\beta$.
Then, for all $f\in L^1$ and all signs $(\varepsilon_n)$,
\[
\Big| \Big\{ \sup_M \big| \sum_{n\leq M} \varepsilon_n \langle
f,f_n\rangle f_n \big| >
\lambda \Big\}\Big| \lesssim_{\gamma,\beta} \frac{\|f\|_1}{\lambda},\qquad \lambda>0.
\]
\end{thm}
As a corollary we obtain, using the Marcinkiewicz interpolation theorem, that
the orthonormal system $(f_n)$ is an unconditional basic sequence in $L^p$ for $1<p<\infty$ under
the conditions on $(\mathscr F_n)$ stated in Theorem~\ref{thm:main}.
\subsection{Analysis of direction regularity}
\begin{lem}\label{lem:dir}Let $(\mathscr F_n)$ be an interval filtration
on a $d$-dimensional rectangle $I$. Let $(\mathscr F_n)$ be $k$-regular
with parameter $\gamma$ and
direction $m$-regular with parameter $\beta$.
For any direction $i=1,\ldots,d$, if $m_i > k_i$ then
$(\mathscr F_n)$ is direction $m'$-regular for some parameter $\beta'$
depending only on $k_i,\gamma$ and $\beta$ where $m' = m - e_i$ and
$e_i=(0,\ldots,0,1,0,\ldots,0)$ is the canonical unit vector in
direction $i$.
\end{lem}
\begin{proof}
Let $i\in \{1,\ldots,d\}$ and $m$, $m'$ be as in the assumptions. We want to prove direction
$m'$-regularity of $(\mathscr F_n)$ for some parameter $\beta'$. Fix $\delta\in
\{1,\ldots,d\}$. If $\delta\neq i$, the condition for direction $m'$-regularity follows
directly from direction $m$-regularity. Thus, assume $\delta=i$ and assume that $(\mathscr
F_n)$ is not direction $m'$-regular with some parameter $C$. This means that there
exists an increasing sequence $n_1 < \cdots < n_C$ of indices and a strictly
decreasing sequence of sets $(A_j)_{j=1}^C$ so that $A_j$ is an atom in
$\mathscr F_{n_j}$ for all $j=1,\ldots,C$ and there exists a B-spline support $B\supset
A_1^\delta$ in $\mathscr F_{n_1}^{\delta}$ of order $m_\delta-1$ that is still a B-spline
support of order $m_\delta-1$ in $\mathscr F_{n_C}^\delta$.
Define $\mathscr G_j = \mathscr F_{n_j}^\delta$ for $j=1,\ldots,C$.
Let $B_1$ be a B-spline support of order $m_\delta$ in $\mathscr G_1$
with $B_1 \supset B\supset A_1^\delta$. Since $(\mathscr F_n)$ is direction
$m$-regular with parameter $\beta$, we know that $B_1$ is not a B-spline support
of order $m_\delta$ in $\mathscr G_\beta$.
Let $B_2\supset B\supset A_1^\delta$ be a B-spline support of order $m_\delta$
in $\mathscr G_{\beta}$ with $B_2 \subset B_1$. Then we know that $B_2$ is not
a B-spline support of order $m_\delta$ in $\mathscr G_{2\beta}$.
Inductively, we get a strictly decreasing sequence $(B_\ell)$ so that $B_{\ell}$
is a B-spline support of order $m_\delta$ in $\mathscr G_{(\ell-1)\beta}$, but not
in $\mathscr G_{\ell\beta}$ and $B_\ell \supset
B\supset A_1^\delta$ for all $\ell$.
Defining $D_\ell = B_\ell \setminus B$, we know that $D_{\ell}\setminus
D_{\ell+k_\delta}$ contains
a B-spline support of order $k_\delta$ in the $\sigma$-algebra $\mathscr
G_{(\ell+k_\delta)\beta}$.
By $k_\delta$-regularity of $(\mathscr G_j)$ in direction $\delta$,
\begin{equation}\label{eq:D}
|B|\leq |B_{\ell + k_\delta}|\leq \gamma | D_\ell \setminus
D_{\ell+k_\delta} | = \gamma |D_\ell| - \gamma
|D_{\ell+k_\delta}|\leq \gamma|D_\ell|,
\end{equation}
Using the estimate $|D_{\ell+k_\delta}|\leq |B_{\ell+k_\delta}|$, this also gives
\begin{equation}\label{eq:geom}
|D_{\ell+k_\delta}|\leq \frac{\gamma}{1+\gamma}|D_\ell|.
\end{equation}
By $k_\delta$-regularity (since $B$ contains a
B-spline support of order $k_\delta\leq m_\delta-1$ in every $\mathscr
G_j$) we have
$|D_1|\leq \gamma |B|$ and $|B|\leq \gamma |D_\ell|$ by \eqref{eq:D}.
Those
inequalities, together with the geometric decay \eqref{eq:geom}, are only possible
if $\ell$ is bounded in terms of $\gamma$ and $k_\delta$, which implies
an upper bound of $C$ in terms of $\gamma, k_\delta$ and $\beta$.
\end{proof}
By induction, this implies that if a $k$-regular filtration $(\mathscr F_n)$ is
direction $m$-regular with $m_i \geq k_i$ for all directions
$i\in\{1,\ldots,d\}$, we obtain that $(\mathscr F_n)$ is also direction
$k$-regular with different constants.
\begin{example}\label{ex:reg}
On the other hand, we now discuss the possibility of a filtration $(\mathscr
F_n)$ that is $k$-regular and direction $k$-regular, but not direction
$m$-regular with $m_i\leq k_i$ for all $i\in\{1,\ldots,d\}$ and $m_\delta <
k_\delta$ for at least one direction $\delta$.
Fix $m_\delta=k_\delta-1$ and let $\varepsilon>0$ arbitrary.
If the $\sigma$-algebra $\mathscr F(\varepsilon)$ on the interval $[-1,1)$
is generated by the intervals $[-1, - \varepsilon),
[\varepsilon, 1)$ and the $m_\delta$ intervals $[- \varepsilon +
2j\varepsilon/m_\delta, - \varepsilon +
2(j+1)\varepsilon/m_\delta)$ for $j=0,\ldots,m_\delta-1$, the $k_\delta$-regularity
parameter of the $\sigma$-algebra $\mathscr F(\varepsilon)$ is smaller than $2$
and we can refine the two intervals $[-1,-\varepsilon)$ and
$[\varepsilon,1)$ in $\mathscr F(\varepsilon)$ a number
of at least $|\log \varepsilon|$ times without increasing
the bound $2$ for the $k_\delta$-regularity
parameter.
This implies that we can give examples of
interval filtrations $(\mathscr F_n)$ (that are
$k$-regular and direction $k$-regular) on
$d$-dimensional rectangles that are not
$m$-regular and not
$m$-direction regular by
using the $\sigma$-algebras $\mathscr F(1/\ell)$
and its refinements described above for all
positive integers $\ell$ in the construction of
the filtration $(\mathscr F_n^\delta)$.
\end{example}
Example~\ref{ex:reg} and Lemma~\ref{lem:dir} explain
the choice of the same order $k$ for
regularity and direction regularity in the
formulation of Theorem~\ref{thm:main}.
If $(\mathscr F_n)$ is an interval filtration on
a $d$-dimensional rectangle in standard form
with $\mathscr F_0$ being the trivial
$\sigma$-algebra, then we denote by $(\mathscr
A_n)$ and $(J_n)$ its rearrangement and the
sequence of characteristic intervals described
in Section~\ref{sec:rearr}, respectively.
\begin{lem}\label{lem:comb}
Let $(\mathscr F_n)$ be an interval filtration on a $d$-dimensional
rectangle $I$ in standard form (with $\mathscr F_0 = \{\emptyset, I\}$)
that is $k$-regular
with parameter $\gamma$ and direction
$k$-regular with parameter $\beta$.
Let $(C_{n})_{n\in\Lambda}$ be a decreasing sequence of sets with
the following properties.
\begin{enumerate}
\item $C_{n}$ is an atom of $\mathscr A_{n}$ for all
$n\in\Lambda$.
\item There exists $s\in\mathbb Z^d$ so that $d_{n}(C_{n},
A_{n}) := d_{\mathscr
A_{n}}(C_{n},
J_{n}) = s$ for all $n\in\Lambda$.
\item There exists a direction $\delta\in \{1,\ldots,d\}$ so
that
\begin{enumerate}
\item $J_{n}^{\delta} \subseteq
J_{m}^{\delta}$ for all $n,m\in\Lambda$
with $n\geq m$,
\item denoting $n_1=\min\Lambda$, there are at least $k_\delta$ atoms of
$\mathscr A_{n_1}^\delta$ between the
sets $C_{n_1}^\delta$ and
$J_{n_1}^\delta$.
\end{enumerate}
\end{enumerate}
Then, the cardinality $\card\Lambda$ of $\Lambda$ admits the bound
\[
\card\Lambda \lesssim_{\gamma,\beta} \sum_{j\neq \delta}(1+ |s_j|).
\]
\end{lem}
\begin{proof}
Assume without restriction that the sequence of
$\sigma$-algebras $(\mathscr A_{n})_{n\in\Lambda}$ is strictly increasing.
This can be done, since we know that $(C_{n})$ is decreasing, which
means that if $n_i < \cdots <n_{i+m}$ with
$n_i,\ldots,n_{i+m}\in\Lambda$ and $\mathscr A_{n_i} = \cdots
=\mathscr A_{n_{i+m}}$ we know that $C_{n_{i+m}} = \cdots =C_{n_i}$. But
also $d_{n_{i+\ell}}(J_{n_{i+\ell}}, C_{n_{i+\ell}}) = s$ is constant
for all $\ell =1,\ldots,m$ which implies that $J_{n_{i+m}} = \cdots =
J_{n_i}$. By construction of the intervals $J_n$ in
Section~\ref{sec:tensor_def}, we get that $m\leq
k_1\cdots k_d$.
Using the notation $d_{n}^\delta = d_{ \mathscr A_{n}^\delta}$, we obtain (by
(2) and (3b))
\[
|s_\delta| = |d_{n}^\delta( C_{n}^\delta, J_{n}^\delta)| \geq
k_\delta+1,\qquad
n\in \Lambda,
\]
which means that there exists a set $\Delta$ between $C_{n_1}^\delta$ and
$J_{n_1}^\delta$ that is a B-spline support of order
$k_\delta$ in all $\sigma$-algebras $\mathscr A_{n}^{\delta}$ for $n\in\Lambda$
and so that the Euclidean distance between $\Delta$ and $J_{n_1}^\delta$ is zero.
This also implies that for all $n\in\Lambda$, the Euclidean distance between
$\Delta$ and $J_{n}^\delta$ is zero.
Since $J_{n}^\delta$ is the largest
atom in $\mathscr A_n^\delta$ contained in some B-spline support of order
$k_\delta$ in $\mathscr A_{n}^\delta$, we know that, by
$k_\delta$-regularity of the
filtration $(\mathscr A_{n}^\delta)$,
\begin{equation}\label{eq:Jdelta}
(k_\delta\gamma)^{-1} |\Delta| \leq| J_{n}^\delta | \leq \gamma
|\Delta|,\qquad n\in\Lambda.
\end{equation}
We split the index set $\Lambda$ into the (not mutually disjoint)
subsets
\[
\Gamma_i = \{ n\in\Lambda : \mathscr A_{n}^i \neq \mathscr
A_{m}^i \text{ for all $m\in\Lambda, m<n$} \},\qquad
i\in\{1,\ldots,d\}.
\]
Every $n\in\Lambda$ is contained in some set $\Gamma_i$ since we assumed
that $(\mathscr A_n)_{n\in\Lambda}$ is strictly increasing.
Additionally, $n_1 = \min\Lambda\in \Gamma_i$ for every
$i\in\{1,\ldots,d\}$.
First we count the indices in the set $\Gamma_\delta$.
By Lemma~\ref{lem:Jint}, inequality \eqref{eq:Jdelta} is only possible a
constant number of times $c_\delta$ depending
on $k_\delta,\gamma$, which implies
$|\Gamma_\delta| \leq c_\delta$.
Next, let $i=1,\ldots,d$ with $i\neq \delta$ arbitrary.
For $n,m\in \Gamma_i$ with $m<n$ we assume that either $J_{n}^i$ is a strict
subset of $J_{m}^i$ or $J_{n}^i \cap J_{m}^i = \emptyset$. This can be
done without restriction since the case $J_{n}^i = J_{m}^i$, by Lemma~\ref{lem:Jint},
can increase the cardinality of $\Gamma_i$ only by a factor depending on $k_i$.
In both cases, there exists an atom of $\mathscr A_{m}^i$ that is not an atom
of $\mathscr A_{n}^i$ and is contained in the set $\conv( J_{n_1}^i,
C_{n_1}^i )$. The number of atoms of $\mathscr
A_{n_1}^i$ contained in $\conv( J_{n_1}^i,
C_{n_1}^i )$ is $1+|s_i|$. If we now assume that $|\Gamma_i| >
2^\beta(1+|s_i|)$, then there exists a strictly decreasing sequence
$(A_{n}^i)_{n\in\Omega}$
with $\Omega \subset \Gamma_i$ and $\card\Omega\geq
\beta$ so that $A_{n}^i$ is an atom of $\mathscr
A_{n}^i$ for all $n\in \Omega$. This implies that there exists a strictly decreasing
sequence of sets $(A_{n})_{n\in\Omega}$ so that
$A_n$ is an atom of $\mathscr A_n$ and $A_n^\delta \subset \Delta$ for all
$n\in\Omega$.
Since $\Delta$ is a B-spline support of order $k_\delta$ in the $\sigma$-algebra
$\mathscr A_{n}^\delta$ for every $n\in\Omega$,
by direction $k$-regularity of $(\mathscr A_n)$ with parameter $\beta$, this is
not possible. Therefore, we have the inequality $|\Gamma_i| \leq 2^\beta
(1+|s_i|)$.
Collecting the estimates above, we obtain
\[
\card \Lambda \leq \sum_{i=1}^d | \Gamma_i | \lesssim_{\gamma,\beta} \sum_{i\neq \delta}
(1+|s_i|),
\]
which is the desired estimate.
\end{proof}
\section{ Proof of Theorem~\ref{thm:main} }\label{sec:proof_main}
This section contains the proof of Theorem~\ref{thm:main}.
Therefore, fix an interval filtration $(\mathscr F_n)_{n\geq 0}$ on a
$d$-dimensional rectangle $I$ in standard
form (with $\mathscr F_0 = \{\emptyset, I\}$) that is $k$-regular with
parameter $\gamma$ and
and direction $k$-regular with parameter $\beta$ and let $(f_n)$ be the
orthonormal system described in
Section~\ref{sec:rearr}.
In order to prove Theorem~\ref{thm:main} it is enough to prove, for arbitrary
integers $N$, all sequences $(\varepsilon_n)_{n\leq N}$ of signs
and all functions $f = \sum_{n\leq N} a_n f_n$ with $T_\varepsilon f = \sum_{n\leq N}
\varepsilon_n a_n f_n$, the following inequality:
\begin{equation}\label{eq:wt11}
| \{ \sup_{M\leq N} | P_M (T_\varepsilon f) | >
\lambda \}| \lesssim_{\gamma,\beta}
\frac{\|f\|_1}{\lambda},\qquad \lambda>0,
\end{equation}
where $P_M$ denotes the orthogonal projector onto $\operatorname{span} (f_n)_{n\leq
M}$, which, by the discussion before Proposition~\ref{prop:proj_max}, is
uniformly bounded on $L^1$.
Fix the index $N$, a function $f = \sum_{n\leq N} a_n f_n$ and a
positive number $\lambda$. If we let the index $\ell$ be such that
$\mathscr F_\ell$ is associated to the function $f_N$, we assume without
restriction that $N$ is chosen sufficiently large so that the $\sigma$-algebra associated
to $f_{N+1}$ is $\mathscr F_{\ell +1}$.
We also assume without restriction that $(\mathscr F_n)$ is of the
form that $(\mathscr F_n)_{n>\ell}$ is a dyadic extension of $\mathscr
F_\ell$.
Based on this interval filtration $(\mathscr F_n)$ we consider the
associated interval filtration $(\mathscr A_n)$ defined in
Section~\ref{sec:rearr} so that the function $f_n$ is associated to the
$\sigma$-algebra $\mathscr A_n=\mathscr A_n^{1}\otimes \cdots\otimes
\mathscr A_{n}^{d}$ for every index $n$.
\subsection{A maximal function}
Let $U_{n}^j$ for $1\leq j\leq d$ be a union of $\ell_j$ neighboring atoms
$A_{n,1},\ldots, A_{n,\ell_j}$ of $\mathscr A_n^{j}$ for $1\leq
\ell_j \leq 3k_j$ so that the lengths of the leftmost atom $A_{n,1}$ and the
rightmost atom $A_{n,\ell_j}$ are
comparable to the length of $U_{n}^j$, i.e.
\begin{equation}\label{eq:bd}
\min ( |A_{n,1}|, |A_{n,\ell_j}| ) \gtrsim_\gamma |U_{n}^j|
\end{equation}
and so that $U_{n}^j$ contains at least one
B-spline support of order $k_j$ in $\mathscr A_n^{j}$.
Observe that if $S$ is a B-spline support of order $k_j$ in $\mathscr
A_n^{j}$ then there exists such a set $U_{n}^j$ with $U_{n}^j\supset S$ by the $k_j$-regularity of the
$\sigma$-algebra $\mathscr A_n^{j}$.
Let $\mathscr C_n$ be the collection of all $U_n= U_{n}^1 \times \cdots
\times U_{n}^d$ arising in this way. Let $a(N)$ be a sufficiently large integer
so that for each atom $A$ of
$\mathscr A_N$ and for every $t\in A$ there exists a set $B\in \mathscr
C_{a(N)}$ with $t\in B\subset A$. (This is possible since $ (\mathscr
F_n)_{n\geq\ell}$
is a dyadic extension of $\mathscr F_\ell$.)
Set $\mathscr C = \cup_{n\leq a(N)} \mathscr C_n$ and
define the maximal function
\[
\mathscr M_{\mathscr C}u(x) = \sup_{B\in\mathscr C, x\in B} \frac{1}{|B|} \int_B
|u(t)|\,\mathrm{d} t.
\]
For $B\in \mathscr C$, define $n(B)$ to be the smallest index $n$ so that $B\in
\mathscr C_n$.
It can be seen that $\mathscr M_{\mathscr C}u(x) \lesssim \mathscr M u(x)$ with
the maximal function $\mathscr M$ defined in \eqref{eq:max_fct}.
Indeed, let $B\in \mathscr C_n$ with $x\in B$ be arbitrary.
Then, we can divide $B$ into at most $(3k_1)\cdots (3k_d)$ atoms $V_j$ of
$\mathscr A_n$ satisfying
$|d_n (A_n(x), V_j)|_1 \leq \sum_{\delta=1}^d 3k_\delta=:C$. Since $\conv(V_j\cup
A_n(x)) \subset B$ for any $j$, we estimate
\begin{align*}
\frac{1}{|B|} \int_B |u(t)|\,\mathrm{d} t &\leq \sum_{j} \frac{1}{|\conv(V_j\cup
A_n(x))|} \int_{V_j} |u(t)|\,\mathrm{d} t \\
&\leq q^{-C}\sum_{j} \frac{q^{|d_n (V_j,
A_n(x))|_1 }}{|\conv(V_j\cup
A_n(x))|} \int_{V_j} |u(t)|\,\mathrm{d} t \lesssim \mathscr M u(x).
\end{align*}
Therefore, by Theorem~\ref{prop:maximal}, $\mathscr M_{\mathscr C}$ is of weak
type $(1,1)$ as well.
\subsection{Decomposition of $f$}
We prove inequality \eqref{eq:wt11}
by splitting the function $f$ using the maximal function $\mathscr M_{\mathscr
C}f$.
Assume that $\|f\|_1 \leq \lambda |I|/2$ since otherwise inequality
\eqref{eq:wt11} is clear.
Define $G_\lambda = \{\mathscr M_{\mathscr C}f > \lambda\}$. For $x\in
G_\lambda$, let $B(x)\in \mathscr C$ be chosen so that
$x\in B(x)$ and with
\begin{equation}\label{eq:condlambda}
\frac{1}{|B(x)|} \int_{B(x)} |f(t)|\,\mathrm{d} t > \lambda
\end{equation}
and also so that for all strictly larger sets $\mathscr C\ni\widetilde{B}\supset B(x)$,
we have the opposite inequality
\[
\frac{1}{|\widetilde{B}|}\int_{\widetilde{B}} |f(t)|\,\mathrm{d} t \leq \lambda.
\]
Note that $B(x)\neq I$ for any $x\in G_\lambda$ because of the assumption
$\|f\|_1 \leq \lambda|I|/2$.
Then, the collection of all those sets
$\{ B(x) : x\in G_\lambda\}$
covers the set $G_\lambda$. Let $(E_j)_j$ be an enumeration of the finitely many different
sets in the collection $\{ B(x) : x\in G_\lambda\}$. Those sets are not
necessarily disjoint, but we will show that
\begin{equation}\label{eq:charfun}
\sum_j \ensuremath{\mathbbm 1}_{E_j} \lesssim 1.
\end{equation}
To see this, let $t\in I$ be arbitrary and
we divide the family $\Gamma(t)= \{E_j : t\in E_j\}$ into a number of subcollections.
First, let for $1\leq \ell_j\leq 3k_j$ and $\ell = (\ell_1,\ldots,\ell_d)$
\[
\Gamma_{\ell}(t) = \{ E = E^1 \times \cdots \times E^d \in\Gamma(t) :
E^\delta \text{ consists of
$\ell_\delta$ atoms of $\mathscr A_{n(E)}^{\delta}$ for all
$\delta$} \}.
\]
Next we divide according to which atom from left to right the point $t$
belongs to. For each choice of $\ell_1,\ldots,\ell_d$, and for each $1\leq
m_\delta
\leq \ell_\delta$ for $\delta=1,\ldots,d$ we define the collection of all sets $E\in
\Gamma_{\ell}(t)$ so that the point $t$ belongs to the $m_\delta$th
atom of $E^\delta$ from left to right as $\Gamma_{\ell,m}(t)$ writing $m=
(m_1,\ldots,m_d)$.
If two sets $E,F\in\Gamma(t)$ are contained in the same collection
$\Gamma_{\ell,m}(t)$, by the nestedness of the
$\sigma$-algebras $(\mathscr A_n)$ we must have that either $E$ is contained
in $F$ or vice versa. But since the sets $E,F$ are chosen
maximally under condition \eqref{eq:condlambda}, we must have $E=F$.
Therefore, each collection $\Gamma_{\ell,m}(t)$
only consists of at most one set and since the number of
collections $\Gamma_{\ell,m}(t)$ is bounded by some constant depending on
$k=(k_1,\ldots,k_d)$, we have
proven \eqref{eq:charfun}.
Next we disjointify the collection $(E_j)$ and set
\[
V_j = E_j \setminus \bigcup_{i<j} E_i.
\]
Obviously $(V_j)$ consists of disjoint sets, $\cup_j V_j = G_\lambda$
and $V_j\subset E_j$ for each $j$.
Based upon this disjoint decomposition of $G_\lambda$, we split the function
$f$ into the following parts:
\begin{align} \label{eq:hdef}
h &= f \cdot \ensuremath{\mathbbm 1}_{G_\lambda^c} + \sum_{j}
Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j}),\\
\label{eq:gdef}
g &= f-h = \sum_{j} \big(f\ensuremath{\mathbbm 1}_{V_j} - Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})
\big),
\end{align}
where the operator $Q_{E_j}$ is given as follows.
For fixed $j\geq 1$, we have $E_j\in \mathscr C_n$ with $n= n(E_j)$.
Then, let $Q_{E_j}$ be the orthogonal projection operator
onto the spline space $S_k(\mathscr A_n \cap E_j)$.
Writing $T_\varepsilon g = \sum_{n\leq N} \varepsilon_n \langle g,f_n\rangle
f_n$ and $T_\varepsilon h = \sum_{n\leq N} \varepsilon_n \langle h,f_n\rangle
f_n$, we obtain
$T_\varepsilon f = T_\varepsilon h + T_\varepsilon g$ and thus
\begin{equation}\label{eq:splitf}
| \{ \sup_{M\leq N} |P_M( T_\varepsilon f) | > \lambda \} | \leq | \{
\sup_{M\leq N} |P_M(T_\varepsilon h)| >
\lambda/2 \} | + | \{ \sup_{M\leq N} |P_M(T_\varepsilon g)| > \lambda/2 \} |.
\end{equation}
\subsection{The function $h$} We start with the estimate
\begin{equation}\label{eq:h0}
| \{ \sup_{M\leq N} |P_M(T_\varepsilon h)| > \lambda/2 \} | \leq
\frac{4}{\lambda^2} \big\|
\sup_{M\leq N} |P_M(T_\varepsilon h)|\big\|_2^2.
\end{equation}
Since the maximal function of the operators $P_M$ is bounded uniformly on $L^2$ by
\eqref{eq:estPn} and Theorem~\ref{prop:maximal}, we estimate further
\begin{equation}\label{eq:h1}
| \{ \sup_{M\leq N} |P_M(T_\varepsilon h)| > \lambda/2 \} | \lesssim
\frac{4}{\lambda^2} \|
T_\varepsilon h\|_2^2 \leq \frac{4}{\lambda^2}\| h\|_2^2,
\end{equation}
where the last equation follows from the orthogonality of the functions $(f_n)$. Therefore, it suffices to estimate
the $L^2$-norm of $h$.
We first estimate
$|f|$ pointwise a.e. on $G_\lambda^c$. Since $f\in S_N$
we let for $t\in G_\lambda^c$
the atom $A$ in $\mathscr A_N$ be such that
$t\in A$.
By definition of $a(N)$ and $\mathscr C$, there exists $B
\in \mathscr C$ with $t\in B\subset
A$. Since $t\in G_\lambda^c$, we know that
\[
\frac{1}{|B|}\int_B |f(s)|\,\mathrm{d} s \leq \lambda.
\]
Since $f$ is a polynomial on $B\subset A$, we invoke Remez' inequality
(Corollary~\ref{cor:remez}) to deduce
$|f(t)|\leq \|f\|_{L^\infty(B)} \lesssim
\lambda$. This argument shows that $|f|\lesssim \lambda$ a.e. on $G_\lambda^c$
and allows us to estimate further
\begin{equation}
\label{eq:h}
\begin{aligned}
\|h\|_2^2 &= \int_{G_\lambda^c} |f|^2 + \int \Big | \sum_{j}
Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})\Big|^2 \lesssim \lambda \int_{G_\lambda^c} |f| + \int
\Big( \sum_{j}
|Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})|\Big)^2 \\
&\lesssim \lambda\int_{G_\lambda^c} |f| +
\sum_{j} \int | Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j}) |^2,
\end{aligned}
\end{equation}
where the last inequality follows from the fact that the non-negative function $u_j :=
|Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})|$ has support contained in $E_j$.
Indeed, let $t\in I$ and let
$j_1(t),\ldots,j_m(t)$ be an enumeration of the indices $j$ so that $t$ is
contained in the support of $u_{j}$. By
\eqref{eq:charfun}, we know that $m\lesssim 1$. Therefore,
\begin{align*}
\Big( \sum_j u_j(t) \Big)^2 = \big( u_{j_1}(t) + \cdots + u_{j_m}(t)
\big)^2
\lesssim \big( \max_\ell u_{j_\ell}(t) \big)^2 \leq \sum_{\ell}
u_{j_\ell}(t)^2 \leq \sum_j u_j(t)^2.
\end{align*}
Next, we will show that, for all $j$, we have the estimate
\begin{equation}\label{eq:localproj}
\int |Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})|^2 \lesssim_\gamma \lambda^2 |E_j|.
\end{equation}
Indeed, setting $n=n(E_j)$, let $(N_i)$ be the B-spline basis of $S_k(\mathscr A_n\cap E_j)$ and denote by
$(N_i^*)$ its dual basis.
Since the linear span of $(N_i)$ is
the range of the operator $Q_{E_j}$,
\[
Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j}) = \sum_i \langle f\ensuremath{\mathbbm 1}_{V_j}, N_i\rangle
N_i^*.
\]
By the properties of the sets $E_j\in \mathscr C_n$ (in particular by the choice
\eqref{eq:bd} of boundary intervals) and the $k$-regularity of the filtration $(\mathscr A_n)$, we get that the dual functions $N_i^*$ satisfy
the estimate
\[
|N_i^*(t)| \lesssim_\gamma \frac{1}{|E_j|},\qquad t\in E_j,
\]
by inequality \eqref{eq:mainestimate}. This implies
\begin{equation}
\label{eq:est1}
\begin{aligned}
\int_{E_j} |Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})|^2 &\lesssim \sum_i |\langle
f\ensuremath{\mathbbm 1}_{V_j}, N_i\rangle|^2 \int_{E_j} N_i^*(t)^2\,\mathrm{d} t \\
&\lesssim_\gamma \sum_i \Big(
\int_{E_j} |f| \Big)^2 \frac{1}{|E_j|} \lesssim \Big(
\int_{E_j} |f|
\Big)^2 \frac{1}{|E_j|}
\end{aligned}
\end{equation}
since the sum over $i$ only contains a constant number of terms (depending on
the order of the splines $k=(k_1,\ldots, k_d)$).
Recall $n=n(E_j)$.
Then, let $A\in\mathscr C_{n-1}$ so that $E_j\subset A$.
Observe that by definition of $n(E_j)$, $A=A^1\times \cdots \times A^d$ is a
strict superset of $E_j=E_j^1\times \cdots \times E_j^d$ in the sense that there
exists precisely one coordinate $\delta=1,\ldots,d$ so that $A^\delta$ is a
strict superset of $E_j^\delta$. This means that one of the atoms of $\mathscr
A_n^\delta$ contained in $E_j^\delta$ is not an atom in $\mathscr
A_{n-1}^\delta$. Nevertheless, $A^\delta$ is a subset of the union of a constant
(depending on $k_\delta$) number of neighbouring B-spline supports in $\mathscr
A_n^\delta$, at least one of them being a subset of $E_j^\delta$. Therefore, by
$k_\delta$-regularity of the $\sigma$-algebra $\mathscr A_n^\delta$, we obtain
$|A^\delta|\lesssim_\gamma |E_j^\delta|$ and therefore, $|A| \lesssim_\gamma |E_j|$.
By the maximality of $E_j$ under condition \eqref{eq:condlambda}, we infer
\[
\frac{1}{|E_j|} \int_{E_j} |f(t)|\,\mathrm{d} t \lesssim_\gamma \frac{1}{|A|}\int_{A}
|f(t)|\,\mathrm{d} t \leq \lambda.
\]
Therefore, we continue the estimate in \eqref{eq:est1} and write
\[
\int |Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})|^2 \lesssim_\gamma \frac{1}{|E_j|}
\Big( \int_{E_j} |f(t)|\,\mathrm{d} t \Big)^2
\lesssim_\gamma \lambda^2 |E_j|,
\]
which shows inequality \eqref{eq:localproj}.
Summing over $j$ yields
\[
\sum_{j} \int | Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j}) |^2 \lesssim_\gamma \lambda^2
|G_\lambda|\lesssim \lambda \|f\|_1
\]
by inequality \eqref{eq:charfun} and
the weak type estimate for the maximal function $\mathscr M_{\mathscr C}$. Inserting this
inequality in the estimate \eqref{eq:h} for the $L^2$ norm of the function $h$
yields
$\| h \|_2^2 \lesssim_\gamma\lambda\| f\|_1$,
which, together with \eqref{eq:h0} and \eqref{eq:h1}, gives the weak type estimate
\begin{equation}\label{eq:hfinal}
| \{ \sup_{M\leq N} |P_M(T_\varepsilon h)| > \lambda/2 \}| \lesssim_\gamma \frac{\|f\|_1}{\lambda}.
\end{equation}
\subsection{The function $g$}
Let $j$ be arbitrary and let $n=n(E_j)$ and thus $E_j= E_j^{1}\times \cdots \times
E_j^{d}\in \mathscr C_n$.
We know that for all $\delta=1,\ldots,d$, the set $E_j^{\delta}$ is a union of $\ell_\delta$ atoms in $\mathscr
A_n^{\delta}$ for some $1\leq \ell_\delta\leq 3k_\delta$.
Then,
define $L_j^{\delta}$ and $H_j^{\delta}$ to be the union of at most $5k_\delta$ and
at most $7k_\delta$ neighboring atoms of
$\mathscr A_{n}^{\delta}$ respectively so that
between $(H_j^\delta)^c$ and $L_j^\delta$ as well as between $(L_j^\delta)^c$
and $E_j^\delta$ are $k_\delta$ atoms of $\mathscr A_n^\delta$.
By $k$-regularity of $(\mathscr A_n)$, this implies that
the distance between $(L_j^{\delta})^c$
and $E_j^{\delta}$ as well as the distance between $(H_j^{\delta})^c$ and
$L_j^{\delta}$ is $\gtrsim_\gamma |E_j^{\delta}|$ and, moreover,
$|H_j^\delta| \lesssim_\gamma |E_j^\delta|$.
Then, set $L_j = L_j^{1}\times \cdots \times L_j^{d}$ and $H_j =
H_j^{1}\times \cdots \times H_j^{d}$.
The letters $L$ and $H$ are chosen here to indicate that $L_j$ and $H_j$ are
large and huge versions
of $E_j$, respectively.
Next, set $\widetilde{G_\lambda} = \cup_j H_j$. Then, we estimate the function
$g$ as follows:
\begin{equation}\label{eq:split_maximal}
\big| \big\{ \sup_{M\leq N} |P_M(T_\varepsilon g)| > \lambda/2 \big\}
\big| \leq |\widetilde{G_\lambda}| +
\big|\big\{ t\in \widetilde{G_\lambda}^c : \sup_{M\leq N} |P_M(T_\varepsilon g)(t)| > \lambda/2
\big\}\big|,
\end{equation}
The term $|\widetilde{G_\lambda}|$ can be estimated by $|G_\lambda|$ if we
use inequality \eqref{eq:charfun}:
\begin{equation}\label{eq:esttilde}
|\widetilde{G_\lambda}| \leq \sum_j |H_j| \lesssim_\gamma \sum_j |E_j| \lesssim
|G_\lambda| \lesssim \frac{ \|f\|_1}{\lambda}.
\end{equation}
where the last inequality follows from the fact that the
maximal function $\mathscr M_{\mathscr C}$ is of weak type $(1,1)$.
Now we
come to the second term of \eqref{eq:split_maximal}, which we estimate as
follows:
\begin{equation}\label{eq:weakTg}
\begin{aligned}
|\{ t\in \widetilde{G_\lambda}^c : \sup_{M\leq N} |P_M(T_\varepsilon g)(t)| > \lambda/2 \}| &\leq
\frac{2}{\lambda} \Big\| \sup_{M\leq N} |P_M(T_\varepsilon g)|
\Big\|_{L^1(\widetilde{G_\lambda}^c)} \\ &= \frac{2}{\lambda} \Big\|\sup_{M\leq
N} \big| \sum_j \sum_{n\leq M}
\varepsilon_n\langle g_j,f_n\rangle f_n \big|
\Big\|_{L^1(\widetilde{G_\lambda}^c)} \\
&\leq \frac{2}{\lambda} \sum_j \Big\| \sup_{M\leq N} \sum_{n\leq M} |\langle g_j,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(\widetilde{G_\lambda}^c)} \\
&=\frac{2}{\lambda} \sum_j \Big\| \sum_{n\leq N} |\langle g_j,f_n\rangle| \cdot
|f_n| \Big\|_{L^1(\widetilde{G_\lambda}^c)}
\end{aligned}
\end{equation}
with $g_j = f\ensuremath{\mathbbm 1}_{V_j} - Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})$.
We will, for fixed index $j$, show the inequality
\begin{equation}\label{eq:crucialtoshow}
\Big\| \sum_n |\langle g_j,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(H_j^c)} \lesssim_{\gamma,\beta} \| g_j \|_{L^1(E_j)}.
\end{equation}
If we know \eqref{eq:crucialtoshow}, we can continue estimate \eqref{eq:weakTg},
since $\widetilde{G_\lambda}^c \subset H_j^c$ for any fixed index~$j$, and
therefore,
\begin{align*}
|\{ t\in \widetilde{G_\lambda}^c : \sup_{M\leq N} |P_M(T_\varepsilon g)(t)| > \lambda/2 \}|
\lesssim_{\gamma,\beta}
\frac{1}{\lambda} \sum_j \| g_j \|_{L^1(E_j)}.
\end{align*}
By the uniform $L^1$ boundedness of the operator $Q_{E_j}$ (a consequence of
Shadrin's theorem~\ref{thm:shadrin}), we infer $\| g_j
\|_{L^1} \leq C \|f\|_{L^1(V_j)}$, which, together with the latter display and
the disjointness of the sets $(V_j)$,
gives us
\[
|\{ t\in \widetilde{G_\lambda}^c : \sup_{M\leq N} |P_M(T_\varepsilon g)(t)| > \lambda/2 \}|
\lesssim_{\gamma,\beta} \frac{1}{\lambda} \| f\|_{L^1}.
\]
Combining this inequality with \eqref{eq:split_maximal}, \eqref{eq:hfinal},
\eqref{eq:esttilde}, and
\eqref{eq:splitf} yields the conclusion of Theorem~\ref{thm:main}. Therefore,
we continue with the proof of \eqref{eq:crucialtoshow}.
Observe first that in order to show \eqref{eq:crucialtoshow}, we restrict the
summation to $n> n(E_j)$ where we recall that $n(E_j)$ is the
smallest index $m$ so that $E_j \in \mathscr C_m$. This is possible, since for $n
\leq n(E_j)$, the function $f_n|_{E_j}$ is contained in the range of
the operator $Q_{E_j}$ (which is the spline space $S_k( E_j \cap \mathscr
A_{n(E_j)})$) and this implies
that $\langle g_j, f_n\rangle = 0$ due to the defining equation $g_j = f\ensuremath{\mathbbm 1}_{V_j} -
Q_{E_j}(f\ensuremath{\mathbbm 1}_{V_j})$.
We slightly change the language, fix the index $j$ and write $E= E_j$,
$L=L_j$, $H=H_j$ and $b$
for a generic function supported on $E$. Thus \eqref{eq:crucialtoshow} is
implied by the estimate
\[
\Big\| \sum_{n> n(E)} |\langle b,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(H^c)} \lesssim_{\gamma,\beta} \|b\|_{L^1(E)}.
\]
Divide the index set $\{n> n(E)\}$ into the
two parts
\[
\Gamma_1 = \{ n> n(E) : J_n \subset L^c \},\qquad
\Gamma_2 = \{ n> n(E) : J_n \subset L \},
\]
where we recall that $J_n$ is the characteristic interval of the function $f_n$
defined in Section~\ref{sec:tensor_def}.
Since $L$ is a union of atoms in $\mathscr A_{n(E)}$ and $J_n$ (for $n>n(E)$) is
an atom in the finer $\sigma$-algebra $\mathscr F_n$,
we have $\{ n> n(E) \} =
\Gamma_1\cup \Gamma_2$ with a disjoint union.
\textsc{Case 1:} First consider the case $n\in\Gamma_1=\{ n> n(E) : J_n
\subset L^c \}$ and use the pointwise estimate \eqref{eq:pw} and its consequence
$\| f_n \|_{L^1} \lesssim |J_n|^{1/2}$ for the functions $f_n$ to deduce
(denoting $d_n = d_{\mathscr A_n}$)
\begin{equation}\label{eq:b_est}
\begin{aligned}
\Big\| \sum_{n\in \Gamma_1} |\langle b,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(H^c)} &\leq \sum_{n\in\Gamma_1}
|\langle b,f_n\rangle | \| f_n \|_{L^1} \\
&\lesssim \sum_{n\in\Gamma_1} \sum_{\substack{\text{$B$ atom of
$\mathscr A_n$:} \\
B\subset E}}
\frac{q^{|d_n(J_n,B)|_1} |J_n| } { |\conv(J_n, B)| }
\int_B |b(y)|\,\mathrm{d} y \\
&= \sum_{s\in\mathbb Z^d} q^{|s|_1} \int_E\Big(\sum_{n\in\Gamma_1}\sum_{\substack{\text{$B$
atom of $\mathscr A_n$:} \\
B\subset E,\ d_n(J_n,B) = s}}
\frac{ |J_n|\ensuremath{\mathbbm 1}_B(y) } { |\conv(J_n, B)| }\Big)
|b(y)|\,\mathrm{d} y.
\end{aligned}
\end{equation}
Let $s\in \mathbb Z^d$ and $y\in E$ be fixed. Define $\Gamma_{1,y,s}$ to be the set of all
$n\in \Gamma_1$ so that there exists an atom $B$ of $\mathscr A_n$ with
$y\in B\subset E$ and $d_n(J_n,B) = s$. Since $s\in \mathbb Z^d$ is fixed, this atom $B$
is given uniquely by the index $n\in \Gamma_{1,y,s}$ and we denote
$B_n = B$.
Now, split the index set $\Gamma_{1,y,s}$ into the not necessarily disjoint union
\[
\Gamma_{1,y,s} = \bigcup_{\delta=1}^d \Gamma_{1,y,s}^{\delta},\qquad
\text{with }\Gamma_{1,y,s}^{\delta} = \{ n\in \Gamma_{1,y,s} : J_{n}^{\delta} \subset
(L^{\delta})^c \}.
\]
We fix the parameter $\delta \in \{1,\ldots,d\}$
and proceed to estimate the sum
\begin{equation}\label{eq:sumtoestimate}
\sum_{n\in\Gamma_{1,y,s}^{\delta}} \frac{|J_{n}|}{|\conv(J_{n},B_{n})|}
\leq \sum_{n\in\Gamma_{1,y,s}^{\delta}}
\frac{|J_{n}^\delta|}{|\conv(J_{n}^\delta,B_{n}^\delta)|}
=
\sum_{n\in\Gamma_{1,y,s}^{\delta}} \int_{J_{n}^\delta}
\frac{1}{|\conv(J_{n}^\delta,B_{n}^\delta)|}\,\mathrm{d} t.
\end{equation}
Let $x$ be the endpoint of $E^\delta$ that is closest to the sets
$J_{n}^\delta$ for $n\in \Gamma_{1,y,s}^\delta$. Then, observe that
$|x-t| \leq |\conv(J_{n}^\delta, B_{n}^\delta)|$ for $t\in J_{n}^\delta$.
Recall that $J_{n}^\delta\subset (L^\delta)^c$ and
$B_{n}^\delta\subset E^\delta$ for all $n\in \Gamma_{1,y,s}^{\delta}$.
Therefore, $|x-t|\geq c |E^\delta|$ for $t\in J_{n}^\delta$ and some constant $c$
depending only on $k$ and $\gamma$.
Moreover, by $k_\delta$-regularity of the filtration in
direction $\delta$, we have $ |\conv(J_{n}^\delta,B_{n}^\delta)| \leq C
\gamma^{|s_\delta|}
|E^\delta|$ for some absolute constant $C$ (we can assume
without restriction that $\gamma \geq 2$).
Those estimates yield
\[
c|E^\delta| \leq |x-t| \leq C\gamma^{|s_\delta|} |E^\delta|,
\qquad t\in J_{n}^\delta.
\]
Let $\Lambda$ be a set of indices $n\in \Gamma_{1,y,s}^{\delta}$ so that for
$i,j\in\Lambda$ with $i < j$ we have $J_{i}^\delta \supseteq
J_{j}^\delta$.
We invoke Lemma \ref{lem:comb} with the setting $C_n = B_{n}$ for $n\in\Lambda$ to deduce that the
cardinality of $\Lambda$ is $\lesssim_{\gamma,\beta} \sum_{j\neq \delta} (1+|s_j|)$.
Those observations imply
\begin{align*}
\sum_{n\in\Gamma_{1,y,s}^{\delta}} \int_{J_{n}^\delta}
\frac{1}{|\conv(J_{n}^\delta,B_{n}^\delta)|}\,\mathrm{d} t &\leq
\sum_{n\in\Gamma_{1,y,s}^{\delta}} \int_{J_{n}^\delta}
\frac{1}{|x-t|}\,\mathrm{d} t \\
&\lesssim_{\gamma,\beta}
\Big(\sum_{j\neq \delta} (1+|s_j|)\Big)
\int_{c|E^\delta|}^{C\gamma^{|s_\delta|} |E^\delta|}
\frac{\,\mathrm{d} u}{u} \lesssim_{\gamma} |s_\delta|\cdot \sum_{j\neq \delta} (1+|s_j|).
\end{align*}
Inserting this estimate, combined with \eqref{eq:sumtoestimate},
in the last line of \eqref{eq:b_est} and summing a geometric series, we obtain that
\[
\Big\| \sum_{n\in \Gamma_1} |\langle b,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(H^c)} \lesssim_{\gamma,\beta} \int_E |b(t)|\,\mathrm{d} t.
\]
\textsc{Case 2: }
Now we consider $n\in\Gamma_2 = \{ n> n(E) : J_n \subset L \}$.
Estimate \eqref{eq:pw} with the notation $d_n = d_{\mathscr
A_n}$ gives
\begin{equation}\label{eq:case2}
\begin{aligned}
\Big\| \sum_{n\in \Gamma_2}& |\langle b,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(H^c)} \leq \sum_{n\in\Gamma_2}
|\langle b,f_n\rangle | \| f_n \|_{L^1(H^c)} \\
&\leq \sum_{n\in\Gamma_2} \sum_{\substack{\text{$A,B$ atom of $\mathscr
A_n$:} \\
B\subset E,\ A\subset H^c}
}
\frac{q^{|d_n(B,J_n)|_1 + |d_n(A,J_n)|_1} |J_n|\cdot |A| } { |\conv(J_n,
B)|\cdot |\conv(J_n,A)| }
\int_B |b(y)|\,\mathrm{d} y \\
&= \sum_{r,s\in\mathbb Z^d} q^{|r|_1 + |s|_1}\int_E \Big(\sum_{n\in\Gamma_2}\sum_{\substack{\text{$A,B$ atom of $\mathscr
A_n$:} \\
B\subset E,\ A\subset H^c, \\ d_n(A, J_n) = r,\
d_n(A,B) = s} }
\frac{ |J_n|\cdot |A|\cdot \ensuremath{\mathbbm 1}_{B}(y) } { |\conv(J_n,
B)|\cdot |\conv(J_n,A)| } \Big)
|b(y)|\,\mathrm{d} y.
\end{aligned}
\end{equation}
For fixed $r,s\in \mathbb Z^d$ and $y\in E$, we let $\Gamma_{2,y,(r,s)}$ be the set of
all $n\in \Gamma_2$ so that there exist two atoms $A,B$ of $\mathscr A_n$ with
$y\in B\subset E$, $A\subset H^c$ and $d_n(A,J_n) = r$, $d_n(A,B) = s$.
For fixed $r,s\in\mathbb Z^d$, those atoms $A,B$ are given uniquely by the index
$n\in \Gamma_{2,y,(r,s)}$ and are denoted by $A_n,B_n$ respectively.
Note that if $n\in \Gamma_{2,y,(r,s)}$ then we know that $d_n(J_n, B_n) =
s-r$.
Split the index set $\Gamma_{2,y,(r,s)}$ further into the (not necessarily disjoint) subcollections
\[
\Gamma_{2,y,(r,s)}^{\delta} = \{ n\in \Gamma_{2,y,(r,s)} :
A^{\delta}
\subset (H^{\delta})^c \},\qquad
\delta\in\{1,\ldots,d\}.
\]
Fixing the parameter $\delta\in \{1,\ldots,d\}$
we estimate the sum
\begin{equation}
\label{eq:deltasplit}
\sum_{n\in
\Gamma_{2,y,(r,s)}^{\delta}} \frac{ |J_n| |A_n| }{ |
\conv(J_n, B_n)| \cdot | \conv(J_n,A_n) |} \leq
\sum_{n\in
\Gamma_{2,y,(r,s)}^{\delta}} \frac{ |J_n^\delta| |A_n^\delta| }{ |
\conv(J_n^\delta, B_n^\delta)| \cdot |
\conv(J_n^\delta,A_n^\delta) |}.
\end{equation}
Consider the indices of all maximal sets $A_n^\delta$ by setting
\[
\Lambda = \{ n \in \Gamma_{2,y,(r,s)}^{\delta} : \text{ there is no
$m\in \Gamma_{2,y,(r,s)}^{\delta}$ with $m < n$ and
$A_m^\delta \supseteq A_n^\delta$} \}.
\]
Moreover, for $m\in\Lambda$, define
\[
\Lambda_m = \{ n\in \Gamma_{2,y,(r,s)}^{\delta} : A_n^\delta
\subseteq A_m^\delta\}
\]
and split the sum on the right hand side of \eqref{eq:deltasplit} into
\begin{equation}\label{eq:deltamax}
\sum_{m\in\Lambda} \sum_{n\in\Lambda_m }\frac{ |J_n^\delta|
|A_n^\delta| }{ |
\conv(J_n^\delta, B_n^\delta)| \cdot |
\conv(J_n^\delta,A_n^\delta) |}.
\end{equation}
Fix $m\in \Lambda$.
If $n\in\Lambda_m$ we have $A_n^\delta \subset A_m^\delta$, and thus we estimate $|A_n^\delta|\leq
|A_m^\delta|$ and also
$|\conv(J_m^\delta,A_m^\delta)| \lesssim_\gamma
|\conv(J_n^\delta,A_n^\delta)|$ since already in the $\sigma$-algebra $\mathscr
A_{m}^\delta$,
we have at least $ k_\delta$ atoms between $J_m^\delta$ and $A_m^\delta$ (recall
$A_m^\delta\subset (H^\delta)^c$, $m>n(E)$,
and $J_n^\delta\subset L^\delta$)
and $k_\delta$-regularity in direction $\delta$ gives this inequality. Therefore, we can
estimate the sum in \eqref{eq:deltamax} from above by
\begin{equation}\label{eq:toshow_double}
\sum_{m\in\Lambda}
\frac{|A_m^\delta|}{|\conv(J_m^\delta,A_m^\delta)|}
\sum_{ n\in\Lambda_m}\frac{ |J_n^\delta| }{ |
\conv(J_n^\delta, B_n^\delta)| }.
\end{equation}
Fix $m\in\Lambda$, fix a B-spline support $\Delta$ of order $k_\delta$ in the
$\sigma$-algebra $\mathscr A_m^\delta$ between
$A_m^\delta$ and $L^\delta$.
For all $n\in\Lambda_m$,
$\Delta$ is a B-spline support of order $k_\delta$ in the $\sigma$-algebra $\mathscr
A_n^\delta$ as well since the number of atoms of $\mathscr A_n^\delta$
between $A_n^\delta$ and $B_n^\delta$ is constant and $A_n^\delta\subset
A_m^\delta$ and also the sets $B_n^\delta$ are decreasing.
If $n\in \Lambda_m$, we know
that between $A_n^\delta$ and $J_n^\delta$ we have $|r_\delta|$
atoms of $\mathscr A_n^\delta$ and this means $|d_n^\delta
(D, J_n^\delta)| \leq |r_\delta|$ for any atom $D\subset \Delta$ in
$\mathscr A_n^\delta$, $n\in\Lambda_m$ (using the notation $d_n^\delta =
d_{\mathscr A_n^\delta}$).
This implies by $k_\delta$-regularity of the $\sigma$-algebra $\mathscr
A_n^\delta$
\begin{equation}\label{eq:reg}
\frac{\gamma^{-|r_\delta|}|\Delta|}{k_\delta} \leq |J_{n}^\delta| \leq
\gamma^{|r_\delta|}
|\Delta|.
\end{equation}
Now we consider two subcases relating the values of $r$ and $s$ for the
analysis of the inner sum in \eqref{eq:toshow_double} for fixed $m\in\Lambda$.
We remark that by definition of $r$ and $s$ and the location of $A_n,B_n,J_n$,
the sign of $r_\delta$ is the same as the sign of $s_\delta$.
\textsc{Case 2a: }$|r_\delta|<|s_\delta|$:
In this case, the set $J_n^\delta$ is strictly between
$A_n^\delta$ and $B_n^\delta$ for all $n\in \Lambda_m$.
This implies that
since $d_n^\delta(A_n^\delta,J_n^\delta) = r_\delta$ and
$d_n^\delta(J_n^\delta,B_n^\delta)= s_\delta-r_\delta$ are both constant
for $n\in \Lambda_m$,
that the sets $J_n^\delta$ have to coincide for all $n\in \Lambda_m$.
Set $C_n = B_n^{1}\times \cdots \times B_n^{\delta-1}
\times A_n^{\delta}\times B_n^{\delta+1}\times \cdots \times B_n^{d}$ for
$n\in \Lambda_m$ and apply Lemma~\ref{lem:comb} to deduce that the cardinality
of $\Lambda_m$ is $\lesssim_{\gamma,\beta} \sum_{j\neq \delta} (1+|s_j - r_j|)$.
This gives the estimate
\[
\sum_{n\in\Lambda_m }\frac{ |J_n^\delta|}
{|\conv(J_n^\delta, B_n^\delta)| } \lesssim_{\gamma,\beta} \sum_{j\neq \delta} (1 + | s_j -r_j|).
\]
\textsc{Case 2b: }$|r_\delta|\geq|s_\delta|$:
Here, $B_n^\delta$ is between $A_n^\delta$ and
$J_n^\delta$ for all $n\in\Lambda_m$.
Let $x$ be the point in $B_m^\delta$ closest to $A_m^\delta$. For
$n\in\Lambda_m$, let $I_n\subset J_n^\delta$ be an interval such that
$|I_n|=|J_n^\delta|/2$ and $\operatorname{dist} (x,I_n) \geq |J_n^\delta|/2$.
Then
\[
\sum_{n\in\Lambda_m}\frac{ |J_n^\delta| }{ |
\conv(J_n^\delta, B_n^\delta)| }
=
2 \sum_{n\in\Lambda_m}
\int_{I_n}
\frac{1}{|\conv(J_n^\delta,B_n^\delta)|} \,\mathrm{d} t.
\]
Note that
$x$ is also an endpoint of $B_n^\delta$ for all
$n\in\Lambda_m$ and thus, if $t\in I_n$,
\begin{equation}\label{eq:x_minus_t}
|J_n^\delta| /2\leq |x-t| \leq |\conv(J_n^\delta,B_n^\delta)|,\qquad
n\in\Lambda_m.
\end{equation}
Let $\Omega \subset \Lambda_m$ be so that for $i,j\in\Omega$
with $i<j$ we have $J_i^\delta \supseteq J_j^\delta$. Then we apply
Lemma~\ref{lem:comb} with $C_n = B_n^{1}\times \cdots \times B_n^{\delta-1}
\times A_n^{\delta}\times B_n^{\delta+1}\times \cdots \times B_n^{d}$ for
$n\in \Omega$ to deduce that the cardinality of $\Omega$ is
$\lesssim_{\gamma,\beta} \sum_{j\neq \delta} (1+|s_j - r_j|)$.
Therefore we estimate further
\[
\sum_{n\in\Lambda_m} \int_{I_n}
\frac{1}{|\conv(J_n^\delta,B_n^\delta)|} \,\mathrm{d} t \lesssim_{\gamma,\beta}
\Big(\sum_{j\neq \delta} (1+|s_j - r_j|) \Big)
\int_{\cup_{n} I_n}
\frac{1}{|x-t|}\,\mathrm{d} t,
\]
which, by inequalites \eqref{eq:x_minus_t} and \eqref{eq:reg}, is smaller than
\[
\Big(\sum_{j\neq \delta} (1+|s_j - r_j|)\Big)
\int_{\gamma^{-|r_\delta|}
|\Delta|/(2k_\delta)}^{K\gamma^{|r_\delta|} |\Delta|}
\frac{\,\mathrm{d} u}{u}\lesssim_{\gamma,\beta}
\Big( \sum_{j\neq \delta}(1+ |s_j - r_j|)\Big) \cdot |r_\delta|,
\]
for some absolute constant $K$.
Thus, we come back to \eqref{eq:toshow_double} and combine the results of
subcases \textsc{2a} and \textsc{2b} to
obtain
\[
\sum_{m\in\Lambda}
\frac{|A_m^\delta|}{|\conv(J_m^\delta,A_m^\delta)|}
\sum_{ n\in\Lambda_m}\frac{ |J_n^\delta| }{ |
\conv(J_n^\delta, B_n^\delta)| }
\lesssim_{\gamma,\beta} \Big( \sum_{j\neq \delta} (1+|s_j - r_j|)\Big)
|r_\delta|
\sum_{m\in\Lambda}
\frac{|A_m^\delta|}{|\conv(J_m^\delta,A_m^\delta)|}.
\]
The next thing is to
estimate the latter sum in terms of $r$ and $s$.
This can be done as follows.
Denote by $x$ the endpoint of $L^\delta$ that is closest
to the sets $A_m^\delta$, $m\in\Lambda$.
Since for all $m\in\Lambda$, $J_m^\delta$ is a subset of $L^\delta$ and
$A_m^\delta$ is a subset
of $(H^\delta)^c$, the distance between $x$ and $t$ is greater than
$c |E^\delta|$ for all $t\in A_m^\delta$ and some constant $c$ depending only on $k,\gamma$. Using
$|d_m^\delta(A_m^\delta, J_m^\delta)| = |r_\delta|$ for all $m\in\Lambda$ and $k_\delta$-regularity
of $(\mathscr A_n^\delta)$, we obtain
\[
c|E^\delta|\leq |x-t| \leq |\conv(J_m^\delta, A_m^\delta)| \leq C
\gamma^{|r_\delta|} |E^\delta|,
\qquad t\in A_m^\delta
\]
for some absolute constant $C$.
This implies, since the maximal sets $A_{m}^\delta$, $m\in\Lambda$, are disjoint,
\[
\sum_{m\in \Lambda} \frac{|A_m^\delta|}{|\conv
(J_m^\delta, A_m^\delta)|} \leq
\sum_{m\in\Lambda} \int_{A_m^\delta} \frac{1}{|x-t|} \,\mathrm{d} t
\leq \int_{c|E^\delta|}^{C\gamma^{|r_\delta|} |E^\delta|}
\frac{1}{u}\,\mathrm{d} u \lesssim_{\gamma} |r_\delta|.
\]
Thus
\[
\sum_{m\in\Lambda}
\frac{|A_m^\delta|}{|\conv(J_m^\delta,A_m^\delta)|}
\sum_{ n\in\Lambda_m}\frac{ |J_n^\delta| }{ |
\conv(J_n^\delta, B_n^\delta)| }
\lesssim_{\gamma,\beta}
\Big( \sum_{j\neq \delta}(1+ |s_j - r_j|)\Big)
|r_\delta|^2.
\]
Therefore, coming back to the very beginning of \textsc{Case 2} and inserting this
estimate into the last line of~\eqref{eq:case2},
\begin{align*}
\Big\| \sum_{n\in \Gamma_2} |\langle b,f_n\rangle| \cdot
|f_n|\Big\|_{L^1(H^c)} &
\lesssim_{\gamma,\beta} \sum_{r,s\in\mathbb Z^d} \sum_{\delta=1}^d
\Big(\sum_{j\neq
\delta} (1+|s_j - r_j|)\Big) \cdot |r_\delta|^2 q^{|s|_1 + |r|_1} \int_E
|b(t)|\,\mathrm{d} t \\
&\lesssim \int_E |b(t)|\,\mathrm{d} t.
\end{align*}
Combining now \textsc{Case 1} and \textsc{Case 2} and setting $b=g_j$, we have proved
\eqref{eq:crucialtoshow}.
This completes the proof of our main Theorem~\ref{thm:main}.
\subsection*{Acknowledgments}
The author is supported by the Austrian Science Fund FWF, project P32342.
\bibliographystyle{plain}
|
\section{Introduction}
An essential submodule of a module is a submodule that intersects any other nonzero submodule nontrivially. It is named by Eckman and Schpof (see [\cite{ES53}]) though the idea behind essential submodule in a left $R$-module $M$ is due to Johnson (see [\cite{RJ51}]). Essential submodules play a crucial role in noncommutative algebra. As an example, a necessary and sufficient condition that \emph{a left ideal in a ring $R$ is essential, is the existence of a nonzero-divisor (that is, a regular element)}, is again a necessary and sufficient condition for a ring $R$ to have a left classical Artin semisimple quotient ring, that is, $R$ is semiprime left Goldie ring. This shows that the self-generating property of essential submodules generate other essential objects. In much of the literature, the definition of an essential submodule includes the parent module. In this paper, we work with proper essential submodules whose definition is the following:
\begin{definition}\label{df}
A proper nontrivial submodule $E$ of a left $R$-module $M$ is said to be a proper essential submodule of $M$ if $E \cap N \neq \{0_{M}\}$ for every nontrivial submodule $N$ of $M$.
\end{definition}
Throughout the paper, unless otherwise stated, we assume $R$ is a ring with unity $1 \neq 0$ which is not a field. Observe that any non-trivial proper ideal $I$ in an integral domain $R$ is a proper essential submodule of the $R$-module $R$. However, it is not necessary for an $R$-module to contain a proper essential submodule. Because, if we consider a finite $\mathbb{Z}$-module $M$ of order $n$, where $n$ is square-free, and if $N_{p}$ denote the submodule of $M$ of order $p$ where $p$ is a prime dividing $n$, then $M = \bigoplus_{p \mid n} N_{p}$ shows that $M$ does not contain any proper essential submodule. The following proposition is our observation and is supposed to prevail in the literature, but we could not trace the result either in direct or indirect form anywhere. So we proved this result and incorporated it here.
\begin{proposition}\label{history}
If an $R$-module $M$ is semisimple, then $M$ does not contain any proper essential submodule. Hence, any $\mathbbm{k}$-vector space $V$ does not contain any proper essential subspace.
\end{proposition}
\begin{proof}
Let $M = \bigoplus_{\alpha \in \Lambda} N_{\alpha}$, where each $N_{\alpha}$ is a simple submodule of $M$. If a submodule $E$ of $M$ intersects every other non-trivial submodule of $M$ non-trivially, then $N_{\alpha} \subseteq E$ for all $\alpha \in \Lambda$. But this implies $\bigoplus_{\alpha \in \Lambda} N_{\alpha} \subseteq E$, that is, $M \subseteq E$.
Let $\{v_{\alpha}\}_{\alpha \in \Lambda}$ be a basis for the $\mathbbm{k}$-vector space $V$. Then we have $V =\bigoplus_{\alpha \in \Lambda} V_{\alpha}$, where $V_{\alpha}$ is the subspace generated by $v_{\alpha}$. As each $V_{\alpha}$ is a simple $\mathbbm{k}$-subspace of $V$, we can use the first part of this proposition to conclude that $V$ does not contain any proper essential subspace.
\end{proof}
A part of the following theorem \eqref{baba} is very much well-known (see [\cite{TL99}]). Therefore, we include proof of the other part of the theorem which we again did not find in the literature.
\begin{theorem}\label{baba}
Let $M$ be a left module over $R$. Then one can show the equivalence
$(\mathrm{a})$ Every submodule of $M$ is a direct summand.
$(\mathrm{b})$ $M$ has no proper essential submodule.
$(\mathrm{c})$ $M$ is semisimple.
$(\mathrm{d})$ $\mathrm{Soc}(M) = M$.
\end{theorem}
\begin{proof}
Equivalency of (a), (c) and (d) are found in [\cite{TL99}]. Here, we show the equivalence of (a) and (b), that is, $\mathrm{(a)} \Leftrightarrow \mathrm{(b)}$.
$\mathrm{(a)} \Rightarrow \mathrm{(b)}:$ Suppose $M$ has a proper essential submodule. Then from the definition, it trivially follows that the proper essential submodule can not be the direct summand.
$\mathrm{(b)} \Rightarrow \mathrm{(a)}:$ Conversely let $N$ be a proper submodule of $M$. Then $N$ is not proper essential. There exists a submodule $T$ such that $N\cap T = \{0\}$. Choose a maximal such $T$, then $N\bigoplus T$ is either proper essential in $M$ or $N \bigoplus T = M$. But as per hypothesis, $N \bigoplus T = M$, so the result follows.
\end{proof}
Theorem \eqref{babama} can be found in [\cite{JL66}]), so, we skip its proof.
\begin{theorem}\label{babama}
The intersection of all essential submodules of a left $R$-module $M$ is equal to its socle $\mathrm{Soc}(M)$.
\end{theorem}
\section{Proper essential submodules over integral and principal ideal domains}
\begin{theorem}\label{maa}
Let $\{M_{\omega}\}_{\omega \in \Lambda}$ be a family of left $R$-modules indexed by a nonempty set $\Lambda$. Then $\bigoplus_{\omega \in \Lambda} M_{\omega}$ has a proper essential submodule if and only if some $M_{\omega}$ has a proper essential submodule.
\end{theorem}
\begin{proof}
Let $M = \bigoplus_{\omega \in \Lambda} M_{\omega}$ and $\omega_{0} \in \Lambda$ be such that $M_{\omega_{0}}$ has a proper essential submodule $E_{\omega_{0}}$. Let $X = \bigoplus_{\omega \in \Lambda} X_{\omega}$, where $X_{\omega_{0}} = E_{\omega_{0}}$ and $X_{\omega} = M_{\omega}$ if $\omega \neq \omega_{0}$. We claim that $X$ is a proper essential submodule of $M$. Let $N$ be any nontrivial submodule of $M$ and let $(n_{\omega})_{\omega \in \Lambda}$ be a nonzero element of $N$. Clearly, $(n_{\omega})_{\omega \in \Lambda} \in X$ if $n_{\omega_{0}} = 0_{M_{\omega_{0}}}$. If $n_{\omega_{0}} \neq 0_{M_{\omega_{0}}}$, consider the cyclic module $\langle n_{\omega_{0}} \rangle$. As $E_{\omega_{0}} \cap \langle n_{\omega_{0}} \rangle \neq \{0_{M_{\omega_{0}}}\}$, we conclude that $X \cap N \neq \{0_{M}\}$.
For the converse, assume that $M$ has a proper essential submodule $E$ but $M_{\omega}$ does not have a proper essential submodule for all $\omega \in \Lambda$. Observe that, for all $\omega \in \Lambda$, we have either $E \cap M_{\omega}$ is proper essential in $M_{\omega}$ or $E \cap M_{\omega} = M_{\omega}$. As $M_{\omega}$ does not have any proper essential submodule, it must be that $E \cap M_{\omega} = M_{\omega}$ for all $\omega \in \Lambda$, that is, $M \subseteq E$, a contradiction.
\end{proof}
Free modules play a central role in algebra, since any module is the homomorphic image of some free module. The following corollary is now immediate from theorem \eqref{maa}.
\begin{theorem}\label{spm}
Let $R$ be an integral domain and let $A$ be a nonempty set. Then the free $R$-module $R(A)$ on the nonempty subset $A$ has a proper essential submodule.
\end{theorem}
\begin{proof}
Since $R(A)$ is free on the subset $A$, we have $R(A) \simeq \bigoplus_{A} R$. The proof is now completed using theorem \eqref{maa}.
\end{proof}
An immediate consequence of theorem \eqref{maa} is the classification of finitely generated modules over principal ideal domains possessing proper essential submodules. Before that, we look the following lemma.
\begin{lemma}\label{ons}
The non-simple left cyclic module $\langle m \rangle$ over an integral domain $R$ has a proper essential submodule if and only if there exists a maximal ideal $\mathcal{M}$ of $R$ containing $\mathrm{Ann}_{R}(m)$ such that for any ideal $I$ of $R$ containing $\mathrm{Ann}_{R}(m)$ we have $\alpha \beta \notin \mathrm{Ann}_{R}(m)$ for some $\alpha \in \mathcal{M} \setminus \mathrm{Ann}_{R}(m)$ and for some $\beta \in I \setminus \mathrm{Ann}_{R}(m)$. Hence, the non-simple cyclic module $\langle m \rangle$ over a principal ideal domain $R$ has a proper essential submodule if and only if the generator of the $\mathrm{Ann}_{R}(m)$ is not square-free.
\end{lemma}
\begin{proof}
Let $\mathcal{M}$ be a maximal ideal of $R$ that contains $\mathrm{Ann}_{R}(m)$ and $I$ be any ideal of $R$ containing $\mathrm{Ann}_{R}(m)$. If for some $\alpha \in \mathcal{M} \setminus \mathrm{Ann}_{R}(m)$ and for some $\beta \in I \setminus \mathrm{Ann}_{R}(m)$ we have $\alpha \beta \notin \mathrm{Ann}_{R}(m)$, then we see that $\mathcal{M}/\mathrm{Ann}_{R}(m)$ is proper essential in $R/\mathrm{Ann}_{R}(m)$. As $R/\mathrm{Ann}_{R}(m) \simeq \langle m \rangle$, we conclude that $\langle m \rangle$ has a proper essential submodule.
Conversely, suppose $\langle m \rangle$ has a proper essential submodule $E$. Since we have $\langle m \rangle \simeq R/\mathrm{Ann}_{R}(m)$, there exists some ideal $J$ of $R$ containing $\mathrm{Ann}_{R}(m)$ such that $J/\mathrm{Ann}_{R}(m) \simeq E$. Put $J$ in some maximal ideal $\mathcal{M}$. As $J/\mathrm{Ann}_{R}(m)$ is proper essential in $R/\mathrm{Ann}_{R}(m)$, for any ideal $I$ of $R$ containing $\mathrm{Ann}_{R}(m)$, we have $J/\mathrm{Ann}_{R}(m) \cap I/\mathrm{Ann}_{R}(m) \neq \{0_{\langle m \rangle}\}$. As $J$ is contained in $\mathcal{M}$, we get that $\mathcal{M}/\mathrm{Ann}_{R}(m) \cap I/\mathrm{Ann}_{R}(m) \neq \{0_{\langle m \rangle}\}$.
Let $\mathrm{Ann}_{R}(m) = \langle a \rangle$ and let $a = up_{1}p_{2} \cdots p_{n}$ where $p_{i}$'s are distinct primes and $u$ is a unit. The only maximal ideals that contain $\langle a \rangle$ are the ideals $\langle p_{i} \rangle$ where $1 \leq i \leq n$. If $\alpha \in \langle p_{i} \rangle \setminus \langle a \rangle$ and $\beta \in \langle q_{i} \rangle \setminus \langle a \rangle$ where $q_{i} = p_{1}p_{2} \cdots p_{i - 1}p_{i + 1} \cdots p_{n}$, we see that $\alpha \beta \in \langle a \rangle$. Now, the first part of this lemma shows that $\langle m \rangle$ does not have a proper essential submodule.
Assume now that $a = up_{1}^{\alpha_{1}}p_{2}^{\alpha_{2}} \cdots p_{n}^{\alpha_{n}}$, where $p_{i}$'s are distinct primes, $u$ is a unit and $\alpha_{i} > 1$ for some $i$. It is now evident that the maximal ideal $\langle p_{i} \rangle$ of $R$ satisfies the property of the first of this lemma.
\end{proof}
\begin{theorem}\label{jg}
A finitely generated left module $M$ over a principal ideal domain $R$ has a proper essential submodule if and only if the betti number of $M$ is at least $1$ or some invariant factor is not square free.
\end{theorem}
\begin{proof}
The fundamental theorem of finitely-generated modules over principal ideal domain yields the invariant-factor decomposition $M = R^{n} \bigoplus R/\langle a_{{1}} \rangle \bigoplus \cdots \bigoplus R/\langle a_{{m}} \rangle$. The proof is now completed using theorems \eqref{maa}, \eqref{spm}, and lemma \eqref{ons}
\end{proof}
\begin{lemma}\label{ram}
A left $R$-module $M$ has a proper essential submodule if and only if the left $R/\mathrm{Ann}_{R}(M)$-module $M$ has a proper essential submodule. Hence, if an $R$-module $M$ has a proper essential submodule, then $\mathrm{Ann}_{R}(M)$ can not be a maximal ideal.
\end{lemma}
\begin{proof}
At first, assume that $R$-module $M$ has a proper essential submodule $E$. Now, let $N$ be any submodule of $R/\mathrm{Ann}_{R}(M)$-module $M$. Consider the cyclic $R/\mathrm{Ann}_{R}(M)$-module $\langle n \rangle$, where $n \in N$ be any element. As, $\langle n \rangle \cap E \neq \{0_{M}\}$, there exists $r \in R$ such that $rn \neq 0$ and $rn \in \langle n \rangle \cap E$. Then $(r + \mathrm{Ann}_{R}(M))n \in \langle n \rangle \cap E$ (as $R/\mathrm{Ann}_{R}(M)$-module). Hence, this shows that $R/\mathrm{Ann}_{R}(M)$-module $M$ has a proper essential submodule $E$.
Conversely, assume that $R/\mathrm{Ann}_{R}(M)$-module $M$ has a proper essential submodule $E$. Let $N$ be any submodule of $M$. Then, $\langle n \rangle \cap E \neq \{0_{M}\}$ for $n \in N$ with $n \neq 0_{M}$. This implies, for some $(r + \mathrm{Ann}_{R}(M)) \in R/\mathrm{Ann}_{R}(M)$ with $r + \mathrm{Ann}_{R}(M) \neq \mathrm{Ann}_{R}(M)$, $(r + \mathrm{Ann}_{R}(M))n \in E$, that is, $rn \in E$. Thus, $E$ is proper essential in $M$.
Now, applying proposition \eqref{history} and the first part of this lemma, we get that $\mathrm{Ann}_{R}(M)$ can not be a maximal ideal.
\end{proof}
Let $M$ be a module over an integral domain $R$. If $p \in R$ is a prime, recall that the $p$-primary component $\mathrm{Ann}^{\ast}_{M}(p)$ of $M$ is the set of all elements of $M$ that are annihilated by some positive power of $p$ and its sub-primary component $\mathrm{Ann}_{M}(p^{m})$ is the set of all elements of $M$ that are annihilated by some fixed power of $p$ (i.e., annihilated by $p^{m}$).
\begin{lemma}\label{jpm}
Let $M$ be a non-faithful left module over a principal ideal domain $R$ and let $p_{1}, p_{2}, \ldots, p_{n}$ be the distinct primes that divide the generator of $\mathrm{Ann}_{R}(M)$. Then $M$ has a proper essential submodule if and only if some sub-primary component of $M$ has a proper essential submodule for some prime $p_{i}$.
\end{lemma}
\begin{proof}
Let $\mathrm{Ann}_{R}(M) = \langle a \rangle$ and $a = up_{1}^{\alpha_{1}}p_{2}^{\alpha_{2}} \cdots p_{n}^{\alpha_{n}}$, where $p_{i}$'s are distinct primes and $u$ is a unit. For $1 \leq i \leq n$, if we set $q_{i} = p_{1}^{\alpha_{1}}p_{2}^{\alpha_{2}} \cdots p_{i - 1}^{\alpha_{i - 1}}p_{i + 1}^{\alpha_{i + 1}} \cdots p_{n}^{\alpha_{n}}$, then the homomorphism $\varphi_{i}: M \rightarrow q_{i}M$ shows that $M/\mathrm{Ker}(\varphi_{i}) \simeq q_{i}M$. Since we have $\mathrm{Ker}(\varphi_{i}) = \langle p_{i}^{\alpha_{i}} \rangle M$ and $q_{i}M = \mathrm{Ann}_{M}(p_{i}^{\alpha_{i}})$, we conclude that $M/\langle p_{i}^{\alpha_{i}} \rangle M \simeq \mathrm{Ann}_{M}(p_{i}^{\alpha_{i}})$. As the ideals $\langle p_{1}^{\alpha_{1}} \rangle$, $\langle p_{2}^{\alpha_{2}} \rangle$, \ldots, $\langle p_{n}^{\alpha_{n}} \rangle$ are pairwise comaximal, the chinese remainder theorem yields $M \simeq \bigoplus_{i = 1}^{n} \mathrm{Ann}_{M}(p_{i}^{\alpha_{i}})$. The result now follows from theorem \eqref{maa}.
\end{proof}
\begin{theorem}\label{bholenath}
Let $M$ be a left torsion module over a principal ideal domain $R$. Then $M$ has a proper essential submodule if and only if $\mathrm{Ann}_{M}(p^{i}) \subsetneq \mathrm{Ann}_{M}(p^{j})$ for some some prime $p$ and for some positive integers $i, j$ with $i < j$.
\end{theorem}
\begin{proof}
Let $m \in M$ be such that $rm = 0$ for some $r \in R$. Let $r = up_{1}^{\alpha_{1}}p_{2}^{\alpha_{2}} \cdots p_{n}^{\alpha_{n}}$ be the prime factorisation of $r$ into primes, where $u$ is some unit. Consider the cyclic submodule $\langle m \rangle$. As the submodule $\langle m \rangle$ is annihilated by the ideal $\langle r \rangle$, we can apply lemma \eqref{jpm} to conclude that $\langle m \rangle = \bigoplus_{i = 1}^{n} \mathrm{Ann}_{\langle m \rangle} (p_{i}^{\alpha_{i}})$. Thus, $M = \sum_{{p{\text -\mathrm{prime}}}} \mathrm{Ann}^{\ast}_{M}(p)$. Now choose finite number of primes $q_{1}, q_{2}, \ldots, q_{l}$ in $R$ and consider the submodules $ \mathrm{Ann}^{\ast}_{M}(q_{1}), \mathrm{Ann}^{\ast}_{M}(q_{2}), \ldots, \mathrm{Ann}^{\ast}_{M}(q_{l})$. If $m \in \mathrm{Ann}^{\ast}_{M}(q_{j}) \bigcap \sum_{i \neq j} \mathrm{Ann}^{\ast}_{M}(q_{i})$ for some fixed $j$ with $1 \leq j \leq l$, then we have $q_{j}^{s_{j}}m = 0$ for some $s_{j} \in \mathbb{N}$ and $m = m_{1} + \cdots +m_{j - 1} + m_{j + 1}+ \cdots + m_{l}$. Also, for all $1 \leq i \leq l$ with $i \neq j$, we have $q_{i}^{s_{i}}m_{i} = 0$ for some $s_{i} \in \mathbb{N}$. As $R$ is a principal ideal domain, there exists $x, y \in R$ such that $q_{j}^{s_{j}}x + \xi y = 1$, where $\xi = q_{1}^{s_{1}}q_{2}^{s_{2}} \cdots q_{j - 1}^{s_{j - 1}}q_{j + 1}^{s_{j + 1}} \cdots q_{l}^{s_{l}}$. But this implies $m = q_{j}^{s_{j}}xm + \xi y(m_{1} + \cdots +m_{j - 1} + m_{j + 1}+ \cdots + m_{l})$, that is, $m = 0$. Hence, we have that $M = \bigoplus_{{p{\text -\mathrm{prime}}}} \mathrm{Ann}^{\ast}_{M}(p)$. Now, if $M$ has a proper essential submodule $E$, then for some prime $p$, $E \cap \mathrm{Ann}^{\ast}_{M}(p)$ is proper essential in $\mathrm{Ann}^{\ast}_{M}(p)$ (see theorem \eqref{maa}). If $\mathrm{Ann}_{M}(p^{i}) = \mathrm{Ann}_{M}(p^{j})$ for all positive integers $i, j$, then observe that $\mathrm{Ann}_{R}(\mathrm{Ann}_{M}(p)) = \langle p \rangle$. Thus, using lemma \eqref{ram}, we conclude $R/\langle p \rangle$-vector space $\mathrm{Ann}_{M}(p)$ has a proper essential subspace $E \cap \mathrm{Ann}_{M}(p)$. Again, applying the second part of lemma \eqref{ram}, we get a contradiction.
Conversely, assume that $\mathrm{Ann}_{M}(p^{i}) \subsetneq \mathrm{Ann}_{M}(p^{j})$ for some prime $p$ and for some positive integers $i, j$ with $i < j$. We claim that $\mathrm{Ann}_{M}(p^{i})$ is proper essential in $\mathrm{Ann}_{M}(p)$. Let $N$ be any submodule of $\mathrm{Ann}_{M}(p)$ and let $n \in N$. If $n \notin \mathrm{Ann}_{M}(p^{i})$, then suppose $l$ be the smallest positive integer such that $n \in \mathrm{Ann}_{M}(p^{l})$. Observe now that $p^{l - i}n \in \mathrm{Ann}_{M}(p^{i})$ and $p^{l - i}n \neq 0$. This implies $p^{l - i}n \in N \cap \mathrm{Ann}_{M}(p^{i})$. Now apply theorem \eqref{maa} to conclude that $M$ has a proper essential submodule.
\end{proof}
If $R$ is an integral domain, then a torsion-free left $R$-module $M$ contains a copy of $R$ as a left $R$-module. Also, the $R$-module $D^{-1}R$ has a proper essential submodule. Therefore, it is natural to consider a torsion-free $R$-module $M$ that contains $D^{-1}R$ as its submodule. In this regard, the question that arises is whether $M$ contains a proper essential submodule. The following theorem answers this question.
\begin{theorem}\label{jpmkj}
Let $M$ be a left torsion-free module over an integral domain $R$ and let $D$ denotes the set of all nonzero elements in $R$. If $M$ contains $D^{-1}R$ as its submodule, then the $R$-module $M$ has a proper essential submodule.
\end{theorem}
\begin{proof}
Let $\frac{r}{d} \in D^{-1}R$ be such that $\frac{r}{d} \neq 0_{M}$. Now, consider the collection $\mathcal{C} = \{N \leq M \mid R \subseteq N, \frac{r}{d} \notin N\}$ of submodules of $M$. Clearly $\mathcal{C}$ is a partially ordered set under inclusion. By Zorn's Lemma, it has a maximal element $\mathbbm{E}$. We claim that $\mathbbm{E}$ is a proper essential submodule of $M$. If not, there exists a non-trivial submodule $K$ that intersects $\mathbbm{E}$ trivially, that is, $\mathbbm{E} \cap K = \{0_{M}\}$. Observe that $K$ must not contain any nonzero element in $D^{-1}R$, because, if $\frac{r^{\prime}}{d^{\prime}} \in D^{-1}R$ is such that $\frac{r^{\prime}}{d^{\prime}} \neq 0_{M}$ and $\frac{r^{\prime}}{d^{\prime}} \in K$, then $r^{\prime} \in K$, contrary to our assumption that $\mathbbm{E} \cap K = \{0_{M}\}$. Let $\alpha \in K$ be such that $\alpha \neq 0_{M}$. Clearly, we have $\mathbbm{E} \cap \langle \alpha \rangle = \{0_{M}\}$. Consider the submodule $\mathbbm{E} \oplus \langle \alpha \rangle$. By maximality of $\mathbb{E}$, it must be that $\frac{r}{d} \in \mathbbm{E} \oplus \langle \alpha \rangle$. This implies $r/d + \alpha \in \mathbbm{E} \oplus \langle \alpha \rangle$. If $\frac{r}{d} + \alpha \in \mathbbm{E}$, then $r + d\alpha \in \mathbbm{E}$. But this gives, $d\alpha \in \mathbbm{E}$, a contradiction. Also, if $\frac{r}{d} + \alpha \in \langle \alpha \rangle$, then $r \in \langle \alpha \rangle$, a contradiction. Thus, we can write $\frac{r}{d} + \alpha = e + r^{\ast}\alpha$, where $e \in \mathbbm{E} \setminus \{0_{M}\}$ and $r^{\ast} \in R \setminus \{0_{R}\}$. But this implies $r + d\alpha = de + dr^{\ast}\alpha$. Hence, we get that $r - de = d(r^{\ast} - 1)\alpha$. Observe that $r - de \in \mathbbm{E}$ whereas $d(r^{\ast} - 1)\alpha \in \langle \alpha \rangle$. Hence, it must be that $r - de = 0_{M}$ and $d(r^{\ast} - 1) = 0_{R}$. As $d \neq 0_{R}$, it must be that $r^{\ast} = 1$. Plugging the value of $r^{\ast}$ in the equation $\frac{r}{d} + \alpha = e + r^{\ast}\alpha$ yields $\frac{r}{d} = e$. But this is absurd because $\mathbbm{E}$ being a member of $\mathcal{C}$ cannot contain $\frac{r}{d}$.
\end{proof}
Another question that arises naturally from the definition of proper essential submodule is: if $M$ has a proper essential submodule, does the quotient module $M/N$ also have a proper essential submodule where $N$ is any proper non-trivial submodule of $M$? In general, the answer is no (example: $\mathbb{Z}$-module $\mathbb{Z}_{p^{2}}$). But under certain constraints, we can give an affirmative answer. The converse, however, always holds.
\begin{lemma}\label{gmkj}
Let $M$ be a left $R$-module and $N$ be a submodule of $M$ such that $M/N$ has a proper essential submodule $E/N$. Then, the submodule $E$ is a proper essential submodule of $M$. Moreover, if a module $M$ has a proper essential submodule $E$ and if $N$ is a submodule of $M$ such that $M/N$ is torsion-free and $N \neq E$, then the submodule $(E + N)/N$ is proper essential in $M/N$.
\end{lemma}
\begin{proof}
If $N = \{0_{M}\}$, there is nothing to show. Assume $N \neq \{0_{M}\}$. Let $L$ be a nontrivial proper submodule of $M$ and $u \in (L + N)/N \cap E/N$. Clearly, $u = l + N = e + N$ for some $l \in L \setminus N$ and $e \in E \setminus N$. But this shows that $l \in E$. Hence, we get that $E$ is proper essential in $M$.
If $N = \{0_{M}\}$, there is nothing to show. Therefore, assume $N \neq \{0_{M}\}$. Let $L/N$ be a non-trivial proper submodule of $M/N$. Let $l \in L \setminus N$ be such $rl \in E$, for some nonzero $r \in R$ and $rl \neq 0_{M}$. As $rl \notin N$, we conclude that $rl + N \in (E + N)/N \cap L/N$, that is, $(E + N)/N$ is proper essential in $M/N$.
\end{proof}
\begin{corollary}
Let $M$ be a left module over an integral domain $R$ that contains $D^{-1}R$ as its submodule and $M \neq \mathrm{Tor}(M)$. Then, the $R$-module $M$ has a proper essential submodule.
\end{corollary}
\begin{proof}
Consider the quotient module $M/\mathrm{Tor}(M)$. As $M/\mathrm{Tor}(M)$ is torsion-free, so applying theorem \eqref{jpmkj} we get that $M/\mathrm{Tor}(M)$ has a proper essential submodule. Again, applying lemma \eqref{gmkj}, we see that $M$ has a proper essential submodule.
\end{proof}
The correspondence between elements in $\mathrm{Hom}_{R}(M, N)$ with matrices over $R$ plays a fundamental role in the theory of modules. Keeping this in mind, the most natural question that we investigate is whether the set of matrices (of some dimension) over $R$ (as an $R$-module) has a proper essential submodule. We obtain an affirmative answer in the following theorem. Let $\mathcal{M}_{m \times n}(R)$ denote the set of all $m \times n$ matrices over $R$.
\begin{theorem}
Let $R$ be an integral domian. Then the left $R$-module $\mathcal{M}_{m \times n}(R)$ possess a proper essential submodule.
\end{theorem}
\begin{proof}
Let $M$ and $N$ are the free $R$ modules on the sets of $n$ and $m$ elements respectively. Clearly, we have $M \simeq R^{n}$ and $N \simeq R^{m}$. As $\mathrm{Hom}_{R}(R^{n}, R^{m}) \simeq \mathrm{Hom}_{R}(M, N)$ and $\mathrm{Hom}_{R}(M, N) \simeq \mathcal{M}_{m \times n}(R)$, we get that $\mathrm{Hom}_{R}(R^{n}, R^{m}) \simeq \mathcal{M}_{m \times n}(R)$. Since we have $\mathrm{Hom}_{R}(R^{n}, R^{m}) \simeq \bigoplus_{i = 1}^{n} \bigoplus_{j = 1}^{m} \mathrm{Hom}_{R}(R, R)$, we obtain $\mathcal{M}_{m \times n}(R) \simeq R^{mn}$. Now, we can apply theorem \eqref{spm} to obtain the desired result.
\end{proof}
\section{Essential socles in a new class of modules}
\begin{definition}
A left $R$-module $M$ is said to be an $\mathrm{SM}$ module if every nontrivial submodule of $M$ contains a simple submodule.
\end{definition}
\begin{proposition}
Let $M$ be a left module over $R$ such that $M$ is either Artinian, or, locally finite, or, is of finite length. Then $M$ is an $\mathrm{SM}$ module.
\end{proposition}
\begin{proof}
Let $M$ be an Artinian module. If $M$ is simple there is nothing to show. Therefore assume $M$ is not simple. Let $N$ be any submodule of $M$. If we consider any strictly descending chain $C_{N}: N_{0} = N \supsetneq N_{1} \supseteq N_{2} \supsetneq \cdots \supsetneq N_{s}$ of submodules of $N$ that terminates with the nontrivial proper submodule $N_{s}$, then $N_{s}$ is a simple module contained in $N$. Because, otherwise one can enlarge the chain $C_{N}$.
Now assume $M$ is a locally finite module. Again, if $M$ is simple then $M$ must be finite and there is nothing to show. So, assume $M$ is not simple. Let $N$ be any submodule of $M$ and let $n \in N$. Let $\mathcal{G}$ be the collection of all submodules of $\langle n \rangle$. As $\langle n \rangle$ is finite, we can choose $S \in \mathcal{G}$ such that $|S|$ is least. Clearly, $S$ can not contain any nontrivial proper submodule $L$, because then $L$ being a member of $\mathcal{G}$ contradicts the minimality of $|S|$. This shows that $S$ is a simple submodule contained in $N$.
Finally, assume that the length of $M$ is finite. If $l(M) = 1$, then $M$ is simple and there is nothing to show. So, assume $l(M) > 1$. Let $N$ be any submodule of $M$. Let $\mathcal{H}$ be the collection of all nontrivial proper submodule of $N$. Choose $K \in \mathcal{H}$ such that $l(K)$ is minimal. If $K$ would not be simple and $P$ is a nontrivial proper submodule of $K$, then $l(P) < l(K)$ contradicts the minimality of $l(K)$. Thus, $K$ is a simple submodule contained in $N$.
\end{proof}
\begin{theorem}
The class of left $\mathrm{SM}$ modules properly contains the class of Artinian modules, locally finite modules, and modules of finite lengths.
\end{theorem}
\begin{proof}
Let $\mathcal{C}_{1}, \mathcal{C}_{2}, \mathcal{C}_{3}$, and $\mathcal{C}_{4}$ denote the class of all $\mathrm{SM}$ modules, modules of finite lengths, locally finite modules, and Artinian modules respectively. Consider the free $\mathbb{Z}_{p^{n}}$-module $M = \bigoplus_{\mathbb{N}} \mathbb{Z}_{p^{n}}$ on the set $\mathbb{N}$, where $p$ is some fixed prime and $n$ is a natural number such that $n > 1$. Let $N$ be any nontrivial proper submodule of $M$ and let $n \in N$. Observe that $|\langle n \rangle| \leq p^{n}$. If we choose a submodule $L$ of $\langle n \rangle$ such that $|L| = p$, then clearly $L$ is simple and contained in $N$. This shows that $M$ is an $\mathrm{SM}$ module.
As $l(\mathbb{Z}_{p^{n}}) = n$, therefore the divergence of the series $\sum_{i = 1}^{\infty} n$ shows that the length of the free $\mathbb{Z}_{p^{n}}$-module $\bigoplus_{\mathbb{N}} \mathbb{Z}_{p^{n}}$ is not finite. Hence, we have that $\mathcal{C}_{2} \subsetneq \mathcal{C}_{1}$.
For each natural number $i$, we define the submodules $L_{i} = \bigoplus_{\mathbb{N}} F_{j}$, where $F_{j} = \mathbb{Z}_{p^{n}}$ if $j > i$, otherwise, $F_{j} = \{0_{\mathbb{Z}_{p^{n}}}\}$. Now, observe that the strictly descending chain of submodules $L_{0} = M \supsetneq L_{1} \supsetneq L_{2} \supsetneq \cdots L_{s} \supsetneq \cdots$ do not terminate. Thus, $\mathcal{C}_{4} \subsetneq \mathcal{C}_{1}$.
Now, consider an infinite $R$-module $W$ of finite length $n$ (for example finite dimensional vector space). If $W_{0} = W \supsetneq W_{1} \supsetneq W_{2} \supseteq \cdots \supsetneq W_{n} = \{0_{W}\}$, then observe that $W = R(A)$, where $A = \{w_{i} \mid w_{i} \in W_{i} \setminus W_{i + 1}, 0 \leq i \leq n - 2\}$ and $R(A)$ is the module generated by the subset $A$. As $W \notin \mathcal{C}_{3}$, therefore, we get that $\mathcal{C}_{3} \subsetneq \mathcal{C}_{1}$.
\end{proof}
\begin{theorem}
Let $M$ be a left $\mathrm{SM}$ module over $R$. Then $\mathrm{Soc}(M)$ is proper essential in $M$ if and only if $M$ is not semisimple.
\end{theorem}
\begin{proof}
Assume first that $\mathrm{Soc}(M)$ is proper essential in $M$. Then, clearly we have $\mathrm{Soc}(M) \subsetneq M$. Hence, applying theorem \eqref{babama}, we conclude that $M$ is not semisimple.
Conversely, assume that $M$ is not semisimple. Again, applying theorem \eqref{babama}, we see that $M$ contains a proper essential submodule. Now, let $N$ be any submodule of $M$. As $N$ contains a simple submodule, we get that $\mathrm{Soc}(M) \cap N \neq \{0_{M}\}$. Further, as $\mathrm{Soc}(M)$ is contained in every proper essential submodule of $M$, we conclude that $\mathrm{Soc}(M)$ is proper essential in $M$.
\end{proof}
\section{Torsion-free modules over integral domains}
In general, it is not clear from theorem \eqref{jpmkj} that under what condition a left torsion-free module $M$ over an integral domain $R$ contains an essential submodule. For this purpose, we define the closure of a submodule $N$ of $M$ that gives some sort of measure in the sense that it is that maximal subspace of $M$ in which $N$ is essential.
\begin{definition}\label{madurga}
Let $M$ be a left torsion-free module over an integral domain $R$. For any submodule $N$ of $M$, we denote $\mathrm{Cl}_{M}(N)$ to mean the closure of $N$ in $M$ defined in the following way: $\mathrm{Cl}_{M}(N) = \{m \in M \mid rm \in N, \text{for some nonzero $r \in R$}\}$.
\end{definition}
Throughout this section, unless otherwise stated, we assume $M$ is a left torsion-free module over an integral domain $R$ and $D$ is the set of nonzero elements in $R$. The following proposition is quite straightforward and follows directly from definition \eqref{madurga}.
\begin{proposition}\label{debanandamaharaj}
Let $N, N_{1}, N_{2}, T$ are submodules of $M$ and let $f: M \rightarrow S$ be any homomorphism of modules. Then, we have the following:
$\mathrm{(a)}$ $\mathrm{Cl}_{M}(\{0_{M}\}) = \{0_{M}\}$ and $\mathrm{Cl}_{M}(N)$ is a submodule of $M$.
$\mathrm{(b)}$ $N \subseteq \mathrm{Cl}_{M}(N)$ and $\mathrm{Cl}_{M}(\mathrm{Cl}_{M}(N)) = \mathrm{Cl}_{M}(N)$.
$\mathrm{(c)}$ If $N_{1} \subseteq N_{2}$, then $\mathrm{Cl}_{M}(N_{1}) \subseteq \mathrm{Cl}_{M}(N_{2})$.
$\mathrm{(d)}$ $f(\mathrm{Cl}_{M}(T)) \leq \mathrm{Cl}_{S}(f(T))$.
\end{proposition}
\begin{proof}
$\mathrm{(a)}:$ $\mathrm{Cl}_{M}(\{0_{M}\})$ is the set of all those elements $m$ in $M$ such that $rm = 0$ for some nonzero $r \in R$. As $M$ is torsion-free, we see that $\mathrm{Cl}_{M}(\{0_{M}\}) = \{0_{M}\}$. If $m_{1}, m_{2}$ are in $\mathrm{Cl}_{M}(N)$, then $r_{1}m_{1} \in N$ and $r_{2}m_{2} \in N$ for some nonzero elements $r_{1}, r_{2}$ in $R$. Since $r_{1}r_{2}(m_{1} - m_{2}) \in N$ and $r_{1}r_{2} \neq 0$, we conclude that $(m_{1} - m_{2}) \in \mathrm{Cl}_{M}(N)$. Now, let $m \in \mathrm{Cl}_{M}(N)$ and $r \in R$ be such that $r \neq 0$ and $rm \in N$. For any $s \in R$, we have $r(sm) \in N$. Thus, $sm \in \mathrm{Cl}_{M}(N)$. Hence, $\mathrm{Cl}_{M}(N)$ is a submodule of $M$.
$\mathrm{(b)}:$ For any element $n \in N$, we have $1n \in N$, so $N \subseteq \mathrm{Cl}_{M}(N)$. It is evident that $\mathrm{Cl}_{M}(N) \subseteq \mathrm{Cl}_{M}(\mathrm{Cl}_{M}(N))$. Now, suppose $m \in \mathrm{Cl}_{M}(\mathrm{Cl}_{M}(N))$. This implies, for some nonzero $r \in R$, $rm$ lie in $\mathrm{Cl}_{M}(N)$. Again, for some nonzero $s \in R$, $(sr)m \in N$. As, $sr \neq 0$, we conclude that $m \in \mathrm{Cl}_{M}(N)$. Hence, $\mathrm{Cl}_{M}(\mathrm{Cl}_{M}(N)) \subseteq \mathrm{Cl}_{M}(N)$.
$\mathrm{(c)}:$ Let $m \in \mathrm{Cl}_{M}(N_{1})$. Then, $rm \in N_{1}$ for some nonzero $r \in R$. As $rm \in N_{2}$ as well, we conclude that $m \in \mathrm{Cl}_{M}(N_{2})$, that is, $\mathrm{Cl}_{M}(N_{1}) \subseteq \mathrm{Cl}_{M}(N_{2})$.
$\mathrm{(d)}:$ Let $s \in f(\mathrm{Cl}_{M}(T))$. Then, there exists $m \in \mathrm{Cl}_{M}(T)$ such that $f(m) = s$. Also, there exists $r \in R$ such that $r \neq 0$ and $rm \in T$. This shows that $rf(m) \in f(T)$, that is, $s \in \mathrm{Cl}_{S}(f(T))$. Hence, we get that $f(\mathrm{Cl}_{M}(T)) \leq \mathrm{Cl}_{S}(f(T))$.
\end{proof}
The question that we investigate is whether the closure operator commutes with arbitrary direct sums of submodules. We see that it is not true in general, but in bigger space, the closure operator commutes with arbitrary direct sum of submodules.
\begin{lemma}\label{joymadurga}
Let $N_{1}, N_{2}, \ldots, N_{n}$ are the submodules of $M$. If $D$ denotes the set of all nonzero elements in $R$, then $\mathrm{Cl}_{D^{-1}M}\big(\bigoplus_{i = 1}^{n} N_{i}\big) = \bigoplus_{i = 1}^{n} \mathrm{Cl}_{D^{-1}M}(N_{i})$.
\end{lemma}
\begin{proof}
It is enough to show that $\mathrm{Cl}_{D^{-1}M}(L \bigoplus S) = \mathrm{Cl}_{D^{-1}M}(L) \bigoplus \mathrm{Cl}_{D^{-1}M}(S)$ for any two submodules $L, S$ of $M$. Observe that if $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(L) \cap \mathrm{Cl}_{D^{-1}M}(S)$, then $\frac{r_{1}m}{d} \in L$ and $\frac{r_{2}m}{d} \in S$ for some nonzero $r_{1}, r_{2} \in R$. As $(r_{1}r_{2})\frac{m}{d} \in L \cap S$, we see that $\frac{m}{d} = 0_{M}$. Now, suppose $(\frac{m_{1}}{d_{1}} + \frac{m_{2}}{d_{2}}) \in \mathrm{Cl}_{D^{-1}M}(L) \bigoplus \mathrm{Cl}_{D^{-1}M}(S)$. We have then $\frac{r_{1}m_{1}}{d_{1}} \in L$ and $\frac{r_{2}m_{2}}{d_{2}} \in S$ for some $r_{1}, r_{2} \in R \setminus \{0\}$. As $(r_{1}r_{2})(\frac{m_{1}}{d_{1}} + \frac{m_{2}}{d_{2}}) \in L \bigoplus S$, we see that $(\frac{m_{1}}{d_{1}} + \frac{m_{2}}{d_{2}}) \in \mathrm{Cl}_{D^{-1}M}(L \bigoplus S)$. Conversely, let $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(L \bigoplus S)$. Therefore, for some nonzero $r \in R$, we have $\frac{rm}{d} \in L \bigoplus S$. Let $\frac{rm}{d} = l + s$. Then, $r(\frac{m}{d} - \frac{l}{r}) = s$, that is, $(\frac{m}{d} - \frac{l}{r}) \in \mathrm{Cl}_{D^{-1}M}(S)$. Also, observe that $\frac{l}{r} \in \mathrm{Cl}_{D^{-1}M}(L)$. Hence, we get that $\frac{m}{d} = \frac{l}{r} + (\frac{m}{d} - \frac{l}{r})$, that is, $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(L) \bigoplus \mathrm{Cl}_{D^{-1}M}(S)$.
\end{proof}
\begin{theorem}\label{joymakamakhyama}
If $\{N_{\alpha}\}_{\alpha \in \Lambda}$ be a family of submodules indexed by a nonempty set $\Lambda$, then $\mathrm{Cl}_{D^{-1}M}\big(\bigoplus_{\alpha \in \Lambda} N_{\alpha}\big) = \bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{D^{-1}M}(N_{\alpha})$. In particular, we have the following, $\mathrm{Cl}_{M}\big(\bigoplus_{\alpha \in \Lambda} N_{\alpha}\big) = M \cap \big(D^{-1}\big(\bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{M}(N_{\alpha})\big)\big)$.
\end{theorem}
\begin{proof}
Let $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}\big(\bigoplus_{\alpha \in \Lambda} N_{\alpha}\big)$. Then, for some $r \neq 0$, we have $\frac{rm}{d} \in \bigoplus_{\alpha \in \Lambda} N_{\alpha}$. Let $\frac{rm}{d} = \sum_{i = 1}^{n} n_{\alpha_{i}}$. This implies, $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(\bigoplus_{i = 1}^{n} N_{\alpha_{i}})$. Now, applying lemma \eqref{joymadurga}, we get that $\mathrm{Cl}_{D^{-1}M}(\bigoplus_{i = 1}^{n} N_{\alpha_{i}}) \leq \bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{D^{-1}M}(N_{\alpha})$. Hence, we obtain that $\frac{m}{d} \in \bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{D^{-1}M}(N_{\alpha})$. Conversely, suppose $\sum_{i = 1}^{n} \frac{m_{\alpha_{i}}}{d_{\alpha_{i}}} \in \bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{D^{-1}M}(N_{\alpha})$, where $\frac{m_{\alpha_{i}}}{d_{\alpha_{i}}} \in \mathrm{Cl}_{D^{-1}M}(N_{\alpha_{i}})$. Then, for each $i$ with $1 \leq i \leq n$, there exists some nonzero $r_{\alpha_{i}}$ such that $\frac{r_{\alpha_{i}}m_{\alpha_{i}}}{d_{\alpha_{i}}} \in N_{\alpha_{i}}$. But then $(r_{\alpha_{1}}r_{\alpha_{2}} \cdots r_{\alpha_{n}})\big(\sum_{i = 1}^{n} \frac{m_{\alpha_{i}}}{d_{\alpha_{i}}}\big) \in \bigoplus_{i = 1}^{n} N_{\alpha_{i}}$ shows that $(r_{\alpha_{1}}r_{\alpha_{2}} \cdots r_{\alpha_{n}})\big(\sum_{i = 1}^{n} \frac{m_{\alpha_{i}}}{d_{\alpha_{i}}}\big) \in \bigoplus_{\alpha \in \Lambda} N_{\alpha}$, that is, $\sum_{i = 1}^{n} \frac{m_{\alpha_{i}}}{d_{\alpha_{i}}} \in \mathrm{Cl}_{D^{-1}M}\big(\bigoplus_{\alpha \in \Lambda} N_{\alpha}\big)$.
For any submodule $N$ of $M$, if we show that $\mathrm{Cl}_{D^{-1}M}(N) = D^{-1}\mathrm{Cl}_{M}(N)$, then we have $\bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{D^{-1}M}(N_{\alpha}) = D^{-1}\big(\bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{M}(N_{\alpha})\big)$, since $D^{-1}$ distributes over arbitrary direct sum. So, suppose $\frac{m}{d} \in D^{-1}\mathrm{Cl}_{M}(N)$. As $m \in \mathrm{Cl}_{M}(N)$, we have $rm \in N$ for some nonzero $r \in R$. But then, we have $(rd)(\frac{m}{d}) \in N$. Hence, $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(N)$. Conversely, suppose $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(N)$. Then, $\frac{rm}{d} \in N$ for some nonzero $r \in R$. This implies, $rm \in N$, that is, $m \in \mathrm{Cl}_{M}(N)$. Hence, we get that $\frac{m}{d} \in D^{-1}\mathrm{Cl}_{M}(N)$. The proof is now completed by taking intersection on both sides of the equation $\mathrm{Cl}_{D^{-1}M}\big(\bigoplus_{\alpha \in \Lambda} N_{\alpha}\big) = D^{-1}\big(\bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{M}(N_{\alpha})\big)$ with $M$.
\end{proof}
\begin{theorem}\label{adisoktiaddyasokti}
If $N$ is a submodule of $M$, then $\mathrm{Soc}\big(\mathrm{Cl}_{M}(N)\big) = \mathrm{Cl}_{M}\big(\mathrm{Soc}(N)\big)$. In particular, they are equal to $\mathrm{Soc}(N)$.
\end{theorem}
\begin{proof}
If $S$ is a simple submodule of $N$ then clearly it is also simle in $\mathrm{Cl}_{M}(N)$. Conversely, if $S = \langle t \rangle$ is a simple submodule of $\mathrm{Cl}_{M}(N)$, then $rt \in N$ for some nonzero $r \in R$. As $\langle rt \rangle = \langle t \rangle$, we conclude that $S$ is a simple submodule of $N$. Hence, we conclude that $\mathrm{Soc}\big(\mathrm{Cl}_{M}(N)\big) = \mathrm{Soc}(N)$. Let $\mathrm{Soc}(N) = \bigoplus_{\alpha \in \Lambda} S_{\alpha}$, where $S_{\alpha}$ is a simple submodule of $N$ for each $\alpha \in \Lambda$. Now, using theorem \eqref{joymakamakhyama}, we get that $\mathrm{Cl}_{D^{-1}M}\big(\mathrm{Soc}(N)\big) = \bigoplus_{\alpha \in \Lambda} \mathrm{Cl}_{D^{-1}M}(S_{\alpha})$. As simple submodules of $M$ are simple submodules of $D^{-1}M$ and vice-versa, we see that $S_{\alpha}$ is a simple submodule of $D^{-1}M$ for each $\alpha \in \Lambda$. In this position, we wish to use the fact that for any left torsion-free module $Z$ (over some integral domain $B$) and for any simple submodule $L$ of $Z$, we have $\mathrm{Cl}_{Z}(L) = L$. To see this, let $z \in \mathrm{Cl}_{Z}(L)$ and $bz \in L$ for some nonzero $b \in B$. Let $L = \langle l \rangle$. Clearly, we have $\langle l \rangle = \langle bl \rangle = \langle bz \rangle$. If $bz = bl$, then $b(z - l) = 0$. As $Z$ is torsion-free and $b \neq 0$, we conclude that $z = l$, that is $z \in L$. If $bz \neq bl$, then for some nonzero $b^{\prime} \in B$, we have $bz = (b^{\prime}b)l$, that is, $b(z - b^{\prime}l) = 0$. Similar arguments as above shows that $z = b^{\prime}l$, that is, $z \in L$. Hence, $\mathrm{Cl}_{Z}(L) \subseteq L$. Now, applying proposition \eqref{debanandamaharaj}, we conclude that $\mathrm{Cl}_{Z}(L) = L$. Thus, we have that $\mathrm{Cl}_{D^{-1}M}\big(\mathrm{Soc}(N)\big) = \bigoplus_{\alpha \in \Lambda} S_{\alpha}$, that is, $\mathrm{Cl}_{D^{-1}M}\big(\mathrm{Soc}(N)\big) = \mathrm{Soc}(N)$. Now, $M \cap \mathrm{Cl}_{D^{-1}M}\big(\mathrm{Soc}(N)\big) = \mathrm{Cl}_{M}\big(\mathrm{Soc}(N)\big)$ shows that $\mathrm{Cl}_{M}\big(\mathrm{Soc}(N)\big) = \mathrm{Soc}(N)$. Hence, $\mathrm{Soc}\big(\mathrm{Cl}_{M}(N)\big) = \mathrm{Cl}_{M}\big(\mathrm{Soc}(N)\big) = \mathrm{Soc}(N)$.
\end{proof}
\begin{corollary}\label{amabjg}
Left torsion-free modules over integral domains are either semisimple or do not possess essential socles.
\end{corollary}
\begin{proof}
If $M$ is not semisimple then $\mathrm{Soc}(M) \neq M$ (see theorem \eqref{baba}). Now, if we set $N = M$ in theorem \eqref{adisoktiaddyasokti}, we see that $\mathrm{Cl}_{M}\big(\mathrm{Soc}(M)\big) = \mathrm{Soc}(M)$. Thus, the socle $\mathrm{Soc}(M)$ is not essential in $M$.
\end{proof}
\begin{theorem}\label{joybabamahadev}
For submodules $U, V$ of $M$, we have the following exact sequence: $0 \rightarrow D^{-1}\mathrm{Cl}_{M}(U) \xrightarrow{\alpha^{\prime}} D^{-1}\mathrm{Cl}_{M}(V) \xrightarrow{\beta^{\prime}} D^{-1}\mathrm{Cl}_{M/U}(V/U) \rightarrow 0$. Moreover, it splits if the sequence $0 \rightarrow U \xrightarrow{\alpha} V \xrightarrow{\beta} V/U \rightarrow 0$ splits.
\end{theorem}
\begin{proof}
Let
$\varphi: \mathrm{Cl}_{D^{-1}M/D^{-1}U}(D^{-1}V/D^{-1}U) \rightarrow \mathrm{Cl}_{D^{-1}M}(D^{-1}V)/\mathrm{Cl}_{D^{-1}M}(D^{-1}U)$
be a map defined by $\varphi: \frac{m}{d} + D^{-1}U \mapsto \frac{m}{d} + \mathrm{Cl}_{D^{-1}M}(D^{-1}U)$. Observe that we have $\frac{m}{d} + \mathrm{Cl}_{D^{-1}M}(D^{-1}U) \in \mathrm{Cl}_{D^{-1}M}(D^{-1}V)/\mathrm{Cl}_{D^{-1}M}(D^{-1}U)$. Because, for some nonzero $r \in R$, we have $\frac{rm}{d} + D^{-1}U = \frac{v}{d^{\prime}} + D^{-1}U$. But this implies $r(\frac{m}{d} - \frac{v}{rd^{\prime}}) \in D^{-1}U$, that is, $(\frac{m}{d} - \frac{v}{rd^{\prime}}) \in \mathrm{Cl}_{D^{-1}M}(D^{-1}U)$. As $\frac{v}{rd^{\prime}} \in \mathrm{Cl}_{D^{-1}M}(D^{-1}V)$, $\frac{m}{d} + \mathrm{Cl}_{D^{-1}M}(D^{-1}U)$ is an element of $\mathrm{Cl}_{D^{-1}M}(D^{-1}V)/\mathrm{Cl}_{D^{-1}M}(D^{-1}U)$. To see that $\varphi$ is well-defined, observe that for elements $\frac{m}{d} + D^{-1}U, \frac{m^{\prime}}{d^{\prime}} + D^{-1}U$ in $\mathrm{Cl}_{D^{-1}M/D^{-1}U}(D^{-1}V/D^{-1}U)$, if we have $(\frac{m}{d} - \frac{m^{\prime}}{d^{\prime}}) \in D^{-1}U$, then $\frac{m}{d} + \mathrm{Cl}_{D^{-1}M}(D^{-1}U) = \frac{m^{\prime}}{d^{\prime}} + \mathrm{Cl}_{D^{-1}M}(D^{-1}U)$, since proposition \eqref{debanandamaharaj} yields $D^{-1}U \subseteq \mathrm{Cl}_{D^{-1}M}(D^{-1}U)$. It is easy to see that $\varphi$ is a module homomorphism. If $\frac{m}{d} + D^{-1}U \in \mathrm{Cl}_{D^{-1}M/D^{-1}U}(D^{-1}V/D^{-1}U)$ be such that $\frac{m}{d} \in \mathrm{Cl}_{D^{-1}M}(D^{-1}U)$, then $\frac{rm}{d} \in D^{-1}U$ for some nonzero $r \in R$, that is, $\frac{m}{d} \in D^{-1}U$. This shows that $\varphi$ is injective. Finally, for any $\frac{m}{d} + \mathrm{Cl}_{D^{-1}M}(D^{-1}U) \in \mathrm{Cl}_{D^{-1}M}(D^{-1}V)/\mathrm{Cl}_{D^{-1}M}(D^{-1}U)$, we have $\frac{rm}{d} \in D^{-1}V$ for some nonzero $r \in R$. As $r(\frac{m}{d} + D^{-1}U) \in D^{-1}V/D^{-1}U$, we conclude that $(\frac{m}{d} + D^{-1}U) \in \mathrm{Cl}_{D^{-1}M/D^{-1}U}(D^{-1}V/D^{-1}U)$, that is, $\varphi$ is surjective. Hence, $\varphi$ is an isomorphism. It is now easy to see that $\mathrm{Cl}_{D^{-1}M/D^{-1}U}(D^{-1}V/D^{-1}U) = D^{-1}\mathrm{Cl}_{M/U}(V/U)$ and $\mathrm{Cl}_{D^{-1}M}(D^{-1}V)/\mathrm{Cl}_{D^{-1}M}(D^{-1}U) = D^{-1}(\mathrm{Cl}_{M}(V)/\mathrm{Cl}_{M}(U))$. This completes the first part of the theorem. Applying lemma \eqref{joymadurga}, we can conclude the second part.
\end{proof}
We close this section with the following theorem that can also be viewed as an application of corollary \eqref{amabjg}.
\begin{theorem}
Left modules over integral domains with nonempty torsion-free parts are either semisimple or do not possess essential socles.
\end{theorem}
\begin{proof}
If $M$ is semisimple, there is nothing to show. So, assume that $M$ is not semisimple. This implies $M$ has a proper essential submodule $E$ (see theorem \eqref{baba}). Now, applying lemma \eqref{gmkj}, we see that $(E + \mathrm{Tor}(M))/\mathrm{Tor}(M)$ is proper essential in $M/\mathrm{Tor}(M)$. Now, if $\mathcal{P}$ be the collection of all essential submodules of $M/\mathrm{Tor}(M)$ and if we denote $\overline{D} = D/\mathrm{Tor}(M)$ for $D/\mathrm{Tor}(M) \in \mathcal{P}$, then from corollary \eqref{amabjg}, we can conclude that the submodule $\bigcap_{\overline{D} \in \mathcal{P}} \overline{D}$ is not proper essential in $M/\mathrm{Tor}(M)$. Since we have $\big(\bigcap_{\overline{D} \in \mathcal{P}} D\big)/\mathrm{Tor}(M) \leq \bigcap_{\overline{D} \in \mathcal{P}} \overline{D}$, therefore, we can conclude from lemma \eqref{gmkj} that the submodule $\bigcap_{\overline{D} \in \mathcal{P}} D$ can not be proper essential in $M$. Further, since $\mathrm{Soc}(M)$ is a submodule of $\bigcap_{\overline{D} \in \mathcal{P}} D$, we conclude that $M$ can not possess an essential socle.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
\IEEEPARstart{I}{n} our pursuit of intelligent machine perception, it is crucial to endow systems, like autonomous vehicles and robots, with an awareness of the scene content beyond their immediately visible field-of-view (FoV). Simply put, the system should be able to hallucinate its surroundings, and associate each prediction with certain confidence, which could help it plan early and decide accordingly.
\marginpar{NEW}{For example, when a moving camera is turning right at a corner in a road-bound scene, the right blind part is completely unobserved but can be reasonably hallucinated. Or when observing a car in a neighboring lane over time, wide FoV synthesis can help to reason about its future position, also beyond its immediately visible FoV. Such wide FoV hallucination ability can benefit vision-based navigation~\cite{wang2019reinforced}, exploration~\cite{jayaraman2018learning}, and augmented-reality telerobotics system~\cite{makhataeva2020augmented}.}
To realize this idea, we can draw inspiration from how humans use vision to relate themselves to the world around them. Humans clearly have a situational awareness that goes beyond their narrow FoV. On the one hand, this is grounded in a capability to propagate local scene content from past observations (\marginpar{NEW}{\eg} anticipate the future position of a previously observed building based on the car's trajectory when driving). On the other hand, it is due to an ability to hallucinate global scene content for unobserved regions based on the scene's context (\marginpar{NEW}{\eg} the unobserved side views in a driving scene are likely to contain trees if the car crosses a forest area). Most importantly, humans can typically assign a degree of confidence in these judgments to quantify their intuition.
Motivated by these observations, in this paper, we tackle the problem of \textit{FoV extrapolation}. The goal is to leverage information from a video sequence with narrow FoV (including the present and few past frames) in order to infer the (present) scene at a wider FoV (see Fig.~\ref{fig:teaser}). There are several challenges associated with this problem. (1) A large image size discrepancy between the input narrow FoV frames and the output wide FoV frame has to be bridged, and the results should be temporally consistent. (2) Some areas in the wide FoV frame may change significantly or even not appear in any of the past narrow FoV frames. For example, far away objects in the past narrow FoV frames need upscaling or novel view synthesis, and some occluded or unobserved regions need to be inpainted. Thus, lots of details need to be hallucinated in the wide FoV frame.
(3) There is ambiguity existing between the narrow FoV observations and the wide FoV ground truth.
The ambiguity mainly comes from two sources: the unobserved information and possible 3D estimation errors.
In particular, the pixels in the wide FoV frame can be roughly divided into four types (see Fig.~\ref{fig:ambiguity}): (a) the observed narrow FoV pixels in the present frame (no ambiguity); (b) the propagated pixels from past frames with accurate propagation (low ambiguity); (c) the propagated pixels from past frames with noisy propagation (medium ambiguity); (d) the unobserved regions (high ambiguity). When the ambiguity is high, strong enforcement of pixel reconstruction may mislead the training process. In contrast, perceptual and adversarial losses can be more suitable to predict a plausible outcome.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{figures/ambiguity5.pdf}\\\
\caption{\marginpar{NEW}{Ambiguity illustration. Best viewed in the digital form.}
}
\label{fig:ambiguity}
\end{figure}
To address these challenges, we propose a temporally consistent FoV extrapolation framework called \textit{FoV-Net}, which consists of two stages (see Fig.~\ref{fig:framework}). A coordinates generation stage propagates past narrow FoV frames into the present wide FoV frame by leveraging 3D scene information (addressing challenge 1). A frame aggregation stage combines the multi-frame propagated information, simultaneously hallucinating fine details and unobserved scene parts (addressing challenges 2\&3). Specifically, in the frame aggregation stage, we introduce an Attention-based Feature Aggregation (AFA) module to better fuse the propagated multi-frame information on the feature level, and a Gated Self-Attention (GSA) module to handle the discussed ambiguities and improve the generation quality (addressing challenges 2\&3). Finally, we introduce an uncertainty mechanism to interpret the hallucination uncertainty at each pixel and guide the learning by reducing supervision ambiguity (addressing challenge 3). Such hallucination uncertainty is rarely discussed in the image synthesis field, but is quite important for practical downstream decision-making applications. Fig.~\ref{fig:teaser} gives an example of our FoV extrapolation with its associated uncertainty.
\section{Related Work}
\label{sec:related}
\myparagraph{Video-based image synthesis.}
This problem exists in various forms in the literature, including video inpainting, video extrapolation, novel view synthesis, future video prediction, video-to-video synthesis, \emph{etc}\onedot Video inpainting~\cite{huang2016temporally,wang2019video,kim2019deep,chang2019free,LGTSM,Gao-ECCV-FGVC} aims to hallucinate the missing pixels through warping or generate the missing pixels conditioned on the neighboring (in spatial or temporal dimensions) visible pixels. The typical setting is to utilize the past, present, and future frames to inpaint the missing pixels in the present frame, all within narrow FoV. Video extrapolation~\cite{lee2019video,choi2020deep,zhang2014parallax,guo2016joint} usually adopts 2D or 3D geometry-based image warping and stitching techniques to blend the observed pixels of adjacent narrow FoV frames in order to extend the FoV, but totally ignores any unobserved pixels and object view changes. Novel view synthesis~\cite{park2017transformation,sun2018multi,hedman2018deep,choi2019extreme,chen2019nvs,flynn2019deepview} aims to generate images of a given object or scene from different viewpoints by blending the observed pixels, as well as hallucinating a few missing pixels mainly for dis-occlusion. When applied to scenes, it is heavily reliant on highly accurate multi-view geometry to produce good results. Future video prediction~\cite{denton2018stochastic,byeon2018contextvp,liu2017voxelflow,lotter2016deep,gao2019disentangling,liu2018dyan} focuses on hallucinating future frames conditioned on the past and present frames, all within narrow FoV. Undoubtedly, this task entails higher uncertainty in the predictions. Video-to-video synthesis~\cite{wang2018video,chen2017coherent,shi2016real,bansal2018recycle} mainly transfers the appearance while preserving the structure of the input (\marginpar{NEW}{\eg} semantic maps). Thus the input and output are usually well aligned and within narrow FoV.
\marginpar{NEW}{Unlike} the existing forms, our goal is to infer the present scene at a wider FoV, including the observed and unobserved pixels, conditioned on the past and present narrow FoV observations.
\marginpar{NEW}{Note that, when the camera is moving forward, most out-of-view regions of the present narrow FoV frame are actually future predictions as far as observations in the past frames are concerned. While if the camera is turning around at a corner, part out-of-view regions may be totally unobserved before, which makes the problem more challenging.}
The object size and view may change significantly, leading to large unobserved regions.
To the best of our knowledge, none of the existing video-based image synthesis forms fully covers the needs of our intricate problem.
\myparagraph{Attention.}
The success of self-attention models in natural language processing has inspired various applications in the computer vision field, such as in image recognition~\cite{hu2019local,SAN}, image synthesis~\cite{SAGAN,brock2018large,parmar2018image}, video prediction~\cite{jia2016dynamic,wang2018non}, and imitation learning~\cite{icra2020attentive}. Self-attention can be formulated as locally adaptable convolutional layers
with different weights for different types of image regions~\cite{SAN}. In our FoV extrapolation problem, we observe that such local adaptability is essential in terms of hallucination quality, since different image regions have different characteristics (see Fig.~\ref{fig:ambiguity}). To improve the hallucination quality, we propose a novel gated self-attention module, motivated by the success of gated convolution in image inpainting~\cite{deepfillv2} and that of self-attention in image recognition~\cite{SAN}.
Note that, the gated self-attention concept is not new and has been introduced in natural language processing~\cite{dhingra2016gated,tran2017generative,zhao2018paragraph,lai2019gated}. However, in this work, we extend it to the video domain.
Additionally, in order to better aggregate the propagated multi-frame information, we propose an Attention-based Feature Aggregation (AFA) module -- which is not to be confused with self-attention used above -- that makes our framework more robust to propagation errors and improves the generation quality (see Sec.~\ref{sec:sa}).
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{figures/framework1.pdf}\\
\caption{The proposed \textit{FoV-Net} framework. Left: the coordinate generation stage (Sec.~\ref{sec:overview}), which estimates the scene-level rigid flow, governed by the camera motion, and uses it to calculate the coordinates (\marginpar{NEW}{\emph{i.e.,}\enspace} pixel displacements). Right: the frame aggregation stage (Sec.~\ref{sec:overview}), which utilizes the generated coordinates to propagate past frames information on a multi-scale feature level -- denoted as \red{red} dots -- and then aggregates the propagated features with an Attention-based Feature Aggregation module (Sec.~\ref{sec:sa}). To synthesize the final result $O_t$, a U-Net architecture is adopted to in/out-paint the missing regions, where a Gated Self-Attention module (Sec.~\ref{sec:sa}) is introduced to handle different ambiguities for better generation quality. Concurrently, an uncertainty map $U_t$ is jointly estimated to interpret the hallucination uncertainty at each pixel and guide the learning by reducing supervision ambiguity (Sec.~\ref{sec:uncertainty}). For $I_t$, we use the identical coordinates as $e_t$, namely, the features are not changed after warping. The coordinate $e_i$ is also used in the GSA blocks to warp the past hidden states, but we omit the arrows here. The discriminator networks \marginpar{NEW}{for adversarial losses} are also omitted for clarity. Due to FoV-Net's recurrent nature, note that previous outputs ${\{O_{t-i}\}}_{i=1,\dots,j}$ and ${\{U_{t-i}\}}_{i=1,\dots,j}$ become future inputs ${\{I_{t-i}\}}_{i=1,\dots,j}$ for temporal coherency purposes. \marginpar{NEW}{Best viewed in the digital form.}
}
\label{fig:framework}
\end{figure*}
\myparagraph{Uncertainty estimation.}
Reasoning about \marginpar{NEW}{the} uncertainty of neural network prediction is essential for practical decision-making applications~\cite{loquercio2020general}. Although uncertainty estimation has been proved to be effective for several computer vision tasks, including object detection~\cite{he2019bounding}, semantic segmentation~\cite{multitaskUncertainty,huang2018efficient}, depth estimation~\cite{depthUncertainty,yang2020d3vo} and optical flow~\cite{flowUncertainty}, it remains largely unexplored in the image synthesis literature. In this paper, we propose a hallucination uncertainty estimation mechanism, which not only enables the prediction of uncertainty, but also guides the learning by reducing supervision ambiguity. In general, there are different ways to estimate the uncertainty, including empirical estimation, predictive estimation, and Bayesian estimation (see ~\cite{depthUncertainty,loquercio2020general} for a comprehensive survey). Among them, the predictive estimation is \marginpar{NEW}{desirable} due to its effectiveness and efficiency, and has been explored in several computer vision tasks~\cite{depthUncertainty,flowUncertainty,multitaskUncertainty}. As a positive side effect, when integrated with the training objective in our problem, the predictive uncertainty can naturally weight the loss functions spatially to reduce the supervision ambiguity.
\section{FoV-Net framework}
\label{sec:framework}
\subsection{System overview}
\label{sec:overview}
Given a present narrow FoV frame $I_t$ and $k$ past narrow FoV frames ${\{I_{t-i}\}}_{i=1,\dots,k}$ \marginpar{NEW}{($k=5$ in our experiment)}, our goal is to synthesize the present wide FoV frame $O_t$ -- close to ground truth $W_t$ -- and predict the hallucination uncertainty $U_t$. In addition, the adjacent synthesized results $O_{t-1}$ and $O_{t}$ should be temporally consistent. To achieve these, we propose a two-stage recurrent framework (Fig.~\ref{fig:framework}) consisting of a coordinate generation stage and a frame aggregation stage, coupled with a hallucination uncertainty mechanism (Sec.~\ref{sec:uncertainty}). The coordinate generation stage is designed to estimate the scene-level rigid flow, governed by the camera motion, and use it to generate coordinates (\marginpar{NEW}{\emph{i.e.,}\enspace} pixel positions) in order to spatially propagate information from the past \marginpar{NEW}{narrow FoV} frames. The frame aggregation stage is designed to aggregate the past narrow FoV frames $I_{t-k},\dots,I_{t-1}$ and present narrow FoV frame $I_{t}$ into one wide FoV image, as well as hallucinate the unobserved missing regions. To enforce temporal coherency, we use a simple recurrent feed-forward strategy: replace the narrow FoV inputs ${\{I_{t-i}\}}_{i=1,\dots,j}$ with the previous outputs ${\{O_{t-i}\}}_{i=1,\dots,j}$, and feed the corresponding previous uncertainty ${\{U_{t-i}\}}_{i=1,\dots,j}$ by channel-wise concatenation ($j=2$ in our settings). We analyze each stage of our framework below.
\myparagraph{Coordinate generation stage}, that builds upon Monodepth2~\cite{monodepth2}, consists of a depth network $\mathcal{D}_{\theta_\mathcal{D}}$ and a relative camera pose network $\mathcal{P}_{\theta_\mathcal{P}}$\footnote{The subscripts $\{\theta_\mathcal{D}, \theta_\mathcal{P}, \theta_\mathcal{A}, \theta_\mathcal{Q}, \theta_\mathcal{T}\}$ are network parameters. We regularly omit the subscripts for brevity.}. During training, \marginpar{NEW}{$\mathcal{P}$} takes a pair of two \marginpar{NEW}{temporally} adjacent frames as input and outputs the relative camera pose, and \marginpar{NEW}{$\mathcal{D}$} takes one frame of the pair as input and outputs its depth. During inference, however, we do not have access to the wide FoV frame which is required by the inverse warping operation~\cite{jaderberg2015spatial} in order to propagate the pixels between two frames.
To address this, we design a forward warping strategy to propagate the past narrow FoV frames to the present wide FoV frame. We first utilize the depth maps from the past narrow FoV frames to calculate the rigid flow $f^{rig}_{t\rightarrow i}(\hat{e}_{t\rightarrow i})$ from the present frame $I_t$ to the past frame $I_i$, using Eq.~\ref{eq:rigid_flow}.
\begin{align}
f^{rig}_{t\rightarrow i}(\hat{e}_{t\rightarrow i})=K\mathrm{T}_{i\rightarrow t}\mathcal{D}_i(c_i)K^{\raisebox{.2ex}{$\scriptscriptstyle-1$}} c_i - c_i,
\label{eq:rigid_flow}
\end{align}
where $K$ denotes the camera intrinsic matrix, $\mathrm{T}_{i\rightarrow t}$ denotes the relative camera pose, and $c_i$ denotes homogeneous coordinates of pixels in frame $I_i$.
Then, using the calculated flow, we compute the spatial mapping, \marginpar{NEW}{\emph{i.e.,}\enspace} coordinate map $\hat{e}_{t\rightarrow i}$, that \marginpar{NEW}{spatially matches the pixel positions of present frame $I_t$ to the corresponding ones of past frame $I_i$}. Finally, we reverse this correspondence, \marginpar{NEW}{\emph{i.e.,}\enspace} $e_i=\text{reverse}(\hat{e}_{t\rightarrow i})$, which now corresponds to the spatial positions from the past frame $I_i$ to the present frame $I_t$. This coordinate map $e_i$ will be used to propagate features in the frame aggregation stage using bilinear sampling.
\myparagraph{Frame aggregation stage} is designed to aggregate the past narrow FoV frames $I_{t-k},\dots,I_{t-1}$ and present narrow FoV frame $I_{t}$ as well as previous wide FoV results ${\{O_{t-i}\}}_{i=1,\dots,j}$, ${\{U_{t-i}\}}_{i=1,\dots,j}$, into one wide FoV image, simultaneously hallucinating unobserved missing regions. It contains aggregation network $\mathcal{A}_{\theta_\mathcal{A}}$, image discriminator network $\mathcal{Q}_{\theta_\mathcal{Q}}$, and temporal discriminator network $\mathcal{T}_{\theta_\mathcal{T}}$. The aggregation network $\mathcal{A}$ first extracts a residual multi-scale feature pyramid (with $N=3$ levels) from each frame using an encoder with shared weights, and propagates the multi-scale features using the computed coordinates $e_i$. Then, an Attention-based Feature Aggregation (AFA) module (Sec.~\ref{sec:sa}) aggregates the propagated features. To synthesize the final result based on the aggregated features, a U-Net decoder is designed, where we introduce the Gated Self-Attention (GSA) module (Sec.~\ref{sec:sa}) to adaptively handle ambiguities and improve the generation quality.
\subsection{Self-attention mechanism}
\label{sec:sa}
\textbf{Attention-based Feature Aggregation (AFA).} In Fig.~\ref{fig:attn_agg_module}, to aggregate the propagated features, each set of propagated \marginpar{NEW}{multi-scale features maps (Fig.~\ref{fig:framework} right)} are fed into a convolution layer followed by softmax normalization to predict frame-wise spatial attention maps (\marginpar{NEW}{\emph{i.e.,}\enspace} one
\begin{wrapfigure}{l}{0.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{figures/attn_module.pdf}\\
\caption{AFA module. Same color coding as Fig.~\ref{fig:framework}}
\label{fig:attn_agg_module}
\end{wrapfigure}
channel attention map for each frame).
Then, the propagated feature maps are multiplied by the attention maps and summed across all frames. This attention-based aggregation module can learn to select the useful features among these frames to address the issues caused by depth/pose errors and frame inconsistency.
\myparagraph{Gated Self-Attention (GSA).}
To make our model adaptable to observations with different ambiguity levels, we encompass self-attention and gating mechanisms to construct a Gated Self-Attention (GSA) module.
Here, we adopt a patch-wise self-attention block introduced in~\cite{SAN}, which efficiently computes local attention weights that vary over spatial coordinates and channels instead of sharing weights to convolve the whole feature maps like conventional CNN.
It has the form:
\begin{align}
\mathbf{y}_i = \sum_{j \in \mathcal{R}(i)} \alpha(\mathbf{x}_{\mathcal{R}(i)})_j \odot \beta(\mathbf{x}_j), \label{eq:SA}\\
\text{where} \hspace{5mm}
\alpha(\mathbf{x}_{\mathcal{R}(i)}) = \gamma(\delta(\mathbf{x}_{\mathcal{R}(i)}))). \nonumber
\end{align}
The function $\beta$ produces the feature vectors $\beta(\mathbf{x}_j)$ that are weighted summarized by the adaptive weight vectors $\alpha(\mathbf{x}_{\mathcal{R}(i)})_j$.
The tensor $\mathbf{x}_{\mathcal{R}(i)}$ is the patch of feature vectors in a 7$\times$7 footprint $\mathcal{R}(i)$. $\alpha(\mathbf{x}_{\mathcal{R}(i)})_j$ is the attention vector at location $j$ in tensor $\alpha(\mathbf{x}_{\mathcal{R}(i)})$, corresponding spatially to the vector $x_j$ in $\mathbf{x}_{\mathcal{R}(i)}$.
Functions $\beta$ and $\gamma$ are mappings implemented via one convolution layer, respectively.
The function $\delta$ combines the feature vectors $x_j$ from the patch $\mathbf{x}_{\mathcal{R}(i)}$ implemented via a concatenation operation $\delta(\mathbf{x}_{\mathcal{R}(i)})) = [\phi(\mathbf{x}_i),[\psi(\mathbf{x}_j)]_{\forall j \in \mathcal{R}(i)}]$, where $\phi$ and $\psi$ are mappings implemented via one convolution layer, respectively. $\odot$ denotes the Hadamard product.
To reduce the impact of vanishing gradients, we wrap the self-attention block in a residual structure, \marginpar{NEW}{\emph{i.e.,}\enspace} $\mathbf{z} = \text{Conv}_r(\mathbf{y}) + \mathbf{x}$. We then equip our self-attention block with a gating mechanism to learn to control the information flows of different ambiguities,
formulated as:
\begin{align}
\mathbf{g} = \text{sigmoid}(\text{Conv}_g(\mathbf{z})) \odot \text{tanh}(\text{Conv}_a(\mathbf{z})),
\label{eq:gate}
\end{align}
where $\text{sigmoid}(\text{Conv}_g(\mathbf{z}))$ and $\text{tanh}(\text{Conv}_a(\mathbf{z})$ denote the gate and feature activation, respectively.
\marginpar{NEW}{$\text{Conv}_r$, $\text{Conv}_g$, $\text{Conv}_a$ are 2D convolution layers, and the subscriptions stand for residual, gate, and attention, respectively.}
\subsection{Uncertainty mechanism}
\label{sec:uncertainty}
We design an uncertainty mechanism to not only predict the interpretable hallucination uncertainty, but also guide the learning by reducing supervision ambiguity.
We draw inspiration from prior work where predicting the data-dependent uncertainty helped in tempering the training objective, \marginpar{NEW}{\eg} by attenuating the effect from erroneous labels in~\cite{flowUncertainty}, or by automatically balancing the loss weighting in~\cite{multitaskUncertainty}.
To realize this idea, one may design a heuristic weighting map, like the spatially discounted reconstruction loss~\cite{yu2018generative}, but such approaches are ad-hoc and cannot adapt to different scenes automatically.
Instead, we opt to jointly learn an uncertainty map during training, which serves as a probabilistic interpretation of our model.
We build upon predictive estimation~\cite{nix1994estimating}, and infer the mean and variance of the distribution $p(O_t|I_{t-i},\mathbb{D})$, where $i=0,\dots,K$, and $\mathbb{D}$ denotes the whole dataset.
The network is trained by log-likelihood maximization (\marginpar{NEW}{\emph{i.e.,}\enspace} negative log-likelihood minimization) and the distribution can be modelled as Laplacian (\marginpar{NEW}{\emph{i.e.,}\enspace} corresponding to L1 loss) or Gaussian (\marginpar{NEW}{\emph{i.e.,}\enspace} corresponding to L2 loss) respectively.
The negative log-likelihood formulation is:
\begin{align}
L_{1Log} = \frac{\|\mu(x)-x^*\|_1}{\sigma(x)} + \log\sigma(x),
\label{eq:loss_u_log}
\end{align}
where $\mu(x)$ and $\sigma(x)$ are the network outputs, encoding mean and variance of the distribution.
Here, we adopt L1 loss for pixel level reconstruction and reformulate Eq.~\ref{eq:loss_u_log} integrating uncertainty $U_t$ as:
\begin{equation}
\begin{split}
L_{1U}^{\theta_A} = \mathbb{E}\left[ \left(\frac{\|O_t-W_t\|_1}{U_t}\right) \odot M \right. \\
\left. + \|O_t-W_t\|_1 \odot (1- M) + U_t \vphantom{\int_1^2}\right],
\label{eq:loss_L1u}
\end{split}
\end{equation}
where $O_t$, $W_t$ are the predicted wide FoV RGB image and the ground truth RGB image (3-channel), respectively. $U_t$ is the estimated hallucination uncertainty map (1-channel). $M$ denotes the mask for out-of-narrow-FoV regions. Within the narrow FoV region $(1- M)$, the L1 is not weighted by uncertainty $U_t$, as new{the} present narrow FoV region has been observed.
Note that, to make the uncertainty term $U_t$ more interpretable and stabilize the training process, we constrain $U_t$ in the range (0,1) with a sigmoid function and modify the regularization term from $\log U_t$ to $U_t$ for gradient stabilization. We found that such modification also leads to better performance.
Additionally, in our recurrent framework, the previous predicted uncertainty $\{U_{t-i}\}_{i=1,\dots,j}$ are also used in future input to act as a confidence signal.
\subsection{Losses}
\label{sec:loss}
\myparagraph{Coordinate generation losses.} Following~\cite{monodepth2}, the objective in this stage is a masked photometric loss $\nu \odot L_{photo}$ and an edge-aware smoothness loss $L_{smooth}$, summarized below:
\begin{align}
L_{CG}^{\theta_D,\theta_P} &= \nu \odot L_{photo}^{\theta_D,\theta_P} + \lambda_s L_{smooth}^{\theta_D,\theta_P},
\label{eq:loss_CG}\\
L_{photo}^{\theta_D,\theta_P} &=\min\limits_{t'}pe(I_t,I_{t'\rightarrow t}), \label{eq:loss_pixel}\\
pe(I_a,I_b) &= \frac{\alpha}{2}(1-\text{SSIM}(I_a,I_b)) + (1-\alpha){\|I_a-I_b\|}_1,
\nonumber\\
L_{smooth}^{\theta_D,\theta_P} &=|\nabla_x d^*_t|e^{-|\nabla_x I_t|} + |\nabla_y d^*_t|e^{-|\nabla_y I_t|},
\label{eq:loss_smooth}
\end{align}
where $\nu=[\min\limits_{t'}pe(I_t,I_{t'\rightarrow t})]$ acts as an auto-mask for suppressing the effect of objects moving at similar speeds to the camera, and $[]$ is the Iverson bracket. $d^*=d_t/\overline{d_t}$ is the mean-normalized inverse depth. \marginpar{NEW}{SSIM is the structural similarity~\cite{ssim}.}
We use hyper-parameters $\alpha=0.85$ and $\lambda_s=0.001$ as in~\cite{monodepth2}.
\myparagraph{Frame aggregation losses.}
The objective of the frame aggregation stage has four loss terms: uncertainty-aware L1 reconstruction loss $L_{1U}^{\theta_A}$, perceptual reconstruction loss $L_{perc}^{\theta_A}$~\cite{lpips}, adversarial loss $L_{adv}^{\theta_A,\theta_Q}$~\cite{lsgan}, and temporal adversarial loss $L_{advT}^{\theta_A,\theta_T}$. The reconstruction losses $L_{1U}^{\theta_A}$ and $L_{perc}^{\theta_A}$ are used to regress the output towards the target ground truth, while the adversarial losses $L_{adv}^{\theta_A,\theta_Q}$ and $L_{advT}^{\theta_A,\theta_T}$ are used to encourage image photo-realism and temporal coherence. The formulations are as follows:
\begin{align}
\begin{split}
L_{FA}^{\theta_\mathcal{A},\theta_\mathcal{Q},\theta_\mathcal{T}} &= \lambda_1 L_{1U}^{\theta_\mathcal{A}}(O_t,W_t) + \lambda_2 L_{perc}^{\theta_\mathcal{A}}(O_t,W_t) \\
&+ L_{adv}^{\theta_\mathcal{A},\theta_\mathcal{Q}}(O_t,W_t) + L_{advT}^{\theta_\mathcal{A},\theta_\mathcal{T}}(O_t,W_t),
\label{eq:loss_FA}
\end{split}\\
L_{perc}^{\theta_\mathcal{A}} &= \mathbb{E}\big[\|\phi(O_t)-\phi(W_t)\|_2^2\odot (1+ M))\big],
\label{eq:loss_LPIPS}\\
\begin{split}
L_{adv}^{\theta_\mathcal{A},\theta_\mathcal{Q}} &= {\mathbb{E}}\big[(Q(W_t, M))^2\odot (1+ M)\big] \\ &+{\mathbb{E}}\big[(Q(O_t, M)-1)^2\odot (1+ M)\big],
\label{eq:loss_GAN}
\end{split}\\
\begin{split}
L_{advT}^{\theta_\mathcal{A},\theta_\mathcal{T}} &= {\mathbb{E}}\big[(\marginpar{NEW}{T}(\{W_{t-i}\}_{i=0,\dots,j}, M))^2\odot (1+ M)\big] \\ &+{\mathbb{E}}\big[(\marginpar{NEW}{T}(\{O_{t-i}\}_{i=0,\dots,j}, M)-1)^2\odot (1+ M)\big],
\label{eq:loss_GAN_temp}
\end{split}
\end{align}
where $\lambda_1=3$, $\lambda_2=10$. $\phi$ is a VGG network~\cite{VGG}.
\section{Experiments}
\label{sec:exp}
To evaluate FoV-Net w.r.t. its FoV extrapolation capabilities, we provide both qualitative and quantitative results.
For more video results and training/implementation details, please visit the supplementary materials.
\myparagraph{Dataset.}
Our method is evaluated on two widely used datasets: raw KITTI sequences~\cite{kitti} using the split from Eigen~\emph{et al}\onedot\cite{eigen2014depth}, and Cityscapes sequences~\cite{Cordts2016Cityscapes}.
To reduce the redundant information in videos, we downsample the frame rate to 1/2 and 1/3 for KITTI and Cityscapes sequences.
Therefore, we have 39350/4382 and 59526/14992 train/val frames on KITTI and Cityscapes, respectively.
During testing, we prepare each video sample with 10 and 5 target frames for KITTI and Cityscapes.
Totally, there are 635 (6350 frames) and 1525 test video samples (7625 frames) in KITTI and Cityscapes, respectively.
We use both monocular left and right camera sequences for training and validation, but only use left camera data for testing.
In each forward pass, we use 6 successive frames ($k=5$ past + 1 present) as input.
The narrow FoV ratio is set to 0.5 in our experiments, \marginpar{NEW}{\emph{i.e.,}\enspace} all 6 frames are cropped 25\% on both left and right sides to mimic the narrow FoV.
\begin{figure*}
\centering
\begin{subfigure}{.12\linewidth}
\centering
\renewcommand{\arraystretch}{2.6}
\setlength\tabcolsep{1pt
\begin{tabular}{c}
Inputs \\
Mono~\cite{monodepth2} \\
VF~\cite{liu2017voxelflow} \\
LGTSM~\cite{LGTSM} \\
\makecell{Mono-LGTSM} \\
Ours\\
\makecell{Uncertainty}\\
\makecell{Target}\\
\end{tabular}
\end{subfigure}
\begin{subfigure}{.41\linewidth}
\raisebox{-\totalheight}{
\includegraphics[height=9.cm]{figures/kitti_result/kitti-0.png}\\
}
\end{subfigure}
\hspace{7mm}
\begin{subfigure}{.41\linewidth}
\raisebox{-\totalheight}{
\includegraphics[height=9.cm]{figures/cs_result/cs-0.png}\\
}
\end{subfigure}
\caption{The qualitative comparisons on KITTI (column 1-2) and Cityscapes (column 3-4). \marginpar{NEW}{The video version is included in our supplementary PDF.}}
\label{fig:kitti_cs_baselines}
\end{figure*}
\myparagraph{Metrics.}
For quantitative evaluation,
four image quality metrics are used: Structural Similarity (SSIM)~\cite{ssim}, Learned Perceptual Image Patch Similarity (LPIPS)~\cite{lpips}, \marginpar{NEW}{Fréchet} Inception Distance (FID)~\cite{fid}, and Fréchet Video Distance (FVD)~\cite{fvd}. SSIM and LPIPS are used to evaluate the similarity between the result and the ground truth. FID and FVD are used for realistic appearance evaluation on the image level and video level, respectively. Thus, FVD can reflect both appearance realism and temporal coherence of the results.
For SSIM, higher scores are better.
For LPIPS, FID, and FVD, lower scores are better.
We use VGG~\cite{VGG} pre-trained on ImageNet as the feature extractor of LPIPS.
To evaluate how significant the modeled hallucination uncertainties are, we use sparsification plots and Area Under the Sparsification Error (AUSE, the lower, the better) which quantify how close the estimate is to the oracle uncertainty, as in~\cite{depthUncertainty}. More details are given in ``Uncertainty results''.
\begin{table*}[t]
\begin{minipage}{0.65\textwidth}
\centering
\setlength{\tabcolsep}{5.5pt}
\caption{The quantitative results on KITTI and Cityscapes.
}
\begin{tabular
{@{\extracolsep{\fill}} l c c c c | c c c c}
\toprule
& \multicolumn{4}{c}{KITTI} & \multicolumn{4}{c}{Cityscapes} \\
\cmidrule{2-5} \cmidrule{6-9}
Model & SSIM$\uparrow$ & LPIPS$\downarrow$ & FID$\downarrow$ & FVD$\downarrow$ & SSIM$\uparrow$ & LPIPS$\downarrow$ & FID$\downarrow$ & FVD$\downarrow$ \\
\midrule[0.6pt]
Mono~\cite{monodepth2} & 0.6803 & 0.2975 & 31.14 & 269.0 & 0.6444 & 0.3375 & 44.90 & 552.8 \\
VF~\cite{liu2017voxelflow} & 0.7184 & 0.3049 & 33.51 & 360.8 & 0.7844 & 0.2936 & 38.78 & 491.5 \\
LGTSM~\cite{LGTSM} & \textbf{0.7369} & 0.2905 & 52.57 & 495.7 & \textbf{0.7959} & 0.2845 & 72.86 & 572.9 \\
Mono-LGTSM & 0.7028 & 0.2798 & 18.95 & 201.6 & 0.7752 & 0.2866 & 51.18 & 457.8 \\
\midrule[0.6pt]
Ours & 0.7162 & \textbf{0.2294} & \textbf{10.94} & \textbf{82.71} & 0.7539 & \textbf{0.2220} & \textbf{9.27} & \textbf{203.4} \\
\bottomrule[1pt]
\end{tabular}
\label{tab:kitti_cs}
\end{minipage}
%
\begin{minipage}{0.35\textwidth}
\centering
\setlength{\tabcolsep}{5.5pt}
\caption{Ablation study results on KITTI.}
\begin{tabular
{@{\extracolsep{\fill}} l c c c c c}
\toprule
Model & SSIM$\uparrow$ & LPIPS$\downarrow$ & FID$\downarrow$ & FVD$\downarrow$ \\
\midrule[0.6pt]
Base & 0.7020 & 0.2407 & 12.26 & 99.75 \\
w/o AFA & 0.7016 & 0.2377 & 11.25 & 91.92 \\
w/o GSA & 0.7160 & 0.2386 & 11.94 & 98.67 \\
w/o U & \textbf{0.7200} & 0.2312 & 11.37 & 89.70 \\
\midrule[0.6pt]
w/o 3D & 0.6830 & 0.2618 & 16.71 & 171.1 \\
w/o Recur & 0.7145 & 0.2331 & 11.11 & 109.5 \\
\midrule[0.6pt]
Ours & 0.7162 & \textbf{0.2294} & \textbf{10.94} & \textbf{82.71} \\
\bottomrule[1pt]
\end{tabular}
\label{tab:ablation}
\end{minipage}
\end{table*}
\subsection{Comparisons}
\label{sec:comparisons}
\myparagraph{Baselines.}
As no prior work is directly comparable to our setting, we compare against geometric method Mono~\cite{monodepth2}, flow-based video prediction method VoxelFlow~\cite{liu2017voxelflow}, and video completion method LGTSM~\cite{LGTSM}, the closest alternatives. We also evaluate the Mono+LGTSM (Mono-LGTSM) setting, where pixels of past frames are first propagated by Mono before fed into LGTSM.
\myparagraph{Hallucination Results.}
Qualitative comparisons with other alternatives are provided in Fig.~\ref{fig:kitti_cs_baselines}. We observe that our method can synthesize more realistic and perceptually appealing results. For example, our method can produce hallucinations with less distortion in the 2nd column result, and better preserve the object appearance with less artifacts in the 3rd and 4th column results. This trend is also reflected from the quantitative measurements (SSIM, LPIPS, and FID scores) in Tab.~\ref{tab:kitti_cs}. Our method generates sharper and more photo-realistic results than VF~\cite{liu2017voxelflow} and LGTSM~\cite{LGTSM}, but the latter have a higher SSIM, probably due to blur, something also argued in~\cite{johnson2016perceptual,Shi-CVPR16-Superres,ma2017pose}.
\marginpar{NEW}{While LPIPS is generally more consistent with human perception and has been wildly adopted in recent image synthesis works ~\cite{ren2020deep,LGTSM}}
Furthermore, our method alleviates the flickering via the simple recurrent strategy which can retain temporal consistency.
Note that, the hallucination of unobserved regions is improved as more surrounding observations become available in later frames.
The FVD score is also consistent with such observations.
In addition to blur and distortion, Mono~\cite{monodepth2} and VF~\cite{liu2017voxelflow} suffer from missing pixels. LGTSM~\cite{LGTSM} cannot propagate the information to the out-of-view regions correctly and thus performs poorly. When using Mono and LGTSM together, we can combine their merits and arrive at better results. This indicates that both 3D cues and hallucination capability are key to the FoV extrapolation problem. FoV-Net gets the best of both worlds, and further addresses ambiguity, leading to the best outcome.
\begin{figure*}[t]
\centering
\footnotesize
\includegraphics[width=1\linewidth]{figures/kittiDepth_csSeg_failure.pdf}\\
\caption{Left: depth extrapolation. Middle: semantic segmentation extrapolation. Right: failure case. Zoom in for more details.}
\label{fig:depth_seg_failure}
\end{figure*}
\begin{figure}[t!]
\begin{subfigure}[t]{{0.49\columnwidth}}
\centering
\includegraphics[height=0.8 in]{figures/sparsification_kitti_v2.pdf}
\caption{KITTI}
\end{subfigure}
\begin{subfigure}[t]{{0.49\columnwidth}}
\centering
\includegraphics[height=0.8 in]{figures/sparsification_cs_v2.pdf}
\caption{Cityscapes}
\end{subfigure}
\caption{\marginpar{NEW}{The sparsification plots. The x-axis denotes the fraction of removed pixels, and the y-axis shows the MSE on the remaining pixels. MSE converges to 0 after removing all pixels in the out-of-view regions. Zoom in for more details.}}
\label{fig:uncertainty}
\end{figure}
\myparagraph{Uncertainty results.}
To evaluate the significance of estimated uncertainty, we adopt a pixel-wise metric Mean-Square-Error (MSE) to sort all pixels in each hallucinated wide FoV image in order of descending uncertainty. Then, we iteratively remove a subset of pixels in the out-of-view regions
(\marginpar{NEW}{\emph{i.e.,}\enspace} 5\% in our experiments) and compute MSE on the remaining to plot a curve that is supposed to shrink if the uncertainty properly encodes the \marginpar{NEW}{hallucinated image's errors} (see Fig.~\ref{fig:uncertainty}).
An ideal sparsification (oracle) is obtained by sorting pixels in descending order of the MSE magnitude. In contrast, a random uncertainty is to remove the pixels randomly each time. \marginpar{NEW}{Besides, we observe that there is usually high uncertainty in the edge part. Therefore, we also construct an edge uncertainty baseline which adopts an estimated soft edge map~\cite{hed_edge} to approximate the hallucination uncertainty.}
As we use 0.5 as our narrow view ratio, the curves decrease to zero when 50\% of
pixels are extracted. The AUSE scores on KITTI are 0.0049, 0.0105, \marginpar{NEW}{0.0149}, and 0.0197 for oracle, uncertainty, \marginpar{NEW}{edge}, and random settings, respectively. The AUSE scores on Cityscapes are 0.0021, 0.0042, \marginpar{NEW}{0.0070}, and 0.0080 for oracle, uncertainty, \marginpar{NEW}{edge}, and random settings, respectively.
As shown in Fig.~\ref{fig:uncertainty}, our method can successfully estimate the hallucination uncertainty which reasonably indicates the prediction errors. For example, in the 2nd column result of Fig.~\ref{fig:kitti_cs_baselines}, the uncertainty is high on the thin pole and car edges, as the depth estimation \marginpar{NEW}{is} noisy there.
In the 4th column result of Fig.~\ref{fig:kitti_cs_baselines}, the uncertainty is high on the car edges and unobserved regions (black holes in the Mono result).
\myparagraph{Ablation study}
is performed on KITTI for each component of FoV-Net. The ablative settings are as follows.
\textbf{Base:} removing the AFA module, GSA module, and the uncertainty mechanism.
\textbf{w/o AFA:} using temporal average pooling to replace the AFA module.
\textbf{w/o GSA:} removing the GSA modules.
\textbf{w/o U:} removing the uncertainty mechanism.
\textbf{w/o 3D:} removing the coordinate generation stage and the feature propagation operation in the frame aggregation network.
\textbf{w/o Recur:} removing temporal modeling, \marginpar{NEW}{\emph{i.e.,}\enspace} no recurrent feed-forward and temporal discriminator $\mathcal{T}$.
The quantitative results are in Tab.~\ref{tab:ablation}. When equipped with our AFA and GSA modules, the results are perceptually closer to the wide FoV targets (in terms of LPIPS score decreasing), as well as more photo-realistic (in terms of FID/FVD scores decreasing) and temporal consistency (in terms of FVD score decreasing). Besides, our interpretable uncertainty mechanism could improve the performance moderately, indicating its ability to guide the learning by reducing supervision ambiguity. In addition, the 3D cues and temporal modeling are also important to good performance.
\subsection{Extended Applications}
\label{sec:extensions}
In this section, we provide several extensions (also see supplementary material).
We further apply our method to depth extrapolation (Fig.~\ref{fig:depth_seg_failure} left) and semantic extrapolation (Fig.~\ref{fig:depth_seg_failure} middle). Results show that our method can benefit other vision perception tasks as well, which are important for robotic motion planning and navigation.
\myparagraph{Depth extrapolation.}
To extrapolate the wide FoV depth, the input will be the narrow FoV depth maps. Our frame aggregation network is then trained with the scale-invariant depth regression loss~\cite{eigen2014depth}.
Both the input and target depth maps are estimated by our depth network $D$.
\myparagraph{Semantic segmentation extrapolation.}
To infer the wide FoV semantic segmentation, the input is the narrow FoV segmentation maps, and our frame aggregation network is then trained with a cross-entropy loss.
Both the input and target segmentation maps are estimated by HRNet~\cite{HRNet}.
\section{Limitations and conclusions}
\marginpar{NEW}{
While FoV-Net has achieved good results, there remain a plethora of avenues for future work.
First, our FoV-Net may be not robust to fast moving objects.
For example, in Fig.~\ref{fig:depth_seg_failure} right, the blurred left region is due to the white moving car which causes serious propagation errors.
One potential solution is to extend the FoV extrapolation in the 3D space with a multi-sensor system and model these moving objects explicitly.}
\marginpar{NEW}{Second, in this work, we focus on extrapolating the present frame to a limited wide view. While 360\degree FoV extrapolation and future prediction could be more helpful in some cases. We plan to extend our method both spatially and temporally to enable 360\degree FoV future prediction.
In conclusion, we present FoV-Net to tackle the \textit{FoV extrapolation} problem. Our framework propagates and aggregates the information observed from the past and current narrow FoV frames to generate the current wide FoV, as well as predicts the hallucination uncertainty. We take a significant step to endow machines with hallucination ability, and we believe such an ability can benefit the robotics community.}
|
\section{Introduction}
The M\"obius function of a poset is an important topic of study in many areas of
mathematics including number theory, topology, and combinatorics. For example,
if the poset is the set of positive integers ordered by the division relation,
then its M\"obius function is the classical M\"obius function in number theory.
The M\"obius function of a poset expresses the reduced Euler characteristic of
the simplicial complex coming from the poset. The principle of inclusion and
exclusion is also generalized in terms of M\"obius functions. See \cite[Chapter
3]{EC1} for more details on M\"obius functions.
The characteristic polynomial of a finite ranked poset is a generating function
for the M\"obius function on the poset. A useful application of characteristic
polynomials is that one can compute the number of regions and bounded regions of
a hyperplane arrangement using the characteristic polynomial of its intersection
poset, see \cite{Stanley2004}.
There are many interesting posets whose characteristic polynomials factor
nicely. For example, the characteristic polynomials of the Boolean poset and the
partition poset are given by \( (q-1)^n \) and \( (q-1)(q-2)\cdots (q-n+1) \)
respectively. See \cite{Blass1997, Sagan1999, Stanley1972} for more
examples.
The main objective of this paper is to study the characteristic polynomials of
the (left) weak order on classical and affine Coxeter groups. The motivation of
this paper was the following observation on the characteristic polynomial of the
poset $\Alt_n$ of alternating permutations ordered by the weak order. Here an
alternating permutation is a permutation $\pi=\pi_1\pi_2\pi_3\dots\pi_n$
satisfying $\pi_1<\pi_2>\pi_3<\cdots$.
\begin{thm}\label{thm:alt}
The characteristic polynomial of $\Alt_n$ is
\[
\chi_{\Alt_n}(q)=q^{\binom{n-1}{2}-\flr{\frac{n}{2}}}(q-1)^{\flr{\frac{n}{2}}}.
\]
\end{thm}
Let $\SG_n$ denote the symmetric group of order \( n \), that is, the set of
permutations on \( [n]:=\{1,2,\dots,n\} \). Note that an element in \( \SG_n \)
is an alternating permutation if and only if its (right) descent set is equal to
\( \{2,4,6,\dots\}\cap [n] \). In Theorem~\ref{thm:fxch} we show that the
characteristic polynomial of the subposet of $\SG_n$ with a fixed descent set
has a simple factorization. Theorem~\ref{thm:alt} then follows immediately from
Theorem~\ref{thm:fxch}.
A Coxeter group is a group defined by generators and certain relations. Coxeter
groups are also studied in many different areas of mathematics, and their
classification is well known. In particular, the symmetric group $\SG_n$ is the
Coxeter group of type $A_{n-1}$, see \cite[Proposition 1.5.4]{BjornerBrenti}.
There are two important orders on Coxeter groups, the Bruhat order and the
(left) weak order. In this paper we will only consider the weak order. See
\cite{Bjorner1984,BjornerBrenti,Jedlivcka2005,Tenner2020} and references therein
for combinatorial properties of the weak order.
In search of a generalization of Theorem~\ref{thm:alt} we study characteristic
polynomials of Coxeter groups under the weak order. By slightly modifying the
definition of the characteristic polynomial so that it can be defined on an
infinite poset, we also compute (modified) characteristic polynomials of affine
Coxeter groups.
The rest of this paper is organized as follows.
In Section~\ref{sec:preliminaries}, we give some definitions and known results
on posets and Coxeter groups. In Section~\ref{sec:char-polyn-an}, we show that
the characteristic polynomial of an interval of a Coxeter group with the weak
order is decomposed into the product of characteristic polynomials of its
subgroups. In Section~\ref{sec:prop-desc-class}, we show that the descent class
of a Coxeter group is an interval. In Section~\ref{sec:perm-with-fixed} we give
a simple product formula for the characteristic polynomial of the set of
permutations with a fixed descent set. In Section~\ref{sec:modif-char-polyn}, we
slightly modify the characteristic polynomial so that it is defined for affine
Coxeter groups. We then compute the generating functions for the modified characteristic
polynomials of the classical Coxeter groups \( A_n, B_n \), and \( D_n \).
Finally, in Section~\ref{sec:modif-char-polyn-1}, we express the modified
characteristic polynomials for affine Coxeter groups \( \tilde{A}_n,
\tilde{B}_n, \tilde{C}_n, \) and \( \tilde{D}_n \) in terms of the corresponding
finite Coxeter groups.
\section{Preliminaries}
\label{sec:preliminaries}
In this section, we give basic definitions and results on posets and Coxeter groups.
\subsection{Posets}
Let $(P,\le)$ be a poset. For \( x,y\in P \) we write \( x<y \) to mean \( x\le
y \) and \( x\ne y \). For $x,y\in P$, if $x< y$ and there is no element $t\in
P$ such that $x<t<y$, then we say that \emph{$y$ covers $x$} and denote by
$x\lessdot y$. The \emph{(closed) interval} $[x,y]$ is the subset $\{t\in P:x\le
t\le y\}$ of $P$. If there is an element $x\in P$ such that $x\le y$
(resp.~$x\ge y$) for all $y\in P$, then \( x \) is called the \emph{bottom}
(resp.~the \emph{top}) of $P$ and is denoted by $\hat{0}$
(resp.~$\hat{1}$).
Let $C=\{t_1,t_2,\dots,t_k\}$ be a subset of $P$. If any two elements of $C$ are
comparable, then $C$ is called a \emph{chain} of $P$. The \emph{length} of a
chain $C$ is defined to be $|C|$. A chain of \( P \) is called a \emph{maximal
chain} if it is not contained in any other chain of \( P \). If every maximal
chain of $P$ has the same length, say $m$, then $P$ is called a \emph{ranked
poset} (or a \emph{graded poset}). In this case, there is a unique \emph{rank
function} $r_P :P\rightarrow\{0,1,2,\dots,m\}$ such that $r_P(t)=0$ for every
minimal element $t\in P$ and $r_P(y)=r_P(x)+1$ if $x\lessdot y$. We say that
$r_P(x)$ is the \emph{rank} of $x\in P$. The length $m$ of a maximal chain of
$P$ is called the \emph{rank} of $P$ and denoted by $r_P(P)$.
For two elements $x,y$ of $P$, a \emph{lower bound} of $x$ and $y$ is an element
$u$ such that $u\le x$ and $u\le y$. If $u$ is a lower bound of $x$ and $y$
satisfying $v\le u$ for any lower bound $v$ of $x$ and $y$, then $u$ is called
the \emph{meet} of $x$ and $y$ and denoted by $x\wedge y$. Similarly, an
\emph{upper bound} of $x$ and $y$ is an element $u$ such that $x\le u$ and $y\le
u$. If \( u \) is an upper bound of $x$ and $y$ satisfying $u\le v$ for any
upper bound $v$ of $x$ and $y$, then $u$ is called the \emph{join} of $x$ and
$y$ and denoted by $x\vee y$. For a subset $A=\{a_1,a_2,\dots,a_i\}$ of $P$, we
denote by $\bigvee A$ the join of all elements in $A$, i.e., $\bigvee A=a_1\vee
a_2\vee\cdots\vee a_i$.
The \emph{M\"{o}bius function} $\mu_P$ of $P$ is the unique function
defined on the pairs \( (x,y) \) of elements \( x,y\in P \) with \( x\le y \) satisfying the following condition:
\[
\sum_{x\le t\le y} \mu_P(x,t)= \delta_{x,y} \qquad \mbox{for all \( x\le y \) in \( P \),}
\]
where \( \delta_{x,y} \) is \( 1 \) if \( x=y \) and \( 0 \) otherwise.
\begin{defn}
Let $P$ be a finite ranked poset with $\hat{0}$. The \emph{characteristic
polynomial} $\chi_P(q)$ of $P$ is defined by
\[
\chi_P(q)=\sum_{t\in P}\mu_P(\hat{0},t)q^{r_P(P)-r_P(t)}.
\]
\end{defn}
\subsection{Coxeter groups}
Let \( S \) be a set and let \( m:S\times S\to \{1,2,\dots,\infty\} \) be a
function such that $m(s,s)=1$ and $m(s,s')=m(s',s)\ge 2$ for $s\neq s'$. Let \(
W \) be the group defined by the generators \( S \) and the relations \(
(ss')^{m(s,s')}=e \) for all \( s,s'\in S \), where \( e \) is the identity
element. In this case we say that $W$ is a \emph{Coxeter group}, $S$ is a set of
\emph{Coxeter generators}, and the pair $(W,S)$ is a \emph{Coxeter system}.
Note that, for any Coxeter generators $s,s'$ with $s\neq s'$, we have
\begin{equation}\label{eq:braid}
s^2=e\mbox{\qquad and\qquad} \overbrace{ss'ss's\cdots}^{m(s,s')}=\overbrace{s'ss'ss'\cdots}^{m(s,s')}.
\end{equation}
In particular, distinct Coxeter generators $s$ and $s'$ commute if and only if $m(s,s')=2$.
A \emph{Coxeter graph} (or a \emph{Coxeter diagram}) is a graph such that the nodes are the Coxeter generators,
nodes $s$ and $s'$ are connected by an edge if $m(s,s')\ge 3$, and the edge has the label $m(s,s')$ if $m(s,s')\ge 4$.
A Coxeter system $(W,S)$ is \emph{irreducible} if the Coxeter graph of $(W,S)$ is connected.
The irreducible Coxeter groups have been classified and every reducible Coxeter group
decomposes uniquely into a product of irreducible Coxeter groups \cite[p.4]{BjornerBrenti}.
From now on, we assume that $(W,S)$ is a Coxeter system.
Since $S$ is a set of generators of $W$, each
element $w\in W$ can be represented as a product of generators, say
$w=s_1s_2\dots s_k$, where $s_i\in S$. Among all such expressions for $w$, the
smallest $k$ is called the \emph{length} of $w$ and denoted by $\ell(w)$. If \(
k=\ell(w) \), then the expression $s_1s_2\dots s_k$ is called a \emph{reduced
word} for $w$.
Let \(\alpha_{s,s'}=\overbrace{ss'ss's\cdots}^{m(s,s')}\). Then the second
equation in \eqref{eq:braid} is rewritten by \(\alpha_{s,s'}=\alpha_{s',s}\).
The replacement of \(\alpha_{s,s'}\) by \(\alpha_{s',s}\) in a word is called a
\emph{braid-move}. It is known that every two reduced words of \(w\in W\) is
connected via a sequence of braid-moves, see \cite{Tits1969}.
\begin{defn}
The Coxeter generators $s\in S$ are called \emph{simple reflections}.
For \( w\in W \), we define the following sets:
\begin{align*}
&D_L(w)=\{s\in S:\ell(sw)<\ell(w)\},\\
&D_R(w)=\{s\in S:\ell(ws)<\ell(w)\}.
\end{align*}
The set $D_L(w)$ (resp.~$D_R(w))$ is called the \emph{left} (resp.~\emph{right})
\emph{descent set} of \( w \) and its elements are called \emph{left}
(resp.~\emph{right}) \emph{descents} of \( w \). Note that $D_L(w)=D_R(w^{-1})$.
\end{defn}
\begin{defn}
For $I\subseteq J\subseteq S$, the set $\MD^J_I=\{w\in W:I\subseteq
D_R(w)\subseteq J\}$ is called a \emph{(right) descent class}. We also denote
$\MD_I=\MD^I_I=\{w\in W:D_R(w)=I\}$ and $W^J=\MD^{S\setminus
J}_\emptyset=\{w\in W:D_R(w)\subseteq S\setminus J\}$.
For
$J\subseteq S$, the subgroup of $W$ generated by $J$ is called a \emph{parabolic
subgroup} and denoted by $W_J$.
\end{defn}
Note that $(W_J,J)$ is also a Coxeter system \cite[Proposition
2.4.1]{BjornerBrenti}.
\begin{defn}
For $u,w\in W$, we denote by $u\le w$ if $w=s_k\dots s_2s_1u$ for some $s_i\in
S$ with $\ell(s_i\dots s_2s_1u)=\ell(u)+i$ for $0\le i\le k$. This partial
order $\le$ is called the \emph{(left) weak order}.
\end{defn}
Throughout this paper we consider a Coxeter group \(W\) as a poset using the
weak order.
If $W$ is finite then $W$ is a lattice \cite{Bjorner1984}. In this case
there exists the top element of \( W \), which we denote by $w_0$. For some
$J\subseteq S$, if $W_J$ is finite, then we denote the top element of $W_J$ by
$w_0(J)$. Since the weak order on $W$ is graded, the rank function is defined on
\( W \). By the definition of the weak order, the rank and the length of each
element are the same.
Let $\mu_W$ denote the M\"{o}bius function of $W$. For $u,w\in W$, it is known
that $\mu_W$ only takes the values \( 0,1 \), and \( -1 \).
\begin{prop}\cite[Corollary 3.2.8]{BjornerBrenti}
\label{prop:mbf}
The M\"{o}bius
function $\mu_W(u,w)$ of $W$ is given by
\[
\mu_W(u,w)=
\begin{cases}
(-1)^{|J|} & \mbox{if $w=w_0(J)u$ for some $J\subseteq S$,}\\
0 & \mbox{otherwise.}
\end{cases}
\]
\end{prop}
If $u=e$ in Proposition~\ref{prop:mbf}, then $\mu_W(e,w)$ is
nonzero if and only if $w=w_0(J)$ for some $J\subseteq S$. Hence the
characteristic polynomial $\chi_W(q)$ of $W$ can be rewritten as
\begin{equation}\label{eq:chrewrt}
\chi_W(q)=\sum_{w\in W}\mu_W(e,w)q^{\ell(w_0)-\ell(w)}=\sum_{J\subseteq S} (-1)^{|J|}q^{\ell(w_0)-\ell(w_0(J))}.
\end{equation}
We note that the M\"{o}bius function \(\mu(u,w)\) for the Bruhat order, which is
another important partial order on a Coxeter group, is given by
\[
\mu(u,w)=(-1)^{\ell(u)+\ell(w)},
\]
see \cite[Theorem]{Verma1971}.
The following proposition is useful in this paper.
\begin{prop}\cite[Proposition 3.1.6]{BjornerBrenti}
\label{prop:loweritv}
If \(u\le w\), then \([u,w]\cong[e,wu^{-1}]\).
\end{prop}
\section{The characteristic polynomial of an interval of a Coxeter group}
\label{sec:char-polyn-an}
In this section we show that the characteristic polynomial for
an interval of a Coxeter group $W$ is decomposed into the product of
the characteristic polynomials of some Coxeter subgroups of $W$.
Recall that since the weak order on a finite Coxeter group is a lattice, we can
define its characteristic polynomial. Even if a Coxeter group is infinite, its
interval can be thought of as a finite graded poset with the bottom and top
elements so the characteristic polynomial of the interval is also defined.
\begin{lem}\label{lem:join}
Let $I, J\subseteq S$ such that $W_I$ and $ W_J$ are finite subgroups of $W$.
Then we have
\[
w_0(I)\vee w_0(J)=w_0(I\cup J).
\]
\end{lem}
\begin{proof}
Since $W_I$ and $W_J$ are finite, we have $\bigvee I=w_0(I)$ and $\bigvee J=w_0(J)$,
see \cite[Lemma 3.2.3]{BjornerBrenti}. Hence
\[
w_0(I)\vee w_0(J)=\left(\bigvee I\right)\vee\left(\bigvee J\right)=\bigvee(I\cup J)=w_0(I\cup J).
\qedhere
\]
\end{proof}
\begin{prop}\label{prop:max}
For $w\in W$, let $\mathcal{M}(w)=\{u\in W:u\le w\mbox{ and } u=w_0(J)\mbox{ for some }J\subseteq S\}$.
Then, $w_0(D_R(w))$ is the unique maximal element of $\mathcal{M}(w)$.
\end{prop}
\begin{proof}
Let $K$ be the union of subsets $J\subseteq S$ such that $w_0(J)\le w$, i.e.,
$K=\bigcup_{w_0(J)\le w} J$. By Lemma~\ref{lem:join} and the definition of
$\mathcal{M}(w)$, we have
\begin{equation}\label{eq:w0K}
w_0(K)=\bigvee\{w_0(J):w_0(J)\le w, J\subseteq S\}=\bigvee \mathcal{M}(w) ,
\end{equation}
which means that $w_0(J)\le w_0(K)$ for all $J$ with \( w_0(J)\le w \). By the
definition of join, we have $w_0(K)=\bigvee_{w_0(J)\le w} w_0(J)\le w$, which
implies $w_0(K)\in\mathcal{M}(w)$. By \eqref{eq:w0K}, this implies that
$w_0(K)$ is the unique maximal element of $\mathcal{M}(w)$.
Now it remains to show that \( K=D_R(w) \).
Suppose $s\in K\subseteq S$.
Since $s\le w_0(K)\le w$ and by the definition of the weak order,
there is a reduced expression of $w$ ending with $s$.
This means that $s\in D_R(w)$.
Conversely, suppose $s\in D_R(w)$.
Then, since $s\le w$ and $s=w_0(\{s\})\in\mathcal{M}(w)$, we have $s\in K$.
Hence we obtain $K=D_R(w)$, which completes the proof.
\end{proof}
To decompose the characteristic polynomial of an interval of $W$, we first
decompose the interval itself into a product of Coxeter subgroups of $W$ as
follows.
\begin{lem}\label{lem:intdecom}
Let $K$ be a subset of $S$ such that $W_K$ is finite.
Let $K_1,K_2,\dots,K_r$ be the subsets of $K$ such that each $K_i$ is a connected component
of the Coxeter graph for $(W_K,K)$ and $\bigcup^r_{i=1} K_i=K$.
Then the interval $[e,w_0(K)]$ of $W$ is isomorphic to
$W_{K_1}\times W_{K_2}\times\cdots\times W_{K_r}$ as posets.
\end{lem}
\begin{proof}
It is well known \cite[Proposition 6.1]{Humphreys1990} that \( W_K \) is
isomorphic to $W_{K_1}\times W_{K_2}\times\cdots\times W_{K_r}$ as Coxeter
groups, hence as posets. Thus it suffices to show that $[e,w_0(K)]=W_K$. For
$u\in W$, let $u\in[e,w_0(K)]$.
Since $u\le w_0(K)$, there is a reduced expression \( s_{i_k}\dots s_{i_1} \) of $w_0(K)$ such that \( s_{i_j}\dots s_{i_1}=u \),
where \( j\le k \). Since every reduced word of $w_0(K)\in W_K$
consists of elements in $K$, we have $u\in W_K$. Conversely, if
$u\in W_K$ then $u\le w_0(K)$ since $w_0(K)$ is the unique maximal element in
$W_K$. Thus $u\in[e,w_0(K)]$, which completes the proof.
\end{proof}
The following lemma shows that the characteristic polynomial of the direct
product of posets is the product of the characteristic polynomials of the
posets.
\begin{lem}\label{lem:prod}
Let \( P \) and \( Q \) be finite graded posets with the bottom and top
elements. Then
\[
\chi_{P\times Q}(q)=\chi_P(q)\chi_Q(q).
\]
\end{lem}
\begin{proof}
It is well known \cite[Proposition 3.8.2]{EC1} that $\mu_{P\times
Q}((s,t),(s',t'))=\mu_P(s,s')\mu_Q(t,t')$ for all $s,s'\in P$ and $t,t'\in
Q$. It is easy to see that $r_{P\times Q}((s,t))=r_P(s)+r_Q(t)$. Hence,
\begin{align*}
\chi_{P\times Q}(q)&=\sum_{(s,t)\in P\times Q} \mu_{P\times Q}(\hat{0}_{P\times Q},(s,t))q^{r_{P\times Q}(P\times Q)-r_{P\times Q}((s,t))}\\
&=\sum_{s\in P,\ t\in Q} \mu_P(\hat{0}_P,s)\mu_Q(\hat{0}_Q,t)q^{(r_P(P)+r_Q(Q))-(r_P(s)+r_Q(t))}\\
&=\chi_P(q)\chi_Q(q). \qedhere
\end{align*}
\end{proof}
Let $u,w\in W$ with $u\le w$. Then the interval $[u,w]$ is
a graded poset with the bottom element $u$ and the top element $w$.
To compute the characteristic polynomial of $[u,w]$,
we need the rank function of the poset $[u,w]$.
Since the rank of $[u,w]$ is the length of maximal chains of $[u,w]$,
it is the same as the difference of the length of $w$ and the length of $u$, i.e., $\ell(w)-\ell(u)$.
Similarly, the rank of an element $t\in [u,w]$ is $\ell(t)-\ell(u)$.
For simplicity, we use the notation $\chi_K$ instead of $\chi_{W_K}$ for the
characteristic polynomial of \(W_K\) for $K\subseteq S$.
\begin{thm}\label{thm:chdecom}
Let $u,w\in W$ with $u\le w$ and let $K=D_R(wu^{-1})$.
Suppose that $K_1,K_2,\dots,K_r$ are the subsets of $K$ such that each $K_i$ is a connected component of the Coxeter graph for $(W_K,K)$ and $\bigcup^r_{i=1} K_i=K$.
Then, the characteristic polynomial of $[u,w]$ is given by
\[
\chi_{[u,w]}(q)=q^{\ell(wu^{-1})-\ell(w_0(K))}\chi_{K_1}(q)\chi_{K_2}(q)\cdots\chi_{K_r}(q).
\]
\end{thm}
\begin{proof}
By Proposition~\ref{prop:loweritv}, we have
\begin{equation}\label{eq:1}
\chi_{[u,w]}(q)=\chi_{[e,wu^{-1}]}(q)=\sum_{t\in [e,wu^{-1}]} \mu_{[e,wu^{-1}]}(e,t)q^{\ell(wu^{-1})-\ell(t)}.
\end{equation}
By Proposition~\ref{prop:mbf}, we have $\mu_{[e,wu^{-1}]}(e,t)\ne 0$ if and only if $t=w_0(J)$ for some
$J\subseteq S$. Therefore we can rewrite \eqref{eq:1} as
\begin{equation}
\label{eq:2}
\chi_{[u,w]}(q)=\sum_{t\in \mathcal{M}({wu^{-1}})} \mu_{[e,wu^{-1}]}(e,t)q^{\ell(wu^{-1})-\ell(t)},
\end{equation}
where
$\mathcal{M}({wu^{-1}})=\{v\in W:v\le wu^{-1}\mbox{ and } v=w_0(J)\mbox{ for some }J\subseteq S\}$.
Let $K=D_R(wu^{-1})$. By Proposition~\ref{prop:max}, we have
$\mathcal{M}({wu^{-1}})\subseteq [e,w_0(K)]$ and \( w_0(K)\le wu^{-1} \). By
Proposition~\ref{prop:mbf} again, if \( t\in [e,w_0(K)]\setminus \mathcal{M}({wu^{-1}}) \),
then \( \mu_{[e,wu^{-1}]}(e,t)= 0 \). Therefore we can rewrite \eqref{eq:1} as
\begin{equation}
\label{eq:3}
\chi_{[u,w]}(q)=\sum_{t\in [e,w_0(K)]} \mu_{[e,wu^{-1}]}(e,t)q^{\ell(wu^{-1})-\ell(t)}.
\end{equation}
Since \( w_0(K)\le wu^{-1} \), we have \( \mu_{[e,wu^{-1}]}(e,t) =
\mu_{[e,w_0(K)]}(e,t) \) for all \( t\in [e,w_0(K)] \).
Therefore the right-hand side of \eqref{eq:3} is equal to
\[
q^{\ell(wu^{-1})-\ell(w_0(K))}\sum_{t\in [e,w_0(K)]} \mu_{[e,w_0(K)]}(e,t)q^{\ell(w_0(K))-\ell(t)}
=q^{\ell(wu^{-1})-\ell(w_0(K))}\chi_{[e,w_0(K)]}(q).
\]
Finally, by Lemmas~\ref{lem:intdecom} and \ref{lem:prod}, we have
\[
\chi_{[e,w_0(K)]}(q) = \chi_{K_1}(q)\chi_{K_2}(q)\cdots\chi_{K_r}(q),
\]
which completes the proof.
\end{proof}
\section{Properties of descent classes}
\label{sec:prop-desc-class}
In this section we show that the descent class $\MD^J_I$ of a Coxeter group $W$
is an interval of $W$ and obtain some properties of the descent classes.
Throughout this section we assume that $(W,S)$ is a finite Coxeter system. Then,
for any $J\subseteq S$, the subgroup $W_J$ is also finite and $w_0(J)$ exists.
Now we show that $\MD^J_I$ is an interval of $W$.
\begin{prop}\label{prop:interval}
For $I\subseteq J\subseteq S$, each descent class $\MD^J_I$ is equal
to the interval $[w_0(I),w_0w_0(J^c)]$ of $W$, where \( J^c = S\setminus J \).
\end{prop}
\begin{proof}
Bj{\"o}rner and Wachs \cite{Bjorner1988} showed that \(\MD^J_I=[w_0(I),w_0^{J^c}]\), where \(w_0^{J^c}\) is the top element of \(W^{J^c}\).
By the relation $w_0=w_0^Jw_0(J)$ in \cite[p.44]{BjornerBrenti}, we are done.
\end{proof}
Note that Theorem~\ref{thm:chdecom} implies that the characteristic polynomial
of an interval $[u,w]$ is determined by $D_R(wu^{-1})$. Therefore the
characteristic polynomial of the descent class $\MD^J_I=[w_0(I),w_0w_0(J^c)]$ is
determined by $D_R(w_0w_0(J^c)w_0(I))$. (Note that $(w_0(I))^{-1}=w_0(I)$.) We
will find a simple description for $D_R(w_0w_0(J^c)w_0(I))$. To do this, we need
the following lemma.
\begin{lem}\label{lem:ndes}
Let $u=u_t\dots u_1$ and $v=v_r\dots v_1$ be reduced expressions of $u,v\in
W$ such that \( u_i\ne v_j \) for all \( 1\le i\le t \) and \( 1\le j\le r \).
Then, for each $v_k\in D_R(v)$, there exists $u_i$ such that $v_ku_i\neq u_iv_k$ if and
only if $v_k\notin D_R(vu)$.
\end{lem}
\begin{proof}
Let $v_k\in D_R(v)$. Then there is a reduced expression \( v'_r \dots v'_1 \)
of \( v \) ending with \( v'_1=v_k \).
For the ``if'' part, suppose that $v_k$ commutes with every $u_i$. Then \(v'_r\dots v'_2
u_t\dots u_1 v'_1\) is a reduced expression of \( vu \) ending with \( v_k
\), which implies $v_k\in D_R(vu)$.
For the ``only if'' part, suppose that there exists $u_i$ that do not commute
with $v_k$. For a contradiction suppose that \(v_k\in D_R(vu)\). Then there is
a reduced expression \( w_{r+t}\dots w_1 \) of \(vu\) ending with \(w_1=v_k\).
Since any two reduced expressions can be obtained from each other via a
sequence of braid-moves, we can find a sequence \( R_1,\dots, R_p \) of
reduced expressions of \( vu \) such that \( R_1=v_r\dots v_1u_t\dots u_1
\), \( R_p=w_{r+t}\dots w_1 \), and each \( R_{j+1} \) is obtained from \( R_j
\) by applying a single braid-move.
Let \( s=v_k \) and \( s'=u_i \). Observe that every simple reflection \( s \)
is to the left of every simple reflection \( s' \) in \( R_1 \).
Therefore we can find the smallest integer \( d \) such
that \( s \) is to the left (resp.~right) of \( s' \) in \( R_d \) (resp.~\(
R_{d+1} \)). This can only happen if the braid-move
$\overbrace{ss'ss'\cdots}^{m}=\overbrace{s'ss's\cdots}^{m}$, for some integer
$m\ge 3$, is used when we obtain \( R_{d+1} \) from \( R_d \). However, since
\( m\ge3 \), we can apply this braid-move only if there is already at least
one \( s \) to the right of \( s' \), which is a contradiction because every \(
s \) is to the left of each \( s' \) in \( R_d \).
Therefore every \( u_i \) must commute with \( v_k \), which completes the proof.
\end{proof}
Now we give a simple expression for the set $D_R(w_0w_0(J^c)w_0(I))$.
\begin{lem}\label{lem:iplus}
Let $I\subseteq J\subseteq S$ and $I^+=\{v\in S:iv\neq vi\mbox{ for some }i\in I\}$.
Then we have $D_R(w_0w_0(J^c)w_0(I))=(J\cup I^+)\setminus I$.
\end{lem}
\begin{proof}
The set \(D_R(w_0w_0(J^c)w_0(I))\) is a subset of $[e,w_0w_0(J^c)w_0(I)]$. By
Proposition~\ref{prop:loweritv}, the interval \([e,w_0w_0(J^c)w_0(I)]\) is
isomorphic to \(\MD^J_I=[w_0(I),w_0w_0(J^c)]\) via $x\mapsto xw_0(I)$ for
\(x\in [e,w_0w_0(J^c)w_0(I)]\). Since \(D_R(w_0w_0(J^c)w_0(I))=\{v\in S: e\le
v\le w_0w_0(J^c)w_0(I)\}\), by the isomorphism \(D_R(w_0w_0(J^c)w_0(I))\) can
also be written as \(\{v\in S:w_0(I)\le vw_0(I)\le w_0w_0(J^c)\}\), say \(K\).
We claim that $K=(J\cup I^+)\setminus I$.
Let $v\in K$. Since $w_0(I)\le vw_0(I)$, we have $v\notin I$. Since
$vw_0(I)\in\MD^J_I$, we have $I\subseteq D_R(vw_0(I))\subseteq J$. Note that
$D_R(w_0(I))=I$ by the maximality of $w_0(I)$. So $D_R(vw_0(I))$ is $I$ or
$I\cup \{v\}$. In the case of $D_R(vw_0(I))=I$, in other words $v\notin
D_R(vw_0(I))$, there exists $i\in I$ such that $iv\neq vi$ by
Lemma~\ref{lem:ndes}. Hence, in this case, we have $v\in I^+$. For the other
case $D_R(vw_0(I))=I\cup \{v\}$, we have $v\in J$ since $I\cup
\{v\}=D_R(vw_0(I))\subseteq J$. So we have $v\in (J\cup I^+)\setminus I$.
Conversely, let $t\in (J\cup I^+)\setminus I$ and consider $tw_0(I)$.
Since $t\notin I$, we have $w_0(I)\le tw_0(I)$. If $t\in J$, then
$D_R(tw_0(I))$ is $I$ or $I\cup \{t\}$ so $I\subseteq D_R(tw_0(I))\subseteq J$.
If $t\in I^+$, then $t\notin D_R(tw_0(I))$ by Lemma~\ref{lem:ndes}.
In other words $D_R(tw_0(I))=I$.
Hence we have $tw_0(I)\in\MD^J_I$, which means that $t\in K$.
\end{proof}
Lemma~\ref{lem:iplus} will be used in later to compute the characteristic
polynomial of $\MD_I$ for the Coxeter group $A_n$.
\section{Permutations with a fixed descent set}
\label{sec:perm-with-fixed}
In this section we give an explicit formula for the characteristic polynomial of
the set of permutations with a fixed descent set ordered by the weak order. As a
corollary we obtain a simple formula for the characteristic polynomial of the
poset of alternating permutations.
An important example of Coxeter groups is the finite irreducible Coxeter group
$A_{n-1}$, which can be identified with the symmetric group as follows. For a
positive integer $n$, let $[n]:=\{1,2,\dots,n\}$. The symmetric group $\SG_n$ is
the set of all bijections from $[n]$ to $[n]$. Each element $\pi\in\SG_n$ is
called a \emph{permutation} and we write $\pi=\pi_1\pi_2\dots\pi_n$ where
$\pi_i=\pi (i)$. The \emph{simple transposition} $s_i$ is the permutation that
exchanges the integers $i$ and $i+1$ and fixes all the other integers. Let $S$
be the set of simple transpositions. Then $(\SG_n,S)$ is the Coxeter group
$A_{n-1}$, see \cite[Proposition 1.5.4]{BjornerBrenti}.
From now on we identify each Coxeter generator \(s_i\) and its index \(i\).
\begin{figure}
\centering
\includegraphics[scale=0.8]{./figures/A_n_graph.pdf}
\caption{The Coxeter graph of $A_n$.}
\label{fig:coxgraph}
\end{figure}
The Coxeter graph of \( A_n \) is shown in Figure~\ref{fig:coxgraph}. Note
that a subgraph of the Coxeter graph of \( A_n \) is connected if and only
if the vertices of the subgraph are consecutive integers. Note also that
$s_is_j\neq s_js_i$ if and only if $|i-j|=1$, i.e., $s_i$ and $s_j$ are not
commutative if and only if they are adjacent.
For the Coxeter group $A_n$, we have an explicit formula for the characteristic
polynomial of $\MD_I$ for $I\subseteq S$.
\begin{thm}\label{thm:fxch}
For $(A_n,S)$ and $I\subseteq S$, let $d$ be the rank of $\MD_I$. Suppose that
$S\setminus I=M_1\cup M_2\cup\cdots\cup M_k$ where each $M_i$ is a maximal set
of consecutive integers. Let $\alpha$, $\beta$ and $\gamma$ be the numbers
of $i$'s such that $|M_i|=1$, $|M_i|=2$ and $|M_i|\ge3$, respectively. Then,
the characteristic polynomial of $\MD_I$ is
\[
\chi_{\MD_I}(q)=q^{d-\alpha-2\gamma-3\beta}(q-1)^{\alpha+2\gamma+\beta}(q^2-q-1)^{\beta}.
\]
\end{thm}
\begin{proof}
By Proposition~\ref{prop:interval}, we have $\MD_I=[w_0(I),w_0w_0(I^c)]$. In
order to use Theorem~\ref{thm:chdecom}, we investigate the structure of
$D_R(w_0w_0(I^c)w_0(I))$. By Lemma~\ref{lem:iplus}, we have
\[
D_R(w_0w_0(I^c)w_0(I))=(I\cup I^+)\setminus I=I^+\setminus I,
\]
where $I^+=\{j\in S:ij\neq ji\mbox{ for some }i\in I\}$. Since
$I^+=\{j\in S:|i-j|=1\mbox{ for some }i\in I\}$, we have
\begin{equation}\label{eq:I+}
I^+\setminus I=\{j\in S\setminus I:|i-j|=1\mbox{ for some }i\in I\}.
\end{equation}
Let $S\setminus I=M_1\cup M_2\cup\cdots\cup M_k$ as in the statement. By
\eqref{eq:I+}, if $|M_i|=1$ or $|M_i|=2$, every element of $M_i$ is contained in
$I^+\setminus I$, and if $|M_i|\ge3$, only the smallest and largest elements of
$M_i$ are contained in $I^+\setminus I$. Hence every maximal subset of
consecutive integers in $I^+\setminus I$ has one or two elements and the
number of subsets of size \( 1 \) (resp.~\( 2 \)) is $\alpha+2\gamma$
(resp.~$\beta$). Therefore, by Lemma~\ref{lem:intdecom}, we have
\begin{equation}\label{eq:A1A2}
[e,w_0(I^+\setminus I)]\cong (\alpha+2\gamma)A_1\times \beta A_2.
\end{equation}
Since the rank of $A_1$ is 1 and the rank of $A_2$ is 3,
the rank of $[e,w_0(I^+\setminus I)]$ is $\alpha+2\gamma+3\beta$.
Thus, by Theorem~\ref{thm:chdecom} and \eqref{eq:A1A2},
\[
\chi_{\MD_I}(q)=q^{d-(\alpha+2\gamma+3\beta)}\chi_{A_1}(q)^{\alpha+2\gamma}\chi_{A_2}(q)^\beta.
\]
Since $\chi_{A_1}(q)=q-1$ and $\chi_{A_2}(q)=(q-1)(q^2-q-1)$, we obtain the
theorem.
\end{proof}
A permutation $\pi=\pi_1\pi_2\dots\pi_n\in\SG_n$ satisfying
$\pi_1<\pi_2>\pi_3<\cdots$ is called an \emph{alternating permutation}. Let
$\Alt_n$ denote the set of alternating permutations in \( \SG_n \). Then one can
easily check that $\Alt_n=\MD_I$ in the Coxeter system \( (A_{n-1},S) \), where
\( I=\{2,4,6,\dots\}\cap [n-1] \). Observe that \( S\setminus I
=\{1,3,5,\dots\}\cap [n-1] \) in which every maximal subset of consecutive
integers has size \( 1 \). Since $|S\setminus I|=\flr{\frac{n}{2}}$ and the rank
of $\Alt_n$ is $\binom{n-1}{2}$, we obtain the characteristic polynomial of
$\Alt_n$ as a corollary of Theorem~\ref{thm:fxch}.
\begin{cor}
The characteristic polynomial of $\Alt_n$ is
\[
\chi_{\Alt_n}(q)=q^{\binom{n-1}{2}-\flr{\frac{n}{2}}}(q-1)^{\flr{\frac{n}{2}}}.
\]
\end{cor}
Stanley \cite{Stanley1972} showed that the characteristic polynomial of a
supersolvable lattice is decomposed into linear factors, see also \cite[Theorem
6.2]{Sagan1999}. Although the characteristic polynomial of \( \Alt_n \) has only
linear factors, one can check that \( \Alt_n \) is not supersolvable.
\section{Modified characteristic polynomials for classical Coxeter groups}
\label{sec:modif-char-polyn}
In this section we find generating functions for modified characteristic
polynomials for the classical Coxeter groups \(A_n\), \(B_n\) and \(D_n\).
\begin{defn}\label{defn:chihat}
Let $W$ be a ranked poset with the bottom element $\hat{0}$. The \emph{modified
characteristic polynomial} \( \hat{\chi}_W(q) \) of \( W \) is defined by
\[
\hat{\chi}_W(q)=\sum_{w\in W} \mu_W(\hat{0},w) q^{r_W(w)},
\]
where \(r_W(w)\) is the rank of \(w\in W\).
\end{defn}
If \( W \) is a finite Coxeter group, there is a simple relation between the
modified characteristic polynomial $\hat{\chi}_W(q)$ and the characteristic
polynomial $\chi_W(q)$ as follows.
\begin{prop}\label{prop:chmodch}
Let $(W,S)$ be a finite Coxeter group and $w_0$ the unique maximal element of $W$. Then we have
\[
\hat{\chi}_W(q)=q^{\ell(w_0)}\chi_W(q^{-1}).
\]
\end{prop}
\begin{proof}
This follows immediately from the definitions of \( \hat{\chi}_W(q)
\) and $\chi_W(q)$.
\end{proof}
In what follows we find an expression for the generating functions for
$\hat{\chi}_{A_n}(q)$, \(\hat{\chi}_{B_n}(q)\), and \(\hat{\chi}_{D_n}(q)\).
First we consider refinements of these generating functions.
\begin{defn}
For the Coxeter system $(A_n,S)$ (resp.~$(B_n,S)$, $(D_n,S)$), let $A_{n,k,l}$
(resp.~$B_{n,k,l}$, $D_{n,k,l}$) be the number of subsets $J\subseteq S$ such that
$|J|=k$ and $\ell(w_0(J))=l$. We define
\begin{align*}
T_A(x,y,z) = \sum_{n\ge0}\sum_{k\ge0}\sum_{l\ge0} A_{n,k,l} x^n y^k z^l,\\
T_B(x,y,z) = \sum_{n\ge0}\sum_{k\ge0}\sum_{l\ge0} B_{n,k,l} x^n y^k z^l,\\
T_D(x,y,z) = \sum_{n\ge2}\sum_{k\ge0}\sum_{l\ge0} D_{n,k,l} x^n y^k z^l,
\end{align*}
where $A_{0,k,l}=B_{0,k,l}=1$ if $k=l=0$ and $A_{0,k,l}=B_{0,k,l}=0$
otherwise.
\end{defn}
In the definition of $T_D(x,y,z)$ the sum is over \( n\ge2 \) for computational
convenience. Note that $B_1\cong A_1$, $D_2\cong A_1\times A_1$, and
$D_3\cong A_3$.
Similar to \eqref{eq:chrewrt}, the polynomial \(\hat{\chi}_W(q)\) can be written
as
\begin{equation}\label{eq:chrewrt1}
\hat{\chi}_W(q)=\sum_{J\subseteq S} (-1)^{|J|}q^{\ell(w_0(J))}.
\end{equation}
If \(W=A_n\), we have
\[
\hat{\chi}_{A_n}(q)=\sum_{k\ge0}\sum_{l\ge0} A_{n,k,l} (-1)^k q^l,
\]
which implies that
\begin{equation}\label{eq:TA}
T_A(x,-1,q)=\sum_{n\ge 0} \left(\sum_{k\ge0}\sum_{l\ge0} A_{n,k,l} (-1)^k q^l\right) x^n=\sum_{n\ge 0} \hat{\chi}_{A_n}(q) x^n.
\end{equation}
By the same arguments we have
\begin{align}
\label{eq:TB}
T_B(x,-1,q)&=\sum_{n\ge 0} \hat{\chi}_{B_n}(q) x^n,\\
\label{eq:TD}
T_D(x,-1,q)&=\sum_{n\ge 2} \hat{\chi}_{D_n}(q) x^n.
\end{align}
We use the fact that $\ell(w_0(A_n))=\binom{n+1}{2}$, $\ell(w_0(B_n))=n^2$,
and $\ell(w_0(D_n))=n(n-1)$, see~\cite[p.16, p.80 Table 2]{Humphreys1990}. Here the
notation $w_0(A_n)$ (resp.~$w_0(B_n)$, \(w_0(D_n)\)) means that the unique maximal
element of $A_n$ (resp.~$B_n$, \(D_n\)).
Now we give explicit formulas for $T_A(x,y,z)$, $T_B(x,y,z)$, and $T_D(x,y,z)$.
\begin{prop}\label{prop:genfc}
We have
\begin{align}
\label{eq:T_A}
& T_A(x,y,z)=\frac{P}{1-xP},\\
\label{eq:T_B}
& T_B(x,y,z)=\frac{Q}{1-xP},\\
\label{eq:T_D}
& T_D(x,y,z)=\frac{x^2P+2x(P-1)+R}{1-xP},
\end{align}
where
\[
P=\sum_{n\ge0} x^n y^n z^{\binom{n+1}2},\ Q=\sum_{n\ge0} x^n y^n z^{n^2},\mbox{ and } R=\sum_{n\ge2} x^n y^n z^{n^2-n}.
\]
\end{prop}
\begin{proof}
Recall that \( A_{n,k,l} \) is the number of subsets $J$ of $S=[n]$ satisfying
$|J|=k$ and $\ell(w_0(J))=l$ in the Coxeter group \( W=A_n \). For such a
subset \( J \), let $J=J_1\cup \cdots\cup J_m$, where each $J_i$ is a
connected component of the Coxeter graph of $(W_J,J)$. Then
$\ell(w_0(J))=\sum_{i=1}^m\ell(w_0(J_i)) =\sum_{i=1}^m \binom{|J_i|+1}{2}$ and
the contribution of \( J\subseteq[n] \) to \( T_A(x,y,z) \) is
\[
\wt(n,J):= x^ny^{|J_1|+\dots+|J_m|}z^{\binom{|J_1|+1}{2} + \cdots + \binom{|J_m|+1}{2}}.
\]
Therefore
\begin{equation}\label{eq:nJ}
T_A(x,y,z) = \sum_{(n,J)\in X} \wt(n,J),
\end{equation}
where \( X \) is the set of all pairs \( (n,J) \) of a nonnegative integer \( n
\) and a set \( J\subseteq[n] \).
We will divide the sum in \eqref{eq:nJ} into two cases \( J=[n] \) and \( J\ne[n] \).
First, note that
\begin{equation}\label{eq:J=[n]}
\sum_{(n,[n])\in X} \wt(n,[n]) = \sum_{n\ge0} x^n y^n z^{\binom{n+1}2} = P.
\end{equation}
Now consider \( (n,J)\in X \) with \( J\ne[n] \). Let \( t \) be the smallest
positive integer not contained in \( J \). Then we have $J=J_1\cup \cdots\cup
J_m$ with \( J_1=[t-1] \) and \( J\setminus J_1 \subseteq \{t+1,t+2,\dots,n\}
\). Let \( J'=\{j-t:j\in J\setminus J_1\} \). Then \( (n-t,J')\in X \) and \(
\wt(n,J)=x^{t}y^{t-1}z^{\binom{t}{2}}\wt(n-t,J') \). Since every \( (n,J)\in X
\) with \( J\ne[n] \) is obtained from \( (n-t,J')\in X \) for some \( t\ge1 \)
and \( J' \) in this way, we have
\begin{equation}\label{eq:J!=[n]}
\sum_{(n,J)\in X, J\ne[n]} \wt(n,J) =
\sum_{t\ge1}x^{t}y^{t-1}z^{\binom{t}{2}}\sum_{(n-t,J')\in X}\wt(n-t,J')
= xP\cdot T_A(x,y,z).
\end{equation}
By \eqref{eq:nJ}, \eqref{eq:J=[n]}, and \eqref{eq:J!=[n]}, we have
$T_A(x,y,z)=P+xP\cdot T_A(x,y,z)$, which implies the first identity \eqref{eq:T_A}.
The second identity \eqref{eq:T_B} can be proved similarly if we consider the
Coxeter system \( (B_n,S) \) with \( S=[n] \) and the Coxeter graph as shown
in Figure~\ref{fig:BD}.
The only difference is that if \( J=[n] \), then \(
\wt(n,J)=x^ny^nz^{n^2} \) since $\ell(w_0(B_n))=n^2$.
\begin{figure}
\centering
\includegraphics[scale=0.8]{./figures/BD.pdf}
\caption{The Coxeter graphs of \( B_n \) and \( D_n \).}
\label{fig:BD}
\end{figure}
For the third identity \eqref{eq:T_D}, we consider Coxeter system \( (D_n,S) \)
with \( S=[n] \) and the Coxeter graph as shown in Figure~\ref{fig:BD}. The
proof is similar to the case of the first identity except that
$\ell(w_0(D_n))=n^2-n$ and we need to consider the cases that the smallest
positive integer \( t \) not contained in \( J \) is \( n-1 \) or \( n \)
separately. It is not hard to check that \( T_D(x,y,z) \) satisfies
$T_D(x,y,z)=R+xPT_D(x,y,z)+2x(P-1)+x^2P$, which implies the third identity.
\end{proof}
By \eqref{eq:TA}, \eqref{eq:TB}, and \eqref{eq:TD}, substituting $y=-1$ and
$z=q$ in the formulas of Proposition~\ref{prop:genfc} gives the following
formulas for the generating functions of \(\hat{\chi}_{A_n}(q)\),
\(\hat{\chi}_{B_n}(q)\), and \(\hat{\chi}_{D_n}(q)\).
\begin{thm}
We have
\begin{align*}
& \sum_{n\ge0} \hat{\chi}_{A_n}(q) x^n=\frac{\sum_{n\ge0} (-1)^n q^{\binom{n+1}2} x^n}{\sum_{n\ge0} (-1)^n q^{\binom{n}2} x^n},\\
& \sum_{n\ge0} \hat{\chi}_{B_n}(q) x^n=\frac{\sum_{n\ge0} (-1)^n q^{n^2} x^n}{\sum_{n\ge0} (-1)^n q^{\binom{n}2} x^n},\\
& \sum_{n\ge2} \hat{\chi}_{D_n}(q) x^n=\frac{\sum_{n\ge2} (-1)^n (q^{\binom{n-1}2}-2q^{\binom{n}2}+q^{2\binom{n}2}) x^n}{\sum_{n\ge0} (-1)^n q^{\binom{n}2} x^n}.
\end{align*}
\end{thm}
\section{Modified characteristic polynomials for affine Coxeter groups}
\label{sec:modif-char-polyn-1}
In this section we express the modified characteristic polynomial \(
\hat{\chi}_W(q) \) of an affine Coxeter group \( W \) using those of finite Coxeter
groups.
Recall that the modified characteristic polynomial $\hat{\chi}_W(q)$ is defined
as a series if $W$ is an infinite poset. However, if \( W \) is an affine
Coxeter group with finite generators, \eqref{eq:chrewrt} shows that
$\hat{\chi}_W(q)$ is a polynomial. Note also that, for an affine Coxeter group,
we only need to consider the elements $w_0(J)$ for $J\subsetneq S$ since
$w_0(S)$ does not exist, i.e.,
\begin{equation}\label{eq:chrewrt2}
\hat{\chi}_W(q)=\sum_{J\subsetneq S} (-1)^{|J|}q^{\ell(w_0(J))}.
\end{equation}
The Coxeter graph of an irreducible affine Coxeter group contains the Coxeter
graph of a finite Coxeter group as a subgraph. Using this observation we can
express the characteristic polynomial of an affine Coxeter group in terms of
those of finite Coxeter groups.
\begin{thm}\label{thm:infintofin}
We have
\begin{align*}
\hat{\chi}_{\tilde{A}_n}(q)&=\hat{\chi}_{A_n}(q)+\sum^n_{k=1} k(-1)^kq^{\binom{k+1}{2}}\hat{\chi}_{A_{n-k-1}}(q),\\
\hat{\chi}_{\tilde{B}_n}(q)&=\hat{\chi}_{B_n}(q)+\sum_{k=0}^{n}(-1)^{k}q^{k(k-1)}\hat{\chi}_{B_{n-k}}(q),\\
\hat{\chi}_{\tilde{C}_n}(q)&=\sum_{k=0}^{n}(-1)^kq^{k^2}\hat{\chi}_{B_{n-k}}(q),\\
\hat{\chi}_{\tilde{D}_n}(q)&=\hat{\chi}_{D_n}(q)+(-1)^nq^{n(n-1)}+\sum_{k=0}^{n}(-1)^kq^{k(k-1)}\hat{\chi}_{D_{n-k}}(q),
\end{align*}
where we assume \( n\ge2 \), \( n\ge3 \), \( n\ge2 \), and \( n\ge 4 \) for \(
\tilde{A}_n \), \( \tilde{B}_n \), \( \tilde{C}_n \), and \( \tilde{D}_n \),
respectively. We define $\hat{\chi}_{A_i} (q) = \hat{\chi}_{B_i} (q)=1$ for $i\le 0$ and $\hat{\chi}_{D_i} (q)=1$ for $i\le 1$.
\end{thm}
\begin{proof}
\begin{figure}
\includegraphics[scale=0.8]{./figures/A_n_tilde.pdf}
\caption{The Coxeter graph of \( \tilde{A}_n \) containing the Coxeter
graph of \( A_n \).}
\label{fig:tA}
\end{figure}
\textbf{Type $\tilde{A}_n$}: The Coxeter graph of $W=\tilde{A}_n$ contains
that of $A_n$ as shown in Figure~\ref{fig:tA}. Let
$S=\{s_1,s_2,\dots,s_{n+1}\}$ be the set of generators. By \eqref{eq:chrewrt2},
we have
\begin{equation}\label{eq:chi(tA)}
\hat{\chi}_{\tilde{A}_n}(q)
=\sum_{J\subsetneq S, s_{n+1}\notin J} (-1)^{|J|} q^{\ell(w_0(J))}
+\sum_{J\subsetneq S, s_{n+1}\in J} (-1)^{|J|} q^{\ell(w_0(J))}.
\end{equation}
Note that the first sum of the right-hand side of \eqref{eq:chi(tA)} is
given by
\begin{equation}\label{eq:chi(tA)1}
\sum_{J\subsetneq S, s_{n+1}\notin J} (-1)^{|J|} q^{\ell(w_0(J))} = \hat{\chi}_{A_n}(q).
\end{equation}
Thus we only need to compute the second sum.
\begin{figure}
\includegraphics[scale=0.8]{./figures/BJdecom.pdf}
\caption{The Coxeter graph of \( \tilde{A}_n \) and the subset \( Q \).}
\label{fig:tA2}
\end{figure}
Suppose \( J\subsetneq S \) with \( s_{n+1}\in J \). Let $Q\subseteq J$ be the
set of generators in the connected component containing $s_{n+1}$ and let
$|Q|=k$. We also define $J'=J\setminus Q$ so that $J=Q\cup J'$. If $k\ge n-1$,
then $J'=\emptyset$. Suppose $k< n-1$ and let $s_a,s_b$ be the generators that
are adjacent to an end vertex of $Q$ in the Coxeter graph of $\tilde{A}_n$, see
Figure~\ref{fig:tA2}. Then $J'$ is a subset of $S'=S\setminus (Q\cup
\{s_a,s_b\})$. Note that $W_{S'}\cong A_{n-k-1}$ since the graph of $S'$ is a
connected path with $|S'|=n-k-1$. Then
\[
(-1)^{|J|} q^{\ell(w_0(J))}=(-1)^{|Q|+|J'|} q^{\ell(w_0(Q))+\ell(w_0(J'))}
= (-1)^{k+|J'|} q^{\binom{k+1}{2}+\ell(w_0(J'))}.
\]
This implies that
\begin{equation}\label{eq:chi(tA)2}
\sum_{s_{n+1}\in J\subsetneq S} (-1)^{|J|} q^{\ell(w_0(J))}
=\sum^n_{k=1} k(-1)^kq^{\binom{k+1}{2}}\hat{\chi}_{A_{n-k-1}}(q).
\end{equation}
By \eqref{eq:chi(tA)}, \eqref{eq:chi(tA)1}, and \eqref{eq:chi(tA)2},
we obtain the formula for \( \hat{\chi}_{\tilde{A}_n}(q) \).
For the rest of the proof we denote by \(S=\{s_0,s_1,\dots,s_n\}\) the set of
generators of \(\tilde{B}_n\), \(\tilde{C}_n\), or \(\tilde{D}_n\).
\begin{figure}
\includegraphics[scale=0.8]{./figures/B_n_tilde.pdf}
\caption{The Coxeter graph of \( \tilde{B}_n \) containing the Coxeter
graph of \( B_n \).}
\label{fig:tB}
\end{figure}
\textbf{Type $\tilde{B}_n$}: The Coxeter graph of $W=\tilde{B}_n$ includes the
graph of $B_n$ as shown in Figure~\ref{fig:tB}.
By \eqref{eq:chrewrt2}, we have
\begin{equation}
\label{eq:chi(tB)}
\hat{\chi}_{\tilde{B}_n}(q)
= \sum_{J\subsetneq S,\{s_0,s_1\}\not\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}
+ \sum_{J\subsetneq S,\{s_0,s_1\}\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}.
\end{equation}
First we observe the case that \(\{s_0,s_1\}\not\subseteq J\).
If $s_0\notin J$, then since $J$ is a subset of the set of generators of $B_n$,
we have $\sum_{J\subsetneq S, s_0\notin J}
(-1)^{|J|}q^{\ell(w_0(J))}=\hat{\chi}_{B_n}(q)$. We have the same identity for
the case $s_1\notin J$. If $s_0, s_1\not\in J$, then we have
$\sum_{J\subsetneq S, s_0,s_1\notin J}
(-1)^{|J|}q^{\ell(w_0(J))}=\hat{\chi}_{B_{n-1}}(q)$. Thus,
\begin{equation}
\label{eq:chiB2}
\sum_{J\subsetneq S,\{s_0,s_1\}\not\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}=
2\hat{\chi}_{B_n}(q)-\hat{\chi}_{B_{n-1}}(q).
\end{equation}
Suppose now that $\{s_0,s_1\}\subseteq J$. Let \( k \) be the smallest
integer such that \( s_k\not\in J \) and let \( Q=\{s_0,\dots,s_{k-1}\} \) and
$J'=J\setminus Q$. Then we have \( 2\le k\le n \), $W_Q\cong D_k$, and
\( J'\subseteq\{s_{k+1},\dots,s_n\} \).
Since $\ell(w_0(Q))=k(k-1)$, we have
\[
(-1)^{|J|}q^{\ell(w_0(J))} = (-1)^{|Q|+|J'|} q^{\ell(w_0(Q))+\ell(w_0(J'))}
=(-1)^k q^{k(k-1)}(-1)^{|J'|} q^{\ell(w_0(J'))}.
\]
Therefore
\begin{align}
\notag
\sum_{J\subsetneq S,\{s_0,s_1\}\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}
&= \sum_{k=2}^{n} (-1)^k q^{k(k-1)}\sum_{J'\subseteq\{s_{k+1},\dots,s_n\}}(-1)^{|J'|} q^{\ell(w_0(J'))}\\
\label{eq:chiB3}
&= \sum_{k=2}^{n} (-1)^k q^{k(k-1)}\hat{\chi}_{B_{n-k}}(q).
\end{align}
Combining \eqref{eq:chi(tB)}, \eqref{eq:chiB2}, and \eqref{eq:chiB3} gives the
desired formula for \( \hat{\chi}_{\tilde{B}_n}(q) \).
\begin{figure}
\includegraphics[scale=0.8]{./figures/C_n_tilde.pdf}
\caption{The Coxeter graph of \( \tilde{C}_n \) containing the Coxeter
graph of \( B_n \).}
\label{fig:tC}
\end{figure}
\textbf{Type $\tilde{C}_n$}: The Coxeter graph of $W=\tilde{C}_n$ includes the
graph of $B_n$ as shown in Figure~\ref{fig:tC}. By \eqref{eq:chrewrt2}, we have
\begin{equation}
\label{eq:chi(tC)}
\hat{\chi}_{\tilde{C}_n}(q)
= \sum_{J\subsetneq S,s_0\not\in J} (-1)^{|J|}q^{\ell(w_0(J))}
+ \sum_{J\subsetneq S,s_0\in J} (-1)^{|J|}q^{\ell(w_0(J))}.
\end{equation}
This can be proved by similar arguments as in the proof of \(\tilde{B}_n\),
where in this case, for \( 1\le k\le n \), we have $W_Q\cong B_k$ and
$\ell(w_0(Q))=k^2$.
\begin{figure}
\includegraphics[scale=0.8]{./figures/D_n_tilde.pdf}
\caption{The Coxeter graph of \( \tilde{D}_n \) containing the Coxeter
graph of \( D_n \).}
\label{fig:tD}
\end{figure}
\textbf{Type $\tilde{D}_n$}: The Coxeter graph of $W=\tilde{D}_n$ includes the graph of $D_n$ as shown in Figure~\ref{fig:tD}.
By \eqref{eq:chrewrt2}, we have
\begin{equation}
\label{eq:chi(tD)}
\hat{\chi}_{\tilde{D}_n}(q)
= \sum_{J\subsetneq S,\{s_0,s_1\}\not\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}
+ \sum_{J\subsetneq S,\{s_0,s_1\}\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}.
\end{equation}
We can obtain the formula by similar arguments as in the proof of $\tilde{B}_n$.
If $\{s_0,s_1\}\nsubseteq J$, we have
\begin{equation}
\label{eq:chiD2}
\sum_{J\subsetneq S,\{s_0,s_1\}\not\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}=2\hat{\chi}_{D_n}(q)-\hat{\chi}_{D_{n-1}}(q).
\end{equation}
Suppose that $\{s_0,s_1\}\subseteq J$.
Let \( k \) be the smallest integer such that \( s_k\not\in J \) and let \( Q=\{s_0,\dots,s_{k-1}\} \) and $J'=J\setminus Q$.
Then, for \( 2\le k\le n-2 \) or \(k=n\), we have $W_Q\cong D_k$ and
\( J'\subseteq\{s_{k+1},\dots,s_n\} \).
For \(k=n-1\), there are two cases such that \(Q\) and \(Q\cup\{s_n\}\).
Then we have \(W_Q\cong D_{n-1}\) and \(W_{Q\cup\{s_n\}}\cong D_n\).
In both cases we have \(J'=\emptyset\).
Therefore
\begin{align}
\notag
\sum_{J\subsetneq S,\{s_0,s_1\}\subseteq J} (-1)^{|J|}q^{\ell(w_0(J))}
&= \sum_{k=2}^{n-2} (-1)^k q^{k(k-1)}\sum_{J'\subseteq\{s_{k+1},\dots,s_n\}}(-1)^{|J'|} q^{\ell(w_0(J'))}\\
\notag
&\qquad +(-1)^{n-1}q^{(n-1)(n-2)}+2(-1)^n q^{n(n-1)}\\
\label{eq:chiD3}
&= \sum_{k=2}^{n-2} (-1)^k q^{k(k-1)}\hat{\chi}_{D_{n-k}}(q)+(-1)^{n-1}q^{(n-1)(n-2)} +2(-1)^n q^{n(n-1)}.
\end{align}
Combining \eqref{eq:chi(tD)}, \eqref{eq:chiD2}, and \eqref{eq:chiD3} gives the
desired formula for \( \hat{\chi}_{\tilde{D}_n}(q) \).
\end{proof}
Note that using Proposition~\ref{prop:chmodch} the polynomials in
Theorem~\ref{thm:infintofin} can also be written as sums of characteristic
polynomials of finite Coxeter groups.
\section*{Acknowledgments}
The authors would like to thank Seung Jin Lee and Yibo Gao for helpful
discussions.
|
\section{Introduction}\setcounter{equation}{0}
In this paper, we consider the composite optimization problem
\begin{equation}\label{comp}
\mbox{minimize} \;\; (g \circ F) (x) \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n,
\end{equation}
where $g:{\rm I\!R}^m \to \Bar{\R} :=(-\infty , + \infty]$ is lower semicontinuous, and $F: {\rm I\!R}^n \to {\rm I\!R}^m$ is semi-differentiable in the point of question; see section \ref{sect01} for definition of semi-differentiability. Problem \eqref{comp} covers constrained optimization problems when $g$ is an extended-real-valued function. Indeed, by redefining $g$ and $F$ as $g(y) + \delta_{X}(z)$ and $(F(x) , G(x))$, problem \eqref{comp} boils down to the constrained optimization problem
\begin{equation}\label{ccomp}
\mbox{minimize} \;\; f(x):= (g \circ F) (x) \quad \mbox{subject to}\;\; G(x) \in X
\end{equation}
where $ G : {\rm I\!R}^n \to {\rm I\!R}^k$, and where $X$ is a closed (possibly non-convex) set in ${\rm I\!R}^k$. The broadness of problems \eqref{comp} and \eqref{ccomp} allows us to cover many optimization problems, in particular, those with amenable composite functions; see \cite{bf, lw, zl, l, dp, ddmp, dil, r21}. Recall from \cite[Definition~10.23]{rw} that the composite function $f =g \circ F$ is called amenable if $g$ is convex, $F$ is continuously differentiable, and a constraint qualification is present. In this paper, we are specially interested in non-amenable composite function which have not been investigated extensively. More precisely, the composite function $g \circ F$ is called non-amenable if either $g$ is non-convex or $F$ is non-differentiable. Our main objectives in this paper are
\begin{itemize}
\item[•]to obtain calculus rules for subderivative and semi-derivative,
\item[•]to characterize a suitable notion of stationary point via the established calculus rules,
\item[•]to design an algorithm by which we can compute such stationary points.
\end{itemize}
\subsection{Our Contribution}
Our approach to handle the composite optimization problem \eqref{comp} is purely primal which allows us to avoid Lagrangian multipliers. More precisely, we use subderivative as a generalized directional derivative \cite[p.~257]{rw}, that is for a given function $f:{\rm I\!R}^n \to \Bar{\R} = (- \infty , \infty]$ and a point $\bar{x}$ with $f(\bar{x})$ finite, the subderivative function ${\rm d} f(\bar{x})\colon{\rm I\!R}^n\to[-\infty,\infty]$ is defined by
\begin{equation}\label{subderivatives}
{\mathrm d}f(\bar{x})(w):=\liminf_{\substack{
t\downarrow 0 \\
w' \to w
}} {\frac{f(\bar{x}+tw')-f(\bar{x})}{t}}.
\end{equation}
Note that when $f$ is differentiable at $\bar{x}$, we have ${\mathrm d}f(\bar{x})(w) = \langle \nabla f(\bar{x}) , w \rangle .$ We establish a chain rule for the subderivative of composite function \eqref{comp}. Utilizing subderivative, we propose a generalized gradient descent method, where we call it the subderivative method:
\vspace{.5cm}
\begin{table}[H]
\renewcommand{\arraystretch}{1}
\centering
\label{alg1}
\begin{tabular}{|
p{0.9\textwidth} |}
\hline
\textbf{Subderivative Method}\\
\begin{enumerate}[label=(\alph*)]
\item[0.] \textbf{(Initialization)} Pick the tolerance $\epsilon \geq 0$, a starting point $x_0 \in {\rm I\!R}^n$, and set $k=0 .$
\item[1.] \textbf{(Termination)} Stop, if a prescribed stopping criterion is satisfied.
\item[2.] \textbf{(Direction Search)} Pick $w_k \in \mbox{arg min}_{\|w\| \leq 1 } {\rm d} f(x_k)(w).$
\item[3.] \textbf{(Line Search)} Choose a step size $\alpha_{k} > 0$ through a line search method.
\item[4.] \textbf{(Update)} Set $x_{k+1} : = x_k + \alpha_k w_k$ and $k+1 \leftarrow k$ and then go to step 1.
\end{enumerate}
\\ \hline
\end{tabular}
\end{table}
\vspace{.5cm}
The subderivative method reduces to the classical gradient descent method if $f$ is differentiable. Indeed, if $x_k$ is a differentiable point of $f$, with $ \nabla f(\bar{x}) \neq 0$, the direction search in the subderivative method is $w_k = - \frac{ \nabla f(\bar{x})}{\| \nabla f(\bar{x})\|}$. We establish a convergence of the subderivative method with the Armijo backtracking line search under the following condition
\begin{equation*}\label{descet_pro}
f(y) \leq f(x) + {\rm d} f(x) (y-x) + \frac{L}{2} \| y -x \|^2
\quad \quad x,y \in {\rm I\!R}^n .\end{equation*}
Notice that when $f$ is $L$ -smooth, that is $\nabla f$ is Lipschitz continuous with constant $L$, the above property follows from the classical descent lemma; see \cite{N}. Therefore, the subderivative method together with its convergence analysis reduces to the classical gradient descent method for $L$-smooth functions. The rest of the paper is organized as follows. In section \ref{sect01} we mention some examples in the framework of problems \eqref{comp} or \eqref{ccomp}, then briefly point out the related works for each of them. Section \ref{sect02} is for the variational analysis of problem \eqref{comp}. In particular, we establish calculus rules for subderivative to identify the directional stationary points of problem \eqref{comp} and consequently obtain the directional stationary points for all examples in section \ref{sect01}. Finally, in section \ref{sec03}, we propose the subderivative method to find an directional stationary point of problem \eqref{comp} with some positive tolerance.
\section{Examples}\setcounter{equation}{0}\label{sect01}
Below we mention several examples of non-amenable composite optimization problems which fall into framework \eqref{comp}. The generalized calculus established in section \ref{sect02} enables us to compute the subderivative of the essential function of each example. Nevertheless, each example requires full treatment of numerical analysis for (approximately) solving the subderivative method's sub-problems. The latter remains for our future work; in \cite{m22}, where we give a full treatment of the numerical analysis of Example \ref{dmax}.
\begin{Example}[\bf training a neural network problem with the ReLU activation]\label{tnn}{\rm
Fix positive integers $N, n_i \in {\rm I\!N}$ for $i=1,...,N$. Let $W^{i} \in {\rm I\!R}^{n_{i-1} \times n_{i}} $ and $b^{i} \in {\rm I\!R}^{n_i}$ for $i=1,2,..N$ stand for the wieghts and biases which are yet to be determined. Fix functions the vector valued functions $\sigma^{i} : {\rm I\!R}^{n_i} \to {\rm I\!R}^{n_{i}}$ for $i=1,2,.., N-1$ for which the j-th component of each $\sigma^i$ is the ReLU activation function acting on the j-th component of its input, i.e., $[\sigma^i]_{j}(x) = \max\{ 0 , x_j\}$. Then, the fully connected ReLU neural network function with $N+1$ layers ( $N-1$ hidden layers) is defined by the following function ($x$ is the input variable)
\begin{equation}\label{nnf}
f(W, b \; ; \; x) = W^N \sigma^{N-1} \big( . \; . \; . \big( W^2 \sigma^1 (W^1 x - b^1) - b^2\big) . \; . \; . \big)-b^N
\end{equation}
where $W =(W^1,...,W^N )$ and $b=(b^1,...,b^N )$. By training a neural network, one means to find the weights $W^i$ and biases $b^i$ such that the function \eqref{nnf} approximately fits a given data set of pairs $(x,y)$. Indeed, by taking $\{ (x^i , y^i) \in {\rm I\!R}^{n_0} \times {\rm I\!R}^{n_{N}} \; | \; i=1,2,...M \}$ as a set of (training) data, the goal is to minimize the following \textit{loss} function
\begin{equation}\label{l2los}
\mbox{minimize} \;\; L(W, b) := \frac{1}{M}\; \sum_{i=1}^{M} \|f(W,b \; ; \; x^i) \; - \; y^i\|^2 \quad \mbox{over all}\;\; W \;, \; b
\end{equation}}
\end{Example}
\begin{Example}[\bf bilevel programming]\label{bilevel}{\rm In this example, we consider a bilevel programming in the following form:
\begin{equation*}\label{bp}
(BP) \quad \quad \mbox{minimize} \;\; \varphi (x,y) \quad \mbox{subject to}\;\; y \in S(x)
\end{equation*}
where, for any given $x$, $S(x)$ denotes the solution set of the lower-level program
\begin{equation*}\label{px}
(P_x) \quad \quad \mbox{minimize} \;\; \psi (x,y) \quad \mbox{subject to}\;\; h(x,y) \leq 0
\end{equation*}
and $\varphi, \psi : {\rm I\!R}^n \times {\rm I\!R}^m \to {\rm I\!R}$, $h:{\rm I\!R}^n \times {\rm I\!R}^m \to {\rm I\!R}^k$ are continuously differentiable. Similar to \cite{by}, one can place more functional constraints in (BP). For simplicity, we do not add more constraints to (BP). A common approach to obtain optimality conditions and solving (BP) numerically, is to reformulate it into a single-level optimization problem, then applying the optimality condition on the latter problem. In the introduction of \cite{by}, there is a comprehensive discussion about different ways of single-level reformulation of (BP). In this paper, we study the value function approach proposed in Outrata \cite{o}. More precisely, in the sense of global minimizers, (BP) is equivalent to the following single-valued problem:
\begin{equation*}\label{v}
(VP) \quad \quad \mbox{minimize} \;\; \varphi (x,y) \quad \mbox{subject to}\;\; \psi(x,y)- V(x) \leq 0, ~ h(x,y)\leq 0
\end{equation*}
where, $V(x) := \inf \{ \psi(x,y)|\; h(x,y)\leq 0 \}$ is the optimal value function of the lower-level problem $P_x$. The value function $V$ is likely non-differentiable, causing problem (VP) not fitting into the amenable framework. However, under some reasonable assumptions, e.g. \cite[Theorem~5]{mins}, the value function is locally Lipschitz continuous and directionally differentiable, hence, it is semi-differentiable at the point of question.
}
\end{Example}
\begin{Example}[\bf mathematical programming with complementary constraints]\label{complec}{\rm
Let $K , H : {\rm I\!R}^n \to {\rm I\!R}^m$ be continuously differentiable functions. Consider the optimization problem
\begin{equation}\label{complec1}
\mbox{minimize} \;\; f(x) \quad \mbox{subject to}\;\; 0 \geq K(x) \perp H(x) \leq 0.
\end{equation}
It is well-know that, due to the complementarity constraint, the Mangasarian–Fromovitz constraint qualification never holds at any point of the feasible set; see \cite[Proposition~1.1]{jzz}. However, a suitable reformulation of the optimization problem \eqref{complec1} may guaranteee a constraint qualification. Namely, the optimization problem \eqref{complec1} falls into framework \eqref{ccomp} by setting $G(x) := (H(x) , K(x))$ and $$ X := \{ (y ,z) \in {\rm I\!R}^{2k} \; | \; \langle x ,y\rangle = 0 ,~ x, y \in {\rm I\!R}^k_{-} \} $$
This reformulation allows us to take advantage of reasonable constraint qualifications, e.g. metric subregularity, to not only drive a necessary optimality condition but also remove the constraint set by an exact penalty. It is worth mentioning that the latter reformulation of \eqref{complec1} is non-amenable, since $X$ is neither convex. More generally, if $C$ is a closed convex cone in the Euclidean space $E$, which is equipped with the inner product $\langle . \rangle $, the following optimization problem falls into framework \eqref{comp} by similar reformulation
\begin{equation}\label{complec2}
\mbox{minimize} \;\; f(x) \quad \mbox{subject to}\;\; C \ni K(x) \perp H(x) \in C.
\end{equation}
}
\end{Example}
\begin{Example}[\bf difference of amenable functions]\label{dame}{\rm
Another class of optimization problems that can follow the framework \eqref{comp} is the difference of convex functions, which are likely non-amenable if the concave part is non-smooth. Difference of convex functions can be considered as a sub-class of difference of amenable functions, which are the functions written in the form
$$ f(x) := g_1 \circ F_1 (x) - g_2 \circ F_2 (x) $$
where each $g_i$ is convex and each $F_i$ is continuously differentiable. Particularly, in \cite{m22}, we investigated the following class of optimization problems which is a far (non-convex) extension of difference of max of convex functions which was studied in \cite{pra}.}
\end{Example}
\begin{Example}[\bf difference of max functions]\label{dmax}{\rm Let $\varphi : {\rm I\!R}^n \to {\rm I\!R}$ be an amenable function and $f_i : {\rm I\!R}^n \to {\rm I\!R}$ be differentiable functions for $i=1,2,...,m.$ Then, the following optimization problem falls into framework \eqref{comp}.
\begin{equation}\label{dmax1}
\mbox{minimize} \;\; \varphi(x) - \max \{ f_1 (x), f_2 (x), \; ... \; , f_m (x) \} \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n
\end{equation}
The optimization problem \eqref{dmax1} can be equivalently written as
\begin{equation}\label{dmax2}
\mbox{minimize} \;\; \min \{ \varphi(x) - f_1 (x), \; ... \; , \varphi(x) - f_m (x) \} \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n .
\end{equation}
In many applications, $m$ is large ; see \cite{asp, pra} and references therein. In the latter case, the pointwise min produces many (subdifferentially) non-regular points, thus the Clarke subdifferential may be unnecessarily huge and provide misleading information.
}
\end{Example}
\begin{Example}[\bf spars optimization]\label{spars}{\rm Let $\Omega \subseteq {\rm I\!R}^n$ bw non-empty set. Finding the sparsest vector in $A \Omega + b$ can be formulated by the following optimization problem
\begin{equation}\label{spars1}
\mbox{minimize} \;\; \| Ax + b \|_{0} \quad \mbox{subject to}\;\; x \in \Omega
\end{equation}
where $\| y \|_0$ counts the number of non-zero components of $y$. If $\Omega$ is convex, then the typical approach to solve the optimization problem \eqref{spars1} is to replace $\| . \|_{0}$ with $\| . \|_1$ as a surrogate function. However, the $\ell_1$ surrogate may not close enough to the actual $\| . \|_0$, thus other types of sparsities can be considered as well; see \cite{cp}.
}
\end{Example}
\subsection{Related works}
In this sub-section, we briefly review previous related works for the composite problem \eqref{comp}. Then, we separately review theoretical and numerical results.
\begin{itemize}
\item \textbf{Variational analysis of the optimization problem \eqref{comp}}
\end{itemize}
In all previous examples, either $g$ is non-convex, or $F$ is non-differentiable which itself may be in the form of compositions of many other non-differentiable functions. The latter irregularities cause a challenge to the study of variational analysis of the problem \eqref{comp} through the different notions of subdifferentials; see \cite{c, i, m18}. To the best of our knowledge, there is no result dealing with the composite optimization problems in such a general framework. The very recently published book \cite{cp} mostly concerns the composite functions $g \circ F$, where both $g$ and $F$ are locally Lipschitz continuous and directionally differentiable. The latter assumptions allow one to employ full limit instead of liminf/sup to define generalized directional derivatives. Consequently, the calculus rules are readily followed akin to the classical differentiable case. The price of such restrictions is to scarify the possibility of involving extended-real-valued functions. Due to this fact, authors in \cite{cp}, had to have a separate analysis of the computation of the tangent cones to constraint sets in various scenarios. Although from the view of exact penalization, Lipschitz continuous functions seem a sufficient ground to study first-order variational analysis of composite structures, the extended-real-valued functions are crucial for the study of second-order variational analysis of composite functions and constrained systems; for instance, the second subderivative of a Lipschitz continuous function may not be a finite-valued function; see \cite[Proposition~13.5]{rw}. The latter restrictions will also not allow one to have a unified analysis of amenable and non-amenable composite functions. Indeed, there is a comprehensive first and second-order order variational analysis on the composite problem \eqref{comp} where $g$ is convex and $F$ is continuously twice differentiable; see \cite{r88, r89, bp, rw, bs, ms, mms1, mms2}, we refer the reader to \cite{ash} for the latest and the strongest results in that topic.
\begin{itemize}
\item \textbf{Numerical methods for the optimization problem \eqref{comp}}
\end{itemize}
Most numerical methods concerning problem \eqref{comp} focus on the amenable cases, where $g$ is convex, and $F$ is smooth. The typical approach to treat amenable cases numerically is to linearize $F$ while keeping $g$ unchanged. The latter technique forces sub-problems for solving \eqref{comp} to become convex. Consequently, solving \eqref{comp} reduces to the solving of a sequence of convex problems. One typical case is the sum of a convex function and a smooth function. There is extensive literature around amenable cases; see e.g. \cite{bf, lw, zl, l, dp, ddmp, dil, r21}. Beyond amenability, the gradient sampling method may asymptotically find a Clarke stationary point for the non-smooth Lipschitz continuous functions \cite{blo, k} . Very recently in \cite{bl}, authors aimed to extend the convergence of the gradient sampling method to the directionally Lipschitz continuous functions. Quite recently, J. Zhang et. al \cite{zljsj} introduced a series of randomized first-order methods and analyzed their complexity in finding a $(\epsilon , \delta) -$ stationary point. The oracle of the algorithm in \cite{zljsj} needs to call a Clarke subgradient in each iteration. Both aforementioned methods are based on sampling (generalized) gradients, thus they do not reduce to the classical gradient descent when iterations happen on differentiable points. Very recently, the monograph \cite{cp} provides a systematic study of solving some classes of non-convex non-differentiable optimization problems by surrogate functions estimating the original function from above at each iteration. The subderivative method that we present in this paper can be viewed in the same line, minimizing an upper surrogate function for the optimization problem \eqref{comp}.
\subsection{Preliminaries}
Throughout this paper, we mostly use the standard notations of variational analysis and generalized differentiation; see, e.g. \cite{m18,rw}. For a nonempty set $\Omega\subset X$, the notation $x\stackrel{\Omega}{\to}\bar{x}$ indicates that $x\to\bar{x}$ with $x\in\Omega$, while $\mbox{\rm co}\,\Omega$ and $\mbox{\rm cone}\,\Omega$ stand for the convex and conic hulls of $\Omega$, respectively. The indicator function $\delta_\Omega$ of a set $\Omega$ is defined by $\delta_\Omega(x):=0$ for $x\in\Omega$ and $\delta_\Omega(x):=\infty$ otherwise, while the Euclidean distance between $x\in{\mathbb X}$ and $\Omega$ is denoted by ${\rm dist}(x;\Omega)$. Sometimes we need to use $\ell_1$-norm for the distance of two sets, in latter case, we again use the same term ''${\rm dist} $", but we clarify the use of $\ell_1$ norm. For a closed set $\Omega$, we denote $\mbox{\rm proj}\,_{\Omega} (x)$, the Euclidean projection on the set $\Omega$. We write $x=o(t)$ with $x\in{\mathbb X}$ and $t\in{\rm I\!R}_+$ indicating that $\frac{\|x\|}{t}\downarrow 0$ as $t\downarrow 0$, where ${\rm I\!R}_+$ (resp.\ ${\rm I\!R}_-$) means the collection of nonnegative (resp.\ nonpositive) real numbers. Recall also that ${\rm I\!N}:=\{1,2,\ldots\}$.
Given a nonempty set $\Omega \subset {\rm I\!R}^n $ with $\bar{x}\in \Omega $, the tangent cone $T_{\Omega}(\bar{x})$ to $\Omega$ at $\bar{x}$ is defined by
\begin{equation*}\label{2.5}
T_{\Omega}(\bar{x})=\big\{w\in {\rm I\!R}^n |\;\exists\,t_k{\downarrow}0,\;\;w_k\to w\;\;\mbox{ as }\;k\to\infty\;\;\mbox{with}\;\;\bar{x}+t_kw_k\in \Omega \big\}.
\end{equation*}
Given the extended-real-valued function $f:{\rm I\!R}^n \to \Bar{\R}:= (-\infty, \infty]$, its domain and epigraph are defined, respectively, by
$$
\mbox{\rm dom}\, f =\big\{ x \in {\rm I\!R}^n | \; f(x) < \infty \big \}\quad \mbox{and}\quad \mbox{\rm epi}\, f=\big \{(x,\alpha)\in X\times {\rm I\!R}|\, f(x)\le \alpha\big \}.
$$
We say that $f \colon{\rm I\!R}^n\to\Bar{\R}$ is {\em piecewise linear-quadratic} (PLQ) if $\mbox{\rm dom}\, f =\cup_{i=1}^{s}\mathbb{P}_i$ with $\mathbb{P}_i$ being polyhedral convex sets for $i=1,\ldots,s$, and if $f$ has a representation of the form
\begin{equation}\label{PWLQ}
f(x)=\hbox{${1\over 2}$}\langle A_ix,x\rangle+\langle a_i,x\rangle+\alpha_i\quad\mbox{for all}\quad x \in\Omega_i,
\end{equation}
where $A_i$ is an $n\times n$ symmetric matrix, $a_i\in{\rm I\!R}^n$, and $\alpha_i\in{\rm I\!R}$ for all $i=1,\ldots,s$. Further, the above function $f$ is called {\em piecewise linear} (PL) if $A_i = 0$for all $i=1,\ldots,s$. Recall from \cite{mms1} that $f \colon{\rm I\!R}^n\to\Bar{\R}$ is {\em Lipschitz continuous} around $\bar{x}\in\mbox{\rm dom}\, f$ {\em relative} to some set $\Omega\subset\mbox{\rm dom}\, f$ if there exist a constant $\ell\in{\rm I\!R}_+$ and a neighborhood $U$ of $\bar{x}$ such that
\begin{equation*}
| f(x)- f(u)|\le\ell\|x-u\|\quad\mbox{for all }\quad x,u\in\Omega\cap U.
\end{equation*}
The main advantage of defining Lipschitz continuity \textbf{relative} to a set, is to not missing the extended-real-valued functions for the study of constrained optimization.
Piecewise linear-quadratic functions and indicator functions of nonempty sets are important examples of extended-real-valued functions that are Lipschitz continuous relative to their domains around any point $\bar{x}\in\mbox{\rm dom}\, f$. Similarly we can define the calmness of $f$ \textbf{at} point $\bar{x}$ relative to a set. In particular, $f$ is called calm at $\bar{x}$ from below if there exist the $\ell > 0$ and a neighborhood $U$ of $\bar{x}$ such that
\begin{equation*}
f(x)- f(\bar{x}) \geq - \ell\|x- \bar{x}\|\quad\mbox{for all }\quad x \in U.
\end{equation*}
If $f$ is Lipschitz continuous around $\bar{x}$ relative to its domain, then $f$ is calm at $\bar{x}$ from below. Given a function $f:{\rm I\!R}^n \to \Bar{\R}$ and a point $\bar{x}$ with $f(\bar{x})$ finite, the subderivative function ${\rm d} f(\bar{x})\colon{\rm I\!R}^n\to[-\infty,\infty]$ is defined by \eqref{subderivatives}. Subderivative and tangent vectors are related by the relation $\mbox{\rm epi}\, {\rm d} f(\bar{x}) = T_{\scriptsize \mbox{\rm epi}\, f} (\bar{x} , f(\bar{x}))$. It is clear, from the latter tangent cone relationship, that ${\rm d} f(\bar{x})(.)$ is always lower semicontinuous, also ${\rm d} f(\bar{x})$ is proper if $f$ is calm at $\bar{x}$ from below. Furthermore, We have $\mbox{\rm dom}\, {\rm d} f(\bar{x}) = T_{\scriptsize \mbox{\rm dom}\, f} (\bar{x} )$ if $f$ is Lipschitz continuous around $\bar{x}$ relative to $\mbox{\rm dom}\, f$, \cite[Lemma~4.2]{mm21}. If $f$ is convex or concave and $\bar{x} \in \mbox{int} \; \mbox{\rm dom}\, f$ then \eqref{subderivatives} reduces to the usual directional derivative i.e,
\begin{equation}\label{dderivative}
{\mathrm d}f(\bar{x})(w)= f' (x ; w) : =\lim_{\substack{
t\downarrow 0 \\
}} {\frac{f(\bar{x}+tw)-f(\bar{x})}{t}}.
\end{equation}
By a slight abuse of notation, we say that $f$ is directionally lower regular at $\bar{x} \in \mbox{\rm dom}\, f$ if $f' (\bar{x} ; w)$, as a number in $[- \infty , +\infty ]$, exists and it is equal to ${\rm d} f(\bar{x})(w)$. Beside the finite-valued-convex (concave) functions, there are many non-convex functions which are directionally lower regular, for instance, $f : = \delta_{X}$ where $X$ is a finite union of polyhedral convex sets, $f$ is piecewise-linear quadratic, and $f(x)=\|x\|_0$, i.e., $\ell_0$-norm; see Example \ref{subzero}. Another example of such functions is the class of \textit{semi-differentiable} functions. Recall from \cite[Definition~7.20]{rw} that a vector-valued function $F: {\rm I\!R}^n \to {\rm I\!R}^m$ is semi-differrentiable at $\bar{x}$ for (direction) $w$, if the following limit exists
\begin{equation}\label{semiderivatives}
{\mathrm d}F(\bar{x})(w):=\lim_{\substack{
t\downarrow 0 \\
w' \to w
}} {\frac{F(\bar{x}+tw')-F(\bar{x})}{t}}.
\end{equation}
$F$ is called semi-differentiable at $\bar{x}$, if the above limit exists for all $w \in {\rm I\!R}^n$, in this case, the continuous mapping ${\rm d} F(\bar{x}) : {\rm I\!R}^n \to {\rm I\!R}^m$ is called the semi-derivative of $F$ at $\bar{x}$. We say $F$ is semi-differentiable (without mentioning $\bar{x}$) if $F$ is semi-differentiable everywhere in ${\rm I\!R}^n$. As it is often clear from context, let us not use different notations for subderivative and semi-derivative; we use ${\rm d}$ for both them. Following \cite[Theorem~7.21]{rw} $F$ is semi-differentiable at $\bar{x}$ if and only if there exists $h: {\rm I\!R}^n \to {\rm I\!R}^m$, continuous and homogeneous of degree one such that
\begin{equation}\label{semidiff-ex}
F(x) = F(\bar{x}) + h(x - \bar{x}) + o (\| x - \bar{x} \|),
\end{equation}
indeed $h = {\rm d} F(\bar{x})$. It is easy to see that, for the locally Lipschitz continuous functions, the semi-differentiability and classical directional differentiability concepts are equivalent. In contrast with \cite[p.~156]{cp}, in the definition of semi-differentiability, we do not require $F$ to be locally Lipschitz continuous. However, one can observe that if $F$ is semi-differentiable at $\bar{x}$, then $F$ is calm at $\bar{x}$, that is there exists a $\ell > 0$ such that for all $x$ sufficiently close to $\bar{x}$
$$ \| F(x) - F(\bar{x})\| \leq \ell \; \| x - \bar{x} \| .$$
\section{First-order variational analysis of non-amenable functions}\setcounter{equation}{0} \label{sect02}
In this section, we study the first-order generalized directional derivative of the composite function $f: = g \circ F $ in problem \eqref{comp}. By developing a chain rule for subderivative, we calculate the subderivative of $f$. The obtained formula is applied to derive a tight first-order necessary optimality condition for problem \eqref{comp}. We first recall the definition of derivability from \cite[Definition~6.1]{rw} which plays an important role in our analysis of non-convex and non-Clarke regular constraint sets.
\begin{Definition}[\bf geometrically derivable sets]
A tangent vector $w \in T_{X}(\bar{x})$ is said to be \textit{derivable}, if there exist $\varepsilon >0 $ and $\zeta : [0 , \varepsilon] \to X$ with $\zeta(0) = \bar{x}$ and $\zeta'_{+} (0) = w$. The set $X$ is called geometrically derivable at $\bar{x}$ if every tangent vector to $X$ at $\bar{x}$ is derivable. We say $X$ is geometrically derivable if it is geometrically derivable at any $x \in X$.
\end{Definition}
The class of geometrically derivable sets is considerably large. It is not difficult to check that the derivability is preserved under finite unions. Therefore, the finite union of convex sets is geometrically derivable as every convex set is geometrically derivable. A consequence of the latter result is that the non-convex polyhedron, graph, and epigraph of PL and PLQ functions are geometrically derivable. Furthermore, under a mild constraint qualification, the derivability is preserved under intersections and pre-image of smooth mappings \cite[Theorem~4.3]{mm21}. The derivability of $X$ at $\bar{x}$ can equivalently be described by the relation $T_{X}(x) = \mbox{\rm Lim}\,_{t \downarrow 0} \frac{X - \bar{x}}{t}$, where the limit means the set-limit in the sense of set-convergence \cite[p.~152]{rw}.\\
\begin{Definition}[\bf metric subregularity qualification condition] Let the composite function $f := g \circ F$ be finite at $\bar{x}$, where $g : {\rm I\!R}^m \to \Bar{\R}$ and $F: {\rm I\!R}^n \to {\rm I\!R}^m $. It is said that the metric subregularity qualification condition (MSQC) holds at $\bar{x}$ for the composite function $f = g \circ F$ if there exist a $\kappa >0$ and $U$ an open neighborhood of $\bar{x}$ such that for all $x \in U$ the following error bound holds :
\begin{equation}\label{msqc}
{\rm dist} ( x \; ; \; \mbox{\rm dom}\, f ) \leq \kappa \; {\rm dist} (F(x) \; ; \; \mbox{\rm dom}\, g )
\end{equation}
Similarly, we say the metric subregularity constraint qualification condition holds for the constraint set $\Omega := \{ x \in {\rm I\!R}^n | \; G(x) \in X \}$ at $\bar{x}$ if MSQC holds for the composite function $\delta_{X} \circ G$, that is there exist a $\kappa > 0$ and $U$, an open neighborhood of $\bar{x}$, such that for all $x \in U$ the following estimates holds:
\begin{equation}\label{msqccons}
{\rm dist} ( x \; ; \; \Omega ) \leq \kappa \; {\rm dist} (G(x) \; ; \; X )
\end{equation}
\end{Definition}
The above version of metric subregularity condition \eqref{msqc} for the composite function $g \circ F$ was first introduced in \cite[Definition~3.2]{mms1} by the author, which only concerns the domain of the functions $f$ and $g$, not their epigraphs as it was considered in \cite{io}. It is clear that the metric subregularity qualification condition is a robust property, in the sense that if \eqref{msqc} holds at $\bar{x}$ then it also holds at any point sufficiently close to $\bar{x}$. Also, observe that the metric subregularity qualification is invariant under changing of norms. The latter is due to the fact that norms are equivalent in finite dimensions. The metric subregularity condition \eqref{msqc} has shown itself a reasonable condition under which the first and second-order calculus rules hold for amenable setting both in finite and infinite-dimensional spaces; see \cite{mm21, mms1, ms, mms2, ash}. Metric subregularity constraint qualification is implied by Robinson constraint qualification. In particular, in nonlinear programming, it is implied by Mangasarian Fromovitz constraint qualification, and in convex programming, it is implied by Slater constraint qualification; for the comparison of various types of constraint qualifications see \cite[Proposition~3.1]{mms1} and \cite[Proposition~3.1]{mm21}. Next, we define a new metric subregularity qualification condition which is even weaker than \eqref{msqc}, and it turns out to be a more reasonable constraint qualification for the bilevel optimization problems; see Example \ref{bilevel}.
\begin{Definition}[\bf directional metric subregularity]\label{dirmscq} Let the composite function $f := g \circ F$ be finite at $\bar{x}$. Take $w \in {\rm I\!R}^n$. We say the directional metric subregularity qualification condition holds at $\bar{x}$ in direction $w$ if \eqref{msqc} holds restricted to the half-line $\{ \bar{x} + tw | \; t\geq 0\}$, that is there exist a $\kappa = \kappa (\bar{x} ,w) > 0$ and a $\epsilon=\varepsilon(\bar{x} , w) >0$ such that for all $t \in [0 ,\varepsilon]$
\begin{equation}\label{dmsqc}
{\rm dist} ( \bar{x} + t w \; ; \; \mbox{\rm dom}\, f ) \leq \kappa \; {\rm dist} (F(\bar{x} + t w) \; ; \; \mbox{\rm dom}\, g )
\end{equation}
We say the directional metric subregularity condition holds $\bar{x}$ (without specifying any direction) if \eqref{dmsqc} holds for all for each $w \in {\rm I\!R}^n$. Similarly, we can define the directional metric subregularity constraint qualification for the constraint set $\Omega := \{ x \in {\rm I\!R}^n | \; G(x) \in X \}$, i.e.,
\begin{equation}\label{dmsqccons}
{\rm dist} ( \bar{x}+tw \; ; \; \Omega ) \leq \kappa \; {\rm dist} (G(\bar{x}+tw) \; ; \; X )
\end{equation}
\end{Definition}
The directional metric subregularity (along all directions) is weaker than (full) metric subregularity \eqref{msqc}, thus any sufficient condition for metric subregularity serves a sufficient condition for directional metric subregularity. If $\kappa $ and $\epsilon$ in \eqref{dmsqc} can be taken uniformly in $w$ then both metric subregularity and directional metric subregularity coincide. Besides sufficient conditions, the checking of \eqref{dmsqc} directly along all $w$ is easier than checking \eqref{msqc}, due to simplicity in dimension one. In Definition \ref{dirmscq}, one can restrict direction $w \in {\rm I\!R}^n$, to $w$ with $\|w\|=1$. If $w$ is feasible direction at point $\bar{x} \in \Omega$, i.e., $\bar{x} + t w \in \Omega$ for sufficiently small but positive $t$, then the directional metric subregularity \eqref{dmsqccons} holds with any constant $\kappa > 0 .$ Therefore, to checking \eqref{dmsqccons}, we only need to establish \eqref{dmsqccons} for infeasible directions. The directional metric subregularity \eqref{dmsqc} is also weaker than the directional metric subregularity introduced by Gfrerer \cite[Definition~2.1]{g} , which requires the estimate \eqref{msqc} holds for all $x$ belong to a \textbf{set} having the set $\{ \bar{x} + t w | t \in ( 0 , \varepsilon ] \}$ within its interior. Therefore, all sufficient conditions for metric subregularity \eqref{msqc}, and the conditions in \cite[Theorem~4.3]{g} and \cite[Theorem~4.1]{byz} can serve as sufficient conditions for \eqref{dmsqc} and \eqref{dmsqccons}. We prefer to keep the term \textit{directional} in definition \eqref{dirmscq}, since \eqref{dmsqc} is metric subregularity merely along direction $w$, not relative to a set which was considered in \cite[Definition~2.1]{g}.
\subsection{Calculus rules via subderivative}
In the following theorem, we establish a subderivative chain rule for the composite function $g \circ F$ where $F$ is semi-differentiable. A non-direction version of the latter chain rule previously obtained by author in \cite{mms1} and \cite{ash} under stronger assumptions; for example, in \cite[Theorem~3.4]{mms1}, $F$ was assumed to be differentiable, and the (full) metric subregularity \eqref{msqc} holds. The latter would allow us to replace $F$ by its first-order taylor expansion to reduce the problem to the problem of the chain rule for the composite function $g \circ A$, where $A$ is an linear map. However, a careful modification of the proof in \cite[Theorem~3.4]{mms1} can be still used here. We give a detailed proof for the sake of completeness.
\begin{Theorem}[\bf subderivative chain rule of non-amenable functions]\label{fcalc}
Let $f: = g \circ F $ be finite at $\bar{x}$ where $g: {\rm I\!R}^m \to \Bar{\R}$ is Lipschitz continuous relative to $\mbox{\rm dom}\, g$, and $F:{\rm I\!R}^n \to {\rm I\!R}^m $ is semi-differentiable at $\bar{x} $ in direction $w$. Further, assume that the directional MSCQ \eqref{dmsqc} holds at $\bar{x}$ for direction $w$. Then, the following chain rule holds:
\begin{equation}\label{fcalc1}
{\rm d} f(\bar{x}) (w) = {\rm d} g ( F(\bar{x})) ( {\rm d} F(\bar{x}) w)
\end{equation}
\end{Theorem}
\begin{proof} Define $\eta(w' ,t) := F(\bar{x}+tw')-F(\bar{x})- t{\rm d} F(\bar{x})(w)$. Deduce from the semi-differentiability of $F$ at $\bar{x}$ for $w$ that $\frac{\eta(w' ,t)}{t}\to 0$ as $t\downarrow 0$ and $w' \to w$. By using the function $\eta$ and definition \eqref{subderivatives}, we get the relationships
\begin{eqnarray*}
\displaystyle{\rm d}(g \circ F)(\bar{x})(w)&=&\liminf_{\substack{t\downarrow 0\\w'\to w}}\frac{g\big(F(\bar{x}+t w')\big)-g \big(F(\bar{x})\big)}{t}\nonumber\\
&=&\liminf_{\substack{t\downarrow 0\\w'\to w}}\frac{g\big(F(\bar{x})+t{\rm d} F(\bar{x})w + \eta(w' , t) \big)-g \big(F(\bar{x})\big)}{t}\nonumber\\
&=&\liminf_{\substack{t\downarrow 0\\w' \to w}}\frac{g\big(F(\bar{x})+t({\rm d} F(\bar{x})w+\frac{\eta(w',t))}{t})\big)-g \big(F(\bar{x})\big)}{t}\nonumber\\
&\ge&{\rm d} g \big(F(\bar{x})\big)\big({\rm d} F(\bar{x}) w \big)\;\mbox{ whenever }\; w\in{\rm I\!R}^n,
\displaystyle
\end{eqnarray*}
which verify the inequality ``$\ge$" in \eqref{fcalc1}. Proceeding next with the proof of the opposite inequality in \eqref{fcalc1}, if ${\rm d} g (F(\bar{x}))({\rm d} F(\bar{x})w)=\infty$, the latter inequality holds obviously. Also observe from the Lipschitz continuity of $g$ around $F(\bar{x})$ relative to its domain that ${\rm d} g (F(\bar{x}))({\rm d} F(\bar{x})w)>-\infty$. Therefore, we may assume that the value ${\rm d} g(F(\bar{x}))({\rm d} F(\bar{x})w)$ is finite, and thus there exist sequences $t_k\downarrow 0$ and $v_k \to {\rm d} F(\bar{x})w$ such that
\begin{equation}\label{subd0}
{\rm d} g \big(F(\bar{x})\big)\big({\rm d} F(\bar{x})w\big)=\lim_{\substack{k\to\infty}}\frac{ g \big(F(\bar{x})+t_k v_k\big)- g \big(F(\bar{x})\big)}{t_k} \in {\rm I\!R}.
\end{equation}
Suppose without loss of generality that $F(\bar{x})+t_k v_k\in\mbox{\rm dom}\, g$ for all $k\in{\rm I\!N}$. Then, the imposed directional MSQC \eqref{dmsqc} at $\bar{x}$ in direction $w$ yields
\begin{equation*}
{\rm dist}(\bar{x}+t_kw \; ; \; \Omega)\le\kappa\,{\rm dist}\big(F(\bar{x}+t_k w)\; ;\; \mbox{\rm dom}\, g \big),\quad k\in{\rm I\!N},
\end{equation*}
which in turn brings us to the relationships
\begin{eqnarray*}
{\rm dist}\Big(w \; ; \; \frac{\Omega-\bar{x}}{t_k}\Big)&\le &\frac{\kappa}{t_k}\,{\rm dist}\big(F(\bar{x})+t_k{\rm d} F(\bar{x})w+ \eta(t_k , w) \; ; \; \mbox{\rm dom}\, g \big)\nonumber \\
&\le&\frac{\kappa}{t_k}\,\big\| F(\bar{x})+t_k{\rm d} F(\bar{x})w+\eta(t_k , w) -F(\bar{x})-t_k v_k\big\|\nonumber\\
&=&\kappa\,\Big\|{\rm d} F(\bar{x})w-v_k+\frac{\eta(t_k , w) }{t_k}\Big\|\;\mbox{ for all }\;k\in{\rm I\!N}.
\displaystyle
\end{eqnarray*}
It allows us to find vectors $w_k\in\frac{\Omega-\bar{x}}{t_k}$ satisfying
$$
\|w-w_k\|\le \kappa\,\Big\|{\rm d} F(\bar{x})w-v_k+\frac{\eta(t_k , w) }{t_k}\Big\| + \frac{1}{k}
$$
and hence, telling us that $\bar{x}+t_k w_k\in\Omega$ for all $k\in{\rm I\!N}$ and that $w_k\to w$ as $k\to\infty$. Since $g$ is Lipschitz continuous relative to its domain around $F(\bar{x})$, there exists $\ell >0$ such that for $k \in {\rm I\!N}$ sufficiently large we have
$$ g\big(F(\bar{x})+t_k v_k)\big)- g\big(F(\bar{x}+t_k w_k)) \geq - \ell \; \| F(\bar{x}+t_k w_k)-F(\bar{x})- t_k v_k \| .$$
Combining above with \eqref{subd0}, we arrive at the relationships
\begin{eqnarray*}
\displaystyle {\rm d} g\big(F(\bar{x})\big)\big({\rm d} F(\bar{x})w\big)&=&\displaystyle\lim_{k\to\infty}\Big[\frac{ g\big(F(\bar{x}+t_k w_k)\big)- g \big(F(\bar{x})\big)}{t_k}+\frac{g\big(F(\bar{x})+t_k v_k)\big)- g\big(F(\bar{x}+t_k w_k)\big)}{t_k}\Big]\\
&\ge& \displaystyle\liminf_{k\to\infty}\frac{g \big(F(\bar{x}+t_kw_k)\big)-g\big(F(\bar{x})\big)}{t_k}-\ell\limsup_{k\to\infty}\Big\|\frac{F(\bar{x}+t_k w_k)-F(\bar{x})}{t_k}-v_k\Big\|\\
&\ge&{\rm d}(g\circ F)(\bar{x})(w)-\ell\limsup_{k\to\infty}\Big\| {\rm d} F(\bar{x})w +\frac{\eta(w_k , t_k)}{t_k}-v_k\Big\|={\rm d} (g\circ F)(\bar{x})(w),\displaystyle
\end{eqnarray*}
This verifies the inequality ``$\ge$" in \eqref{fcalc1} and completes the proof of the theorem.
\end{proof}
\begin{Example}[\bf subderivative of zero norm]\label{subzero}{\rm Let $A \in {\rm I\!R}^{m \times n}$ be a matrix and $b \in {\rm I\!R}^m$ be a column vector. For each $x \in {\rm I\!R}^n,$ define $f(x) := \| Ax + b\|_{0}$, where $\|y\|_{0}$ counts the number of nonzero components of $y$. Then, $f$ is directionally lower regular at each $x \in {\rm I\!R}^n$, and the subderivative of $f$ is calculated by
\begin{eqnarray*}
\displaystyle
{\rm d} f(x)(w) = \left\{\begin{matrix}
0 & \;S(Aw) \subseteq S(Ax + b) \\
\infty & \; S(Aw) \nsubseteq S(Ax + b)
\end{matrix}\right.
\displaystyle
\end{eqnarray*}
where $S(y)$ denotes the support of $y \in {\rm I\!R}^m$, that is the set of all $i \in \{1,2,...,n\}$ with $y_i \neq 0 $.
}
\end{Example}
\begin{proof}
Take $x, \;w \in {\rm I\!R}^n$, and $W \subset {\rm I\!R}^n$, a bounded neighborhood of $w$. It is not difficult to show that there exists a $\varepsilon >0$ such that for all $t \in (0 , \varepsilon)$ and all $w' \in W$ one has
$$ f(x+tw')= \| Ax +b +t Aw' \|_{0} = \|Ax + b \|_{0} + \mbox{Card}\; \big( S(Aw') \setminus S(Ax + b) \big), $$
where ''Card(A)" stands for the number of elements in the set $A$. It is not difficult to check that $S(Aw) \subseteq S(Aw')$, for all $w'$ sufficiently close to $w$. The latter implies that
$$f(x+tw) \leq f(x+tw') \quad \mbox{for all}~ t\in (0,\varepsilon) $$
Therefore, for all $t \in \in (0,\varepsilon)$ and $w'$ close to $w$ we have
\begin{equation*}
\frac{f(x+tw) - f(x)}{t} \leq \frac{f(x+tw') - f(x)}{t} = \frac{\mbox{Card}\; \big( S(Aw') \setminus S(Ax + b) \big)}{t}
\end{equation*}
Hence,
\begin{eqnarray*}
\displaystyle
{\rm d} f(x)(w) &=& \liminf_{t \downarrow 0} \; \frac{f(x+tw) - f(x)}{t} \\\nonumber
&=& \lim_{t \downarrow 0} \; \frac{\mbox{Card}\; \big( S(Aw) \setminus S(Ax + b) \big)}{t} = \left\{\begin{matrix}
0 & \;S(Aw) \subseteq S(Ax + b) \\
\infty & \; S(Aw) \nsubseteq S(Ax + b)
\end{matrix}\right.
\displaystyle
\end{eqnarray*}
The above relationships show that $f$ is directionally lower regular.
\end{proof}
In Theorem \ref{fcalc}, if further $g$ is semi-differentiable at $F(\bar{x})$ for the direction ${\rm d} F(\bar{x})(w)$, then $g \circ F$ is semi-differentiable at $\bar{x}$ for the direction $w$. In particular, if both $F$ and $g$ are semi-differentiable then so is $g \circ F$ and thus it is directionally lower regular. Note that in the latter case, the proof can get significantly simpler as one deal with (full) \textit{limit} instead of \textit{liminf }, similar to the clasical differentiable chain rule. Moreover, in the semi-differentiable case, due to the existence of the full limit, we can take $g$ a vector-valued function and drop the relative Lipschitz continuity assumption on $g$. For the sake of completeness, we give a short proof for the semi-differentiable chain rule, as it seems to be absent in the literature.
\begin{Proposition}[\bf chain rule for semi-differentiable functions]\label{semidiff-chain}
Let $F: {\rm I\!R}^n \to {\rm I\!R}^m$ be semi-differentiable at $\bar{x}$ for direction $w$ and $G: {\rm I\!R}^m \to {\rm I\!R}^p$ be semi-differentiable at $F(\bar{x})$ for the direction ${\rm d} F(\bar{x})(w)$ then $G \circ F$ is semi-differentiable at $\bar{x}$ for the direction $w$ with
\begin{equation}\label{semidiff-chain1}
{\rm d} (G \circ F)(\bar{x}) (w) = {\rm d} G(F(\bar{x})) ({\rm d} F(\bar{x})w).
\end{equation}
In particular, $G \circ F$ is semi-differentiable at $\bar{x}$ if $F$ and $G$ are semi-differentiable at $\bar{x}$ and $F(\bar{x})$ respectively.
\end{Proposition}
\begin{proof}
Define $\eta(w' ,t) := F(\bar{x}+tw')-F(\bar{x})- t{\rm d} F(\bar{x})(w)$. Deduce from the semi-differentiability of $F$ at $\bar{x}$ for $w$ that $\frac{\eta(w' ,t)}{t}\to 0$ as $t\downarrow 0$ and $w' \to w$. Following relationships hold
\begin{eqnarray*}
\displaystyle{\rm d}(G \circ F)(\bar{x})(w)&=&\lim_{\substack{t\downarrow 0\\w'\to w}}\frac{G\big(F(\bar{x}+t w')\big)-G \big(F(\bar{x})\big)}{t}\nonumber\\
&=&\lim_{\substack{t\downarrow 0\\w'\to w}}\frac{G\big(F(\bar{x})+t{\rm d} F(\bar{x})w + \eta(w' , t) \big)-G \big(F(\bar{x})\big)}{t}\nonumber\\
&=&\lim_{\substack{t\downarrow 0\\w' \to w}}\frac{G\big(F(\bar{x})+t({\rm d} F(\bar{x})w+\frac{\eta(w',t))}{t})\big)-G\big(F(\bar{x})\big)}{t}\nonumber\\
&=&{\rm d} G \big(F(\bar{x})\big)\big({\rm d} F(\bar{x}) w \big),
\displaystyle
\end{eqnarray*}
where the last equality holds due to the semi-differentiability of $G$ at $F(\bar{x})$ for the direction ${\rm d} F(\bar{x})(w).$
\end{proof}
In contrast with subdifferential chain rules, both chain rules \eqref{fcalc1} and \eqref{semidiff-chain1} hold in the form of equality. These major pros allow us to apply these chain rules over the arbitrary number of functions composed of each other. The latter is particularly useful to compute the subderivative of the loss function of a deep neural network in \eqref{l2los}. In this regard, we have the following corollary for which we omit its proof.
\begin{Corollary}[\bf auto-differentiation via semi-derivative]
Let $F_i$, $i=1,...k$ be semi-differentiable functions on ${\rm I\!R}^n$. Let $x , w \in {\rm I\!R}^n$. The the semi-derivative of the composition $F_k \circ F_{k-1} \circ ... F_{1}$ can be calculated through the following recursive way:
\begin{equation*}
\left\{\begin{matrix}
x_0 := x \\
x_i := F_{i}(x_{i-1})& i=1,...,k \\
u_i (.) := {\rm d} F_{i}(x_{i-1})(.)& i=1,...,k
\end{matrix}\right.
\end{equation*}
The semi-derivative of the above composite function is $u_k$, i.e.,
$$ {\rm d} \big( F_k \circ F_{k-1} \circ ... F_{1} \big)(x)(.) = u_k(.) $$
\end{Corollary}
\begin{Proposition}[\bf semi-differentiability of the distance functions]\label{semidist}{\rm Let $X$ be a nonempty closed set. Set $f(x):= {\rm dist} (x ; X)$. Then, the subderivative of $f$ is computed by
\begin{equation}\label{subdist}
{\rm d} f(x)(w) =\left\{\begin{matrix}
{\rm dist} (w \; , \; T_{X} (x)) & x \in X \\\\
\min \{ \frac{\langle x - y \; , \; w\rangle}{f(x)} \big| \;y \in \mbox{\rm proj}\,_{X}(x) \}& x \notin X
\end{matrix}\right.
\end{equation}
Furthermore, $f$ is semi-differentiable if $X$ is geometrically derivable.
}
\end{Proposition}
\begin{proof}
The expression for the subderivative of $f$ in \eqref{subdist} was obtained in \cite[Example~8.53]{rw}. We only give a proof to the second part. Fix $\bar{x} , w \in {\rm I\!R}^n$. Since $f$ is (globally) Lipschitz continuous, it suffices to show that
$$ {\rm d} f(\bar{x})(w) = \limsup_{t \downarrow 0} \frac{f(\bar{x} + t w) - f(\bar{x})}{t}. $$
According to the definition of subderivative we always have $\leq$. To prove the opposite inequality, first, we assume $\bar{x} \in X$, thus $f(\bar{x}) = 0$. Since $X$ is geometrically derivable, we have $$ T_{X}(\bar{x}) =\mbox{\rm Lim}\,_{t \downarrow 0}\frac{X - \bar{x}}{t}.$$ By the connection of limits and distance functions in \cite[Corollary~4.7]{rw}, we can write
\begin{eqnarray*}
\displaystyle
{\rm d} f(\bar{x})(w) &=& {\rm dist} (w \; ; \; T_{X} (\bar{x})) = {\rm dist} \big( w \; ; \; \mbox{\rm Lim}\,_{t \downarrow 0}\frac{X - \bar{x}}{t} \big) \\\nonumber
&=& \lim_{t \downarrow 0} {\rm dist} \big( w \; ; \; \frac{X - \bar{x}}{t} \big) = \lim_{t \downarrow 0} \frac{f(x + t w) - f(x)}{t}.
\displaystyle
\end{eqnarray*}
Now we suppose $\bar{x} \notin X$. By \eqref{subdist}, we can find $\bar{y} \in \mbox{\rm proj}\,_{X} (\bar{x})$ such that ${\rm d} f(\bar{x})(w) = \frac{\langle \bar{x} - \bar{y} \; , \; w\rangle}{f(\bar{x})} $. Define $\eta (t):= \|\bar{x} + tw - \bar{y} \|$. The function $\eta$ is differentiable at $t = 0$, because $\bar{x} \neq \bar{y}$. Since $f(\bar{x} + tw) \leq \| \bar{x} + tw - \bar{y} \|$ for all $t \geq 0$, we have
$$ \frac{f(\bar{x} + t w) - f(\bar{x})}{t} \leq \frac{\eta(t) - \eta(0)}{t} \quad \quad t>0 $$
Hence
$$ \limsup_{t \downarrow 0} \frac{f(\bar{x} + t w) - f(\bar{x})}{t} \leq \eta' (0) = \frac{\langle \bar{x} - \bar{y} \; , \; w\rangle}{\| \bar{x} -\bar{y} \|} = {\rm d} f(\bar{x}) (w) $$
which proves the semi-differentiability of $f$ at $\bar{x}$.
\end{proof}
\begin{Remark}{\rm
In the light of Theorem \ref{fcalc} and Proposition \ref{semidist} the the penalized function associated with the constraint $G(x) \in X$ is semi-differentiable if $X$ is geometrically derivable and $G$ is semi-differentiable. More precisely, in the constrained optimization \eqref{ccomp}, if $\varphi := g \circ F$ and $G$ are semi-differentiable at $x$, and $X$ is geometrically derivable at $G(x)$, then the function $ f := \varphi + \rho \; {\rm dist} (G(.) ; X)$ is semi-differentiable, consequently, it is directionally lower regular at $x$. These conditions are general enough to cover interesting classes of non-amenable functions, such as examples mentioned in section \ref{sect01}. We will see that directional lower regularity, in particular, semi-differentiability of $f$, is a key to establish first-order descent methods for the penalized problems.}
\end{Remark}
An immediate consequence of a chain rule is sum rule that can be obtained by Theorem \ref{fcalc}. A non-directional version of this sum rule had been previously obtained in \cite[Theorem~4.4]{mm21}.
\begin{Corollary}[\bf sum rule via subderivative]\label{fsum}
Let functions $f, h : {\rm I\!R}^n \to \Bar{\R}$ be finite at $\bar{x}$. Let $f$ and $h$ be Lipschitz continuous around $\bar{x}$ relative to their domains. Furthermore, assume that the metric subregularity condition holds at $\bar{x}$ in direction $w$ for $\mbox{\rm dom}\, f \cap \mbox{\rm dom}\, g$ that is, there exist a $\kappa = \kappa(\bar{x} , w) > 0 $ and a $\varepsilon = \varepsilon (\bar{x} , w) >0$ such that for all $t \in [0 , \epsilon]$, the following estimate holds:
\begin{equation}\label{summsqc}
{\rm dist} (\bar{x} + tw \; ; \; \mbox{\rm dom}\, f \cap \mbox{\rm dom}\, h ) \leq \kappa \; \big( {\rm dist} (\bar{x} + tw \ ; \mbox{\rm dom}\, f) + {\rm dist} (\bar{x} +tw \; ; \; \mbox{\rm dom}\, h) \big).
\end{equation}
Then, the subderivative sum rule holds at $\bar{x}$ for the direction $w$:
\begin{equation}\label{fsum1}
{\rm d} (f+h)(\bar{x})(w) = {\rm d} f(\bar{x})(w) + {\rm d} h(\bar{x})(w)
\end{equation}
\end{Corollary}
\begin{proof}
Define $g: {\rm I\!R}^n \times {\rm I\!R}^n \to \Bar{\R}$ with $g(x ,y):= f(x) + h(y)$ and $F: {\rm I\!R}^n \to {\rm I\!R}^n \times {\rm I\!R}^n$ with $F (x) = (x ,x)$. It is easy to check that the estimate \eqref{summsqc} reduces to the directional metric subregularity \eqref{dmsqc} for the composite function $g \circ F$ at $\bar{x}$ for direction $w$. By applying chain rule \eqref{fcalc1} on previous composite function, we arrive at \eqref{fsum1}.
\end{proof}
Note that if one of the functions $f$ or $h$ be finite-valued, then \eqref{summsqc} holds automatically for all direction $w \in {\rm I\!R}^n$. In particular, if one the functions $f$ or $h$ is semi-differentiable at $\bar{x}$, then the sum rule \eqref{fsum1} holds for all direction $w \in {\rm I\!R}^n$ without assuming relative Lipschitz continuity. The both chain rule \eqref{semidiff-chain1} and sum rule \eqref{fsum1} will lead us to the calculation of tangent vectors of constraint sets, which is a key to (primal) first-order necessary optimality condition of the problem \eqref{ccomp}.
\begin{Corollary}[\bf calculation of tangent vectors to the constraint sets]\label{calctan} Let $\bar{x} \in \Omega:= \{x \big|\; G(x) \in X \}$ where $G: {\rm I\!R}^n \to {\rm I\!R}^m$ is semi-differentiable at $\bar{x}$ for all directions $w \in C:= \{ w \big|\; {\rm d} G(\bar{x})(w) \in T_{X} (G(\bar{x})) \}$, and the directional metric subregularity constraint qualification \eqref{dmsqccons} holds at $\bar{x}$ for all such directions, then
\begin{equation}\label{calctan1}
T_{\Omega}(\bar{x}) = \{ w \in {\rm I\!R}^n \big| \; {\rm d} G(\bar{x})(w) \in T_{X} (G(\bar{x})) \}
\end{equation}
\end{Corollary}
\begin{proof}
Observe that we always have $T_{\Omega}(\bar{x}) \subseteq \{ w \in {\rm I\!R}^n \big| \; {\rm d} G(\bar{x})(w) \in T_{X} (G(\bar{x})) \} = C$ without assuming any constraint qualification. To show the opposite inclusion, pick $ w \in C$. By chain rule \eqref{fcalc1}, we have $${\rm d} \delta_{\Omega}(\bar{x})(w) = {\rm d} \delta_{C} ( G(\bar{x})) ({\rm d} G(\bar{x}) w) = 0$$ which forces $w \in \mbox{\rm dom}\, {\rm d} \delta_{\Omega}(\bar{x}) = T_{\Omega}(\bar{x}).$
\end{proof}
\begin{Proposition}[\bf first-order necessary optimality condition]\label{fonc}
Let $\bar{x}$ be a local optimal solution for problem \eqref{ccomp}, where $X$ is a closed set and $g$ is a locally Lipschitz continuous function. Let $C \subseteq {\rm I\!R}^n$ be a nonempty set. Assume that $F$ and $G$ are semi-differentiable at $\bar{x}$ respectively for all $w \in C$. Finally suppose that the directional metric subregularity constraint qualification \eqref{dmsqccons} holds at $\bar{x}$ for all $ w \in C$. Then,
\begin{equation}\label{fonc1}
{\rm d} g ( F(\bar{x})) ( {\rm d} F(\bar{x})w ) \geq 0 \quad \quad \text{for all} ~ w \in C ~ \mbox{with}~ {\rm d} G(\bar{x}) w \in T_{X}(G(\bar{x})).
\end{equation}
\end{Proposition}
\begin{proof}
Let $f : {\rm I\!R}^n \to \Bar{\R}$ be the essential function of the optimization problem \eqref{ccomp}, i.e., $f = g \circ F + \delta_{X} \circ G$. Since $\bar{x}$ is a local minimizer of the problem $\min_{x \in {\rm I\!R}^n} ~ f(x) ,$ we have ${\rm d} f (\bar{x})(w) \geq 0$ for all $w \in {\rm I\!R}^n . $ On the other hand, by the sum rule \eqref{fsum1} and the chain rule \eqref{fcalc1}, for all $w \in C$ we have
\begin{eqnarray*}
\displaystyle
{\rm d} f(\bar{x}) (w) &=& {\rm d} (g \circ F) (\bar{x})(w) + {\rm d} (\delta_{X} \circ G)(\bar{x})(w) \\\nonumber
&=& {\rm d} g (F(\bar{x})) ({\rm d} F(\bar{x})w) + {\rm d} \delta_X (G(\bar{x}))({\rm d} G(\bar{x})w)\\\nonumber
&=& {\rm d} g (F(\bar{x})) ({\rm d} F(\bar{x})w) + \delta_{\scriptsize T_{X}(G(\bar{x}))} ({\rm d} G(\bar{x})w )
\displaystyle
\end{eqnarray*}
This verifies \eqref{fcalc1}.
\end{proof}
The bigger set $C$ is, the sharper the necessary condition \eqref{fonc1} is. Particularly, we are interested in the case $C = {\rm I\!R}^n .$ Since we alwas have $T_{\Omega} \subseteq \{w| \; {\rm d} G(\bar{x})(w) \in T_{X}(G(\bar{x}))\}$, the set $C$ does not need to be larger than $\{w| \; {\rm d} G(\bar{x})(w) \in T_{X}(G(\bar{x}))\}$. Hence, one does need to worry about semi-differentiability or directional metric subregularity in all directions, as any differentiability property constraint qualification comes with a cost in the bilevel programming; see Example \ref{obilevel}. Motivated by the latter observation, we have the following definition of stationarity relative to a given set.
\begin{Definition}[\bf directional stationary points relative to sets]\label{ds} We say that a feasible point $\bar{x}$ for problem \eqref{ccomp} is a \textit{(d)irectional-stationary} point relative to the set $C$, if it satisfies the first-order necessary condition \eqref{fonc1}. The feasible point $\bar{x}$ is called a d-stationary point (without mentioning any set), if it is a d-stationary point relative to the set $C:=\{w| \; {\rm d} G(\bar{x})(w) \in T_{X}(G(\bar{x}))\}$.
\end{Definition}
It is well-known that the d-stationary points are the closest candidates to the local minimizers among other stationary points, this can be illustrated in a ReLU neural network function; indeed, for the function $f(x) = - \max \{0 , x \}$, the point $\bar{x} =0$ is identified as a Clarke- even a limiting-stationary point, whereas, it is not a d-stationary point which correctly rejects its optimal candidacy. To compare the different types of stationary points we refer the reader to \cite[Proposition~1]{cps}. When an optimization problem lacks a constriant qualification or direffrentiability on its datas, e.g. bilevel programming and opimization with complimentarity constraints, the d-stationarity relative to smaller sets can be used. To illustrait the latter situation, in the following example and remark, we compute the d-stationary points of the bilevel programming formulated in Example \ref{bilevel}.
\begin{Example}[\bf optimality condition of bilevel programming]\label{obilevel}{\rm Let $(\bar{x} , \bar{y}) \in {\rm I\!R}^n \times {\rm I\!R}^m$, and let $D \subset {\rm I\!R}^n $ be a nonempty set. In the setting of the Example \ref{bilevel}, assume that the value function $V$ is semi-differentiable for all directions $w \in D .$ If $(\bar{x} , \bar{y})$ solves the problem (BP), it solves (VP) as well. Furthermore, we assume that in (VP) the directional metric subregularity CQ \eqref{dmsqccons} holds at $\bar{x}$ for the all directions $(w , u) \in D \times {\rm I\!R}^m $. Theforefore, by Proposition \ref{fonc}, we have $(\bar{x} , \bar{y})$ is a $d$-stationary point for (VP) relative to the set $D \times {\rm I\!R}^m$ that is
$$ \nabla \varphi (\bar{x} , \bar{y}) (w , u) \geq 0 $$
whenever $(w , u) \in D \times {\rm I\!R}^m$ and
\begin{equation}\label{obilevel1}
\langle \nabla \psi (\bar{x} , \bar{y}) \; , \; (w , u) \rangle - V' (\bar{x} ; w) \leq 0, \quad \langle \nabla h(\bar{x} , \bar{y}) \; , \; (w,u) \rangle \leq 0
\end{equation}
In the above, we implicitly assume $h(\bar{x} , \bar{y}) =0$, otherwise we could neglect the constraint $h(x , y) \leq 0 $, thus drop the related inequality in \eqref{obilevel1}. For the calculation of $V' (\bar{x} ; w)$ in terms of the initial dates $h$ and $\psi$, one can see e.g., \cite[Proposition~4.3]{by} or \cite[Theorem~30]{mins}.
}
\end{Example}
\begin{Remark}[\bf comparison with other results]{\rm In Example \ref{obilevel}, we only require $V$ to be semi-differentiable at $\bar{x}$ for directions $w \in D$. The latter is implied by directional differentiability and \textit{directional Lipschitz continuity}, defined in \cite{by}, of the value function $V$; see \cite[Theorem~4.1]{by} and \cite[Proposition~4.3]{by}. The optimality condition obtained in Example \ref{obilevel} is in a primal version which is generally sharper than a dual version obtained by a subdifferential. In this paper, we only stick to the primal argument. However, one can obtain a dual optimality condition for the constrained problem \eqref{ccomp} and Example \ref{obilevel} by adapting a suitable directional subdifferential. Alternatively, under the assumption of the existence of critical direction made in \cite[Theorem~5.1]{by}, we can derive a dual-directional optimality condition. Indeed, in Example \ref{obilevel}, if $w \in D $ and there exists $v \in {\rm I\!R}^m$ such that $(w , v)$ satisfying \eqref{obilevel1} with $ \nabla \varphi (\bar{x} , \bar{y}) (w , v) \leq 0 $, then our directional optimality condition tells us that the following linear programming (in variable $u$) has the optimal solution $v$ with the optimal value $- \langle \nabla_{x} \varphi (\bar{x} , \bar{y}) , w \rangle$:
\begin{equation}\label{lp}
\mbox{minimize} \;\; \langle \nabla_{y} \varphi (\bar{x} , \bar{y}) , u \rangle\quad \mbox{subject to} \quad u ~ \mbox{satisfies} ~ \eqref{obilevel1}
\end{equation}
Now one by taking a dual of the above linear programming will arrive to a dual-version of the optimality condition for the bilevel problem \eqref{bilevel}. Indeed the dual of problem \eqref{lp} is
\begin{eqnarray}\label{dlp}
\displaystyle
&\mbox{maximize}& - \lambda_1 V'(\bar{x} ; w) + \lambda_1 \langle \nabla_{x} \psi (\bar{x} , \bar{y}) , w \rangle + \lambda_2 \langle \nabla_{x} h (\bar{x} , \bar{y}) , w \rangle \\\nonumber
&\mbox{subject to}& \lambda_1 \nabla_{y} \psi (\bar{x} , \bar{y}) + \lambda_2 \nabla_{y} h (\bar{x} , \bar{y})= - \nabla_{y} \varphi(\bar{x} , \bar{y}) ,\\\nonumber
&& \lambda_1 \geq 0 , \lambda_2 \geq 0
\end{eqnarray}
Thanks to the strong duality argument in linear programming, problems \eqref{lp} and \eqref{dlp} have the same optimal value which is $- \langle \nabla_{x} \varphi (\bar{x} , \bar{y}) , w \rangle$. Therefore, the dual necessary optimality condition for the bilevel problem \eqref{bilevel} in direction $w \in D$ amounts to existence of $\lambda_1 , \lambda_2 \in {\rm I\!R}$ such that
\begin{equation}\label{sipeqop1}
\left\{\begin{matrix}
\displaystyle \lambda_1 V'(\bar{x} ; w) - \lambda_1 \langle \nabla_{x} \psi (\bar{x} , \bar{y}) , w \rangle - \lambda_2 \langle \nabla_{x} h (\bar{x} , \bar{y}) , w \rangle = \langle \nabla_{x} \varphi (\bar{x} , \bar{y}) , w \rangle ,\\\\
\lambda_1 \nabla_{y} \psi (\bar{x} , \bar{y}) + \lambda_2 \nabla_{y} h (\bar{x} , \bar{y})= - \nabla_{y} \varphi(\bar{x} , \bar{y}),\\\\
\lambda_1 \geq 0 , \lambda_2 \geq 0.
\end{matrix}\right.
\end{equation}
However, in this paper, we are not dealing with dual optimality conditions as we believe that in non-amenable structures dual elements might not carry helpful information. To the best of our knowledge the primal optimality condition obtained in Example \ref{obilevel}, firstly, is the sharpest optimality condition for a general bilevel problem, secondly, it requires minimal assumptions compared to the other similar results obtained in the literature, in particular, the ones in the very recent paper \cite[Theorems~ 5.2, 5,3]{by}.
}
\end{Remark}
\section{Subderivative Method}\setcounter{equation}{0} \label{sec03}
In the sequel, we focus on the following unconstrained optimization problem
\begin{equation}\label{uncons}
\mbox{minimize} \;\; f(x) \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n ,
\end{equation}
where $f: {\rm I\!R}^n \to \Bar{\R} $ is calm at all $x \in {\rm I\!R}^n$ from below. We are going to propose a first-order algorithm to obtain a d-stationary point for problem \eqref{uncons} up to some positive tolerance $\varepsilon > 0$. To make the latter precise, fix $\varepsilon > 0 $. We say $\bar{x}$ is a $\varepsilon$-d(irectional) stationary point of problem \eqref{uncons} if ${\rm d} f(\bar{x})(w) \geq - \epsilon$ for all $w \in {\rm I\!R}^n$ with $\|w\| \leq 1$. Since we do not require $f$ takes only finite values, we can incorporate the constraint set into objective funtion of \eqref{uncons} with indicator function. However, from computation purpose, it is better we incorporate constraint sets into objective function by a finite exact penalty. More precisly, in the constrained optimization problem
\begin{equation}\label{conso}
\mbox{minimize} \;\; \varphi(x) \quad \mbox{subject to}\;\; G(x) \in X ,
\end{equation}
the constraint set can be incorprated into objective function by the penalty term $\rho \; {\rm dist} (G(x) \; ; \; X)$ where $\rho > 0$ is the penalty constant. Therefore, setting $f(x) := \varphi(x) + \rho \; {\rm dist} (G(x) \; ; \; X)$ in \eqref{uncons} we can cover constrained optimization problems.
\vspace{.5cm}
\begin{table}[H]
\renewcommand{\arraystretch}{1}
\centering
\label{alg1}
\begin{tabular}{|
p{0.9\textwidth} |}
\hline
\textbf{Subderivative Method}\\
\begin{enumerate}[label=(\alph*)]
\item[0.] \textbf{(Initialization)} Pick the tolerance $\epsilon \geq 0$, starting point $x_0 \in {\rm I\!R}^n$, and set $k=0 .$
\item[1.] \textbf{(Termination)} If $\min_{\|w\| \leq 1} {\rm d} f(x_k)(w) \geq - \varepsilon $ then Stop.
\item[2.] \textbf{(Direction Search)} Pick $w_k \in \mbox{arg min}_{\|w\| \leq 1 } {\rm d} f(x_k)(w).$
\item[3.] \textbf{(Line Search)} Choose the step size $\alpha_{k} > 0$ through a line search method.
\item[4.] \textbf{(Update)} Set $x_{k+1} : = x_k + \alpha_k w_k$ and $k+1 \leftarrow k$ then go to step 1.
\end{enumerate}
\\ \hline
\end{tabular}
\end{table}
\vspace{.3cm}
\begin{Remark}[\bf comments on the subderivative method]{\rm
\begin{itemize}
\item[]
\item[\textbullet] The subderivative method does not require any more assumptions on $f$ in order to be well-defined. Indeed, for any $f: {\rm I\!R}^n \to \Bar{\R}$ calm at each $x \in {\rm I\!R}^n $ from below, the subderivative ${\rm d} f(x)(.)$ is proper and lower semicontinuous for each $x \in {\rm I\!R}^n$. Therefore, the minimization sub-problem in step 2 admits a solution.
\item[\textbullet] The rich calculus of the subderivative established in section \ref{sect02} eases the computation of ${\rm d} f$ for most functions, in particular, the non-amenable composition. These calculus are not available for other primal tools such as Clarke directional derivatives. The latter is the prime reason we assign and specify the name ``subderivative method" over ``a generalized gradient method".
\item[\textbullet] The minimization sub-problem in step 2 does not need to be solved exactly, we only need to find an $w_k$ with ${\rm d} f(x_k)(w_k) < - \varepsilon .$ If $f$ is Clarke regular at $x_k$, then the sub-problems with subderivative method are convex programmings, indeed, in this case ${\rm d} f(x_k)(.)$ is a sub-linear function. In another case, if $f$ is differentiable at $x_k$ (not necessarly Clarke regular) the solution of the sub-problem in step 2 is $w_k = - \frac{\nabla f(x_k)}{\| \nabla f(x_k) \|}$, and step 1 is simply checking if $\| \nabla f(x_k) \| \leq \varepsilon $. Luckily, we mostly deal with locally Lipschitz continuous functions; hence, it is likely that on many iterations we just have $w_k = - \frac{\nabla f(x_k)}{\| \nabla f(x_k) \|}$, as Lipschitz continuous functions are differentiable almost everywhere by the Rademacher theorem.
\item[\textbullet] If $f$ is semi-differentiable, as it is in the all examples in section \ref{sect01}, then the subderivative method becomes a descent method; see Lemma \ref{backt}. In particular, in the constrained optimization \eqref{conso}, the penalized function $f(x) := \varphi (x) + \rho \; {\rm dist}(G(x) , X)$ is semi-differentiable, if $X$ is derivable, and $G$ and $\varphi$ are semi-differentiable; See Proposition \ref{semidiff-chain} and Example \ref{semidist}. Hence, the subderivative method applied on $\varphi (x) + \rho \; {\rm dist}(G(x) , X)$ is a descent method.
\end{itemize}
}
\end{Remark}
\subsection{Convergence Analysis}
In order to establish the convergence of the subderivative method, we need that the generated directions (eventually) decrease the objective function sufficiently with a step size under consideration. Therefore, we propose the following condition.
\begin{Definition}[\bf descent property]\label{def_descent_pro}
Let $f : {\rm I\!R}^n \to \Bar{\R}$ and $\Omega \subseteq {\rm I\!R}^n$. We say $f$ has a descent property on $\Omega$ if there exists a $L:=L(f , \Omega) \geq 0$ such that for all $x , y \in \Omega$ we have
\begin{equation}\label{descet_pro}
f(y) \leq f(x) + {\rm d} f(x) (y-x) + \frac{L}{2} \| y -x \|^2
\end{equation}
\end{Definition}
In the above definition, we do not require $f$ to be differentiable, not even Lipschitz continuous. If $f$ is continuously differentiable and its derivative is Lipschitz continuous, then \eqref{descet_pro} holds on $\Omega = {\rm I\!R}^n$. The latter is known as the \textit{descent lemma}; see \cite[Lemma~4.22]{b}. An important class of non-amenable functions that enjoys the above descent property is the class of non-smooth concave functions. Furthermore, it is not difficult to verify that the descent property is preserved under addition and non-negative scalar multiplication. The composite function $f = g \circ F$ satisfies descent property \eqref{descet_pro}, if $g$ is concave, and $F$ is $L-$smooth function, i.e. $\nabla F(.)$ is Lipschitz continuous with a constant $L$. A striking observation is that the Moreau envelope of any bounded-below function enjoys the descent property \eqref{descet_pro} without assuming convexity or weak convexity. More precisely, if $f : {\rm I\!R}^n \to \Bar{\R}$ is prox-bounded with a constant $c > 0$, i.e., $f + \frac{1}{2c} \| . \|^2$ is bounded below \cite[Definition~1.23]{rw} , then $e_{r} f(x) := \inf_{y\in {\rm I\!R}^n}\{ \frac{1}{2r} \| x - y \|^2 + f(y) \}$ satisfies the descent property for all $r \in ( 0 , c )$. We gather the above observations in the following proposition.
\begin{Proposition}[\bf sufficient conditions for the descent property]\label{sufdes}
Let $f, \; g: {\rm I\!R}^n \to \Bar{\R}$, $\Omega \subseteq {\rm I\!R}^n$, and $F: {\rm I\!R}^n \to {\rm I\!R}^m$. Then, the following assertions hold
\begin{itemize}
\item[1.] if $f$ is $L$-smooth, then $f$ has the descent property \eqref{descet_pro} on ${\rm I\!R}^n$.
\item[2.] if $\lambda > 0$, and $f$ satisfies the descent property \eqref{descet_pro} on $\Omega$, then $\lambda f$ has the descent property on $\Omega$.
\item[3.] if $f$ and $g$ satisfy the descent propert on $\Omega$ so is $f + g$.
\item[4.] if $g$ is a finite-valued concave function, then $g$ satisfies the descent property on ${\rm I\!R}^n$ with the constant $L = 0$.
\item[5.]if $\Omega$ is a convex-compact set, $F$ is $L$-smooth on $\Omega$, and $g$ is a finite-valued concave function, then $g \circ F$ has the descent property on $\Omega$ with a constant $L$.
\item[6.] if $f$ is prox-bouded with a constant $c$, then for all $r \in (0, c), $ $e_{r} f$ has the descent property with a constant $L = \frac{1}{r}$.
\item [7.] if $\Omega$ is a convex-compact set, $F$ is $L$-smooth on $\Omega$, and $g$ is prox-bounded with a constant $c$, then for each $r \in (0 , c)$, $[e_r g] \circ F$ has the descent property on $\Omega$ with a constant $\frac{L}{r} .$
\item [8.] Let $g_i : {\rm I\!R}^{m_i} \to \Bar{\R} $ be convex functions, and let $F_i: {\rm I\!R}^{n} \to {\rm I\!R}^{m_i}$ be $L$-smooth functions for each $i=1,2$, then for each $r \in (0 , \infty)$, $f:= [\mbox{\rm e}\,_r g_1 ] \circ F_1 - g_2 \circ F_2 $ has the descent property on ${\rm I\!R}^n$ with a constant $\frac{L(1+r)}{r}.$
\end{itemize}
\end{Proposition}
\begin{proof}
(1) is a classical result known as the descent lemma \cite[Lemma~4.22]{b} . (2) is obvious. To prove (3), observe from descent property that $ - \infty < {\rm d} f(x)(w)$ for all $x , w \in {\rm I\!R}^n$. Hence, by defnition of ``$\liminf$", we get ${\rm d} f(x)(w) + {\rm d} g(x)(w) \leq {\rm d} (f+g)(x)(w)$ for all $x , w \in {\rm I\!R}^n $, without imposing any qualification condition for sum. Now the descent property for $f+g$ follows from adding the corresponging inequalities in \eqref{descet_pro} for both $f$ and $g$. Turning to (4), note that finite-valued concave functions are locally Lipshitz continuous and directionally differentiable, thus they are semi-differentiable with $g' (x; w) = {\rm d} g(x)(w) $ for all $x ,w \in {\rm I\!R}^n$. Therfore, (4) immedietly follows from the inequality for concave functions
$$ g(y) - g(x) \leq g' (x ; y -x) .$$
To prove (5), first we recall some properties of ${\rm d} g$. Ineed, since $g$ is concave, ${\rm d} g(z)(.)$ is concave and positively homeogeneous for all $z \in R^m$. Hence, for all $a , b \in {\rm I\!R}^m$ one has
${\rm d} g(z)(a) \leq {\rm d} g(z)(a+b) - {\rm d} g(z)(b)$. Moreover, $| {\rm d} g(z)(w) |\leq \ell \|w\|$, where $\ell > 0$ is any Lipschitz constant of $g$ on any neighborhood of $z$. We use the latter facts shortly. Now set $\eta (x , y) := F(y) - F(x) - \nabla F(x) (y -x) $. Since $F$ is $L$-smooth on $\Omega$, we have $\|\eta (x ,y)\| \leq \frac{ L}{2} \; \|y -x\|^2 $ for all $x ,y \in \Omega$. Now fix $x ,y \in \Omega$, by concavity of $g$ we have
\begin{eqnarray*}
\displaystyle
g(F(y))& \leq & g(F(x)) + {\rm d} g(F(x))(F(y)-F(x))\\\nonumber
&=&g(F(x)) + {\rm d} g(F(x))\big(\nabla F(x) (y -x) + \eta (x , y)\big)\\\nonumber
&\leq&g(F(x)) + {\rm d} g(F(x))\big(\nabla F(x) (y -x)\big) - {\rm d} g(F(x)) \big(-\eta (x , y) \big)\\\nonumber
&\leq& g(F(x)) + {\rm d} g(F(x))\big(\nabla F(x) (y -x)\big) + \ell \| - \eta (x,y) \| \\\nonumber
&\leq& g(F(x)) + {\rm d} (g \circ F) (x)(y -x) + \frac{\ell L}{2} \| y - x \|^2 \\\nonumber
\displaystyle
\end{eqnarray*}
where $\ell > 0$ is any Lipschitz constant of $g$ over the set $\Omega + \mathbb{B}$. Note that in the third and forth inequalities we used the some properties of ${\rm d} g$ we discussed earlier. In the last inequality, we used the chain rule \eqref{fcalc1} . Part (6), immediately follows from part (3) and (4) and the fact that the Moreau envelope of any function is a sum of a $L$-smooth function and a concave function. Indeed, $e_r f (x) > - \infty$ for all $r \in (0 , c)$, and we have
$$ e_r f (x) = \frac{1}{2r} \| x \|^2 + \inf_{y} \{ - \frac{1}{r} \langle x , y \rangle + \frac{1}{2r} \|y\|^2 + f(y) \} $$
noting that the infimum over an arbitrary number affine functions produces a concave function. To prove (7), we write $ e_r g (z) = \varphi (z) + h(z)$, where $\varphi$ is a $\frac{1}{r}$-smooth function and $h$ is a concave function.
Hence, we have $([e_r g] \circ F) (x) = (\varphi \circ F)(x) + (h \circ F)(x) $. Observe that $\varphi \circ F$ is $\frac{L}{r}$-smooth function, and $h \circ F$ satisfies the descent property by (5), thus (7) follows from (2). Assertion (8) follows from $(7)$, $(5)$, and $(3).$
\end{proof}
Observe that in (5), the compactness of $\Omega$ is redundant if $g$ is Lipschitz continuous on $F(\Omega)$. It is well-known that the Moreau envelope of convex functions is $L$-smooth. Thus (6) can be read as a far generalization of the smoothening results to any prox-bounded functions, in particular, functions that are bounded from below. Assertion (8) pertains to the difference of amenable functions, in particular, the difference of convex functions when $F_i$ are taken identity maps; see Example \ref{dame}. The difference of amenable functions contains a broad class of non-amenable functions. In \cite{m22}, we show that the sub-problems of the subderivative method can be accurately solved for some interesting sub-class of difference of amenable functions, in particular, \eqref{dmax}. As we will see later, the convergence of the subderivative method solely depends on the descent property, thus applicable to the Moreau envelope of any bounded-below function, no matter if a function is highly discontinuous or irregular. For instance, the Moreau envelope of the norm-zero function, which has a closed-form; see Example \ref{spars}. The following result shows a rudimentary convergence of the subderivative method with a pre-defined line search.
\begin{Lemma}[\bf convergence of subderivative method with a diminishing line search]\label{Th_easyconve}
Suppose $f: {\rm I\!R}^n \to \Bar{\R}$ satisfy the descent property \eqref{descet_pro} on ${\rm I\!R}^n$. Then, the subderavative method with the positive tolerance $\varepsilon > 0$ and the step-size $\alpha_k > 0$ defined by any sequence $\{\alpha_k\}_{k=0}^{\infty}$ satisfying $\sum_{k} \alpha_k = \infty ,$ and $\sum_{k} \alpha_k^2 < \infty $, either stops after finite iterations, or it generates a sequence $\{x_k\}_{k=0}^{\infty}$ for which $f(x_k) \to -\infty$.
\end{Lemma}
\begin{proof}
We suppose the subderivative method does not stop at any iteration, meaning that for the generated vector $x_k$, the corresponding descent direction satisfies ${\rm d} f(x_k) (w_k) < - \varepsilon .$ Plugging $x_{k+1} = x_k + \alpha_k w_k$ into \eqref{descet_pro} we get
\begin{eqnarray*}
\displaystyle
f(x_{k+1}) - f(x_k) &\leq & {\rm d} f(x_k) (x_{k+1} - x_k) + \frac{L}{2} \| x_{k+1} - x_k \|^2 \\\nonumber
&=& \alpha_k {\rm d} f(x_k)(w_k) + \frac{ L}{2} \alpha_k^2\\\nonumber
&\leq & - \varepsilon \alpha_k + \frac{L}{2} \alpha_k^2 \\\nonumber
\displaystyle
\end{eqnarray*}
By summing up both sides of the above inequality over $k=0,1, ...$ we get
$$ \lim_{k \to \infty }f(x_{k+1}) - f(x_0) \leq - \varepsilon \sum_{k=0}^{\infty}\alpha_k + \frac{L}{2} \sum_{k=0}^{\infty} \alpha_k^2 = - \infty $$
This finishes the proof.
\end{proof}
\begin{Remark}[\bf Armijo backtracking line search] {\rm
As the proof of Lemma \ref{Th_easyconve} confirms, in the subderivative method, one does not need to solve the minimization sub-problems accurately. Any direction $w_k$ with ${\rm d} f(x_k)(w_k) < -\varepsilon$ will do the job. Although the step-size chosen in Lemma \eqref{Th_easyconve} is simple and can be determined independently from $x_k$ and $w_k$, it may cause slow convergence. One of the most efficient and practically used line searches is \textit{Armijo backtracking} method which can be adapted to the subderivative version; indeed, fix the parameter $\mu \in (0 , 1)$, called a reduction multiple, and assume that we are in the $k^{th} -$iteration. The Armijo backtracking line search determines the step-size $\alpha_k >0$ in the following way:
If the following inequality holds for $\alpha_k = 1$, then the step-size is chosen $\alpha_k =1$.
\begin{equation}\label{backtrine}
f(x_k + \alpha_k w_k) - f(x_k) <\frac{\alpha_k}{2} {\rm d} f(x_k) (w_k).
\end{equation}
Otherwise, keep updating $\alpha_k$ by multiplying it by $\mu$ until the above inequality holds. In the next lemma, we show that backtracking method terminates after finite numbers of updates onder some mild assumptions.}
\end{Remark}
\begin{Lemma}[\bf finitness of Armijo backtracking process]\label{backt}
In the each of the following situations, the Armijo backtracking process to find the step-size $\alpha_k$, within the subderivative algorithm, terminates after finite number of iterations.
\end{Lemma}
\begin{itemize}
\item[(i)] $f$ is directionally lower regular (in particular, semi-differentiable) at $x_k$ for each $k=1,2...$
\item[(ii)] $f$ satisfies the descent property \eqref{descet_pro} on ${\rm I\!R}^n .$
\end{itemize}
\begin{proof}
Let $x_k , w_k \in {\rm I\!R}^n$ with ${\rm d} f(x_k) (w_k) < 0$. Hence, ${\rm d} f(x_k) (w_k) < \frac{1}{2} {\rm d} f(x_k) (w_k).$ In situation (i), we have
$$ {\rm d} f(x_k) (w_k) = \lim_{ t \downarrow 0} \frac{f(x_k + t w_k) - f(x_k)}{t} < \frac{1}{2} {\rm d} f(x_k) (w_k) $$
hence, there exists $m \in {\rm I\!N}$ sufficiently large such that
$$ f(x_k + \mu^{m} w_k) - f(x_k) <\frac{\mu^{m}}{2} {\rm d} f(x_k) (w_k) .$$
In situation (ii), for all $t >0 $, the descent property \eqref{descet_pro} gives us
\begin{eqnarray*}
\displaystyle
\frac{f(x_k + t w_k) - f(x_k)}{t} \leq {\rm d} f(x_k)(w_k) + \frac{ L}{2} t .\\\nonumber
\displaystyle
\end{eqnarray*}
Define $m$ by the smallest non-negative integer number such that $$ {\rm d} f(x_k)(w_k) + \frac{ L}{2} \mu^m < \frac{1}{2} {\rm d} f(x_k)(w_k) $$
where $\mu \in (0 ,1)$ is the reduction multiple in the Armijo backtracking method. Thus, it can be observed that \eqref{backt} holds for $\alpha_k = \mu^{m}$.
\end{proof}
In the following theorem, we prove that the subderivative method with Armijo backtracking line search, intialized on an aribitrary vector in ${\rm I\!R}^n$, finds a $\varepsilon$-$d$-stationary point after atmost $O(\varepsilon^{-2})$ iterations.
\begin{Theorem}[\bf convergence rate of the subderivative method]\label{rate_TH}
Let $\{x_k\}_{k =0}^{\infty}$ be a sequence generated by the subderivative method with the Armijo backtracking line search. Assume that $f$ is bounded below and has the descent property \eqref{descet_pro} on ${\rm I\!R}^n$. Then, for sufficiently large number $N \in {\rm I\!N}$, the following inequality holds
\begin{equation}\label{rate}
\min_{ 0 \leq k \leq N} |{\rm d} f(x_k)(w_k) | \leq \sqrt{\frac{f(x_0) - f^{*}}{M (N+1)}}
\end{equation}
where $f^*$ is a lower bound of $f$ and $M:=\min\{\frac{1}{2} , \frac{\mu}{2L}\} $.
\end{Theorem}
\begin{proof} Before proving \eqref{rate}, let us first prove the following claim that guarantees the sufficient decrease of the subderivative method under the descent propety assuption.
\textbf{Claim:} For each $k\in {\rm I\!N} \cup \{0\}$, the follwing inequality holds
\begin{equation}\label{suf-dec1}
f(x_{k+1}) - f(x_k) \leq - M \; \min\{ |{\rm d} f(x_k)(w_k)| \; , \; [{\rm d} f(x_k)(w_k)]^2 \}
\end{equation}
proof: set $\delta_k := \min\{ |{\rm d} f(x_k)(w_k)| , [{\rm d} f(x_k)(w_k)]^2 \}$. We suppose we are in the $k^{th}-$ iteration of the subderivative method and moving toward to the next generated point. This ensures the existence of the direction $w_k \in {\rm I\!R}^n$ with ${\rm d} f(x_k)(w_k) < 0$. By the descent property \eqref{descet_pro} for all $t >0$ we have
\begin{eqnarray}\label{sufeq1}
\displaystyle
f(x_{k} + t w_k) - f(x_k) &\leq & {\rm d} f(x_k) (tw_k) + \frac{Lt}{2} \| w_k \|^2 \\\nonumber
&=& t {\rm d} f(x_k)(w_k) + \frac{ L t^{2}}{2}
\displaystyle
\end{eqnarray}
Let $\alpha_k$ be determined by the backtracking procedure, meaning that the stepsize $\frac{\alpha_k}{\mu}$ is not acceptable and does not satisfy \eqref{backtrine}:
\begin{equation}\label{sufdec3}
f(x_k + \frac{\alpha_k}{\mu} w_k ) - f(x_k) > \frac{\alpha_k}{2\mu} {\rm d} f(x_k)(w_k).
\end{equation}
Combining the above inequality with \eqref{sufeq1} we get
\begin{eqnarray*}
\displaystyle
& & \frac{\alpha_k}{2\mu} {\rm d} f(x_k)(w_k) < \frac{\alpha_k}{ \mu} {\rm d} f(x_k)(w_k) + \frac{ L \alpha_k^{2}}{ 2\mu^2}\\\nonumber
&\rightarrow & {\rm d} f(x_k)(w_k) < 2 {\rm d} f(x_k)(w_k) + \frac{ L \alpha_k}{ \mu}\\\nonumber
&\rightarrow & - \frac{ \mu {\rm d} f(x_k)(w_k)}{L} < \alpha_k \\\nonumber
&\rightarrow & -\frac{\mu}{2L} [{\rm d} f(x_k)(w_k)]^2 > \frac{\alpha_k}{2} {\rm d} f(x_k)(w_k). \\\nonumber
\displaystyle
\end{eqnarray*}
Combining the above last inequality with \eqref{backtrine} we get
$$ f(x_k + {\alpha_k} w_k ) - f(x_k) \leq \frac{\alpha_k}{2} {\rm d} f(x_k)(w_k) < -\frac{\mu}{2L} [{\rm d} f(x_k)(w_k)]^2 \leq -M \delta_k .$$
This completes the proof of the claim. Turning to the proof of \eqref{rate}, by summing up the both sides of \eqref{suf-dec1} we get
\begin{eqnarray}\label{conv1}
\displaystyle
&& M \sum_{k=0}^{N} \delta_k \leq \sum_{k=0}^{N} f(x_k) - f(x_{k+1})\\\nonumber
&\rightarrow& (N+1)M \min_{ 0 \leq k \leq N} \{\delta_k \} \leq f(x_0) - f(x_{N+1}) \leq f(x_0) - f^*\\\nonumber
& \rightarrow & \min_{ 0 \leq k \leq N} \{\delta_k\} \leq \frac{f(x_0) - f^* }{M (N+1)}
\displaystyle
\end{eqnarray}
Above inequality holds for all non-negative integer $N$. Additionally, if $N$ is sufficiently large, we have $\min_{ 0 \leq k \leq N} \{\delta_k\} < 1$. Considering the definition of $\delta_k$, for such $N$ we have
$$ \min_{ 0 \leq k \leq N} \{\delta_k\} = \min_{ 0 \leq k \leq N} \{ [{\rm d} f(x_k)(w_k)]^2 \} $$
this together with last inequality obtained in \eqref{conv1} proves \eqref{rate}.
\end{proof}
In Theorem \eqref{rate_TH}, we really do not need to impose the descent property on entire ${\rm I\!R}^n$. For instance, if $f$ is semi-differentiable and coercive, we can assume the descent property locally on ${\rm I\!R}^n$. In the following theorem, similar to the gradient descent method for smooth minimization, we show that we only need to have the descent property on a sublevel set containing the iterations of the algorithm.
\begin{Theorem}[\bf convergence of subderivative descent method]\label{rate_TH2}
Let $\{x_k\}_{k =0}^{\infty}$ be a sequence generated by the subderivative method with the Armijo backtracking line search. Set $\Omega := \{ x \big| \; f(x) \leq f(x_0)\} $. Assume that $f$ is bounded-below, semi-differentiable on $\Omega$, and has the descent property \eqref{descet_pro} on $\Omega + B$. Then, for sufficiently large number $N \in {\rm I\!N}$, \eqref{rate} holds.
\end{Theorem}
\begin{proof}
First, we prove inequality \eqref{suf-dec1} when the descent property \eqref{descet_pro} is only assumed on $\Omega + B.$ To prove the latter, by Lemma \eqref{backt} (i), the subderivative method is a descent method, because $f$ is semi-differentiable on $\Omega$. Hence, for each $k \in {\rm I\!N}$, we have $x_k \in \Omega$. Indeed, since $f$ is semi-differentiable, for each $ k \in {\rm I\!N}$, we have ${\rm d} f(x_k)(w_k) = f' (x_k ; w_k) <0 $. Therefore, for each $k \in {\rm I\!N}$, there exists a $\varepsilon_k >0 $ such that $x_k + t w_k \in \Omega$ for all $t \in [0, \varepsilon_k]$, hence, $x_k + t w_k \in \Omega + \mathbb{B}$ for each $t \in [0, 1 + \varepsilon_k]$. Now proceeding with the proof of inequality \eqref{suf-dec1}, inequality \eqref{sufeq1} holds for all $t \in [0, 1 + \varepsilon_k]$. Now if $\alpha_k \neq 1$ is determined by the Armijo backtracking procedure, inequality \eqref{sufdec3} holds with $\alpha_k = \mu^m$ for some positive integer $m$. Since $ \frac{\alpha_k}{\mu} \in (0 ,1 + \varepsilon_k]$, we get $x_k + \frac{\alpha_k}{\mu} w_k $. Form here, the rest of the proof, including the proof of \eqref{rate}, agrees line by line of the proof of Theorem \ref{rate_TH} after inequality \eqref{sufdec3}.
\end{proof}
\begin{Remark}[\bf comments on convergence results]{\rm
Since norms are equivalent in finite dimensions, the descent property \eqref{descet_pro} is invariant under the choice of the norm, i.e., \eqref{descet_pro} holds if and only it holds with a quadratic term $\|y-x\|_{*}^2$ where $\| . \|_{*}$ is any norm on ${\rm I\!R}^n$. Therefore, if in the subderivative method, one solves the sub-problems with respect to any norm $\|.\|_{*}$, i.e., $\| w\|_{*} \leq 1$, the convergence results, Theorems \ref{rate_TH} and \ref{rate_TH2}, remain true. The only thing that will change is the constant $M$ in \eqref{rate}. As we will see shortly, changing the norm sometimes leads to solving the sub-problems of the subderivative method accurately. A close look at the proof of Theorem \ref{rate_TH2}, reveals that the assumption of semi-differentiability can be replaced by directional lower regularity, as we only used the ${\rm d} f(x)(w) = f' (x;w)$ as result of semi-differentiability. The latter improvement might seem minor, yet there are functions which are directionally lower regular but not semi-differentiable; e.g. $f(x) = \| x \|_{0}$, Example \ref{subzero}.}
\end{Remark}
\begin{Example}[\bf problems with separable-variable subderivative]\label{sep} {\rm Let $\varphi : {\rm I\!R}^n \to {\rm I\!R}$ be a differentiable function, and $ g : {\rm I\!R}^n \to {\rm I\!R}$ be a function with separable-variable subderivative, i.e, there exist one-variable functions $g_i : {\rm I\!R} \to {\rm I\!R}$ , $i =1,...,n$ such that for each $x \in {\rm I\!R}^n$ and each $w \in {\rm I\!R}^n$ one has $ {\rm d} g(x)(w) = \sum^{n}_{i=1} g_i (w_i ) $. In this example, we consider the following minimization problem:
\begin{equation}\label{sep1}
\mbox{minimize} \;\; f(x):= \varphi (x) + g(x) \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n.
\end{equation}
For each $x \in {\rm I\!R}^n$ and each $i \in \{ 1,...,n \}$, let $s_i$ and $w_i$ be the optimal value and an optimal solution of the following one-variable minimization problem respectively
\begin{equation}\label{sep2}
P(x , i): \quad \quad \mbox{minimize} \;\; \frac{\partial \varphi (x)}{ \partial x_i} t + h_i(t) \quad \mbox{subject to}\;\; t \in [-1 , +1].
\end{equation}
Then, the subderivative method for problem \eqref{sep1}, with the norm-choice $\| . \|_{\infty}$ in sub-problems, reduces to the following process
\begin{table}[H]
\renewcommand{\arraystretch}{0}
\centering
\label{alg1}
\begin{tabular}{|
p{0.9\textwidth} |}
\hline
\textbf{}\\
\begin{enumerate}[label=(\alph*)]
\item[0.] \textbf{(Initialization)} Pick the tolerance $\varepsilon > 0$, starting point $x_0 \in {\rm I\!R}^n$, and set $k=0 .$
\item[1.] \textbf{(Termination)} $\sum^{n}_{i=1} s^k_i \geq - \varepsilon $ where each $s^k_i$ is optimal value of the problem $P(x_k , i)$ in \eqref{sep2}.
\item[2.] \textbf{(Direction Search)} The components of the descent direction $w_k$ are $w_k^i$, an optimal solution of problem $P(x_k , i)$.
\item[3.] \textbf{(Line Search)} Choose the step size $\alpha_{k} > 0$ through a line search method.
\item[4.] \textbf{(Update)} Set $x_{k+1} : = x_k + \alpha_k w_k$ and $k+1 \leftarrow k$ then go to step 1.
\end{enumerate}
\\ \hline
\end{tabular}
\end{table}
}
\end{Example}
Detail for the direction search: given the iteration $x_k$, the subderivative of $f$ at $x_k$ in direction $w$ is $$ {\rm d} f(x_k)(w) = \langle \nabla \varphi (x) , w \rangle + {\rm d} g(x_k) (w) = \sum_{i = 1}^{n} \big( \frac{\partial \varphi (x_k)}{ \partial x_i} w_i + g_i(w_i) \big) .$$
On the other hand, the constraint $\| w \|_{\infty} \leq 1$ is fully separable, and it is equivalent to $w_i \in [-1 , +1]$ for all $i =1,...,n$. The latter makes the sub-problem of direction search fully separable in $w$. Indeed,
$$ \min_{\| w \|_{\infty} \leq 1} {\rm d} f(x_k)(w) =\sum_{i = 1}^{n} \big( \min_{| w_i |\leq 1} \frac{\partial \varphi (x_k)}{ \partial x_i} w_i + g_i(w_i) \big) = \sum_{i = 1}^{n} s_i^k .$$
Examples of non-differentiable functions with separable-variable subderivative are $g(x) = \| x\|_{1}$ and $g(x) = - \|x\|_{1}$. Indeed, in the case $g(x) = \lambda \| x\|_1$ where $\lambda >0$, we have
$$ {\rm d} g(x)(w) = \sum_{i \in I_1 \cup I_2} \lambda w_i + \sum_{i \in I_3 \cup I_4} (-\lambda w_i) + \sum_{i \in I_5 \cup I_6 \cup I_7} \lambda |w_i| $$
where
$$I_1 (x) = \{ i| \; x_i >0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) + \lambda \geq 0 \}, \quad I_2 (x) =\{ i| \; x_i >0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) + \lambda < 0 \}, $$
$$ I_3 (x) = \{ i| \; x_i < 0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) - \lambda \geq 0 \}, \quad I_4 (x) = \{ i| \; x_i < 0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) - \lambda < 0 \}$$
$$ I_5 (x) =\{ i| \; x_i = 0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) - \lambda \geq 0 \}, \quad I_6 (x) = \{ i| \; x_i = 0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) + \lambda \leq 0 \}$$
$$ I_7 (x) = \{ i| \; x_i = 0 \; ,\; \frac{\partial \varphi}{\partial x_i} (x) + \lambda > 0, \quad \frac{\partial \varphi}{\partial x_i} (x) - \lambda < 0 \} $$
Therefore, by Example \ref{sep}, the vector $w_k=(w_k^1,w_k^2,...,w_k^n)$ defined by
$$ w_{k}^i := \begin{cases}
-1 & i \in I_1 \cup I_3 \cup I_5 \\
1& i \in I_2 \cup I_4 \cup I_6 \\
0& i \in I_7
\end{cases} $$
solves the sub-problems of the subderivative method.
\begin{Example}[\bf a sub-class of difference of amenable functions]\label{cc} {\rm Let $\varphi : {\rm I\!R}^n \to R$ and $F:{\rm I\!R}^n \to {\rm I\!R}^m$ be $L$-smooth functions. Let $g: {\rm I\!R}^m \to {\rm I\!R}$ be a convex function. In this example, we consider a non-regular composite minimization problem in the following framework
\begin{equation}\label{cc1}
\mbox{minimize} \;\; f(x):= \varphi (x) - (g \circ F)(x) \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n.
\end{equation}
For each $i \in \{1,2,...,n\}$, let $e_i \in {\rm I\!R}^n$ be the vector whose $i$-th component is $1$ and zero for other components. Define the finite set $E \subset {\rm I\!R}^n$ with $2n$ elements by $E:=\{\pm e_i |\; i=1,2,...,n\}.$ Then, the subderivative method, with the norm-choice $\|.\|_{1}$, for problem \eqref{cc1} reduces to the following process
\begin{table}[H]
\renewcommand{\arraystretch}{0}
\centering
\label{alg1}
\begin{tabular}{|
p{0.9\textwidth} |}
\hline
\textbf{}\\
\begin{enumerate}[label=(\alph*)]
\item[0.] \textbf{(Initialization)} Pick the tolerance $\varepsilon > 0$, starting point $x_0 \in {\rm I\!R}^n$, and set $k=0 .$
\item[1.] \textbf{(Termination)} $$s_k :=\min \{ \pm \frac{\partial \varphi (x_k)}{x_i} \mp {\rm d} g(F(x_k))(\nabla F(x_k)e_i) | \; i=1,2,...,n \} \geq - \varepsilon$$.
\item[2.] \textbf{(Direction Search)} Pick $w_k \in E$ with the minimum value
$$ \langle \nabla \varphi (x_k) , w_k \rangle -{\rm d} g(F(x_k))(\nabla F(x_k)w_k) $$
\item[3.] \textbf{(Line Search)} Choose the step size $\alpha_{k} > 0$ through a line search method.
\item[4.] \textbf{(Update)} Set $x_{k+1} : = x_k + \alpha_k w_k$ and $k+1 \leftarrow k$ then go to step 1.
\end{enumerate}
\\ \hline
\end{tabular}
\end{table}
}
\end{Example}
Detail for the direction search: given the iteration $x_k$, the subderivative of $f$ at $x_k$ in direction $w$ is $$ {\rm d} f(x_k)(w) = \langle \nabla \varphi (x_k) , w\rangle -{\rm d} g(F(x_k))(\nabla F(x_k)w) .$$
Observe that ${\rm d} f(x_k)(.)$ is a concave function, thus at least one of its global minimizers subject to the polytube $P:= \{w| \; \|w\|_{1} \leq 1 \}$ happens on an extreme point of $P$. The set of all extreme points of $P$ is $E$ defined above. Therefore, the sub-problems of the subderivative method in the direction search are accurately solved by the procedure in the above table.
\begin{Remark}[\bf difference of amenable functions]{\rm In Example \ref{cc}, the sub-problems of the subderivative method reduced to the evaluating of the subderivative function, ${\rm d} f(x_k)(.)$, at $2n$ elements of the set $E$, where $n$ is the number of variables. It is worth to mention that we could reduce these $2n$ evaluations to $n+1$ evaluations by evaluating ${\rm d} f(x_k)(.)$ at vectors $\{e_i |\; i=1,...,n\} \cup \{\ - e \} $ where $e \in {\rm I\!R}^n$ is the vector with its components equals to $1$. Indeed, $P:= \mbox{\rm co}\, \big(\{e_i |\; i=1,...,n\} \cup \{\ - e \} \big) $ is a convex polytube in ${\rm I\!R}^n$ having origin in its interior. Therefore, there exists a norm $\|.\|_{*}$ such that $P=\{w| \; \| w \|_{*} \leq 1 \}$. Now it is enough to consider the subderivative method with the norm-choice $\|.\|_{*}.$ Although Example \ref{cc} deals with a sub-class of the difference of amenable functions, it can be used to solve all difference of amenable functions. Indeed, by applying a Moreau envelope on the convex part, the general case can be fitted into the framework \ref{cc1}. More precisely, in order to solve
\begin{equation*}\label{cc2}
\mbox{minimize} \;\; (g_1 \circ F_1)(x) - (g_2 \circ F_2)(x) \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n,
\end{equation*}
one can apply a Moreau envelope only on the convex function $g_1$ to smoothening $g_1 \circ F_1$, i.e.,
\begin{equation*}\label{cc3}
\mbox{minimize} \;\; \varphi(x) - (g_2 \circ F_2)(x) \quad \mbox{subject to}\;\; x \in {\rm I\!R}^n,
\end{equation*}
where $\varphi := [e_r g_1] \circ F_1$. The convergence analysis of the latter problem was discussed in Proposition \ref{sufdes} (8). In \cite{m22}, we investigate the difference of amenable functions in more detail and establish a relationship between the stationary points of the last two problems.
}
\end{Remark}
\textbf{Concluding remarks:} In this paper, we studied first-order variational analysis of non-regular functions, mainly the non-amenable composite functions. The approach was taken in this paper is purely primal. By establishing rich calculus rules for a suitable notion of directional derivatives, we could identify the stationary points in a large class of optimization problems, including the ones in the example section. We adopted a generalized gradient descent method to search for such stationary points. In particular, we showed that our algorithm applied on the Moreau envelope of any bounded-below function finds a $\varepsilon$-stationary point with the same rate that the smooth gradient descent does. Our future goal is to treat some examples in section \ref{sect01} numerically and solve the sub-problems of the subderivative method for some class of functions. In \cite{m22}, we give a full treatment of the numerical analysis of Example \ref{dmax}.
\small
|
\section{Introduction}
Let $f : \mathscr{M} \to \mathbb{C}$ be a holomorphic function on a domain $\mathscr{M} $ in $\mathbb{C}^{n+1}$ and assume that $f$ has no critical value different from $0$.
Let $\alpha$ be a complex number with a non negative real part and $N$ a positive integer. There are two natural ways to define a distribution on $\mathscr{M} $ whose restriction to $ \mathscr{M} \setminus \{f = 0\}$ is equal to $\vert f\vert^{2\alpha}f^{-N}$. The first one is given by the principal value method (see for instance \cite{H-L}):\\
Let $\xi$ be a $\mathscr{C}^\infty_c(\mathscr{M} )$ differential form on $\mathscr{M} $ of type $(n+1, n+1)$ and define for $\varepsilon > 0$ the distribution $T_{\alpha,N}^\varepsilon $ by
$$ \langle T_{\alpha,N}^\varepsilon, \xi \rangle := \int_{\vert f\vert \geq \varepsilon} \vert f\vert^{2\alpha}f^{-N}\xi.$$
We give in section 2 a rather short proof of the fact that the limit $T_{\alpha,N}$ when $\varepsilon$ goes to $0$ of $ \langle T_{\alpha,N}^\varepsilon, \xi \rangle $ exists and defines a distribution $T_{\alpha,N}$ on $\mathscr{M} $ whose restriction to $ \mathscr{M} \setminus \{f = 0\}$ is equal to $\vert f\vert^{2\alpha} f^{-N}$. \\
The second method is to show that for $\Re(\lambda)$ large enough and for any test differential form $\xi \in \mathscr{C}^\infty_c(\mathscr{M} )^{(n+1,n+1)}$ the function
$$ \lambda \mapsto \int_{\mathscr{M} } \vert f\vert^{2\lambda}f^{-N}\xi$$
is holomorphic and defines a holomorphic family of distributions on $\mathscr{M} $. Moreover this holomorphic family of distributions admits a meromorphic extension to the complex $\lambda$-plane with no pole when $\Re(\lambda) \geq 0$. Then,
$\alpha $ is not a pole and we define the distribution $S_{\alpha,N}$ on $\mathscr{M} $ as the value at $\lambda = \alpha$ of this meromorphic extension.\\
We give a proof of the existence of $T_{\alpha,N}$ and $S_{\alpha,N}$ and we prove the equality in $Db_{\mathscr{M}}$, the sheaf of distributions on $\mathscr{M}$, $T_{\alpha,N} = S_{\alpha,N}$.\\
Our main tools for these proofs are
\begin{enumerate}
\item The Asymptotic Expansion Theorem of fiber-integrals given in \cite{B2} (which uses Hironaka's Desingularization Theorem); see also \cite{BMa}.
\item The existence of a local Bernstein identity for $f$, due to J.E. Bjork in the analytic case (see \cite{Be} and \cite{Bj1}).
\end{enumerate}
As an application we first deduce of this result the absence of torsion for some $\mathcal{D}$-modules generated by the distributions constructed in the first part. In fact we prove more : any distribution in such a $\mathcal{D}$-module has the Standard Extension Property (compare with the result of \cite{B-K}). We recall the definition of the Standard Extension Property in the begining of the section 2.\\
Then we apply this Theorem to the determination of generators of the conjugate $\mathcal{D}$-modules of the $\mathcal{D}$-modules $\mathcal{N}_\lambda$ associated to $z(\sigma)^\lambda$, the power $\lambda $, where $\lambda$ is any complex number, of the (multivalued) root of the universal equation of degree $k$, $z^k + \sum_{j=1}^k (-1)^h\sigma_hz^{k-h} = 0$ whose structure is studied in \cite{B4}.\\
\section{Existence of the Principal Value}
\subsection{The Standard Extension Property}
First, we recall the definition of this property (see \cite{B3} in Appendix and \cite{B-K} in Paragraph 3.1 or \cite{Bj2})
\begin{defn} \label{SEP}
Let $\mathscr{M} $ be a complex manifold of pure dimension $n+1$ and let $T$ be a distribution on $\mathscr{M} $. We shall say that $T$ has the {\bf Standard Extension Property} if the following condition are fulfilled:
\begin{enumerate}
\item Outside a hypersurface $H$ in $\mathscr{M} $ the distribution $T$ is a $\mathscr{C}^\infty$ function.
\item For each point $x$ in $H$ there exists an open neighborhood $U$ of $x$, a local holomorphic equation $\{f = 0\} = H \cap U$ of $H$ in $U$ such that for any test differential form $\xi$ in $\mathscr{C}_c^\infty(U)^{(n+1,n+1)}$ we have
$$ \langle T, \xi \rangle = \lim_{\varepsilon \to 0} \int_{\vert f\vert \geq \varepsilon} T\xi .$$
\end{enumerate}
\end{defn}
It is easy to see that if the condition 2 of the previous definition is satisfied for some choice of local equation of $H$, it is satisfied for any other choice.\\
Note also that this property is clearly stable by multiplication by a $\mathscr{C}^\infty$ function but is not stable, in general, by the action of $\mathcal{D}_{\mathscr{M}}$. For instance the locally integrable function $1/\bar z$ defines a distribution on $\mathbb{C}$ which has the standard extension property, but $\partial_z(1/\bar z) = i\pi\delta_0$ is a non zero torsion element in $Db_{\mathbb{C}}$.\\
The following result is proved in \cite{B-K}
\begin{thm}\label{B-K}
Let $\mathcal{M}$ be a regular holonomic $\mathcal{D}_{\mathscr{M}}$-module and let $H$ be a hypersurface in $\mathscr{M} $ such that $\mathcal{M}_{\mathscr{M} \setminus H}$ is $\mathcal{O}_{\mathscr{M} \setminus H}$-coherent. Let $i : \mathcal{M} \to Db_{\mathscr{M}}$
be a $\mathcal{D}_{\mathscr{M}}$-linear morphism such that $i $ belongs to $L^2(c(\mathcal{M}))$, the $L^2$ lattice (see \cite{B-K}) of the conjugate module of $\mathcal{M}$ (see \cite{Ka1}). Then any distribution in $i(\mathcal{M})$ has the Standard Extension Property (relatively to $H$).
\end{thm}
The main difficulty to use this theorem is the identification of the $L^2$-lattice. When the $\mathcal{D}_{\mathscr{M}}$-module has no $\mathcal{O}_{\mathscr{M}}$-torsion, the $L^2$-lattice is given by a local square integrability condition along $H$ which is often rather easy to verify. But we must know "a priori" that the $\mathcal{D}_{\mathscr{M}}$-module we are considering has no $\mathcal{O}_{\mathscr{M}}$-torsion to use this simple characterization of the $L^2$-lattice. We shall see in section 5 that the main point is to prove that there is no torsion in some of the regular holonomic $\mathcal{D}_{\mathscr{M}}$-modules involved in our computations.
\bigskip
\subsection{Principal Value}
Let us begin by giving the precise Asymptotic Expansion Theorem for fiber-integrals proved in \cite{B1} which we shall need below. We keep the situation and the notations introduced in the beginning of section 1.
\begin{thm}\label{AS.EX.1}
Let $\varphi \in \mathscr{C}^\infty_c(\mathscr{M} )$ be a differential form of type $(n, n)$ on $\mathscr{M} $. Define the function $\Theta(s) := \int_{f=s} \varphi $ for each $s \in \mathbb{C}$. This function, which is $\mathscr{C}^\infty$ on $\mathbb{C} \setminus \{0\}$, admits when $s \to 0$ a asymptotic expansion of the form
\begin{equation}
\Theta(s) \simeq \sum_{m,m'}^{r, j} a_{m,m'}^{r,j}\vert s\vert^{2r}(Log\, \vert s\vert)^js^m\bar s^{m'}
\end{equation}
where $m, m'$ are non negative integers, $r$ describes a finite set $R \subset [0, 1[ \cap \mathbb{Q}$ and where $j$ is an integer in $[0, n+1]$. The finite set $R$ is independent of the choice of $\varphi$. Moreover, this asymptotic expansion is term-wise differentiable at any order.$\hfill \blacksquare$\\
\end{thm}
\parag{Remarks}\begin{enumerate}
\item The continuity of the function $\Theta$ at the point $s = 0$ implies that for $r = 0$ and $j \geq 1$ we have $m+m' \geq 1$. In fact it is proved in Proposition 6 of {\it loc. cit.} that for $j \geq 1$ we have $a_{m, 0}^{0, j} = 0$ (resp. $a_{0,m'}^{0, j} = 0$) which implies that for $r= 0$ and $j \geq 1$ we have $m \geq 1$ and $m' \geq 1$ when $a_{m,m'}^{0,j} \not= 0$. This shows that no new type of term appears in such an expansion when we apply $s\partial_s$ or $\bar s\partial_{\bar s}$. And, of course, neither a constant term nor a term like $s^m(Log\vert s\vert)^j$ or $\bar s^{m'}(Log\vert s\vert)^j$ with $j \geq 1$ may appear in these expansions. So the functions $s\partial_s\Theta, \bar s\partial_{\bar s}\Theta$ and $s\bar s\partial_s\partial_{\bar s}\Theta$ are bounded by $O(\vert s\vert^\gamma)$ for some $\gamma > 0$ when $s$ goes to $0$.
\item It is also proved, in {\it loc. cit.} that the linear map $\varphi \mapsto a_{m,m'}^{r,j}(\varphi)$ is a $(1,1)$-current on $\mathscr{M} $. This current is supported by the singular set of the hypersurface $\{f = 0\}$ when $(r, j) \not= (0, 0)$, that is to say for terms which do not correspond to an usual term of a Taylor expansion at the origin of a $\mathscr{C}^\infty$ function.\\
\end{enumerate}
It will be important in the sequel to use the following corollary of this theorem.
\begin{cor}\label{AS.EX.2}
In the situation of the previous theorem, for any differential $\mathscr{C}^\infty$ form $\psi$ and $\xi$ respectively of type $(n,n+1)$ and $(n+1,n+1)$ with $f$-proper supports in $\mathscr{M}$ the fiber-integrals
$$ \eta(s) := \int_{f=s} \bar f\frac{\psi}{d\bar f} = \bar s \int_{f=s} \frac{\psi}{d\bar f} \quad {\rm and} \quad \zeta(s) := \int_{f=s} f\bar f\frac{\xi}{df\wedge d\bar f} = s\bar s\int_{f=s} \frac{\xi}{df\wedge d\bar f} $$
admit asymptotic expansions of the same type as above when $s$ goes to $0$ (but see the previous remark 1).
\end{cor}
\parag{Remark} For any compact set $K$ in $\{f = 0\}$ there exists an integer $N$ such that we have the inclusion of sheaves in an open neighborhood of $K$ in $\mathscr{M} $
$$ f^N\Omega^{n+1}_\mathscr{M} \subset df \wedge \Omega^n_\mathscr{M} \quad {\rm and} \quad f^N\bar f^N \Omega^{n+1}_\mathscr{M} \wedge \overline{\Omega}^{n+1}_\mathscr{M} \subset \Omega^n_\mathscr{M} \wedge \overline{\Omega}^n_\mathscr{M} \wedge df\wedge \bar df .$$
So the only new information in the previous corollary is that we may keep $m$ and $m'$ non negative in the expansions.
\parag{Proof} Remark first that the previous theorem is in fact local around the hypersurface $\{f = 0\}$ in $\mathscr{M}$ and so the asymptotic expansion is valid for any $\mathscr{C}^\infty$ differential form of type $(n,n)$ with f-proper support in $\mathscr{M} $.\\
Then using the fact that for such a form $\varphi$ we have the same type of asymptotic expansion for
$$s\partial_s\Theta(s) = \int_{f=s} \frac{fd'\varphi}{df}, \quad \bar s\partial_{\bar s}\Theta(s) = \int_{f=s} \frac{\bar fd''\varphi}{d\bar f} \quad {\rm and\ for} \quad s\bar s\partial_s\partial_{\bar s} \Theta(s) = \int_{f=s} \frac{f\bar fd'd''\varphi}{df\wedge d\bar f} $$
we see that it is enough to show that we have the local surjectivity of $d', d'', d'd''$ with f-proper supports for the type $(n+1, n), (n, n+1)$ and $(n+1, n+1)$ respectively.\\
Using a partition of unity, this is consequence of the following lemma and the local parametrization theorem for the hypersurface $\{f = 0 \}$\footnote{In order that we may choose, near each point in $\{ f = 0 \}$, $U$ and $D$ such that $\bar U \times \partial D$ does not meet $\{f= 0\}$ in our next lemma.}.
\begin{lemma}\label{Dolbeault}
Let $U$ be an open polydisc in $\mathbb{C}^n$ and $D$ a disc in $\mathbb{C}$. Let $\psi$ and $\xi$ be $\mathscr{C}^\infty$ forms respectively of type $(n,n+1)$ and $(n+1, n+1)$ with support in $K \times D$ where $K \subset U$ is a compact set. Fix a relatively compact open disc $D'$ in $D$. Then there exists $\varphi_1$ and $\varphi_2$ which are $\mathscr{C}^\infty$ forms of type $(n,n)$ and with support in $K \times D$ such that
$$ d''\varphi_1 = \psi \quad {\rm and} \quad d'd''\varphi_2 = \xi $$
on $U\times D'$.$\hfill \blacksquare$
\end{lemma}
\parag{Proof} This an easy consequence of the fact that $\partial_z$ and $\partial_z\partial_{\bar z}$ are surjective on $\mathscr{C}^\infty(D)$ and that we may solve the corresponding equations with $\mathscr{C}^\infty$ dependence of a parameter on a relatively compact open disc $D' \subset\subset D$ using a fundamental solution of the corresponding operator ( see for instance \cite{B-M} chapter IV Proposition 5.2.4 for details). $\hfill \blacksquare$
\parag{Remark} Note that $d'\bar \varphi_1 = \bar\psi$ gives also the $d'-$case.
\begin{thm}\label{PV.1}
In the situation $f : \mathscr{M} \to \mathbb{C}$ introduced in section 1, let $\alpha$ be a complex number such that its real part $\Re(\alpha)$ is non negative. Then, for any positive integer $N$ and for any $\mathscr{C}^\infty$ differential form $\xi$ of type $(n+1,n+1)$ with compact support in $\mathscr{M} $ the limit when $\varepsilon > 0$ goes to $0$ of
$$ \langle T_{\alpha,N}^\varepsilon, \xi \rangle := \int_{\vert f\vert \geq \varepsilon} \vert f\vert^{2\alpha}f^{-N}\xi$$
exists and defines a distribution (that is to say a $(0,0)$ current) $T_{\alpha,N}$ on $\mathscr{M} $.
\end{thm}
The main argument to prove this result uses the following consequence of the previous corollary of the Asymptotic Expansion Theorem (see also \cite{H-L} for a proof of the proposition below using a direct computation in a desingularization of the hypersurface $\{f = 0 \}$).
\begin{prop}\label{P.V.2}
Let $\psi$ be a $\mathscr{C}^\infty$ differential form of type $(n, n+1)$ with f-proper support in $\mathscr{M} $. Then we have for any positive integer $N$
$$ \lim_{\varepsilon \to 0} \ \int_{\vert f\vert = \varepsilon} f^{-N}\psi= 0 .$$
\end{prop}
\parag{Proof} Note first that the result is local and is obvious near a point where $f$ does not vanish. Around a point where $f$ vanishes, Milnor's Fibration Theorem allows to use Fubini's Theorem to compute this integral for $\varepsilon$ small enough as follows:
$$ \int_{\vert f\vert = \varepsilon} f^{-N}\psi = \int_{0}^{2\pi} d\theta \int_{f = s} f^{-N}\frac{\psi}{d\theta} $$
where $\psi /d\theta$ on $\{f = s = \varepsilon e^{i\theta} \}$ is equal to \ $\bar f\psi/id\bar f$ \ because we have $d\theta = id\bar f/\bar f$ and taking in account the type of $\psi$ and the fact that $\{f = s\}$ is a complex $n$-dimensional sub-manifold in $\mathscr{M}$ for $s \not= 0$. This gives with $s := \varepsilon e^{i\theta}$
$$ \int_{\vert f\vert = \varepsilon} f^{-N}\psi = \int_{0}^{2\pi} d\theta \int_{f = s} f^{-N}\frac{\bar f\psi}{id\bar f}. $$
When the integral $\int_{f = s} \bar f\psi/d\bar f$ is $\mathcal{O}(\varepsilon^{N+1})$ it is clear that the limit when $\varepsilon$ goes to $0$ exists and vanishes. So it is enough to show the result when we replace the fiber-integral
$$-i\ s^{-N}\bar s\partial_{\bar s}\Theta = \int_{f=s} f^{-N} \bar f\frac{\psi}{id\bar f}$$
by its asymptotic expansion at a high enough order, and then, by linearity, to prove the assertion when we replace the fiber-integral by $ \vert s\vert^{2r}(Log \vert s\vert)^js^{m-N}\bar s^{m'} $. In this case the exponent of $e^{i\theta}$ is given, for $s = \varepsilon e^{i\theta}$, by :
$$ -N + m - m'$$
and the corresponding exponent of $\varepsilon$ is given by
$$ 2r +m +m' - N.$$
In order to find a non zero integral between $0$ and $2\pi$ we need that $m = m' + N$ and then the exponent of $\varepsilon$ is then equal to $2r + 2m'$. So the only terms where the limit is not clearly equal to $0$ appear when $r = m' = 0$. But in this case the limit is again zero thanks to Remark 1 following Theorem \ref{AS.EX.1}.$\hfill \blacksquare$
\parag{Proof of Theorem \ref{PV.1}} The assertion is local near each point of $\{f = 0\}$ so, using Lemma \ref{Dolbeault} (in fact the remark following it) we may assume that $\xi = d'\psi$ where $\psi$ is a $\mathscr{C}^\infty$ differential form of type $(n, n+1)$ with f-proper support. Then Stokes' Formula gives
$$(\alpha - N) \int_{\vert f\vert \geq \varepsilon} \vert f\vert^{2\alpha}f^{-N}\frac{df\wedge\psi}{f} + \int_{\vert f\vert \geq \varepsilon} \vert f\vert^{2\alpha}f^{-N}\xi = \int_{\vert f\vert = \varepsilon} \vert f\vert^{2\alpha}f^{-N}\psi.$$
We know that the limit of the right hand-side is $0$ when $\varepsilon$ goes to $0$ thanks to Proposition \ref{P.V.2}.
Now write the first integral in the left hand-side as
$$ \int_{\vert s\vert \geq \varepsilon} \vert s\vert ^{2\alpha}s^{-N} \frac{ds\wedge d\bar s}{s\bar s} \int_{f= s} \frac{\bar f\psi}{d\bar f} $$
and using polar coordinates $s = \rho e^{i\theta}$ so $ds \wedge d\bar s = -2i \rho d\rho d\theta$ and the asymptotic expansion of the fiber-integral of $\bar f \psi / d\bar f $ we may replace this fiber-integral by its asymptotic expansion at the origin at a sufficient large order. Then, using linearity, it is enough to consider the integrability at $0$ of each non zero term which only depends on the real part of $\alpha$. So we have only to consider the terms for which
$$ -N + m-m' = 0 .$$
The corresponding real part of the power of $\rho$ is given by
$$-N - 1 + 2r + m + m' + 2\Re(\alpha) = -1 + 2r + 2m' + 2\Re(\alpha).$$
This is at least equal to $-1$ for $\Re(\alpha ) \geq 0$ because $m' + r \geq 0$. Then either we have $r + m' + \Re(\alpha) > 0$ and this implies the integrability at $0$, or $r = m' = \Re(\alpha) = 0$. And in this case, thanks to Remark 1 following Theorem \ref{AS.EX.1}, the term to integrate is $O(\vert s\vert^{\gamma-1})$ for some $\gamma > 0$, giving again the local integrability at $0$. So the limit of $\langle T_{\alpha,N}^\varepsilon , \xi \rangle$ exists for each test form $\xi$ on $\mathscr{M} $ when $\varepsilon$ goes to $0$.\\
The fact that the so obtained linear form on test differential forms is a distribution is an easy exercise left to the reader.$\hfill \blacksquare$\\
Note that in the previous proof, we may conclude with the weaker hypothesis asking that $\Re(\alpha) + r > 0$ for any $r$ in $R \cup \{1\} \setminus \{0\}$. Remark that with this hypothesis $\Re(\alpha)$ is not a pole for the meromorphic extension of $F_{ N,\xi}(\lambda), \forall N \in \mathbb{N}$, defined below.
\parag{Remark $R1$}
If we replace $\vert f\vert^{2\alpha}$ by $\vert f\vert^{2\alpha}(Log\vert f\vert^2)^\beta$ where $\beta$ is any positive real number, the same result holds true with the same proof.
\section{Bernstein identity and meromorphic extension}
We first recall the fundamental theorem of Bernstein \cite{Be}, generalized to the local analytic case by Bjork \cite{Bj1}.
\begin{thm}\label{Bernstein 0}
Let $f : (\mathbb{C}^{n+1}, 0) \to (\mathbb{C}, 0)$ be a non zero holomorphic germ. Then there exists a monic polynomial \ $b \in \mathbb{C}[\lambda]$ and, near $0$ in $\mathbb{C}^{n+1}$, a holomorphic partial differential operator $P$ with polynomial coefficients in $\lambda$ such that the identity
\begin{equation}
P(z, \partial_z, \lambda)f^{\lambda+1} = b(\lambda)f^\lambda
\end{equation}
holds locally outside the hypersurface $\{f = 0 \}$ on some open neighborhood of the origin.
\end{thm}
The minimal monic polynomial $b \in \mathbb{C}[\lambda]$ such that and identity of this kind holds near $0$ is called the {\bf Bernstein polynomial of $f$ at the origin}.\\
Recall also the fundamental result of Kashiwara \cite{K}.
\begin{thm}\label{Bernstein 1}
For any non zero holomorphic germ $f$ at the origin, the roots of the Bernstein polynomial of $f$ at the origin are rational and negative.$\hfill \blacksquare$
\end{thm}
\begin{cor}\label{Bernstein 2}
Let $f : (\mathbb{C}^{n+1}, 0) \to (\mathbb{C}, 0)$ be a non zero holomorphic germ and let $\mathscr{M} $ be an open neighborhood of the origin on which the Bernstein identity of $f$ is valid\footnote{This means precisely that the Bernstein identity for $f$ is valid in the universal cover of \\ $\mathscr{M} \setminus \{f = 0 \}$, on which $f^\lambda $ is defined as $\exp(\lambda Log\, f)$ for any given determination of $Log\, f$.}. For any $N \in \mathbb{N}$ and any test differential form $\xi \in \mathscr{C}^{\infty}_c(\mathscr{M} )^{(n+1,n+1)}$ the function defined for $\lambda \in \mathbb{C}$ such that $\Re(\lambda) \gg N$ by
$$ F_{N, \xi}(\lambda) := \int_\mathscr{M} \vert f\vert^{2\lambda}f^{-N}\xi$$
is holomorphic and admits a meromorphic extension to the complex $\lambda$-plane with poles of order at most $n+1$ at points in $\cup_r \ \{r - \mathbb{N}\}$ where $r$ is a root of the Bernstein polynomial of $f$ at the origin.\\
Moreover, for any $\alpha \in \mathbb{C}$, the linear forms on $\mathscr{C}^{\infty}_c(\mathscr{M} )^{(n+1,n+1)}$ given by the coefficient $P_k(\lambda = \alpha, F_{N,\xi}(\lambda))$ of $(\lambda - \alpha)^{-k}, k \in \mathbb{Z}, k \leq n+1$, in the Laurent expansion at $\lambda = \alpha$ of the meromorphic extension of $F_{N, \xi}$ is a distribution on $\mathscr{M} $.\\
For $k \in [1, n+1]$ this distribution has support in $\{f = 0 \}$ and, for $k \in [2, n+1]$ or for $k \in [1, n+1]$ and $\alpha \not\in \mathbb{Z}$, the support of this distribution is contained in the singular set of $\{f = 0 \}$.
\end{cor}
\parag{proof} The equation $(2)$ implies, as $b$ has rational coefficients, that we may find for any positive integer $M$ a anti-holomorphic differential operator $P_M$ depending polynomially of $\lambda$ such that we have
\begin{equation}
P_M(\bar f^{\lambda+M}) = b(\lambda)\dots b(\lambda+M-1)\bar f ^{\lambda}.
\end{equation}
Note that the roots of $B_M(\lambda) := b(\lambda)\dots b(\lambda+M-1)$ are negative rational numbers. So for $M \gg N$ we obtain the equality of continuous functions for $\Re(\lambda) \gg N$
$$ \vert f\vert^{2\lambda}f^{-N} = \frac{1}{B_M(\lambda)}P_M(\vert f\vert^{2\lambda}f^{-N}\bar f^M) $$
If $P_M^*$ is the adjoint of $P_M$ this implies for $\Re(\lambda) \gg N$ and $M$ large enough, that for any test differential form $\xi \in \mathscr{C}^{\infty}_c(\mathscr{M} )^{(n+1,n+1)}$ we obtain
\begin{equation}
\int_\mathscr{M} \vert f\vert^{2\lambda}f^{-N}\xi = \frac{1}{B_M(\lambda)} \int_\mathscr{M} \vert f\vert^{2\lambda}f^{-N}\bar f^MP_M^*(\xi) .
\end{equation}
This gives the meromorphic extension to the complex $\lambda$-plane of the holomorphic distribution $\xi \mapsto \int_M \vert f\vert^{2\lambda}f^{-N}\xi $ defined for $\Re(\lambda) \gg N$, because $P_M^*(\xi)$ is $\mathscr{C}^\infty_c$ in $\mathscr{M} $ and depends polynomially on $\lambda$ and because the right hand-side of the above formula is holomorphic on the open set $\Re(\lambda) > -m$ for any given positive integer $m$ as soon as $M$ is large enough compare to $N+m$.\\
Moreover this meromorphic extension has no pole at points which are not inside the union of the sets $r- \mathbb{N}$ where $r$ is a root of $b$.\\
It is easy to see that near points where $f$ does not vanish, this meromorphic extension has no pole and that near points where $f= 0$ but where $df$ does not vanish the poles of this meromorphic extension are at most simple poles at negative integers. This complete the proof. $\hfill \blacksquare$\\
\parag{Remark $R2$} Let $q$ be a positive integer. The $q$-th derivative in $\lambda$ of the holomorphic function $F_{N, \xi}(\lambda) $ is given, for $\Re(\lambda) $ large enough by the absolutely converging integral
$$ \int_\mathscr{M} \vert f\vert^{2\lambda}(Log \vert f\vert^2)^qf^{-N}\xi$$
and the meromorphic extension of these functions allows to define, for each integer $q$, a meromorphic distributions on $\mathscr{M} $ which has no pole for $\Re(\lambda) \geq 0$.\\
Now analog arguments as above give easily a generalization of the previous theorem to these cases.
\section{Equality of the Principal Value with the value of the Meromorphic Extension }
\subsection{The equality theorem}
We keep the notations of the introduction.
\begin{defn} In the situation above we define the distribution $S_{\alpha,N}$ on $\mathscr{M} $ by the formula
$$ \langle S_{\alpha,N} , \xi \rangle := P_0(\lambda = \alpha, \int_\mathscr{M} \vert f\vert^{2\lambda}f^{-N}\xi ).$$
\end{defn}
The aim of this paragraph is to prove the following result:
\begin{thm}\label{coincide 1}
Assume that $\Re(\alpha) \geq 0$. Then for any positive integer $N$ we have for any test differential form $\xi$:
$$ \langle S_{\alpha,N}, \xi \rangle = \langle T_{\alpha,N}, \xi \rangle = \lim_{\varepsilon \to 0} \int_{\vert f\vert \geq \varepsilon} \vert f\vert^{2\alpha}f^{-N}\xi .$$
\end{thm}
\parag{Proof} We want to show that the analog of the equality $(4)$ for $\lambda = \alpha$ holds if we perform the integration only on the subset $\{\vert f\vert \geq \varepsilon\}$ with an error which goes to zero when
$\varepsilon $ goes to $0$. This would be enough to complete the proof. But, of course, the error comes from the boundary terms which are integrals on $\{\vert f \vert = \varepsilon \}$ appearing in the various Stokes Formulas necessary to pass from $P_M$ to its adjoint $P_M^*$. It is easy to see that such "error" terms have the following shape: a polynomial in $\lambda$ with coefficient like
$$ \int_{\vert f\vert = \varepsilon} \vert f\vert^{2\alpha} f^{-N}\bar f^{M'}\psi $$
where $\psi$ is in $\mathscr{C}^{\infty}_c(\mathscr{M} )^{(n+1,n)}$ and $M'$ is an integer in $[0,M]$. Now using the same arguments than in the proof of Theorem \ref{PV.1} we see that the only non zero term in such an integral comes from the coefficient of $s^m\bar s^{m'}$ in the asymptotic expansion at $s = 0$ of the function $s \mapsto \int_{f=s} \bar f\psi/d\bar f$ such that
$$ m - m' - N - M' = 0 .$$
And this non zero term comes with a power of $\varepsilon$ which is at least equal to
$$2\Re(\alpha) + m + m' - N + M' + 2r = 2\Re(\alpha) + 2m' +2M' + 2r \geq 0 $$
and may be some $(Log\, \varepsilon)^q$ factor. So such term goes to $0$ when $\varepsilon$ goes to $0$ when $m'+r+M' > 0$ for $\Re(\alpha) \geq 0$ but also in the case where $\Re(\alpha) + m' + M' + r = 0$, thanks again to Remark 1 following Theorem \ref{AS.EX.1}. This concludes the proof.$\hfill \blacksquare$\\
Again in the previous proof, we may conclude with the weaker hypothesis asking that $\Re(\alpha) + r > 0$ for any $r$ in $R \cup \{1\} \setminus \{0\}$.
\parag{Remark $R3$} Using Remarks $R1$ and $R2$ we obtain again with the same proof, that the previous theorem is still valid if we replace $\vert f\vert^{2\alpha}$ by $\vert f\vert^{2\alpha}(Log \vert f\vert^2)^q$ for any positive integer $q$.
\subsection{Non torsion of the corresponding $\mathcal{D}$-modules}
We shall deduce from Theorem \ref{coincide 1} an important corollary. To formulate this result we need the following definition, where we keep the situation described in the introduction.
\begin{defn}\label{formal action}Let $\alpha$ be a complex number with a non negative real part. Let $V$ be an holomorphic vector field on $\mathscr{M} $. We define the {\bf formal action} of $V$ on $\vert f\vert^{2\alpha} f^{-N}$ by the formula
$$ \langle V, \vert f\vert^{2\alpha} f^{-N} \rangle := (\alpha - N)V(f)\vert f\vert^{2\alpha}f^{-N-1} .$$
Then this defines a "formal action" of $\mathcal{D}_{\mathscr{M}}$ on the $\mathcal{O}_{\mathscr{M}}$-module $\mathcal{O}_{\mathscr{M}}\vert f\vert^{2\alpha}[ f^{-1}]$.
\end{defn}
Remark that, thanks to Theorem \ref{coincide 1}, each element in $\mathcal{O}_{\mathscr{M}}\vert f\vert^{2\alpha}[ f^{-1}]$ defines an unique distribution on $\mathscr{M}$ having the standard extension property along the hypersurface $\{ f = 0 \}$. This gives a natural $\mathcal{O}_{\mathscr{M}}$-linear embedding of this $\mathcal{O}_{\mathscr{M}}$-module in $Db_{\mathscr{M}}$.
\begin{cor}\label{no torsion}
The action of any $P \in \mathcal{D}_{\mathscr{M}}$ on any element in the sub-$\mathcal{O}_{\mathscr{M}}$-module $\mathcal{O}_{\mathscr{M}}\vert f\vert^{2\alpha}[ f^{-1}] \subset Db_{\mathscr{M}}$ coincides with the formal action defined above.\\
In particular, the sub-$\mathcal{D}_{\mathscr{M}}$-module generated by $Z_\alpha$ in $Db_{\mathscr{M}}$ has no torsion, where $Z_\alpha$ is the distribution on $\mathscr{M}$ associated to the locally bounded function $\vert f\vert^{2\alpha}$ on $\mathscr{M} $.
\end{cor}
\parag{Proof} Let $V$ be a holomorphic vector field on $\mathscr{M} $. Then let $T_{\alpha,N}$ be the distribution defined by
$$ \langle T_{\alpha,N} , \xi \rangle := P_0(\lambda = \alpha, \int_\mathscr{M} \vert f\vert^{2\alpha} f^{-N} \xi) $$
and let $V^*$ be the adjoint of $V$; we obtain:
\begin{align*}
& \langle V(T_{\alpha,N}) , \xi \rangle = \langle T_{\alpha,N} , V^*(\xi) \rangle = \lim_{\varepsilon \to 0} \int_{\vert f\vert \geq \varepsilon} \vert f\vert^{2\alpha} f^{-N} V^*(\xi) \\
& \quad = \lim_{\varepsilon \to 0} \int_{\vert f\vert \geq \varepsilon} V(\vert f\vert^{2\alpha} f^{-N})\xi
\end{align*}
because the boundary term on $\{ \vert f\vert = \varepsilon \}$ has limit $0$ when $\varepsilon$ goes to $0$ using the fact that
$$V(\vert f\vert^{2\alpha} f^{-N}) = (\alpha - N) V(f) \vert f\vert^{2\alpha} f^{-N-1}, $$
the same argument than in the proof of Theorem \ref{coincide 1} and the remark that $V(f)\xi$ is in $\mathscr{C}_c^\infty(\mathscr{M} )^{(n+1,n+1)}$.\\
This gives, using again Theorem \ref{coincide 1}, that $V(T_{\alpha,N}) = V(f)T_{\alpha,N+1}$ in $Db_{\mathscr{M}}$.\\
As holomorphic vector fields generate the $\mathcal{O}_\mathscr{M} $-algebra $\mathcal{D}_{\mathscr{M}}$, this is enough to complete the proof.$\hfill \blacksquare$\\
\parag{Remark $R4$} The generalization of the previous corollary to the cases where we replace $\vert f\vert^{2\alpha}$ by $\vert f\vert^{2\alpha}(Log \vert f\vert^2)^q$ for any positive integer $q$ is again an easy exercise.
\subsection{The case $\alpha < 0$}
We shall explain now how to define, when $\alpha$ is a negative real number\footnote{We leave the case where $\alpha$ is a complex number with a negative real part as an exercise.} a "Finite Part " of the integral
$$ \int_{\mathscr{M}} \vert f\vert^{2\alpha}f^{-N}\xi$$
using the Asymptotic Expansion Theorem.
For a differential form $\xi \in \mathscr{C}^\infty_c(\mathscr{M})^{(n+1, n+1)}$, we shall write as follows the asymptotic expansion when $s$ goes to $0$ of the fiber-integral
$$s \mapsto \int_{f=s} f\bar f \xi/df\wedge d\bar f \simeq \sum_{r,j,m,m'} \tilde{T}_{m,m'}^{r,j}(\xi) \vert s\vert^{2r} s^m \bar s^{m'} (Log\vert s\vert)^j $$
where $\xi \mapsto \tilde{T}(\xi) $ is a distribution with support in $\{f = 0 \}$ for each $r \in R$ a finite subset in $ [0, 1[ \cap \mathbb{Q}$, for $ j \in [0, n+1]$ and for $ m, m' \in \mathbb{N}$ (see the Corollary \ref{AS.EX.2} and also also Remark 1 following Theorem \ref{AS.EX.1}).
\begin{thm}\label{PV with pole}
For $\alpha$ a negative real number such that $-\alpha \not\in R + \mathbb{N}$ the following limit exists and defines a distribution on $\mathscr{M}$
\begin{align*}
& \lim_{\varepsilon \to 0}\Big( \int_{\{\vert f\vert \geq \varepsilon \}} \vert f\vert^{2\alpha} f^{-N}\xi + 2i\pi \sum_{m = m'+N}^{ \alpha+r +m' < 0} \tilde{T}_{m,m'}^{r, j}(\xi) \frac{\varepsilon^{2(\alpha + r + m')}}{2(\alpha+r+m')}(-Log\varepsilon)^j\Big)
\end{align*}
which extends to $\mathscr{M}$ the function $\vert f\vert^{2\alpha}f^{-N}$ on $\{ f \not= 0\}$.\\
If $-\alpha$ is in $R + \mathbb{N}$, write $-\alpha = r + m', r \in R$ and $m' \in \mathbb{N}$ and add to the sum inside the limit in the left hand-side above the sum
$$ 2i\pi\sum_{j = 0}^{n+1} \tilde{T}_{m'+N,m'}^{r, j}(\xi)(-Log\varepsilon)^{j+1}/(j+1) .$$
Then the result is analogous.
\end{thm}
\parag{Proof} Using the asymptotic expansion for the fiber-integral $\int_{f=s} f\bar f \xi/df\wedge d\bar f $ recalled above, the proof that the limit in the left hand-side exists is analogous to the proof of Theorem \ref{PV.1}.$\hfill \blacksquare$ \\
Of course, the constant term of the Laurent development at $\lambda = \alpha$ of the meromorphic extension of the function $ \lambda \mapsto \int_{\mathscr{M}} \vert f\vert^{2\lambda}f^{-N}\xi$ gives also such a distribution.
But in the case $\alpha < 0$ the relation between the distribution defined by
$$\xi \mapsto P_0\big(\lambda = \alpha, \int_\mathscr{M} \vert f \vert^{2\lambda}f^{-N}\xi\big)$$
and the distribution defined in the previous theorem is not so clear in general, even if the difference between these two distributions has clearly its support in $\{ f = 0 \}$. \\
Of course, when $\alpha$ satisfies the condition $\alpha + r > 0$ for each $r$ in $R \cup \{1\} \setminus \{0\}$, not only there is no term in the sum where $\alpha +r +m' \leq 0$ (using Remark 1 following Theorem \ref{AS.EX.1} for the case $0$) and the equality of the two distributions follows from the remark following Theorem \ref{coincide 1}.
\subsection{An easy generalization and an example}
Consider now $p \geq 2$ domains $\mathscr{M} _i$ in $\mathbb{C}^{n_i+1}$ for $i \in [1,p]$ and $p$ holomorphic functions $f_i : \mathscr{M} _i \to \mathbb{C}$. Then on $\mathscr{M} := \prod_{i=1}^p \mathscr{M} _i$ it is easy to prove, for $\alpha := (\alpha_1, \dots, \alpha_p)$ satisfying $\Re(\alpha_i) \geq 0, \forall i \in [1,p]$, using Fubini' s Theorem, the existence of the distribution
$$ \langle T_{\alpha,N}, \xi \rangle := \lim_{\varepsilon \to 0} \langle T_\alpha^\varepsilon , \xi \rangle $$
where $\varepsilon := (\varepsilon_1, \dots, \varepsilon_p)$ is in $(\mathbb{R}^{+*})^p$, $N := (N_1, \dots, N_p) $ is in $ \mathbb{N}^p$, $\xi $ is in $\mathscr{C}^{\infty}_c(\mathscr{M} )^{(n+1, n+1)}$, $n+1 := \sum_{i=1}^p n_i+1$ and where
$$ \langle T_{\alpha,N}^\varepsilon , \xi \rangle := \int_{\prod_{i=1}^p \{\vert f_i\vert \geq \varepsilon_i \}} \quad \big( \prod_{i=1}^p \vert f_i\vert^{2\alpha_i} f_i^{-N_i} \big)\xi .$$
It is also easy to make the meromorphic extension to $\lambda := (\lambda_1, \dots, \lambda_p) \in \mathbb{C}^p$ of the holomorphic distributions on $\mathscr{M} $ defined for $ \prod_{i=1}^p \{\Re(\lambda_i) \gg N_i \}$ and then to prove
the following generalization of Theorem \ref{coincide 1} to this "product" case:
\begin{thm}\label{generalized}
In the product situation described above the meromorphic extension of the distribution
$$ \xi \mapsto \int_\mathscr{M} \big(\prod_{i=1}^p \vert f_i\vert^{2\lambda_i} f_i^{-N_i}\big) \xi $$
is holomorphic near the point $\alpha \in \mathbb{C}^p$ satisfying $\Re(\alpha_i) \geq 0, \forall i \in [1,p]$, and we have the equality
$$ P_0\Big(\lambda = \alpha, \int_\mathscr{M} \big( \prod_{i=1}^p \vert f_i\vert^{2\lambda_i} f_i^{-N_i}\big) \xi\Big) = \langle T_{\alpha,N}, \xi \rangle $$
where the left hand-side denotes the value at $\lambda = \alpha$ of the meromorphic extension.
\end{thm}
\parag{Proof} The only point to precise in order to apply the $p = 1$ case and Fubini's Theorem successively to prove this generalization it the following remark:
\begin{itemize}
\item Let $\mathscr{M} _1\times \mathscr{M} _2 $ be the product of two complex manifolds $\mathscr{M} _1\times \mathscr{M} _2$ and let $\xi$ be a $\mathscr{C}_c^\infty $ test differential form on $\mathscr{M} _1\times \mathscr{M} _2$ and $T_2$ a distribution on $\mathscr{M} _2$.
Then the test differential form defined on $\mathscr{M} _1$ by $\langle T_2, \xi \rangle $ is a $\mathscr{C}_c^\infty $ test differential form on $\mathscr{M} _1$. So for any distribution $T_1$ on $\mathscr{M} _1$ the distribution $T_1\boxtimes T_2$ is well defined on $\mathscr{M} _1\times \mathscr{M} _2$ by the rule
$$\langle T_1\boxtimes T_2, \xi \rangle := \langle T_1, \langle T_2, \xi \rangle \rangle.$$
\end{itemize}
\parag{Remark} Note that the definition of $T_{\alpha,N}$ implies that we have
$$ \langle T_{\alpha,N}, \xi \rangle = \lim_{\varepsilon \to 0} \int_{\vert \prod_{i=1}^p f_i \vert \geq \varepsilon } \big(\prod_{i=1}^p \vert f_i\vert^{2\alpha_i} f_i^{-N_i}\big) \xi . $$
Then, using the argument in Corollary \ref{no torsion} on each $\mathscr{M} _i, i \in [1, p]$ we obtain that for any $P \in \mathcal{D}_{\mathscr{M}}$ the distribution $PT_{\alpha,N}$ has the standard extension property for the hypersurface $\{\prod_{i=1}^p f_i = 0 \}$. So the sub-$\mathcal{D}_{\mathscr{M}}$-module of $Db_{\mathscr{M}}$ generated by $T_{\alpha,N}$ has no torsion.\\
Note also that this implies that the sub-$\mathcal{D}_{\mathscr{M}}$-module generated by $T_{\alpha,0}$ is contained in the $\mathcal{O}_\mathscr{M}$-module generated by the $T_{\alpha,N}$ when $N$ is in
$\mathbb{N}^p$.
\parag{Remark $R5$} Again it is easy to generalize the previous theorem to the cases where we replace $\vert f_i\vert^{2\alpha_i}$ by $\vert f_i\vert^{2\alpha_i}(Log \vert f_i\vert^2)^{q_i}$ for any non negative integers $q_i, i \in [1, p]$.
\parag{An example}
\parag{Notations} Let $\mathscr{M} := \mathbb{C}^k$ with coordinates $z_1, \dots, z_k$ and let $\pi: \mathscr{M} \to \mathscr{N} \simeq \mathbb{C}^k$ be the quotient by the action of the permutation group $\mathfrak{S}_k$ on $\mathscr{M} $. Let $\sigma_1, \dots, \sigma_k$ be the elementary symmetric polynomials in $z_1, \dots, z_k$ which give a coordinate system on $\mathscr{N} $. Let also $\Delta := \prod_{1\leq i < j \leq k} (z_i - z_j)^2 $ be the discriminant which is a polynomial in $\sigma_1, \dots, \sigma_k$.\\
\begin{lemma}\label{simplet}
Consider on $\mathscr{M} := \mathbb{C}^k$ with coordinates $z_1, \dots, z_k$ the holomorphic functions $f := z_1$ and $\Delta := \prod_{1\leq i < j \leq k} (z_i - z_j)^2 $. Then defining new coordinates $x_1 := z_1$ and $x_h := z_h - z_1$ for $h \in [2, k]$ we obtain a decomposition $\mathscr{M} = \mathscr{M}_1\times \mathscr{M}_2$ where $\mathscr{M}_1 := \mathbb{C}$ with coordinate $x_1$ and $\mathscr{M}_2 := \mathbb{C}^{k-1}$ with coordinates $x_h, h \in [2, k]$. Moreover $f$ is in $\mathbb{C}[x_1]$ and $\Delta$ is in $\mathbb{C}[x_2, \dots, x_k]$.
\end{lemma}
\parag{Proof} The reader will see easily that $\Delta = \prod_{h=2}^k x_h^2 \prod_{2\leq i < j \leq k} (x_i - x_j)^2$.$\hfill \blacksquare$\\
This easy lemma allows to apply Theorem \ref{generalized} to the situation described in the previous lemma.
\begin{cor}\label{no torsion M}
For any $\alpha \in \mathbb{C}$ such that $\Re(\alpha) \geq 0$ and any non negative integers $q$ and $N_1, N_2$ the distribution on $\mathscr{M}$ associated to the locally integrable function
$$X_{\alpha,N_1,N_2,q} := \big(\sum_{j=1}^k \vert z_j\vert^{2\alpha}(Log\vert z_j\vert^2)^q z_j^{-N_1} \big) \Delta^{-N_2}$$
generates a sub-$\mathcal{D}_{\mathscr{M}}$-module of the $\mathcal{D}_{\mathscr{M}}$-module $Db_{\mathscr{M}}$ which has no torsion. Moreover, any distribution in this sub-$\mathcal{D}_{\mathscr{M}}$-module has the standard extension property along the divisor $\{\pi^{-1}(\sigma_k\Delta(\sigma) )= 0 \}$ in $\mathscr{M}$ and is $\mathscr{C}^\infty$(in fact real analytic) outside this divisor.\\
The same properties are true along the divisor $\{\sigma_k\Delta(\sigma) = 0\}$, for the sub-$\mathcal{D}_{\mathscr{N}}$-module of $Db_{\mathscr{N}}$ generated by the distribution $\pi_*(X_{\alpha,N_1,N_2,q})$.\\
\end{cor}
\parag{Proof} For our assertion on $\mathscr{M}$ it is enough to prove that for any holomorphic vector field $V$ on $\mathscr{M}$ the distribution $V(\vert z_1\vert^{2\alpha}z_1^{-N_1}\Delta^{-N_2})$ has the standard extension property and this is clear from the remark following Theorem \ref{generalized} and Remark $R5$. \\
The assertion on the corresponding $\mathcal{D}_{\mathscr{N}}$-module is consequence of the fact that if a distribution $T$ on $\mathscr{M}$ has the standard extension property along the divisor $\{\pi^{-1}(\sigma_k\Delta(\sigma)) = 0 \}$ then the distribution $\pi_*(T)$ has the standard extension property along the divisor $\{ (\sigma_k\Delta(\sigma) = 0 \}$ because the quotient map $\pi$ is a finite \'etale map outside $\{\Delta(\sigma) = 0\}$. $\hfill \blacksquare$
\parag{Remark} If $S$ and $T$ are two distributions on a complex manifold $\mathscr{N}$ such that $\mathcal{D}_{\mathscr{N}}S \subset Db_{\mathscr{N}}$ and $\mathcal{D}_{\mathscr{N}}T \subset Db_{\mathscr{N}}$ have no $\mathcal{O}_\mathscr{N}$-torsion, it is not clear that there is no $\mathcal{O}_\mathscr{N}$-torsion in $D_\mathscr{N}(S+T) \subset Db_{\mathscr{N}}$ . But if we know that $\mathcal{D}_{\mathscr{N}}S$ and $\mathcal{D}_{\mathscr{N}}T$ contain only distributions having the standard extension property for a given hypersurface $H$, then this also the case for any distribution
in $\mathcal{D}_{\mathscr{N}}S + \mathcal{D}_{\mathscr{N}}T \subset Db_{\mathscr{N}}$ which contains $\mathcal{D}_\mathscr{N}(S+T)$.
\section{Some conjugate D-modules}
The aim of this section is to give some examples of explicit computations of the conjugate module (in Kashiwara sense, see \cite{Ka1}) of some regular holonomic $\mathcal{D}$-modules. We consider here the case of the $\mathcal{D}$-modules associated to the multivalued functions $z(\sigma)^\lambda$ for $\lambda \in \mathbb{C}$ where $z(\sigma)$ is the root of the universal degree $k$ equation
$$ z^k + \sum_{h=1}^k (-1)^h\sigma_hz^{k-h} $$
where $\sigma := (\sigma_1, \dots, \sigma_k)$ is in $N := \mathbb{C}^k$. The structure of these regular holonomic $\mathcal{D}_{\mathscr{N}}$-modules has been described in \cite{B4} and we describe here for each such $\mathcal{D}$-module and for each simple factor which appears in its decomposition a distribution $T$ on $\mathscr{N}$ which generates the sub-$\mathcal{D}_{\bar{\mathscr{N}}}$-module of $Db_{\mathscr{N}}$ which is the conjugate of the module we consider.\\
The reader will notice that even if in each case the corresponding distribution $T$ is rather easily constructed from the horizontal (multivalued) basis of the corresponding vector bundle with a simple pole connection associated to the $\mathcal{D}_{\mathscr{N}}$-module under consideration, the proof that this distribution generates the conjugate module uses in a crucial way the non trivial argument of "non torsion" which is proved in the previous section (see Corollary \ref{no torsion M}).\\
We begin by recalling the basic results on the conjugation functor of M. Kashiwara. The following theorem is proved in \cite{Ka1}:
\begin{thm}[Kashiwara conjugation functor]
Let $\mathscr{N}$ be a complex manifold, $\mathcal{D}_{\mathscr{N}}$ the sheaf of holomorphic partial differential operators on $\mathscr{N}$ and $\mathcal{D}_{\bar{\mathscr{N}}}$ the sheaf of anti-holomorphic partial differential operator on $\mathscr{N}$. Note $Db_{\mathscr{N}}$ the sheaf of distributions on $\mathscr{N}$. It is a left-$\mathcal{D}_{\mathscr{N}}$-module but also a left-$\mathcal{D}_{\bar{\mathscr{N}}}$-module and these two actions commute.\\
For each regular holonomic $\mathcal{D}_{\mathscr{N}}$-module $\mathcal{N}$ the sub-$\mathcal{D}_{\bar{\mathscr{N}}}$-module
$$c_\mathscr{N}(\mathcal{N}) := Hom_{\mathcal{D}_{\mathscr{N}}}(\mathcal{N}, Db_{\mathscr{N}})$$
is regular holonomic (as a $\mathcal{D}_{\bar{\mathscr{N}}}$-module) and the contra-variant functor $c$ is an anti-equivalence of categories which satisfies $c_{\bar{\mathscr{N}}}\circ c_\mathscr{N} = Id$.$\hfill \blacksquare$
\end{thm}
Moreover, M. Kashiwara also obtains the following proposition which will be useful to describe the $\mathcal{D}_{\bar{\mathscr{N}}}$-modules $c(\mathcal{N}_\lambda)$:
\begin{prop}[ see \cite{Ka1} Prop. 5]\label{K2}
If $\mathcal{N}$ is a regular holonomic $\mathcal{D}_{\mathscr{N}}$-module on a complex manifold $\mathscr{N}$ and if $T$ is in $c(\mathcal{N})$, the following conditions are equivalent:
\begin{enumerate}[a)]
\item $T$ is an injective sheaf homomorphism of $\mathcal{N}$ to $Db_{\mathscr{N}}$.
\item $T$ generates $c(\mathcal{N})$ as a $\mathcal{D}_{\bar{\mathscr{N}}}$-module.$\hfill \blacksquare$
\end{enumerate}
\end{prop}
\parag{Notations} We consider the quotient map $ \pi : \mathscr{M} := \mathbb{C}^k \to \mathscr{N}:= \mathbb{C}^k\big/\mathfrak{S}_k \simeq \mathbb{C}^k$ with respective coordinates $z_1, \dots, z_k$ and $\sigma_1,\dots, \sigma_k$, where $\sigma_h$ is the $h$-th elementary symmetric function of $(z_1, \dots, z_k)$.\\
The vector fields on $\mathscr{M}$ associated to partial derivatives in $z_1, \dots, z_k$ are denoted by $\partial_{z_1}, \dots, \partial_{z_k}$ and the vector fields on $\mathscr{N}$ associated to partial derivatives in $\sigma_1, \dots, \sigma_k$ are denoted by $\partial_1, \dots, \partial_k$.\\
We note $\mathcal{I}$ the left ideal in $\mathcal{D}_{\mathscr{N}}$ generated by the following global sections:
\begin{align*}
& A_{p,q} = \partial_p\partial_q - \partial_{p+1}\partial_{q-1} \quad {\rm where} \quad (p,q) \in [1,k-1]\times [2,k] \\
& \mathcal{T}^m := \partial_1\partial_{m-1} + \partial_mE \quad {\rm where} \quad E := \sum_{h=1}^k \sigma_h\partial_h \quad {\rm and} \quad m \in [2, k]
\end{align*}
Recall that a {\bf trace function} on $\mathscr{N}$ is a holomorphic function $F$ such that there exists a holomorphic function $f$ in one variable $z$ such that
$$ F(\sigma) = \sum_{j=1}^k f(z_j) $$
where $\sigma := ( \sigma_1, \dots, \sigma_k)$ are the elementary symmetric functions of $z_1, \dots, z_k$. \\
It is proved in \cite{B1} that the trace functions are annihilated by the left ideal $\mathcal{I}$ in $\mathcal{D}_{\mathscr{N}}$ and this characterizes the trace functions.\\
We note $U_{-1}$ the vector field on $\mathscr{N}$ given by $U_{-1} := k\partial_1 + \sum_{h=1}^{k-1} (k-h)\sigma_h\partial_{h+1}$ which is the image by the tangent map $T_\pi$ to $\pi$ of the vector field $V_{-1} := \sum_{j=1}^k \partial_{z_j}$ on $\mathscr{M}$.\\
We note $U_{0}$ the vector field on $\mathscr{N}$ given by $U_0 := \sum_{h=1}^k h\sigma_h\partial_h $ which is the image by $T_\pi$ of the vector field $V_{0} := \sum_{j=1}^k z_j\partial_{z_j}$on $\mathscr{M}$.\\
We note $U_1$ the vector field on $\mathscr{N}$ given by $U_1 := \sum_{h=1}^k (\sigma_1\sigma_h -(h+1)\sigma_{h+1})\partial_h$ which is the image by $T_\pi$ of the vector field $V_{1} := \sum_{j=1}^k z_j^2\partial_{z_j}$ on $\mathscr{M}$.\\
The left ideal $\mathcal{J}_\lambda$ in $\mathcal{D}_{\mathscr{N}}$ is, by definition, the sum $\mathcal{I} + \mathcal{D}_{\mathscr{N}}(U_0 - \lambda)$ and we define $\mathcal{N}_\lambda := \mathcal{D}\big/\mathcal{J}_\lambda$.\\
The following results which give the structure of the $\mathcal{D}_{\mathscr{N}}$-module $\mathcal{N}_\lambda$ for each $\lambda \in \mathbb{C}$ are proved in \cite{B4}:\\
\begin{enumerate}
\item For each complex number $\lambda$, the $\mathcal{D}_{\mathscr{N}}$-module $\mathcal{N}_\lambda$ is holonomic and regular.
\item The right multiplication by $U_{-1}$ induces a $\mathcal{D}_{\mathscr{N}}$-linear map
$$ \square U_{-1} : \mathcal{N}_\lambda \to \mathcal{N}_{\lambda+1}$$
which is an isomorphism for each $\lambda \not= -1, 0$. Moreover the right multiplication by $U_1$ induces an isomorphism $ \square U_1 : \mathcal{N}_{\lambda +1} \to \mathcal{N}_{\lambda }$ for any $\lambda \not= 0, -1$ and we have $\square U_1\circ \square U_{-1} = \lambda(\lambda+1)$ on $\mathcal{N}_{\lambda}$.
\item For $\lambda \not\in \mathbb{Z}$ the $\mathcal{D}_{\mathscr{N}}$-module $\mathcal{N}_\lambda$ is simple.
\item The kernel $\mathcal{N}_{-1}^\square$ of the $\mathcal{D}_{\mathscr{N}}$-linear map $\varphi_{-1} : \mathcal{N}_{-1} \to \mathcal{O}_N(\star \sigma_k)$ defined by $\varphi_{-1}(1) = \sigma_{k-1}/\sigma_k$ is simple.
\item The sub-module $\mathcal{N}_0^\square$ generated by $U_1$ in $\mathcal{N}_0$ is simple and the quotient $\mathcal{N}_0\big/\mathcal{N}_0^\square$ is isomorphic to $\mathcal{O}_N(\star \sigma_k)$
\item The torsion sub-module $\mathfrak{T}$ in $\mathcal{N}_1$ is isomorphic to $H^1_{[\sigma_k = 0]}(\mathcal{O}_N)$; it is generated by the class of $\partial_kU_{-1}$ in $\mathcal{N}_1$.
\item The sub-module $Im(\square U_{-1})$in $\mathcal{N}_1$ (which is generated by $U_{-1}$) contains $\mathfrak{T}$ and the quotient $Im(\square U_{-1})\big/\mathfrak{T}$ is isomorphic to $\mathcal{O}_\mathscr{N}$ via the map $\varphi_1 :\mathcal{N}_1 \to \mathcal{O}_N$ defined by $\varphi_1(1) = \sigma_1$ (and then $[U_{-1}] \mapsto k$).
\item The quotient $\mathcal{N}_1^\square := \mathcal{N}_1\big/Im(\square U_{-1})$ is simple and isomorphic to the quotient
$$\mathcal{D}_{\mathscr{N}}\big/ \mathcal{I} + \mathcal{D}_{\mathscr{N}}(U_0-1) + \mathcal{D}_{\mathscr{N}}U_{-1}.$$
\item The right multiplication by $U_1$ which sends $\mathcal{N}_1$ in $\mathcal{N}_0^\square$ vanishes\footnote{It is proved in \cite{B4} Formula $(19)$ page 20, that we have $U_{-1}U_1 = (U_0+1)U_0 \quad {\rm modulo} \quad \mathcal{I}$.} on $Im(\square U_{-1})$ and induces an isomorphism of $\mathcal{N}_1^\square$ onto $\mathcal{N}_0^\square$.
\item The right multiplication by $U_{-1}$ which sends $\mathcal{N}_{-1}$ to $\mathcal{N}_0 $ induces an isomorphism of $\mathcal{N}_{-1}^\square$ onto $\mathcal{N}_0^\square$.
\end{enumerate}
\parag{Important remark} The point 2 recalled above shows that to study the $\mathcal{D}_{\mathscr{N}}$-modules $\mathcal{N}_\lambda$ it is enough to consider the following cases:
\begin{itemize}
\item The cases $\Re(\lambda) \in [0, 1[$ and $\lambda \not= 0$. We call it the case $G$.
\item The cases $\lambda = -1, 0, 1$. We call them the case $\lambda = -1, 0, 1$ respectively.
\end{itemize}
Then for any $\lambda \in \mathbb{C}$ we reach one of these previous cases using an isomorphism given either by $U_1^N$ or by $U_{-1}^N$ for a suitable $N \in \mathbb{N}$.\\
Now define the following distributions on $\mathscr{M}$:
\begin{enumerate}
\item For $\Re(\lambda) \in [0, 1[, \lambda \not= 0$ define $X_\lambda := \sum_{j=1}^k \vert z_j\vert^{2\lambda} $.
\item $X_1 := \sum_{j=1}^k \vert z_j - \sigma_1/k\vert^2 = \sum_{j=1}^k \vert z_j\vert^2 -\vert\sigma_1\vert^2/k$.
\item $X_0 := \sum_{j=1}^k (\bar z_j-\bar \sigma_1/k)Log \vert z_j\vert^2 = \sum_{j=1}^k \vert z_j\vert^2z_j^{-1}Log \vert z_j\vert^2 - (\bar \sigma_1/k) Log\vert \sigma_k\vert^2$.
\item $X_{-1} = \sum_{j=1}^k (\bar z_j- \bar\sigma_1/k) z_j^{-1} = \sum_{j=1}^k \vert z_j \vert^2z_j^{-2} - (\bar\sigma_1/k)\sigma_{k-1}/\sigma_k$.
\end{enumerate}
For each case we define the distribution $\mathcal{X}_\lambda := \pi_*(X_\lambda)$ on
$\mathscr{N}$.\\
Define also the following distributions on $\mathscr{N}$:
\begin{enumerate}
\item For the case $G$ :\quad $\mathcal{Y}_\lambda := 0$.
\item For the case $\lambda = 1$: \quad $\mathcal{Y}_1 := \sigma_1/\bar\sigma_k$
\item For the case $\lambda = 0 $: \quad $\mathcal{Y}_0 := 1/\bar \sigma_k$.
\item For the case $\lambda = -1$: \quad $\mathcal{Y}_{-1} := \sigma_{k-1}/\sigma_k$.
\end{enumerate}
Then we have the following results:
\begin{thm}\label{conj.1}[conjugate modules]\\
In all cases $G, -1, 0, 1$ the distribution $\mathcal{X}_\lambda$ defines an element in $c(\mathcal{N}_\lambda)$ via the $\mathcal{D}_{\mathscr{N}}$-linear map sending $1$ to $\mathcal{X}_\lambda$.\\
In case $G$ the distribution $\mathcal{X}_\lambda$ defines a generator of the (simple) $\mathcal{D}_{\bar{\mathscr{N}}}$-module $c(\mathcal{N}_\lambda)$.\\
In cases $\lambda =1$ the distribution $\mathcal{X}_1$ defines an element in $c(\mathcal{N}_1^\square) \subset c(\mathcal{N}_1)$ and gives a generator of this (simple) $\mathcal{D}_{\bar{ \mathscr{N}}}$sub-module.\\
In cases $\lambda = -1, 0$ the simple module $c(\mathcal{N}_\lambda^\square)$ is a quotient of $c(\mathcal{N}_\lambda)$ and the image of $\mathcal{X}_\lambda $ in $c(\mathcal{N}_\lambda^\square)$ gives a generator of this (simple) $\mathcal{D}_{\bar{ \mathscr{N}}}$-module.\\
In all cases $\lambda = -1, 0, 1$ the map sending $1$ to $\mathcal{Y}_\lambda$ is in $c(\mathcal{N}_\lambda)$ and the $\mathcal{D}_{\bar{\mathscr{N}}}$-module $c(\mathcal{N}_\lambda)$ is generated by $\mathcal{X}_\lambda$ and $\mathcal{Y}_\lambda$.
\end{thm}
\parag{Proof} Consider the first point. We have to prove that the left ideal $\mathcal{J}_\lambda$ annihilates the distribution $\mathcal{X}_\lambda$ for each case.\\
Corollary \ref{no torsion M} and Remark R5 give that for $\Re(\lambda) \geq 0$ any distribution in the $\mathcal{D}_\mathscr{M}$-module generated by the distribution
$$ \vert z_j\vert^{2\lambda}(Log \vert z_j\vert^2)^q z_j^{-N}$$
has the standard extension property. So the same property holds for any distribution in the $\mathcal{D}_\mathscr{N}$-module generated by the distribution
$$ \sum_{j=1}^k \vert z_j\vert^{2\lambda}(Log \vert z_j\vert^2)^q z_j^{-N}$$
along the hypersurface $\{\sigma_k\Delta(\sigma) = 0 \}$ thanks to Corollary \ref{no torsion M} .\\
Remark also that the generator of the left ideal $\mathcal{I}$ and the vector field $U_0$ are in the left ideal of $\mathcal{D}_\mathscr{N}$ generated by $\partial_h, h \in [1, k-1]$ and $\sigma_k\partial_k$ and that these vector fields annihilate the distribution $1/\bar \sigma_k$ (see Lemma \ref{facile} below).\\
We have also $\sigma_k\partial_k(Log \vert\sigma_k\vert^2) = 1$ in $Db_{\mathscr{N}}$. \\
Now the verification that in each case the distribution $\mathcal{X}_\lambda$ is annihilated by the generator of the left ideal $\mathcal{J}_\lambda$ is "formal" (see Corollary \ref{no torsion M} ) because $\mathcal{X}_\lambda$ is locally on the open set $\{ \sigma_k\Delta(\sigma) \not= 0\}$ a linear combination with anti-holomorphic coefficients of the holomorphic trace functions $z_j(\sigma)^\lambda, j \in [1, k]$.\\
Moreover it is clear that $z_j(\sigma)^\lambda$ is homogeneous of degree $\lambda$ on $\mathscr{M}$ and the vector field $U_0$ is equal to $ \pi_*(V_0)$ where $V_0$ is the Euler vector field on $\mathscr{M}$. \\
In the case $G$ we know that $c(\mathcal{N}_\lambda)$ is a simple $\mathcal{D}_{\bar{\mathscr{N}}}$-module. So the second point of the theorem is proved as $\mathcal{X}_\lambda$ is clearly not $0$ in $Db_\mathscr{N}$.\\
For $\lambda = 1$, as we know that ideal $\mathcal{J}_1 = \mathcal{I} + \mathcal{D}_{\mathscr{N}}(U_0-1)$ already annihilates $\mathcal{X}_1$ it is enough, thanks to the point 8 recalled above, to check that $U_{-1}(\mathcal{X}_1) = 0$ in $Db_{\mathscr{N}}$ to conclude because we know that $\mathcal{N}_1^\square$ is simple and $\mathcal{X}_1$ is not $0$ in $Db_{\mathscr{N}}$. Again the fact that it is enough to check this on the open set $\{\Delta(\sigma) \not= 0\}$ makes the computation easy :
$$ V_{-1}(\vert z_j\vert^2) = \bar z_j \quad {\rm and} \quad V_{-1}(\vert \sigma_1\vert^2) = k\bar \sigma_1 \quad {\rm so} \quad U_{-1}(\mathcal{X}_1) = 0.$$
\begin{lemma}\label{facile}
Let $\mathcal{M}$ be the $\mathcal{D}_{\mathbb{C}}$-module quotient of $\mathcal{D}_{\mathbb{C}}$ by the left ideal generated by $z\partial_z$. Then $c(\mathcal{M})$ is isomorphic to $\mathcal{O}_{\bar{\mathbb{C}}}(*\bar z)$.
\end{lemma}
\parag{Proof} First recall that the distribution on $\mathbb{C}$ associated to the locally integrable function $1/\bar z$ satisfies $\partial_z(1/\bar z) = i\pi \delta_0$, where $\delta_0$ is the Dirac mass at the origin.
So the $\mathcal{D}_{\overline{\mathbb{C}}}$-sub-module of $Db_{\mathbb{C}}$ generated by $1/\bar z$ is contained in $c(\mathcal{M})$. To prove that it is equal to $c(\mathcal{M})$, thanks to Proposition \ref{K2}, it is enough to prove that the kernel of the $\mathcal{D}_{\mathbb{C}}$-linear map $\mathcal{M} \to Db_\mathbb{C} $ defined by $1 \mapsto 1/\bar z$ is injective. But modulo the left ideal generated by $z\partial_z$ any
$P \in \mathcal{D}_{\mathbb{C}}$ may be written as $a_0 + \sum_{p=1}^m a_p\partial_z^p$ where $a_0, \dots, a_p$ are complex numbers. Then $a_0$ has to vanish and the $\mathbb{C}$-linear independence of the distributions $\partial_z^{p-1}\delta_0, p \geq 1$ allows to conclude.$\hfill \blacksquare$
\begin{cor}\label{utile}
The conjugate of the sub-$\mathcal{D}_{\mathscr{N}}$-module $Im(\square U_{-1}) := \mathcal{D}_{\mathscr{N}}U_{-1} \subset \mathcal{N}_1$ is equal to $\mathcal{D}_{\bar{\mathscr{N}}}\sigma_1/\bar\sigma_k \subset Db_{\mathscr{N}}$ which is isomorphic to the $\mathcal{D}_{\bar{\mathscr{N}}}$-module $\mathcal{O}_{\bar{\mathscr{N}}}(*\bar \sigma_k)$ . It is generated by the $\mathcal{D}_{\mathscr{N}}$-linear map sending the class of $U_{-1}$ in $Im(\square U_{-1}) \subset \mathcal{N}_1$ to the distribution $\mathcal{Y}_1 = \sigma_1/\bar \sigma_k$. So the distribution $\mathcal{X}_1$ and $ \mathcal{Y}_1$ in $Db_{\mathscr{N}}$ generate $c(\mathcal{N}_1)$ as a $\mathcal{D}_{\bar{\mathscr{N}}}$-sub-module of $Db_{\mathscr{N}}$.
\end{cor}
\parag{Proof} It is proved in Proposition 4.1.7 and Lemma 4.1.8 of \cite{B4} that the annihilator of the class of $U_{-1}$ in $\mathcal{N}_1$ is generated by $\partial_h, h \in [1, k-1]$ and $\sigma_k\partial_k$. Note that this is also the annihilator in $\mathcal{D}_{\mathscr{N}}$ of the distribution $1/\bar \sigma_k$ in $Db_{\mathscr{N}}$ thanks to the previous lemma. This shows that the $\mathcal{D}_{\mathscr{N}}$-linear map defined by $[U_{-1}] \mapsto 1/\bar\sigma_k \in Db_{\mathscr{N}}$ is injective, proving the isomorphism of $\mathcal{D}_{\bar{\mathscr{N}}}$-modules $c(Im(\square U_{-1})) \simeq \mathcal{D}_{\bar{\mathscr{N}}}1/\bar\sigma_k \simeq \mathcal{O}_{\bar{\mathscr{N}}}(*\bar \sigma_k) $ thanks to Proposition \ref{K2}.\\
Then the exact sequence of $\mathcal{D}_{\mathscr{N}}$-modules
$$ 0 \to Im(\square U_{-1}) \to \mathcal{N}_1 \to \mathcal{N}_1^\square \to 0 $$
gives the exact sequence of $\mathcal{D}_{\bar{\mathscr{N}}}$-modules
$$ 0 \to c(\mathcal{N}_1^\square) \to c(\mathcal{N}_1) \overset{\alpha}{\to}
c(Im(\square U_{-1}) \to 0 $$
where the map $\alpha$ is the quotient map. Now, let us prove that the distribution
$\mathcal{Y}_1 = \sigma_1/\bar \sigma_k$ is a generator of $c(Im(\square U_{-1}))$ via the $\mathcal{D}_{\mathscr{N}}$-linear map $U_{-1} \mapsto \mathcal{Y}_1$.
As $\sigma_1$ commutes with the action of $\mathcal{D}_{\bar{\mathscr{N}}}$, the annihilator of $\sigma_1/\bar \sigma_k$ in $\mathcal{D}_{\bar{\mathscr{N}}}$ is the same than the annihilator of $1/\bar \sigma_k$ proving our claim. \\
But we know that $\mathcal{X}_1$ is a generator of $c(\mathcal{N}_1^\square)$. So, it is enough to prove that $\mathcal{Y}_1$ is in $c(\mathcal{N}_1)$ to see that with $\mathcal{X}_1$ they generate this $\mathcal{D}_{\bar{\mathscr{N}}}$-module. This is proved in our next lemma and completes the proof of this corollary.$\hfill \blacksquare$\\
\begin{lemma}\label{ajout}
For $\lambda = 1, 0, -1$ the distribution $\mathcal{Y}_\lambda$ defines an element in $c(\mathcal{N}_\lambda)$ via the $\mathcal{D}_{\mathscr{N}}$-linear map $1 \mapsto \mathcal{Y}_\lambda$.\\
\end{lemma}
\parag{Proof} The fact that $\mathcal{Y}_1$ in in $c(\mathcal{N}_1)$ is easy because $\sigma_1$ is a trace function which satisfies $U_0(\sigma_1) = \sigma_1$, because $\partial_h, h \in [1, k-1]$ and
$ \sigma_k\partial_k$ annihilate the distribution $1/\bar\sigma_k$ and it is easy to check that the generators of $\mathcal{I}$ and $U_0$ are in this left ideal of $\mathcal{D}_{\mathscr{N}}$ generated by these vector fields.\\
Now we have also, using the same argument, that $\mathcal{Y}_0$ is in $c(\mathcal{N}_0)$.\\
To see that $\mathcal{Y}_{-1}$ is annihilated by $\mathcal{J}_{-1}$, remark first that $\mathcal{D}_{\mathscr{N}}\mathcal{Y}_{-1} \subset Db_{\mathscr{N}}$ has no torsion (see Corollary \ref{no torsion M} for $\alpha = 0, N = -1, m = 0$).
Then note that
$$\sigma_{k-1}/\sigma_k = \pi_*(\sum_{j=1}^k 1/z_j)$$
is a trace function on the open set $\{\sigma_k\Delta(\sigma) \not= 0 \}$ and that $U_0 + 1$ kills also this holomorphic function on this open set. This is enough to complete the proof of the lemma using again the absence of torsion.$\hfill \blacksquare$\\
\begin{lemma}\label{autre ajout}
The $\mathcal{D}_{\mathscr{N}}$-linear map defined by $[1] \mapsto \mathcal{X}_0 \in Db_{\mathscr{N}}$ is a generator of $c(\mathcal{N}_0^\square)$ and the $\mathcal{D}_{\mathscr{N}}$-linear map defined by $[1] \mapsto \mathcal{X}_{-1} \in Db_{\mathscr{N}}$ is a generator of $c(\mathcal{N}_{-1}^\square)$.
\end{lemma}
\parag{Proof} As we already know that the distribution $\mathcal{X}_1 $ is a generator of $c(\mathcal{N}_1^\square)$ the point 9 above implies that a distribution $Z$ in $c(\mathcal{N}_0)$ which satisfies $U_1(Z) = \mathcal{X}_1$ must induce a generator in $c(\mathcal{N}_0^\square)$. But $\mathcal{X}_0$ is such a distribution.\\
Thanks to point 10 above and the previous result, the distribution $U_{-1}(\mathcal{X}_0)$ gives a generator in $c(\mathcal{N}_{-1}^\square)$ and we have, using as above the absence of torsion the equality $U_{-1}(\mathcal{X}_0) = \mathcal{X}_{-1}$ in $Db_{\mathscr{N}}$.$\hfill \blacksquare$\\
\parag{End of proof of Theorem \ref{conj.1}} For $\lambda = 0$ we have the exact sequence of $\mathcal{D}_{\mathscr{N}}$-modules (see point 5 above)
$$ 0 \to \mathcal{N}_{0}^\square \to \mathcal{N}_0 \to \mathcal{O}_N(*\sigma_k) \to 0 $$
where the simple sub-module $\mathcal{N}_0^\square$ is generated by $U_1$. So we have an exact sequence of $\mathcal{D}_{\bar{\mathscr{N}}}$-modules
$$ 0 \to c( \mathcal{O}_N(*\sigma_k)) \to c(\mathcal{N}_0) \to c(\mathcal{N}_0^\square) \to 0 $$
and if we find a distribution $Z$ in $c(\mathcal{N}_0) $ such that its image in $c(\mathcal{N}_0^\square)$ is equal to $\mathcal{X}_{-1}$, and a distribution $T$ which generates $c( \mathcal{O}_N(*\sigma_k)) $, then $Z$ and $ T$ will generate $c(\mathcal{N}_0)$. \\
But $U_1(\mathcal{Y}_{0}) = 0$ in $Db_{\mathscr{N}}$ because the coefficient of $\partial_k$ in $U_1$ is $\sigma_1\sigma_k$ and $\sigma_k\partial_k$ kills the distribution $1/\bar\sigma_k$. So $\mathcal{X}_0$ and $ \mathcal{Y}_0$ generates $c(\mathcal{N}_0)$.\\
The situation is analogous for $\lambda = -1$ because $\mathcal{X}_{-1}$ is in $c(\mathcal{N}_{-1})$ and its image generates $c(\mathcal{N}_{-1}^\square)$ thanks to the previous lemma. \\
Now the distribution $\mathcal{Y}_{-1}$ is in $c(\mathcal{N}_{-1})$ and generates a sub-$\mathcal{D}_{\bar N}$-module isomorphic to $\mathcal{O}_{\bar N}(*\bar \sigma_k)$ because its annihilator is generated by $\bar\partial_h, h \in [1, k-1]$ and $\bar\sigma_k\bar\partial_k$. Its image in $c(\mathcal{N}_{-1}^\square)$ is either a generator or $0$. But it has to be non $0$ because there is no non zero morphism between $\mathcal{O}_{\bar N}(*\bar \sigma_k)$ and $c(\mathcal{N}_{-1}^\square)$. So $\mathcal{X}_{-1}$ and $ \mathcal{Y}_{-1}$ generate $c(\mathcal{N}_{-1})$.$\hfill \blacksquare$\\
\parag{Complement}
We give in the following lemma, for each case $\lambda = -1, 0, 1$, a distribution which generates the $\mathcal{D}_{\bar{\mathscr{N}}}$-module $c(\mathcal{N}_\lambda)$.
\begin{lemma}
Consider the distribution $Z_{-1} := \sum_{j=1}^k \bar z_j/z_j$ on $\mathscr{M}$ and define \\ $\mathcal{Z}_{-1} := \pi_*(Z_{-1})$. Then $\mathcal{Z}_{-1}$ is a generator of $c(\mathcal{N}_{-1})$.
\end{lemma}
\parag{Proof} First write $Z_{-1} = \sum_{j=1}^k \vert z_j\vert^2/z_j^2 $ in order to show that any distribution in the $\mathcal{D}_{\mathcal{M}}$-module generated by $Z_{-1}$ has the standard extension property,
by applying Corollary \ref{no torsion M}. Then the same is true for any distribution in the
$\mathcal{D}_{\mathcal{N}}$-module generated by $\mathcal{Z}_{-1}$.
Then it is easy to see that $\mathcal{Z}_{-1}$ is in $c(\mathcal{N}_{-1})$.\\
Then we have $\bar U_{-1}(\mathcal{Z}_{-1}) = \pi_*(\sum_{j=1}^k 1/z_j )= \mathcal{Y}_{-1}$ and
$$ \mathcal{X}_{-1} = \mathcal{Z}_{-1} - (\bar\sigma_1/k) \sigma_{k-1}/\sigma_k = (1 -( \bar \sigma_1/k)\bar U_{-1})( \mathcal{Z}_{-1}) \in \mathcal{D}_{\bar{\mathscr{N}}} \mathcal{Z}_{-1}.$$
So $\mathcal{D}_{\bar{\mathscr{N}}} \mathcal{Z}_{-1}$ is equal to $c(\mathcal{N}_{-1})$ thanks to Theorem \ref{conj.1}.$\hfill \blacksquare$\\
\begin{lemma}
Let $Z_1 := \sum_{j=1}^k \vert z_j \vert^2 - \sigma_1/\sigma_k$ in $Db_{\mathscr{M}}$ and $\mathcal{Z}_1 := \pi_*(Z_1) \in Db_{\mathscr{N}}$. Then $\mathcal{Z}_{1}$ is a generator of $c(\mathcal{N}_{1})$.
\end{lemma}
\parag{Proof} Any distribution in $\mathcal{D}_{\mathscr{N}} \mathcal{Z}_1$ has the standard extension property because this is true for the $\mathcal{D}_{\mathscr{N}}$-module generated by $\pi_*(\sum_{j=1}^k \vert z_j \vert^2 )$ and and by $1/\sigma_k$.
So $\mathcal{Z}_1 $ is in $c(\mathcal{N}_1)$. Moreover we have
$$ (\bar U_0 +k)(\mathcal{Z}_1 ) = (k+1)\pi_*(\sum_{j=1}^k \vert z_j \vert^2 )= (k+1) \mathcal{X}_1 + \frac{k+1}{k}\bar\sigma_1\bar\sigma_k \mathcal{Y}_1 $$
and
$$ (\bar U_0 -1)(\mathcal{Z}_1 ) = (k-1)\mathcal{Y}_1.$$
So $\mathcal{D}_{\bar{\mathscr{N}}} \mathcal{Z}_{1}$ is equal to $c(\mathcal{N}_{1})$ thanks to Theorem \ref{conj.1}.$\hfill \blacksquare$\\
\begin{lemma}
$\mathcal{X}_{0} + \mathcal{Y}_0 $ is a generator of $c(\mathcal{N}_{0})$.
\end{lemma}
\parag{Proof} We already know that $\mathcal{X}_0$ and $\mathcal{Y}_0$ are in $c(\mathcal{N}_{0})$. Moreover we have
$$ (\bar U_0 -1)(\mathcal{X}_0 + \mathcal{Y}_0) = -(k+1)/\bar \sigma_k = -(k+1)\mathcal{Y}_0 $$
and
$$ (\bar U_0 + k )(\mathcal{X}_0 + 1/\bar \sigma_k) = \mathcal{X}_0 .$$
This is enough to conclude, thanks to Theorem \ref{conj.1}.$\hfill \blacksquare$\\
\newpage
|
\section{Introduction}
Manifolds with nonnegative curvature and finite fundamental group are rare and significant in the study of Riemannian geometry. Most constructions involve Lie groups. Homogeneous spaces, biquotients, and certain manifolds admitting actions with cohomogeneity one admit metrics with nonnegative curvature. Such constructions often produce families of metrics, sometimes on a single manifold \(M\). We then investigate the moduli space
\[\modse{sec}(M)=\mathfrak{R}_{\text{sec}
\geq 0}(M)/\text{Diff}(M)\]
where \(\mathfrak{R}_{\text{sec}\geq0}(M)\) is the space of nonnegatively curved metrics on \(M\) and the diffeomorphism group Diff\((M)\) acts on \(\mathfrak{R}_{\text{sec}\geq0}(M)\) by pulling back metrics. We define \(\mods{Ric}(M)\) similarly for metrics of positive Ricci curvature.
If \(M\) is 2- or 3-dimensional, \(\modse{sec}(M)\) and \(\mods{Ric}(M)\) are connected, as can be established using the uniformization theorem or the Ricci flow. Little is known for 4-manifolds, but for \(n\geq 5\), we find examples of \(M^n\) for which the moduli spaces are not only disconnected, but have infinitely many components. Specifically, Dessai and Gonzalez-Alvaro \cite{DGA21} showed that if \(M^5\) is homotopy equivalent to \(\ar{}P^5,\) them \(\modse{sec}(M)\) and \(\mods{Ric}(M)\) have infinitely many path components. The first named author \cite{Goo20b} described an infinite class of total spaces \(N^5\) of nontrivial principal \(S^1\) bundles over simply connected 4 manifolds such that \(\mods{Ric}(N^5)\) has infinitely many path components. Indeed, for every \(k\geq 2\) authors have found examples of \(M^{2k+1}\) with finite fundamental group such that \(\modse{sec}(M^{2k+1})\) and \(\mods{Ric}(M^{2k+1})\) have infinitely many path components, see \cite{DKT18, De17, De20, Goo20a, We21, Wr11}. In this paper, we identify a new collection of 5-manifolds with that property:
\begin{main}\label{thm}
Let \(M^5\) be an orientable, non-spin 5-manifold with \(\pi_1(M)=\f{Z}_2\) and universal cover diffeomorphic to \(S^2\times S^3.\) Then \(\modse{sec}(M)\) and \(\mods{Ric}(M)\) have infinitely many path components.
\end{main}
There are 10 diffeomorphism types of manifolds satisfying the hypotheses of \tref{thm}. The only closed simply connected 5-manifolds known to admit metrics of nonnegative curvature are \(S^5,\) \(S^2\times S^3,\) \(\ensuremath{\operatorname{SU}}(3)/\ensuremath{\operatorname{SO}}(3)\) and the total space of the nontrivial \(S^3\) bundle over \(S^2,\) which we denote by \(S^2\tilde\times S^3.\) Those 4 diffeomorphism types include all five dimensional homogeneous spaces and biquotients, and spaces admitting effective actions of a Lie group which is non-abelian or has rank \(\geq2\) \cite{ADF96, DV14, GGS11, H10, , Si16}. The known methods to distinguish components of \(\modse{sec}(M^5)\) require \(M\) to be an orientable non-spin manifold with a spin universal cover. \(\ensuremath{\operatorname{SU}}(3)/\ensuremath{\operatorname{SO}}(3)\) and \(S^2\tilde\times S^3\) are non-spin. Combining \tref{thm} with the results of \cite{DGA21}, we conclude that \(\modse{sec}\) has infinitely many components for all orientable, non-spin \(\f{Z}_2\) quotients of \(S^5\) and \(S^2\times S^3\). That is, the property holds for all known examples within reach of current methods.
In \sref{manifolds} we describe three families of nonnegatively curved 5-manifolds, and compute topological invariants. The first two families, \(X_{k,l}\) and \(\overline X_{k,l},\) are quotients of \(S^3\times S^3\) by \(S^1\), where \(S^1\) is embedded with weights \(k\) and \(l\) into a \(T^2\) action on \(S^3\times S^3\). The families \(X_{k,l}\) and \(\overline{X}_{k,l}\) differ by the \(T^2\) action used. The manifolds are also total spaces of principal \(S^1\) bundles over \(\f{C}P^2\#\f{C}P^2\) or \(\f{C}P^2\#\overline{\f{C}P}^2;\) it was shown in \cite{Goo20b} that for those manifolds, \(\mods{Ric}\) has infinitely many components.
The third family, \(Q^5_0(d),\) are \(\f{Z}_2\) quotients of Brieskorn variteties, namely the intersection of \(S^7\) with the variety in \(\f{C}^4\) defined by \(z_0^d+z_1^2+z_2^2+z_3^2=0\) for \(d\) even. They can be equipped with a metric of nonnegative curvature using a cohomogeneity one action by \(S^1\times \O(3)\) as in \cite{GVWZ06}.
In \sref{diffeo}, we sort those families into diffeomorphism types using Su's classification \cite{Su12} along with a computation of the primary diffeomorphism invariant for the families \(X_{k,l}\) and \(\overline{X}_{k,l}\) from \cite{Goo20b}. In particular, we see that any manifold satisfying the hypotheses of \tref{thm} is diffeomorphic to infinitely many members of one of the three families. To prove \tref{thm} we pull back nonnegatively curved metrics using those diffeomorphisms and show that the metrics obtained represent infinitely many components of \(\modse{sec}.\) Furthermore, we use the Ricci flow to show that each metric has a neighbor with positive Ricci curvature, and those neighbors represent infinitely many components of \(\mods{Ric}\).
We use the relative \(\eta\) invariant of a spin\(^c\) Dirac operator to distinguish connected components of \(\modse{sec}\) and \(\mods{Ric}.\) The nonnegatively curved metric on each manifold \(M\) described in \sref{manifolds} can be lifted to a \(\f{Z}_2\) invariant metric on the universal cover \(\tilde M\) which can in turn be extended to a \(\f{Z}_2\) invariant metric of positive scalar curvature on a 6-manifold \(W\) such that \(\partial W=\tilde M.\) The relative \(\eta\) invariant of \(M\) can then be computed using a Lefschetz fixed point theorem for manifolds with boundary, developed by Donnely in \cite{Do78} based on the index theorem for manifolds with boundary of \cite{APSI75}. The invariant is given in terms of the index of a Dirac operator, which vanishes because of the positive scalar curvature, and an integral over the fixed point set of the \(\f{Z}_2\) action on \(W.\)
\section{Preliminaries}
\subsection{Diffeomorphism Classification}\label{subsec:diffeomorphismclassification}
Our result depends on a diffeomorphism classification of quotients of \(S^2\times S^3\) due to Su. The key diffeomorphism invariant is the cobordism class of a characteristic submanifold. Given a manifold \(X^n\) with \(\pi_1(X^n)=\f{Z}_2,\) let \(f:X^n\to\ar{}P^N\) be a map which is an isomorphim on \(\pi_1\) and transverse to \(\ar{}P^{N-1}\subset \ar{}P^{N}.\) Then \(P=f^{-1}(\ar{}P^{N-1})\) is a characteristic submanifold of \(X^n.\) If \(X^n\) is non-spin but has spin universal cover, \(P\) admits two Pin\(^+\) structures which represent inverse classes in \(\Omega_{n-1}^{\text{Pin}^+}\). The class \([P]\in\Omega_{n-1}^{\text{Pin}^+}/\pm\) is a diffeomorphism invariant. \(\Omega_4^{\text{Pin}^+}\) is isomorphic to \(\f{Z}_{16},\) generated by \([\ar{}P^4],\) and we use the identification \(\Omega_4^{\text{Pin}^+}/\pm\cong\{0,1,...,8\}\). See \cite{GT98, KT90, LdM71, Su12} for details.
The classification in \cite{Su12} uses two families of model manifolds. The family \(X(q),\) \(q\in\{0,2,4,6,8\}\) are constructed from pairs of homotopy \(\ar{}P^5\)'s by removing tubular neighborhoods of generators of the fundamental group and gluing along the boundaries. The second family \(Q_0^5(d)\), \(d\in\f{N}_0,\) are quotients of Brieskorn varieties and will be described in \sref{nontrivial}.
\begin{thm}\cite{Su12}\label{su}
Let $M^5$ be a smooth, orientable, non-spin 5-manifold with $\pi_1(M)\cong {\mathbb{Z}}_2$ and universal cover diffeomorphic to $S^2\times S^3$. Let $P\subset M$ be a characteristic submanifold.
\begin{enumerate}
\item If $\pi_1(M)$ acts trivially on $\pi_2(M)$, then $M$ is diffeomorphic to $X(q)$ for $q\in\{0,2,4,6,8\}$ such that $[P]=q\in \Omega^{Pin^+}_4/\pm$.
\item If $\pi_1(M)$ acts non-trivially on $\pi_2(M)$, then $M$ is diffeomorphic to $Q^5_0(d)$ for $d\in \{0,2,4,6,8\}$ such that $[P]=d\in\Omega^{Pin^+}_4/\pm$.
\end{enumerate}
\end{thm}
\subsection{Relative \(\eta\) Invariants}
The relative \(\eta\) invariant allows us to distinguish path components in \(\modse{sec}\) and \(\mods{Ric}\). Specifically, the relative \(\eta\) invariant is constant on each path component of the space of positive scalar curvature metrics on a closed manifold. Here we define the relative \(\eta\) invariant and present some of its properties. See \cite{We21, DGA21, APSII75, Do78} for more details.
Let $Q^{2n-1}$ be a closed $spin^c$ manifold with finite fundamental group, Riemannian metric $g$ and a flat connection on the principal \(\ensuremath{\operatorname{U}}(1)\) bundle associated to the spin$^c$ structure. Let $M$ be the universal cover of \(Q\) with the lifted spin\(^c\) structure. Let $\alpha:\pi_1(Q)\rightarrow U(k)$ be a unitary representation and $E_\alpha:=M\times_{\alpha}\mathbb{C}^k$ the corresponding flat vector bundle over $Q$. Let \(D_Q\) be the spin\(^c\) Dirac operator on \(Q\), \(D_Q\otimes E_\alpha\) the twisted spin\(^c\) Dirac operator, and \(D_M\) the spin\(^c\) Dirac operator on \(M\). For a self adjoint elliptic operator \(D\) with spectrum \(\{\lambda\}\), counted with multiplicity, $\eta(D):=\eta(0)$ is the analytic continuation to \(z=0\) of the function \(\eta:\f{C}\to\f{C}\), $$\eta(z):=\sum_{\lambda\neq 0} \frac{\text{sign}(\lambda)}{|\lambda|^z}$$ which is analytic when the real part of \(z\) is large.
The \emph{relative $\eta$ invariant} of $Q$ is defined as
\begin{equation}\label{def:relative-eta-invariant}
\eta_\alpha(Q,g):=\eta(D_Q\otimes E_\alpha)-k\cdot\eta(D_Q).
\end{equation} The following result is the key to distinguish path components in the moduli spaces. See \cite[Proposition 3.3]{DGA21} for the proof.
\begin{prop}\label{prop:etainvsame}
Let $Q^{2n-1}$ be a closed connected $spin^c$ manifold, $\alpha:\pi_1(Q)\rightarrow U(k)$ a unitary representation and $E_\alpha$ the associated flat complex vector bundle. Suppose that the principal $U(1)$-bundle associated to the $spin^c$ structure on $Q$ is given a flat connection. Let $g_0$ and $g_1$ be two metrics of $scal>0$ which lie in the same path component in $\mathfrak{R}_{scal>0}(Q)$. Then ${\eta}_\alpha(Q,g_0)={\eta}_\alpha(Q,g_1)$.
\end{prop}
By \cite[(2$\cdot$14)]{APSII75}, the twisted $\eta$ invariant on $Q$ can be computed in terms of the equivariant $\eta$ invariant of $M$
\begin{equation}\label{twistedeta}
\eta(D_Q\otimes E_\alpha)=\frac{1}{|\pi_1(Q)|}\sum_{g\in \pi_1(Q)}\eta_g(D_M)\cdot\chi_\alpha(g).
\end{equation} Here $\chi_\alpha$ is the character of $\alpha$ and the \emph{equivariant eta-invariant} is defined as $\eta_g(D_M):=\eta_g(0),$ the analytic continuation to \(z=0\) of $$\eta_g(z):=\sum_{\lambda\neq 0}\frac{\text{sign}(\lambda)tr(g^\#_\lambda)}{|\lambda|^z},$$ where the eigenvalues are counted without multiplicity, and $g^\#_\lambda$ is the map induced by $g$ on the eigenspace of $D_M$ corresponding to eigenvalue \(\lambda\).
We now discuss how to compute the relative $\eta$ invariant in the context of this paper. Let $W^{2n}$ be a compact $spin^c$ manifold with simply connected boundary $M^{2n-1}=\partial W$. Assume that $W$ is of product form near the boundary, that is, there is a an neighborhood of $M$ which is isometric to $I\times M$, for an interval I. Equip the principal \(\ensuremath{\operatorname{U}}(1)\) bundle associated to the spin\(^c\) structure with a connection which near the boundary is constant in the direction determined by the interval \(I.\) The complex spinor bundle associated to the $spin^c$ structure on $W$ decomposes as $S=S^+\oplus S^-$. Let $D^+_W:\Gamma(S^+)\rightarrow \Gamma(S^-)$ be the \(spin^c\) Dirac operator on $W$ and $D_M$ the $spin^c$ Dirac operator on $M$.
Next, let $\tau$ be an involution on $W$ which is fixed point free on $M$. Define $Q^{2n-1}:=M/\tau$ and let $\alpha:{\mathbb{Z}}_2\rightarrow U(1)$ be the nontrivial unitary representation of the fundamental group of $Q$. By \eqref{twistedeta}
\begin{equation*}
\eta(D_Q\otimes E_\alpha)=\frac{1}{2}(\eta(D_M)-\eta_\tau(D_M))
\end{equation*} whereas for the trivial representation we have
\begin{equation*}
\eta(D_Q)=\frac{1}{2}(\eta(D_M)+\eta_\tau(D_M)).
\end{equation*}
The relative $\eta$ invariant is then given by \(\eta_\alpha(Q,g)=-\eta_\tau(D_M).\)
By Donnelly's equivariant index theory \cite{Do78}
$$\text{index}(D_W,\tau)=\sum_{N\subset W^\tau}a(N)-\frac{h_\tau(D_M)+\eta_\tau(D_M)}{2},$$ where $\text{index}(D_W,\tau):=tr(\tau|_{ker D_W})-tr(\tau|_{ker D^*_W})$, $N$ is a component of the fixed point set \(W^\tau\), $h_\tau(D_M):=tr(\tau|_{ker D_M})$, and $a(N)$ is a so-called \emph{local contribution}, the integral over \(N\) of a differential form depending on the metric on \(N\) and the action of \(\tau\) on the normal bundle to \(N.\) \footnote{For the precise definition of the local contribution $a(N)$ see \cite{Do78}. When \(N\) is a closed manifold, \(a(N)\) is the integrand in the equivariant index theorem for closed manifolds; see \cite[\S3]{ASIII68} and \cite[\mathbb{S} III.14]{LM89}. }
If we assume that the metric on $W$ has scal $\geq 0$ everywhere, the connection on the principal \(\ensuremath{\operatorname{U}}(1)\) bundle associated to the spin\(^c\) structure is flat, and the metrics on $M$ and $Q$ have scal $>0$, then by the usual Lichnerowicz argument (see \cite{Li63}, \cite{LM89}, and \cite{APSII75}) it follows that $\text{index}(D_W,\tau)$ and $h_\tau(D_M)$ vanish. The relative $\eta$ invariant is then given by
\begin{equation}\label{fixedpointformula}
\eta_\alpha(Q,g)=-\eta_\tau(D_M)=-2\sum_{N\subset W^\tau}a(N).
\end{equation}
The computation of the local contribution depends on the situation and will be given in the proofs of \lref{Qeta} and \lref{etavalues}.
\section{Families of nonnegatively curved manifolds}\label{manifolds}
In this section, we describe three families of nonnegatively curved five manifolds satisfying the hypotheses of \tref{thm}. We calculate topological invariants of those manifolds, and the bundles used to describe them, as required for the diffeomorphism classification in \sref{diffeo} and the computation of the relative \(\eta\) invariants in \sref{eta}.
\subsection{The fundamental group acts trivially on higher homotopy groups}
Given relatively prime integers \(k,l\), define the homomorphism
\[i_{k,l}:U(1)\to T^2,\quad z\mapsto (z^k,z^l).\]
We describe two free actions of \(T^2\) on \(S^3\times S^3\) by isometries of the product of round metrics. To describe the actions we consider \(S^3\times S^3\subset \f{C}^2\times\f{C}^2.\)
\subsection*{Case I}
Let \(T^2\) act on \(S^3\times S^3\) by means of the homomorphism
\begin{equation}\label{action1}T^2\to T^2\times T^2\subset U(2)\times U(2)\end{equation}
\[(z,w)\mapsto((z,zw),(w,w)).\]
The action is free. Define \(B^4=S^3\times S^3/T^2.\) For integers \(k,l\), where \(k\) is odd, \(l\) is even, and gcd\((k,l)=1\), define \(N_{k,l}^5=S^3\times S^3/i_{k,l}(U(1))\). One can use the results of \cite{DV14} to identify \(B^4\cong\f{C}P^2\#\overline{\f{C}P^2}\) and \(N^5_{k,l}\cong S^2\times S^3.\) \(N^5_{k,l}\) admits a free action by \(T^2/i_{k,l}(U(1))\cong U(1)\) with quotient \(B^4\). Let \(c_{k,l}\in\zco{B^4}{2}\) be the first Chern class of the principal \(U(1)\) bundle \(N^5_{k,l}\to B^4.\) Let \(X_{k,l}^5=N_{k,l}^5/\f{Z}_2\) be the quotient of \(N_{k,l}\) by the action of \(\f{Z}_2=\{\pm1\}\subset U(1).\) \(X_{k,l}^5\) admits a free action by \(U(1)/\f{Z}_2\cong U(1)\) with quotient \(B^4\). The first Chern class of that bundle is \(2c_{k,l}.\) By O'Neil's formula, \(N_{k,l}^5\) admits an \(U(1)\) invariant Riemmanian metric of nonnegative sectional and positive scalar curvature which descends to such a metric on \(X_{k,l}^5\).
\subsection*{Case II} Let \(T^2\) act on \(S^3\times S^3\) by means of the homomorphism
\[T^2\to T^2\times T^2\subset U(2)\times U(2)\]
\[(z,w)\mapsto((z,zw),(w,z^2w)).\] Define \(\overline{B}^4=S^3\times S^3/T^2,\) \(\overline{N}_{k,l}^5=S^3\times S^3/i_{k,l}(U(1)),\) and \(\overline{X}_{k,l}^5=\overline{N}_{k,l}^5/\f{Z}_2\) exactly as in Case I, but with this distinct \(T^2\) action. One can use the results of \cite{DV14, To02} to identify \(\overline{B}^4\cong\f{C}P^2\#{\f{C}P^2}\) and \(\overline{N}^5_{k,l}\cong S^2\times S^3.\) Let \(\overline{c}_{k,l}\in\zco{\overline{B}^4}{2}\) be the first Chern class of the principal \(U(1)\) bundle \(\overline{N}^5_{k,l}\to \overline{B}^4.\) Again we have a principal \(U(1)\) bundle \(\overline X_{k,l}^5\to \overline B^4\)with first Chern class \(2\overline{c}_{k,l}.\) By O'Neil's formula, \(\overline{N}_{k,l}^5\) admits an \(S^1\) invariant Riemmanian metric of nonnegative sectional and positive scalar curvature which descends to a such a metric on \(\overline{X}_{k,l}^5\).
We compute the cohomology rings of \(B^4\) and \(\overline{B}^4\) in terms of generators with which it will be straighforward to identify \(c_{k,l}\) and \(\overline{c}_{k,l}.\) Let \(p_1\) and \(w_2\) denote the first Pontryagin and second Stiefel-Whitney classes respectively.
\begin{lem}\label{invariants}
\
\begin{enumerate}
\item \(\zco{B^4}{*}=\f{Z}[u,v]/(u^2+uv,v^2)\) \item \(p_1(TB^4)=0,\) \(w_2(TB^4)=v\text{\normalfont\ mod } 2\), and \(c_{k,l}=-lu+kv.\) \item \(w_2(N^5_{k,l})=0\) and \(w_2(X^5_{k,l})\neq0.\)
\item \(\zco{\overline{B}^4}{*}=\f{Z}[\bar u,\bar v]/(\bar u^2+\bar u\bar v,\bar v^2+2\bar u\bar v)\)
\item \(p_1(T\overline B^4)=6\bar u^2,\) \(w_2(T\overline B^4)=\bar v\text{\normalfont\ mod } 2\), and \(c_{k,l}=-l\bar u+k\bar v.\) \item \(w_2(\overline N^5_{k,l})=0\) and \(w_2(\overline X^5_{k,l})\neq0.\)
\end{enumerate}
\end{lem}
\begin{proof} Let \(\phi:B^4\to BT^2\) be the classifying map of the bundle \(T^2\to S^3\times S^3\to B^4\). The sequence of maps \(S^3\times S^3\to B^4\to BT^2\) is homotopy equivalent to the fibration \[S^3\times S^3\to ET^2\times_{T^2}(S^3\times S^3)\to BT^2.\] The representation \eqref{action1} is the product of representations
\[\rho_1:T^2\to T^2\subset U(2):(z,w)\to(z,zw)\]
\[\rho_2:T^2\to T^2\subset U(2):(z,w)\to(w,w).\]
For \(i=1,2\) we have bundle maps
\[\begin{tikzcd}
S^3\times S^3\arrow{r}\arrow{d} & ET^2\times_{T^2}(S^3\times S^3)\arrow{r}\arrow{d}& BT^2\arrow{d} \\
S^3\arrow{r}& ET^2\times_{\rho_i}S^3\arrow{r} & BT^2
\end{tikzcd}\]
induced by projections onto the first or second factor of \(S^3\times S^3\). Let \(e_i\) be the Euler class of the sphere bundle \(ET^2\times_{\rho_i}S^3\to BT^2\). Using the spectral sequences of the fibrations above one sees that \(\phi^*\) induces an isomorphism
\[H^*(B,\f{Z})\cong H^*(BT^2)/(e_1,e_2).\]
Let \(\tau:H^*(T^2,\f{Z})\to H^*(BT^2,\f{Z})\) be the transgression of the universal bundle. The weights of \(\rho_1\) are \((1,0)\) and \((1,1)\in\) Hom\((\f{Z}^2,\f{Z})=\text{Hom}(\pi_1(T^2),\f{Z})=H^1(T^2,\f{Z})\). By \cite{BH58}, \(e_1=(-\tau((1,0)))(-\tau((1,1)))\). Let \(u=-\tau((1,0))\) and \(v=-\tau((0,1))\). Then \(e_1=u^2+uv\). Similarly, the weights of \(\rho_2\) are both \((0,1)\) and \(e_2=v^2\). We reuse the notation \(u,v\) for their images under \(\phi^*\). That completes the proof of statement (1) of the lemma.
It follows immediately that the signature of \(B\) is 0, and by Hirzebruch's signature theorem \(p_1(TB)=0\). Since \(w_2(TB)^2=p_1(TB)\) mod 2, we can further conclude that \(w_2(TB)=v\) mod 2 or 0. However, \(B\) is not spin (this follows from the intersection form) and thus \(w_2(TB)=v\) mod 2.
One checks that \(U(1)\to N_{k,l}\to B\) is the \(U(1)\)-bundle associated to \(T^2\to S^3\times S^3\to B\) by the representation
\[T^2\to U(1): (z,w)\mapsto z^{-l}w^k\]
which has weight \((-l,k)\).
Let \(\tau_\phi:H^*(T^2,\f{Z})\to H^*(B,\f{Z})\) be the transgression of \(T^2\to S^3\times S^3\to B\). Then again by \cite{BH58} we have \(c_{k,l}=\tau_\phi((l,-k)))\). By naturality of the transgression, \(\tau_\phi=\phi^*\tau\), so \(c_{k,l}=-lu+kv.\) Since \(k\) is odd and \(l\) is even, \(v\) mod 2 is in the kernel of the mod 2 Gysin sequence for \(U(1)\to N_{k,l}\to B\), and thus \(w_2(TN_{k,l})=0\). By the same argument, the first Chern class of \(U(1)\to X_{k,l}\to B\) is \(2c_{k,l}\), and the corresponding map in the Gysin sequence is an isomorphism, so \(w_2(TX_{k,l})\neq0\).
The case of \(\overline{B}\) follows similarly, with \(\overline{u}\) and \(\overline{v}\) the images of generators of \(H^*(T^2,\f{Z})\) under minus the transgression of \(T^2\to S^3\times S^3\to \overline{B}.\) If we orient \(\overline{B}\) such that \(\left<\overline{u}^2,[\overline{B}]\right>=1\) the cohmology ring implies that the signature of \(\overline{B}\) is 2. So \(p_1(T\overline{B})=6\overline u^2\) and \(w_2(T\overline{B})=\overline{v}\) mod 2. The rest of the lemma follows.
\end{proof}
We remark that existence of a diffeomorphism between \(N_{k,l}\) or \(\overline N_{k,l}\) and \(S^2\times S^3\) follows immediately from the cohomology and second Stiefel Whitney class, by the work of Smale \cite{Sm62}.
\subsection{The fundamental group acts non-trivially on higher homotopy groups}\label{nontrivial}
The following spaces were first studied by Brieskorn \cite{Bri66}. See also \cite{HM68}, \cite{Bre72} and \cite[Chapter 9]{BoGa07} for more details.
Let $D^8:=\{z\in {\mathbb{C}}^4||z_0|^2+|z_1|^2+|z_2|^2+|z_3|^2\leq 1\}$ and $S^7:=\{z\in {\mathbb{C}}^4||z_0|^2+|z_1|^2+|z_2|^2+|z_3|^2= 1\}$ be the unit disk and unit sphere in ${\mathbb{C}}^4$ respectively. Let $f_d:{\mathbb{C}}^4\rightarrow {\mathbb{C}}$ be defined as $f_d(z):=z^{d}_0+z^2_1+z^2_2+z^2_3$, for $d\in {\mathbb{N}}_0$. For $\epsilon\in {\mathbb{R}}_{\geq 0}$, we define the \emph{Brieskorn varieties}\label{def:brieskornvar}
\vspace{-0.5mm}
$$W^6_\epsilon(d):=D^8\cap f_d^{-1}(\epsilon),$$
\vspace{-5mm}
$$M^5_\epsilon(d):=S^7\cap f_d^{-1}(\epsilon).$$
For $\epsilon > 0$, $W^6_\epsilon(d)$ is a smooth complex manifold with boundary $\partial W^6_\epsilon(d)=M^5_\epsilon(d)$, whereas for $\epsilon=0$, $M^5_0(d)$ is a smooth manifold but $W^6_0(d)$ is a variety with an isolated singular point at $z=0$. $M^5_\epsilon(d)$ comes with a natural orientation as a link and is sometimes also called a \emph{Brieskorn manifold}. We summarize some properties of these spaces in the following theorem; see \cite{HM68} and \cite{GT98} for details.
\begin{thm}\hfill\label{thm:brieskvarprop}
\begin{enumerate}
\item $W^6_\epsilon(d)$ is homotopy equivalent to a bouquet $S^3\vee ... \vee S^3$ with $d-1$ summands.
\item For $\epsilon$ sufficiently small, $M^5_\epsilon(d)$ is diffeomorphic to $M^5_0(d)$.
\item If $d$ is odd, then $M^5_0(d)$ is diffeomorphic to $S^5$.
\item If $d$ is even, then $M^5_0(d)$ is diffeomorphic to $S^2\times S^3$.
\end{enumerate}
\end{thm}
\subsection*{Cohomogeneity one action}\label{sec:cohomonebries}
An action of a compact Lie group $G$ on a smooth manifold $M$ is said to be of \emph{cohomogeneity one} if the orbit space $M/G$ is one-dimensional (see for example \cite[\textsection 6.3]{AlBe15} for more details on such actions).
For more details on the following cohomogeneity one action, see also \cite[\textsection 4.2]{DGA21} and \cite[\textsection 1]{GVWZ06}. Let $U(1)\times O(3)$ act on ${\mathbb{C}}^4$ in the following way. For $(w,A)\in U(1)\times O(3)$ and $z\in {\mathbb{C}}^4$, we set $$(w,A)\cdot z=(w^2z_0,(A(w^dz_1,w^dz_2,w^dz_3)^T)^T).$$ This restricts to an action by ${\mathbb{Z}}_{2d}\times O(3)$ on $W^6_\epsilon(d)$ and $M^5_\epsilon(d)$ ($\epsilon\neq 0$), and by $U(1)\times O(3)$ on $W^6_0(d)$ and $M^5_0(d)$. The action on $M^5_0(d)$ is of cohomogeneity one. Indeed, each orbit is uniquely identified by the value of \(|z_0|\), and the range of \(|z_0|\) is \([0,t]\), where \(t\) is the unique positive solution to \(t^d+t^2=1\). The principal isotropy type, when \(|z_0|\in(0,1)\), is ${\mathbb{Z}}_2\times O(1)$. The singular isotropy types are $U(1)\times O(1)$ when \(|z_0|=0\) and ${\mathbb{Z}}_2\times O(2)$ when \(|z_0|=t\). The corresponding singular orbits are both of codimension two.
Now consider $\tau=(1,-\text{Id})\in U(1)\times O(3)$. \(\tau\) is a holomorphic, orientation preserving involution on $W^6_\epsilon(d)$, acting without fixed points on $M^5_\epsilon(d)$ for $0\leq\epsilon<1$. For $0<\epsilon<1$, the fixed points of the action of $\tau$ on $W^6_\epsilon(d)$ are $p_j=(\lambda_j,0,0,0)$, $1\leq j\leq d$, where $\lambda_j$ is a complex $d$-root of $\epsilon$ for all $j$. These isolated fixed points lie in the interior of $W^6_\epsilon(d)$.
We call the quotient manifold $Q^5_\epsilon(d):=M^5_\epsilon(d)/\tau$ a \emph{Brieskorn quotient}\label{def:brieskornquotient}. \(Q^5_\epsilon(d)\) admits an action by \(O(3)/(\pm\text{Id})=SO(3)\). Some important properties of these quotients are summarized in the following; see \cite{Bro67} and \cite{GT98}.
\begin{thm}\label{thm:briesquoproperties}\hfill
\begin{enumerate}
\item For all $\epsilon$ sufficiently small, $Q^5_\epsilon(d)$ is diffeomorphic to $Q^5_0(d)$.
\item For $d$ odd, $Q^5_0(d)$ is homotopy equivalent to ${\mathbb{R}}\emph{P}^5$.
\item For \(d\) even, $\pi_1(Q^5_0(d))$ acts non-trivially on $\pi_2(Q^5_0(d))\cong {\mathbb{Z}}$, $H_2(Q^5_0(d);{\mathbb{Z}})=0$, and $w_2(Q^5_0(d))\neq 0$
\end{enumerate}
\end{thm}
From now on, we will always assume that $d$ is even, unless otherwise stated.
We can now use the construction of Grove-Ziller \cite{GZ00} to equip \(M^5_0(d)\) with an \(U(1)\times O(3)\) invariant metric of nonnegative sectional curvature, which descends to nonnegatively curved \(U(1)\times \ensuremath{\operatorname{SO}}(3)\) invariant metric on \(Q^5_0(d)\). Using Cheeger deformations, one obtains a metric which simultaneously has nonnegative sectional and positive scalar curvature on \(Q^5_0(d)\) (see \cite[\textsection 6.1.4]{We21}).
\section{Diffeomorphism Types}\label{diffeo}
In this section we identify the diffeomorphism types of the manifolds described in \sref{manifolds} using \tref{su}. For the family \(Q^5_0(d),\) the following is immediate.
\begin{lem}\label{Qdiff}
Let \(M\) satisfy the hypotheses of \tref{thm}. Assume further that \(\pi_1(M)\) acts non-trivially on \(\pi_2(M)\). For some $d\in\{0,2,4,6,8\}$, \(M\) is diffeomorphic to $Q^5_0(d+16k)$ for all \(k\in\f{N}_0\).
\end{lem}
For the families \(X_{k,l}\) and \(\overline X_{k,l}\), since \(\f{Z}_2\subset U(1)\) acts trivially on \(\pi_2(N_{k,l})\) and, \(\pi_2(\overline N_{k,l})\), \(\pi_1(X_{k,l})\) and \(\pi_1(\overline X_{k,l})\) act trivially on \(\pi_2(X_{k,l})\) and \(\pi_2(\overline X_{k,l})\) respectively. It follows from \lref{invariants} and \tref{su} that the diffeomorphism types of \(X_{k,l}\) and \(\overline X_{k,l}\) are determined by the cobordism classes of characteristic submanifolds. Those cobordism classes can in turn be determined by a result in \cite{Goo20b} using topological invariants of the bundles \(X_{k,l}\to B\) and \(\overline X_{k,l}\to\overline B.\)
\begin{lem}\label{characteristics}
Let \(P\subset X^5_{k,l}\) and \(\overline{P}\subset \overline X^5_{k,l}\) be a characteristic submanifolds. Identifying \(\Omega_4^{\text{Pin}^+}\) with \(\f{Z}_{16}\), \(P\) and \(\overline{P}\) admit Pin\(^+\) structures such that
\[[P]=\left(1+\frac{\ep}{2}\right)\left(l^2+2kl\right)\text{ \normalfont mod }16\]
and
\[[\overline{P}]=\left(1+\frac{\ep}{2}\right)\left(l^2+2kl+2k^2\right)-\ep\text{ \normalfont mod }16\]
where \(\ep\in\{1,-1\}\) is an unknown constant.
\end{lem}
\begin{proof}
As we saw in the proof of \lref{invariants}, the principal bundle \(U(1)\to X_{k,l}\to B\) has first Chern class \(2c_{k,l}\). Furthermore, since \(l\) is even and \(k\) is odd, \(c_{k,l}=v\) mod 2, so \(w_2(TB)=c_{k,l}\) mod 2. Lemmas 1.5 and 1.7 in \cite{Goo20b} then imply that a characteristic submanifold \(P\subset X_{k,l}\) admits a Pin\(^+\) structure such that
\[[P]=\left(1+\frac{\ep}{2}\right)\left<c_{k,l}^2,[B]\right>-\frac{\ep}{2}\ensuremath{\operatorname{sign}}(B).\] Using \lref{invariants} we can orient \(B\) such that \(\langle uv,B\rangle=1\), and compute
\(\langle c_{k,l}^2,[B]\rangle =-l^2-2lk\). The formula for \([\overline{P}]\) follows using the orientation of \(\overline{B}\) such that \(\langle \overline{u}^2,[\overline{B}]\rangle=1\) and sign\(([\overline{B}])=2.\)
\end{proof}
We combine \lref{characteristics} and \tref{su} to prove
\begin{lem}\label{infinitelists}
Let \(M\) satisfy the hypotheses of \tref{thm}. Assume further that \(\pi_1(M)\) acts trivially on \(\pi_2(M)\). Then \(M\) is diffeomorphic to either \(X_{k,2}\) or \(\overline X_{k,4}\) for an infinitely many \(k\in\f{Z}.\)
\end{lem}
\begin{proof}
By \tref{su}, \(M\) is diffeomorphic to \(X(q)\) for some \(q\in\{0,2,4,6,8\}.\) So we need only prove the lemma for those manifolds. Applying \lref{characteristics} and \tref{su}, we see that for all \(r\in\f{Z}\)
\begin{align*} X(0)\cong X^5_{8r-1,2} && X(2)\cong \overline X^5_{8r+2\ep+3,4} \\
X(4)\cong X^5_{8r+1,2} && X(6)\cong \overline X^5_{8r+2\ep-1,4} \\
X(8)\cong X^5_{8r+3,2}
\end{align*}
where \(\ep\) is the unkown constant in \lref{characteristics}
\end{proof}
We note that each \(X(q)\) is diffeomorphic to infinitely many \(X_{k,l}^5\) or \(\overline X_{k,l}^5\) with \(l\) different from \(2\) or \(4\), but the lists above suffice for our purpose.
\section{Computation of the \(\eta\) invariants}\label{eta}
In this section, we compute relative \(\eta\) invariants for the manifolds \(Q^5_0(d)\), \(X^5_{k,l}\) and \(\overline X^5_{k,l}\) and prove \tref{thm}. That proof will utilize the following lemma, which follows arguments of \cite{DGA21} and \cite{We21}.
\begin{lem}\label{modulispaces}
Let \(M^{2n+1}\) be a closed spin\(^c\) manifold with finite fundamental group \(\pi_1(M)\) and \(\alpha:\pi_1(M)\to U(k)\) a unitary representation. Let \(\{M_i\}\) be an infinite set of spin\(^c\) manifolds, diffeomorphic to \(M\), each equipped with a flat connection on the principal \(U(1)\) bundle associated to the spin\(^c\) structure. For each \(i\) let \(\alpha_i:\pi_1(M_i)\to U(k)\) be a unitary representation which pulls back to \(\alpha\) and \(g_i\) a nonnegatively curved Riemannian metric on \(M_i\) with positive scalar curvature. If \(\{\eta_{\alpha_i}(M_i,g_i)\}\) is infinite, then \(\modse{sec}(M)\) and \(\mods{Ric}(M)\) have infinitely many path components.
\end{lem}
\begin{proof}
We first make a reduction to facilitate the proof. Since the connections described in the lemma are all flat, each spin\(^c\) structure pulls back to a spin\(^c\) structure on \(M\) with flat associated \ensuremath{\operatorname{U}}(1) bundle. There are finitely many isomorphism types of such bundles and finitely many spin\(^c\) structures corresponding to each type. Thus we can give \(M\) a spin\(^c\) structure such that, restricting the set \(\{M_i\}\) if necessary, we may assume the diffeomorphisms between \(M\) and each \(M_i\) are spin\(^c\) structure preserving, without changing the other hypotheses. We equip the principal \(U(1)\) bundle associated to that spin\(^c\) structure on \(M\) with a flat connection.
Next, let \(H\subset\text{Diff}(M)\) be the group of diffeomorphisms which preserve \(\alpha\) and the spin\(^c\) structure. By the previous discussion the orbit of the spin\(^c\) structure under Diff\((M)\) is finite. Since \(\pi_1(M)\) is finite, the orbit of \(\alpha\) is also finite. Thus \(H\) has finite index in Diff\((M)\). It follows that \(\modse{sec}\) and \(\mods{Ric}\) have infinitely many components if and only if \(\mathfrak{R}_{\text{sec}\geq0}(M)/H\) and \(\mathfrak{R}_{\text{Ric}>0}(M)/H\) respectively have infinitely many components.
For each \(i\) reuse the notation \(g_i\) for the pullback of that metric to \(M\). Since the diffeomorphism from \(M\) to \(M_i\) is spin\(^c\) preserving, \(\alpha_i\) pulls back to \(\alpha\), and any two flat connections on a principal \(U(1)\) over a manifold with finite fundamental group are isomorphic, \(\eta_\alpha(M,g_i)=\eta_{\alpha_i}(M_i,g_i)\). Denote by \([g_i],[g_j]\) the images of \(g_i,g_j\) in \(\mathfrak{R}_{\text{sec}\geq0}(M)/H\). Suppose \([g_i]\) and \([g_j]\) lie in the same path component. By Ebin's slice theorem, there exists a path \(\gamma:[0,1]\to\mathfrak{R}_{\text{sec}\geq0}(M)\) such that \(\gamma(0)=g_i\) and and \(\gamma(1)=\phi^*g_j\) for some \(\phi\in H\). As demonstrated in \cite{BW07}, if we apply the Ricci flow for any small positive amount of time to \(\gamma\), we obtain a path \(\tilde\gamma:[0,1]\to\mathfrak{R}_{\text{Ric}>0}(M).\) Concatenating \(\tilde\gamma\) with the path formed by applying the Ricci flow to \(g_i\) and \(\phi^*g_j\) we get a path in \(\mathfrak{R}_{\text{scal}>0}(M)\) connecting \(g_i\) and \(\phi^*g_j\). Thus by \pref{prop:etainvsame} \[\eta_{\alpha_i}(M_i,g_i)=\eta_\alpha(M,g_i)=\eta_\alpha(M,\phi^*g_j)=\eta_\alpha(M,g_j)=\eta_{\alpha_j}(M_j,g_j).\]
Thus since \(\{\eta_{\alpha_i}(M_i,g_i)\}\) is infinite, the metrics \(g_i\) must represent infinitely many path components of \(\mathfrak{R}_{\text{sec}\geq0}(M)/H\). The metrics of positive Ricci curvature obtained by applying the Ricci flow to \(g_i\) must represent infinitely many components of \(\mathfrak{R}_{\text{Ric}>0}(M)/H.\)
\end{proof}
The strategy in \cite{DGA21} applies to compute the relative \(\eta\) invariants of \(Q^5_0(d)\) (see also \cite{We21}). For $\epsilon>0$, the coboundary $W^6_\epsilon(d)$ comes with a canonical $spin^c$ structure induced by the complex manifold structure. By \tref{thm:brieskvarprop}, \(H^2(W_\ep^6(d),\f{Z})=0\), so the principal \(\ensuremath{\operatorname{U}}(1)\)-bundle associated to the spin\(^c\) structure is isomorphic to the trivial bundle and admits a flat connection. $M^5_\epsilon(d)$ inherits a $spin^c$ structure as the boundary of $W^6_\epsilon(d)$. The action of $\tau$, which is holomorphic on \(W^6_\epsilon(d)\), preserves the $spin^c$ structure and induces a $spin^c$ structure on $Q^5_\epsilon(d)=M^5_\epsilon(d)/\tau$. The principal \(\ensuremath{\operatorname{U}}(1)\) bundles associated to those spin\(^c\) structures inherit flat connections from the connection on \(W_\ep^6(d)\). Let \(\alpha:\pi_1(Q_\ep^5(d))\to \ensuremath{\operatorname{U}}(1)\) be the nontrivial unitary representation.
\begin{lem}\label{Qeta}
Let \(g\) be an \(SO(3)\) invariant metric of positive scalar curvature on \(Q^5_0(d)\). Then \[\eta_\alpha(Q^5_0(d),g)=-\frac{d}{4}.\]
\end{lem}
\begin{proof}
The lemma follows by the proof of claim 6.1 in \cite{DGA21}. We give a short summary.
Consider the non-trivial flat line bundle $E_\alpha = M^5_\epsilon(d)\times_{\alpha} {\mathbb{C}}$ over $Q^5_\epsilon(d)$. One shows using Cheeger deformations, as described in \cite[Proposition 6.1.7]{We21}, \cite[Proposition 7.1.3]{We21} and \cite[\textsection 5 and pp.22-23]{DGA21}), that \(g\) lies in the same path component of \(\mathfrak{R}_{\text{scal}>0}(Q_0^5(d))\) as a metric \(\phi^*g_\ep,\) where \(\phi:Q_0^5(d)\to Q_\ep^5(d)\) is a diffeomorphism, and \(g_\ep\) is a metric of positive scalar curvature which lifts to a metric on \(M_\ep^5\) which in turn extends to a metric of nonnegtive scalar curvature on \(W_\ep^6(d).\) If we pull back the \(spin^c\) structure and associated connection with \(\phi,\) then \(\eta_\alpha(Q_0^5(d),g)=\eta_\alpha(Q_\ep^5(d),g_\ep)\). \(g_\ep\) satisfies the conditions such that the relative \(\eta\) invariant is given by \eqref{fixedpointformula}. \(\tau\) has fixed points \(p_1,...,p_d\in W_\ep^6(d),\) and the local contribution of each is \(a(p_i)=\ov{8}\) as computed via the Dolbeault operator in\cite[Proposition 3.5]{DGA21} and \cite[Proposition 2.3.13]{We21}, yielding
$$\eta_\alpha(Q^5_\epsilon(d),g_\epsilon)=-2\sum_{j=1}^da({p_j})=-\frac{d}{4}.$$
\end{proof}
It remains to compute the relative \(\eta\) invariant for \(X_{k,l}^5\) and \(\overline X_{k,l}^5.\) Let \(E^6_{k,l}\to B^4\) be the complex line bundle associated to the principal \(U(1)\) bundle \(N^5_{k,l}\to B^4\). Let \(W^6_{k,l}\subset E^6_{k,l}\) be the disc bundle, so \(\partial W_{k,l}=N_{k,l}\). The tangent bundle of \(W_{k,l}\) is the pullback of \(TB\oplus E_{k,l}\). Since \(w_2(TB)=c_{k,l}\) mod 2, \(W_{k,l}\) is spin. Since \(W_{k,l}\) is simply connected, the spin structure is unique. The action of \(-1\in S^1\) on \(N_{k,l}\) extends to an involution \(\tau\) on \(W_{k,l}\) which acts as \(-1\) on each \(D^2\) fiber. Since the spin structure is unique, the derivative \(d\tau\) acting on the frame bunlde \(P_{\ensuremath{\operatorname{SO}}}\) of \(W_{k,l}\) lifts to an isomorphism \(\tilde d\tau\) of the Spin bundle \(P_\text{Spin}\). Since the fixed point set of \(\tau\), which is the zero section of \(W_{k,l}\to B\), has codiminseion two, \(\tilde d\tau^2\) must be the fiberwise action of \(-1\) (see for instance \cite[Proposition 8.46]{AtBo68}).
Let \(P_{U(1)}=W_{k,l}\times U(1)\) be the trivial principal \(U(1)\) bundle over \(W_{k,l}\) equipped with the trivial connection. Define an isomorphism
\(\tau'\) of \(P_{U(1)}\), covering \(\tau\), by \(\tau'(x,z)=(\tau(x),-z)\). Then \(P_{\text{Spin}^c}=P_\text{Spin}\times_{\f{Z}_2}P_{\ensuremath{\operatorname{U}}(1)}\to P_{\ensuremath{\operatorname{SO}}}\times P_{U(1)}\) is a \(spin^c\) structure for \(W_{k,l}\) with associated bundle \(P_{U(1)}\). Here \(\f{Z}_2\) acts on \(P_{\text{Spin}}\times P_{\ensuremath{\operatorname{U}}(1)}\) diagonally by the fiberwise action of \(\pm1\). Furthermore, we can define an involution \(\tau''\) on \(P_{\text{Spin}^c}\):
\[\tau''([p,x,z])=[\tilde d\tau (p),\tau(x),iz]\] which covers the involution \(d\tau\times \tau'\) on \(P_{\ensuremath{\operatorname{SO}}}\times P_{U(1)}\). One checks that \(\tau''^2=\)id, so \(\tau''\) generates a \(\f{Z}_2\) action on \(P_{\text{Spin}^c}\) which covers the \(\f{Z}_2\) action on \(W_{k,l}\) generated by \(\tau\).
\(P_{\text{Spin}^c}\) restricts to a \(spin^c\) structure on \(N_{k,l}\) on which \(\f{Z}_2\) acts freely, and therefore induces a \(spin^c\) structure on \(X_{k,l}.\) Each \(spin^c\) structure is equipped with a flat connection on the associated complex line bundle. Identically, the trivial connection on \(P_{U(1)}\) induces the same on the restriction of \(P_{\ensuremath{\operatorname{U}}(1)}\) to \(N_{k,l},\) which in turn induces a flat connection on the quotient by \(\f{Z}_2,\) which is the line bundle over \(X_{k,l}\) associated to the nontrivial representation \(\alpha:\pi_1(X_{k,l})=\f{Z}_2\to U(1).\) Using the \(spin^c\) structure and connections defined in this way we have
\begin{lem}\label{etavalues}
Let \(g\) be an \(U(1)\) invariant metric of positive scalar curvature on \(X_{k,l}\). Let \(\overline g\) be an \(S^1\) invariant metric of positive scalar curvature on \(\overline X_{k,l}\). Then
\[\eta_\alpha(X_{k,l},g)=\pm\ov{8}\left(l^2+2kl\right)\]
\[\eta_\alpha(\overline X_{k,l},\overline{g})=\pm\ov{8}\left(2+l^2+2kl+2k^2\right)\]
\end{lem}
\begin{proof}
We give the argument for $X_{k,l}$, the reasoning being identical for $\overline{X}_{k,l}$. See also \cite[Proposition 7.1.5]{We21} for more details.
The metric \(g\) lifts to an \(U(1)\)-invariant metric \(\tilde g\) on \(N_{k,l}\). Equip $N_{k,l}\times D^2$ with the product of $\tilde g$ and a \(U(1)\)-invariant metric of nonnegative curvature on \(D^2\) which is product like near the boundary. Give $W_{k,l}$ the metric which turns the projection $N_{k,l}\times D^2\rightarrow N_{k,l}\times_{S^1} D^2=W_{k,l}$ into a Riemannian submersion. This metric is of product form near the boundary, has $scal\geq 0$ everywhere and $scal>0$ on the boundary. The metric induced on the boundary of \(W_{k,l}\) can be obtained from \(\tilde g\) by shrinking the \(U(1)\) fibers, and induces a metric on \(X_{k,l}\) which lies in the same component of \(\mathfrak{R}_{\text{scal}>0}(X_{k,l})\) as \(g\).
Thus by \eqref{fixedpointformula}, the relative $\eta$ invariant is given by
$$\eta_\alpha(X_{k,l},g)=-2a(B^4),$$ where $B^4$ represents the fixed point set of the action of $\tau$ on $W_{k,l}$, the zero section of the disk bundle \(W_{k,l}^6\to B^4\). The local contribution is given by\footnote{See for example \cite[Theorem 2.3.11]{We21}.}
$$a(B^4)=\pm i \int_{B^4}e^{\iota^*(c)/2}\hat{A}_\pi(L_{k,l})\hat{A}(B^4),$$ where $\iota$ is the inclusion of the zero section into the disk bundle, $c$ is the canonical class of the $spin^c$ structure on $W_{k,l}$, $\hat{A}(B^4)$ is the usual Hirzebruch $\hat{A}$ genus and
$$\hat{A}_\pi(L_{k,l})=\frac{1}{2i}\frac{1}{\cosh(c_{k,l}/2)}.$$ Recall that $c_{k,l}=-lu+kv$ by Lemma \ref{invariants}. Evaluation of the above integral amounts to determining the coefficient of $u^2$ in the above polynomial, using the relations $uv=-u^2$ and $v^2$ from Lemma \ref{invariants}. Note that \(c=0\) since the corresponding connection is flat. Thus, we obtain $a(B^4)=\pm \frac{1}{16}\left(-l^2-2kl\right)$ and the result follows.
\end{proof}
We can now prove \tref{thm}. Let \(M\) satisfy the hypotheses of \tref{thm}. If \(\pi_1(M)\) acts nontrivialy on \(\pi_2(M)\), then \(M\) is diffeomorphic to \(Q_0^5(d+16k)\) for all \(k\in\f{N}_0\) by \lref{Qdiff}. As discussed at the end of \sref{manifolds}, \(Q_0^5(d+16k)\) admits a \(U(1)\times SO(3)\) invariant metric \(g_k\) of nonnegative sectional and positive scalar curvature to which we can apply \lref{Qeta}. Since \(\eta_\alpha(Q_0^5(d+16k),g_k)=-\frac{d}{4}-4k\) takes infinitely many values, \(\modse{sec}(M)\) and \(\mods{Ric}(M)\) have infinitely many components by \lref{modulispaces}. If \(\pi_1(M)\) acts trivially on \(\pi_2(M)\), by \lref{infinitelists}, \(M\) is diffeomorphic to either \(X_{k,2}^5\) or \(\overline X_{k,4}^5\) for infinitely many values of \(k.\) In either case, each manifold admits a \(U(1)\) invariant metric of nonnegative sectional and positive scalar curvature inherited from \(S^3\times S^3\). By \lref{etavalues}, the relative $\eta$ invariant, computed for those metrics, is a nontrivial polynomial in \(k\) and again takes infinitely many values. \tref{thm} then follows from \lref{modulispaces}.
\printbibliography
\end{document} |
\section{Introduction}
Over the past two decades, several works have demonstrated that impedance matching between the receive antenna and front-end significantly impacts channel capacity in wireless channels\cite{domi,domi2,lau,wall,gans,gans2}. In order to implement capacity-optimal matching, the receiver must know the antenna impedance. However, this impedance depends on time-varying near-field loading conditions and often changes in an unpredictable manner. To mitigate such variation, researchers have proposed antenna impedance estimation techniques\cite{wu,wu2,wu2021,wu_PCA,wu3,ali,hass,moha,vasi}.
In previous works, impedance estimation at single-antenna receivers has been studied \cite{wu,wu2,wu_PCA, wu_cgr}. However, modern receivers are often equipped with multiple antennas for multiplexing and/or diversity benefits. Therefore, in this paper, we investigate the general and more important problem of estimating the antenna impedance matrix at MIMO receivers. Hassan and Wittneben investigated joint MIMO impedance and channel estimation using least squares\cite{hass}, and Wu solved a similar problem using hybrid estimation\cite{wu2021}. However, it remains unclear if either of these aforementioned approaches leads to the optimal MIMO impedance estimator. We fill in this gap in this paper.
This paper considers antenna impedance estimation algorithms using training data for multiple-input, multiple-output (MIMO) communication systems. We assume the receiver switches its impedance in a predetermined fashion during each training sequence. In i.i.d. Rayleigh fading channels, the maximum-likelihood (ML) estimator is derived for the impedance matrix as a function of the top block eigen-vector of the sample covariance matrix. This ML estimator is shown to be a method of moments (MM) when the fading channel is temporally correlated. Fundamental lower bounds, e.g., Cram\'er-Rao bounds (CRB), on these estimators are derived and important properties of these estimators, e.g., bias and efficiency, are explored through numerical simulations. The trade-off between channel and impedance estimation is demonstrated empirically.
The rest of the paper is organized as follows. We present our system model in Sec.~\ref{4secII}, derive a set of maximum-likelihood estimators for the MIMO antenna impedance and channel covariance matrix in Sec.~\ref{4secIII}, and derive method of moments (MM) estimators of these matrices under multiple packets scenarios and discuss ML estimators in Sec.~\ref{4secIV}. We then explore important properties of the estimators through numerical examples in Sec.~\ref{4secV}, and summarize our conclusions in Sec.~\ref{4secVI}.
\section{System Model}\label{4secII}
Consider a narrowband multiple-input multiple-output (MIMO) communications link with $M$ receive antennas and $N$ transmit antennas. The receiver model is illustrated in Fig.~\ref{fig_SystemModel_MIMO}. This circuit model is identical to the ones widely used to model a scenario, where amplifier noise dominates \cite{gans2,lau,wall}. This model is also a special case of the more general and complex models, which include additional noise sources, e.g., sky-noise and downstream noise \cite{domi,domi2,ivrl}.
\iflatex
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.6\textwidth, keepaspectratio=true]{./figs/SystemModel}
\end{center}
\vspace{-12pt}
\caption{Circuit model of a multiple-antenna receiver}
\vspace{-18pt}
\label{fig_SystemModel_MIMO}
\end{figure*}
\fi
In Fig.~\ref{fig_SystemModel_MIMO}, we model the antenna array by its Thevenin equivalent,
\beq\label{antenna_MIMO}
\bv \ = \ \bZ_\bA \bi + \bv_o \ ,
\eeq
where $\bv, \bi \in \mathbb{C}^M$ are the voltage across, and current into, the antenna array terminals. In particular, the antenna impedance is a symmetric matrix in $\mbbC^{M\times M}$ due to the reciprocity theorem of electromagnetics \cite{bala},
\beq\label{impd}
\bZ_\bA \ = \ \bR_\bA + j \bX_\bA \ ,
\eeq
where $\bR_\bA$ and $\bX_\bA$ are the resistance and reactance matrices, respectively. The incident electromagnetic field induces open-circuit voltage $\bv_o \in \mathbb{C}^M$ in \er{antenna_MIMO}. Under flat-fading conditions, the open-circuit voltage $\bv_o$ is modeled as \cite{domi2}
\beq\label{oc_volt}
\bv_o \ = \ \bG\bx \ ,
\eeq
where $\bx\in\mbbC^N$ is the transmitted symbol and $\bG\in\mbbC^{M\times N}$ is the matrix of fading path gains. Similar to the previous two papers, we consider a Rayleigh fading environment, where transmit antennas are sufficiently separated. Thus, columns of $\bG$ are modeled as i.i.d. zero-mean, complex Gaussian random vectors, $\bg_i\sim\mccn(\bzro, \bSig_\bg)$.
As shown in Fig.~\ref{fig_SystemModel_MIMO}, noisy voltage signal $\bv_L\in\mbbC^M$ across load impedance $\bZ_L\in\mbbC^{M\times M}$ is observed \cite{gans2,lau,wall},
\beq
\bv_L \ = \ \bZ_L\left({\bZ_\bA + \bZ_L}\right)^{-1}\bG\bx+\bn_L \ ,
\eeq
where the noise $\bn_L\in\mbbC^M$ is a zero-mean, circularly-symmetric, complex Gaussian random vector with covariance $E[\bn_L\bn_L^H]=\bSig_L$, which is hereafter denoted by $\bn_L\sim \mccn(\bzro,\bSig_L)$.
As mentioned in the previous paper, performance of estimators typically depend on the signal-to-noise ratio (SNR) in estimation theory, which is conventionally defined, for example, as $\Trace[\bv_L\bv_L^H\bSig_L^{-1}]$ \cite[Sec.~II-A]{domi2}.
In circuit theory, however, power depends on both voltage and current \cite[eq.~22]{ivrl}.
This estimation-theory SNR formula does not correctly predict the ratio of the physical signal power and noise power in the receiver front-end. For a given $\bv_o$, this ratio of physical signal power to noise power across the load is given by
\beq
\rho \ \define \ \frac{1}{\sigma_{n}^2}\Trace\left[\bR_L^{1/2}\left(\bZ_\bA + \bZ_1\right)^{-1} \bv_o\bv_o^H\left(\bZ_\bA + \bZ_1\right)^{-H}\bR_L^{1/2}\right] \ ,
\eeq
where $\sigma_n^2$ represents the noise power at the output of the amplifier and $\bR_L=\Real\{\bZ_L\}$ is the load resistance. As in the previous paper, we correct this discrepancy by defining $\bSig_L$ in a way that ensures the SNR and physical power ratio $\rho$ coincide:
\[
\bSig_L\ \define \sigma_n^2 \bZ_L\bR_L^{-1} \bZ_L^H\ .
\]
With this definition, it is convenient to redefine the observed signal as
\beq\label{sgl_model_MIMO}
\bw \ \define \ \bR_L^{1/2}\bZ_L^{-1}\bv_L \ = \ {\bR_L^{1/2}}\left({\bZ_\bA + \bZ_L}\right)^{-1}\bG\bx+\bn \ ,
\eeq
where $\bn\sim\mccn(\bzro,\sigma_n^2\bI_M)$ represents physical noise power referred to the amplifier output.
This signal model \eqref{sgl_model_MIMO} correctly connects the estimator performance to physical signal-to-noise power ratio. As mentioned before in the previous paper, this connection is essential to accurately predict the impact of impedance mismatch at MIMO receivers on important system-level metrics, such as channel capacity.
Suppose the channel gain matrix and antenna impedance matrix are unknown to the receiver. As in the previous two papers, our objective is to jointly estimate these two matrices using observations of known training sequences.
Suppose the transmitter sends a known training sequence of length $T$, i.e., $\bx_1, \dots, \bx_T \in\mbbC^N$ to the receiver, during which the receiver synchronously shifts its impedance as $\bZ_{L,1}, \dots, \bZ_{L,T}$. Also assume both the fading path gain $\bG$ and impedance $\bZ_A$ remain fixed during each transmission. The received observations take the following form,
\beq
\bw_t \ = \ {\bR_{L,t}^{1/2}}\left({\bZ_\bA + \bZ_{L,t}}\right)^{-1}\bG\bx_t+\bn_t \ ,
\eeq
where $t=1,2,\dots,T$ and the additive noises $\bn_t\sim\mccn(\bzro,\sigma_n^2\bI_M)$ are independent and identically distributed (i.i.d.).
We again assume the load impedance takes on two possible matrices,
\beq\label{imp_shift_MIMO}
\bZ_L \ = \ \begin{cases}
\bZ_1 \ ,& 1\leq t\leq K \ ,\\
\bZ_2 \ , & K\leq t \leq T \ .
\end{cases}
\eeq
Similar to the previous two papers, we assume $\bZ_L=\bZ_1$ is the load impedance used to receive the transmitted data, and is matched to our best estimate of $\bZ_A$; additionally $\bZ_L=\bZ_2$ is an impedance variation introduced in order to make $\bZ_A$ observable. To estimate $\bZ_A$, $\bZ_1\neq \bZ_2$ is required.
Note that in order the perform optimal detection of the transmitted symbols in \er{sgl_model_MIMO}, an accurate estimate of the entire matrix coefficient of $\bx$ is ideal, not simply the fading path gain matrix $\bG$. This motivates the definition of
an effective channel matrix that communication algorithms need,
\beq\label{H_MIMO}
\bH \ \define \ \bR_1^{1/2}\left(\bZ_\bA + \bZ_1\right)^{-1}\bG\in\mbbC^{M\times N}\ ,
\eeq
whose columns are also i.i.d. zero-mean, complex Gaussian, $\bh_i\sim\mccn(\bzro,\bSig_\bh)$, and
\beq\label{Sig_h_MIMO}
\bSig_\bh \ = \ \bR_1^{1/2}\left(\bZ_\bA + \bZ_1\right)^{-1} \bSig_\bg \left(\bZ_\bA + \bZ_1\right)^{-H}\bR_1^{1/2} \ .
\eeq
With this choice of load impedance in \er{imp_shift_MIMO} and definition of $\bH$ \er{H_MIMO}, we can express the observations in a simpler, bilinear form.
The voltage observation \er{sgl_model_MIMO} at the load is then,
\beq\label{4observations}
\bw_t \ = \ \begin{cases}
\bH \bx_t + \bn_{t} \ , & 1\leq t \leq K \\
\bF \bH \bx_t + \bn_{t} \ , & K+1\leq t \leq T
\end{cases}
\eeq
where $\bn_t\sim\mccn(\bzro,\sigma_n^2\bI_M)$ are independent over time $t$ and we define $\bF\in\mbbC^{M\times M}$ as a one-to-one mapping of $\bZ_\bA$ for mathematical convenience, conditioned on $\bZ_1\neq\bZ_2$,
\bea\label{F_MIMO}
\bF &=& \bR_2^{1/2}\left(\bZ_2 + \bZ_\bA\right)^{-1}\left(\bZ_1 + \bZ_\bA\right)\bR_1^{-1/2}\ .
\eea
We present this paper as a generalization of the previous paper to MIMO receivers. Here the goal is to derive maximum-likelihood (ML) estimators for $\bH, \bZ_A$ and $\bSig_\bh$ based on the observations \er{4observations}.
From the invariance principle of maximum-likelihood estimation (MLE) \cite[pg. 185]{kay}, knowing the MLE of $\bF$ is equivalent to knowing that of $\bZ_A$ and vice versa. Theoretically it suffices to derive estimators for $\bH, \bF$ and $\bSig_\bh$. Specifically, we follow the two-step procedure as described in the last paper: First, we consider joint maximum-likelihood estimation of $\bF$ and $\bSig_\bh$, treating $\bH$ as a nuisance parameter. Second, given estimates of $\bF$ and $\bSig_\bh$, we then estimate $\bH$ using minimum mean-squared error estimation. Again we focus exclusively on estimators for $\bF$ and $\bSig_\bh$ in the next two sections; estimators for $\bH$ will be explored through numerical examples in Sec.~\ref{4secV}.
\section{Maximum-Likelihood Estimators}\label{4secIII}
In this section, we derive maximum-likelihood (ML) estimators for $\bF$ and $\bSig_\bh$ based on observations in \er{4observations}. It is often convenient to find sufficient statistics before deriving the ML estimators.
We write \eqref{4observations} in matrix form, after defining $\bX_1\define[\bx_1,\bx_2,\dots,\bx_K]\in\mbbC^{N\times K}$ and $\bX_2\define[\bx_{K+1},\bx_{K+2},\dots,\bx_T]\in\mbbC^{N\times (T-K)}$,
\bea\label{Ws}
\bW_1&=& \bH \bX_1 + \bN_1 \ , ~~~~
\bW_2 \ = \ \bF \bH \bX_2 + \bN_2 \ ,
\eea
where $\bN_1$ and $\bN_2$ are analogously defined, independent and have i.i.d. entries $\mccn(0,\sigma_n^2)$.
The known training sequences for MIMO channel estimation are often equal-energy and orthogonal. We further assume $K=T/2$ and equal-energy and orthogonal training over the first and last $K$ symbols i.e.,
\beq\label{Xs_def}
\bX_1\bX_1^H \ = \ \bX_2\bX_2^H \ = \ \frac{PK}{N}\bI_N \ ,
\eeq
This can be achieved by using a normalized discrete Fourier transform (DFT) matrix, e.g., \cite[eq. 10]{bigu}.
We present a sufficient statistic in the next lemma.
\begin{lemma}[Sufficient Statistic]\label{4lem_ss}
Consider the observations $\bW_1$ and $\bW_2$ defined in \er{Ws} and known training sequences in \er{Xs_def}. Then
\beq\label{Ys_MIMO}
\bY_1 \ =\ \left(\frac{2N}{PT}\right) \bW_1\bX_1^H\ , ~~~~ \bY_2 \ = \ \left(\frac{2N}{PT}\right)\bW_2\bX_2^H\ ,
\eeq
are sufficient for estimating unknown matrices $\bF$ and $\bSig_\bh$. Moreover, $\bY_1 - \bH$ and $\bY_2 -\bF\bH$ are independent random matrices with i.i.d. $\mccn(0,\sigma^2)$ entries, where $\sigma^2\define 2N\sigma_n^2/PT$.
$\hfill\diamond$
\end{lemma}
\begin{proof}
From \er{Ws} and \er{Xs_def}, we have $\bY_1 = \bH + \left(\frac{2N}{PT}\right)\bN_1\bX_1^H$. To show the entries of the last matrix are i.i.d., we vectorize it,
\beq
\left(\frac{2N}{PT}\right) \vc \left(\bN_1\bX_1^H\right) \ = \ \left(\frac{2N}{PT}\right) \left(\bX_1^*\otimes \bI_M\right) \vc \bN_1 \in\mbbC^{MN} \ ,
\eeq
which is zero-mean and has covariance matrix $\left(\frac{2N}{PT}\right)^2 \left(\bX_1^*\bX_1^T\otimes \bI_M\right) \sigma_n^2\bI_{MN} =\sigma^2\bI_{MN}$, where by definition $\sigma^2= 2N\sigma_n^2/PT$. Note an identity of Kronecker product
$\vc \left(\bA\bB\bC\right) = \left(\bC^T\otimes\bA\right)\vc \bB$ is used \cite{brew}. This shows that $\bY_1 - \bH$ is a random matrices with i.i.d. $\mccn(0,\sigma^2)$ entries. Similarly, $\bY_2 -\bF\bH = \left(\frac{2N}{PT}\right)\bN_2\bX_2^H$ is also a random matrices with i.i.d. $\mccn(0,\sigma^2)$ entries. The independence between these two matrices follow from that noises are independent over time \er{Ws}.
From the Neyman-Fisher theorem \cite[pg. 117]{kay}, to prove sufficiency of \er{Ys_MIMO} it suffices to show that $p\left(\bW_1,\bW_2;\bF,\bSig_\bh\right)$ factors into a product $g\left(\bY_1,\bY_2, \bF,\bSig_\bh\right) f\left(\bW_1,\bW_2\right)$, where $f$ does not depend on $\bY_1,\bY_2, \bF,\bSig_\bh$ and $g$ does not depend on $\bW_1,\bW_2$.
We prove this using the conditional pdf
\beq
p\left(\bW_1,\bW_2;\bF,\bSig_\bh\right) \ = \ E_\bH\left[p\left(\bW_1,\bW_2|\bH;\bF,\bSig_\bh\right)\right]\ ,
\eeq
where the expectation $E_\bH[\cdot]$ is with respect to $\bH$ \er{H_MIMO}. Since $\bW_1$ and $\bW_2$ are conditionally independent given $\bH$, we have
\bea\label{4lem_ss_pdf}
&&(\pi\sigma_n^2)^{NT} p\left(\bW_1,\bW_2;\bF,\bSig_\bh\right) \nn\\
&=& E_\bH\left[\exp\left(-\frac{1}{\sigma_n^2}\left\lVert\bW_1-\bH\bX_1\right\rVert^2-\frac{1}{\sigma_n^2}\left\lVert\bW_2-\bF\bH\bX_2\right\rVert^2\right)\right]\nn\\
&=& E_\bH\left[\exp\left(\frac{2\Real\Trace\left[\bH^H\bW_1\bX_1^H\right]}{\sigma_n^2}-\frac{\Trace\left[\bH^H\bH\bX_1\bX_1^H\right]}{\sigma_n^2}+ \frac{2\Real\Trace\left[\bH^H\bF^H\bW_2\bX_2^H\right]}{\sigma_n^2}\right. \right.\nn\\
&& \left. \left. -\frac{\Trace\left[\bH^H\bF^H\bF\bH\bX_2\bX_2^H\right]}{\sigma_n^2}\right)\right] \exp\left(-\frac{1}{\sigma_n^2}\left\lVert\bW_1\right\rVert^2-\frac{1}{\sigma_n^2}\left\lVert\bW_2\right\rVert^2\right) \nn\\
&=&E_\bH\left[\exp\left(\frac{2\Real\Trace\left[\bH^H\bY_1+\bH^H\bF^H\bY_2\right]}{\sigma^2}-\frac{\Trace\left[\bH^H\bH+\bH^H\bF^H\bF\bH\right]}{\sigma^2}\right) \right]\nn\\
&&\exp\left(-\frac{1}{\sigma_n^2}\left\lVert\bW_1\right\rVert^2-\frac{1}{\sigma_n^2}\left\lVert\bW_2\right\rVert^2\right)\ ,
\eea
where $\lVert\bA \rVert^2=\Trace[\bA^H\bA]$ denotes the Frobenius norm. Also, the third equality follows from the identities $2\Real\Trace[\bA] = \Trace[\bA] + \Trace[\bA^H]$ and $\Trace[\bA\bB] = \Trace[\bB\bA]$, and the fourth equality follows from \er{Xs_def} and the definition of $\sigma^2$.
In \er{4lem_ss_pdf}, denote the first factor by $(\pi\sigma_n^2)^{NT}g\left(\bY_1,\bY_2, \bF,\bSig_\bh\right)$ and the second by $ f\left(\bW_1,\bW_2\right)$. Note $g$ depends on
$\bY_1,\bY_2, \bF$ and $\bSig_\bh$ (through the expectation) but not on $\bW_1,\bW_2$. And $f$ depends on $\bW_1,\bW_2$ only, not $\bY_1,\bY_2, \bF,\bSig_\bh$. This completes the proof.
\end{proof}
Based on the sufficient statistics in \er{Ys_MIMO}, we want to estimate the following complex parameters
\beq\label{theta_MIMO}
\btt \ \define \vc \begin{bmatrix}
\bF & \bSig_\bh
\end{bmatrix} \ ,
\eeq
where $\bF$ is defined in \er{F_MIMO} and $\bSig_\bh$ in \er{Sig_h_MIMO}. Here we present the maximum-likelihood (ML) estimator, such that
\beq
\hat{\btt}_{ML} \ \define \ \arg\max_{\btt} p\left(\bY_1,\bY_2;\btt\right) \ .
\eeq
The next theorem shows the ML estimator can be calculated via block eigen-decomposition using the sufficient statistic given in \er{Ys_MIMO}.
\begin{theorem}[Single-Packet ML Estimators]\label{MLE_ch4}
Let $\bY_1$ and $\bY_2$ be the sufficient statistics in \er{Ys_MIMO}. Suppose $\bF$ and $\bSig_\bh$ are unknown. Consider the sample covariance matrix,
\beq\label{S_MIMO}
\bS \ \define \ \frac{1}{N} \begin{bmatrix}
\bY_1\bY_1^H & \bY_1\bY_2^H\\
\bY_2\bY_1^H & \bY_2\bY_2^H
\end{bmatrix} \in\mbbC^{2M\times 2M}\ .
\eeq
The eigen-decomposition of $\bS$ can be written as
\beq
\bS \bU_\bs \ = \ \bU_\bs \diag(\mu_1,\dots,\mu_{2M}) \ ,
\eeq
where $\diag(\cdot)$ denotes a square diagonal matrix with its input as diagonal entries, and the eigen-values $\mu_k\geq 0$ are in descending order. Define the unitary eigen-vector matrix $\bU_\bs$ as a 2 by 2 block matrix, i.e.,
\beq
\bU_\bs \ \define \ \begin{bmatrix}
\bU_{\bs11} & \bU_{\bs12} \\
\bU_{\bs21} & \bU_{\bs22}
\end{bmatrix} \ ,
\eeq
where $\bU_{\bs ij}\in\mbbC^{M\times M}$ and $i,j=1,2$.
Then, the maximum-likelihood estimate of $\btt$ is,
\beq
\hat{\btt}_{ML} \define \vc \left[\hat{\bF}_{ML} ~~ \hat{\bSig_\bh}\right] \ ,
\eeq
where $\sigma^2\define 2N\sigma_n^2/PT$ and, conditioned on $\bU_{\bs11}$ is non-singular,
\beq
\hat{\bF}_{ML} \ = \ \bU_{\bs21} \bU_{\bs11}^{-1} \ , ~~~~ \hat{\bSig_\bh} \ = \ \bU_{\bs11}\left(\diag(\mu_1,\dots,\mu_M) - \sigma^2\bI_M\right)^+ \bU_{\bs11}^H \ .
\eeq
Here $(\cdot)^+$ is an element-wise operator on real matrices, such that $[(\bA)^+]_{ij}\define \max\{[\bA]_{ij},0\}$.
\end{theorem}
\begin{proof}
Consider the sufficient statistic in \eqref{Ys_MIMO}, and define
\beq\label{V_MIMO}
\bV \ \define \ \begin{bmatrix}
\bY_1\\
\bY_2
\end{bmatrix} \ = \ \begin{bmatrix}
\bH + \bN_1\\
\bF\bH + \bN_2
\end{bmatrix}\ = \ \begin{bmatrix}
\bh_1 + \bn_{1,1}& \cdots &\bh_N + \bn_{1,N}\\
\bF\bh_1 + \bn_{2,1}& \cdots & \bF\bh_N + \bn_{2,N}
\end{bmatrix} \ .
\eeq
Due to uncoupled transmit antennas, the $N$ columns of $\bH$ are also i.i.d. zero-mean, circularly-symmetric, complex Gaussian random vectors, i.e., $\bh_k \sim\mccn\left(\bzro_M,\bSig_\bh\right)$ for all $1\leq k\leq N$, where $\bSig_\bh\define E\left[\bh_k\bh_k^H\right]$. Thus, the prior information for channel $\bH$ can be written as,
\beq\label{pdf_H_MIMO}
p(\bH) = \frac{1}{\det\left(\pi \bI_N\otimes\bSig_\bh\right)} \exp\left[-\left(\vc \bH\right)^H\left( \bI_N\otimes\bSig_\bh^{-1}\right)\vc \bH\right] \ .
\eeq
The covariance matrix of $\bV$ is defined as,
\bea\label{CovV}
\bSig & \define & \frac{1}{N} E[\bV\bV^H] \ = \ \begin{bmatrix}
\bSig_\bh + \sigma^2\bI_M& \bSig_\bh \bF^H\\
\bF\bSig_\bh & \bF\bSig_\bh\bF^H + \sigma^2\bI_M
\end{bmatrix} \ .
\eea
Since $\bSig$ is Hermitian, it decomposes into the following block eigen-system \cite{pere},
\bea\label{block_eigen}
\bSig &=& \begin{bmatrix}
\bI_M\\
\bF
\end{bmatrix} \bSig_\bh \begin{bmatrix}
\bI_M&\bF^H
\end{bmatrix} + \sigma^2\bI_{2M} \ = \ \bB_1\bD_1\bB_1^H + \bB_2\bD_2\bB_2^H \ ,
\eea
where $\bD_i\in\mbbC^{M\times M}$ and $\bB_i\in\mbbC^{2M\times M}$ (for $i=1,2$) are the block eigenvalues and orthonormal block eigenvectors of $\bSig$,
\bea
\bB_1 &=&
\begin{bmatrix}
\bI_M\\
\bF
\end{bmatrix} \bA_1^{-\frac{1}{2}} \ , ~~~ \bB_2 \ = \ \begin{bmatrix}
-\bF^H\\
\bI_M
\end{bmatrix} \bA_2^{-\frac{1}{2}} \ , \nn\\
\bD_1 &=& \bA_1^{\frac{1}{2}}\bSig_\bh\bA_1^{\frac{1}{2}} + \sigma^2\bI_M \ , ~~~ \bD_2 \ = \ \sigma^2\bI_M \ ,
\eea
and $\bA_1$ and $\bA_2$ are Hermitian matrices defined as,
\beq\label{As}
\bA_1 \ \define \ \bF^H\bF + \bI_M \ , ~~~ \bA_2 \ \define \ \bF\bF^H + \bI_M \ .
\eeq
Note both $\bA_1$ and $\bA_2$ are positive definite; also $\bF^H\bA_2=\bA_1\bF^H$ and $\bF\bA_1=\bA_2\bF$.
Since eigenvalues of $\bD_i$ are also eigenvalues of $\bSig$ \cite[Th. 1.1]{pere}, and observe from \eqref{block_eigen} that $\bD_2$ is already diagonal, we write down the (scalar) eigenvalue decomposition of $\bSig$,
\beq\label{Sig_eig}
\bSig \ = \ \bU \bLam\bU^H \ \define \ \begin{bmatrix}
\bB_1 \bQ & \bB_2
\end{bmatrix} \begin{bmatrix}
\bLam_1 & \bzro_{M\times M} \\
\bzro_{M\times M} & \bLam_2
\end{bmatrix} \begin{bmatrix}
\bQ^H\bB_1^H\\
\bB_2^H
\end{bmatrix}\ ,
\eeq
where $\bD_1\bQ = \bQ\bLam_1$, $\bLam_1=\diag(\lambda_1,\dots,\lambda_M)$ is diagonal, $\bQ\in\mbbC^{M\times M}$ is unitary, and $\bLam_2 = \sigma^2\bI_M$. Note $\lambda_j \geq \sigma^2$ for all $1\leq j \leq M$. In other words, the covariance matrix $\bSig$ has $2M$ real eigenvalues, where the largest $M$ eigenvalues are greater than or equal to $\sigma^2$, while the smallest $M$ eigenvalues are exactly $\sigma^2$.
Under the assumptions above, we assume $\sigma^2$ is known but $\bLam_1$ and $\bU$ are unknown. If we find the joint ML estimator for $\bLam_1$ and $\bU$, the ML estimators for $\btt$ \eqref{theta_MIMO} follows by the invariance principle of MLE \cite[Th. 7.4, pg. 185]{kay}. From \eqref{Sig_eig}, we firstly find the inverse of covariance $\bSig$ using it scalar eigen system,
\beq
\bSig^{-1} \ = \ \bU\bLam^{-1}\bU^H \ = \ \begin{bmatrix}
\bB_1 \bQ & \bB_2
\end{bmatrix} \begin{bmatrix}
\bLam_1^{-1} & \bzro_{M\times M} \\
\bzro_{M\times M} & \bLam_2^{-1}
\end{bmatrix} \begin{bmatrix}
\bQ^H\bB_1^H\\
\bB_2^H
\end{bmatrix} \ .
\eeq
We then write the density function of $\bV\define[\bv_1,\dots,\bv_N]$ in \eqref{V_MIMO} as,
\bea\label{likelihood}
&&p(\bV; \btt) \ = \ \det(\pi \bSig)^{-N} \exp \left(-\sum_{i=1}^{N}\bv_i^H\bSig^{-1}\bv_i\right) \nn\\
& = & \det(\pi \bSig)^{-N} \exp \left(-N\Trace[\bS\bSig^{-1}]\right) \nn\\
&= & \left(\pi \prod_{j=1}^{2M} \lambda_j\right)^{-N} \exp\left(-N \sum_{j=1}^{2M} \frac{\bu_j^H\bS\bu_j}{\lambda_j}\right) \nn\\
&=& \left(\pi \prod_{j=1}^{2M} \lambda_j\right)^{-N} \exp \left[N \sum_{j=1}^{M}\left(\frac{1}{\sigma^2}-\frac{1}{\lambda_j}\right)\bu_j^H\bS\bu_j - \frac{N\Trace(\bS)}{\sigma^2}\right] \ ,
\eea
where, in the second equality, sample covariance matrix defined in \er{S_MIMO} can be equivalently written in terms of $\bV=[\bv_1,\dots,\bv_N]$,
\beq
\bS \ = \ \frac{1}{N} \bV\bV^H \ = \ \frac{1}{N} \sum_{i=1}^{N} \bv_i\bv_i^H\ ,
\eeq
and in the third equality we define $\bU\define[\bu_1,\dots,\bu_N]$.
Since we have $\lambda_j\geq \sigma^2$ for $1\leq j \leq M$, the coefficients $\frac{1}{\sigma^2}-\frac{1}{\lambda_j}$ are non-negative and in decreasing order. The likelihood function is maximized, when
$\bu_j$ is chosen to be the eigenvector corresponding to the $j$th largest eigenvalue of the sample covariance $\bS$. Assume the eigenvalues of $\bS$ are ordered in descending order, $[\mu_1,\dots,\mu_{2M}]$, the maximum of \eqref{likelihood} is,
\bea\label{likelihood_max}
\max_{\lambda_j, 1\leq j\leq M} p(\bV) & = & \left(\pi \prod_{j=1}^{2M} \lambda_j\right)^{-N} \exp \left(-N\sum_{j=1}^{M}\frac{\mu_j}{\lambda_j} - \frac{N}{\sigma^2}\sum_{k=M+1}^{2M}\mu_k\right) \nn\\
&\leq & \left(\pi \sigma^{2M}\prod_{j=1}^{M} \mu_j\right)^{-N} \exp\left(-NM- \frac{N}{\sigma^2}\sum_{k=M+1}^{2M}\mu_k\right)\ ,
\eea
where the last step is because $xe^{-x}$ is uniquely maximized at $x=1$, and equality is achieved by letting
\beq
\hat{\lambda}_j = \max \{\mu_j, \sigma^2 \}\ , ~~~1\leq j \leq M \ .
\eeq
Now it remains to show that \eqref{likelihood} can be achieved by actual estimates of $\bF$ and $\bSig_\bh$. Assume that the $\bS\bU_\bs = \bU_\bs\diag(\mu_1,\dots,\mu_{2M})$, where
\beq
\bU_\bs \ \define \ \begin{bmatrix}
\bU_{\bs11}& \bU_{\bs12}\\
\bU_{\bs21}&\bU_{\bs22}
\end{bmatrix} \ .
\eeq
Then we verify that the following estimates achieve the maximum in \eqref{likelihood},
\beq
\hat{\bSig} \ = \ \bU_\bs \diag(\hat{\lambda}_1,\dots,\hat{\lambda}_M,\sigma^2,\dots,\sigma^2)\bU_\bs^H \ .
\eeq
Thus, the joint ML estimators for $\bF$ and $\bSig_\bh$ are, conditioned on $\bU_{\bs11}$ is non-singular,
\bea\label{mle_iid}
\hat{\bF}_{ML} &=& \bU_{\bs21} \bU_{\bs11}^{-1} \ , ~~~~ \hat{\bSig_\bh} \ = \ \bU_{\bs11}\left(\diag(\mu_1,\dots,\mu_M) - \sigma^2\bI_M\right)^+ \bU_{\bs11}^H \ .
\eea
The derivation for $\hat{\bF}_{ML}=\bU_{\bs21} \bU_{\bs11}^{-1}$ is straightforward from \er{block_eigen}, as $\bF$ only depends on the eigen-vector matrix and not the eigen-values.
The derivation of $\hat{\bSig_\bh}$ follows also \er{block_eigen},
\bea
&&\hat{\bSig} -\sigma^2\bI_{2M}\ = \ \begin{bmatrix}
\bI_M\\
\hat{\bF}
\end{bmatrix} \hat{\bSig_\bh} \begin{bmatrix}
\bI_M&\hat{\bF}^H
\end{bmatrix} \nn\\
&&\begin{bmatrix}
\bI_M&\hat{\bF}^H
\end{bmatrix}\left(\hat{\bSig} -\sigma^2\bI_{2M}\right)\begin{bmatrix}
\bI_M\\
\hat{\bF}
\end{bmatrix} \ = \ \left(\bI_M + \hat{\bF}^H\hat{\bF}\right)\hat{\bSig_\bh} \left(\bI_M + \hat{\bF}^H\hat{\bF}\right) \nn\\
\hat{\bSig_\bh} & = & \left(\bI_M + \hat{\bF}^H\hat{\bF}\right)^{-1}\begin{bmatrix}
\bI_M&\hat{\bF}^H
\end{bmatrix}\left(\hat{\bSig} -\sigma^2\bI_{2M}\right)\begin{bmatrix}
\bI_M\\
\hat{\bF}
\end{bmatrix}\left(\bI_M + \hat{\bF}^H\hat{\bF}\right)^{-1} \ ,
\eea
which reduces to \eqref{mle_iid} after plugging in $\hat{\bF}_{ML}$'s formula and some simplification. This completes the proof.
\end{proof}
In order to study the efficiency of the joint ML estimators for $\bF$ and $\bSig_\bh$, a natural next step is to find tight fundamental lower bounds on these estimators. Two fundamental lower bounds are investigated next.
It has been shown that the (complex) Fisher information matrix (FIM) on $\btt$ is given by \cite[eq.~37]{wu3}, which extended results on real parameters to complex ones \cite[eq. 15.52]{kay},
\bea\label{FIM_MIMO}
&&[\boldsymbol{\mathcal{I}}(\btt)]_{ij} \ = \ N\Trace\left[\bSig^{-1}\frac{\partial \bSig}{\partial \theta_i^*}\bSig^{-1}\frac{\partial \bSig}{\partial \theta_j}\right] \ ,
\eea
where $\bSig$ is given in \er{block_eigen}. The error covariance matrix $\bC_{\hat{\btt}}$ of any unbiased estimator $\hat{\btt}$ is lower bounded by the Cram\'er-Rao bound (CRB), i.e., the inverse of $\boldsymbol{\mathcal{I}}(\btt)$,
\beq\label{crb_ch4}
\bC_{\hat{\btt}} \ \define \ E_{\bY_1,\bY_2;\btt}\left[\left(\hat{\btt}-\btt\right)\left(\hat{\btt}-\btt\right)^H\right] \ \geq \ \boldsymbol{\mathcal{I}}(\btt)^{-1}\ ,
\eeq
where the expectation is with respect to pdf in \er{likelihood} and $\bA\geq \bB$ means $\bA-\bB$ is positive semi-definite.
The formula in \er{crb_ch4} is unlikely to simplify further without additional assumptions. Hence we evaluate the CRB numerically in simulations.
Another useful lower bound is the Miller-Chang bound (MCB) \cite{mill}. The formula for this bound is given below, but details of its derivation are given in the Appendix of \cite{wu3}. It can be shown that square of Frobenius norm of any unbiased estimator for $\bF$, for all $\bH\in\mbbC^{M\times N}$ \eqref{H_MIMO}, is lower bounded by the MCB,
\bea\label{MCB}
\mathcal{M}(\bF) & = & \sigma^2\Trace \left(E\left[\left(\bH^*\bH^T\right)^{-1}\right]\right) \Trace \left(\bF\bF^H + \bI_M\right) \nn\\
&=&\begin{cases}
\infty \ , & N \leq M \ ,\\
\frac{\sigma^2}{N-M}\Trace\left(\bSig_\bh^{-1}\right) \Trace \left(\bF\bF^H + \bI_M\right) \ , & N > M \ ,
\end{cases}
\eea
where $N>M$ means the number of antennas at the transmitter is greater than that at the receiver, and the final expression follows from that $\bH^*\bH^T$ is a complex $M\times M$ Wishart matrix of degree $N$ \eqref{pdf_H_MIMO} and the mean of its inverse is derived by Maiwald and Kraus \cite[eq. 39]{maiw}. Note the independence between columns of $\bH$ is essential. If $N\leq M$, the inverse mean of $\bH^*\bH^T$ is unbounded, then one packet is likely insufficient for any unbiased estimator of $\bF$ to have finite error in Frobenius norm. This motivates finding estimators when observations from multiple packets are available.
\section{Estimators for Multiple Packets}\label{4secIV}
In the last section, we derived the maximum-likelihood estimators for $\bF$ and $\bSig_\bh$ using observations of training sequences from one packet.
In this section, we consider estimators based on multiple packets, where channel varies from packet to packet.
As in the previous paper, suppose the transmitter sends $L$ identical training packets to the receiver. During transmission of each training packet, the receiver shifts it load impedance as described in \er{imp_shift_MIMO}. Similar to previous papers and last section, block fading is assumed, i.e., the channel remains constant within a packet but randomly varies from packet to packet. Similarly to \er{4observations} the signal observations at the $l$-th packet can be described as
\beq\label{4observations_mp}
\bw_{l, t} \ = \ \begin{cases}
\bH_l \bx_t + \bn_{l,t} \ , & 1\leq t \leq K \\
\bF \bH_l \bx_t + \bn_{l,t} \ , & K+1\leq t \leq T
\end{cases}
\eeq
where the random noise vectors $\bn_{l,t}\sim\mccn(\bzro,\sigma_n^2\bI_M)$ are independent over packets $1\leq l\leq L$ and time $1\leq t\leq T$. We can express above observations in a compact matrix form, with a slight abuse of notation\footnote{Ideally, we would use notations like $\bW_{mp,1}$, $\bH_{mp}$, $\bN_{mp,1}$ and etc to distinguish them from their single-packet counterparts. We hereafter drop the subscript $mp$ for simplicity when confusion is unlikely to occur.},
\beq\label{Ws_mp}
\bW_{1} \ \define \ \bH(\bI_L\otimes \bX_1) + \bN_{1} \ , ~~\bW_{2} \ \define \ \bF\bH(\bI_L\otimes \bX_2) + \bN_{2} \ ,
\eeq
where $\bW_{1}\in\mbbC^{M\times LK}$, $\bW_{2}\in\mbbC^{M\times L(T-K)}$, $\bX_1$ and $\bX_2$ are defined above \er{Ws}, $\bN_{1}$ and $\bN_{2}$ are independent random matrices each with i.i.d. entries $\mccn(0,\sigma_n^2)$, and we define the multi-packet channel as, again slightly abusing notation,
\beq\label{eq_Hmp_def}
\bH \ \define \ [\bH_1~ \cdots~\bH_L] \in\mbbC^{M\times NL} \ .
\eeq
Here $\bH_l\in\mbbC^{M\times N}$ is the channel matrix for the $l$-th packet, whose columns are spatially i.i.d. complex Gaussian $\mccn(\bzro, \bSig_\bh)$ across transmit antennas but temporally correlated across packets. If the normalized channel correlation is $\bC_\bh$, then the space-time correlation of $\bH$ is can be shown as
\beq\label{vec_Hmp}
\vc \bH \sim\mccn\left(\bzro_{MNL}, \bC_\bh\otimes\bI_N\otimes\bSig_\bh\right) \ .
\eeq
In this section, we assume $\bC_\bh$ is known.
Similar to the previous section, the goal of this section is to derive estimators for both $\bF$ and $\bSig_\bh$, or $\btt$ as defined in \er{theta_MIMO}, treating $\bH$ as a nuisance parameter. Then, we explore estimators for $\bH$ given $\bF$ and $\bSig_\bh$ through numerical examples. The following lemma generalizes Lemma \ref{4lem_ss} to multiple packets.
\begin{lemma}[Multi-Packet Sufficient Statistic]\label{4lem_ss_mp}
Consider the observations $\bW_{1}$ and $\bW_{2}$ defined in \er{Ws_mp} and known training sequences in \er{Xs_def}. Then
\beq\label{Ys_MIMO_mp}
\bY_{1} \ =\ \left(\frac{2N}{PT}\right) \bW_{1}\left(\bI_L\otimes\bX_1^H\right)\ , ~~\bY_2 \ = \ \left(\frac{2N}{PT}\right)\bW_{2}\left(\bI_L\otimes\bX_2^H\right)\ ,
\eeq
are sufficient for estimating unknown matrices $\bF$ and $\bSig_\bh$. Moreover, $\bY_{1} - \bH$ and $\bY_{2} -\bF\bH$ are independent random matrices with i.i.d. $\mccn(0,\sigma^2)$ entries, where $\sigma^2\define 2N\sigma_n^2/PT$.
$\hfill\diamond$
\end{lemma}
\begin{proof}
From the definition of $\bW_{1}$ in \er{Ws_mp} along with $\bX_1$ in \er{Xs_def}, we can readily express $\bY_{1}$ in \er{Ys_MIMO_mp} as
\beq
\bY_1 \ = \ \bH + \left(\frac{2N}{PT}\right)\bN_1\left(\bI_L\otimes\bX_1^H\right) \ .\nn
\eeq To show the entries of the last matrix are i.i.d., we vectorize it,
\beq
\left(\frac{2N}{PT}\right) \vc \left[\bN_1\cdot\left(\bI_L\otimes\bX_1^H\right)\right] \ = \ \left(\frac{2N}{PT}\right) \left[\left(\bI_L\otimes\bX_1^*\right)\otimes \bI_M\right]\vc \bN_1 \in\mbbC^{MNL} \ ,
\eeq
which is zero-mean and has covariance matrix
\beq
\left(\frac{2N}{PT}\right)^2 \left(\bI_L\otimes\bX_1^*\bX_1^T\otimes \bI_M\right) \sigma_n^2\bI_{MNL} \ =\ \sigma^2\bI_{MNL} \ .\nn
\eeq
Note Kronecker product is associative and
$\vc \left(\bA\bB\bC\right) = \left(\bC^T\otimes\bA\right)\vc \bB$ is used \cite{brew}. This shows that $\bY_1 - \bH$ is a random matrices with i.i.d. $\mccn(0,\sigma^2)$ entries. Similarly, we can show that
\beq
\bY_2 -\bF\bH \ = \ \left(\frac{2N}{PT}\right)\bN_2\left(\bI_L\otimes\bX_2^H\right) \ ,\nn
\eeq
which is also a random matrices with i.i.d. $\mccn(0,\sigma^2)$ entries. The independence between these two matrices follow from that noises are independent over time and across packets \er{4observations_mp}.
From the Neyman-Fisher theorem \cite[pg. 117]{kay}, to prove sufficiency of \er{Ys_MIMO_mp} it suffices to show that $p\left(\bW_1,\bW_2;\bF,\bSig_\bh\right)$ factors into a product $g\left(\bY_1,\bY_2, \bF,\bSig_\bh\right) f\left(\bW_1,\bW_2\right)$, where $f$ does not depend on $\bY_1,\bY_2, \bF,\bSig_\bh$ and $g$ does not depend on $\bW_1,\bW_2$.
We prove this using the conditional pdf
\beq
p\left(\bW_1,\bW_2;\bF,\bSig_\bh\right) \ = \ E_\bH\left[p\left(\bW_1,\bW_2|\bH;\bF,\bSig_\bh\right)\right]\ ,
\eeq
where the expectation $E_\bH[\cdot]$ is with respect to $\bH$ as defined in \eqref{eq_Hmp_def}. Since $\bW_1$ and $\bW_2$ are conditionally independent given $\bH$, we have
\bea\label{4lem_ss_mp_pdf}
&&(\pi\sigma_n^2)^{MLT} p\left(\bW_1,\bW_2;\bF,\bSig_\bh\right) \nn\\
&=& E_\bH\left[\exp\left(-\frac{1}{\sigma_n^2}\left\lVert\bW_1-\bH\left(\bI_L\otimes\bX_1\right)\right\rVert^2-\frac{1}{\sigma_n^2}\left\lVert\bW_2-\bF\bH\left(\bI_L\otimes\bX_2\right)\right\rVert^2\right)\right]\nn\\
&=& E_\bH\left[\exp\left(\frac{2\Real\Trace\left[\bH^H\bW_1\left(\bI_L\otimes\bX_1^H\right)\right]}{\sigma_n^2}-\frac{\Trace\left[\bH^H\bH\left(\bI_L\otimes\bX_1\bX_1^H\right)\right]}{\sigma_n^2}\right. \right.\nn\\
&& \left. \left. + \frac{2\Real\Trace\left[\bH^H\bF^H\bW_2\left(\bI_L\otimes\bX_2^H\right)\right]}{\sigma_n^2} -\frac{\Trace\left[\bH^H\bF^H\bF\bH\left(\bI_L\otimes\bX_2\bX_2^H\right)\right]}{\sigma_n^2}\right)\right]\nn\\
&&\exp\left(-\frac{1}{\sigma_n^2}\left\lVert\bW_1\right\rVert^2-\frac{1}{\sigma_n^2}\left\lVert\bW_2\right\rVert^2\right) \nn\\
&=&E_\bH\left[\exp\left(\frac{2\Real\Trace\left[\bH^H\bY_1+\bH^H\bF^H\bY_2\right]}{\sigma^2}-\frac{\Trace\left[\bH^H\bH+\bH^H\bF^H\bF\bH\right]}{\sigma^2}\right) \right]\nn\\
&&\exp\left(-\frac{1}{\sigma_n^2}\left\lVert\bW_1\right\rVert^2-\frac{1}{\sigma_n^2}\left\lVert\bW_2\right\rVert^2\right)\ ,
\eea
where $\lVert\bA \rVert^2=\Trace[\bA^H\bA]$ denotes the Frobenius norm. Also, the third equality follows from the identities $2\Real\Trace[\bA] = \Trace[\bA] + \Trace[\bA^H]$ and $\Trace[\bA\bB] = \Trace[\bB\bA]$, and the fourth equality follows from \er{Xs_def} and the definition of $\sigma^2$.
In the final expression of \er{4lem_ss_mp_pdf}, we denote the first factor by $(\pi\sigma_n^2)^{MLT}g\left(\bY_1,\bY_2, \bF,\bSig_\bh\right)$ and the second by $ f\left(\bW_1,\bW_2\right)$. Note through the expectation over $\bH$, $g$ only depends on
$\bY_1,\bY_2, \bF, \bC_\bh$ (which is assumed known), and $\bSig_\bh$ but not on $\bW_1,\bW_2$. And $f$ only depends on $\bW_1,\bW_2$, but not $\bY_1,\bY_2, \bF,\bSig_\bh$. Thus, the Neyman-Fisher theorem applies\cite[pg. 117]{kay}, and this completes the proof.
\end{proof}
Note $\bY_1$ and $\bY_2$ in \er{Ys_MIMO_mp} are a sufficient statistic regardless what the correlation matrix $\bC_\bh$ is. But $\bC_\bh$ will play a role in the PDF after the expectation over $\bH$.
As in the last section, our ultimate goal is to find the maximum-likelihood (ML) estimators for $\btt$,
\beq
\btt \ \define \vc \begin{bmatrix}
\bF & \bSig_\bh
\end{bmatrix} \ ,
\eeq
where $\bF$ is defined in \er{F_MIMO} and $\bSig_\bh$ in \er{Sig_h_MIMO}.
Using the multi-packet sufficient statistics in \er{Ys_MIMO_mp}, the multi-packet ML estimators for $\btt$ shall satisfy the following optimal criteria,
\beq\label{4mle_criteria_mp}
\hat{\btt}_{ML} \ \define \ \arg\max_{\btt} p\left(\bY_1,\bY_2;\btt\right) \ .
\eeq
However, as we learned from its prequel \cite{wu_PCA}, these ML estimators
are unlikely in closed-form in general. Thus, we defer discussion on the ML estimators but first seek another set of estimators, i.e., the \textit{method of moments} estimators \cite[Ch. 9]{kay}.
\begin{lemma}[Method of Moments Estimators]
Let $\bY_1$ and $\bY_2$ be the sufficient statistics in \er{Ys_MIMO_mp}. Suppose $\bF$ and $\bSig_\bh$ are unknown. Consider the sample covariance matrix,
\beq\label{S_mp_MIMO}
\bS_{mp} \ \define \ \frac{1}{NL} \begin{bmatrix}
\bY_1\bY_1^H & \bY_1\bY_2^H\\
\bY_2\bY_1^H & \bY_2\bY_2^H
\end{bmatrix} \in\mbbC^{2M\times 2M}\ .
\eeq
The eigen-decomposition of $\bS$ can be written as
\beq
\bS_{mp} \bU_\bs \ = \ \bU_\bs \diag(\mu_1,\dots,\mu_{2M}) \ ,
\eeq
where $\diag(\cdot)$ denotes a square diagonal matrix with its input as diagonal entries, and the eigen-values $\mu_k\geq 0$ are in descending order. Define the unitary eigen-vector matrix $\bU_\bs$ as a 2 by 2 block matrix, i.e.,
\beq
\bU_\bs \ \define \ \begin{bmatrix}
\bU_{\bs11} & \bU_{\bs12} \\
\bU_{\bs21} & \bU_{\bs22}
\end{bmatrix} \ ,
\eeq
where $\bU_{\bs ij}\in\mbbC^{M\times M}$ and $i,j=1,2$.
Then, the method of moments (MM) estimators of $\bF$ and $\bSig_\bh$ are, respectively, conditioned on $\bU_{\bs11}$ is non-singular,
\beq\label{MME_ch4}
\hat{\bF}_{MM} \ = \ \bU_{\bs21} \bU_{\bs11}^{-1} \ , ~~~~ \hat{\bSig_\bh}_{MM} \ = \ \bU_{\bs11}\left(\diag(\mu_1,\dots,\mu_M) - \sigma^2\bI_M\right)^+ \bU_{\bs11}^H \ ,
\eeq
where $\sigma^2\define 2N\sigma_n^2/PT$ and the function $(\cdot)^+$ is defined in Theorem \ref{MLE_ch4}.
\end{lemma}
\begin{proof}
The multi-packet sufficient statistics can be collected as
\beq\label{Vs_MIMO}
\bV_s \ \define \ \begin{bmatrix}
\bY_{1}\\
\bY_{2}
\end{bmatrix} \ = \ \begin{bmatrix}
\bH\\
\bF\bH
\end{bmatrix} + \bN_s\in\mbbC^{2M\times NL} \ ,
\eeq
where the noise is i.i.d., that is $\vc \bN_s\sim\mccn\left(\bzro_{2MNL}, \sigma^2\bI_{2MNL}\right)$, as proven in Lemma \ref{4lem_ss_mp}.
It is straightforward to show that
\beq
E\left[\bS_{mp}\right] \ = \ \frac{1}{NL}E\left[\bV_s\bV_s^H\right] \ = \ \begin{bmatrix}
\bSig_\bh + \sigma^2\bI_M& \bSig_\bh \bF^H\\
\bF\bSig_\bh & \bF\bSig_\bh\bF^H + \sigma^2\bI_M
\end{bmatrix}\ .
\eeq
Apparently, the second moments of sufficient statistics $\bY_1$ and $\bY_2$ are functions of unknown parameters $\btt$, or $\bF$ and $\bSig_\bh$ \er{theta_MIMO},
\beq
E[\bS_{mp}] \ = \ \bT\left(\btt\right) \ = \ \begin{bmatrix}
\bI_M\\
\bF
\end{bmatrix} \bSig_\bh \begin{bmatrix}
\bI_M&\bF^H
\end{bmatrix} + \sigma^2\bI_{2M}\ ,
\eeq
where $\bT:\mbbC^{2M^2}\rightarrow\mbbC^{2M\times 2M}$ denotes a mapping.
Then, from basic principles of MM estimation \cite[Sec.~9.4]{kay}, we find $\hat{\btt}_{MM}$ by the inverse of aforementioned mapping,
\beq
\hat{\btt}_{MM} \ = \ \bT^{-1}\left(\bS_{mp}\right) \ .
\eeq
The formula of MM estimators in \er{MME_ch4} follows directly from the proof of Theorem \ref{MLE_ch4}. This completes the proof.
\end{proof}
The MM estimators are provably consistent, easy to determine, and does not require knowing $\bC_\bh$. However, they are generally sub-optimal to the corresponding ML estimators \cite[Ch.~9]{kay}. Next we show a special case where the ML estimators and the MM estimators coincide, and then discuss how to find the ML estimators in general fading conditions if $\bC_\bh$ is known.
\begin{corollary}[ML Estimators for Fast Fading]
If the Rayleigh fading channel is temporally i.i.d., that is $\bC_\bh=\bI_L$ in \er{vec_Hmp}, then the multi-packet MM estimators given in \er{MME_ch4} for $\bF$ and $\bSig_\bh$ satisfy \er{4mle_criteria_mp} and hence are the maximum-likelihood estimators.
\end{corollary}
\begin{proof}
Conditioned on $\bC_\bh=\bI_L$, the channel matrix $\bH$ in \er{vec_Hmp} would satisfy,
\beq
\vc \bH \sim\mccn\left(\bzro_{MNL}, \bI_{NL} \otimes\bSig_\bh\right) \ .
\eeq
Note the similarity between the distribution of this multi-packet $\bH$ and that of the single-packet channel matrix in \er{pdf_H_MIMO}. The proof follows Theorem \ref{MLE_ch4}, except the number of i.i.d. columns in $\bH$ is $NL$ rather $N$.
\end{proof}
Next we discuss finding the ML estimator $\hat{\btt}_{ML}$ under general fading conditions.
We assume the temporal correlation $\bC_\bh$ is known, and $L$ packets can be decorrelated by its eigen-vector matrix $\bQ$, i.e.,
\beq\label{Ch_eig_ch4}
\bQ^H\bC_\bh\bQ \ = \ \bD \ \define \ \diag(d_1,\dots,d_L) \ .
\eeq
Note $\bC_\bh$ is normalized with 1's on its diagonal, i.e., $\Trace[\bC_\bh] = \Trace[\bD] =L$.
Consider the decorrelated observation, i.e,.
\beq\label{V_mp}
\bV \ \define \ \bV_s \left(\bQ^*\otimes \bI_N\right) \ = \ \begin{bmatrix}
\bH_{d} \\
\bF\bH_{d}
\end{bmatrix} + \bN \in\mbbC^{2M\times NL}\ ,
\eeq
where $\vc \bN = [\left(\bQ^*\otimes \bI_N\right)^T\otimes\bI_{2M}]\vc\bN_s\sim\mccn\left(\bzro_{2MNL}, \sigma^2\bI_{2MNL}\right)$ is i.i.d. and,
\beq
\vc\bH_{d}\sim\mccn\left(\bzro_{MNL}, \bD \otimes\bI_N \otimes\bSig_\bh\right) .
\eeq
The understand this, for each $1\leq k\leq L$, we have $N$ i.i.d. complex Gaussian random vectors that follow $\mccn\left(\bzro_M, d_k\bSig_\bh\right)$, where $d_k$ are defined in \er{Ch_eig_ch4}.
We define the log-likelihood function based on the pdf of $\bV\define [\bv_1, \dots, \bv_{NL}]$ is,
\bea\label{pdf_mp_ch4}
\mathcal{L}(\btt) & \define & \ln p\left(\btt;\bV\right)p(\bV; \btt) \nn \\
& = & C - N \sum_{k=1}^{L} \left(\ln \det \bSig_k + \sum_{i=1}^{N}\bv_{(k-1)N+i}^H\bSig_k^{-1}\bv_{(k-1)N+i}\right) \nn\\
& = & C - N \sum_{k=1}^{L} \left(\ln \det \bSig_k+\Trace[\bS_k\bSig_k^{-1}]\right) \ ,
\eea
where $C$ is a constant independent of $\btt$ and we define for $1\leq k \leq L$,
\beq
\bS_k \ \define \ \frac{1}{N}\sum_{i=1}^{N} \bv_{(k-1)N+i}\bv_{(k-1)N+i}^H \ ,
\eeq
and
\bea\label{CovYk}
\bSig_k & \define & \begin{bmatrix}
d_k\bSig_\bh + \sigma^2\bI_M& d_k\bSig_\bh \bF^H\\
d_k\bF\bSig_\bh & d_k\bF\bSig_\bh\bF^H + \sigma^2\bI_M
\end{bmatrix} \nn\\
&=& d_k
\begin{bmatrix}
\bI_M\\
\bF
\end{bmatrix} \bSig_\bh \begin{bmatrix}
\bI_M&\bF^H
\end{bmatrix} + \sigma^2\bI_{2M} \ = \ \bB_1\bD_{1,k}\bB_1^H + \bB_2\bD_2\bB_2^H \ ,
\eea
where $\bD_{1,k}\in\mbbC^{M\times M}$ and $\bB_i\in\mbbC^{2M\times M}$ (for $i=1,2$) are the block eigenvalues and orthonormal block eigenvectors of $\bSig$,
\bea\label{eq_Dk}
\bB_1 &=&
\begin{bmatrix}
\bI_M\\
\bF
\end{bmatrix} \bA_1^{-\frac{1}{2}} \ , ~~~ \bB_2 \ = \ \begin{bmatrix}
-\bF^H\\
\bI_M
\end{bmatrix} \bA_2^{-\frac{1}{2}} \ , \nn\\
\bD_{1,k} &=& d_k \bA_1^{\frac{1}{2}}\bSig_\bh\bA_1^{\frac{1}{2}} + \sigma^2\bI_M \ , ~~~ \bD_2 \ = \ \sigma^2\bI_M \ ,
\eea
and $\bA_1$ and $\bA_2$ are positive-definite matrix defined in \er{As}.
Since eigenvalues of $\bD_{1,k}$ are also eigenvalues of $\bSig$ \cite[Th. 1.1]{pere}, and observe that $\bD_2$ is already diagonal, we write down the (scalar) eigenvalue decomposition of $\bSig$,
\beq\label{Sig_eig_mp}
\bSig_k \ = \ \bU \begin{bmatrix}
\bLam_{1,k} & \bzro_{M\times M} \\
\bzro_{M\times M} & \sigma^2\bI_M
\end{bmatrix} \bU^H \ \ ,
\eeq
where we define
\beq\label{eq_U_mp}
\bU \ \define \ \begin{bmatrix}
\bB_1 \bQ & \bB_2
\end{bmatrix} \ ,
\eeq
and we have
\bea
\bD_{1,k}\bQ &=& \bQ\bLam_{1,k} \ , \nn\\
\bLam_{1,k}&=&\diag(\lambda_{1,k},\dots,\lambda_{M,k}) \ ,
\eea
with $\bQ\in\mbbC^{M\times M}$ unitary. Note $\lambda_{j,k} \geq \sigma^2$ for all $1\leq j \leq M$ and $1\leq k\leq L$. In other words, the covariance matrix $\bSig_k$ has $2M$ real eigenvalues, where the largest $M$ eigenvalues are greater than or equal to $\sigma^2$, while the smallest $M$ eigenvalues are exactly $\sigma^2$.
Note that quality of virtual channels ($d_k$) after decorrelation is buried in $\lambda_{j,k}$. Because of the freedom of $d_k$ as in definition of $\bD_{1,k}$ in \er{eq_Dk}, a closed-form expression for the ML estimator is mathematically intractable. Since seeking a fast algorithm that allows real-time impedance estimation is our goal, we consider numerical methods based on iterations (e.g., gradient descent or Newton's method) are out of the scope of this paper. Instead, we use the fundamental lower bound in CRB as a reference when we evaluate the ML or MM estimators in the numeral section.
The multi-packet FIM follows directly from \eqref{FIM_MIMO},
\beq\label{FIMmp}
[\boldsymbol{\mathcal{I}}_{mp}(\btt)]_{ij} \ = \ N\cdot \sum_{k=1}^{L}\Trace\left[\bSig_k^{-1}\frac{\partial \bSig_k}{\partial \theta_i^*}\bSig_k^{-1}\frac{\partial \bSig_k}{\partial \theta_j}\right] \ ,
\eeq
where $\bSig_k$ is the $k$-th covariance matrix defined in \er{CovYk}. Similarly, the error covariance matrix $\bC_{\hat{\btt}}$ of any unbiased estimator $\hat{\btt}$ is lower bounded by the Cram\'er-Rao bound (CRB), which is the inverse of $\boldsymbol{\mathcal{I}}_{mp}(\btt)$,
\beq\label{crb_mp_ch4}
\bC_{\hat{\btt}} \ \define \ E_{\bY_1,\bY_2;\btt}\left[\left(\hat{\btt}-\btt\right)\left(\hat{\btt}-\btt\right)^H\right] \ \geq \ \boldsymbol{\mathcal{I}}_{mp}(\btt)^{-1}\ ,
\eeq
where the expectation is with respect to pdf in \er{pdf_mp_ch4} and $\bA\geq \bB$ means $\bA-\bB$ is positive semi-definite.
For any estimator of $\bF$, we find an estimator for antenna impedance via \er{F_MIMO}, i.e.,
\beq\label{zAhat}
\hat{\bZ}_\bA \ = \ \left(\bZ_1-\bZ_2\bR_2^{-1/2}\hat{\bF}_{MM}\bR_1^{1/2}\right)\left(\bR_2^{-1/2}\hat{\bF}_{MM}\bR_1^{1/2}-\bI_M\right)^{-1}\ .
\eeq
However, due to the reciprocity theorem of electromagnetics\cite[pg. 144]{bala}, $\bZ_\bA$ is symmetric and so should any reasonable estimate of it. Here we replace $\hat{\bZ}_\bA$ by its nearest symmetric matrix, i.e.,
\beq\label{zA_est_MIMO}
\tilde{\bZ}_\bA \ \define \ \frac{1}{2}\left(\hat{\bZ}_\bA + \hat{\bZ}_\bA^T \right) \ .
\eeq
Based on this new estimate, the receiver matches its load impedance for minimum noise-figure, which reduces to maximum power transfer under our noise model \cite[eq.~10]{deso}, $\hat{\bZ}_L = \tilde{\bZ}_\bA^*$.
Consequently, we calculate an excess (transmit) power needed for this matching compared to the truly optimal one, i.e., $\bZ_{L,opt} = \bZ_\bA^*$,
\beq\label{ex_pwr}
10\log_{10}\left(\Trace\left[\left(4\bR_\bA\right)^{-1}\bSig_\bg\right]/E\Trace\left[\left(\tilde{\bZ}_\bA^*+\bZ_\bA\right)^{-H}\tilde{\bR}_\bA\left(\tilde{\bZ}_\bA^*+\bZ_\bA\right)^{-1}\bSig_\bg\right] \right) \ ,
\eeq
where $\tilde{\bR}_\bA \define \Real\{\tilde{\bZ}_\bA\}$.
In the next section, we compare the performance of estimators derived in this paper against their corresponding lower bounds, and explore the potential benefits of these estimators on system-level metrics, such as channel capacity.
\section{Numerical Results}\label{4secV}
In this section, we explore the performance of estimators in the previous section through numerical examples. Consider a narrow-band MIMO communications system with $N=4$ transmit antennas and $M=2$ receive antennas, whose carrier frequency is 2.1 GHz. This frequency is chosen based on the first E-UTRA down-link operating band in LTE specifications\cite{3gpp_TS36101}. The duration of each data packet equals to a subframe of LTE, i.e., $T_s=1$ ms. Block-fading channel is assumed, such that during one data packet, the channel information remains the same, but it generally varies from packet to packet \cite{bigu}.
For each data packet, a training sequence precedes data sequence \cite[Fig. 1(a)]{liu}.
We take the two partitions of the training sequence $\bX=[\bX_1, \bX_2]$ in \er{Xs_def} from a normalized discrete Fourier transform
(DFT) matrix of dimension $K=T/2=32$, e.g., \cite[eq. 10]{bigu}. In particular, the first part $\bX_1$ is chosen as the first $N$ rows, while $\bX_2$ the next $N$ rows, and $\bX_i\bX_i^H=K\bI_N$ for $i=1,2$. The unknown antenna impedance is that of a uniform linear array (ULA) \cite{domi2}, i.e.,
\beq
\bZ_\bA \ = \ \begin{bmatrix}
72.8521 + j1.6869 &-15.7457 -j27.8393\\
-15.7457 -j27.8393 & 72.8521 + j1.6869
\end{bmatrix} \,\Omega \ .
\eeq The load impedance is $\bZ_1=50\bI_M \,\Omega$ for the first $K=32$ symbols of each training sequence, and $\bZ_2=(50+j20)\bI_M+10\times{\bf 1} \,\Omega$ for the remaining $T-K=32$ symbols, where ${\bf 1}$ is the $M$ by $M$ all one matrix. From \er{F_MIMO}, it follows that
\beq
\bF \ = \ \begin{bmatrix}
0.9804 - j0.1613& 0.0261 - j0.0334\\
0.0261 - j0.0334& 0.9804 - j0.1613
\end{bmatrix} \ .
\eeq
In this section, we explore important properties of the estimators derived in previous sections.
The average post-detection SNR of a received symbol is defined from \er{sgl_model_MIMO} as \cite[Sec. VIII]{bigu},
\beq\label{symbolSNR}
\gamma \ \define \ \frac{E\Trace\left[\bH\bx\bx^H\bH^H\right]}{E\Trace\left[\bn_L\bn_L^H\right]} \ = \ \frac{P\sigma_H^2}{\sigma_{n}^2}\ ,
\eeq
where $\sigma_{n}^2$ is the noise variance at each port of the $M$-port receiver and $\sigma_H^2$ is the mean of diagonal entries of $\bSig_\bh$ in \er{pdf_H_MIMO},
\beq\label{sgmH2_MIMO_ch5}
\sigma_H^2 \ \define \ \frac{1}{M}\Trace[\bSig_\bh] \ .
\eeq
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.6\textwidth, keepaspectratio=true]{./figs/Fml_iid_fd_Tx_diversity_M2_L5_12_17_18}
\end{center}
\vspace{-12pt}
\caption{Relative MSE of $\hat{\bF}_{ML}$ versus SNR in i.i.d. Fading, $L=5$. }
\vspace{-18pt}
\label{fig_iid_F_MIMO}
\end{figure}
As shown is Fig. \ref{fig_iid_F_MIMO}, the relative root mean-square error (RMSE) is plotted against SNR \er{symbolSNR}. The ML estimator in $\hat{\bF}_{ML}$, for a given $L=5$ packets, becomes efficient as the number of transmit antenna increases, i.e., more spatial diversity. We also observe that the Miller-Chang bound (MCB) is tighter than the CRB and touches the RMSE for all values of $L$ and SNR plotted in Fig.~\ref{fig_iid_F_MIMO}. Although the ML estimators are asymptotically unbiased and efficient, i.e., it achieves its corresponding CRB, the MCB (if exists) better predicts the RMSE of $\hat{\bF}_{ML}$ for finite sample size in $L$. For different $N$ the CRB generally have different values as indicated by the formulas of Fisher information matrix in \er{FIM_MIMO}, but their numerical evaluations
seem indistinguishable in Fig.~\ref{fig_iid_F_MIMO}.
Next, we investigate the performance of estimators derived previously under different Rayleigh fading conditions, i.e., fast, medium, and slow fading \cite{wu3}. In particular, Clarke's model is assumed \cite{badd,zhen} and the normalized channel correlation matrix is,
\beq
\bC_\bh \ = \ \begin{bmatrix}
R[0] & R[-1] & \cdots & R[-L+1]\\
R[1] & R[0] & \cdots& R[-L+2] \\
\vdots & \ddots & \ddots & \vdots \\
R[L-1] & R[L-2]& \cdots & R[0]
\end{bmatrix} \ ,
\eeq
where $R[l] \ = \ J_0(2\pi f_dT_s |l|)$,
$J_0(\cdot)$ is the zeroth-order Bessel function of the first kind, $T_s=1~ms$ is the sampling interval, and $l$ is the sample difference. The fading frequency (maximum Doppler frequency) is
$f_d\define v/\lambda$,
where $v$ is the velocity of the fasting moving scatterer and $\lambda$ the wave-length of the carrier frequency.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.6\textwidth, keepaspectratio=true]{./figs/MIMO_Fmm_MSE_L10_4by2_12_17_18}
\end{center}
\vspace{-12pt}
\caption{Relative MSE of $\hat{\bF}_{MM}$ for a 4 by 2 MIMO, $L=10$}
\vspace{-18pt}
\label{fig_Fmm_fd_4by2}
\end{figure}
In Fig.~\ref{fig_Fmm_fd_4by2}, we plot the relative RMSE of the method of moments (MM) estimators $\hat{\bF}_{MM}$ \er{MME_ch4}, for a MIMO with $N=4$ transmit and $M=2$ receiver antennas. The velocity of the fastest moving scatterer is $v=300, 50$ and $5$ km/h, which represents a fast, medium, and slow fading scenario, respectively. The MM estimator $\hat{\bF}_{MM}$ is about 3 dB aways from its CRB under slow fading, and this gaps narrows to less 1 dB for medium and fast fading. Thus, a faster fading results in improved impedance estimation accuracy. This is reasonable as fast fading means more temporal diversity and less correlation between observations. Similar to Fig.~\ref{fig_iid_F_MIMO}, the CRB depends very little on fading conditions; the CRB's of three cases considered are indistinguishable.
Next we evaluate the excess power defined in \er{ex_pwr}. A faster fading channel results in a smaller excess power. This means the transmitter may save power for an intended receive SNR, due to an improved match between antenna and load after impedance estimation and mismatch compensation. For example, the gain between the fast and slow fading cases is about 3 dB at low SNR. If a 0.5 dB excess power or less is considered a good match in practice, then it is achieved at relatively low SNR for all fading conditions. Further, the excess power vanishes at high SNR. Next we give two examples which demonstrates the benefits of this impedance estimation algorithm in terms of ergodic capacity.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.6\textwidth, keepaspectratio=true]{./figs/MIMO_pwr_fds_4by2_12_19_18}
\end{center}
\vspace{-12pt}
\caption{Excess Power of $\hat{\bF}_{MM}$ for 4 by 2 MIMO.}
\vspace{-18pt}
\label{fig_pwr_4by2}
\end{figure}
As derived by Hassibi and Hochwald, a lower bound on (ergodic) capacity exists, which incorporates the MMSE channel estimation error\cite[eq.~21]{Hassibi}, i.e.,
\beq
\ C_l \ = \ E \left[\log_2\det\left(\bI_M +\gamma_\text{eff}\cdot \frac{1}{N} \bSig_{\tilde{\bh}} \bH_w\bH_w^H\right)\right] \ ,
\eeq
where $\vc \bH_w\sim\mccn(\bzro, \bI_{MN})$, $\bSig_{\tilde{\bh}}$ is the normalized version of $\bSig_{\bh}$ such that $\Trace[\bSig_{\tilde{\bh}}]/M=1$, and with \er{symbolSNR} the effective SNR is defined as,
\beq
\gamma_\text{eff} \ = \ \frac{P\sigma_{H}^2}{\sigma_{n}^2} \frac{PT\sigma_{H}^2}{PT\sigma_{H}^2+N(P\sigma_{H}^2+\sigma_{n}^2)} \ = \ \gamma \,\frac{1}{1+(1+1/\gamma)N/T}\ .
\eeq
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.6\textwidth, keepaspectratio=true]{./figs/MIMO_capacity_fds_5dBls_4by2_12_19_18}
\end{center}
\vspace{-12pt}
\caption{Ergodic Capacity over SNR for 4 by 2 MIMO.}
\vspace{-18pt}
\label{fig_capacity_gain}
\end{figure}
Shin and Lee derived an upper bound for this ergodic capacity in closed-form, putting the expectation between $\log_2(\cdot)$ and $\det(\cdot)$ by Jensen's inequality \cite[Th.~III.2]{shin}, i.e.,
\beq
C_l \ \leq \ \log_2\left(\sum_{k=0}^{M} \left[\left(\frac{\gamma_\text{eff}}{N}\right)^kk! \,\sigma_k\left(\bSig_{\tilde{\bh}}\right)\cdot \sigma_k\left(\bI_N\right)\right]\right) \ ,
\eeq
where $M\leq N$ is assumed and $\sigma_k(\bA)$ denotes the sum of all the $k$-rowed principal minor determinants of a square matrix $\bA$ \cite[pg. 17]{horn}. In particular, we have \cite[Th.~II.3]{shin},
\beq
\sigma_k\left(\bI_N\right) \ = \ {N \choose k} \ = \ \frac{N!}{k! (N-k)!} \ .
\eeq
Consider a 4 by 2 MIMO system again, i.e., $N=4$ and $M=2$. This ergodic capacity upper bound boils down to
\beq\label{capacity_MIMO}
C_l \ \leq \ \log_2\left[1 + 2\gamma_\text{eff} + \frac{3}{4} \gamma_\text{eff}^2\cdot \det(\bSig_{\tilde{\bh}})\right] \ .
\eeq
Although calculating an upper bound instead of the exact ergodic capacity is less than ideal, it should qualitatively demonstrate the capacity boost using our proposed antenna impedance estimation algorithm.
In Fig.~\ref{fig_capacity_gain}, ergodic capacity upper bound \er{capacity_MIMO} are plotted against SNR of an originally mismatched receiver. The power loss due to mismatch is chosen as $5$dB. After applying our algorithm and matching to the estimate of $\bZ_\bA$ \er{zA_est_MIMO}, a significant gain on ergodic capacity $C(\hat{\bF}_{MM})$ is observed, compared to the mismatched receiver (the black solid line). This gain ranges from about 50\% at low SNR to 20\% at high SNR. The black dash line represents an upper bound on ergodic capacity, where the receiver is always optimally matched and observes the channel without errors. This upper bound, although unachievable by any practical system, is closed in by $C(\hat{\bF}_{MM})$ to around 1 dB or less for all SNR and fading conditions considered. Also note at low SNR, faster fading leads to a marginally capacity boost, which vanishes as SNR increases.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.6\textwidth, keepaspectratio=true]{./figs/MIMO_capacity_d_fds_4by2_12_19_18}
\end{center}
\vspace{-12pt}
\caption{Ergodic Capacity over $d/\lambda $ for 4 by 2 MIMO, SNR = 10dB.}
\vspace{-18pt}
\label{fig_capacity_gain_over_d}
\end{figure}
Plotted in Fig.~\ref{fig_capacity_gain_over_d} is the ergodic capacity upper bound in \er{capacity_MIMO} versus antenna element-separation $d/\lambda$. The SNR for the originally mismatched receiver is fixed at 10 dB, while other settings remain identical as in Fig.~\ref{fig_capacity_gain}. Similar observations are also made here, as the $C(\hat{\bF}_{MM})$'s hone in the practically unachievable upper bound (the black dash line) within a fraction of 1 bit/s/Hz. This upper bound seems to depend very little on antenna spacing, yet the other capacity curves tend to drop for closely spaced arrays. Compared to the mismatched receiver, our algorithm improves capacity by over 30\% for all data points.
\section{Conclusion}\label{4secVI}
In this paper, we derived the maximum-likelihood (ML) and method of moments (MM) estimators for MIMO antenna impedance using training sequences in various fading conditions. In particular, under i.i.d. fading, the ML estimator was derived as the ratio out of the top block eigen-vector of the sample covariance matrix. This ML estimator was shown to be a MM estimator under temporally correlated Rayleigh fading. We also derived two fundamental lower bounds on these estimators, and explored the performance of these estimators through numerical examples. The ML and MM estimators become efficient (to CRB) when sufficient spatial and/or temporal diversity exists. A typical rule of thumb is the number of diversity is four times the number of receive antennas. Additionally, trade-off between channel correlation and impedance estimation accuracy was investigated. Our numerical results indicate that the MIMO antenna impedance can be accurately estimated in a matter of milliseconds. This estimate is able to compensate power losses due to impedance mismatch partially at low SNR and almost all at high SNR. In the example of ergodic capacity, if the original mismatch at the receiver is significant, large capacity boost can be observed in general.
\bibliographystyle{unsrt}
|
\section{QUANTUM NEURAL NETWORKS}
Machine learning is a way of extracting hidden patterns from data by learning a set of optimal parameters for a mathematical expression that most closely match the data. The mathematical expression used for pattern extraction is called a machine learning algorithm. An algorithm with an optimal set of parameters, learned via training, is called a model. With near-term devices available on cloud, execution of quantum machine learning (QML) algorithms on quantum computers or simulators is now feasible.
QML algorithms are built with variational circuits i.e., parametrized circuits, composed of quantum gates whose actions are defined by parameters. Training is the process of "learning" optimal parameters of the gates which produce as accurate inferences as possible for new data samples. The measurement results from a variational circuit run on a QPU are sent to a CPU for parameter optimization, i.e., computation of objective function, gradients, and new parameters. The updated parameters are fed back to the quantum circuit to adjust the parameterized gates for next iteration. The illustration of the process is shown in the figure.
\begin{figure}[H]
\centering
\includegraphics[scale=0.47]{Figures/var.pdf}
\caption{Variational quantum circuit parameter update}
\label{fig:var}
\end{figure}
Google and Xanadu offer Python based software packages specifically for quantum machine learning: Tensorflow Quantum (Google) and Pennylane (Xanadu) \cite{Penny}.
Neural network is one of the subsets of machine learning algorithms, defined by a stack of layers, each composed of an affine transformation $Wx+b$ and a non-linear activation function $\phi(\cdot)$. Each layer of a neural network can be mathematically described as $L(x) = \phi(Wx+b)$. The output from one layer is then fed as input into the subsequent layer and the entire network is a composition of different layers: $L(x) = L_m \circ L_{n-1} \circ \ldots L_1(x)$. The entries of the matrix $W$ and the bias vector $b$ for each layer are learned as parameters through an iterative training process given an objective function. The goal is to find an optimal set of parameters $\{W_1, W_2, \ldots, W_m, b_1, b_2, \ldots, b_m \}$ for a network of $m$ layers.
In quantum neural networks, the objective is to implement the classical mathematical expression $L(x)=\phi(Wx+b)$ as a quantum state $L(\ket{x})=\ket{\phi(Wx+b)}$. In converting classical neural networks into quantum circuits, the key components are:
\begin{itemize}
\item data encoding: $x \rightarrow \ket{\psi(x)}$
\item affine transformation $W\ket{\psi(x)}+\ket{b}$
\item non-linear activation function $\phi(\ket{\cdot})$
\end{itemize}
In the qubit model, all available unitary gates are linear. Hence a direct way of implementing the bias addition component and the non-linear activation function component of classical neural networks into quantum is absent in the model.
In the CV model, however, the displacement gate and the Kerr gate allow for a direct translation from classical to quantum.
\subsection{Continuous variable QNN}
Naturally embedded in the CV model are quantum gates to directly implement the expression $L(\ket{x})=\ket{\phi(Wx+b)}$.
The affine transformation $Wx + b$ is implemented by the composition $\mathit{D} \circ \mathit{U}_2 \circ \mathit{S} \circ \mathit{U}_1$, where $U_k$ denotes the $k^{th}$ interferometer, $S$ a set of $m$ squeezers, $D$ a set of $m$ displacement gates. The activation function $\phi(\ket{\cdot})$ is implemented by a set of Kerr gates, which are non-linear. The composition $\phi \circ \mathit{D} \circ \mathit{U}_2 \circ \mathit{S} \circ \mathit{U}_1$ acting on a quantum state $\ket{x}$ gives us the desired state $L(\ket{x})= \ket{\phi(Wx+b)}$. The schematic of the circuit is shown below.
\begin{figure}[H]
\centering
\includegraphics[scale=0.38]{Figures/qnn.pdf}
\caption{CV quantum neural network architecture \cite{cv_19}}
\label{fig:qnn}
\end{figure}
The interferometer $U_k$ on an $m-$qumode system is composed of $m-1$ beamsplitters and $m$ rotation gates as shown in the figure.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{Figures/inter.pdf}
\caption{Make-up of an interferometer}
\label{fig:inter}
\end{figure}
The action of a phaseless interferometer $\mathit{U}_k$ on the quantum state $\ket{x} = \otimes_{k=1}^m \ket{x_k}$ has an effect of an orthogonal matrix acting on $\ket{x}$ \cite{cv_19}. Orthogonal matrices are just unitary matrices with real entries, inducing length-preserving rotations. Then the transpose of an orthogonal matrix represents the reverse rotation of the original matrix, thus orthogonal. Then the composition $U_2 \circ S \circ U_1$ can be considered as the composition $O_2 \circ S \circ \left(O_1^T \right)^T$, where $O_2$ and $O_1^T$ are orthogonal.
Let $W$ be the linear transformation matrix we want to implement with a quantum circuit. Any matrix $W$ can be factorized using singular value decomposition (SVD) as $W = U \Sigma V^{\ast}$, where $U$ and $V$ are orthogonal and $\Sigma$ is diagonal \cite{SVD}. The parameterized squeezer $S(r_k)$ acts on the quantum state $\ket{x_k}$ of each $k^{th}$ qumode as $S(r_k)\ket{x_k} = \sqrt{e^{-r_k}}\ket{e^{-r_k} x_k}$. Collectively they have an effect of a diagonal matrix $\mathit{S}= S(r_1) \otimes S(r_1) \otimes \ldots \otimes S(r_m)$ acting on $\ket{x} = \otimes_{i=1}^m \ket{x_k}$. The composition $\mathit{U}_2 \circ \mathit{S} \circ \mathit{U}_1$ implements a quantum version of the linear transformation matrix $W$ \cite{cv_19}.
The bias addition is realized with displacement gates $D$.
The displacement gate has an effect
\[
D(\alpha_k )\ket{\psi_k}=\ket{\psi_k +\sqrt{2}\alpha_k}
\]
for each $k^{th}$ qumode. Then $D(\alpha)\ket{\psi}=\ket{\psi +\sqrt{2}\alpha}$ collectively for $\alpha^T = [\alpha_1, \alpha_2, \ldots, \alpha_m]$. For some desired bias $b$, let $\alpha = \frac{b}{\sqrt{2}}$, then the collection of displacement gates implements the bias addition. The composition $D \circ U_2 \circ S \circ U_1$ acting on the quantum state $\ket{x}$ gives us the affine transformation
\[
D \circ U_2 \circ S \circ U_1 \ket{x}=\ket{O_2 \Sigma O_1 x + b}=\ket{Wx+b}.
\]
The non-linear activation function $\phi(\cdot)$ is realized with Kerr gates. The Kerr gate, parameterized by the parameter $\kappa$, is a non-linear transformation gate. Let $n$ be the cutoff dimension and $m$ the number of qumodes. For the quantum state $\ket{\psi}$ of one qumode, which is a superposition of $n$ Fock basis states, the Kerr gate with parameter $\kappa$ has an effect
\[
\begin{aligned}
K(\kappa)\ket{\psi} &=
\begin{bmatrix}
e^{i \kappa 0^2} & 0 & \ldots & 0\\
0 & e^{i \kappa 1^2} & \ldots & 0\\
\vdots & &\ddots & \vdots\\
0 & 0 &\ldots & e^{i \kappa (n-1)^2}\\
\end{bmatrix}
\begin{bmatrix}
\psi_0\\
\psi_1\\
\vdots \\
\psi_{n-1}\\
\end{bmatrix}\\
&=
\begin{bmatrix}
\psi_0\\
e^{i \kappa 1^2}\psi_1\\
\vdots \\
e^{i \kappa (n-1)^2}\psi_{n-1}\\
\end{bmatrix},
\end{aligned}
\]
which is non-linear.
Together, the circuit $L=\Phi \circ D \circ U_2 \circ S \circ U_1$ gives us a quantum version $L \left( \ket{x} \right)=\ket{\Phi(Wx+b)}$ of a classical neural network $L(x)=\Phi(Wx+b)$.
\subsection{CV Binary Classifier}
The binary classifier outlined in "Continuous-variable quantum neural networks" is a classical and quantum hybrid network \cite{cv_19}.
The dataset used contains 284,806 genuine and fraudulent credit card transactions with 29 features, out of which only 492 are fraudulent. The dataset is truncated to 10 features as per the paper and 1,968 samples with 1:3 ratio of fraudulent vs. genuine.
The proposed classical-quantum hybrid model has a classical neural network taking input vectors of size 10 and outputting vectors of size 14, quantum encoding circuit, and a 2-qumode quantum neural network which outputs vectors of size 2. We can regard the output vectors as one-hot encoding of binary classification of fraudulent vs. genuine. The architecture of the hybrid network is
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{Figures/binary.pdf}
\caption{Binary hybrid classifier circuit \cite{cv_19}}
\label{fig:binary}
\end{figure}
The data flow of the circuit is
\begin{itemize}
\item Classical network: 2 hidden layers with 10 neurons, each using Exponential Linear Units (ELU) as activation function. Output layer with 14 neurons.
\item Data encoding: Output vector from the classical network is converted to a quantum state by the circuit - squeezers, interferometer, displacement gates, and Kerr gates
\item Quantum network: 4 layers of QNN. Each layer is composed of interferometer 1, squeezers, interferometer 2, displacment gates, and Kerr gates.
\item Measurement: The expectation value of the Pauli$-X$ gate $\bra{\phi_k}X \ket{\phi_k}$ is evaluated for each qumode state$ \ket{\phi_k}$ for the $k^{th}$ qumode.
\end{itemize}
The experiment yields $97\%$ training accuracy.
\href{https://github.com/sophchoe/Binary_Classification_Pennylane_Keras}{Code: Keras-Pennylane implementation}
\end{document}
\section*{APENDIX}
\subsection{Qubit based QNN}
In this section, I analyze two qubit-based quantum neural network MNIST classifiers: one by Google using Tensorflow Quantum and the other by IBM using Qiskit and PyTorch. Both models are binary classifiers, performing measurements on only one qubit with the observed computational result after measurement being either $0$ or $1$. Hence, they work with a reduced dataset of two classes. The Google version is a pure quantum network and the IBM version is a classical-quantum hybrid network with the classical network being a convolutional neural network.
For quantum data encoding, Google uses the Pauli-X gate and IBM uses the Hadamard followed by a parameterized $R_y$ gate. The Google approach is to prepare qubits with the Pauli$-X$ gate for the values above a certain threshold and applying controlled rotation gates using the ancillary bit as the control.
\subsubsection{Tensorflow Quantum}
Google's QNN binary classifier implementation on MNIST code is listed on the Tensorflow Quantum website. The network is a pure quantum circuit composed of Pauli-$X$ gates, Hadamard gates, paramterized Pauli$-XX$ gates and parametrized Pauli$-ZZ$ gates. The circuit is implemented on Google's $63-$qubit Sycamore, whose architecture is of a lattice shape as shown in the figure. To select $n-$qubits from the chip, it is customary to pick a $log_2 (n) \times log_2 (n)$ block.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{Figures/Sycamore.pdf}
\caption{Google's Sycamore QPU architecture \href{https://techcrunch.com/wp-content/uploads/2019/10/qubit-lattice.png}{[image source]}}
\label{fig:sycamore}
\end{figure}
The dataset used is of 10,338 samples of the classes 3 and 6, with each image matrix being $28 \times 28$ pixels ranging from $0$ (darkest) to $255$ (lightest). The pre-processing steps are:
\begin{itemize}
\item Normalize the pixel values to $[0,1]$.
\item Reduce the image size from $28 \times 28$ to $4 \times 4$.
\item Convert pixel values higher than $0.5$ to $1$ and the rest to $0$.
\end{itemize}
The result of the pre-processing process are $4 \times 4$ image matrices with entries either 0 or 1 as shown in the figure.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Figures/3_6_matrix.pdf}
\caption{images of class 3 and class 6}
\label{fig:3_6}
\end{figure}
For the circuit, $4 \times 4$ grid qubits are used mirroring $4 \times 4$ input matrices. An additional read-out qubit is added, which is the only qubit used for measurement after computation. The quantum gates used for the circuit are Pauli-$X$ gates for data encoding, parametrized $XX-$gates and parametrized $ZZ-$gates for QNN, and the Pauli-$X$ and the Hadamard gate for the read-out qubit. A simplified version of the architecture, with the first column of grid qubits is depicted in the figure.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Figures/tensorq.pdf}
\caption{images of class 3 and class 6}
\label{fig:3_6}
\end{figure}
The data-flow of the circuit is
\begin{itemize}
\item Initialized all the qubits to $\ket{0}$.
\item Apply the Pauli$-X$ gate and the Hadamard gate to the read-out qubit, obtaining $\frac{\ket{0}-\ket{1}}{\sqrt{2}}$.
\item Data encoding: Convert those qubits whose corresponding pixel values are $1$ to $\ket{1}$ by applying Pauli$-X$ gates.
\item QNN: Apply parametrized $XX-$gates to (read-out qubit, $k^{th} qubit)$ pair followed by parametrized $ZZ-$gates.
\item Apply the Hadamard gate to the read-out qubit.
\item Perform measumrement on the read-out qubit, obtaining the expectation value $\bra{\psi}Z\ket{\psi}$.
\end{itemize}
When the composition $H \circ X$ is applied to $\ket{0}$, we get $\frac{\ket{0}-\ket{1}}{\sqrt{2}}$.
\begin{equation*}
\begin{aligned}
H \circ X \ket{0} &=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 & 1 \\
1 & -1 \end{bmatrix}
\begin{bmatrix}
0 & 1 \\
1 & 0
\end{bmatrix}
\begin{bmatrix}
1 \\ 0
\end{bmatrix} \\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 & 1 \\
1 & -1 \end{bmatrix}
\begin{bmatrix}
0 \\ 1
\end{bmatrix} \\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\ -1
\end{bmatrix} \\
&= \frac{1}{\sqrt{2}}
\left(
\begin{bmatrix}
1 \\ 0
\end{bmatrix}
-
\begin{bmatrix}
0 \\ 1
\end{bmatrix}
\right) \\
&= \frac{\ket{0}-\ket{1}}{\sqrt{2}}
\end{aligned}
\end{equation*}
For those qubits whose corresponding pixels values of the image matrix is not above the threshold value $0.5$, the quantum state with the ancilla qubit is
\begin{equation*}
\begin{aligned}
\frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{0} &=
\frac{1}{\sqrt{2}} \begin{bmatrix} 1 \\ -1 \end{bmatrix} \otimes \begin{bmatrix} 1 \\ 0 \end{bmatrix}\\
&= \frac{1}{\sqrt{2}} \begin{bmatrix} 1 \begin{bmatrix} 1 \\ 0 \end{bmatrix}\\ -1\begin{bmatrix} 1 \\ 0 \end{bmatrix} \end{bmatrix}\\
&=\frac{1}{\sqrt{2}} \begin{bmatrix}1 \\ 0 \\ -1 \\ 0 \end{bmatrix}
\end{aligned}
\end{equation*}
For those qubits whose corresponding pixels values of the image matrix is above the threshold value $0.5$, the Pauli$-X$ gate is applied and the state is changed from $\ket{0}$ to $\ket{1}$. Then the quantum state with the ancilla qubit is
\begin{equation*}
\frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{1}
= \frac{1}{\sqrt{2}} \begin{bmatrix}0 \\ 1 \\ 0 \\ -1 \end{bmatrix}
\end{equation*}
Thereafter, the circuit has a layer of parametrized $XX$ gates followed by a layer of parametrized $ZZ$ gates for each pair (readout qubit, $k^{th}$ qubit). Suppose the first qubit has its corresponding pixel value not exceeding the threshold, hence in state $\frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{0}$. Application of the $XX$ gate with parameter $t_0$ is then
\begin{equation*}
\begin{aligned}
& (X \otimes X)^{t_0} \left( \frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{0} \right) =\\
& \begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) & 0 & 0 &-i \sin \left( \frac{\pi t_0}{4} \right) \\
0 & \cos \left( \frac{\pi t_0}{2} \right) & -i \sin \left( \frac{\pi t_0}{2} \right) & 0 \\
0 & -i \sin \left( \frac{\pi t_0}{2} \right) & \cos \left( \frac{\pi t_0}{2} \right) & 0 \\
-i \sin \left( \frac{\pi t_0}{2} \right) & 0 & 0 &\cos \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}
\frac{1}{\sqrt{2}} \begin{bmatrix}1 \\ 0 \\ -1 \\ 0 \end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
-\cos{\left(\frac{\pi t_0}{2} \right)} \\
- i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix}
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}
\end{aligned}
\end{equation*}
Notice applying the $XX$ gate parametrized by $t_k$ does not change the state of the read-out qubit. For those qubits whose corresponding pixel values are 1, applicataion of the $XX-$gate gives us
\begin{equation*}
(X \otimes X)^{t_1} \left( \frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{1} \right) = \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix}
\otimes
\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
\cos \left( \frac{\pi t_1}{2} \right) \\
\end{bmatrix}
\end{equation*}
As the circuit repeats the $XX$ operation on each pair of (readout qubit, $k^{th}$ qubit), the state of the readout qubit remains the same, while the the state of the $k^{th}$ qubit is either $\cos \left( \frac{\pi t_0}{2} \right) \ket{0} + i \sin \left( \frac{\pi t_1}{2} \right) \ket{1}$ or $i \sin \left( \frac{\pi t_1}{2} \right) \ket{0} + \cos \left( \frac{\pi t_0}{2} \right) \ket{1}$ depending on its corresponding pixel value.
Denote the state of the $0^{th}$ qubit $\ket{\psi_0}$. Then the parametrized $ZZ$ gate with parameter $s_0$ is applied.
\begin{equation*}
\begin{aligned}
(Z \otimes Z)^{s_0} & \left(\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix}
\otimes
\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
\cos \left( \frac{\pi t_1}{2} \right) \\
\end{bmatrix} \right) \\
&=
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & e^{i \pi s_0} & 0 & 0 \\
0 & 0 & e^{i \pi s_0} & 0 \\
0 & 0 & 0 & 1
\end{bmatrix}
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
-\cos{\left(\frac{\pi t_0}{2} \right)} \\
-i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0} i \sin \left( \frac{\pi t_0}{2} \right) \\
-e^{i \pi s_0} \cos{\left(\frac{\pi t_0}{2} \right)} \\
-i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&=
\begin{bmatrix}
1 \\
-e^{i \pi s_0}
\end{bmatrix}
\otimes
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0}i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}
\end{aligned}
\end{equation*}
by the property of tensor products. Now the state of the readout qubit is $\ket{0}-e^{i \pi s_0} \ket{1}$. Then applying the $ZZ$ gate on the (readout qubit, $1^{st}$ qubit) pair gives us
\begin{equation*}
\begin{aligned}
(Z \otimes Z)^{s_1} & \left(
\begin{bmatrix}
1 \\
-e^{i \pi s_0}
\end{bmatrix}
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_1}{2} \right) \\
i \sin \left( \frac{\pi t_1}{2} \right) \\
\end{bmatrix} \right) \\
&=
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & e^{i \pi s_1} & 0 & 0 \\
0 & 0 & e^{i \pi s_1} & 0 \\
0 & 0 & 0 & 1
\end{bmatrix}
\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
\cos \left( \frac{\pi t_1}{2} \right) \\
-e^{i \pi s_0}i \sin \left( \frac{\pi t_1}{2} \right) \\
-e^{i \pi s_0}\cos{\left(\frac{\pi t_1}{2} \right)}
\end{bmatrix}\\
&=
\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
e^{i \pi s_1}\cos \left( \frac{\pi t_1}{2} \right) \\
-e^{i \pi s_1}e^{i \pi s_0}i \sin \left( \frac{\pi t_1}{2} \right) \\
-e^{i \pi s_0}\cos{\left(\frac{\pi t_1}{2} \right)}
\end{bmatrix}\\
&=
\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
e^{i \pi s_1}\cos \left( \frac{\pi t_1}{2} \right) \\
-e^{i \pi s_1}e^{i \pi s_0}i \sin \left( \frac{\pi t_1}{2} \right) \\
-e^{i \pi s_1}e^{i \pi s_1}e^{i \pi s_0}\cos{\left(\frac{\pi t_1}{2} \right)}
\end{bmatrix}\\
&=
\begin{bmatrix}
1 \begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
e^{i \pi s_1}\cos \left( \frac{\pi t_1}{2} \right) \\
\end{bmatrix}\\
-e^{i \pi s_1}e^{i \pi s_0}\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
e^{i \pi s_1}\cos \left( \frac{\pi t_1}{2} \right)
\end{bmatrix}
\end{bmatrix}\\
&=
\begin{bmatrix}
1 \\
-e^{i \pi s_1}e^{i \pi s_0}
\end{bmatrix}
\otimes
\begin{bmatrix}
i \sin \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_1} \cos \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}\\
\end{aligned}
\end{equation*}
As the circuit continues the $ZZ$ gate iteration on all $16$ qubits, the state of the readout qubit becomes
\begin{equation*}
\begin{bmatrix}
1 \\
- e^{i \pi s_{15}}e^{i \pi s_{14}} \ldots e^{i \pi s_0}
\end{bmatrix}
=
\begin{bmatrix}
1 \\
- e^{i \pi s_{15}+i \pi s_{14}+\ldots + i \pi s_0}
\end{bmatrix}
=
\begin{bmatrix}
1 \\
- e^{i \pi \sum_0^{15} s_k}
\end{bmatrix}
\end{equation*}
Then the circuit applies the Hadamard gate on the ancilla qubit while no operation is applied the data qubit. This is expressed as
\begin{equation*}
\begin{aligned}
H \otimes I &= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 & 1 \\
1 & -1
\end{bmatrix}
\otimes
\begin{bmatrix}
1 & 0 \\
0 & 1
\end{bmatrix}\\
&=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \begin{bmatrix}
1 & 0 \\
0 & 1
\end{bmatrix}&
1 \begin{bmatrix}
1 & 0 \\
0 & 1
\end{bmatrix}\\
1 \begin{bmatrix}
1 & 0 \\
0 & 1
\end{bmatrix}
& -1 \begin{bmatrix}
1 & 0 \\
0 & 1
\end{bmatrix}
\end{bmatrix} \\
&=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 & 0 &1 & 0\\
0 & 1 & 0 &1\\
1 & 0 &-1 & 0\\
0 & 1 & 0 &-1
\end{bmatrix}
\end{aligned}
\end{equation*}
Applying it to the previous state gives us the final state before measurement
\begin{equation*}
\begin{aligned}
(H \otimes & I)
\left(\begin{bmatrix}
1 \\
-e^{i \pi \sum_0^{15} s_j}
\end{bmatrix}
\otimes
\begin{bmatrix}
i \sin \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right) \\
\end{bmatrix}
\right)\\
&= \frac{1}{\sqrt{2}} \begin{bmatrix}
1 & 0 &1 & 0\\
0 & 1 & 0 &1\\
1 & 0 &-1 & 0\\
0 & 1 & 0 &-1
\end{bmatrix}
\begin{bmatrix}
i \sin \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right) \\
-e^{i \pi \sum_0^{15} s_j}i \sin \left( \frac{\pi t_0}{2} \right) \\
-e^{i \pi \sum_0^{15} s_j}e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right) \\
\end{bmatrix}\\
&=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
i \sin \left( \frac{\pi t_k}{2} \right) - e^{i \pi \sum_0^{15} s_j} i \sin \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
- e^{i \pi \sum_0^{15} s_j}
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)\\
i \sin \left( \frac{\pi t_k}{2} \right) + e^{i \pi \sum_0^{15} s_j} i \sin \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
+ e^{i \pi \sum_0^{15} s_j}
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
(1- e^{i \pi \sum_0^{15} s_j})i \sin \left( \frac{\pi t_k}{2} \right) \\
(1 - e^{i \pi \sum_0^{15} s_j})
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)\\
(1+ e^{i \pi \sum_0^{15} s_j})i \sin \left( \frac{\pi t_k}{2} \right) \\
(1 + e^{i \pi \sum_0^{15} s_j})
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 - e^{i \pi \sum_0^{15} s_j} \\
1 + e^{i \pi \sum_0^{15} s_j}
\end{bmatrix}
\otimes
\begin{bmatrix}
i \sin \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&=
\begin{bmatrix}
1 - e^{i \pi \sum_0^{15} s_j} \\
1 + e^{i \pi \sum_0^{15} s_j}
\end{bmatrix}
\otimes
\frac{1}{\sqrt{2}}
\begin{bmatrix}
i \sin \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}
\end{aligned}
\end{equation*}
by properties of tensor products. Then the final state of the readout qubit is $\begin{bmatrix}
1 - e^{i \pi \sum_0^{15} s_j} \\
1 + e^{i \pi \sum_0^{15} s_j}
\end{bmatrix}$
and of the $k^{th}$ image qubit $\begin{bmatrix}
i \sin \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_{k+1}} \cos \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}$
up to global phase $\frac{1}{\sqrt{2}}$.
The measurement operator outputs the expectation value of the Pauli$-Z$ gate $\bra{\psi} Z \ket{\psi}$ where $\ket{\psi}$ is the final state of the readout qubit. Then up to global phase
\begin{equation*}
\begin{aligned}
&\bra{\psi} Z \ket{\psi}\\
&=
\begin{bmatrix}
\left(1 - e^{i \pi \sum_0^{15} s_j}\right)^{\ast} &
\left(1 + e^{i \pi \sum_0^{15} s_j} \right)^{\ast}
\end{bmatrix}
\begin{bmatrix}
1 & 0 \\
0 & -1
\end{bmatrix}
\begin{bmatrix}
1 - e^{i \pi \sum_0^{15} s_j} \\
1 + e^{i \pi \sum_0^{15} s_j}
\end{bmatrix}\\
&=
\begin{bmatrix}
\left(1 - e^{i \pi \sum_0^{15} s_j}\right)^{\ast} &
\left(1 + e^{i \pi \sum_0^{15} s_j} \right)^{\ast}
\end{bmatrix}
\begin{bmatrix}
1 - e^{i \pi \sum_0^{15} s_j} \\
- \left( 1 + e^{i \pi \sum_0^{15} s_j} \right)
\end{bmatrix}\\
&=
\left(1 - e^{i \pi \sum_0^{15} s_j}\right)^{\ast}
\left(1 - e^{i \pi \sum_0^{15} s_j}\right) -
\left(1 + e^{i \pi \sum_0^{15} s_j}\right)^{\ast}
\left(1 + e^{i \pi \sum_0^{15} s_j}\right)\\
&=
\Vert 1 - e^{i \pi \sum_0^{15} s_j}\Vert^2 -
\Vert 1 + e^{i \pi \sum_0^{15} s_j}\Vert^2
\end{aligned}
\end{equation*}
Notice this is a real value. As we identify lower values to class $3$ and higher values to class $6$, we want
\[
\begin{aligned}
&\Vert 1 - e^{i \pi \sum_0^{15} s_j}\Vert^2 < \Vert 1 + e^{i \pi \sum_0^{15} s_j}\Vert^2\\
&\Rightarrow \Vert 1 - e^{i \pi \sum_0^{15} s_j}\Vert
< \Vert 1 + e^{i \pi \sum_0^{15} s_j}\Vert\\
\end{aligned}
\]
for class 3 and $\Vert 1 - e^{i \pi \sum_0^{15} s_j}\Vert^2 > \Vert 1 + e^{i \pi \sum_0^{15} s_j}\Vert^2$ for class 6.
The training process involves finding optimal values of $\theta_k$ and $\phi_k$. Notice that there are $16 \times 2=32$ parameters. The resulting quantum circuit is converted to a tensor, wrapped as a Keras layer, and is run using Hinge loss and Adam optimizer. The network achieves training $89.92\%$ accuracy. In order to build a multiclass classification network of 10 classes, we will need $Ceil \left(log_{2} 10\right)= 4$ ancilla qubits.
\href{}{}
\subsubsection{Qiskit PyTorch}
The IBM approach is to use a classical convolutional network with $1-$dimensional output and using it as the parameter for an $R_y$ gate on the uniform superposition of $\ket{0}$ and $\ket{1}$. Thereafter, measurement is performed without any additional transformation, hence the model can be viewed as a quantum data encoding model, not necessarily a neural network.
The Qiskit implementation of MNIST binary classifier is a classical-hybrid network using convoluional neural network for the classical component. The quantum component of the network is composed of one qubit on whose uniform superposition, a parametrized $R_y \left( \theta \right)$ is applied before measurement. The parameter $\theta$ of the $\R_y$ gate is the output from the classical convolutional neural network. In that sense, the quantum circuit in a strict sense is a quantum data encoding circuit, not a quantum neural network. The data encoding circuit is showing in the firgure.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{Figures/qiskit.pdf}
\caption{Qiskit PyTorch implementation of MNIST binary classifier}
\label{fig:tensorq}
\end{figure}
PyTorch and quantum encoding and measurement using Qiskit. The classical convolutional network is composed of two convolutional layers each with kernel size (5,5), max pool, and activation function Relu. The resulting image matrices are flattened, and then one hidden layer of feed-forward neural network is applied with activation function Relu. The output of the classical network is a vector of length one for each image and it is encoded into quantum state, by being used as the parameter for an $R_y$ gate. Then without any quantum neural network circuit, measurement is performed directly on the quantum encoded data.
The composition operation of the Hadamard gate and a parametrized $R_y$ gate is
\begin{equation*}
\begin{aligned}
R_y (\theta) \circ H \ket{0} &=
\begin{bmatrix}
\cos{\left(\frac{\theta}{2} \right)} & -\sin{\left(\frac{\theta}{2}\right)} \\
\sin{\left(\frac{\theta}{2}\right)} &
\cos{\left(\frac{\theta}{2} \right)}
\end{bmatrix}
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 & 1 \\
1 & -1
\end{bmatrix}
\begin{bmatrix}
1 \\
0
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos{\left(\frac{\theta}{2} \right)}-\sin{\left(\frac{\theta}{2}\right)} \\
\cos{\left(\frac{\theta}{2} \right)}+\sin{\left(\frac{\theta}{2}\right)}
\end{bmatrix}\\
\end{aligned}
\end{equation*}
To get classification $\ket{0}=\begin{bmatrix} 1 \\ 0\end{bmatrix}$, we want
\begin{equation*}
\cos{\left(\frac{\theta}{2}\right)}
-\sin{\left(\frac{\theta}{2}\right)} >
\cos{\left(\frac{\theta}{2} \right)}
+\sin{\left(\frac{\theta}{2}\right)}
\end{equation*}, which means
$-\sin{\left(\frac{\theta}{2}\right)} > \sin{\left(\frac{\theta}{2}\right)}$
Then we want $\frac{/theta}{2}$ to lie in the third or fourth quadrant of the unit circle: $\pi < \frac{\theta}{2} < 2\pi$, which gives us $2 \pi <\theta < 4 \pi$. By the same token, we want $0 < \theta < 2 \pi$ to get the label $1.$ The target range for the parameter $\theta$ as output from the classical network is very clear. Hence, most of heavy weight lifting is done by the classical network not by the quantum circuit.
For the computation of the gradient, the network employees the analytic differentiation method (finite difference formula), given by
\begin{equation*}
\resizebox{1\hsize}{!}
$
\begin{aligned}
\frac{dy}{d\theta}
&= meas\left(R_y \left( \theta + \delta \right)
\frac{1}{\sqrt{2}}
\begin{bmatrix} 1\\1
\end{bmatrix}\right)
-meas\left(R_y \left( \theta - \delta \right)
\frac{1}{\sqrt{2}}
\begin{bmatrix} 1\\1
\end{bmatrix}\right)\\
&= meas\left(
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos{ \left( \frac{\theta +\delta}{2}\right)}
-\sin{ \left( \frac{\theta +\delta}{2}\right)}\\
\cos{ \left( \frac{\theta +\delta}{2}\right)}
+\sin{ \left( \frac{\theta +\delta}{2}\right)}
\end{bmatrix}\right)
-meas\left(
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos{ \left( \frac{\theta -\delta}{2}\right)}
-\sin{ \left( \frac{\theta -\delta}{2}\right)}\\
\cos{ \left( \frac{\theta -\delta}{2}\right)}
+\sin{ \left( \frac{\theta -\delta}{2}\right)}
\end{bmatrix}\right)
\end{aligned}$
}
\end{equation*}
for some small $\delta \in \mathbb{R}$. Hence measurements of the same circuit are performed twice with slight variation of the parameter value $\theta$ by $\delta$ and the difference is considered as the gradient of the circuit parametrized by $\theta$. This method is called "phase shift". Optimization is carried out by the PyTorch in-built Adam optimizer with learning rate $0.001$. The model achieves $100\%$ training accuracy on 100 samples. Loss function NLLLoss
\href{https://qiskit.org/textbook/ch-machine-learning/machine-learning-qiskit-pytorch.html}{Online code}
\begin{equation*}
\begin{aligned}
& (X \otimes X)^{t_0} \left( \frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{0} \right) =\\
& \begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) & 0 & 0 &-i \sin \left( \frac{\pi t_0}{4} \right) \\
0 & \cos \left( \frac{\pi t_0}{2} \right) & -i \sin \left( \frac{\pi t_0}{2} \right) & 0 \\
0 & -i \sin \left( \frac{\pi t_0}{2} \right) & \cos \left( \frac{\pi t_0}{2} \right) & 0 \\
-i \sin \left( \frac{\pi t_0}{2} \right) & 0 & 0 &\cos \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}
\frac{1}{\sqrt{2}} \begin{bmatrix}1 \\ 0 \\ -1 \\ 0 \end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
-\cos{\left(\frac{\pi t_0}{2} \right)} \\
- i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix} \\
-1 \begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}\\
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix}
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}\\
&=
I \left( \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix} \right)
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) & i \sin \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) & \cos \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}
\begin{bmatrix}
1 \\
0
\end{bmatrix}\\
&=
I \left( \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix} \right)
\otimes
\begin{bmatrix}
\cos \left(- \frac{\pi t_0}{2} \right) & -i \sin \left( -\frac{\pi t_0}{2} \right) \\
-i \sin \left( -\frac{\pi t_0}{2} \right) & \cos \left(- \frac{\pi t_0}{2} \right) \\
\end{bmatrix}
\begin{bmatrix}
1 \\
0
\end{bmatrix}\\
&=
\left(I \otimes X^{-t_k} \right)
\left( \frac{\ket{0}-\ket{1}}{\sqrt{2}} \otimes \ket{0} \right)\
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
(Z \otimes Z)^{s_0} & \left(\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-1
\end{bmatrix}
\otimes
\begin{bmatrix}
i \sin \left( \frac{\pi t_1}{2} \right) \\
\cos \left( \frac{\pi t_1}{2} \right) \\
\end{bmatrix} \right) \\
&=
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & e^{i \pi s_0} & 0 & 0 \\
0 & 0 & e^{i \pi s_0} & 0 \\
0 & 0 & 0 & 1
\end{bmatrix}
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
i \sin \left( \frac{\pi t_0}{2} \right) \\
-\cos{\left(\frac{\pi t_0}{2} \right)} \\
-i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0} i \sin \left( \frac{\pi t_0}{2} \right) \\
-e^{i \pi s_0} \cos{\left(\frac{\pi t_0}{2} \right)} \\
-i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0} i \sin \left( \frac{\pi t_0}{2} \right) \\
-e^{i \pi s_0} \cos{\left(\frac{\pi t_0}{2} \right)} \\
-e^{i \pi s_0}e^{i \pi s_0}i \sin \left( \frac{\pi t_0}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0}i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}\\
-e^{i \pi s_0}\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0}i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}
\end{bmatrix}\\
&=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1 \\
-e^{i \pi s_0}
\end{bmatrix}
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0}i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}\\
&=
\begin{bmatrix}
1 \\
-e^{i \pi s_0}
\end{bmatrix}
\otimes
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_0}{2} \right) \\
e^{i \pi s_0}i \sin \left( \frac{\pi t_0}{2} \right) \\
\end{bmatrix}
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
&\ket{0} \bra{0} \otimes I + \ket{1} \bra{1} \otimes U
= \ket{0} \bra{0} \otimes I + \ket{1} \bra{1} \otimes e^{i \alpha} I\\
&=
\begin{bmatrix} 1 \\ 0 \end{bmatrix}
\begin{bmatrix} 1 & 0 \end{bmatrix}
\otimes
\begin{bmatrix} 1 & 0 \\0 & 1 \end{bmatrix}
+ \begin{bmatrix} 0 \\ 1 \end{bmatrix}
\begin{bmatrix} 0 & 1 \end{bmatrix}
\otimes
\begin{bmatrix} e^{i \alpha} & 0 \\0 & e^{i \alpha} \end{bmatrix} \\
&= \begin{bmatrix} 1 & 0 \\0 & 0 \end{bmatrix}
\otimes
\begin{bmatrix} 1 & 0 \\0 & 1 \end{bmatrix}
+
\begin{bmatrix} 0 & 0 \\0 & 1 \end{bmatrix}
\otimes
\begin{bmatrix} e^{i \alpha} & 0 \\0 & e^{i \alpha} \end{bmatrix} \\
&= \begin{bmatrix} 1\begin{bmatrix} 1 & 0 \\0 & 1 \end{bmatrix} &
0\begin{bmatrix} 1 & 0 \\0 & 1 \end{bmatrix} \\
0\begin{bmatrix} 1 & 0 \\0 & 1 \end{bmatrix} &
0\begin{bmatrix} 1 & 0 \\0 & 1 \end{bmatrix} \end{bmatrix}
+ \begin{bmatrix} 0\begin{bmatrix} e^{i \alpha} & 0 \\0 & e^{i \alpha} \end{bmatrix} &
0\begin{bmatrix} e^{i \alpha} & 0 \\0 & e^{i \alpha} \end{bmatrix} \\
0\begin{bmatrix} e^{i \alpha} & 0 \\0 & e^{i \alpha} \end{bmatrix} &
1\begin{bmatrix} e^{i \alpha} & 0 \\0 & e^{i \alpha} \end{bmatrix} \end{bmatrix}\\
&= \begin{bmatrix} 1 & 0 & 0 & 0\\ 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0\end{bmatrix}
+ \begin{bmatrix} 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 \\
0 & 0 & e^{i \alpha} & 0\\ 0 & 0 & 0 & e^{i \alpha}\end{bmatrix} \\
&=\begin{bmatrix} 1 & 0 & 0 & 0\\ 0 & 1 & 0 & 0 \\
0 & 0 & e^{i \alpha} & 0\\ 0 & 0 & 0 & e^{i \alpha}\end{bmatrix}\\
&=\begin{bmatrix} 1\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} &
0\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \\
0\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} &
e^{i \alpha}\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}\end{bmatrix} \\
&=\begin{bmatrix} 1 & 0 \\ 0 & e^{i \alpha} \end{bmatrix}
\otimes
\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
X^{\frac{t}{2}} \otimes X^{\frac{t}{2}} &= e^{\frac{i \pi t}{4}}
\begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) & -i \sin \left( \frac{\pi t}{4} \right) \\
-i \sin \left( \frac{\pi t}{4} \right) & \cos \left( \frac{\pi t}{4} \right)
\end{bmatrix}
\otimes
e^{\frac{i \pi t}{4}}
\begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) & -i \sin \left( \frac{\pi t}{4} \right) \\
-i \sin \left( \frac{\pi t}{4} \right) & \cos \left( \frac{\pi t}{4} \right)
\end{bmatrix}\\
&=
e^{\frac{i \pi t}{4}}e^{\frac{i \pi t}{4}}
\begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) \begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) & -i \sin \left( \frac{\pi t}{4} \right) \\
-i \sin \left( \frac{\pi t}{4} \right) & \cos \left( \frac{\pi t}{4} \right)
\end{bmatrix}
& -i \sin \left( \frac{\pi t}{4} \right) \begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) & -i \sin \left( \frac{\pi t}{4} \right) \\
-i \sin \left( \frac{\pi t}{4} \right) & \cos \left( \frac{\pi t}{4} \right)
\end{bmatrix} \\
-i \sin \left( \frac{\pi t}{4} \right)
\begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) & -i \sin \left( \frac{\pi t}{4} \right) \\
-i \sin \left( \frac{\pi t}{4} \right) & \cos \left( \frac{\pi t}{4} \right)
\end{bmatrix}
& \cos \left( \frac{\pi t}{4} \right)
\begin{bmatrix}
\cos \left( \frac{\pi t}{4} \right) & -i \sin \left( \frac{\pi t}{4} \right) \\
-i \sin \left( \frac{\pi t}{4} \right) & \cos \left( \frac{\pi t}{4} \right)
\end{bmatrix}
\end{bmatrix}
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
(Z \otimes Z)^{s_k} \ket{\psi_k} &=
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & e^{i \pi s_k} & 0 & 0 \\
0 & 0 & e^{i \pi s_k} & 0 \\
0 & 0 & 0 & 1
\end{bmatrix}
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
i \sin \left( \frac{\pi t_k}{2} \right) \\
-\cos{\left(\frac{\pi t_k}{2} \right)} \\
-i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right) \\
-e^{i \pi s_k} \cos{\left(\frac{\pi t_k}{2} \right)} \\
-i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right) \\
-e^{i \pi s_k} \cos{\left(\frac{\pi t_k}{2} \right)} \\
-e^{i \pi s_k}e^{i \pi s_k}i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}} \left(
\begin{bmatrix}
1 \\ -e^{i \pi s_k}
\end{bmatrix}
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix} \right)\\
&=
\begin{bmatrix}
1 \\ -e^{i \pi s_k}
\end{bmatrix}
\otimes
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}
\end{aligned}
\end{equation*}
Denote the state $\ket{\psi_k}$. Then the parametrized $ZZ$ gate with parameter $s_k$ is applied.
\begin{equation*}
\begin{aligned}
(Z \otimes Z)^{s_k} \ket{\psi_k} &=
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & e^{i \pi s_k} & 0 & 0 \\
0 & 0 & e^{i \pi s_k} & 0 \\
0 & 0 & 0 & 1
\end{bmatrix}
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
i \sin \left( \frac{\pi t_k}{2} \right) \\
-\cos{\left(\frac{\pi t_k}{2} \right)} \\
-i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right) \\
-e^{i \pi s_k} \cos{\left(\frac{\pi t_k}{2} \right)} \\
-i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right) \\
-e^{i \pi s_k} \cos{\left(\frac{\pi t_k}{2} \right)} \\
-e^{i \pi s_k}e^{i \pi s_k}i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}\\
&= \frac{1}{\sqrt{2}} \left(
\begin{bmatrix}
1 \\ -e^{i \pi s_k}
\end{bmatrix}
\otimes
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix} \right)\\
&=
\begin{bmatrix}
1 \\ -e^{i \pi s_k}
\end{bmatrix}
\otimes
\frac{1}{\sqrt{2}}
\begin{bmatrix}
\cos \left( \frac{\pi t_k}{2} \right) \\
e^{i \pi s_k} i \sin \left( \frac{\pi t_k}{2} \right)
\end{bmatrix}
\end{aligned}
\end{equation*}
\end{document}
\section{CONCLUSION}
In this pape, classical and CV quantum hybrid classifiers using different number of qumodes and cutoff dimensions were examined. The quantum gates available in the CV model allow a natural implementation of a quantum neural network layer $L(\ket{x})=\ket{\phi(Wx+b)}$. The flexibility of different measurement methods to output vectors of different lengths allows the networks to produce results that are interpreted as one-hot encoded labels of MNIST image dataset.
There is a limitation in encoding classical data into quantum states on near-term devices due to the number of qumodes, which is currently eight on Xanadu's X8 photonic QPU. Classical networks were used to reduce the number of entries in the image matrices for quantum data encoding. Although the role of the classical network is for pre-processing, the majority of parameters that are learned through the training process is on the classical network side. One way of encoding all the entries of an image matrix would be iterating through smaller sections of the image, which naturally segues to convolutional operations. A combination of quantum convolutional network layers and quantum neural network layers is a way of implementing a purely quantum network.
In implementing machine learning algorithms in quantum, the CV model offers many advantages over the qubit-based model. The quantum computational state space of the CV model is infinite while it is finite in the qubit-based model. The ability to define the dimension of the CV quantum computational space in approximating the original infinite computational space avails users of added freedom and flexibility in experimenting quantum algorithms. Photonic quantum computers, which are a version of physical implementation of the CV model can be easily incorportated into the current computing infrastructure because of their operability at room temperature.
\end{document}
\section{CV MNIST CLASSIFIERS}
The multi-classifier models presented in this section are inspired by the CV binary classifier architecture described in the section above \cite{cv_19}. They are run on Xanadu's $X8$ photonic quantum simulator, made up of $8-$qumodes. There are two models classifying 8 classes: digits $0, 1, \ldots, 7$ and five models classifying 10 classes: digits $0, 1, \ldots, 9$. The measurement outputs from these models are interpreted as one-hot encoded predictions of image labels.
The classifiers on 8 classes are:
\begin{tabular}[h]{c| ccc}
\hline
num of & cutoff & measurement & output \\
qumodes & dim & method & size \\ \hline
3 & 2 & probability & $2^3=8$ \\
8 & 2 & expectation value &$1 \times 8 = 8$ \\ \hline
\end{tabular}
A classical neural network is used as a pre-processing step to reduce the original image matrices of size $28 \times 28 = 784$ to vectors of lower length that the data encoding circuit can accommodate. For the data encoding circuit, squeezers, an interferometer, displacement gates, and Kerr gates are used. For the quantum neural network, the quantum circuit implementing $L(\ket{x})=\ket{\phi(Wx+b)}$ as in the binary classifier is used.
The 10-class classifiers are built on $2, 3, \ldots, 6-$qumodes. The label $k \in \{0, 1, \ldots, 9\}$ of an image matrix, when converted into a one-hot encoded vector, becomes a vector of length $10$ with all zeros but the $k^{th}$ entry as $1$. The cutoff dimensions are selected so that the output vectors exceed $10$ in length. For each classifier, a different number of zeros are padded to the one-hot encoded labels to match the output size of the circuit. The cutoff dimension used for each classifier and the size of output vectors are shown in the table.
\begin{tabular}[h]{c| ccc}
\hline
num of & cutoff &output & number of \\
qumodes & dim & size & padding 0's \\ \hline
2 & 4 & $4^2=16$ & 6\\
3 & 3 & $3^3=27$ & 17\\
4 & 2 & $2^4=16$ & 6\\
5 & 2 & $2^5=32$ & 22 \\
6 & 2 & $2^6=64$ & 54\\
\hline
\end{tabular}
The architecture in Figure 6 depicts the data flow of image matrix $\rightarrow$ classical layers $\rightarrow$ reduced output vectors $\rightarrow$ data encoding $\rightarrow$ QNN $\rightarrow$ measurement $\rightarrow$ output vectors as one-hot encoded labels.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Figures/architecture.pdf}
\caption{MNIST multi-classifier architecture}
\label{fig:arch}
\end{figure}
The boxes $U, S, D,$ and $K$ represent interferometers, a set of $m$ squeezers, a set of $m$ displacement gates, and a set of $m$ Kerr gates respectively, where $m=$the number of qumodes.
\subsection{Classical layers}
Classical feed forward neural networks are used for pre-processing of data images to reduce the size of $28 \times 28 = 784$ to smaller size vectors fitting the number of parameters available for data encoding. The image matrices are flattened to vectors of size $28 \times 28 = 784$ and then reduced to vectors of smaller sizes through Keras dense layer operations with activation function "ELU". The output vectors are then encoded as quantum states by the data encoding quantum circuit.
\subsection{Data encoding}
The output vectors from the classical neural network are in classical states. The quantum data encoding circuit converts classical states into quantum states. The quantum gates used for data encoding are squeezers, interferometer, displacement gates, and Kerr gates. The entries of a classical vector are used as the parameters of these parameterized quantum gates.
An interferometer is composed of beamsplitters on pairs of adjacent qumodes and rotation matrices. Squeezers $S(z)$ and displacement gates $D(\alpha)$ can be considered either one parameter gates or two parameter gates when we view the parameters $z, \alpha \in \mathbb{C}$ in Euler formula $z = a + bi = r \left(\cos{\phi} + i \sin{\phi} \right) = re^{i\phi}.$ For the purpose of data encoding, we use them as two-parameter gates.
The number of parameters that these gates can accommodate for $m-$qumode circuits are $8m - 2$, where $m$ is the number of qumodes.
\begin{tabular}[H]{ccccc}
\hline
squeezers & BS & rotation & displacement & Kerr \\ \hline
$2m$ & $2(m-1)$ & $m$ & $2m$ & $m$ \\ \hline
\end{tabular}
Based on these values, the size of the classical network output vectors was determined.
\subsection{Quantum circuit}
The QNN circuit implements a quantum version of the classical neural network
$L(x) = \phi(Wx+b)$ as
\[
L(\ket{x})=\ket{\phi(Wx+b)}=\phi \circ D \circ U_2 \circ S \circ U_i \ket{x}
\]
where $\ket{x}$ is a quantum data encoded state of the original data vector $x$, $U_k$ interferometers, $S$ squeezers, $D$ displacement gates, and $\phi(\cdot)$ Kerr gates respectively. The number of parameters per a collection of gates on $m-$qumodes are
\begin{tabular}[H]{cccc |c}
\hline
$U_k$ & $S$ & $D$ & $\phi(\cdot)$ & total \\ \hline
$2(m-1)+m$ & $m$ & $m$ & $m$ & $2 \cdot 2(m-1)+5m$\\
& & & & $= 9m - 4$\\ \hline
\end{tabular}
On the 8-qumode classifier, 2 layers of QNN are applied. On the rest of classifiers, 4 layers are applied. The number of parameters for the varying number of qumodes are shown in the table.
\begin{tabular}[H]{c|cc}
\hline
num of qumodes & per layer & total \\ \hline
2 & 14 & 56 \\
3 & 23 & 92 \\
4 & 32 & 128 \\
5 & 41 & 164 \\
6 & 50 & 200\\
8 & 68 & 136 \\ \hline
\end{tabular}
\subsection{Measurement}
Pennylane offers three different measurement methods for CV-based computations: expectation value, variance, and probabilities. The expectation value and variance methods yield a single-valued result for each qumode. The probability method yields vectors of size $n^m$, where $n=$cutoff dimension and $m=$the number of qumodes.
For the eight-qumode model, the expectation value method was used. The result is a vector of length $8: [\bra{\psi_0}X\ket{\psi_0}, \bra{\psi_1}X\ket{\psi_1}, \ldots, \bra{\psi_7}X\ket{\psi_7}]$ where $\bra{\psi_k}X\ket{\psi_k}$ represents the expectation value measurement result of the $k^{th}$ qumode. It is then translated as a one-hot encoded label vector of the corresponding image.
For the rest of the models, the probability method was used.
\subsection{Parameter update}
With the Pennylane Tensorflow plug-in functionality, the quantum circuit is converted into a Keras layer and added to the classical layers. Then Keras' built-in loss functions and optimizers can be used for parameter update. For most of the models, Categorical Crossentropy is used for loss function and Stochastic Gradient Descent is used for optimizer. For the 8-qumode classifier, the Mean Squared Error loss function performed better. The updated parameters are then used as the parameters of the quantum gates for the subsequent iteration of training.
\subsection{Experimetal results}
The $4-$qumode classifier yielded the best result of $100\%$ training accuracy on $600$ data samples. Accuracy comparison with the qubit-based binary classifiers using Tensorflow Quantum and Qiskit is listed in the table below.
\scalebox{0.72}{
\begin{tabular}{ |c|c|c|c|}
\hline
& Tensorflow Q & Qiskit - PyTorch & Pennylane - Keras \\ \hline
Number of classes & 2 & 2 & 10 \\ \hline
Number of samples & 10,338 & 100 & 600 \\ \hline
Training accuracy & $89.92\%$ & $100\%$ & $100\%$ \\ \hline
\end{tabular}
\label{Tab:MNIST}}
All the classifiers tested achieve above $95\%$ training accuracy.
For the 8-qumode classifier, 300 samples and 2 layers of QNN were used with 50 epochs. For the rest of the classifiers, 600 samples and 4 layers of QNN were used. With the 4-qumode classifier, training accuracy of $100\%$ is achieved in 70 epochs.
The loss and accuracy for the models with varying number of qumodes are listed below.
\begin{tabular}[h!]{c| ccc}
\hline
num of & learning & training & training \\
qumodes & rate & loss & accuracy \\ \hline
2 & 0.02 & 0.4966 & $97.14\%$ \\
$3^{\ast}$ & 0.05 & 0.0563 & $99.90\%$ \\
3 & 0.05 & 0.4567 & $98.04\%$\\
4 & 0.03 & 0.1199 & $100.00\%$\\
5 & 0.02 & 0.2057 & $98.40\%$ \\
6 & 0.02 & 0.2055 & $98.22\%$\\
8 & 0.1 & 0.0199 & $99.77\%$ \\\hline
\end{tabular}
The second line $3^{\ast}$ indicates an $8-$class classifier.
All of these classifiers followed the typical loss and accuracy graphs depicted in Figure 7. The validation accuracy is not as high as the training accuracy, indicating "over-fitting". Further experiments with different hyper-parameters or the use of regularizer are called for.
\href{https://github.com/sophchoe/Continous-Variable-Quantum-MNIST-Classifiers}{Code: Keras-Pennylane implementation}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Figures/MNIST.pdf}
\caption{MNIST multi classifier experimental results}
\label{fig:MNIST_exp}
\end{figure}
\end{document}
\section{INTRODUCTION}
\justifying
Unlike the original assumption that quantum computers would replace classical computers, quantum processing units (QPUs) are emerging as task-specific special purpose processing units much like Graphics Processing Units. The currently available QPUs are called near-term quantum devices because they are not yet fully fault-tolerant and are characterized by shallow and short quantum circuits. Nonetheless, the availability of these devices allows for active research on quantum algorithms specific for these devices, especially in quantum chemistry, Gaussian boson sampling, graph optimization, and quantum machine learning.
The QPUs are based on two different theoretical models of quantum computing: the discrete variable (qubit-based) model and the continuous variable (CV) model \cite{qi_12}. The discrete variable model is an extension of the computational space from $\{0,1\}$ to a complex projective Hilbert space $\mathbb{C}P^1$, which is the surface of a projective sphere. The computational basis is composed of two elements $\{\ket{0}, \ket{1}\}$ \cite{qi_10}. The CV model is an extension of the computational space to an infinite dimensional Hilbert space whose computational basis is infinite, i.e., $\{\ket{0}, \ket{1}, \ldots, \ket{n}, \ldots\}$ \cite{qi_12}. The quantum state of an information channel (qumode) is represented by an infinite complex projective linear combination of these basis states. In practice, it is approximated by by a finite linear combination using only a finite basis $\{\ket{0}, \ket{1}, \ldots, \ket{n}$. The number of basis states used for the approximation is called cutoff dimension. On an $m-$qumode system with cutoff dimension $n$, the computational state space is of dimension $n^m$. By varying cutoff dimension and the number of qumodes used for computation, we can control the dimension of the computational space.
Implementing machine learning algorithms on quantum computers is an active areaa of research. Quantum machine learning algorithms can be implemented on variational quantum circuits with parametrized quantum gates \cite{circuit_18}. A quantum circuit is a collection of quantum gates and the change of states induced by the circuit on the initial quantum state is considered quantum computation. The results of quantum computing extracted via measurement are incorporated into optimization and parameter update computations performed on classical circuits \cite{circuit_18}.
Quantum neural networks (QNN), a subset of quantum machine learning, follow the same architectural frame work: QNN on a QPU and optimization on a CPU. The main components of classical neural networks described as $L(x)=\phi(Wx+b)$ are the affine transformation $Wx+b$ and the non-linear activation function $\phi(\cdot)$. In the qubit model, all the available unitary gates are linear in nature hence a direct implementation of bias addition and non-linear activation function is not feasible. In the CV model, the displacement gate implements bias addition and the Kerr gate, non-linear activation function, hence a direct translation of $L(x)=Wx+b$ to its quantum version $L(\ket{x})= \ket{\phi(Wx+b)}$ is naturally embedded in the model.
The proposed CV MNIST classifiers are built on Xanadu's X8 simulator, which simulates quantum computation on an $8-$qumode photonic quantum computer. The ability to control the size of output vectors based on the number of qumodes and the notion of cutoff dimension allows for producing one-hot encoded labels of MNIST dataset. Different architectures on $2-$qumodes, $3-$qumodes, up to $8-$qumodes are introduced. They are classical-quantum hybrid networks with classical feed-forward neural networks, quantum data encoding circuits, and CV quantum neural network circuits according to the CV binary classifier proposed by Killoran et al \cite{cv_19}. The quantum machine learning software library, Pennylane \cite{Penny}, is used for the quantum circuit and Tensorflow Keras is used for the classical circuit and optimization. The classifiers achieve above $95\%$ training accuracy.
This paper is organized in the following manner: In Section 2, the CV model of quantum computing is discussed, especially the infinite dimensionality of the computation state space and how the notion of of cutoff dimension allows us to determine the dimension of the actual computational state. In Section 3, we examine how CV quantum neural networks naturally implement classical neural networks. In Section 3.2, the architecture of the binary classifier in "Continuous variable quantum neural networks" is examined \cite{cv_19}. In Section 4, the structure of the 8 hybrid classifiers and the experimental results are presented.
\end{document}
\section{CV QUANTUM COMPUTING}
\begin{figure}[b]
\centering
\includegraphics[scale=0.8]{Figures/Fock_basis.pdf}
\caption{Fock basis: Gaussian states \href{https://upload.wikimedia.org/wikipedia/commons/7/71/Wignerfunction_fock_0.png}{image source}}
\label{fig:Fock}
\end{figure}
The quantum state space of the CV model is an infinite dimensional Hilbert space, which is approximated by a finite dimensional space for practical information processing purposes. Under phase space representation of the model, the computational basis states are Gaussian states as opposed to single points. Therefore the quantum gates taking these Gaussian states to Gaussian states offer richer arrays of transformations than just unitary (linear) transformations as in the qubit model.
The CV model is based on the wave-like property of nature. It uses the quantum state of the electromagnetic field of bosonic modes (qumodes) as information carriers \cite{qi_cv_03}. Its physical implementation is done using linear optics, containing multiple qumodes \cite{optics_01}. Each qumode contains a certain number of photons probablistically, which can be manipulated with quantum gates.
A photon is a form of electromagnetic radiation. The position wave function $\Psi(x)$ describes the light wave electromagnetic strength of the qumode depending on the number of photons present. It is a complex valued function on real valued variables: $\Psi: \mathbb{R} \rightarrow \mathbb{C}: x \mapsto \alpha$, where $x$ is the position(s) of the photons. The wave function with more than one photon displays the constructive and destructive interactions between the light waves of the photons.
Phase space representation of the quantum state of a linear optical system describes the state using the position $x$ and momentum $p$ variables of the photons in the given qumode system. The quasi-probablity distribution of $x$ and $p$ are plotted on the $xp-$plane, using the Wigner function. It is given by
\[
W(x,p)=\frac{~{p}}{h}=\frac{1}{h} \int_{-\infty}^{\infty}e^{-\frac{ipy}{\hbar}}\Psi \left(x+\frac{y}{2}\right)\Psi^{\ast}\left(x-\frac{y}{2}\right) dy
\]
where $h=6.62607015 \times 10^{-34}$ is the Plank constant and $\hbar = 6.582119569 \times 10^{-16}$ the reduced Plank constant \cite{kitten_08}.
This Wigner function is applied to the position wave function $\Psi_k(x)$, where $k$ denotes the number of photons present in a qumode, to create Fock basis, also known as number basis. The image of the Wigner functions $W_0(x_0,p_0), W_1(x_1,p_1), \ldots, W_5(x_5, p_5)$ where $x_k$ and $p_k$ represent the position and momentum variables with $k$ photons present in the system is shown in Figure 1.
\newpage
They are used as computational basis states expressing the quantum state of a system. The quantum state $\ket{\psi}$ of a qumode is expressed as a superposition of Fock basis states:
\[
\ket{\psi}=c_0 \ket{0}+c_1 \ket{1}+\hdots+c_n \ket{n}+\hdots, \hspace{2pt} \sum_{k=0}^{\infty} \Vert c_k \Vert ^2=1
\]
where $c_k$ is the probability amplitude of basis $\ket{k}$.
In practice, there are not going to be an infinite number of photons physically present in a qumode. We approximate $\ket{\psi}$ with $\ket{\hat{\psi}}$ by cutting off the trailing terms. The number of Fock basis states we use to approximate the true state $\ket{\psi}$ is called "cutoff dimension". Let $n$ be cutoff dimension. Then the approximating state $\ket{\hat{\psi}}$ is in superposition of $\ket{0}, \ket{1}, \ldots, \ket{n-1}$, i.e.,
\[
\ket{\psi}=c_0 \ket{0}+c_1 \ket{1}+\hdots+c_{n-1} \ket{n-1} \text{where} \sum_{k=0}^{n-1} \Vert c_k \Vert ^2=1.
\]
In vector representation, the state is expressed as a column vector of size $n=$cutoff dimension, with the $k^{th}$ entry being $c_k$.
A multi-qumode system is represented by the tensor product of individual qumode states: $\ket{\psi_0} \otimes \ket{\psi_1} \otimes \ldots \otimes \ket{\psi_{m-1}}$, where $m$ is the number of qumodes. When these states are approximated by cutoff dimension $n$, the computational basis of the resulting quantum state is of size $n^m$.
\[
\begin{aligned}
& \ket{\psi_0} \otimes \ket{\psi_1} \otimes \ldots \otimes \ket{\psi_{m-1}}=\\
&d_0 \ket{00 \ldots 0} + d_1 \ket{00 \ldots 1} + d_{n^m-1}\ket{n-1,n-1,\ldots,n-1}
\end{aligned}
\]
Fock basis states constitute the eigenstates of the number operator $\hat{n} := \hat{a}^{\dagger}\hat{a}$ where $\hat{a}^{\dagger}$ is called the constructor and $\hat{a}$ is called the annihilator \cite{qi_12}. Their matrix representation is given by
\[
\hat{a}^{\dagger} =
\begin{bmatrix}
0 & 0 & 0 & \hdots & 0 & 0\\
\sqrt{1} & 0 & 0 & \ldots & 0 & 0\\
0 & \sqrt{2} & 0 & \ldots & 0 & 0\\
0 & 0 & \sqrt{3} & \ldots & 0 & 0\\
\vdots & & \ldots & \ddots & & \vdots\\
0 & 0 & 0 & \ldots & \sqrt{n-1}& 0\\
\end{bmatrix}
\text{and}
\]
\[
\hat{a} =
\begin{bmatrix}
0 & \sqrt{1} & 0 &0 & \ldots & 0\\
0 & 0 & \sqrt{2}& 0 & \ldots & 0\\
0 & 0 & 0& \sqrt{3} & \ldots & 0\\
\vdots & & \ldots & \ddots & & \vdots\\
0 & 0 & 0 & 0 &\ldots & \sqrt{n-1}\\
0 & 0 & 0 & 0 &\ldots & 0\\
\end{bmatrix}
\]
Note that the constructor $\hat{a}^{\dagger} \ket{k} = \sqrt{k+1} \ket{k+1} \text{ for }k \ge 0$ indeed constructs the next level Fock basis state and the annihilator $\hat{a} \ket{0} = 0$, $\hat{a} \ket{k} = \sqrt{k} \ket{k-1} \text{ for }k \ge 1$ annihilates the the current level Fock basis state to the lower.
The product of $\hat{a}^{\dagger}$ and $\hat{a}$ returns a matrix
\[
\begin{bmatrix}
0 & 0 & 0 &0 & \ldots & 0\\
0 & 1 & 0 & 0 & \ldots & 0\\
0 & 0 & 2& 0 & \ldots & 0\\
\vdots & & \ldots & \ddots & & \vdots\\
0 & 0 & 0 & 0 & n-2 & 0\\
0 & 0 & 0 & 0 & 0 & n-1
\end{bmatrix}
\]
which we denote as "number operator" $\hat{n}$. The Fock basis states $\ket{k}$ are indeed eigenstates of the number operator as shown in $\hat{n}\ket{k} = k \ket{k}$ where $k \in \{ 0, 1, \ldots, n-1 \}$ for cutoff dimension$=n$.
Standard Gaussian gates in the CV model are of the form $U= e^{-iHt}$ where $H$ represents the Hamiltonian of the system, describing the total energy as the sum of kinetic and potential energy. In the finite quantum state space of a qumode approximated by cutoff dimension $n$, this infinite sum representing a quantum gate is also approximated by a finite matrix exponential.
The matrix exponential $e^{-iHt}$ is of the form
\[
\hat{U} = \sum_{k=0}^{n-1}\frac{({-iHt})^k}{k!}=I -iHt + \frac{(-iHt)^2}{2!}+ \hdots + \frac{{-iHt}^{n-1}}{(n-1)!}
\]
Some of the standard Gaussian gates taking Gaussian states to Gaussian states are listed below.
Squeezer with parameter $z: S(z) = exp \left( \frac{z^{\ast} \hat{a}^2+z \hat{a}^{{\dagger}^2}}{2} \right)$
Rotation with parameter $\phi: R(\phi) = exp\left(i \phi \hat{a}^{\dagger}\hat{a} \right)$
Displacement with parameter $\alpha: D(\alpha)=exp\left(\alpha \hat{a}^{\dagger}-\alpha^{\ast} \hat{a} \right)$
Another Gaussian gate, the beamsplitter, is a two-qumode gate with parameters $\theta$ and $\phi$:
\[
B(\theta, \phi) =
exp\left( \theta \left( e^{i\phi} \hat{a} \hat{b}^{\dagger} + e^{-i \phi}
\hat{a}^{\dagger}\hat{b} \right) \right)
\]
where $\hat{b}^{\dagger}$ and $\hat{b}$ are constructor and annihilator of the 2nd qumode respectively.
Measurement is done via counting the number of photons present in each qumode with a photon detector. Xanadu's Pennylane offers a suite of measurement methods as outlined in Table 1. The size of the output vectors are affected by the number of qumodes used for computation and the notion of cutoff dimension, denoted by $n$. Currently available QPU by Xanadu, $X8$, offers $8$ qumodes.
\begin{table}
\centering
\begin{tabular}[b]{lcc}
\hline
Measurement method & output size & output size \\
& per qumode & for $m-$qumodes \\ \hline
expectation value & $1$ & $m$\\
variance & $1$ & $m$\\
probability & $n$ & $n^m$\\
\hline
\end{tabular}
\end{table}
The measurement methods used in the proposed MNIST classifiers are the expectation value method and the probability memthod.
The expectation value method produces a single real-valued output. Let $\ket{\psi} = \begin{bmatrix} \psi_0 \\ \psi_1 \end{bmatrix}$ be the quantum state of a multi-qumode system after desired quantum computational operations are performed. The expectation value measurement method returns $\bra{\psi} A \ket{\psi}$ where the operator $A$ is usually the Pauli$-X$, Pauli$-Y$, or Pauli$-Z$ gate. The expectation value of the Pauli$-X$ matrix is
\[
\begin{aligned}
\bra{\psi} X \ket{\psi} &=
\begin{bmatrix} \psi_0^{\ast} & \psi_1^{\ast} \end{bmatrix}
\begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}
\begin{bmatrix} \psi_0 \\ \psi_1 \end{bmatrix} \\
&=
\begin{bmatrix} \psi_0^{\ast} & \psi_1^{\ast} \end{bmatrix}
\begin{bmatrix} \psi_1 \\ \psi_0 \end{bmatrix}\\
&=
\psi_0^{\ast} \psi_1 + \psi_1^{\ast} \psi_0 \\
&=
2\left( Re(\psi_0)Re(\psi_1) + Im(\psi_0)Im(\psi_1) \right)\in \mathbb{R},
\end{aligned}
\]
which is a real number. In a multi-qumode system with $m$ qumodes, we can get a vector of size $m$ by getting the expectation value for each qumode
The probabilities method returns the probability of each computation basis state. Suppose an $m-$qumode system has cutoff dimension $n$ for each qumode. Then, there are $n^m$ computational basis states, hence we get a vector of size $n^m$.
\end{document}
|
\section{Introduction}
\label{sec:intro}
Capsule Networks (CapsNets)~\cite{hinton-dynamic,hinton-em,hinton-stacked-caps} were recently introduced to overcome the shortcomings of Convolutional Neural Networks (CNNs).
CNNs loose the spatial relationships between its parts because of max pooling layers, which progressively drop spatial information~\cite{hinton-dynamic}. Furthermore, CNNs are also commonly known as ``black-box'' models: most of the techniques providing interpretation over the model are \textit{post-hoc}: they produce localized maps that highlight important regions in the image for predicting objects~\cite{gradcam}. CapsNets attempt to preserve and leverage an image representation as a hierarchy of parts, \textit{carving-out} a parse tree from the networks. This is possible thanks to the iterative routing mechanism~\cite{hinton-dynamic} which models the connections between capsules. This can be seen as a parallel attention mechanism, where each active capsule can choose a capsule in the layer above to be its parent in the tree~\cite{hinton-dynamic}. Therefore, CapsNets can produce explainable representations encoded in the architecture itself~\cite{hinton-dynamic} yet can be still successfully applied to a number of applicative tasks\cite{zhao20193d,paoletti2018capsule,afshar2018brain}.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{figures/CapsNet+REM_new.pdf}
\caption{\centering{With REM, a Capsule Network builds stronger relationships with lower entropy between capsules.\label{fig:capsnet_vs_remcaps}}}
\end{figure}
\indent However, understanding what really happens inside a CapsNet is still an open challenge. Indeed, for a given input image, there are too many active (and each-other coupled) capsules, making the routing algorithm connections still difficult to understand, as the coupling coefficients typically have each-other similar values (Fig.~\ref{fig:capsnet_vs_remcaps}), non exploiting the routing algorithm potential and making architectures without routing performing comparably~\cite{gu}. Furthermore, backward and forward passes of a CapsNet come at an enormous computational cost, since the number of trainable parameters is very high. For example, the CapsNet model deployed on the MNIST dataset by Sabour et al.\cite{hinton-dynamic} is composed by an encoder and a decoder part. The full architecture has $8.2$M of parameters.
Do we really need that amount of trainable parameters to achieve competitive results on such a task? Recently, many pruning methods were applied to CNNs in order to reduce the complexity of the networks, enforcing sparse topologies~\cite{enzo-neurips, sparsevd, louizos}: is it possible to tailor one of these approaches with not only the purpose of lowering the parameters, but aiding the model's interpretability?
\indent This work introduces REM (Routing Entropy Minimization) for CapsNets, which aims to improve the explainabilty of the routing algorithm of CapsNets. Pruning can effectively reduce the overall entropy of the connections of the parse tree-like structure encoded in a CapsNet, because in low pruning regimes it removes noisy couplings which cause the entropy to increase considerably.
We collect the coupling coefficients studying their frequency and cardinality, observing lower per-class conditional entropy on these: the pruned version adds a missing explicit prior in the routing mechanism, grounding the coupling of the unused primary capsules disallowing fluctuations under the same baseline performance on the validation/test set. This implies that the parse trees are more stable for the pruned models, providing higher explainability on the relevant features (per-class) selected from the routing mechanism.
\indent This paper is organized as follows: in Section \ref{sec:sota} we introduce some of the basic concepts of CapsNets and their related works, in Section \ref{sec:method} we describe our techinque called REM, in Section \ref{sec:results} we investigate the effectivess of our method testing it on many datasets. The last two sections discuss the limitations of our approach and the conclusion of our work.
\section{Background and Related Work}
\begin{figure*}[b]
\centering
\includegraphics[width=1.5\columnwidth]{figures/capsnet_rem_architecture.pdf}
\caption{Capsule Network architecture (encoder network).\label{fig:capsnet_encoder}}
\end{figure*}
\label{sec:sota}
This section first describes the fundamental aspects of CapsNets and their routing algorithm introduced by \cite{hinton-dynamic}. Then, we review the literature especially related to how introduce sparsity into CapsNets. The notation used in our work is summarized in Table~\ref{tab:notation}.
\begin{table}
\renewcommand{\arraystretch}{1.2}
\footnotesize
\centering
\begin{tabular}{ll}
\toprule
\textbf{Symbols} & \textbf{Meaning}\\ \midrule
$\Xi$ & Cardinality of the dataset\\\hline
$M, N$ & Spatial dimensions of the PrimaryCaps layer\\\hline
$T $ & Primary capsules types\\\hline
$I $ & Primary capsules \\ \hline
$J$ & Output (digit) capsules \\\hline
$\hat{y}^\xi$ & Target class for the $\xi$-th sample \\\hline
$y^\xi$ & Predicted class for the $\xi$-th sample \\\hline
$\theta$ & Generic parameter for the model\\\hline
$\boldsymbol{u}_i$ & Pose of the $i$-th primary capsule (normalized) \\\hlin
${\boldsymbol{s}}_{j}$ & Pose of the $j$-th output capsule (non-normalized)\\\hline
$\boldsymbol{v}_j$ & Pose of the $j$-th output capsule (normalized)\\\hline
$c_{ij}^{\xi}$ & \makecell[l]{Coupling coefficient for the $\xi$-th input linking\\ the $i$-th primary capsule to the $j$-th output capsule}\\\hline
$\boldsymbol{c}_{-j}$ & Couplings to the same $j$-th output caps\\\hline
$\boldsymbol{W}$ & Transformation matrix\\ \cline{1-2}
$\mathbb{1}_{\hat{y}^\xi}$ & One-hot encoding for the target class $y^\xi$\\ \cline{1-2}
$\Theta(\cdot)$ & One-step function\\ \cline{1-2}
$q_K(\cdot)$ & Uniform quantizer on $K$ levels\\ \cline{1-2}
$\tilde{x}$ & Quantized representation of $x$\\ \cline{1-2}
$\mathbb{P}(x)$ & Probability of $x$ \\\cline{1-2}
$\mathbb{H}_j$ & Entropy of the couplings toward $j$-th output capsule \\ \cline{1-2}
$\mathbb{H}$ & Entropy of the couplings for the CapsNet\\
\bottomrule
\end{tabular}
\caption{Summary of the notation.}
\label{tab:notation}
\end{table}
\\\indent\textbf{Capsule Networks Fundamentals.}
CapsNets group neurons into \textit{capsules}, namely activity vectors, where each capsule accounts for an object of one of its parts. Each element of these vectors accounts for different properties of the object such as its pose and other properties like color, deformation, etc.~\cite{hinton-dynamic}. The magnitude of a capsule stands for the probability of existence of that object in the image~\cite{hinton-dynamic}. Fig.~\ref{fig:capsnet_encoder} shows a standard architecture of a CapsNet with two capsule layers, PrimaryCaps and DigitCaps (also called OutputCaps).
The poses of $L$-th capsules $\boldsymbol{u_i}$, called \textit{primary capsules}, are built upon convolutional layers. In order to compute the poses of the capsules of the next layer $L+1$, an iterative routing mechanism is performed. Each capsule $\boldsymbol{u_i}$ makes a prediction $\hat{\boldsymbol{u}}_{{j|i}}$, thanks to a transformation matrix $\boldsymbol{W}_{ij}$, for the pose of an upper layer capsule $j$
\begin{equation}
\hat{\boldsymbol{u}}_{{j|i}} = \boldsymbol{W}_{{ij}}\boldsymbol{u}_{i}.
\end{equation}
Then, the total input ${\boldsymbol{s}}_{j}$ of capsule $j$ of the DigitCaps layer is computed as the weighted average of votes $\hat{{\boldsymbol{u}}}_{j|i}$
\begin{equation}\label{eq:sj}
{\boldsymbol{s}}_{j} = \sum\limits_{i} c_{ij} \hat{{\boldsymbol{u}}}_{j|i},
\end{equation}
where $c_{ij}$ are the coupling coefficients between a primary capsule $i$ and an output capsule $j$.
The pose ${\boldsymbol{v}}_j$ of an output capsule $j$ is then defined as the normalized ``squashed'' $s_j$
\begin{equation}\label{eq:squashing}
{\boldsymbol{v}}_{j} = \textrm{squash}({\boldsymbol{s}}_{j}) = \frac{\|{\boldsymbol{s}}_{j}\|^{2}}{1 + \|{\boldsymbol{s}}_{j}\|^{2}} \frac{{\boldsymbol{s}}_{j}}{\|{\boldsymbol{s}}_{j}\|}.
\end{equation}
So the routing algorithm computes the poses of output capsules and the connections between capsules of consecutive layers. The coupling coefficients are computed dynamically by the routing algorithm and they are dependent on the input~\cite{hinton-dynamic}.
The coupling coefficients are determined by a ``routing softmax'' activation function, whose initial logits $b_{ij}$ are the log prior probabilities the $i$-th capsule should be coupled to the $j$-th one
\begin{equation}\label{eq:softmax}
c_{ij} = \textrm{softmax}(b_{ij}) = \frac{e^{b_{ij}}}{\sum\limits_{k}. e^{b_{ik}}} \\
\end{equation}
At the first step of the routing algorithm they are equals and then they are refined by measuring the agreement between the output ${\boldsymbol{v}}_{j}$ of the $j$-th capsule and the prediction $\hat{{\boldsymbol{u}}}_{j|i}$ for a given input. The agreement is defined as the scalar product $a_{ij} = {\boldsymbol{v}}_{j} \cdot \hat{{\boldsymbol{u}}}_{j|i}$.
At each iteration, the update rule for the logits is
\begin{equation}\label{eq:update-rule}
b_{ij} \gets b_{ij} + a_{ij}.
\end{equation}
The steps defined in \eqref{eq:sj}, \eqref{eq:squashing}, \eqref{eq:softmax}, \eqref{eq:update-rule} are repeated for the $t$ iterations of the routing algorithm. In order to train the network, Sabour~\emph{et~al.}~\cite{hinton-dynamic} replaced the cross entropy loss with the \textit{margin loss}.
\subsection{Related work}
\indent\textbf{Capsule Networks.} They were first introduced by Sabour~\emph{et~al.}~\cite{hinton-dynamic} and since then a lot of work has been done, both to improve the routing mechanism and to build deeper models. Regarding the routing algorithm, Hinton~\emph{et~al.}~\cite{hinton-em} replace the dynamic routing with Expectation-Maximization, adopting matrix capsules instead of vector capsules. Wang~\emph{et~al.}~\cite{Wang2018AnOV} model the routing strategy as an optimization problem. Li~\emph{et~al.}~\cite{nn-encaps} use master and aide branches to reduce the complexity of the routing process. Hahn~\emph{et~al.}~\cite{self-routing} incorporates a \textit{self-routing} method such that capsule models do not require agreements anymore. De Sousa Ribeiro~\emph{et~al.}~\cite{desousa} replace the routing algorithm with variational inference of part-object connections in a probabilistic capsule network, leading to a significant speedup without sacrificing performance. Since the CapsNet model introduced by Sabour~\emph{et~al.}~\cite{hinton-dynamic} is a shallow network, several works attempted to build deep CapsNets. Rajasegaran~\emph{et~al.}~\cite{deepcaps} propose a deep capsule network architecture which uses a novel 3D convolution based dynamic routing algorithm aiming at improving the performance of CapsNets for more complex image datasets. Gugglberger~\emph{et~al.}~\cite{res-caps} introduce residual connections to train deeper capsule networks.
\indent\textbf{Sparse Capsule Networks.} A naive solution to reduce uncertainty within the routing algorithm is to simply running more iterations. As shown by Paik~\emph{et~al.}~\cite{paik} and Gu~\emph{et~al.}~\cite{gu}, the routing algorithms tends to overly polarize the link strengths, namely a simple route in which each input capsule sends its output to only one capsule and all other routes are suppressed. On the one hand, this behavior is desirable because it makes the routing algorithm more explainable, by making it possible to extract a parse tree thanks to this coupling coefficients. On the other hand, running many iterations is only useful in the case of networks with few parameters, as demonstrated by Renzulli~\emph{et~al.}~\cite{renzulli}, otherwise the performance will drop. Rawlinson ~\emph{et~al.}~\cite{sparse-caps} trained CapsNets in an unsupervised setting, showing that in unsupervised learning the routing algorithm does not discriminate among capsule anymore: the coupling coefficients collapse to the same value. Therefore, they sparsify latent capsule layers activities by masking output capsules according to a custom ranking function.
Kosiorek~\emph{et~al.}~\cite{hinton-stacked-caps} impose sparsity and entropy constraints into capsules, but they do not employ an iterative routing mechanism. Jeong~\emph{et~al.}~\cite{ladder-caps} introduced a structured pruning layer called ladder capsule layers, which removes irrelevant capsules, namely capsules with low activities. Kakillioglu~\emph{et~al.}~\cite{3D-Caps} solve the task of 3D object classification on point clouds with pruned Capsule Networks. Their objective was to compress robust capsule models in order to deploy them on resource-constrained devices.
The main contribution of our work relies on the fact that we regularize and prune the parameters in a CapsNet as a way to minimize the entropy of the connections computed by the routing algorithm. In fact, we show that relationships between objects and their parts in a standard CapsNets described by Sabour et al.~\cite{hinton-dynamic} have high entropies. We minimize these entropies so that we can extract stronger parse trees. This allows us to effectively build dictionaries upon the input datasets and understand which are the shared object parts and transformations between different the entities in the images.
\section{Routing Entropy Minimization}
\label{sec:method}
The coupling coefficients computed by the routing mechanism model the part-whole relationships between capsules of two consecutive capsule layers, as displayed in Fig.~\ref{fig:capsnet_vs_remcaps}. Assigning parts to objects (namely learning how each object is composed), is a challenging task.
One of the main goal of the routing algorithm is to extract a parse tree of these relationships.
Given the $\xi$-th input of class $j$, an ideal parse tree for a primary capsule $i$ detecting one of the parts of the entity in the input $\xi$ would ideally lead to
\begin{equation}
\boldsymbol{c}_{i-}^\xi = \mathbb{1}_{\hat{y}^\xi}.
\end{equation}
This means that the routing process is able to carve a parse tree out of the CapsNet which explains perfectly the relationships between parts and wholes.
One of the problems of this routing procedure is that there is no constraint on how strong a parse tree should be.
In this section we present our techinque REM, first showing how to extract a parse tree and then how to build stronger parse trees.
The pipeline of our method is depicted in Fig.~\ref{fig:pipeline}.
\begin{figure*}
\centering
\includegraphics[scale=0.8]{figures/pipeline.pdf}
\caption{Pipeline\label{fig:pipeline} of REM.}
\end{figure*}
\subsection{Parse trees extraction} \label{sec:parse-tree-extraction}
Once we have a trained CapsNets model, in order to interpret the routing mechanism, we extract all the possible routing coupling coefficients and build a parse tree. Towards this end, we want to define a metric which helps us deciding if the relationships captured by the routing algorithm resemble or not a parse tree. Therefore, we organize the coupling coefficients into \textit{associative arrays}, we can compute the number of occurrences of each coupling sequence in order to compute the entropy of the whole coupling dictionary to measure the simplicity of the parse tree.
In the next paragraphs, we explain how to generate these sequences by discretizing the coupling coefficients and how to create the dictionary.
\\\indent \textbf{Quantization.} During the quantization stage, we first compute the \textit{continuous} coupling coefficients $c_{ij}^{\xi}$ for each $\xi$-th input example. It should be noticed that $c_{ij}^{\xi}$ are the coupling coefficients obtained after the forward pass of the last routing iteration. Then, we quantize every $c_{ij}^{\xi}$ into $K$ discrete levels through the uniform quantizer $q_K(\cdot)$, obtaining
\begin{equation}
\tilde{c}_{ij}^{\xi} = q_K({c}_{ij}^{\xi}).
\end{equation}
We choose the lowest $K$ such that the accuracy is not deteriorated. We will here on refer to CapsNet+Q as trained CapsNet where the coupling coefficients are quantized.
\indent\textbf{Extracting the parse tree.} Given the quantized coupling coefficients of a CapsNet+Q, we can extract the parse tree (and create a dictionary of parse trees) for each class $j$, where each entry is a string composed by the quantization indices of the coupling coefficients. We will extract the coupling coefficients $\boldsymbol{\tilde{c}}_{-j}^\xi$ between the primary capsules $I$ and the predicted $j$-th output capsule.
Given a dictionary for the coupling coefficients of a CapsNet+Q, we can compute the entropy for each class as
\begin{equation}
\label{eq::Hj}
\mathbb{H}_j = - \sum\limits_{\xi}\left\{ \mathbb{P}(\boldsymbol{\tilde{c}}_{-j}^\xi \mid y^\xi = j) \cdot \log_2 \left[\mathbb{P}(\boldsymbol{\tilde{c}}_{-j}^\xi \mid y^\xi = j)\right]\right\}
\end{equation}
where $\mathbb{P}(\boldsymbol{\tilde{c}}_{-j}^\xi \mid y^\xi = j)$ is the frequency of occurrences of a generic string $\xi$ for each \emph{predicted} class $y^\xi$.
Finally, the entropy of a dictionary for a CapsNet+Q on a given dataset is the average of the entropies $H_j$ of each class
\begin{equation}
\label{eq::H}
\mathbb{H} = \frac{1}{J}\sum\limits_{j} \mathbb{H}_j.
\end{equation}
Intuitively, the lower \eqref{eq::H}, the stronger the connections between capsules are in the parse tree-like structure carved-out from the routing algorithm.
We also target to obtain the distribution of these coupling coefficients. In general, we know we have $\Xi \times I \times J$ coupling coefficients for the full dataset (with potential redundancies). Given the $i$-th primary capsule, however, we are only interested to $\boldsymbol{c}_{-j}^{\xi}|y^\xi=j$. In this way, we reduce the coupling coefficients space to $I \times J$. We compute then the average of all the inputs belonging to an object class in order to output just $I \times J$ coupling coefficients.
\\\indent\textbf{Preliminary result.} Let us now generate the parse trees on the CapsNet model trained on the MNIST dataset as in ~\cite{hinton-dynamic}, investigating their evolution during training.
To this end, we compute the distribution of the coupling coefficients on the test set for two models: the model trained after the first epoch and the one that achieved the lowest loss value on the validation set.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figures/cij_distributions_unpruned.pdf}
\caption{Coupling coefficients distributions of two CapsNets+Q at epochs 1 and 190 on MNIST (test set).\label{fig:cij_nopruning}}
\end{figure}
Fig.~\ref{fig:cij_nopruning} shows the distributions for the two models and Table~\ref{tab:acc_nopruing} reports the corresponding accuracy and entropy values. We sampled a subset of the distributions for all the object classes.
It can be observed that after the first epoch CapsNet is clearly far from optimality, both in term of performance and parse tree explainability: indeed all coupling coefficients are almost all equals to the value selected at initialization, i.e. $1/J$.
This effect is almost the same when we train a CapsNet with $t=1$ as Gu et al.~\cite{gu}, where its entropy is exactly zero but capsules are uniformly
coupled.
On the other hand, a well trained CapsNet achieves higher generalization but its entropy is getting higher, so the routing mechanism is difficult to interpret and to explain. As a matter of fact, we can see in Table~\ref{tab:dictionary_capsnetq} the dictionary for the $\text{CapsNet}_{190}\text{+Q}$: the number of keys for each class is too high to make the routing algorithm explainable. Notice that the MNIST test set has 10k images and the total number of keys is 8147.
We target a model that has high generalization but a dictionary where number of keys is low.
\begin{table}
\footnotesize
\centering
\begin{tabular}{ccc}
\hline
\textbf{Model} & \textbf{Accuracy} & \textbf{Entropy} \\ \hline
$\text{CapsNet}_{1}\text{+Q}$ & $97.53$ & $1.83$ \\
$\text{CapsNet}_{190}\text{+Q}$ & $99.55$ & $9.41$ \\ \hline
\end{tabular}
\caption{\centering\label{tab:acc_nopruing}Classification results (\%) and entropy of two CapsNets+Q at epochs 1 and 190 on MNIST (test set).}
\end{table}
\begin{table*}[t]
\footnotesize
\centering
\begin{tabular}{*{13}{c}}
\hline
\textbf{Model} & $\mathbf{\textbf{\#}_0}$ & $\mathbf{\textbf{\#}_1}$ & $\mathbf{\textbf{\#}_2}$ & $\mathbf{\textbf{\#}_3}$ & $\mathbf{\textbf{\#}_4}$ & $\mathbf{\textbf{\#}_5}$ & $\mathbf{\textbf{\#}_6}$ & $\mathbf{\textbf{\#}_7}$ & $\mathbf{\textbf{\#}_8}$ & $\mathbf{\textbf{\#}_9}$ \\ \hline
$\text{CapsNet}_{190}\text{+Q}$ & $974$ & $358$ & $968$ & $967$ & $859$ & $856$ & $939$ & $670$ & $965$ & $591$\\ \hline
\end{tabular}
\caption{\centering\label{tab:dictionary_capsnetq}Dictionary sizes for the $\text{CapsNet}_{190}\text{+Q}$ model on MNIST (test set).}
\end{table*}
\subsection{Unconstrained routing entropy}
In this subsection we are going to more-formally analyze the distribution of the coupling coefficients
\begin{equation}
\label{eq::standardcij}
c_{ij} = \frac{e^{b_{ij} + \sum_{r=1}^t \boldsymbol{v}_j^r \boldsymbol{u}_{j} \boldsymbol{W}_{ij}}}{\sum_k e^{b_{ik} + \sum_{r=1}^t \boldsymbol{v}_k^r \boldsymbol{u}_k \boldsymbol{W}_{ik}}}
\end{equation}
where $t$ indicates the target routing iterations.\footnote{for abuse of notation, in this subsection we suppress the index $\xi$} Let us evaluate the $c_{ij}$ over a non-yet trained model: as we saw also in Section \ref{sec:parse-tree-extraction}, we have
\begin{equation}
\label{eq::uniformcij}
c_{ij} \approx \frac{1}{J}\ \forall i,j.
\end{equation}
When updating the parameters, following~\cite{gu}, we have
\begin{align}
\frac{\partial L}{\partial \boldsymbol{W}_{ij}} =& \left[\frac{\partial L}{\partial \boldsymbol{v}_j} \frac{\partial \boldsymbol{v}_j}{\partial \boldsymbol{s}_j} \cdot c_{ij} +\right. \nonumber\\
&+\left. \sum_{m=1}^M \left( \frac{\partial L}{\partial \boldsymbol{v}_m} \frac{\partial \boldsymbol{v}_m}{\partial \boldsymbol{s}_m} \cdot \boldsymbol{\hat{u}}_{m|i}\frac{\partial c_{im}}{\boldsymbol{\hat{u}}_{m|i}}\right) \right]\cdot \boldsymbol{u}_i
\end{align}
where we can have the gradient for $\boldsymbol{W}_{ij}\approx 0$ in a potentially-high number of scenarios, despite $c_{ij}\neq\{0, 1\}$. Let us analyze the simple case in which we have perfect outputs, matching the ground truth, hence we are close to a local (or potentially the global) minimum of the loss function:
\begin{equation}
\left\|\frac{\partial L}{\partial \boldsymbol{v}_m} \right\|_2 \approx 0\ \forall m.
\end{equation}
Looking at \eqref{eq:softmax}, we see that the right class is chosen, but given the squashing function, we have as an explicit constraint that, given the $j$-th class as the target one, we require
\begin{equation}
\label{eq::constraint}
\left\|\boldsymbol{v}_j\right\|_2 \gg \left\|\boldsymbol{v}_m\right\|_2\ \forall m\neq j
\end{equation}
on the $\boldsymbol{W}_{ij}$, which can be accomplished in many ways, including:
\begin{itemize}
\item having sparse activation for the primary capsules $\boldsymbol{u}_i$: in this case, we have constant $\boldsymbol{W}_{ij}$ (typically associated to no-routing based approaches); however, we need heavier deep neural networks as they have to force sparse signals already at the output of the primary capsules. In this case, the coupling coefficients $c_{ij}$ are also constant by definition;
\item having sparse votes $\boldsymbol{\hat{u}}_{j|i}$: this is a combination of having both primary capsules and weights $\boldsymbol{W}_{ij}$ enforcing sparsity in the votes, and the typical scenario with many routing iterations.
\end{itemize}
Having sparse votes, however, does not necessarily result in having sparse coupling coefficients: according to \eqref{eq:update-rule}, the coupling coefficients are multiplied with the votes, obtaining the output capsules. The distribution of the coupling coefficients requires \eqref{eq::constraint} to be satisfied only: if $\boldsymbol{W}_{ij}$ is not sparsely distributed, we can still have sparse votes. However, this is the main reason we observe high entropy in the coupling coefficient distributions (Fig.~\ref{fig:cij_nopruning}): as the votes $\boldsymbol{\hat{u}}_{j|i}$ are implicitly sparse (yet also disordered, as we are not explicitly imposing any structure in the coupling coefficients distribution), the model is still able to learn but it finds a typical solution where $c_{ij}$ are not sparse. However, we would like to have sparsely distributed, recurrent $\boldsymbol{c}_{-j}$, establishing stable relationships between the features extracted at primary capsules layer.\\
Minimizing explicitly the entropy term \eqref{eq::Hj} is an intractable problem due to the non-differentiability of the entropy term and of the quantization step (in our considered setup) and due to the huge computational complexity to be introduced at training time. Hence, we can try to implicitly enforce routing entropy minimization by forcing a sparse and organized structure in the coupling coefficients. Towards this end, one efficient solution relies in enforcing sparsity in the $\boldsymbol{W}_{ij}$ representation: enforcing a vote between the $i$-th primary capsule and the $j$-th output caps to be exactly zero for any input, according to \eqref{eq::standardcij}
\begin{equation}
c_{ij} = \frac{1}{\sum_k e^{b_{ik} + \sum_{r=1}^t \boldsymbol{v}_k^r \boldsymbol{u}_k \boldsymbol{W}_{ik}}}.
\end{equation}
In this way, having a lower variability in the $c_{ij}$ values (and hence building more stable relationships between primary and output capsules), straightforwardly we are also explicitly minimizing the entropy of the quantized representations for the coupling coefficients. In the next subsection, we are going to tailor a sparsity technique to accomplish such a goal.
\subsection{Enforcing REM with pruning}
CapsNets are trained via standard back-propagation learning, minimizing some loss function like margin loss. Our ultimate goal is to assess to which extent a variation of the value of some parameter $\theta$ would affect the error on the network output. In particular, the parameters not affecting the network output can be pushed to zero in a soft manner, meaning that we can apply an $\ell_2$ penalty term. A number of approaches have been proposed, especially in the recent years~\cite{louizos2017learning,molchanov2019importance,lee2018snip}. One recent state-of-the-art approach, LOBSTER~\cite{lobster} proposes to penalize the parameters by their gradient-weighted $\ell_2$ norm, leading to the update rule
\begin{align}
\theta^{t+1} = &\theta^t - \eta G\left[\frac{\partial L}{\partial \theta^t}\right] - \lambda \theta^t \textrm{ReLU}\left[1- \left | \frac{\partial L}{\partial \theta^t}\right| \right],
\label{eq:oleupdateRule}
\end{align}
where $G\left[\frac{\partial L}{\partial \theta^t}\right]$ is any gradient-based optimization update (for SGD it is the plain gradient, but other optimization strategies like Adam can be plugged) and $\eta, \lambda$ are two positive hyper-parameters.\\
Such a strategy is particularly effective on standard convolutional neural networks, and easy to plug in any back-propagation based learning system. Furthermore, LOBSTER is a regularization strategy which can be plugged at any learning stage, as it self-tunes the penalty introduced according to the learning phase: for this non-intrusiveness in the complex and delicate routing mechanism for CapsNets, it resulted a fair choice to enforce REM.
\section{Experiments and Results}
\label{sec:results}
In this section we report the experiments and the results that we performed to test REM. We first show the results on the MNIST dataset, reporting also how the entropy and the accuracy values change during training. Then, we test REM on more complex datasets such as Fashion-MNIST, CIFAR10 and Tiny ImageNet. We also performed experiments to test the robustness to affine transformations of CapsNets+REM. Except for Tiny ImageNet, we used the same architectures configurations and augmentations described in \cite{hinton-dynamic}.\footnote{we have removed the decoder part of the network, see Sec.~\ref{sec:limitations} for more details.}
We trained models with five random seeds. We report the classification accuracy (\%) and entropy (averages and standard deviations), the \textit{sparsity} (percentage of pruned parameters, median) and the number of keys in the dictionary (median).
The code will be open-source released upon acceptance of the paper, to guarantee the reproducibility of the results. The experiments were run on a Nvidia Ampere A40 equipped with 48GB RAM, and the code uses PyTorch~1.9
\subsection{Preliminary results on MNIST}
In order to assess our REM technique, we analyze in-depth the benefits of pruning towards REM on the MNIST dataset. Nowadays, despite its outdatedness, MNIST remains an omni-present benchmark for CapsNets~\cite{hinton-dynamic, gu, sparse-caps}.
\begin{figure}[h]
\includegraphics[width=1.0\columnwidth]{figures/plot_entropy_pp.pdf}
\captionsetup{justification=centering}
\caption{\centering{Accuracy and entropy curves vs pruned parameters on MNIST (test set).}}
\label{fig:acc_vs_pp}
\end{figure}
Fig.~\ref{fig:acc_vs_pp} shows how the entropy (red line) and classification accuracy (blue dotted line) changes as the sparsity increases during training. We can see that at the beginning of the training stage the entropy is low because the routing algorithm has not learned yet to correctly discriminate the relationships between the capsules; but at the end of the training process REM lowers the entropy of the network.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figures/cij_distributions_prunedvsunpruned.pdf}
\caption{\centering{Coupling coefficients distributions for each class of a CapsNets+Q and a CapsNet+REM on MNIST (test set).\label{fig:capsnetq_vscapsnetrem}}}
\end{figure}
In Fig.~\ref{fig:capsnetq_vscapsnetrem} we plot the distributions of the coupling coefficients for a CapsNet+Q and a CapsNet+REM following the method described in Sec.~\ref{sec:parse-tree-extraction}. We can see that the distributions of the CapsNet+REM model are sparser that the ones for the CapsNet+Q model, namely we carve a stronger parse tree while achieving high generalization. Table~\ref{tab:mnist-pruned-params} shows that there is no performance loss when pruning a CapsNet, even when REM is applied with its quantization levels.
\begin{table*}
\footnotesize
\centering
\begin{tabular}{*{3}{c}}
\hline
\textbf{Model} & \textbf{Accuracy} & \textbf{Sparsity} \\ \hline
CapsNet & $99.57\pm_{0.0002}$ & $0$ \\
CapsNet & $99.58\pm_{0.0003}$ & $85.53$ \\\hline
CapsNet+Q & $99.56\pm_{0.0003}$ & $0$ \\
CapsNet+REM & $99.56\pm_{0.0002}$ & $85.53$ \\\hline
\end{tabular}
\caption{\centering\label{tab:mnist-pruned-params}Results for CapsNets, CapsNets+Q, CapsNets+REM on MNIST (test set).}
\end{table*}
\begin{table*}[]
\footnotesize
\centering
\begin{tabular}{*{13}{c}}
\hline
\textbf{Model} & \textbf{Sparsity} & $\mathbf{\textbf{\#}_0}$ & $\mathbf{\textbf{\#}_1}$ & $\mathbf{\textbf{\#}_2}$ & $\mathbf{\textbf{\#}_3}$ & $\mathbf{\textbf{\#}_4}$ & $\mathbf{\textbf{\#}_5}$ & $\mathbf{\textbf{\#}_6}$ & $\mathbf{\textbf{\#}_7}$ & $\mathbf{\textbf{\#}_8}$ & $\mathbf{\textbf{\#}_9}$ & \textbf{Entropy} \\ \hline
CapsNet+Q & 0 & $972$ & $489$ & $1007$ & $985$ & $904$ & $856$ & $939$ & $670$ & $968$ & $957$ & $9.53\pm_{0.54}$\\
CapsNet+REM & 85.53 & $40$ & $46$ & $15$ & $39$ & $12$ & $117$ & $182$ & $14$ & $310$ & $9$ & $4.16\pm_{1.59}$\\ \hline
\end{tabular}
\caption{\centering\label{tab:mnist-dictionary-test}(Quantized) CapsNets number of keys for each class on MNIST (test set).}
\end{table*}
Table~\ref{tab:mnist-dictionary-test} shows the number of keys in each class for CapsNets+REM and a CapsNets+Q, namely a CapsNet where the quantization is applied without pruning the network during training. We also report the number of keys for each class and the entropy of the dictionary for the same quantized models shown in Table~\ref{tab:mnist-pruned-params}. We can see that the dimension of the dictionary for CapsNets+REM is lower than the one for CapsNets+Q. Also the entropy measure for CapsNets+REM is lower compared to CapsNets+Q, namely, REM has successufully built a parse tree between primary and class capsules with stronger relationships. We now show ho we can improve the explainability of such connections. Each key can be seen as a \textit{cluster} of images with shared features or parts. For each key of the predicted class, we take all the images of the input dataset that belong to that key, sorted according to the size of that cluster. Then, we overlap all the images in a cluster. Figure \ref{fig:mnist_keys} depicts a visual representation of some of these clusters for a subset of object classes. We can see that a CapsNet+REM can not only predict which digit is present in an input image, but it can also show which are the shared object parts (such as different localized parts for the digit 7 and 9) and their transformations (rotation and translation for the digit 0) in a cluster.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{figures/mnist_heatmaps.pdf}
\caption{\centering{Each digit is a cluster composed by the input images. Shared object parts and transformations are extracted with the dictionary of a CapsNet+REM.}}
\label{fig:mnist_keys}
\end{figure}
\subsection{Results}
We trained and tested CapsNets+REM on more complex datasets such as:
\begin{itemize}
\item Fashion-MNIST, 28x28 grayscale images (10 classes);
\item CIFAR10, 32x32 RGB images (10 classes);
\item Tiny ImageNet, 64x64 RGB images (200 classes).
\end{itemize}
\begin{table*}[]
\centering
\footnotesize
\begin{tabular}{cccccccccc}
\toprule
& \multicolumn{3}{c}{\textbf{F-MNIST}} & \multicolumn{3}{c}{\textbf{CIFAR10}} & \multicolumn{3}{c}{\textbf{T-ImageNet}} \\
\textbf{Model} & \textbf{Accuracy} & \textbf{\begin{tabular}[c]{@{}c@{}}Sparsity\end{tabular}} & \textbf{Entropy} & \textbf{Accuracy} & \textbf{\begin{tabular}[c]{@{}c@{}}Sparsity\end{tabular}} & \textbf{Entropy} & \textbf{Accuracy} & \textbf{\begin{tabular}[c]{@{}c@{}}Sparsity\end{tabular}} & \textbf{Entropy} \\ \midrule
CapsNet+Q & $92.46\pm_{0.002}$ & $0$ & $8.64\pm_{1.15}$ & $78.42\pm_{0.026}$ & $0$ & $6.26\pm_{0.61}$ & $58.50\pm_{0.25}$ & $0$ & $5.18\pm_{0.67}$\\
CapsNet+REM & $92.62\pm_{0.001}$ & $80.71$ & $4.80\pm_{1.70}$ & $79.25\pm_{0.005}$ & $81.17$ & $4.15\pm_{0.62}$ & $54.02\pm_{0.14}$ & $44.27$ & $3.15\pm_{0.81}$ \\ \bottomrule
\end{tabular}\caption{\centering\label{tab:results-otherdatasets}Accuracy, sparsity and entropy results on on Fashion-MNIST, CIFAR10 and Tiny ImageNet (test set).}
\end{table*}
\begin{table*}[]
\footnotesize
\centering
\begin{tabular}{ccccc}
\toprule
& \textbf{expanded MNIST} & \multicolumn{3}{c}{\textbf{affNIST}} \\
\textbf{Model} & \textbf{Accuracy} & \textbf{Accuracy} & \textbf{\begin{tabular}[c]{@{}c@{}}Sparsity\end{tabular}} & \textbf{Entropy} \\ \hline
CapsNet+Q & $99.22$ & $77.93\pm_{0.005}$ & $0$ & $8.64\pm_{1.15}$ \\
CapsNet+REM & $99.22$ & $81.81\pm_{0.008}$ & $71.26$ & $8.45\pm_{1.10}$ \\ \hline
CapsNet+Q & $99.36\pm_{0.0005}$ & $83.14\pm_{0.002}$ & $0$ & $8.45\pm_{0.99}$ \\
CapsNet+REM & $99.48\pm_{0.0002}$ & $85.23\pm_{0.001}$ & $87.32$ & $5.93\pm_{1.39}$ \\ \hline
\end{tabular}\caption{\centering\label{tab:results-affnist}Accuracy, sparsity and entropy results on affNIST test set for under and well-trained models.}
\end{table*}
\indent \textbf{Specific setup for Tiny ImageNet.} In order to test also in transfer learning scenarios, we train CapsNets on the Tiny ImageNet dataset, using in this case only a pretrained ResNet18 model as back-bone and fine-tuning the CapsNet parameters only. We used 10\% of the training set as validation set and the original validation set as test set.\\%We also changed to number of discretization levels to 128.
\indent \textbf{Discussion.} As we can see in Tables \ref{tab:results-otherdatasets}, a CapsNet+REM has an high percentage of pruned parameters with a minimal performance loss. So this confirms our hypothesis that CapsNets are over-parametrized.
Furthermore, for the Fashion-MNIST dataset the entropy of the CapsNet+REM is much lower that the CapsNet+Q, while for the other datasets the differences between the two approaches is less evident. This is an expected behavior: for CIFAR10 and Tiny ImageNet, both CapsNets achieved poor classification results and this introduces noise on the coupling coefficients, leading to higher uncertainty in the classification task hence in the features selection.
\indent \textbf{Robustness to affine transformations.} To test the robustness to affine transformations of CapsNets+REM, we used expanded MNIST: a dataset composed by padded and translated MNIST, in which each example is an MNIST digit placed randomly on a black background of 40×40 pixels. We used the affNIST dataset as test set, in which each example is an MNIST digit with a random small affine transformation. We tested an under-trained CapsNet with early stopping which achieved 99.22\% accuracy on the expanded MNIST test set as in ~\cite{hinton-dynamic, gu}. We also trained these models until convergence.
We can see in Table~\ref{tab:results-affnist} that the under-trained networks entropies are high. Instead, a well-trained CapsNet+REM can be robust to affine transformations and have a low entropy.
\subsection{Limitation}\label{sec:limitations}
A CapsNet is typically composed of an encoder and a decoder part, where the latter is a reconstruction network with 3 fully connected layers \cite{hinton-dynamic}. In the previously-discussed experiments, we have removed the decoder. The main limitation of our work arise when computing the entropy of CapsNets trained with the decoder. We observed that the entropy of a CapsNets+REM is almost the same as that of a CapsNet+Q shown in Tables \ref{tab:mnist-dictionary-test} and \ref{tab:results-otherdatasets}.
Indeed, when the decoder is used, the activity vector of an output capsule encodes richer representations of the input.
Sabour \textit{et al}.~\cite{hinton-dynamic} introduced the decoder to boost the routing performance on MNIST by enforcing the pose encoding a capsule. They also show that, when a perturbed activity vector is feeded to the decoder, such perturbation affects the reconstruction. So capsules representations are \emph{equivariant}, meaning that transformations applied to the input are described by continuous changes in the output vector. In order to verify if output capsules of a trained CapsNet+REM without the decoder (so with low entropy) are still equivariant, we stacked on top of it the reconstruction network, without training the encoder. We can see in Fig.~\ref{fig:reconstruction} that the output capsules are still equivariant to many transformations so we successfully obtained a CapsNet+REM with low entropy but with equivariant representations encoded into capsules.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figures/mnist_reconstructions.pdf}
\caption{MNIST perturbation reconstructions of a freezed CapsNet+REM with a decoder stacked on top of it.}
\label{fig:reconstruction}
\end{figure}
\section{Conclusion}
\label{sec:conlcusion}
This paper investigated and improved the explainability of the routing algorithm in CapsNets with REM (Routing Entropy Minimization), which drives the model parameters distribution towards low entropy configurations. We first showed how to extract the parse tree of a CapsNet by discretizing its connections and then collecting the possible parse trees in associative arrays.
Standard CapsNets show high entropy in the parse trees structures, as an explicit prior on the coupling coefficients distribution is missing. We showed how pruning methods, in low pruning regimes, naturally reduce such entropy as well as the cardinality over the possible parse trees, testing such a phenomenon on several datasets. Furthermore, we empirically observe that a CapsNet+REM model remains robust to affine transformations and its output capsules are still equivariant. REM opens research pathways towards the distillation of parse trees and model interpretability, including the design of a pruning technique specifically-designed for REM. |
\section{Introduction}\label{intro-sec}
It is well known that the $N=1$ superconformal superspace, in its complexified version \cite{ma1,va}, is the superflag $Fl(2|0,2|1,4|1)$, on which the conformal supergroup $\rSL(4|1)$ acts naturally. The space $\C^{4|1}$,
{underlying}
the defining representation of
$\rSL(4|1)$, is the space of supertwistors \cite{pe,fe}.
Dealing with the complexified version has the advantage of seeing this structure, while the {conditions for the} real form can be imposed later on \cite{va}. It is also well known, and differently from the non super case,
that not all the superflags are projective superspaces (take for example the super Grassmannian $Gr(1|1, 2|2)$ \cite {ma1}) and indeed the projective cases are rare among these superspaces, {though a new approach to this question was taken
in \cite{sv}.}
For the super Grassmannians only the extreme cases
$Gr(p|0, m|n)$ or $Gr(p|n, m|n)$ are superprojective and are both embedded into the projective superspace {$\mathbb{P}^{M|N}$ for suitable $M$ and $N$
see \cite{fl}}. These super Grassmannians are dual to each other and are the antichiral and chiral superspaces respectively.
The superflag $Fl(2|0,2|1,4|1)$ can be embedded in the product
$$Fl(2|0,2|1,4|1)\subset Gr(2|0,4|1)\times Gr(2|1,4|1)\,,$$ and using the super Segre embedding \cite{flln} the superflag is embedded into the projective superspace $\mathbb{P}^{80|64}$ \cite{cfl,fl}.
For $N=2$ we can reproduce the same situation with
$$Fl(2|0,2|2,4|2)\subset Gr(2|0,4|2)\times Gr(2|2,4|2)\,,$$ but this superflag is too big. The scalar superfields associated to it have too many field components to be useful in the formulation of supersymmetric field theories. Still, the antichiral $Gr(2|0,4|2)$ and chiral $Gr(2|2,4|2)$ superspaces do have physical applications so it is useful to study them. They are both embedded in $\mathbb{P}^{8|8}$.
There is a third super Grassmannian, $Gr(2|1,4|2)$, which is not projective but that has physical applications. It is the harmonic superspace of \cite{gikos,gios}. This example, together with more general ones, were studied from this point of view in the series of papers \cite{hl,hh,hah,hho,heho,he,haho}.
Here we will consider only the (anti)chiral superspace (also considered in \cite{me}). Our aim is to quantize it by substituting the supergroup $\rSL(4|2)$ by the quantum group $\rSL_q(4|2)$ (in the sense of Manin \cite{ma2}) and trying to define appropriately the quantum super Grassmannian as an homogeneous superspace. This was done for $N=1$ in \cite{cfl,fl}. As we will see, the $N=2$ case has its own peculiarities.
This program could, in principle be proposed for general homogeneous superspaces, not necessarily superprojective. But the projectivity gives us an advantage: the algebra associated to the projective embedding (super Pl\"{u}cker embedding) can be seen both, as a quotient algebra of the projective superspace
$\mathbb{P}^{8|8}$ modulo some homogeneous polynomial relations (super Pl\"{u}cker relations) as well as a {graded} subring of the superring $\C[\rSL(4|2)]$,
{encoding its projective embedding (see also \cite{fi1, cfg, afl}).}
We will see this in detail in Section \ref{classical-sec}.
One can then define a quantum super Grassmannian as a certain subalgebra of the super Hopf algebra $\rSL_q(4|2)$. If done correctly, the subalgebra must represent a quantum homogeneous superspace for $\rSL_q(4|2)$, that is, the coproduct in $\rSL_q(4|2)$ induces a coaction on the quantum super Grassmannian.
The chiral Minkowski $N=2$ superspace $\bM$ emerges naturally in
this context as the big cell in the Grassmannian $Gr(2|0,4|2)$.
The $N=1$ case was extensively studied in \cite{fl}, Chapter 4. However, as
remarked above, the $N=2$ SUSY has its own peculiarities, which make
the theory richer. We view the big cell in $Gr(2|0,4|2)$ as
the subsupermanifold containing certain $2|0$ subspaces and we realize it
as the set $S$ of pairs of vectors
in $\C^{4|2}$ modulo the natural right $\rGL(2)$ action, which accounts
for basis change.
Hence, we construct $\bM$ as the quotient of $S$ modulo
the ordinary general linear group $\rGL(2)$.
The quantization of $\bM$ is obtained, as expected,
as the subsuperring of a localization of $\rSL_q(4|2)$,
generated by the quantum coinvariants with respect to the
coaction of quantum $\rGL_q(2)$ (see \cite{fl}, Chapter 4 for
the $N=1$ case). The presentation
of this quantum superring via generators and relations, makes an essential
use of the commutation relations among the quantum determinants
appearing
in the definition of the quantum $Gr(2|0,4|2)$ and the
Pl\"{u}cker relations. Moreover, the quantum Minkowski space, $\bM_q$, is isomorphic to
the quantum Manin superalgebra, that is, the quantum super bialgebra
of matrices, as described in \cite{ma2}. This fact is highly non
obvious, it depends on the quite involved
commutation relations of quantum determinants
and it shows how this framework is natural and suitable for
more exploration, as we detail below.
\medskip
The chiral Minkowski $N=2$ superspace, being
a quotient, appears then naturally also
as a principal
bundle for the action of $\rGL(2)$. There is an extensive literature
regarding the quantization of principal bundles (see \cite{afl, hajac, hajac1,
hajac2, buachalla} and references therein). In particular the notion of
Hopf-Galois extension \cite{mont}
appears to be the right one to formulate, in the
affine setting, the
theory of principal bundles
to obtain their quantum deformations.
We hence proceed to define Hopf-Galois extensions in the
SUSY framework and prove that the chiral Minkowski $N=2$
superspace $\bM$ is {the base for
a principal bundle $S$} for the supergroup $\rGL(2)$,
by realizing it as a trivial Hopf-Galois extension
(see also \cite{kessler} for a more geometric, yet equivalent, view
on super principal bundles).
Next, we construct a quantum deformation $\bM_q$ of $\bM$, by taking
advantage of our previous realization and show
that $\bM_q$ is the quantum space, {base for the
quantum principal bundle $S_q$}, for $\rGL_q(2)$.
\medskip
We plan to explore, in a forthcoming paper, the construction
of covariant differential calculi on the quantum chiral Minkowski $N=2$
superspace and then proceed towards the realization of
a theory in a curved background.
\medskip
The paper is organized as follows.
In Section \ref{plucker-sec} we describe the super Pl\"{u}cker embedding of the super Grassmannian and its presentation in terms of generators and relations.
In Section \ref{classical-sec} we give the classical super Grassmannian as a subalgebra of the coordinate superalgebra of $\rSL(4|2)$.
In Section \ref{bigcell-sec} we briefly describe the big cell of the super Grassmannian, the $N=2$, $D=4$ Minkowski superspace.
In Section \ref{quantumgrassmannian-sec} we pass to define the super Grassmannian as a subsuperalgebra of $\rSL_q(4|2)$, computing the commutation relations of the generators and the quantum super Pl\"{u}cker relations that they satisfy.
In Section \ref{homogenous-sec} we give the coaction of $\rSL_q(4|2)$ on the quantum super Grassmannian defined in the previous section, proving that it is a quantum homogeneous superspace.
In the last section, Section \ref{quantumbundle-sec}, we construct the $N=2$ ordinary chiral Minkowski superspace $\bM$ and its quantization $\bM_q$, realized both first as
homogeneous spaces for the action of the ordinary (quantum) general
linear group in dimension $2$, then as {bases for} (quantum)
principal bundles for $\rGL(2)$ and $\rGL_q(2)$ respectively.
\section{The super Pl\"{u}cker embedding}\label{plucker-sec}
We are going to give the embedding of $Gr(2|0,4|2)$ in the projective superspace $\mathbb{P}^{8|8}$.
Let $E=\bigwedge^2\C^{4|2}$ and $\{e_1,\dots,e_4,\epsilon_5,\epsilon_6\}$ an homogeneous basis for $\C^{4|2}$, we then have a basis for $E$ as
\begin{align*}
&e_i\wedge e_j \quad 1\leq i<j\leq 4, \quad \epsilon_5\wedge\epsilon _5,\quad \epsilon_6\wedge\epsilon _6,\quad \epsilon_5\wedge\epsilon _6,&\hbox{(even)},\\
&e_k\wedge \epsilon _5,\quad e_k\wedge\epsilon_6\quad 1\leq k\leq 4, &\hbox{(odd)}\,.
\end{align*}
So $E\simeq \C^{9|8}$ and $\mathbb{P}(E)\simeq\mathbb{P}^{8|8}$. An element of $E$ is given as
$$Q=q+\lambda_5\wedge\epsilon_5+\lambda_6\wedge\epsilon _6+a_{55}\epsilon_5\wedge\epsilon_5+a_{66}\epsilon_6\wedge\epsilon_6+ a_{56}\epsilon_5\wedge\epsilon_6\,,$$
with
$$q=q_{ij}e_i\wedge e_j,\quad \lambda_m=\lambda_{im}e_i,\qquad i,j=1,\dots, 4,\quad m=5,6\,.$$
The element $q$ is decomposable if $q=a\wedge b$, where
$$a=r+\xi_5\epsilon_5+\xi_6\epsilon_6,\qquad b=s+\eta_5\epsilon_5+\eta_6\epsilon_6,$$ with
$r=r_ie_i$, $s=s_ie_i$.
One obtains the following relations
\begin{align}
&q=r\wedge s&& &&\,\nonumber\\
&\lambda_5=\xi_5s-\eta_5 r,\qquad &&\lambda_6=\xi_6s-\eta_6 r,&&\nonumber\\
&a_{55}=\xi_5\eta_5,
\qquad &&a_{66}=\xi_6\eta_6,
\qquad &&a_{56}=\xi_5\eta_6+
\xi_6\eta_5\,,\label{wedgeproduct}
\end{align}
which imply
\begin{align}
&q\wedge q=0,\nonumber\\
&q\wedge\lambda_5=0,\qquad &&q\wedge\lambda_6=0,&&\nonumber\\
&\lambda_5\wedge\lambda_5=-2a_{55}q,\qquad &&\lambda_6\wedge\lambda_6=-2a_{66}q,\qquad &&\lambda_5\wedge\lambda_6=-a_{56}q,\nonumber\\
&\lambda_5 a_{55}=0,\qquad &&\lambda_6 a_{66}=0,\qquad&&\nonumber\\
&\lambda_5 a_{66}=-\lambda_6a_{56},\qquad &&\lambda_6 a_{55}=-\lambda_5a_{56},\qquad&&\nonumber
\\&a_{55}^2=0,\qquad &&a_{66}^2=0,\qquad && a_{56}a_{56}=-2a_{55}a_{66},\nonumber
\\&a_{55}a_{56}=0,\qquad &&a_{66}a_{56}=0\,.\qquad && \label{superplucker}
\end{align}
Relations (\ref{superplucker}) are the super Pl\"{u}cker relations.
We can write them in coordinates in the following way (always $1\leq i<j<k\leq 4$ and $5\leq n\leq 6$):
\begin{align}
&q_{12}q_{34}- q_{13}q_{24}+q_{14}q_{23}=0,&&\hbox{(Pl\"{u}cker relation)}
\nonumber\\
&q_{ij}\lambda_{kn}-q_{ik}\lambda_{jn}+q_{jk}\lambda_{in}=0,
\nonumber\\
&\lambda_{in}\lambda_{jn}=a_{nn}q_{ij}, &&\lambda_{i5}\lambda_{j6}+\lambda_{i6}\lambda_{j5}=a_{56}q_{ij},
\nonumber\\
&\lambda_{in}a_{nn}=0,&&\lambda_{i5}a_{66}=-\lambda_{i6}a_{56}&&\lambda_{i6}a_{55}=-\lambda_{i5}a_{56}
\nonumber\\&a_{nn}^2=0&&a_{55}a_{56}=0,&&a_{66}a_{56}=0\nonumber\\
&a_{56}a_{56}=-2a_{55}a_{66}\,. &&\label{superpluckercoor}
\end{align}
For $N=1$ these relations were given in \cite{cfl}, were the relation $a_{55}^2=0$ was missing but implicitly assumed. As we can see, for $N=2$ extra relations appear. The super Pl\"{u}cker relations for arbitrary $N$ are given in \cite{sv}, page 17. They coincide with ours by a change in the notation
due to the appearance of a sign, because of a different convention
on row/column vectors, hence the consequent change of sign
of the supertranspose.
We will denote as $\cI_P$ the ideal generated by them in the affine superspace
{$\mathbb{A}^{9|8}$}
(with generators $q_{ij}, a_{nm}, \lambda_{kn}$). They are homogeneous quadratic equations, so they are defined in the projective space $\mathbb{P}^{8|8}$.
Let us denote $\rGr= Gr(2|0,4|2)$ and consider the super Pl\"{u}cker map
$$\begin{CD}\rGr @>>>\mathbb{P}^{8|8}\\
\rspan\{a, b\}@>>>[a\wedge b]\,,\end{CD}$$
We have the following
\begin{proposition}\label{superring-prop}The superring associated to the image of $\rGr$ under the super Pl\"{u}cker embedding is
$$\C[\rGr]\cong\C[q_{ij}, a_{nm}, \lambda_{kn}]/\cI_P\,,$$
that is, the relations in $\cI_P$ are all the relations satisfied by the generators $q_{ij}, a_{nm}, \lambda_{kn}$. Then $Gr(2|0,4|2)$, is a projective supervariety.\end{proposition}
{\it Proof.} This is proven for arbitrary $N$ (and further generalizations) in \cite{sv}, Theorem 6, {denoted as the ``algebraic case''}.
\hfill$\blacksquare$
\section{ The classical picture}\label{classical-sec}
As stated in the introduction, one can see $\C[\rGr]$ as a subalgebra of
$\C[\rSL(4|2)]$. Let us display the generators of this algebra in matrix form
\be\begin{pmatrix}
g_{11}&g_{12}&g_{13}&g_{14}&\vline&\gamma_{15}&\gamma_{16}\\
g_{21}&g_{22}&g_{23}&g_{24}&\vline&\gamma_{25}&\gamma_{26}\\
g_{31}&g_{32}&g_{33}&g_{34}&\vline&\gamma_{35}&\gamma_{36}\\
g_{41}&g_{42}&g_{43}&g_{44}&\vline&\gamma_{45}&\gamma_{46}\\
\hline
\gamma_{51}&\gamma_{52}&\gamma_{53}&\gamma_{54}&\vline&g_{55}&g_{56}\\
\gamma_{61}&\gamma_{62}&\gamma_{63}&\gamma_{64}&\vline&g_{65}&g_{66}
\end{pmatrix}\,\label{groupmatrix}\ee then
$$\C[\rSL(4|2)]=\C[g_{ij},g_{mn}, \gamma_{im}, \gamma_{nj}]/(\mathrm{Ber}-1)\,,
$$
where $\mathrm{Ber}$ is the Berezinian of the matrix and $1\leq i,j\leq 4$ and $5\leq m,n\leq 6$.
\begin{proposition}The superring\label{generators-prop}
$$\C[\rGr]\cong\C[q_{ij}, a_{nm}, \lambda_{kn}]/\cI_P\,,$$
is generated, as a subring of $\C[\rSL(4|2)]$ by the elements
\begin{align*}
&y_{ij}=g_{i1}g_{j2}-g_{i2}g_{j1},&&\eta_{kn}=g_{k1}\gamma_{n2}-g_{k2}\gamma_{n1}\\
&x_{55}=\gamma_{51}\gamma_{52}, &&x_{66}=\gamma_{61}\gamma_{62}&&x_{56}=\gamma_{51}\gamma_{62}+\gamma_{61}\gamma_{52}
\end{align*} with the homomorphism
$$\begin{CD}
\C[\rGr]@>>>\C[\rSL(4|2)]\\
q_{ij}, \lambda_{kn}@>>>y_{ij}, \eta_{kn}\\
a_{55},a_{66},a_{56}@>>>x_{55},x_{66},x_{56}
\end{CD}
$$
\end{proposition}
{\it Proof.} The proof uses an argument similar to the one used to obtain (\ref{wedgeproduct}). Instead of taking the vectors $a$ and $b$ we have to take the first two columns of the matrix (\ref{groupmatrix}).
\hfill$\blacksquare$
\section{The big cell}\label{bigcell-sec}
A $(2|0)$ subspace of $\C^{4|2}$ is given as the linear span of two even vectors\footnote{Here we are using implicitly the formalism of the {\it functor of points} to describe a super variety \cite{dm}.}
\be\label{foptsV}
V(A)= \rspan\begin{pmatrix}u_1&v_1\\u_2&v_2\\u_3&v_3\\u_4 &v_4\\\hline\mu_1&\nu_1\\\mu_2&\nu_2\end{pmatrix}\qquad u_i, v_i\in A_0,\; \mu_i,\nu_i\in A_1\,,
\ee
where $A$ is any superalgebra. Clearly there is a right action of $\rGL_2(A)$ over $V(A)$ (change of basis). We assume now that
\be\det\begin{pmatrix}u_1&v_1\\u_2&v_2\end{pmatrix} \quad \hbox{invertible in}
\quad A_0
\label{bigcell}\ee
{This is a supervariety, which is
an affine open set into the affine superspace $\mathbb{A}^{8|4}$.
It is represented by the superring:
$$
\C[S]=\C[a_{ij},\al_{kl}][T]/((a_{11}a_{22}-a_{12}a_{21})T-1), \,\,
1 \leq i<j \leq 4, \, 5 \leq k<l \leq 6
$$
The condition of invertibility of the determinant function
$a_{11}a_{22}-a_{12}a_{21}$ accounts for the condition in (\ref{bigcell}).}
\medskip
Then, using the right action of $\rGL_2(A)$ we can bring
{(\ref{foptsV})} to the standard form
$$V(A)=\rspan\begin{pmatrix}\id_{2\times 2}\\P_{2\times 2}\\\psi_{2\times 2}^t\end{pmatrix}\,.
$$
$P$ and $\psi$ are even and odd coordinates in the open subset of $\rGr$ characterized by
(\ref{bigcell}) called the big cell of $\rGr$.
As for the $N=1$ case, one can show that the subgroup of $\rSL(4|2)$ that leaves invariant the big cell contains the (complexified) $N=2$ super Poincar\'{e} group times the $R$-symmetry (dilations for $N=1$). In fact, we call $\rGr$ the antichiral\footnote{See \cite{fl} to see why this space is the antichiral one.} conformal superspace, while the big cell is the antichiral Minkowski superspace. In this respect we do not coincide with the notation of \cite{hl}-\cite{haho}, where they call directly Minkowski superspace to the Grassmannian.
We do not extend here on this construction, but the condition (\ref{bigcell}) will be also used in the quantum setting.
\section{The quantum Grassmannian} \label{quantumgrassmannian-sec}
We follow \cite{ma2} to define the quantum group $\rSL_q(r|s)$.
\begin{definition} \label{Manin-def} The quantum matrix superalgebra $\rM_q(r|s)$ is defined as
$$
\rM_q(r|s)=_{\mathrm{def}}\C_q \langle z_{ij},\xi_{kl}\rangle/\cI_M
$$
where $\C_q\langle z_{ij},\xi_{kl}\rangle$ denotes the free
superalgebra over $\C_q=\C[q,q^{-1}]$
generated by the even variables
$$z_{ij},\qquad \hbox{ for }\quad 1 \leq i,j \leq r \quad \hbox{ or } \qquad r+1 \leq i,j \leq r+s.$$
and by the odd variables
\begin{align*}&\xi_{kl}&& \hbox{for }\quad 1 \leq k \leq r, \quad r+1 \leq l \leq r+s \\&&&\hbox{or }\, r+1 \leq k \leq r+s, \quad 1 \leq l \leq r,\end{align*}
satisfying the relations $\xi_{kl}^2=0$ and $\cI_M$ is an ideal that we describe below. We can visualize the generators as a matrix
\be\begin{pmatrix}z_{r\times r}&\xi_{r\times s}\\
\xi_{s\times r}&z_{s\times s}\end{pmatrix}\,.\label{generatorsmatrix}\ee
It is convenient sometimes to have a common notation for even and odd variables.
$$
a_{ij}=\begin{cases} z_{ij} & 1 \leq i,j \leq r, \, \hbox{ or } \quad
r+1 \leq i,j \leq r+s ,\\ \\
\xi_{ij} & 1 \leq i \leq r,\quad r+1 \leq j \leq r+s, \, \hbox{ or } \\&
r+1 \leq i \leq r+s, \quad 1 \leq j \leq r\,.
\end{cases}
$$
We assign a parity to the indices: $p(i)=0$ if $1 \leq i \leq r$ and $p(i)=1$ if $r+1 \leq i \leq r+s$. The parity of $a_{ij}$ is $\pi(a_{ij})=p(i)+p(j)$ mod 2. Then, the ideal $\cI_M$ is generated by the relations \cite{ma2}:
\begin{align}
&a_{ij}a_{il}=(-1)^{\pi(a_{ij})\pi(a_{il})}
q^{(-1)^{p(i)+1}}a_{il}a_{ij}, && \hbox{for } j < l \nonumber\\&&& \nonumber\\
&a_{ij}a_{kj}=(-1)^{\pi(a_{ij})\pi(a_{kj})}
q^{(-1)^{p(j)+1}}a_{kj}a_{ij}, && \hbox{for } i < k \nonumber\\ &&\nonumber&\\
&a_{ij}a_{kl}=(-1)^{\pi(a_{ij})\pi(a_{kl})}a_{kl}a_{ij}, && \hbox{for }
i< k,\;j > l \nonumber\\&&&\hbox{or } i > k,\; j < l \nonumber\\&&& \nonumber\\
&a_{ij}a_{kl}-(-1)^{\pi(a_{ij})\pi(a_{kl})}a_{kl}a_{ij}=(-1)^{\pi(a_{ij})\pi(a_{kl})}(q^{-1}-q)
a_{kj}a_{il},&\nonumber&\\
&&& \hbox{for }i<k,\nonumbe
j<l\nonumber
\\&&&\label{ManinCR}
\end{align}
$\hbs$
\end{definition}
There is also a comultiplication
$$\begin{CD}\rM_q(m|n)@>\Delta>>\rM_q(m|n)\otimes \rM_q(m|n)\end{CD}$$
$\Delta( a_{ij})=\sum_ka_{ik}\otimes a_{kj}$ and a counit $\varepsilon (a_{ij})=\delta_{ij}$.
One can restrict to $\rSL_q(r|s)$ by setting the quantum Berezinian to 1. The antipode is the usual one (see \cite{ma2} or \cite{fl}, Appendix E). Then $\rSL_q(m|n)$ is a super Hopf algebra.
We can now define the quantum Grassmannian $\rGr_q$ mimicking Proposition \ref{generators-prop}:
\begin{definition}The quantum super Grassmannian $\rGr_q:=Gr_q(2|0,4|2)$ is the subalgebra of $\rSL_q(4|2)$ generated by the elements\label{qgrassmannian-def}
\begin{align*}
&D_{ij}:=a_{i1}a_{j2}-q^{-1}a_{i2}a_{j1} &&D_{in}:=a_{i1}a_{n2}-q^{-1}a_{i2}a_{n1}&&\\
&D_{55}:=a_{51}a_{52} &&D_{66}:=a_{61}a_{62}\\&D_{56}=a_{51}a_{62}-q^{-1}a_{52}a_{61}\\
\end{align*}
with $1\leq i<j\leq 4$ and $\;n=5,6 $.
\end{definition}
\hfill$\blacksquare$
We want to give a presentation in terms of generators and relations, as in Proposition \ref{superring-prop} for the classical case. Note that, first of all, we have to compute the commutation rules among the $D$'s. After some (tedious) calculations we arrive at:
\begin{itemize}
\item Let $1\leq i,j,k,l\leq6$ be not all distinct, and $D_{ij}$, $D_{kl}$ not both odd. Then
\be D_{ij}D_{kl}=q^{-1}D_{kl}D_{ij},\qquad (i,j)<(k,l),\; i<j, k<l\,,\label{CR3}\ee where the ordering `$<$' of pairs is the lexicographical ordering.
\item Let $1\leq i,j,k,l\leq6$ be all distinct, and $D_{ij}$, $D_{kl}$ not both odd and $D_{ij},D_{kl}\neq D_{56}$. Then
\begin{align}
&D_{ij}D_{kl}=q^{-2}D_{kl}D_{ij}, &&1\leq i<j<k<l\leq 6,\nonumber\\
&D_{ij}D_{kl}=q^{-2}D_{kl}D_{ij}-(q^{-1}-q)D_{ik}D_{jl} &&1\leq i<k<j<l\leq 6,\nonumber\\
&D_{ij}D_{kl}=D_{kl}D_{ij}&&1\leq i<k<l<j\leq 6,\label{CR1}
\end{align}
\item Let $1\leq i<j\leq4$, $5\leq n\leq m\leq 6$. Then
\begin{align}
&D_{in}D_{jn}= -q^{-1}D_{jn}D_{in}-(q^{-1}-q)D_{ij}D_{nn}=-qD_{jn}D_{in},\nonumber\\
&D_{ij}D_{nm}=q^{-2}D_{nm}D_{ij},\nonumber \\
&D_{i5}D_{j6}=-q^{-2}D_{j6}D_{i5}-(q^{-1}-q)D_{ij}D_{56},\nonumber\\
&D_{i6}D_{j5}=-D_{j5}D_{i6},\nonumber\\
&D_{i5}D_{i6}=-q^{-1}D_{i6}D_{i5},\nonumber\\
&D_{i5}D_{i6}=-q^{-1}D_{i6}D_{i5},
\nonumber\\
&D_{55}D_{66}=q^{-2}D_{66}D_{55}\,,\nonumber \\
&D_{55}D_{56}=0\,. \label{CR2}
\end{align}
\end{itemize}
The Pl\"{u}cker relations are modified. One has for $1\leq i<j<k\leq 4$ and $ n=5, 6$:
\begin{align}
&D_{12}D_{34}-q^{-1}D_{13}D_{24}+q^{-2}D_{14}D_{23}=0, \nonumber\\
&D_{ij}D_{kn}-q^{-1}D_{ik}D_{jn}+q^{-2}D_{jk}D_{in}=0, \nonumber
\\
&D_{i5}D_{j6}+q^{-1}D_{i6}D_{j5}=qD_{ij}D_{56}, \nonumber\\
&D_{in}D_{jn}=qD_{ij}D_{nn}, \nonumber \\
&D_{in}D_{nn}=0, \nonumber\\&D_{i5}D_{66}=-q^{-1}D_{i6}D_{56},\nonumber\\
&D_{i6}D_{55}=-q^{2}D_{i5}D_{56} \nonumber\\
&D_{nn}^2=0,\nonumber\\&D_{55}D_{56}=0, \nonumber \\&D_{66}D_{56}=0\nonumber\\\
&D_{56}D_{56}=(q^{-1}-3q)D_{55}D_{66}\,.\label{quantumplucker}
\end{align}
The first relation in (\ref{CR2}) has been simplified with the use of the fourth relation in (\ref{quantumplucker}).
We have the following
\begin{proposition}The quantum Grassmannian superring $\rGr_q=Gr_q(2|0,4|2)$ is given in terms of generators and relations as
$$\rGr_q=\C_q\langle X_{ij}, X_{mn}, X_{im} \rangle, \cI_q,\qquad 1\leq i<j\leq 4;\quad 5\leq m\leq n\leq 6 \,,$$
where $\cI_q$ is the ideal generated by the commutation relations (\ref{CR3}),(\ref{CR1}),(\ref{CR2}) and the quantum super Pl\"{u}cker relations (\ref{quantumplucker}).
\end{proposition}
{\it Proof.}
We give a sketch of the argument, whose idea is expressed
in \cite{fi1} Theorem 5.4 and also in \cite{fl} Chapter 4.
The super Pl\"{u}cker relations are all the relations satisfied by the quantum determinants: suppose that there is an extra relation R. Then $R=(q-1)R^{(1)}$. Then $R^{(1)}$ may be of the form $R^{(1)}=(q-1)R^{(2)}$ or $R^{(1)}$ mod $(q-1)$ not identically 0. In the second case, since $\rSL_q(m|n)$ is an algebra without torsion, we would have an additional classical Pl\"{u}cker relation, which cannot be. In the first case we have the same possibilities for $R^{(2)}$. At the end of the procedure we will obtain $R^{(n)}=0$, that would be a new classical Pl\"{u}cker condition. But we know that this is not possible.
\hfill$\blacksquare$
\section{The quantum super Grassmannian as a quantum homogeneous superspace}\label{homogenous-sec}
To finish the interpretation of the quantum super Grassmannian as an homogeneous superspace under the quantum supergroup $\rSL_q(4|2)$ we have to see how it is the coaction on $\rGr_q$.
This is done in the following
\begin{proposition} The restriction of the comultiplication in $\rSL_q(4|2)$
$$\begin{CD}
\rSL_q(4|2)@> \Delta>> \rSL_q(4|2)\otimes\rSL_q(4|2)\\
a_{ij}@>>> \Delta(a_{ij})=\sum_{k=1}^6a_{ik}\otimes a_{kj}
\end{CD}
$$
to the subalgebra $\rGr_q$ is of the form\footnote{We denote with $\Delta$ both, the comultiplication and its restriction to $\rGr_q$ in order not to burden the notation. The meaning should be clear from the context.}
$$\begin{CD}
\rGr_q@> \Delta>> \rSL_q(4|2)\otimes\rGr_q\,.
\end{CD}
$$
\end{proposition}
{\it Proof.} The coaction property is guaranteed by the associativity of the coproduct, so we only have to check that
$$D_{ij}, D_{im}, D_{mn}\in \rSL_q(4|2)\otimes\rGr_q\,.$$
Let us denote as $D_{ij}^{kl}=a_{ik}a_{jl}-q^{-1}a_{il}a_{jk}$, so in the previous notation $D_{ij}=D_{ij}^{12}$. After some calculations one can prove
\begin{enumerate}
\item Let us call $P$ the condition $1\leq k,l\leq 6$ and at least one of the two indices is less that 5. For $1\leq i<j \leq 4$:
\begin{align*}\Delta(D_{ij})&= \sum_{P \cap (k<l)} D_{ij}^{kl} \otimes D_{kl}^{12}- (a_{i5}a_{j6}+q^{-1}a_{i6}a_{j5}) \otimes D_{56}\\&- (1+q^{-2})\sum_{5\leq k \leq 6}a_{ik}a_{jk} \otimes D_{kk}\,.\end{align*}
\item For $1\leq i \leq 4$ and $5\leq m \leq 6$:
\begin{align*}
\Delta(D_{im})&=\sum_{\substack{k<5 \\ k<l}} a_{ik}a_{ml} \otimes D_{kl} -q^{-1}\sum_{\substack{k<5 \\ l<k}} a_{ik}a_{ml} \otimes D_{lk}\\& + (a_{i5}a_{m6}+q^{-1}a_{i6}a_{m5}) \otimes D_{56}\\
& + (1+q^{-2})\sum_{5\leq k \leq 6} a_{ik}a_{mk} \otimes D_{kk}+q^{-1} \sum_{\substack{k\geq5 \\ l<5}} a_{ik}a_{ml} \otimes D_{lk}\,.
\end{align*}
\item For $5 \leq m,n \leq 6$:
\begin{align*}
\Delta(D_{56})&=\sum_{\substack{k<5 \\ k<l}} a_{5k}a_{6l} \otimes D_{kl} -q^{-1}\sum_{\substack{k<5 \\ l<k}} a_{5k}a_{6l} \otimes D_{lk}\\& + (a_{55}a_{66}+q^{-1}a_{56}a_{65}) \otimes D_{56}\\
& + (1+q^{-2})\sum_{5\leq k \leq 6} a_{5k}a_{6k} \otimes D_{kk}+q^{-1} \sum_{\substack{k\geq5 \\ l<5}} a_{5k}a_{6l} \otimes D_{lk}\,,
\end{align*}
and
\begin{align*}
\Delta(D_{nn})= \sum_{1\leq k<l \leq 6} a_{nk}a_{nl} \otimes D_{kl} +\sum_{5\leq k \leq 6} a_{nk}^{2} \otimes D_{kk}\,.
\end{align*}
\end{enumerate}
This proves our statement.
\hfill$\blacksquare$
\section{Quantum super bundles: quantum chiral Minkowski
superspace}\label{quantumbundle-sec}
In this section we want to reinterpret our construction in
the framework of quantum principal bundles, as in \cite{afl} and references therein.
We shall concentrate our attention on the local picture, that is,
we want to look at the quantization of a super
bundle $S \lra \C^{4|4}$, with base space
the chiral Minkowski superspace $\C^{4|4}$, which we interpret as the
big cell into the Grassmannian supermanifold $\mathrm{Gr}$
{(see also Sec. \ref{bigcell-sec}).}
We shall not develop a full theory of quantum principal super bundles,
but we will recall the key definitions in order to put in the correct
framework our construction.
We start with the classical definition.
\begin{definition}\label{p-princ}
Let $X$ and $M$ be topological spaces, $P$ a topological group and
$\wp: X \lra M$ a continuous function. We say
that $(X, M, \wp, P)$ is a \textit{$P$-principal bundle}
(or {\it principal bundle} for short) with total space
$X$ and base $M$, if the following conditions hold
\begin{enumerate}
\item $\wp$ is surjective.
\item $P$ acts freely from the right on $X$.
\item $P$ acts transitively on the fiber $\wp^{-1}(m)$
of each point $m \in M$.
\item $X$ is locally trivial over $M$, i.e. there is an open covering
$M=\cup U_i$ and homeomorphisms
$\sigma_i:\wp^{-1}(U_i) \lra U_i \times P$
that are $P$-equivariant i.e.,
$\sigma_i(up)=\sigma_i(u)p$, $u \in U_i$, $p \in P$.
\end{enumerate}
\medskip
If $X\cong M\times P$ we say that the bundle is globally trivial.
\hfill$\blacksquare$
\end{definition}
We can then define {algebraic, analytic or smooth} $P$-principal
bundles, by the taking objects and morphisms in the appropriate
categories. There is clearly no obstacle in writing the same
definition in the super context, provided we exert some care
in the definition of surjectivity
(see \cite{ccf}, Section 8.1 for details).
We would like, however, to take a different
route.
We turn to the notion of Hopf-Galois extension, that is most fruitful
for the quantization. Our definition in the super category is
the same as for the ordinary one (see \cite{mont} for more details in the ordinary category).
\begin{definition} \label{hg-def}
Let $(H, \Delta, \epsilon, S)$ be a Hopf superalgebra and
$A$ be an $H$-comodule superalgebra with coaction
$\delta : A \lra A \otimes H$.
Let
\beq \label{coinvariant-def}
B:=A^{\coinv\,H} := \{a \in A \,|\, \delta(a) = a \otimes 1\}~.
\eeq
The extension $A$ of the superalgebra $B$ is called
$H$-{Hopf-Galois} (or simply {Hopf-Galois}) if the
map
$$
\chi:A \otimes_B A \lra A \otimes H, \qquad \chi=(m_A \otimes
\mathrm{id})(\mathrm{id}
\otimes_B \delta)
$$
called the {canonical map}, is bijective ($m_A$ denotes the
multiplication in $A$).
The extension $B=A^{\coinv\,H}\subset A$ is called
$H$-\textit{principal comodule superalgebra} if it is Hopf-Galois and
$A$ is $H$-equivariantly projective as a left $B$-supermodule, i.e., there
exists a left $B$-supermodule and right $H$-comodule morphism $s : A \to B
\otimes A$ that is a section of the (restricted) product $m : B
\otimes A \to A$.
In particular if $H$ is a Hopf algebra with bijective antipode
over a field, the condition of equivariant projectivity of $A$ is
equivalent to that of faithful flatness of $A$ (see \cite{ss},
\cite{abps}).
\hfill$\blacksquare$
\end{definition}
{We now follow \cite{afl} Sec. 2, in giving the definition of quantum
principal bundle, though it differs slightly from the one given
in the literature, which also requires the existence of a differential
structure (see e.g. \cite{bm} Ch. 5). We plan to explore such
structures in a forthcoming paper.
\begin{definition}\label{pb-def}
We define \textit{quantum principal bundle}
a pair $(A,B)$, where $A$ is an $H$-Hopf Galois extension and
$A$ is $H$-equivariantly projective as a left $B$-supermodule, that is,
$A$ is an $H$-principal comodule superalgebra.
\hfill$\blacksquare$
\end{definition}}
In the ordinary case, the notion of affine $H$-principal bundle
coincides with Definition \ref{p-princ} when we take $H=\cO(P)$,
$A=\cO(X)$ and $B=\cO(M)$, where $\cO(X)$ denotes the algebra
of functions on $X$ (algebraic, differential, holomorphic, etc).
The Hopf-Galois condition is equivalent to saying that the action of $P$
on $X$ is free and the equivariance property means that the bundle
is locally trivial.
We assume, in the algebraic setting, that all our varieties are affine.
\medskip
There is a special case of Hopf-Galois extensions, corresponding to
a globally trivial principal bundle. In this case the
technical conditions of Definition \ref{pb-def} are
automatically satisfied. We shall focus on this case
leaving aside the general one.
\begin{definition}\label{cleft-def}
Let $H$ be a Hopf superalgebra and $A$ an $H$-comodule superalgebra.
The algebra extension $A^{\coinv\,H}\subset A$ is called
a \textit{cleft extension} if
there is a right $H$-comodule map $j : H \to A$, called \textit{cleaving map},
that is convolution invertible, i.e. there exists a map $h:H\to A$
such that the convolution product $j\star h$ satisfies,
$$j\star h:=m_A\circ(j\otimes h)\circ\Delta(f)=\epsilon(f)\cdot 1\, $$
or, in Sweedler notation
$\Delta(f)=\sum f_1 \otimes f_2$,
$$\sum j(f_1)h(f_2)=\epsilon(f)\cdot 1\,$$
for all $f \in H$. The map $h$ is the convolution inverse of $j$.
An extension $A^{\coinv\,H}\subset A$ is called a
{trivial extension} if there exists such map.
Notice that when $j$ is an algebra map,
its convolution inverse is just $h=j \circ S^{-1}$.
\hfill $\blacksquare$
\end{definition}
A trivial extension is actually a Hopf-Galois extension and a
principal bundle. {When $j$ is an algebra map,
we have an algebra isomorphism $A \cong B \# H$} (see
\cite{mont}, Sections 4.1 and 7.2 for the smash product `$\,\#\,$'),
which in the classical case means that we have a trivial bundle
(see \cite{mont} Chapter 8 and \cite{bm} Sec. 5.1.2).
\medskip
We now examine an example with physical significance coming
from our previous treatment. Consider the set
of $4 \times 2 \, | \, 2 \times 2$
supermatrices with complex entries
\beq
\begin{pmatrix} a_{11} & a_{12} \\a_{21} &a_{22}\\ a_{31} &a_{32}\\a_{41}&a_{41}\\
\alpha_{51}&\alpha_{52}\\\alpha_{61}&\alpha_{62}
\end{pmatrix}\,.
\eeq
This can be seen as the affine superspace
$\bA^{8|4}$ described by the coordinate superalgebra
$\C[a_{ij},\al_{kl}]$ with $ i=1,\dots, 4$, $\,j,l=1,2$, $\,k=5,6$.
As in the ordinary
setting, we can view elements in $\bA^{8|4}$ as $2|0$ subspaces
of $\C^{4|2}$:
$$
W=\Span\{a_1,a_2\} \subset \C^{4|2}
$$
In this way,
$W$ may also be viewed as an element in $\mathrm{Gr}$.
\medskip
In the superspace $\bA^{8|4}$ consider the open subset $S$ consisting
of matrices such that the minor formed with $a_{ij}$, $i,j=1,2$ is invertible.
This open subset $S$ is described by its coordinate superalgebra:
$$
\C[S]=\C[a_{ij},\al_{kl}][T]/((a_{11}a_{22}-a_{12}a_{21})T-1)
$$
We have a right action of $\rGL_2(\C)$ on $S$ corresponding to
the change of basis of such subspaces:
$$
\Span\{a_1,a_2\}, g \mapsto \Span\{a_1 \cdot g, a_2 \cdot g\},\qquad g\in \rGL_2(\C)\,.
$$
\begin{proposition}\label{affine-prop}
Let the notation be as above.
The quotient of $S$
by the right $\rGL_2(\C)$ action is
an affine superspace $\bM$ of dimension $4|4$, the $N=2$ chiral Minkowski superspace $\bM$.
\end{proposition}
{\it Proof.}
We can write:
\beq
\bM=\left\{ (a_1,a_2),\, a_1,a_2 \in \C^{4|2} \;|\; \det\begin{pmatrix} a_{11} & a_{12} \\ a_{21} & a_{22}
\end{pmatrix} \neq 0 \right\}
\Big/\rGL_2(\C) \qquad
\eeq
In the quotient $\bM$ we can choose a (unique) representative
$(u,v)$ for $(a_1,a_2)$ of the form:
\beq
\left\{
\begin{pmatrix} 1 \\ 0 \\u_{1} \\u_{2}\\ \nu_{3} \\ \nu_{4}\end{pmatrix},
\begin{pmatrix} 0 \\ 1 \\v_{1} \\v_{2} \\ \eta_{3} \\ \eta_{4}
\end{pmatrix}\right\}\,,
\eeq
so $\bM$ is $\C^{4|4}$.
\hfill$\blacksquare$
We notice that $\bM$ is naturally identified with the dense open
set of the Grassmannian
$\mathrm{Gr}$ in the Pl\"ucker embedding, determined by the
invertibility of the coordinate $q_{12}$ in $\mathbb{P}^{8|8}$
\medskip
We now would like to retrieve a set of global coordinates for
$\bM$ starting from the global coordinates $a_{ij}$ for $S$.
Let $\C[\rGL_2]=\C[g_{ij}][T]/((g_{11}g_{22}-g_{12}g_{21})T-1)$ be the
coordinate algebra for the algebraic group $\rGL_2(\C)$.
Let us write heuristically the equation relating the generators
of $\C[S]$, $\C[\rGL_2]$ and the polynomial
superalgebra $\C[\bM]:=\C[u_{ij}, \nu_{kl}]$
\beq\label{mink-coord}
\begin{pmatrix}
a_{11} & a_{12} \\
a_{21} & a_{22} \\
a_{31} & a_{32} \\
a_{41} & a_{42} \\
\al_{51} & \al_{52} \\
\al_{61} & \al_{62}
\end{pmatrix}
=
\begin{pmatrix} 1 & 0 \\0 & 1 \\
u_{31} & u_{32}\\
u_{41} & u_{42}\\
\nu_{51}& \nu_{52} \\
\nu_{61} & \nu_{62} \end{pmatrix}
\begin{pmatrix} g_{11} & g_{12} \\ g_{21} & g_{22}
\end{pmatrix}
\eeq
We obtain immediately:
$$
\begin{pmatrix} g_{11} & g_{12} \\ g_{21} & g_{22}
\end{pmatrix} = \begin{pmatrix} a_{11} & a_{12} \\ a_{21} & a_{22}
\end{pmatrix}
$$
and then with a short calculation,
$$
u_{i1}= -d_{2i}d_{12}^{-1}\qquad u_{i2}= d_{1i}d_{12}^{-1}$$
$$\nu_{k1}= -d_{2k}d_{12}^{-1}\qquad \nu_{k2}= d_{1k}d_{12}^{-1}
$$
for $i=3,4$ and $k=5,6$, where
$$d_{rs}:=a_{r1}a_{s2}-a_{r2}a_{s1} \qquad r<s\,.$$
\begin{proposition}Let the notation be as above.
\begin{enumerate}
\item The complex supermanifold $S$ is diffeomorphic to the supermanifold
$\C^{4|4} \times \rGL_2(\C)$:
$$\begin{CD} S @>\psi>> \C^{4|4} \times \rGL_2(\C)\,,\end{CD}$$
with
\begin{align*}
& \psi^*(g_{ij})=a_{ij} \\
&\psi^*(u_{i1})= -d_{2i}d_{12}^{-1}, &&
\psi^*(u_{i2})= d_{1i}d_{12}^{-1}.\\
&\psi^*(\nu_{k1})= -d_{2k}d_{12}^{-1} &&
\psi^*(\nu_{k2})= d_{1k}d_{12}^{-1}
\end{align*}
\item The diffeomorphism $\psi$ is $\rGL_2(\C)$-equivariant
with respect to the right $\rGL_2(\C)$ action, hence
$S/\rGL_2(\C) \cong \C^{4|4}$.
\end{enumerate}
\end{proposition}
{\it Proof.} We notice that $\psi$ is invertible,
$\psi^{-1}$ is given by:
$$
(\psi^{-1})^*(a_{ij})=g_{ij}
$$
and the rest follows from equation (\ref{mink-coord}).
The right equivariance of $\psi$ is a simple calculation, taking into account that the determinants $d_{ij}$ transform as $d_{ij}\det g'$, were $g'\in\rGL_2(\C)$.
\hfill$\blacksquare$
\begin{lemma}\label{coinvariant-lemma}
The coordinate superalgebra $\C[\bM]:=\C[u_{ij}, \nu_{kl}]$
is isomorphic to $\C[S]^{\coinv \,\C[\rGL_2]}$ the coinvariants in
$\C[S]$ with respect to the $\C[\rGL_2]$ right coaction $\de$:
$$
\begin{CD} \C[S] @>\de>> \C[S] \otimes \C[\rGL_2] \\
\begin{pmatrix}
a_{11} & a_{12} \\
a_{21} & a_{22} \\
a_{31} & a_{32} \\
a_{41} & a_{42} \\
\al_{51} & \al_{52} \\
\al_{61} & \al_{62}
\end{pmatrix} @>>>\begin{pmatrix}
a_{11} & a_{12} \\
a_{21} & a_{22} \\
a_{31} & a_{32} \\
a_{41} & a_{42} \\
\al_{51} & \al_{52} \\
\al_{61} & \al_{62}
\end{pmatrix} \otimes \begin{pmatrix} g_{11} & g_{12} \\ g_{21} & g_{22}
\end{pmatrix}
\end{CD}
$$
\end{lemma}
{\it Proof}.
In our heuristic calculation we computed expressions for
the coordinates on $\bM$. We claim that
these are coinvariants, so we need to show $\de(c)=c\otimes 1$
for any $c\in\{u_{ij}, \nu_{kl}\}$. A little calculation gives us
$$\de(d_{rs})= d_{rs} \otimes (g_{11}g_{22}-g_{12}g_{21})\; \Rightarrow\;\de(d_{rs}d_{12}^{-1})= d_{rs}d_{12}^{-1} \otimes 1$$
which proves our claim.
On the other hand, the space of functions on $S$ that are invariant under the right translation of the group can be identified with the space of functions on the quotient $S/\rGL_2(\C)$. Since we have global coordinates in $\bM$, any other invariant will be a function of these coordinates. In the Hopf algebra language, this means that $\{u_{ij}, \nu_{kl}\}$ are the only independent coinvariants.
\hfill$\blacksquare$
\begin{proposition}\label{cleaving-prop}
Let the notation be as above.
The natural projection $p:S \lra S/\rGL_2(\C)$ is a trivial principal
bundle on the chiral superspace.
\end{proposition}
{\it Proof.} We will show that $\C[S]$ is a trivial $\C[\rGL_2]$-Hopf
Galois extension of $\C[\bM]$.
In our previous lemma we proved that $\C[\bM]\cong
\C[S]^{\coinv \,\C[\rGL_2]}$, so if we give an algebra
cleaving map we will have proven our proposition. We define
$$\begin{CD}
\C[\rGL_2] @>j>> \C[S]\\
g_{ij}@>>>a_{ij}\,.\end{CD}
$$
We leave to the reader the easy check that $j$ is convolution invertible with convolution inverse
$$h=j\circ S\,.$$ Moreover, the calculation below shows that $j$ is a $\C[\rGL_2(\C)]$-comodule map,
$$(\delta \circ j) (g_{ij})=\delta(a_{ij})=\sum a_{ik}\otimes g_{kj}.$$
$$((j \otimes \rid) \circ \Delta) (g_{ij})=(j \otimes \rid)(\sum g_{ik} \otimes g_{kj})=\sum a_{ik}\otimes g_{kj}.$$
$$\Rightarrow~~~\delta \circ j= (j \otimes \rid) \circ \Delta.$$
This proves the result.
\hfill$\blacksquare$
We now go to the quantum setting, where
we lose the geometric
interpretation and we retain only the algebraic point of view.
Hence a quantum principal super bundle over an affine base
is just understood as a Hopf-Galois extension with the properties mentioned in Definition \ref{pb-def}.
\medskip
We want to study the quantization of the example studied above.
Let $\C_q[S]$ be the quantization of the superalgebra $\C[S]$
obtained by taking the Manin relations (\ref{ManinCR}) among the entries still
denoted as $a_{ij}$ and $\al_{kl}$, with $i,j=1,\dots,4$ and $k,l=5,6$.
\begin{definition} The $N=2$ quantum chiral Mikowski superspace,
$\C_q[\bM]$, is the superalgebra generated by the elements
$$
\Tilde{u}_{i1}= -q^{-1}D_{2i}D_{12}^{-1},\qquad \Tilde{u}_{i2}= D_{1i}D_{12}^{-1},$$
$$\Tilde{\nu}_{k1}= -q^{-1}D_{2k}D_{12}^{-1},\qquad \Tilde{\nu}_{k2}= D_{1k}D_{12}^{-1}\,,
$$
for $i=3,4$ and $k=5,6$ in $\C_q[\rGr]$, where
$$
D_{rs}:= a_{r1}a_{s2}-q^{-1}a_{r2}a_{s1}, \qquad r<s.
$$
\end{definition}
\hfill$\blacksquare$
Using our previous computations for commutation relations among $D_{rs}$'s we get the following commutation relations among $\Tilde{u}_{ij}$'s and $\Tilde{\nu}_{kl}$'s:
\begin{align}
&\Tilde{u}_{i2}\Tilde{u}_{i1}= q^{-1}\Tilde{u}_{i1}\Tilde{u}_{i2}, \qquad &&i=3,4, \nonumber \\
&\Tilde{\nu}_{k1}\Tilde{\nu}_{k2}=-q^{-1}\Tilde{\nu}_{k2}\Tilde{\nu}_{k1}, \qquad &&k=5,6, \nonumber \\
&\Tilde{\nu}_{5l}\Tilde{\nu}_{6l}=-q^{-1}\Tilde{\nu}_{6l}\Tilde{\nu}_{5l}, \qquad &&l=1,2,\nonumber \\
&\Tilde{u}_{3j}\Tilde{u}_{4j}=q^{-1}\Tilde{u}_{4j}\Tilde{u}_{3j}, \qquad &&j=1,2,\nonumber \\
&\Tilde{u}_{ij}\Tilde{\nu}_{kj}=q^{-1}\Tilde{\nu}_{kj}\Tilde{u}_{ij}, \qquad &&j=1,2~~i=3,4~~k=5,6,\nonumber \\
&\Tilde{u}_{i1}\Tilde{\nu}_{k2}=\Tilde{\nu}_{k2}\Tilde{u}_{i1}, \qquad &&i=3,4~~k=5,6,\nonumber\\
&\Tilde{u}_{31}\Tilde{u}_{42}=\Tilde{u}_{42}\Tilde{u}_{31},\nonumber\\
&\Tilde{\nu}_{51}\Tilde{\nu}_{62}=-\Tilde{\nu}_{62}\Tilde{\nu}_{51}, \nonumber \\
&\Tilde{u}_{32}\Tilde{u}_{41}-\Tilde{u}_{41}\Tilde{u}_{32}= (q^{-1}-q)\Tilde{u}_{42}\Tilde{u}_{31}, \nonumber \\
&\Tilde{u}_{i2}\Tilde{\nu}_{k1}-\Tilde{\nu}_{k1}\Tilde{u}_{i2}= (q^{-1}-q)\Tilde{\nu}_{k2}\Tilde{u}_{i1},\qquad &&i=3,4~~k=5,6,\nonumber\\
&\Tilde{\nu}_{52}\Tilde{\nu}_{61}+\Tilde{\nu}_{61}\Tilde{\nu}_{52}=-(q^{-1}-q)\Tilde{\nu}_{62}\Tilde{\nu}_{51}\,.\label{CRMinkowski}
\end{align}
We have the following
\begin{proposition}
The quantum chiral Minkowski superspace $\C_{q}[M]$ is isomorphic to the quantum superalgebra of matrices $\rM_q(2|2)$ (Definition \ref{Manin-def}).
\end{proposition}
{\it Proof.}
We define the map $\beta:\C_{q}[M] \longrightarrow \rM_q(2|2)$ by giving it on the generators as follows:
\begin{eqnarray*}
\beta(\tilde{u}_{ij}):=z_{rs}~~~~~~\hbox{where}~~r=i-2~~\hbox{and}~~s=\begin{cases} 1~~\hbox{if}~~j=2,\\
2~~\hbox{if}~~j=1,
\end{cases}
\end{eqnarray*}
\begin{eqnarray*}
\beta(\tilde{\nu}_{kl}):= \xi_{mn}~~~~~~\hbox{where}~~m=k-2~~\hbox{and}~~n=\begin{cases} 1~~\hbox{if}~~l=2,\\
2~~\hbox{if}~~l=1\,.
\end{cases}
\end{eqnarray*}
It is clearly bijective. Comparing the commutation relations (\ref{ManinCR}) with (\ref{CRMinkowski}) it follows that $\beta$ is an isomorphism.
\hfill$\blacksquare$
We now want to show the main result for this section.
\begin{theorem}
The quantum superalgebra $\C_q[S]$ is a trivial quantum
principal super bundle on the quantum chiral Minkowski superspace.
\end{theorem}
{\it Proof.}
We need to show that $\C_q[S]$ is a trivial Hopf-Galois extension of $\C_q[M]$. We will proceed similarly to the classical case. It is easy to see that the quantum version of Lemma \ref{coinvariant-lemma} also holds. It is enough to check that
$$\de_{q}(D_{rs})= D_{rs} \otimes (g_{11}g_{22}-q^{-1}g_{12}g_{21})\,.$$
Therefore, we need to give an algebra cleaving map $j_{q}: \C_{q}[\rGL_2(\C)] \lra \C_q[S]$. Define:
$$
j_{q}(g_{ij}):=a_{ij},\qquad h_q=j_q\circ S_q,
$$
so $h_{q}:\C_{q}[\rGL_2(\C)] \lra \C_{q}[S]$,
$$h_{q}(g_{11}):=D^{-1}a_{22} \qquad h_{q}(g_{12}):=-qD^{-1}a_{12}$$
$$h_{q}(g_{21}):=-q^{-1}D^{-1}a_{21} \qquad h_{q}(g_{22}):= D^{-1}a_{11},$$
where $D:=a_{11}a_{22}-q^{-1}a_{12}a_{21}$. One can observe that:
$$j_{q} \star h_{q}= \varepsilon.1= h_{q}\star j_{q},$$
where $\star$ denotes the convolution product, i.e $j_{q}$ is convolution invertible. Moreover, a similar calculation to the one given in Proposition \ref{cleaving-prop} shows that $j_{q}$ is a $\C_{q}[\rGL_{2}]$-comodule map, i.e. $\delta_{q} \circ j_{q}= (j_{q} \otimes\rid) \circ \Delta$. Therefore, $j_{q}$ is an algebra cleaving map and $\C_{q}[M]\subset \C_{q}[S]$ is a trivial extension. Hence the theorem is proven.\hfill$\blacksquare$
\section*{Acknowledgements}
We thank P. Aschieri, E. Latini, T. Weber, R. Buachalla, A. Carotenuto, E. Shemyakova and T. Voronov for helpful observations.
This work is supported by the Spanish Grants FIS2017-84440-C2- 1-P funded by MCIN/AEI/10.13039/501100011033 “ERDF A way of making Europe”, PID2020-116567GB-C21 funded by MCIN/AEI/10.13039/501100011033 and the project PROMETEO/2020/079 (Generalitat Valenciana).
\hfill\eject
|
\part{First Part}
\chapter{Introduction}
This is the first part of the lecture notes that grew out of the special course given by the author to the audience of exceptional undergraduate and first and second year graduate students at UCLA interested in geometry during the 2021/2022 academic year.
In these lecture notes we present an approach to the fundamental structures in differential geometry that uses the vernacular of sheaves, differential operators and horizontal subbundles.
This language is now standard in other disciplines, notably algebraic and complex geometry, but has been slow to enter the textbooks for students aspiring to work in differential geometry or use differential geometry in other geometric disciplines.
A notable exception is the recent textbook \cite{wedhorn} that uses the language of sheaves and cohomology to introduce the basics of smooth manifolds. This textbook, however, does not tap into the differential geometry territory per se, as it does not discuss metrics, connections, curvature, parallel transport or holonomy groups.
Another feature of this book is the treatment of the notion of connection from various different viewpoints: as a Koszul connection, as a covariant derivative, as an Ehresmann connection, as a parallel transport, etc. A similar extensive exposition of the notion of connection can be found in \cite{Spivak} and in \cite{postnikov}, but those expositions do not use the modern language that we try to develop in these notes.
A byproduct of the chosen approach is a relatively simple and conceptually clear construction of canonical connections (Levi-Civita, Chern, Bismut, etc.) without a heavy reliance on explicit local expressions as a starting point.
Finally, these lecture notes were heavily inspired by the course \cite{verbit} in differential geometry given by Misha Verbitsky in 2013.
\subsubsection*{Prerequisites}
This book assumes familiarity with the language of sheaves and the basics of smooth manifolds on the level of first year graduate courses.
Both of these topics are covered in the already mentioned \cite{wedhorn}. For a more extensive modern treatment of the basics of smooth manifolds see the wonderful textbook \cite{JLee}. Other references include \cite{JohnLee}, \cite{postnikov} or the classic \cite{kobayashi1996foundations}.
Familiarity with scheme theory and complex geometry will also be very helpful.
\subsubsection*{Acknowledgments}
First, we would like to thank Misha Verbitsky for a wonderful and broad education in geometry. The author took $5$ geometric courses and sat on numerous seminars organized by Misha while being an undergraduate student in the National Research University Higher School of Economics.
Second, I want to thank my scientific advisors Jack Huizenga and Alexei Pirkovskii for the ongoing support and interest in our work.
Last but not least, these lecture notes were prepared during the ongoing Russian invasion of Ukraine. Our heart is with those who had to bear the unnecessary and intolerable suffering one encounters during the war time. We also thank UCLA for the accommodations provided to the author during this rough period of life.
\chapter{Differential Operators}
\section{Differential operators over a ring}
\begin{definition}
A {\bf commutator} of two linear operators $A,B: V \to V$ is a linear operator defined as$$[A,B]:= A \circ B-B \circ A.$$
\end{definition}
The next key definition is due to A. Grothendieck.
\begin{definition}\label{def diff op} Let $R$ be a (commutative) $k$-algebra. A {\bf differential operator of order} $\mathbf{0}$ is a morphism of $R$-modules $D^0:R \to R$. Denote differential operators of order $0$ by $Diff^0(R)$.
More generally, a $k$-linear morphism $D^i: R \to R$ is a {\bf differential operator of order} $\mathbf{i}$, $D^i \in Diff^i(R)$, if for any $D^0 \in Diff^0(R)$ the commutator $[D^i, D^0]$ lies in $Diff^{i-1}(R)$.
We have $$Diff^0(R) \subset Diff^1(R) \subset Diff^2(R) \subset ...,$$
and the union $Diff^{\bullet}(R):=\cup_{i\geq0}Diff^i(R)$ is called the {\bf algebra of differential operators on} $\mathbf{R}$.
\end{definition}
\begin{example}[Differential operators of order $0$]
It follows from the above definition that $$Diff^0(R) \iso R$$ via $D^0 \mapsto D^0(1)$. The inverse morphism is given by $a \mapsto (m_a: b \mapsto ab).$ We will consistently use this identification throughout this chapter.
\end{example}
\begin{example}[Differential operators of order $1$] \label{derivations}
Let us show that $$Diff^1(R) \iso Der(R) \oplus R$$ as $k$-vector spaces. First note that for any derivation $X \in Der(R)$, any $m_a \in Diff^0(R)$ and any $b \in R$ we have $$[X, m_a](b) = X(ab) - aX(b) = aX(b) + bX(a) - aX(b) = X(a)b,$$ so we see that $[X, m_a]$ is just $m_{X(a)} \in Diff^0(R)$. Thus $Der(R) \subset Diff^1(R).$
Now consider $D^1 \in Diff^1(R)$. Define $$X(a) := D^1(a) - D^1(1) a, \quad \forall a \in R.$$
\begin{exercise}
Check that $X$ defined above is a derivation.
\end{exercise}
This exercise shows that any $D^1 \in Diff^1(R)$ can be written as $$D_1 = X + m_{D^1(1)}.$$
\end{example}
Note that if $D^i \in Diff^i(R)$ and $D^j \in Diff^j(R)$, then for any $m_a \in Diff^0(R)$:
\begin{align*} Diff^{i+j-1} &\ni [m_a, D^i]\circ D^j + D^i \circ [m_a, D^j]
=m_a \circ D^i \circ D^j - D^i \circ m_a \circ D^j \\& + D^i \circ m_a \circ D^j - D^i \circ D^j \circ m_a = m_a \circ D^i \circ D^j - D^i \circ D^j \circ m_a = [m_a, D^i \circ D^j] \
\end{align*}
which shows that $D^i \circ D^j \in Diff^{i+j}(R)$. This way, we get a {\bf filtered algebra} $\mathbf{Diff^{\bullet}(R)}$ of differential operators.
\begin{exercise}\label{ideals}
Let $D \in Diff^k(R)$.
\begin{enumerate}[(a)]
\item Prove that $D(a_1a_2...a_k) = D'(a_2a_3...a_k)+a_1D(a_2a_3...a_k),$ where $D'=[D, m_{a_1}] \in Diff^{k-1}(R)$.
\item Let $I \subset R$ be an ideal. Use induction to prove $D(I^{k+1}) \subset I.$
\end{enumerate}
\end{exercise}
\begin{exercise}\label{ass gr comm}
Use induction and the Jacobi identity $$[A, [B,C]]=[[A, B], C]+[B, [A,C]], \quad A,B,C \ - \text{linear operators},$$to prove that if $D^i \in Diff^i(R), D^j\in Diff^j(R)$, then the commutator $[D^i, D^j]$ lies in $Diff^{i+j-1}(R)$.
\end{exercise}
\subsection{Differential operators over the polynomial ring}
For this subsection, let $R:=\RR[t_1, ..., t_n]$.
\begin{lemma}\label{extension by zero}
If $D \in Diff^k(R)$ is a differential operator that vanishes on all polynomials of degree $\leq k$, then $D=0$.
\end{lemma}
\begin{proof} Note that if $t_1^{i_1}\cdot ... \cdot t_n^{i_n}$ is a monomial of degree $|I|=i_1+...+i_n = k+1$, then by Exercise \ref{ideals} (a) $$D(t_1^{i_1}\cdot ... \cdot t_n^{i_n}) = D'(t^{|I|-1}) + t_iD(t^{|I|-1}), \quad D'=[D, m_{t_i}] \in Diff^{k-1}(R), \quad |I|-1=k.$$
The first term is zero by induction on the order of a differential operator and the fact that $D'$ vanishes on polynomials of degree $\leq k-1$. The second term is zero by the assumption of the lemma.
The inductive step for monomials $t_1^{i_1}\cdot ... \cdot t_n^{i_n}$ of degree $|I|=i_1+...+i_n > k+1$ is handled using the same formula.
\end{proof}
\begin{lemma}\label{restriction}
Let $R_{\leq k} =: \RR[t_1, ..., t_n]_{\leq k}$ be the space of polynomials of degree $\leq k$. Let $A: R_{\leq k} \to R$ be a linear operator. Then there exists $D \in Diff^k(R)$ such that $D|_{R_{\leq k}}: R_{\leq k} \to R$ coincides with $A$.
\end{lemma}
\begin{proof}
Consider a differential operator $D_{f,i_1...i_n}=\frac{f(t_1, ..., t_n)}{i_1!\cdot ... \cdot i_n!}\left( \frac{d}{dt_1}\right)^{\circ i_1} \circ ... \circ \left( \frac{d}{dt_n} \right)^{\circ i_n}$ for $f \in R$. Note that $D_{f,i_1...i_n}$ is zero on all monomials of degree $<|I|=i_1+...+i_n$, $$D_{f,i_1 ... i_n} (t_1^{i_1}...t_n^{i_n}) = f(t_1, ..., t_n),$$ and is zero on all other monomials of degree $|I|$.
Thus $$D = \sum_{i_1+...+i_n \leq k} D_{A(t_1^{i_n}...t_n^{i_n}),i_1...i_n}$$is the desired differential operator as monomials $A(t_1^{i_n}...t_n^{i_n})$ with $i_1+...+i_n \leq k$ form the basis of $R_{\leq k}$.
\end{proof}
\begin{proposition}\label{poly generated}
The algebra of differential operators $Diff^{\bullet} \left(\RR[t_1, ..., t_n] \right)$ is generated by $t_i$ and $\frac{d}{dt_j}$ for $i,j=1,...,n$.
\end{proposition}
\begin{proof}
Let $D$ be a differential operator of order $k$. Consider $D|_{R_{\leq k}}: R_{\leq k} \to R$. By Lemma \ref{restriction}, there is $\tilde{D}\in Diff^k(R)$ that lies in the subalgebra generated by $t_i$ and $\frac{d}{dt_j}$ and that satisfies $D|_{R_{\leq k}} = \tilde{D}|_{R_{\leq k}}$. The differential operator $D - \tilde{D} \in Diff^k(R)$ vanishes on all polynomials of degree $\leq k$ by construction and therefore, by Lemma \ref{extension by zero}, $D = \tilde{D}$.
\end{proof}
\begin{exercise}
Prove that $Diff^{\bullet}(R)$ is an $\RR$-algebra generated by $t_1, ..., t_n, \frac{d}{dt_1}, ..., \frac{d}{dt_n}$ with relations
$$[t_i, t_j]=0, \quad \left[\frac{d}{dt_i}, \frac{d}{dt_j}\right]=0, \quad \forall i,j,$$
$$\left[t_i, \frac{d}{dt_i} \right]=1,$$
$$\left[ \frac{d}{dt_i}, t_j \right] = 0, \quad i \neq j.$$
\end{exercise}
\subsection{Differential operators over the ring of smooth functions}
In this subsection, we consider differential operators over the ring $C^{\infty}(\RR^n)$ of smooth function on $\RR^n$. We denote by $x_i$ the standard coordinate functions $x_i: (p_1,...p_n) \mapsto p_i.$
Let $m_x = \{ f \in R \ | f(x) =0 \}$ be the maximal ideal of smooth functions vanishing at point $x$. By Exercise \ref{ideals} (b), for any differential operator of order $i$ and $f \in m_x^{i+1}$ we have $D(f) =0$. This observation allows us to prove the following key lemma.
\begin{lemma}\label{lemma vanish}
Any differential operator $D \in Diff^i(C^{\infty}(\RR^n))$ that vanishes on polynomials of degree $\leq i$ is trivial.
\end{lemma}
\begin{proof}
Consider $D \in Diff^i(C^{\infty}(\RR^n))$ that vanishes on $\RR[x_1, ..., x_n]_{\leq i}$. Take any $f \in C^{\infty}(\RR^n)$ and any point $x \in \RR^n$. Take the Taylor polynomial $P$ for function $f$ of degree $i$ at point $x$. We have $$f-P \in m_x^{i+1}.$$ By the observation above and the assumption that $D$ vanishes on $\RR[x_1, ..., x_n]_{\leq i}$ we conclude $$D(f-P) \in m_x \quad \implies \quad D(f-P) =0 \quad \implies \quad D(f) = D(P) + D(f-P) =0 .$$
\end{proof}
Next we prove an analogue of Proposition \ref{poly generated} for $Diff^{\bullet}(C^{\infty}(\RR^n)$.
\begin{proposition}\label{classic}
The algebra $Diff^{\bullet}(C^{\infty}(\RR^n))$ is generated by $\frac{d}{dx_i}$ as a $C^{\infty}(\RR^n)$-algebra.
\end{proposition}
\begin{proof} Similar to the proof of Lemma \ref{restriction}, given an $\RR$-linear map $$A: \RR[x_1, ..., x_n]_{\leq k} \to C^{\infty}(\RR^n)$$ we can explicitly construct a differential operator $$D \in C^{\infty}(\RR^n)\left<\frac{\partial}{\partial x_1}, ..., \frac{\partial}{\partial x_n}\right> \subset Diff^k(C^{\infty}(\RR^n))$$ whose restriction to $\RR[x_1, ..., x_n]_{\leq k}$ coincides with $A$ as a sum of differential operators
$$D = \sum_{i_1+...+i_n \leq k} D_{A(x_1^{i_n}...x_n^{i_n}),i_1...i_n} \ ,$$
$$D_{f,i_1...i_n}=\frac{f(x_1, ..., x_n)}{i_1!\cdot ... \cdot i_n!}\left( \frac{\partial}{\partial x_1}\right)^{\circ i_1} \circ ... \circ \left( \frac{\partial}{\partial x_n} \right)^{\circ i_n}, \quad f \in C^{\infty}(\RR^n).$$
Now consider an arbitrary differential operator $D$ of order $k$. Consider the restriction $$D|_{\RR[x_1, ..., x_n]_{\leq k}} : \RR[x_1, ..., x_n]_{\leq k} \to C^{\infty}(\RR^n).$$ As above, construct a differential operator $\tilde{D}$ of order $i$ as a sum of $D_{f, i_1,..., i_n}$ with the property $D=\tilde{D}$ on $\RR[x_1, ..., x_n]_{\leq k}$.
Consider $\overline{D} = D - \tilde{D} \in Diff^k(C^{\infty}(\RR^n)).$ By construction $\overline{D}$ vanishes on all polynomials of order $\leq k$. By Lemma \ref{lemma vanish} $\overline{D} = 0$ or, equivalently $$D = \tilde{D} \in C^{\infty}(\RR^n)\left<\frac{\partial}{\partial x_1}, ..., \frac{\partial}{\partial x_n}\right> \subset Diff^k(C^{\infty}(\RR^n)).$$
\end{proof}
This way, in the case $R = C^{\infty}(\RR^n)$ Proposition \ref{classic} allows to relate Definition \ref{def diff op} to the historic analytical definition of a differential operator $D \in Diff^k(C^{\infty}(\RR^n))$ as an expression $$D= \sum_{i_1+...+i_n \leq k} f_{i_1, ..., i_n}(x_1, ..., x_n) \frac{\partial^{i_1+...+i_n}}{\partial x_1^{i_1}... x_n^{i_n}}, \quad f_{i_1, ..., i_n}(x_1, ..., x_n) \in C^{\infty}(\RR^n).$$
The key Definition \ref{def diff op} can be sheafified if one replaces a ring $R$ with a \emph{sheaf} of algebras $\mathcal{R}$ on a topological space $X$.
\begin{definition}\label{sheaf def} Let $\mathcal{R}$ be a sheaf of (commutative) $k$-algebras on a topological space $X$. A {\bf sheaf of differential operators on} $\bm{\mathcal{R}}$ is a functor
$$\{U \subset X \ \text{open} \}^{op} \to \text{Filtered $k$-algebras},$$
$$U \mapsto Diff^{\bullet}(\mathcal{R}(U)),$$
where $Diff^{\bullet}(\mathcal{R}(U))$ is the filtered algebra of differential operators over $\mathcal{R}(U)$ in the sense of Definition \ref{def diff op}. One readily checks that this assignment satisfies the sheaf axioms.
We will denote this sheaf by $Diff^{\bullet}_{\mathcal{R}}$ or simply by $Diff^{\bullet}$ if the underlying sheaf $\mathcal{R}$ is clear from the context. Observe that after forgetting the filtration the sheaf of differential operators is a subsheaf of the sheaf of internal homomorphisms
$$ Diff^{\bullet}_{\mathcal{R}} \subset \sheafhom_{k}(\mathcal{R}, \mathcal{R}),$$ where homomorphisms are taken with respect to the $k$-vector space structure of $\mathcal{R}$.
Finally, if $D \in Diff^i_{\mathcal{R}}(U) = Diff^i(\mathcal{R}(U))$, then we say that $D$ is a {\bf differential operator of order $\bm{i}$ on $\bm{U}$}. \end{definition}
\section{Algebra of symbols of differential operators over a ring}\label{symbols section}
Given a $k$-algebra $R$ or a sheaf of $k$-algebras $\Rsheaf$, consider the filtered algebra of differential operators $$Diff^0(R) \subset Diff^1(R) \subset Diff^2(R) \subset ... \subset Diff^{\bullet}(R)$$and the filtered sheaf of algebras of differential operators $$Diff^0_{\Rsheaf} \subset Diff^1_{\Rsheaf} \subset Diff^2_{\Rsheaf} \subset ... \subset Diff^{\bullet}_{\Rsheaf}.$$
\begin{proposition}\label{commutative}
The associated graded algebra $\oplus_{i \geq 0} Diff^i(R)/Diff^{i-1}(R)$ and the associated graded sheaf of algebras $\oplus_{i \geq 0} Diff^i_{\Rsheaf}/Diff^{i-1}_{\Rsheaf}$ are commutative.
\end{proposition}
\begin{proof}
By Exercise \ref{ass gr comm}, for $D^i \in Diff^i(R), D^j \in Diff^j(R)$, we have $$[D^i, D^j] = D^i \circ D^j - D^j \circ D^i \in Diff^{i+j-1}(R).$$Hence as elements of $Diff^{i+j}(R)$ $$D^i\circ D^j \equiv D^j \circ D^i \quad \mod Diff^{i+j-1}(R).$$
\end{proof}
\begin{definition}
The associated graded algebra $$ Symb^* (Diff^{\bullet}(R)):=\oplus_{i \geq 0} Symb^i :=gr(Diff^{\bullet}(R)) = \oplus_{i \geq 0}Diff^i(R)/Diff^{i-1}(R)$$ is called the {\bf algebra of symbols of differential operators over} $\bm{R}$. Note that $S^0 = Diff^0(R)$ is canonically isomorphic to $R$, so $Symb (Diff^{\bullet}(R))$ is a (commutative) $R$-algebra.
We can similarly define a {\bf sheaf of (commutative) $\bm{\mathcal{R}}$-algebras of symbols of differential operators over $\bm{\mathcal{R}}$}: $$ Symb^* (Diff^{\bullet}_\Rsheaf):=\oplus_{i \geq 0} Symb^i :=gr(Diff^{\bullet}_\Rsheaf) = \oplus_{i \geq 0}Diff^{i}_\Rsheaf/Diff^{i-1}_\Rsheaf.$$
\end{definition}
Here
\begin{example}\label{symb 1}
Let us show that $Symb^1 = Diff^1(R)/Diff^0(R)$ is isomorphic to the $R$-module $Der(R)$ of derivations of $R$.
\end{example}
\begin{proof}
In Example \ref{derivations} we showed that every derivation is also a differential operator of order $1$. This way we get a morphism $$Der(R) \inj{} Diff^1(R) \surj Diff^1(R)/Diff^0(R).$$
The inverse morphism is defined by $$ Diff^1(R) \ni D \to D - m_{D(1)} \in Der(R).$$ One easily checks that $D-m_{D(1)}$ is indeed a derivation. Note that for $D=m_a \in Diff^0(R)$ the difference $D - m_{D(1)} = m_a - m_{a \cdot 1} = 0$, so the inverse morphism is in fact defined as a map $$Diff^1(R)/ Diff^0(R) \to Der(R).$$
\end{proof}
\begin{proposition}\label{key homo}
There is a natural homomorphism of graded commutative $R$-algebras $$Sym^*_R(Der(R)) = \oplus_{i \geq 0}Sym^i_R(Der(R)) \to \oplus_{i \geq 0} Symb^i = Symb^*(Diff^{\bullet}(R))$$that extends the isomorphism $Der(R) \stackrel{\iso}{\to} Symb^1$,
and an analogous natural homomorphism of sheaves of graded commutative $\Rsheaf$-algebras
$$Sym^*(Der_{\Rsheaf}) \to Symb^*(Diff^{\bullet}_{\Rsheaf}).$$
\end{proposition}
\begin{proof}
Consider the following diagram of solid arrows:
\begin{center}\begin{tikzcd}
Sym^*(Der(R)) \arrow[r, dashed] & Symb^*(Diff^{\bullet}(R)) \\
Der(R) \arrow[u, hook] \arrow[r, "\iso"] & Symb^1 \arrow[u, hook].
\end{tikzcd} \end{center}
By Proposition \ref{commutative} the target algebra $Symb^*(Diff^{\bullet}(R))$ is commutative.
The upper horizontal dashed arrow is then just the universal property of the symmetric algebra applied to the bottom isomorphism of Example \ref{symb 1} followed by the inclusion ${Symb^1 \inj{} Symb^*(Diff^{\bullet}(R))}$ on the right of the commutative diagram.
\end{proof}
\section{Algebra of symbols of differential operators over the ring of smooth functions on a manifold}
In this section we will specialize the results of \S \ref{symbols section} on the symbols of differential operators to the setting of differential operators on a smooth manifold $M$.
Consider the following objects:
\begin{itemize}
\item the ring $R=C^{\infty}(M)$ of smooth functions on a smooth manifold $M$,
\item the sheaf $\Rsheaf=C^{\infty}_M$ of smooth functions on $M$ (so that $R=C^{\infty}(M)=\Gamma(M, C^{\infty}_M)=\Gamma(M, \Rsheaf)$),
\item the algebra $Diff^{\bullet}(M):=Diff^{\bullet}(C^{\infty}(M))$ of global differential operators on $M$,
\item the sheaf of algebras $Diff^{\bullet}_M:=Diff^{\bullet}_{C^{\infty}_M}$ of differential operators on $M$ in the sense of Definition \ref{sheaf def} (so that ${Diff^{\bullet}(M) = \Gamma(M, Diff^{\bullet}_M))}$.
\end{itemize}
While the similarity of notation for the global and local objects might appear slightly confusing at first, the reader will soon find that working with sheafified objects is almost identical to working with global objects. This justifies our choice of notation.
The well known equivalence between vector fields and derivations of the sheaf of smooth functions $C^{\infty}_M$ yields an isomorphism of sheaves
$$Der(C^{\infty}_M) \iso TM,$$where we identify the tangent bundle $TM$ with the locally free sheaf of its sections (vector fields). In this setting, the morphism of Proposition \ref{key homo} becomes $$Sym^*(TM) \to Symb^*(Diff^{\bullet}_M).$$
The first goal of this section will be to prove the following key theorem.
\begin{theorem} Let $M$ be a smooth manifold. The natural homomorphism of sheaves
$$Sym^*(TM) \to Symb^*(Diff^{\bullet}_M)$$ is an isomorphism.
\end{theorem}
An immediate corollary of this theorem is that the sheaf of symbols of differential operators on a manifold is \emph{locally free}. We will prove the above theorem as a series of lemmas.
\newpage
\addcontentsline{toc}{chapter}{Bibliography}
\bibliographystyle{alpha}
|
\section{Introduction}
\input{intro}
\section{Related Work}
\input{related}
\section{Proposed Co-Teaching Method}
\input{method}
\section{Experiments}
\input{experiment}
\vspace{-10pt}
\section{Summary and Conclusions}
\label{sec:conclusion}
We have presented Co-Teaching (CT) as new method for both UDA and UDE. Our multi-class image classification experiments on two public benchmarks, \ie Office-Home and DomainNet, and semantic image segmentation experiments on another two public sets, \ie Cityscapes and ACDC, support our conclusions as follows. \xredit{Due to the existence of cross-domain visual ambiguity, a domain-specific model is not universally applicable to handle samples from its targeted domain.
With its ability to resolve such visual ambiguity, CT beats strong baselines} on UDA, \ie GSFDA for image classification and FDA for driving scene segmentation. CT is also better than the prior art on UDE, \ie KDDE, for both tasks.
\clearpage
\bibliographystyle{splncs04}
\subsection{Task 1. Multi-Class Image Classification} \label{ssec:exp-img-clf}
\subsubsection{Experimental Setup} \label{sssec:img-clf-setup}
We adopt two public collections, Office-Home \cite{officehome} and DomainNet\cite{domainnet}. Office-Home contains 15,588 images of 65 object classes common in office and home scenes, \eg chair, table and TV. There are four different domains, \ie artistic images (A), clip art (C), product images (P), and real-world images (R). DomainNet, previously used in the Visual Domain Adaptation Challenge at ICCV 2019\footnote{\url{http://ai.bu.edu/visda-2019}}, has 362,470 images of 345 object classes from four domains, \ie clipart, painting, real, and sketch.
In order to evaluate a model's performance on both source and target domains, we adopt the data split provided by \cite{tomm21kdde}, where images per domain have been divided at random into two disjoint subsets, one for training and the other for test\footnote{\url{https://github.com/li-xirong/ude}}.
A specific UDE task is defined with one domain as $D_s$ and another domain as $D_t$. Per collection, by pairing its individual domains, we define 12 tasks in total.
\textbf{Competitor methods}. We include as a baseline ResNet-50 trained by standard supervised learning on $D_s$. As mentioned above, we compare with existing methods from the following three groups: \\
\indent$\bullet$ Method targeted at UDE: KDDE \cite{tomm21kdde}. \\
\indent$\bullet$ Methods targeted at UDA: DDC \cite{ddc}, DANN\cite{jmlr16-dann}, DAAN \cite{icdm19-daan}, CDAN \cite{nips18-cdan}, SRDC \cite{tang2020unsupervised}, \rb{PDA\cite{lv2021pareto}} and GSFDA\cite{yang2021generalized}. \\
\indent$\bullet$ Method technically related: MultiT \cite{you2017learning}.
As each method is provided with the same training data and eventually yields a specific ResNet-50 network for inference, such an experimental setup enables a head-to-head comparison between the different methods.
\textbf{Details of implementation}.
Following \cite{tomm21kdde}, we train networks by SGD with a momentum of $0.9$, an initial learning rateof $0.005$, and a weight decay of $0.0005$. The learning rate is decayed by $0.1$ every $30$ epochs on Office-Home and every $10$ epochs on DomainNet. A fixed number of training epochs is used, which is 100 for Office-Home (with batch size 32) and 30 on DomainNet (with batch size 96) as the latter is much larger. Models obtained at the last epoch are used for evaluation. Each method is run independently three times with averaged performance reported.
\textbf{Performance metric}. We report accuracy (\%), \ie the percentage of test images correctly classified.
\subsubsection{Results on Office-Home}
\input{sota}
\subsubsection{Ablation Study}
\input{ablation}
\subsection{Task 2. Driving Scene Segmentation}
\input{expseg}
\subsubsection{Experimental Setup}
We follow the setup of \cite{acdc}, using Cityscapes \cite{cityscapes} as $D_s$ and ACDC \cite{acdc} as $D_t$. Both datasets have pixel-level ground truth \wrt 19 traffic-related labels such as bicycle, road and sidewalk. Different from Cityscapes consisting of normal lighttime driving scenes, ACDC have four adverse conditions, \ie fog, nighttime, rain and snow, see Fig. \ref{fig:segmentation}. For both datasets, we adopt their official data splits, \ie 2,975 training and 500 test images in Cityscapes and 1,600 training and 406 test images in ACDC.
\textbf{Competitor methods}. We again compare with KDDE \cite{tomm21kdde}. Our choice of $\mathcal{N}_s$ is DeepLabv2\cite{chen2017deeplab} with ResNet-101 as its backbone, as has been used in the previous work \cite{acdc}. As for $\mathcal{N}_t$, we adopt the classic AdaSegNet \cite{tsai2018learning} and the more recent FDA\cite{yang2020fda}, which is found to be the most effective on ACDC \cite{acdc}.
\textbf{Details of Implementation}.
Following \cite{chen2017deeplab}, we adopt SGD with an initial learning rate of 0.001, a momentum of 0.9, and a weight decay of 0.0005. We train 75000 iterations with batch size 1. Previous work on semantic image segmentation \cite{panfilov2019improving} reports the Mixup technique has an adverse effect on the performance, which is also observed in our preliminary experiment on driving scene segmentation, we therefore uses CT without miCT for this task.
\textbf{Performance metric}. We report IoU per class, and mean IoU (mIoU) for measuring the overall performance.
\subsubsection{Results}
On the source domain $D_s$, we observe that the performance of the AdaSegNet and FDA decreases in \ref{tab:segmentation}, which confirms the necessity of the UDE for semantic segmentation.
Two-stage methods can alleviate the decline, and CT outperforms KDDE.
Under the UDA setting, CT boosts both UDA methods and achieves the best performance on the target domain $D_t$. More visually in \ref{fig:segmentation}, for the adverse conditions, we can obviously observe that CT can produce more accurate segmentation and improve the generalization, such as eliminating the misclassification of the sky into buildings.
On the expanded domain, CT achieves the best performance on the most categories in \ref{tab:segmentation}
and CT(FDA) achieves the best average performance for UDE. This demonstrates our method’s applicability to cross-condition driving scene segmentation.
\xredit{\cref{tab:seg_ambiguity} shows pixel-level accuracy. CT better handles pixels that have inconsistent predictions by $\mathcal{N}_s$ and $\mathcal{N}_t$}.
\input{tables/table_seg_ambiguity}
\input{tables/table_segmentation}
\begin{figure*}[tbh!]
\centering
\includegraphics[width=\columnwidth]{figures/segmentation.pdf}
\caption{
\textbf{Qualitative results of driving scene segmentation}. The first row is from $D_s$ (normal condition in the sunlight), while the other rows are from $D_t$ (adverse conditions in the nighttime, fog, snow and rain). Important difference between the results is marked out by white bounding boxes.}
\label{fig:segmentation}
\end{figure*}
\subsection{Problem Formalization}
We use $x$ to indicate a specific sample. For a manually labeled sample, we use $y$ to indicate its manual annotation, which can be a one-hot vector for multi-class image classification or a multi-dimensional binary mask for semantic image segmentation. Let $\mathcal{N}$ be a deep neural network, which is supposed to output $\mathcal{N}(x)$ that well matches the (unknown) label of a novel sample. Following \cite{tomm21kdde}, we formalize the UDE task as follows. Given a set of $n_s$ \emph{labeled} samples $\{(x_s,y_s)\}$ randomly sampled from a source domain $D_s$ and a set of $n_t$ \emph{unlabeled} samples $\{x_t\}$ sampled randomly from a target domain $D_t$, the goal of UDE is to train $\mathcal{N}$ that works for an expanded domain covering both $D_s$ and $D_t$. Accordingly, we use $D_{s+t}$ to indicate the expanded domain.
The previous approach to UDE, as aforementioned, is Knowledge Distillation Domain Expansion (KDDE) \cite{tomm21kdde}. At a high level, KDDE works in two stages. In the first stage, two domain-specific teacher networks $\mathcal{N}_s$ and $\mathcal{N}_t$ are trained, where $\mathcal{N}_s$ for $D_s$ is learned from the labeled set $\{(x_s, y_s)\}$ by standard supervised learning, while $\mathcal{N}_t$ for $D_t$ is trained on $\{(x_s, y_s)\}$ and $\{x_t\}$ by an off-the-shelf UDA method. In the second stage, knowledge distillation (KD) is performed to inject the dark knowledge of the teacher networks into a student network $\mathcal{N}_u$, which will be eventually used for inference. Depending on the domain identity of a training sample, KDDE selectively uses the two teachers, \ie $\mathcal{N}_s$ to deal with samples from $D_s$ and $\mathcal{N}_t$ for samples from $D_t$, see Eq. \cref{eq:kdde}.
\begin{equation} \label{eq:kdde}
\left\{
\begin{array}{ll}
\mathcal{N}_s & \leftarrow \mbox{supervised-learning}(\{(x_s,y_s)\}) \\
\mathcal{N}_t & \leftarrow \mbox{UDA}(\{(x_s,y_s)\}, \{x_t\}) \\
\mathcal{N}_u & \leftarrow
\left\{
\begin{array}{l}
\mbox{KD}(\mathcal{N}_s,\{x\}), \quad x\in D_s \\
\mbox{KD}(\mathcal{N}_t,\{x\}), \quad x\in D_t
\end{array}
\right.
\end{array}
\right.
\end{equation}
The two-stage property of KDDE ensures flexibility in choosing UDA methods for implementing $\mathcal{N}_t$. We thecrefore opt to inherit this property, but introduce a novel Co-Teaching (CT) method into the second stage.
\subsection{Framework}
\label{subsec:method_stage2}
As illustrated in Fig. \cref{fig:overview}, CT consists of knowledge distillation based CT (kdCT) and mixup based CT (miCT). Both kdCT and miCT are introduced to exploit the domain-specific advantages of the two teacher networks.
In particular, kdCT allows the student network to simultaneously learn the two teacher networks' dark knowledge about every training sample.
Meanwhile, miCT improves the generalization ability of the student network by using the mixup technique in a cross-domain manner. As the two implementations of CT are orthogonal to each other, they can be used either alone or jointly.
\subsubsection{Knowledge Distillation based Co-Teaching}
We depart from a standard KD process with one student network ($\mathcal{N}_u$) and one teacher network, either $\mathcal{N}_s$ or $\mathcal{N}_t$. Let us consider $\mathcal{N}_s$ for instance.
Given a set of samples $\{x\}$, KD from $\mathcal{N}_s$ to $\mathcal{N}_u$ is achieved by minimizing the Kullback-Leibler (KL) divergence between $\mathcal{N}_s(\{x\})$ and $\mathcal{N}_u(\{x\})$, defined as $KL(\mathcal{N}_s(\{x\}),\mathcal{N}_u(\{x\}))$.
In a similar manner, we define the loss of KD from $\mathcal{N}_t$ to $\mathcal{N}_u$ $KL(\mathcal{N}_t(\{x\}),\mathcal{N}_u(\{x\}))$. To perform multi-teacher KD, a straightforward solution is to average the two losses.
In the context of UDA / UDE, however, the two teacher networks are supposed to specialize in handling samples from their targeted domains.
Thecrefore, we shall not treat them equally.
To be more precise, for training samples from $D_s$, we expect that $\mathcal{N}_s$ leads the teaching process, while the $\mathcal{N}_t$ acts as an assistant. The opposite is true when exploiting training examples from $D_t$. To that end, we introduce a parameter $\gamma$ to weigh the importance of two teachers in the kdCT process as:
\begin{equation} \label{eq:kdct}
\begin{array}{l}
L_{kdct}(\{x\})=
\left\{
\begin{array}{l}
\gamma \cdot KL(\mathcal{N}_{s},\mathcal{N}_u)+
(1-\gamma) \cdot KL(\mathcal{N}_{t},\mathcal{N}_u) \quad x \in D_s
\\
\gamma \cdot KL(\mathcal{N}_{t},\mathcal{N}_u)+
(1-\gamma) \cdot KL(\mathcal{N}_{s},\mathcal{N}_u) \quad x \in D_t
\end{array} \right.
\end{array}
\end{equation}
Given \cref{eq:kdct}, KDDE can now be viewed as a special case of kdCT with $\gamma=1$. \rb{Such a mechanism provides the \emph{possibility} of making correct decision for the samples, which subjectively divided into the host domain but have more opposite domain styles. By contrast, KDDE, exclusively using one teacher network for one domain, lacks such a compensation mechanism by definition. }
Concerning the choice of $\gamma$, in principle we shall a value larger than $0.5$ to emphasize the leading-teacher network, which is $\mathcal{N}_s$ for samples from $D_s$ and $\mathcal{N}_t$ for samples from $D_t$.
It has been recognized that introducing certain randomness into the process of deep network training can make the network more resistant to noise and thus improve its robustness \cite{srivastava2014dropout,fawzi2016robustness,he2019parametric,zhang2018mixup}.
Thecrefore, we let $\gamma$ follow a probability distribution, instead of using a fixed value.
Since the beta distribution can produce diversified probability distributions with ease by adjusting two positive shape parameters denoted by $\alpha$ an $\beta$, we choose to use $\gamma \sim Beta(\alpha,\beta)$. Our randomness is introduced at a mini-batch level, by sampling randomly a specific $\gamma$ from the beta distribution per batch.
Given the two domain-specific losses $L_{kdct}(\{x_s\})$ and $L_{kdct}(\{x_t\})$ comuted by \cref{eq:kdct}, we define the overall loss $L_{kdct}$ as their sum.
\subsubsection{Mixup based Co-Teaching}
The Mixup technique \cite{zhang2018mixup}, which synthesizes a new training sample by a convex combination of two real samples, is shown to be effective for improving image classification networks. We thus re-purpose this technique to generate new domain-expanded samples denoted by $\{x_m\}$. In particular, $x_m$ is obtained by blending $x_s$ randomly chosen from $D_s$ with $x_t$ randomly chosen from $D_t$.
Our mixup based Co-Teaching (miCT) is implemented by transferring the two-teacher knowledge via the mixed samples to the student network. The teachers' joint knowledge \wrt $x_m$ is naturally creflected by their combined prediction denoted as $\hat{y}_m$.
Accordingly, the loss of miCT $L_{mict}$ is computed as
\begin{equation} \label{eq:mict}
\left\{ \begin{array}{ll}
x_m & = \lambda \cdot x_s + (1-\lambda) \cdot x_t, \\
\hat{y}_m &= \lambda \cdot \mathcal{N}_s(\{x_s\})+(1-\lambda) \cdot \mathcal{N}_t(\{x_t\}), \\
L_{mict} & =KL(\hat{y}_m , \mathcal{N}_u(\{x_m\})).
\end{array}\right.
\end{equation}
with $\lambda \sim Beta(1,1)$.
Both $L_{kdct}$ and $L_{mict}$ are KL-divergence based losses for knowledge distillation. So they can be directly added and minimized together for the joint use of kdCT and miCT.
\subsection{Applications} \label{subsec:method_stage1}
As \cref{fig:overview} shows, CT is a generic method for UDA and UDE. Depending on the choice of the network and the loss function, the method can be used with ease for multi-class image classification and semantic image segmentation. For the former task, a classification network, \eg ResNet \cite{he2016deep}, shall be used with the KD losses computed at the image level. For the latter, a segmentation network, \eg DeepLabv2 \cite{chen2017deeplab}, shall be adopted with the KD losses computed at the pixel level.
\subsection{Unsupervised Domain Adaptation(UDA)}
The major line of research on UDA is to learn domain-invariant feature representations,
either by domain discrepancy reduction \cite{ddc,icml17-jan,eccvw16-deep-coral,kang2019contrastive} or by adversarial training \cite{jmlr16-dann,nips18-cdan,cvpr2018-mcd,icdm19-daan,mm19-jada}. In Deep Domain Confusion (DDC)\cite{ddc}, the domain discrepancy between features from the source / target domains is defined as the maximum mean discrepancy (MMD) on the last layers of a deep image classification network. Contrastive Adaptation Network (CAN)\cite{kang2019contrastive} takes class information into account, measuring both intra-class and inter-class domain discrepancy.
Domain Adversarial Neural Network (DANN)\cite{jmlr16-dann} is among the first to introduce adversarial training into the context of UDA.
CDAN\cite{nips18-cdan} extends DANN by taking the multiplicative interaction of feature representations and class predictions as the input of its discriminator.
SRDC\cite{tang2020unsupervised} enhances its discriminator by clustering features from intermediate layers of the network. All the above is conducted for image classification.
We notice an increasing interest in extending UDA from the image level to the pixel level for cross-domain semantic image segmentation. AdaSegNet \cite{tsai2018learning} and Advent \cite{Vu_2019_CVPR} use adversarial training at the output space, while pixel-level adaptation is performed in DCAN \cite{wu2018dcan} and Cycada \cite{hoffman2018cycada}.
The state-of-the-art FDA \cite{yang2020fda} uses self-predicted pseudo labels for self-supervised training.
Our proposed CT method conceptually differs from the existing works as it essentially performs meta learning on top of a specific UDA method. Moreover, in contrast to the prior art which cares only the \xredit{target-domain} performance, CT aims for a broader scope that covers the source and target domains both. While GSFDA \cite{yang2021generalized} tries to prevent performance deterioration on the source domain by continual learning, \xredit{while PDA \cite{lv2021pareto} attempts to balance the source supervised loss and the cross-domain alignment loss by Pareto optimal solution}. Our experiment indicates that the deterioration remains.
CT works better for both image classification and semantic segmentation.
\textbf{Progress on UDE}.
CT is inline with KDDE \cite{tomm21kdde}, as both aim for training a model that suits the expanded domain by knowledge distillation (KD).
However, KDDE uses its teacher networks in a relatively limited manner, where the knowledge of the teacher network trained for the source (target) domain is transferred to the student network via the source (target) samples exclusively.
\xredit{Consequently, KDDE lacks the ability to exploit the potential of the teacher network derived from one domain in handling cross-domain visual ambiguity for the other domain}.
\textbf{Multi-Teacher Knowledge Distillation (MTKD)}.
MTKD has been used in the context of multi-source UDA, where multiple domain-adapted teacher models are trained by pairing the target domain with each of the multiple source domains \cite{kdadapt,belal2021knowledge}. Such a tactic is inapplicable to single-source UDA as this paper works on.
In the context of standard supervised learning, MultiT \cite{you2017learning} transfers the dark knowledge of multiple pre-trained teacher networks to a student network by minimizing the KL divergence between the averaged prediction of the teachers and the prediction of the student. Treating the teachers equally, MultiT is suboptimal for UDA/UDE.
\xredit{\textbf{Co-Teaching}. It is worth pointing out that the term co-teaching has been used in other contexts which conceptually and technically differ from ours. For supervised learning, Decoupling \cite{malach2017decoupling} and Co-Teach \cite{han2018co} aim for label denoising within a single domain. Decoupling simultaneously trains two models $h_1$ and $h_2$, updating them with samples having $h_1(x) \neq h_2(x)$ in a given mini-batch. Co-Teach alternately uses samples correctly classified by one model to train the other model. Since both methods are fully supervised, they are inapplicable for UDA/UDE where the target domain is unlabeled. For multi-target domain adaptation, CGCT \cite{roy2021curriculum} proposes a co-teaching strategy that uses a dual classifier head to provide pseudo labels for unlabeled target-domain samples. As our Co-Teaching is model-agnostic, any UDA method including CGCT can in principle be used to instantiate the UDA module of our method.}
\subsubsection{Results on DomainNet}
As DomainNet is much larger than Office-Home in terms of classes and samples, we are only able to train the following UDA methods, \ie CDAN, DANN and DDC.
As \cref{tab:total} shows,
using either DDC or CDAN as $\mathcal{N}_t$, our CT method again achieves the best performance on $D_s$, $D_t$ and $D_{s+t}$.
|
\section{Introduction}
Recently, many studies are dealing with diverse information from multiple sources finding relationships among them \cite{ramachandram2017multimodalsurvey}. Especially, deep learning based multi-modal learning has drawn big attention with its powerful performance. While the classic approaches \cite{huang2012trad1, cao2015trad2, wang2009trad3, wang2012trad4} need to design each modal feature manually, using Deep Neural Networks (DNNs) has the advantage of automatically learning meaningful representation from each modality.
Many applications including action recognition \cite{gao2019findfusion, kim2021multispectraldetection}, object detection \cite{eitel2015multimodalod}, and image/text retrieval \cite{zhen2019dscmr} have shown the effectiveness of multi-modal learning through DNNs by analyzing a phenomenon in multi-view.
\begin{figure}[t]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=8.5cm]{./Figure/Fig1.pdf}}
\end{minipage}
\vspace{-0.4cm}
\caption{Illustration of audio-visual multi-modal learning. (a) Fusion of two modalities.
(b) Learning from a common latent space of two modalities.
(c) The proposed framework provides an associative bridge between two modalities through memory. The audio (\ie, target) modality is recalled from memory by querying the visual (\ie, source) modality. Then, both the visual and the recalled audio modalities are utilized for a downstream task.}
\label{fig:1}
\vspace{-0.6cm}
\end{figure}
Audio-visual data is one of the main ingredients for multi-modal applications such as synchronization \cite{Chung16sync,chung2019perfect}, speech recognition \cite{afouras2018deep, petridis2018end}, and speech reconstruction from silent video \cite{l2w,akbari2018lip2audspec}. Along with the rapid increase of the demands for audio-visual applications, research efforts on how to efficiently handle audio-visual data have been made. There are two main streams on handling audio-visual data. First is to extract features from the two modalities and fuse them to achieve complementary effect, as shown in Fig.\ref{fig:1} (a). Such researches \cite{petridis2018end, afouras2018deep, noda2015avsr3} try to find the most suitable architecture of DNNs to fuse the modalities. Commonly used methods are early fusion, late fusion, and intermediate fusion. These fusion methods are known to be simple, yet effectively improve the performance of a given task. However, since both modalities are necessary for the fusion, these methods cannot work when one of the modalities is missing. Second is finding a common hidden representation of two modalities by training DNNs (Fig.\ref{fig:1} (b)). Different from the first method, it can utilize the shared information of both modalities from the learned cross-modal representation with uni-modal inputs. This can be achieved by finding the common latent space of different modalities using metric learning \cite{Chung16sync,chung2019perfect} or resembling the other modality which contains rich information for a given task using knowledge distillation \cite{zhao2020hearing}. However, reducing the heterogeneity gap \cite{huang2018crossmedia}, induced by inconsistent distribution of different modalities, is still considered as a challenging problem \cite{hu2019sdml,peng2019cmgan}.
In this paper, we propose a novel multi-modal bridging framework, especially in audio speech modality and visual face modality. The proposed framework brings the advantages of the two aforementioned audio-visual multi-modal learning methods, while alleviating the problems that each method contains. That is, it can obtain both audio and visual contexts during inference even when the uni-modal input is provided only. This gives explicit complementary knowledge with the multi-modal information to uni-modal tasks which could suffer from information insufficiency. Furthermore, our work can be free from finding a common representation of different modalities, as shown in Fig.\ref{fig:1}(c).
To this end, we propose to handle the audio-visual data through memory network \cite{weston2014memory, miller2016keyvalue} which contains two modality-specific memories: source-key memory and target-value memory. Each memory stores visual and audio features arranged in pairs, respectively. Then, an associative bridge is constructed between the two modality-specific memories, to access the target-value memory by querying the source-key memory with source modal representation. Thus, when one modality (\ie, source) is given, the proposed framework can recall the other saved modality (\ie, target) from target-value memory through the associative bridge. This enables it to complement the information of uni-modal input with the recalled target modal information. Therefore, we can enrich the task-solving ability of a downstream task. The proposed framework is verified on two applications using audio-visual data: lip reading, and speech reconstruction from silent video by using visual modality as source modality and audio modality as target modality.
In summary, the major contributions of this paper are as follows:
\begin{itemize}
\item We propose a novel audio-visual multi-modal bridging framework that enables it to utilize the information of multi-modality (\ie, audio and visual modalities) with uni-modal (\ie, visual) input during inference.
\item We verify the effectiveness of the proposed framework on two applications: lip reading and speech reconstruction from silent video and achieve state-of-the-art performances. Moreover, we visualize that the associative bridge adequately relates the source and target memories.
\item Through the proposed modality-specific memory operation (\ie, querying by source modality and recalling target modality), it does not need to find a common latent space of different modalities. We analyze it by comparing the proposed framework with the methods finding a common latent space of multi-modal data.
\end{itemize}
\section{Related Work}
\subsection{Multi-modal learning with audio-visual data}
Audio-visual multi-modal learning is one of the active research areas. There are two categories of audio-visual multi-modal learning using DNNs: fusion and finding a common latent space of cross-modal representation. The fusion methods \cite{tao2020avsr_multi, nefian2002avsr4, dupont2000avsr1} aim to exploit the complementary information of different modalities and achieve high performance compared to the uni-modal methods. They try to find the best fusion architecture of a given task \cite{afouras2018deep,neti2000avsr2,noda2015avsr3,lee2008avsr5}. However, as the fusion methods receive all modalities as inputs, they could not properly work if one of them is not available. The learning methods finding a common latent space from multi-modal data \cite{ngiam2011multimodal,hu2019sdml,andrew2013dcca,kan2016mvdn} aim to reduce the heterogeneity gap between the two modalities. Several works \cite{Chung16sync,chung2019perfect,feng2014crmr1} have proposed metric learning methods and adversarial learning methods to find the common representation. Other works \cite{zhao2020hearing, afouras2020asr} have proposed to learn from superior modality for a given task using knowledge distillation \cite{hinton2015distilling} which guides the learned feature to resemble the superior modal feature.
Although finding a shared latent space or guiding one modal representation to resemble the other has the advantage of using the common information between the two modalities with uni-modal inputs, reducing the heterogeneity gap between the multi-modal data is considered as a challenging problem \cite{zhen2019dscmr,peng2019cmgan}.
In this paper, we try to not only take the advantages of both methods, but also alleviate the problems of each method. We propose to handle the audio-visual data using two modality-specific memory networks connected with an associative bridge. During inference, the proposed framework can exploit both source and the recalled target modal contexts even when the input is uni-modal. Moreover, since each modality works on its corresponding modality-specific module, we can bypass the difficulty of finding a shared latent space.
\subsection{Memory network}
Memory network is a scheme to augment neural networks using external memory \cite{weston2014memory, sukhbaatar2015endmem}. They have shown the effectiveness of memory network on modeling long-term dependencies in sequential data \cite{lee2021videomemory}. Miller \etal \cite{miller2016keyvalue} introduce key-value paired memory structure where key memory is firstly used to address relevant memories with respect to a query, extracting addressed values from the value memory.
We utilize the key-value memory network \cite{miller2016keyvalue}, where the key memory is for saving the source modal features, and the value memory is for saving the target modal features. Thus, we can access both source and target modal contexts by recalling the saved target modal feature from the value memory when only source modality is available.
The memory network is also used in multi-modal modeling. Song \etal \cite{song2018cmmn} introduce a cross-modal memory network for cross-modal retrieval.
Huang \etal \cite{huang2019acmm} propose an aligned cross-modal memory network for few-shot image and sentence matching.
Using a shared memory, they encode memory-enhanced features, which will be used for image/text matching. Distinct from the previous methods, our proposed framework uses modality-specific memory network where source-key memory saves the source modality and target-value memory saves the target modality.
\subsection{Lip reading}
Lip reading is a task that recognizes speech as text from lip movements. Chung \etal \cite{chung2016lrw} propose word-level audio-visual corpus data and a baseline architecture. The performance of word-level lip reading is significantly improved with the architecture \cite{stafylakis2017reslstm, petridis2018end} of a 3D convolution layer, a ResNet-34, and Bi-RNNs. Some works \cite{weng2019twostream, xiao2020deformation} use both optical flow and video frames to capture fine-grained motion. Xu \etal \cite{xu2020discriminative} suggest a pseudo-3D CNN for the frontend which is more efficient compared to vanilla 3D CNN. Zhang \etal \cite{zhang2020cutout} show that the lip reading can be made beyond the lips by utilizing entire face as inputs. Martinez \etal \cite{martinez2020mstcn} improve the backend by changing the Bi-RNN into multi-scale temporal CNN.
It is widely known that the audio modality has superior knowledge for speech recognition than the visual modality by showing outstanding performance. In this paper, we try to complement the lip visual information by recalling the speech audio information from the proposed multi-modal bridging framework.
\begin{figure*}[t!]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=14.9cm]{./Figure/Fig2.pdf}}
\end{minipage}
\vspace{-0.3cm}
\caption{Overview of the proposed multi-modal bridging framework with an example of visual modality as a source and the audio modality as a target. The source-key memory is for saving source modal feature, and the target-value memory is for memorizing the target modal representations.}
\label{fig:2}
\vspace{-0.5cm}
\end{figure*}
\subsection{Speech reconstruction from silent video}
Speech reconstruction from silent video aims to generate acoustic speech signal from silent talking face video. Ephrat \etal \cite{ephrat2017vid2speech, ephrat2017improved} firstly generate speech using CNN and they improve it with a two tower CNN-based encoder-decoder architecture whose inputs are both optical flow and video frames. Akbari \etal \cite{akbari2018lip2audspec} propose to pretrain an auto-encoder to reconstruct the speech, whose decoder part is used to generate the speech from a face video. Vougioukas \etal \cite{ganbased} propose GAN based approach which maps video directly to audio waveform. Prajwal \etal \cite{l2w} attempt to learn on unconstrained single-speaker dataset. They present a model that consists of stacked 3D convolutions and an attention-based speech decoder, formulating the task as a sequence-to-sequence problem.
The speech reconstruction from silent video is considered as a challenging problem, due to the information insufficiency of face visual movement to fully represent the speech audio. We try to provide complementary information with the recalled audio representation through the proposed associative bridge with memory, and enhance its performance. With both the visual and the recalled audio contexts, we can generate high-quality speech in both speaker-dependent and speaker-independent settings.
\section{Multi-modality Associative Bridging}
The main objective of the proposed framework is to recall the target modal representation with only source modal inputs. To this end, (1) each modality-specific memory is guided to save the representative features of each modality, and (2) an associative bridge is constructed which enables it to recall the target modal representation by querying the source-key memory with source modal feature. As shown in Fig.\ref{fig:2}, the proposed multi-modality associative bridging framework is composed of two modality-specific memory networks: source-key memory $\mathbf{M}_{src}\in\mathbb{R}^{N \times C}$ and target-value memory $\mathbf{M}_{tgt}\in\mathbb{R}^{N \times D}$, where $N$ represents the number of memory slots, and $C$ and $D$ are the dimension of each modal feature, respectively. From the following subsections, we will describe the details of the proposed framework with examples of visual modality as source modality and the audio modality as target modality.
\subsection{Embedding modality-specific representations}
Each memory network inside the proposed framework saves generic representations of each modality. The generic visual and audio representations are produced from the respective modality-specific deep embedding modules. The visual (\ie, source modal) representation $f_{src}\in\mathbb{R}^{T \times C}$ is extracted by using spatio-temporal CNN that captures both spatial and temporal information, and the audio (\ie, target modal) representation $f_{tgt}\in\mathbb{R}^{T \times D}$ is embedded from 2D CNN whose input is preprocessed mel-spectrogram from raw audio signal, where $T$ represents the temporal length of each representation. Since the paired audio-video inputs are synchronized in time, the two embedding modules can be designed to output the same temporal length.
\subsection{Addressing modality-specific memory}
Based on the modality-specific representations, we firstly introduce how the source and target addressing vectors are formulated. The addressing vector refers to the guidance that determines where to assign weights on memory slots for a given query. Suppose that the source modal representation $f_{src}$ is given as a query, then the cosine similarity with source-key memory $\mathbf{M}_{src}$ is obtained,
\begin{align}
\label{eq:1}
s^{i,j}_{src}=\frac{\textbf{M}_{src}^i \cdot f^j_{src} }{||\textbf{M}_{src}^i||_2 \cdot ||f^j_{src}||_2 },
\end{align}
where $s_{src}^{i,j}$ represents the cosine similarity between $i$-th memory slot of source-key memory and source modal feature in $j$-th temporal step.
Next, the relevance probability is obtained using Softmax function as follows,
\begin{align}
\label{eq:2}
\alpha^{i,j}_{src}=\frac{\exp{(r \cdot s^{i,j}_{src})}}{\sum_{k=1}^N{\exp{(r \cdot s^{k,j}_{src})}}},
\end{align}
where $r$ is a scaling factor for similarity. By calculating the probability over the entire memory slot, the source addressing vector for the $j$-th temporal step $A^j_{src}=\{\alpha^{1,j}_{src},\alpha^{2,j}_{src},\dots \alpha^{N,j}_{src}\}$ can be obtained.
The same procedure is applied for target modal representation $f_{tgt}$ and target-value memory $\mathbf{M}_{tgt}$ to produce the target addressing vector, $A^j_{tgt}=\{\alpha^{1,j}_{tgt},\alpha^{2,j}_{tgt},\dots \alpha^{N,j}_{tgt}\}$ of $j$-th temporal step. The addressing vectors will be utilized in recalling the saved representations inside memory and connecting the two modality-specific memories, in the following subsections.
\subsection{Memorizing the target modal representations}
The obtained target addressing vector $A^j_{tgt}$ is to correctly match the target-value memory $\mathbf{M}_{tgt}$ for reconstructing target representation $\hat{f}^j_{tgt}$. To do so, the target-value memory $\mathbf{M}_{tgt}$ is trained to memorize the proper target modal representation ${f}^j_{tgt}$. We firstly obatin the reconstructed target representation $\hat{f}^j_{tgt}$ as follows,
\begin{align}
\label{eq:4}
\hat{f}^j_{tgt} = A^j_{tgt} \cdot \mathbf{M}_{tgt}.
\end{align}
Then, we design the reconstruction loss function to guide the target-value memory $\mathbf{M}_{tgt}$ to save the proper representation. We minimize the Euclidean distance between the target representation and the reconstructed representation,
\begin{align}
\label{eq:5}
\mathcal{L}_{save} = \mathbb{E}_j[||f^j_{tgt} - \hat{f}^j_{tgt}||^2_2].
\end{align}
With the saving loss, the target-value memory $\mathbf{M}_{tgt}$ saves the representative features of the target modality. Therefore, the recalled target modal representation $\hat{f}^j_{tgt}$ from target-value memory $M_{tgt}$ is able to represent the original target modal representation $f_{tgt}$.
\subsection{Bridging source and target memories}
\vspace{-0.11cm}
To recall the target modal representation from the target-value memory by using the source-key memory and source modal inputs, we construct an associative bridge between the two modality-specific memories. Specifically, the source-key memory is utilized to provide the bridge between source and target modalities in the form of the source addressing vector. That is, through the source addressing vector $A^j_{src}$, the corresponding saved target representation is recalled.
To achieve this, the source addressing vector $A^j_{src}$ is guided to match to the target addressing vector $A^j_{tgt}$ with the following bridging loss,
\begin{align}
\label{eq:6}
\mathcal{L}_{bridge} = \mathbb{E}_j[D_{KL}(A^j_{tgt}||A^j_{src})],
\end{align}
where $D_{KL}(\cdot)$ represents Kullback–Leibler divergence \cite{kullback1951kld}.
With the bridging loss, the source-key memory saves the source modal representations in the same location, where the target-value memory saves the corresponding target modal features. Therefore, when a source modal representation is given, the source-key memory provides the location information of the corresponding saved target modal representation in the target-value memory, using the source addressing vector.
\subsection{Applying for downstream tasks}
\vspace{-0.11cm}
Through the associative bridge and the modality-specific memories, we can obtain the recalled target modal feature $v_{tgt}$ by using source addressing vector $A_{src}$ as follows,
\begin{align}
\label{eq:3}
v^j_{tgt} = A^j_{src} \cdot \mathbf{M}_{tgt}.
\end{align}
Here, the target modal feature $v_{tgt}$ is recalled by querying the source-key memory $\mathbf{M}_{src}$ with the source modal representation $f_{src}$. Thus, we do not need the target modal inputs for recalling the target modal feature. Then, we can apply the recalled target modal feature for a downstream task in addition to the source modality, improving task performance by exploiting the complementary information.
\subsection{End-to-End training}
The proposed framework is trainable in an end-to-end manner, including the modality-specific embedding modules, memory networks, and the downstream sub-networks. To this end, the following task loss is applied,
\begin{align}
\label{eq:7}
\mathcal{L}_{task} = g(h(f_{src}\oplus v_{tgt});y) + g(h(f_{src}\oplus f_{tgt});y),
\end{align}
where $g(\cdot)$ is a loss function corresponding to the downstream task, $h(\cdot)$ is a fusion layer such as a linear layer, $y$ represents label, and $\oplus$ represents concatenation. The first term of the loss function is related to the performance on a given task that utilizes both the source and the recalled target modalities. The second term guarantees that the target modal embedding module learns the meaningful representations that will be saved into target-value memory in an end-to-end manner.
Finally, the total loss function is defined as a sum of the all loss functions,
\begin{align}
\label{eq:8}
\mathcal{L}_{total} = \mathcal{L}_{save} + \mathcal{L}_{bridge} + \mathcal{L}_{task}.
\end{align}
The pseudo code for training the proposed framework is shown at Algorithm \ref{alg:1}.
\makeatletter
\renewcommand{\ALG@beginalgorithmic}{\small}
\algrenewcommand\alglinenumber[1]{\small #1:}
\makeatother
\begin{algorithm}[t!]
\caption{Training algorithm of the proposed framework}
\label{alg:1}
\begin{algorithmic}[1]
\State {\bf Inputs}: The training pairs of source and target modal inputs $(X_{src}, X_{tgt})$ and label $y$, where $X_{src} = \{x^{l}_{src}\}_{l=1}^L$, $X_{tgt}=\{x^{s}_{tgt}\}_{s=1}^S$. The learning rate $\eta$.
\State {\bf Output}: The optimized parameters of the network $\Phi$
\vspace{0.05in}
\hrule
\vspace{0.05in}
\State Randomly initialize parameters of the network $\Phi$
\For{each iteration}
\State ${f}_{src}=\{{f}^j_{src}\}_{j=1}^T=$Source\_embed$(X_{src})$
\State ${f}_{tgt}=\{{f}^j_{tgt}\}_{j=1}^T=$Target\_embed$(X_{tgt})$
\For{$j = 1,2,...,T$}
\State $A^j_{src} = $Softmax$(r\cdot$CosineSim$(\mathbf{M}_{src},f^j_{src}))$
\State $A^j_{tgt} = $Softmax$(r\cdot$CosineSim$(\mathbf{M}_{tgt},f^j_{tgt}))$
\State $\hat{f}^j_{tgt} = A^j_{tgt} \cdot \mathbf{M}_{tgt}$
\State $v^j_{tgt} = A^j_{src} \cdot \mathbf{M}_{tgt}$
\EndFor
\State $\mathcal{L}_{save}=\sum_{j=1}^{T}||f^j_{tgt} - \hat{f}^j_{tgt}||^2_2$
\State $\mathcal{L}_{bridge}=\sum_{j=1}^{T}D_{KL}(A^j_{tgt}||A^j_{src})$
\State $\mathcal{L}_{task}=g(h(f_{src}\oplus v_{tgt});y) + g(h(f_{src}\oplus f_{tgt});y)$
\State $\mathcal{L}_{tot} = \mathcal{L}_{save} / T + \mathcal{L}_{bridge} / T + \mathcal{L}_{task}$
\State Update $\Phi \leftarrow \Phi - \eta \nabla_{\Phi} \mathcal{L}_{tot}$
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Experiments}
The main strength of the proposed audio-visual bridging framework is that it is possible to use multi-modal representation even if only one modal input is available. Therefore, we can enhance the uni-modal downstream tasks by exploiting complementary information from the recalled modal features. We show the effectiveness of the proposed framework on two applications, lip reading and speech reconstruction from silent video, each of which takes visual modality as an input. Therefore, visual modality is utilized as a source modality and audio modality is used as a target modality.
\subsection{Application 1: Lip reading}
\label{sec:4.1}
Lip reading is a task that recognizes speech by solely depending on lip movements. We apply the proposed multi-modal bridging framework to the lip reading to complement the visual context by bringing superior knowledge of the audio through the associative bridge and to enhance the performance.
\vspace{-0.3cm}
\subsubsection{Dataset}
\vspace{-0.2cm}
We utilize two public benchmark databases for word-level lip reading, LRW \cite{chung2016lrw} and LRW-1000 \cite{yang2019lrw1000}. Both datasets are composed of 25 fps video and 16kHz audio.
{\bf LRW} \cite{chung2016lrw} is a large-scale word-level English audio-visual dataset. It includes 500 words with a maximum of 1,000 training videos each.
For the preprocessing, the video is cropped into 136 $\times$ 136 size centered at the lip, resized into 112 $\times$ 112, and converted into grayscale. For the data augmentation, we use random horizontal flipping and random erasing for all frames in a video consistently. The audio is preprocessed using window size of 400, hop size of 160, and 80 mel-filter banks. Thus, the preprocessed mel-spectrogram has 100 fps with 80 spectral dimensional features. We use SGD optimizer, batch size of 320, and initial learning rate of 0.03.
{\bf LRW-1000} \cite{yang2019lrw1000} is Mandarin words audio-visual dataset. It consists of 718,018 video samples with 1,000 word classes.
The same preprocessing and data augmentation are applied as in LRW preprocessing except for cropping because the dataset is already cropped. Moreover, since the audio provided from the dataset is longer than the word boundary by 0.4-sec, we use the video as the same length as the audio. We use Adam \cite{kingma2014adam} optimizer, batch size of 60, and initial learning rate of 0.0001.
\vspace{-0.3cm}
\subsubsection{Architecture}
For the baseline architecture, we follow the typical architecture \cite{petridis2018end, stafylakis2017reslstm} whose visual embedding module consists of one 3D convolution layer and ResNet-18 \cite{he2016resnet}, and backend module is composed of 2 layered Bi-GRU \cite{schuster1997bidirectional}. We design the audio embedding module to output the same sequence length as that of the visual embedding module. For the task loss $g(\cdot)$, cross entropy loss is applied.
The details of the network architecture can be found in supplementary.
\vspace{-0.3cm}
\subsubsection{Results}
In order to verify the effectiveness of the proposed multi-modal bridging framework on complementing the visual modality with recalled audio modality, we compare the word-level lip reading using only visual modal inputs on benchmark datasets with the state-of-the-art methods.
Table \ref{table:1} shows the overall lip reading performances on LRW and LRW-1000 datasets. Our proposed framework achieves the highest accuracies among the previous approaches on both datasets. Especially for LRW-1000, which is known to be a difficult dataset due to unbalanced training samples, the proposed method attains a large improvement of $5.58\%$ from the previous state-of-the-art method \cite{zhang2020cutout}. From this result, we can confirm that the proposed framework is even more effective for the difficult task with the ability of complementing the insufficient visual information with the recalled audio. Moreover, since our multi-modal associative bridging framework is not dependent on the downstream architecture, deep architecture such as temporal CNN can be adopted to the proposed method to improve word prediction performance.
We also conduct an ablation study with four different models for each language (\ie, $N$=0, 44, 88, 132 for English and $N$=0, 56, 112, 168 for Mandarin) to examine the effect of the number of memory slots. The ablation results on memory slot size are reported in supplementary material. For LRW, the best word accuracy of 85.41\% is achieved when $N$=88. The proposed framework improves the baseline with a margin of 1.27\%. For LRW-1000, the best word accuracy is 50.82\% when $N$=112 by improving the baseline performance with 5.89\%.
The proposed framework improves the performance regardless of the number of memory slots from the baseline in both languages.
By employing the recalled audio feature as complementary information of the visual context, the proposed framework successfully refines the word prediction achieving state-of-the-art performance.
\begin{table}[]
\renewcommand{\arraystretch}{1.2}
\renewcommand{\tabcolsep}{7mm}
\resizebox{0.9999\linewidth}{!}{
\begin{tabular}{ccc}
\hline \hline
\textbf{Method} & \textbf{LRW} & \textbf{LRW-1000} \\ \hline
Yang \etal \cite{yang2019lrw1000} & 83.0 & 38.19 \\
Multi-Grained \cite{wang2019multigrained} & 83.3 & 36.91 \\
PCPG \cite{luo2020pcpg} & 83.5 & 38.70 \\
Deformation Flow \cite{xiao2020deformation} & 84.1 & 41.93 \\
MI Maximization \cite{zhao2020mi} & 84.4 & 38.79 \\
Face Cutout \cite{zhang2020cutout} & 85.0 & 45.24 \\
MS-TCN \cite{martinez2020mstcn} & 85.3 & 41.40 \\ \hline
\textbf{Proposed Method} & \textbf{85.4} & \textbf{50.82} \\ \hline \hline
\end{tabular}}
\vspace{0cm}
\caption{Lip reading word accuracy comparison with visual modal inputs on LRW and LRW-1000 dataset.}
\vspace{-0.5cm}
\label{table:1}
\end{table}
\subsection{Application 2: Speech reconstruction from silent video}
\label{sec:4.2}
Speech reconstruction from silent video is a task of inferring the speech audio signal by watching the facial video. To demonstrate the effectiveness of the proposed multi-modal bridging framework, we apply the proposed framework to the speech reconstruction from silent video task to provide the recalled audio context in an early stage of decoding for generating a high quality speech.
\vspace{-0.2cm}
\subsubsection{Dataset}
\vspace{-0.2cm}
{\bf GRID} dataset \cite{cooke2006grid} contains short English phrases with 6 words from predefined dictionary. The video and audio are sampled with rate of 25fps and 16kHz, respectively. Following \cite{ganbased,l2w}, subjects 1, 2, 4, and 29 are taken for speaker-dependent task. For the speaker-independent setting, we follow the same split as \cite{ganbased} which uses 15 subjects for training, 5 for validation, and 5 for testing. For the preprocessing, the face is detected, cropped and resized into 96 $\times$ 96 size. The audio is preprocessed with window size of 800, hop size of 160, and 80 mel-filter banks, becoming 80-dimensional mel-spectrogram in 100 fps. We use Adam optimizer, batch size of 64, and initial learning rate of 0.001.
\vspace{-0.2cm}
\subsubsection{Architecture}
\vspace{-0.2cm}
For the baseline architecture, we follow the state-of-the-art method \cite{l2w} whose visual embedding module is composed of 3D CNN and Bi-LSTM. We adopt the backend module as the decoder part of Tacotron2 \cite{tacotron2}. We utilize the same architecture of audio embedding module as lip reading experiment except for additional one convolution layer with kernel size of 5 before the Residual block. We adopt Griffin-Lim \cite{griffinlim} algorithm for audio waveform conversion. For the task loss $g(\cdot)$, L1 distance loss is applied.
More details of the network architecture can be found in supplementary material.
\begin{figure*}[t!]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=16.5cm]{./Figure/Fig3.pdf}}
\end{minipage}
\caption{Face video clips (source modality) and corresponding addressing vectors for recalling audio modality (target modality) from learned representations inside memory: (a) results from lip reading and (b) results from speech reconstruction from silent video.}
\label{fig:3}
\vspace{-0.3cm}
\end{figure*}
\vspace{-0.2cm}
\subsubsection{Results}
\vspace{-0.2cm}
We use three standard speech quality metrics for quantitative evaluation: STOI \cite{stoi}, ESTOI \cite{estoi}, and PESQ \cite{pesq}. Table \ref{table:4} shows the performance comparison on GRID dataset in speaker-dependent setting. We report the average test scores for 4 speakers with the same setting of the previous works \cite{akbari2018lip2audspec, ganbased, ephrat2017improved, l2w, yadav2020speech}. The table clearly indicates that our model outperforms previous methods, including state-of-the-art performance. These improvements are from recalling the audio representations in the early stage of the backend which enables it to refine the generated mel-spectrogram.
\begin{table}[]
\renewcommand{\arraystretch}{1.2}
\renewcommand{\tabcolsep}{4mm}
\centering
\resizebox{0.89\linewidth}{!}{
\begin{tabular}{cccc}
\hline \hline
Method & STOI & ESTOI & PESQ \\ \hline
Vid2Speech \cite{ephrat2017vid2speech} & 0.491 & 0.335 & 1.734 \\
Lip2AudSpec \cite{akbari2018lip2audspec} & 0.513 & 0.352 & 1.673 \\
Vougioukas \etal \cite{ganbased} & 0.564 & 0.361 & 1.684 \\
Ephrat \etal \cite{ephrat2017improved} & 0.659 & 0.376 & 1.825 \\
Lip2Wav \cite{l2w} & 0.731 & 0.535 & 1.772\\
Yadav \etal \cite{yadav2020speech} & 0.724 & 0.540 & 1.932\\\hline
\textbf{Proposed Method} & \textbf{0.738} & \textbf{0.579} & \textbf{1.984} \\
\hline \hline
\end{tabular}}
\vspace{0.1cm}
\caption{Performance of speech reconstruction comparison with visual modal inputs in a speaker-dependent setting on GRID.}
\vspace{-0.6cm}
\label{table:4}
\end{table}
Moreover, we ask 25 human participants to rate the Naturalness and Intelligibility. Naturalness is evaluating how natural the synthetic speech is compared to the actual human voice, and intelligibility is how clearly the words sound in the synthetic speech compared to the actual transcription. 6 samples of generated speech for each of 4 speakers of GRID are used. The human subjective evaluation results are reported at Table \ref{table:6}. Compared to the previous works \cite{ephrat2017vid2speech, l2w}, the proposed method achieves better scores on both Naturalness and Intelligibility. Moreover, with WaveNet \cite{wavenet} vocoder instead of Griffin-Lim, we can improve the scores as close to that of the ground truth. This indicates the reconstructed mel-spectrogram is of high-quality so that we can further improve the audio quality by using the state-of-the-art vocoder.
We also conduct an experiment on the speaker-independent setting, which is known to be a complex setting, of the GRID dataset to verify the effectiveness of the proposed method. As shown in Table \ref{table:5}, compared to \cite{ganbased, l2w}, the proposed framework achieves the highest performance. It can be inferred that even in a complex setting, the proposed framework can achieve meaningful outcomes by bringing the additional information through the associative bridge and memory. We visualize the examples of the generated mel-spectrogram in supplementary material.
\begin{table}[]
\centering
\renewcommand{\arraystretch}{1.2}
\renewcommand{\tabcolsep}{4.5mm}
\resizebox{0.95\linewidth}{!}{
\begin{tabular}{ccc}
\hline \hline
Method & Naturalness & Intelligibility \\ \hline
Vid2Speech \cite{ephrat2017vid2speech} & 1.31 $\pm$ 0.24 & 1.42 $\pm$ 0.23 \\
Lip2Wav \cite{l2w} & 2.83 $\pm$ 0.21 & 2.94 $\pm$ 0.19 \\ \hline
\textbf{Proposed Method} & \textbf{2.93} $\pm$ \textbf{0.21} & \textbf{3.56} $\pm$ \textbf{0.19} \\ \hline
\makecell{\textbf{Proposed Method}\\ (+WaveNet vocoder \cite{wavenet})} & 4.37 $\pm$ 0.16 & 4.27 $\pm$ 0.14 \\
Ground Truth & 4.62 $\pm$ 0.13 & 4.57 $\pm$ 0.14 \\
\hline \hline
\end{tabular}
}
\vspace{0.1cm}
\caption{Mean opinion scores for human evaluation on GRID.}
\vspace{-0.1cm}
\label{table:6}
\end{table}
\begin{table}[]
\renewcommand{\arraystretch}{1.2}
\renewcommand{\tabcolsep}{5mm}
\centering
\resizebox{0.9999\linewidth}{!}{
\begin{tabular}{cccc}
\hline \hline
Method & STOI & ESTOI & PESQ \\ \hline
Vougioukas \etal \cite{ganbased} & 0.445 & - & 1.240 \\
Lip2Wav \cite{l2w} & 0.565 & 0.279 & 1.279\\ \hline
\textbf{Proposed Method} &\textbf{ 0.600} & \textbf{0.315} & \textbf{1.332}\\ \hline \hline
\end{tabular}}
\vspace{0cm}
\caption{Performance of speech reconstruction comparison with visual modal inputs on the speaker-independent setting on GRID.}
\vspace{-0.3cm}
\label{table:5}
\end{table}
We conduct an ablation study on different memory slot size, which is shown in supplementary material. It shows the best scores of 0.738 STOI, 0.579 ESTOI, and 1.984 PESQ when $N$=150. Moreover, the performance of the proposed framework improves regardless of the number of memory slots, which verifies its effectiveness.
\subsection{Learned representation inside memory}
\label{sec:4.3}
\begin{figure}[t!]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=7.8cm]{./Figure/Fig4.pdf}}
\end{minipage}
\caption{Examples of similarity between memory addressing vectors of different video clips in LRW. Note source addressing vector is for bridging video and audio modal features in memory.}
\label{fig:4}
\vspace{-0.5cm}
\end{figure}
In this section, we visualize the addressing vectors of both lip reading and speech reconstruction model in speaker-independent setting. Fig.\ref{fig:3} (a) shows the video clips of LRW dataset with consecutive 5 frames and the corresponding addressing vectors of lip reading model. From the addressing vectors of different speakers speaking the same pronunciation, we observe the similar tendency of the addressing vectors. For example, when the face video is saying ``sta'' in words \textit{started} and \textit{start}, similar memory slots are highly addressed. The same tendency can be observed in the speech reconstruction model shown in Fig.\ref{fig:3} (b). This shows that source-key memory consistently finds the corresponding saved audio location in the target-value memory by using the talking face video clips as a query, which means the associative bridge is meaningfully constructed.
In addition, we compare addressing vectors of facial video clips with different pronunciations. Fig.\ref{fig:4} shows the consecutive video frames with its corresponding pronunciation, and the comparison results. We can observe that the source addressing vectors of saying similar pronunciation have high similarity, while differently pronouncing videos have low similarity. For example, video clips of pronunciation ``a\textipa{\textupsilon}'' of word \textit{about} and \textit{amount} have 0.906 cosine similarity. In contrast, the similarity between ``ri'' of word \textit{period} and ``a\textipa{\textupsilon}'' of word \textit{about} is low with 0.404.
\subsection{Comparison with methods finding a common latent space of multi-modality}
\label{sec:4.4}
We examine that the proposed framework can bypass the difficulty of finding a common representation of different modalities while bridging them. We compare the performance of word-level lip reading with the previous multi-modal learning methods that can exploit shared information of audio-visual modalities with uni-modal inference input by finding a common latent space. We build two multi-modal adaptation methods: cross-modal adaptation method \cite{Chung16sync} and knowledge distillation method \cite{hinton2015distilling}. The first is pretrained to synchronize the audio-visual modalities, and then trained for lip reading. The second method is additionally trained so that the features from lip reading model resemble the features from the automatic speech recognition model.
We show the word-level lip reading accuracies on LRW dataset in Table \ref{table:8}.
By utilizing multi-modality with visual modal inputs only, all of the methods show the performance improvements from the baseline, and the proposed framework achieves the best performance. The comparison shows the efficiency of the proposed framework, where it does not need to find a common latent space of two modalities by dealing with each modality in a modality-specific memory.
\begin{figure}[t!]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=8cm]{./Figure/Fig5.pdf}}
\end{minipage}
\caption{t-SNE \cite{van2008tsne} visualization of learned representation of (a) visual and audio modality, and (b) the recalled audio from visual modality and the actual audio modality.}
\label{fig:5}
\vspace{-0.2cm}
\end{figure}
\begin{table}
\centering
\renewcommand{\tabcolsep}{1.5mm}
\resizebox{0.9999\linewidth}{!}{
\begin{tabular}{ccccc}
\hline \hline
Method & Baseline & \makecell{Cross-modal\\ Adaptation \cite{Chung16sync}} & \makecell{Knowledge \\Distillation \cite{hinton2015distilling}} & \makecell{\textbf{Proposed}\\ \textbf{Method}} \\\hline \\[-0.5em]
ACC(\%) & 84.14 & 84.20 & 84.50 & \textbf{85.41} \\[0.5em] \hline \hline
\end{tabular}}
\vspace{0cm}
\caption{Lip reading word accuracy comparison with learning methods of finding a common representation of multi-modality.}
\vspace{-0.5cm}
\label{table:8}
\end{table}
Lastly, we visualize the representations of visual modality, audio modality, and recalled audio modality from visual modality, by mapping them into 2D space. Fig.\ref{fig:5} shows t-SNE \cite{van2008tsne} visualization of learned representations of visual and audio modalities, and the recalled audio from visual modality and the actual audio modality. Since we handle each modality with modality-specific embedding module and memory, the two modalities have separate representations in the latent space (Fig.\ref{fig:5} (a)). However, as Fig.\ref{fig:5} (b) shows, the recalled audio from the visual modality through the associative bridge shares a representation similar to the audio modal representation. Thus, we can utilize both audio and visual contexts while maintaining their own modal representations. This visualization demonstrates that we can effectively bridging the multi-modal representations without suffering from the cross-modal adaptation by dealing with each modality in modality-specific modules.
\section{Conclusion}
In this paper, we have introduced the multi-modality associative bridging framework that connects both audio and visual context through source-key memory and target-value memory. Thus, it can utilize both audio and visual information even if only one modality is available. We have verified the effectiveness of the proposed framework on two applications: lip reading and speech reconstruction from silent video, and achieved state-of-the-art performances. Furthermore, we have shown that the proposed framework can bridge the two modalities while maintaining separate latent space for each.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}\vspace{-2mm}
Human pose
recovery has garnered a lot of research interest in the vision community. It finds extensive use in a wide range of applications such as, augmented reality, virtual shopping, human-robot interaction, \emph{etc}\onedot. Recent advances in human pose
recovery is largely attributed to the availability of 3D pose supervision from large-scale datasets~\cite{ionescu2013human3,mehta2017monocular,varol2017learning}. Although, the models achieve superior performance on in-studio benchmarks, they usually suffer from
poor cross-dataset performance.
Training on synthetic and in-studio dataset
, which lack diversity in subject appearance, lighting, background variation, among more, can inherently induce a domain-bias \cite{pmlr-v37-long15, kundu2019_um_adapt} thereby restricting generalizability. This calls for the question: how to bridge performance gaps across diverse
data domains?
Some works use weaker forms of supervision such as 2D pose keypoints,
either directly \cite{kanazawa2018end,VNect_SIGGRAPH2017} from the manually annotated large-scale datasets \cite{andriluka20142d,johnsonclustered} or indirectly \cite{bogo2016keep,kolotouros2019learning} via an off-the-shelf 2D pose detector network. Further, multi-view \cite{rhodin2018learning,Iqbal_2020_CVPR} cues have also been used to learn a reliable 2D-to-3D mapping. However, acquiring such additional supervision often comes at a cost, such as laborious manual skeletal-joint annotation or cumbersome calibrated multi-camera setups. This severely limits the deployability of a model in a novel, in-the-wild target environment. Images in a target deployment scenario can vary from those used in
training setup in several ways, from simple changes in lighting or weather, to large domain shifts such as thermal or near-infrared (NIR) imagery. It is highly impractical to gather pose annotations
for every novel deployment scenario.
\begin{figure}[!tbhp
\begin{center}
\vspace{-5mm}
\includegraphics[width=1.00\linewidth]{figures_topo/NuerIPS_Overview_fig_iccv.pdf}
\vspace{-4mm}
\caption{ \small
The proposed self-supervised adaptation relies only on silhouette supervision (unaffected by domain-shift) to adapt the source trained pose regressor to novel target deployment scenarios. Topology extraction from raw foreground silhouettes is one of the key design aspects of our framework.
}
\vspace{-6mm}
\label{fig:concept}
\end{center}
\end{figure}
Motivated by these observations, we aim to realize a self-adaptive monocular 3D human pose recovery framework that does not rely on any strong ground-truth (2D/3D pose) or auxiliary target supervision (multi-view, or depth).
The proposed adaptation framework (Fig. \ref{fig:concept}{\color{red}}) aims to transfer the task knowledge from a labeled source domain
to an unlabeled target (Fig. \ref{fig:concept}{\color{red}B}), while relying only on foreground (FG) silhouettes.
We observe that the traditional FG segmentation techniques are mostly unaffected by input domain shifts (Fig.~\ref{fig:concept}{\color{red}C}) and thus the FG masks are usually robust and easy to obtain in diverse target scenarios. Consider a deployment scenario where a static camera is capturing a moving object. Here, one can rely on classical vision based background subtraction techniques to obtain a FG silhouette while being considerably robust to a wide range of domain shifts (including gray-scale, thermal, NIR imaging, etc). Certain deep learning based class-agnostic FG segmentation methods~\cite{yang2019anchor,lu2019see} (motion
and salient-object segmentation methods) also exhibit reasonable robustness to domain-shifts. Hence, an adaptation
framework that can effectively use silhouette supervision would open up the scope for
adaptation to diverse deployment scenarios.
Prior human mesh recovery works \cite{kolotouros2019learning,Iqbal_2020_CVPR,kanazawa2018end,kanazawa2019learning} have demonstrated considerable performance gains by training with
articulation-centric annotations such as 3D and 2D pose supervision. At the same time, silhouettes have found a relegated use, mostly towards enforcing auxiliary shape-centric objectives \cite{8100067,pavlakos2018learning
. Although these weaker objectives employ silhouettes, they do not offer
strong supervisory signal
in isolation, hence are not self-sufficient. In this paper, we aim to realize a stronger, self-sufficient and articulation-centric supervision from raw silhouettes. The key challenge is that the articulation information is usually lost in direct pixel-level losses, as raw silhouettes mostly provide 2D shape information. Our key insight is to extract a skeleton-like 2D topology information from silhouettes (see Fig.~\ref{fig:concept}{\color{red}D}), which can provide a stronger articulation-centric supervision.
We further propose a topological alignment objective that is inspired from 3D Chamfer distance~\cite{fan2017point} and can directly work on binary images.
We leverage this objective to drive our target adaptation training using just silhouettes, and demonstrate the applicability of our framework across various domains.
We make the following main contributions:
\begin{itemize}[leftmargin=*]
\vspace{-1.5mm}
\item We develop a series of convolution-friendly spatial transformations in order to disentangle a topological-skeleton representation from the raw silhouette.
\vspace{-1.5mm}
\item We devise a Chamfer-inspired spatial topology alignment loss via distance field computation, while effectively avoiding any gradient hindering spatial-to-pointset mapping.
\vspace{-1.5mm}
\item
Our self-supervised adaptation framework achieves state-of-the-art performance
on multiple challenging
domains, such as low-resolution domain (LR), universal adversary perturbed domain (UAP), in-studio Human3.6M~\cite{ionescu2013human3}, and in-the-wild 3DPW~\cite{von2018recovering}.
\end{itemize}
\vspace{-2mm}
\section{Related work}\vspace{-2mm}
\textbf{Human mesh recovery.} Supervised human mesh recovery on monocular RGB images has been well explored in recent years, works such as \cite{kolotouros2019learning,kanazawa2018end,kanazawa2019learning,pavlakos2018learning} achieve impressive performance on standard
benchmarks. These works heavily rely on large-scale annotated in-studio datasets~\cite{ionescu2013human3,mehta2017monocular} for the bulk of their training, however they report and agree on the lack of generalized performance on unseen in-the-wild data.
Most works~\cite{kanazawa2018end,kolotouros2019learning,texturepose,pavlakos2018learning,kanazawa2019learning, LassnerClosing,kundu_human_mesh,bogo2016keep} try to address the issue by utilizing manually annotated 2D supervision from in-the-wild datasets~\cite{johnson2011learning,andriluka20142d,lin2014microsoft}, and some additionally fine-tune their models by enforcing temporal~\cite{kanazawa2019learning,huang2017towards} or multiview~\cite{liang2019shape,texturepose,huang2017towards} constraints.
Kolotouros~\emph{et al}\onedot~\cite{kolotouros2019learning} has also explored integrating pre-trained off-the-shelf 2D pose detectors into the training loop, aimed towards improving in-the-wild performance.
Despite the diversity in approaches,
prior works mainly
focus on applicability to a single target domain i.e. the ideal in-the-wild RGB data. Due to such a viewpoint, even the
weakly-supervised approaches~\cite{NEURIPS2019_d4a93297,tan2018indirect,pavlakos2018learning} render themselves irrelevant towards extending to new unlabeled target domains, as their methods
implicitly assume
access to significant amount of annotated samples from the final target domain.
Xu~\emph{et al}\onedot~\cite{xu20203d} proposed to overcome the performance-gap due to low-resolution (LR) imagery, but required labeled targets with simultaneous access to high-resolution source data.
Later,
Zhang~\emph{et al}\onedot~\cite{zhang2020inference} proposed to address the domain induced
performance degradation via self-supervised inference stage optimization. But their requirement of 2D pose keypoints as inpu
, all the while only addressing 3D pose distribution shifts, severely restricts their applicability.
Recognising this deficiency in prior works, our aim is to provide a self-supervised framework which can facilitate the overcoming of
performance gaps across a wide
array of domains.
\begin{wraptable}[12]{r}{7cm}
\vspace{-4.5mm}
\footnotesize
\caption{
Characteristic comparison of our method against prior
arts, separated by
access to direct (paired) supervision for adaptation to new domains.
MV: multi-view, Sil.: silhouette
}
\vspace{-3mm}
\centering
\setlength\tabcolsep{4.5pt}
\resizebox{0.5\textwidth}{!}{
\begin{tabular}{l|ccc|c}
\hline
\multirow{2}{*}{\makecell{\vspace{-1.5mm} \\ Model-based \\ Methods}} & \multicolumn{3}{c|}{\makecell{Paired sup.}} &
\multirow{2}{*}{\makecell{\vspace{-1.5mm}\\ Self-sup.\\ Adaptation }}\\%Annotation \\ or MV$^*$ setup \\for adaptation }} \\
\cline{2-4}
& \makecell{ \vspace{-2mm}\\2D/3D \\pose} & \makecell{ \vspace{-2mm}\\MV \\ cams} & \makecell{ \vspace{-2mm}\\Sil.} &
\\ \hline
\cite{bogo2016keep,kanazawa2018end,kolotouros2019learning} &\cmark &\xmark &\xmark &\xmark
\\
\cite{liang2019shape,texturepose,huang2017towards} &\cmark &\cmark &\xmark &\xmark \\
\cite{LassnerClosing,guan2009estimating,pavlakos2018learning,tan2018indirect} &\cmark &\xmark &\cmark &\xmark \\
Ours &\xmark &\xmark &\cmark &\cmark \\
\hline
\end{tabular}}
\vspace{-2mm}
\label{tab:char}
\end{wraptable}
\textbf{Use of silhouettes.}
Silhouettes have found extensive use in works which aim to obtain 3D human body shape from images. Sigal~\emph{et al}\onedot~\cite{sigal2008combined} is one of the earliest works attempting to estimate the high quality 3D human shape, via fitting a parametric human-body model (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, SCAPE~\cite{anguelov2005scape}) to ground truth silhouettes. Subsequently, Guan~\emph{et al}\onedot~\cite{guan2009estimating} used silhouettes, shading and edges as cues towards the body-shape fitting process, but required a user-given 2D pose initialization.
Lassner~\emph{et al}\onedot~\cite{LassnerClosing} proposed a silhouette based fitting approach for SMPL~\cite{loper2015smpl} human-body model, by using an off-the-shelf
segmentation network. These works primarily aim for
shape fitting based optimization and attempt to fit a 3D parametric mesh to a set of 2D evidence. However, solving these iterative optimization formulations for each and every instance can be prohibitively slow at test-time, and are also prone to getting stuck in local minimas.
In contrast to these shape-centric works, we primarily aim to recover accurate posture of the person and consider shape estimation as a useful by-product of our formulation.
A few works~\cite{tan2018indirect,pavlakos2018learning,kundu2020self} formulate encoder-decoder based architectures with silhouette and 2D keypoint reconstruction
as primary training objectives for their final regressor network. Silhouettes have also been used~\cite{tung2017self,kundu_human_mesh,NEURIPS2019_d4a93297,Sminchisescu2002HumanPE,Biggs2018CreaturesGA} as weak forms of auxiliary supervision usually in tandem with optical-flow, motion or other self-supervised objectives, some of these even relying on synthetic data for pre-training their networks. Although these works remain relevant in terms of performance on ideal in-the-wild RGB data, they fail to offer a fast, self-supervised
end-to-end
framework capable of adapting to diverse unseen target domains (see Table~\ref{tab:char}).
\vspace{-1mm}
\section{Approach}\vspace{-1.5mm}
We aim to accurately recover the 3D human pose for unlabeled target domain samples, while relying only on silhouette supervision to adapt a source-trained model-based regressor.
\vspace{-1.5mm}
\subsection{Background and motivation}\label{sec:3_1}
\vspace{-2mm}
Traditionally, weakly-supervised human mesh estimation has been framed as a bundle-adjustment problem \cite{bogo2016keep}. Given the SMPL body model~\cite{loper2015smpl}, an iterative fitting algorithm aims to adjust the pose, shape and camera parameters to minimize the reprojection loss against the given 2D observations.
\noindent
\textbf{3.1.1\hspace{1mm} Model-based pose regression network}.\label{subsec:3_1_1}\hspace{2mm}
Given an input image $I$, a CNN regressor outputs the SMPL~\cite{loper2015smpl} parameters (pose $\theta$, shape $\beta$) alongside the camera parameters, $R_c=\{R, s, t\}$ (as shown in Fig.~\ref{fig:main1}{\color{red}A}).
Here, $R_c$ includes global orientation $R\in\mathbb{R}^{3\times 3}$, scale $s\in \mathbb{R}$, and
translation $t\in \mathbb{R}^2$.
Following this, the SMPL module generates the ordered mesh vertices in the canonical space, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\hat{V}\in\mathbb{R}^{6890\times 3} = \mathcal{M}(\theta, \beta)$.
The mesh vertices $\hat{V}$ are mapped to 3D pose $\hat{Y}$
(\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot \textit{vertex-to-3DPose})
with $k$ joints using a pre-defined linear regression: $\hat{Y}\!\in\mathbb{R}^{k\times3}=W_p\hat{V}$.
We can then project 3D pose $\hat{Y}$ to 2D pose $\hat{Z} \in \mathbb{R}^{k \times 2}$ using weak-perspective camera projection:
$\hat{Z}=\pi(R\hat{Y}, s, t)$.
\begin{figure*
\begin{center}
\vspace{2mm}
\includegraphics[width=1.0\linewidth]{figures_topo/Approach_fig_NIPS.pdf}
\caption{\small
Framework overview highlighting the modules and the series of transformations to realize the topology alignment objective.
\textbf{A.} Image to SMPL regression (Sec. \ref{subsec:3_1_1}{\color{red}.1}).
\textbf{B.} Extracting topological-skeleton via inwards distance-map of silhouette mask (Sec. \ref{sec:3_2}).
\textbf{C.} and \textbf{D.} Obtaining outwards distance-map of topological-skeleton to support spatial-chamfer (Sec. \ref{sec:3_3}).
\textbf{E.} Chamfer-inspired topology-alignment objective (Sec. \ref{sec:3_3}).
}
\label{fig:main1}
\vspace{-9mm}
\end{center}
\end{figure*}
\textbf{3.1.2\hspace{1mm} 2D pose supervision.}\hspace{2mm}
In the presence of 2D pose ground-truth (GT), prior approaches~\cite{kanazawa2018end,pavlakos2018learning,bogo2016keep,LassnerClosing} aim to obtain the same representation for the prediction branch in order to employ a direct loss.
Here, the \textit{vertex-to-3DPose} mapping is already well specified in the SMPL module via fixed regression parameters $W_p$.
Such approaches heavily benefit from this pre-specified body-joint structure. As a result, the mesh fitting problem boils down to minimizing the distance between two ordered point-sets (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot a one-to-one mapping between the 2D joint locations) which is easily achieved via a direct $L1/L2$ loss. Note that in such cases, post SMPL-regression, all the intermediate mappings operate on pre-specified ordered point-set representations, bypassing the need for a self-sufficient articulation-centric gradient via a rendering pipeline (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot a mapping from point-set to spatial-maps).
\textbf{3.1.3\hspace{1mm} Silhouette supervision.}\hspace{2mm}
Unlike 2D pose supervision, learning from raw silhouette GT comes with a multitude of challenges. Though one can easily obtain the spatial silhouette from the predicted SMPL mesh via differentiable rendering~\cite{ravi2020pytorch3d},
devising an effective
self-sufficient loss stands as the major bottleneck.
A
trivial loss to align the two binary silhouette masks would be a pixel-level $L1$ or cross-entropy loss, as employed in general segmentation tasks. However, such a loss fails
to capture any gradient information along the spatial direction (see Fig.~\ref{fig:main2}{\color{red}A}).
Certain prior-arts \cite{huang2017towards,LassnerClosing,Gavrila19963DMT} employ a chamfer based distance between two point-sets, termed as \textit{fitting-loss}. Here, the two point-sets are basically the FG pixel-locations obtained from the GT and predicted silhouettes. However, this is meant for a better shape fitting and is almost always used in tandem with direct 2D or 3D pose supervision. When employed in isolation, it often lead to sub-optimal degenerate solutions, specifically in the absence of any pose-related articulation
component.
Based on the above observations, we propose
a new representation space, termed as \textit{topological-skeleton}, which can facilitate seamless morphological silhouette alignment
via effective spatial distance computations. To this end, we introduce the following:
\textbf{a) Distance-map, $D$.}
For a given silhouette-mask $S$, its \textit{distance-map} is defined as a spatial map $D(u)$, whose intensity at each pixel-location $u\!\in\!\mathbb{U}$ represents its distance from the closest
mask-outline
pixel of $S$ (see Fig.~\ref{fig:main1}{\color{red}B}). We utilize two different distance-maps: a) inwards distance map, $D_\textit{in}$, and b) outwards distance map, $D_\textit{out}$. The \textit{inwards distance-map} holds zero intensity for BG pixels (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot pixels outside the active region of $S$), and vice-versa for the \textit{outwards distance-map}.
Mathematically,
\noindent
\begin{equation} \label{eq_1}
D_\textit{in}(u)=
\begin{dcases*}
\min_{\{u^\prime:S(u^\prime)=0\}} \vert u-u^\prime \vert, & \text{if} \hspace{0.3mm} S(u)=1, \\
0, & \text{otherwise}.
\end{dcases*} \quad
D_\textit{out}(u)=
\begin{dcases*}
\min_{\{u^\prime:S(u^\prime)=1\}} \vert u-u^\prime \vert, & \text{if} \hspace{0.3mm} S(u)=0, \\
0, & \text{otherwise}.
\end{dcases*}
\end{equation}
\textbf{b) Topological-skeleton, $T$.}
\textit{Topological-skeleton} is a thin-lined pattern that represents the geometric and structural core of a silhouette mask. This can also be interpreted as the ridges-lines~\cite{blum1967transformation,855792,Chang2007ExtractingSF} of $D_\textit{in}$ extracted for a given silhouette (see Fig.~\ref{fig:main1}{\color{red}B}). For example, people with arms wide open would have the same structural topology, irrespective of their body-build type. In the absence of direct pose-centric supervision (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot 2D or 3D pose annotations), we propose to treat \textit{topological-skeleton} $T$ as a form of
pose-centric influence that can be extracted from the raw silhouette. Further, we aim to realize $T$ in the form of a binary spatial map, unlike the point-set form of 2D pose keypoints. Here, $T(u)\!=\!1$ indicates that pixel $u\!\in\!\mathbb{U}$ is present in the topological ridge.
We rely on distance-map to;
a) construct the topological-skeleton via inwards distance-map of the silhouette (see Fig.~\ref{fig:main1}{\color{red}B}) as discussed in Sec. \ref{subsec:3_2_1}{\color{red}.1}, and b) devise a Chamfer-inspired training objective via outwards distance-map of the topological-skeleton (see Fig.~\ref{fig:main1}{\color{red}C}) as discussed in Sec. \ref{sec:3_3}. Note that, the outwards distance-map of the topological-skeleton is defined by replacing $S$ with $T$ in Eq.\ref{eq_1}.
\vspace{-1mm}
\subsection{Silhouette topology extraction}\label{sec:3_2}
\vspace{-2mm}
Here, the primary objective is to formalize a series of transformations in order to extract the respective topological-skeletons for the predicted and GT silhouette masks, as shown in Fig.~\ref{fig:main1}.
In the prediction pathway, the silhouette is obtained via a fast differentiable renderer~\cite{ravi2020pytorch3d} which operates on the mesh vertices $\hat{V}$. However, a non-binary silhouette is not suitable for distance-map and topology definitions as discussed in Sec.~\ref{sec:3_1}. Thus, we employ a binary activation function~\cite{hubara2016binarized} as shown in Fig.~\ref{fig:main2}{\color{red}B}. The final outcome is represented as $\hat{S}\!=\!\texttt{M2S}(\hat{V})$. While formulating the later transformation modules, one must pay special attention to design constraints discussed below.
\textbf{Design constraints.} We intend to update the CNN by back-propagating a loss which would be computed at the topological representation, $\hat{T}$ (see Fig.~\ref{fig:main1}). An ineffective
regressor output $\hat{\Theta}_\textit{reg}$ to $\hat{T}$ mapping results in vanishing gradient issues.
Additionally, parallelizable operations remain critical and desirable for faster training on modern GPUs. Hence, we aim to formulate each transformation via convolution-friendly
spatial operations,
thereby avoiding any spatial-map to point-set mapping
while implementing the topology-extraction and alignment objectives.
\textbf{3.2.1\hspace{2mm} Silhouette to distance-map, \texttt{S2D}.}\label{subsec:3_2_1} \hspace{2mm}
Restricting to spatial convolution-friendly computation, we devise a
recursive erosion (or thinning) operation in order to gradually shave off the boundary regions of the silhouette mask. To this end, we compute the number of active neighbours of a pixel by convolving the binary silhouette input with an $n\!\times\!n$ neighbourhood kernel $N$, as shown in Fig.~\ref{fig:main2}{\color{red}D}. The boundary pixels are expected to have less than maximum neighbours (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\!<\!8$ neighbours for $n\!=\!3$), these non-maximal pixels are then purged via the
$\text{ReLU}$ activation to output a thinned binary silhouette map, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $S_{i+1}\!=\!\text{ReLU}(N \circledast S_i-m)$. Here, $\circledast$ denotes a convolution operator and $m=n^{2}\!-\!2$. Following this, all the series of thinned
binary maps are added to realize a distance map equivalent,
\noindent
\begin{equation} \label{eq_2}
\begin{array}{ll}
D_\textit{in}=\texttt{S2D}(S)=\sum_{i=0}^{l} S_{i} \hspace{4mm}
\text{where:} \hspace{2mm} S_{0}=S \hspace{4mm} \text{and}\hspace{4mm} S_{i+1}=\text{ReLU}(N \circledast S_i-m)
\end{array}
\end{equation}
\noindent
Interestingly, the \textit{outwards distance-map} of $S$ can also be computed as $D_\textit{out}\!=\!\texttt{S2D}(1\!-\!S)$.
\begin{figure*
\begin{center}
\vspace{-4mm}
\includegraphics[width=1.0\linewidth]{figures_topo/Training_procedure_fig_NIPS.pdf}
\vspace{-5mm}
\caption{\small
\textbf{A.} A representative spatial-gradient landscape
for different kind of silhouette losses (left: Z-axis indicates gradient magnitude, right: intensity indicates gradient magnitude) such as, i) Pixel-level $L1$ or cross-entropy loss, ii) Chamfer inspired silhouette loss, iii) Chamfer-inspired topology-alignment loss.
\textbf{B.} Details of the \texttt{M2S} module (see Sec. \ref{sec:3_2}). \textbf{C.} Details of the \texttt{S2D} module (see Sec. \ref{subsec:3_2_1}{\color{red}.1}) where \textbf{D.} shows details of a single convolutional thinning. \textbf{E.} Details of the \texttt{D2T} module (see Sec. \ref{subsec:3_2_2}{\color{red}.2}). Pixel intensity range shown on top-right.
}
\label{fig:main2}
\vspace{-6mm}
\end{center}
\end{figure*}
\textbf{3.2.2\hspace{2mm} Distance-map to topological-skeleton, \texttt{D2T}.}\label{subsec:3_2_2} \hspace{2mm}
We aim to realize the thin-lined topological-skeleton as the ridge-line of the distance-map $D_\textit{in}$. In other words, the ridge-line pixels are the local maxima of $D_\textit{in}$. In view of convolution-friendly
backpropable implementation, a pixel would be selected as a ridge-pixel if its distance-map intensity matches with the local MaxPooled map of the same (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot where $D_\textit{in}\!-\!\text{MaxPool}(D_\textit{in})\!=\!0$). Following this, the topological-skeleton is realized as:
\noindent
\begin{equation} \label{eq_3}
T=\texttt{D2T}(D_\textit{in},S)=\text{ReLU}(D_\textit{in}-Maxpool(D_\textit{in})+1)\odot S
\end{equation}
Where $\odot$ represents element-wise dot product. Note that, intensities of $D_\textit{in}$ are discrete values in the range $[0,l]$ \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, ${0,1,2,..,l}$. Thus, the $+1$ term followed by $\text{ReLU}$ selects the ridge-pixels along with the false-positives from the BG region, which are later pruned by masking using
$S$ (see Fig.~\ref{fig:main2}{\color{red}E}). The resultant $T$ is also a binary map by formulation, i.e $T(u)\!\in\mathbb\!\{0,1\}$.
In summary, the prediction pathway involves a series of transformations, $\texttt{M2S}\!\!\shortrightarrow \!\!\texttt{S2D}\!\!\shortrightarrow \!\!\texttt{D2T}$ to realize $\hat{V}\!\shortrightarrow \!\hat{S}\!\shortrightarrow \!\hat{D}_{in}\!\shortrightarrow \!\hat{T}$ (see Fig.~\ref{fig:main1}). Similarly, the GT pathways involves
$\texttt{S2D}\!\!\shortrightarrow \!\!\texttt{D2T}$ to realize $S\!\shortrightarrow \!D_{in}\!\shortrightarrow \!T$.
\vspace{-1mm}
\subsection{Topological loss}\label{sec:3_3}
\vspace{-2mm}
Here, we
discuss the formulation of
our topological loss, which aims to quantify the misalignment between $\hat{T}$ and $T$.
We draw inspiration from Chamfer distance~\cite{fan2017point} which quantifies the misalignment between the predicted and GT point-sets.
Let $T_p$ and $\hat{T}_p$ be the point-set representation of $T$ and $\hat{T}$ respectively, implying $T_p\!=\!\{u:T(u)\!=\!1\}$. Note that, unlike 2D pose keypoints,
$|T_p|\!\neq\!|\hat{T}_p|$ and there exist no point-to-point
correspondence. One-way-Chamfer measures the sum
of the shortest distance of each point in $T_p$ against all points in $\hat{T}_p$. Thus, the two-way Chamfer is expressed as,
\vspace{-2mm}
\begin{equation} \label{eq_4}
\mathcal{L}_\textit{T}^{(p)}=\mathcal{L}^{(p)}_{T_p\shortrightarrow \hat{T}_p} + \mathcal{L}^{(p)}_{\hat{T}_p\shortrightarrow {T}_p} =
\sum_{u\in T_p} \min_{u^\prime\in\hat{T}_p}\Vert u-u^\prime \Vert_{2}^{2} +
\sum_{u\in \hat{T}_p} \min_{u^\prime\in {T}_p}\Vert u-u^\prime \Vert_{2}^{2}
\end{equation}
\vspace{-2mm}
However, in the view of convolution-friendly implementation we aim to avoid any spatial-map
to point-set mapping.
Implying, we can not obtain $T_p$ or $\hat{T}_p$ as required to compute the above loss term. To this end, we propose to make use of the distance-map representation to formalize a similar loss term that would directly operate on spatial-maps. We recognize that the outwards distance-map of $T$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $D_\textit{out}^T(u)$, stores the shortest distance of each BG pixel $u$ against the active pixels in $T$. Thus, an
equivalent form of $\mathcal{L}^{(p)}_{\hat{T}_p\shortrightarrow T_p}$ can be computed as; $\Vert (D_\textit{out}^T\odot \hat{T})\Vert_{1,1}$ where $\Vert . \Vert_{1,1}$ computes the $L1$-norm of the vectorized matrix. Thus, the final Chamfer-inspired topology-alignment loss is realized as,
\noindent
\begin{equation} \label{eq_5}
\mathcal{L}_{T}^{(sp)}=\Vert\hat{D}_\textit{out}^{T}\odot {T}\Vert_{1,1} + \Vert D_\textit{out}^T\odot \hat{T}\Vert_{1,1} \hspace{4mm} \text{where:}\hspace{2mm} \hat{D}_\textit{out}^{T} = \texttt{S2D}(1-\hat{T}) \hspace{2mm} \text{and} \hspace{2mm} D_\textit{out}^{T} = \texttt{S2D}(1-T)
\end{equation}
Note that, $\texttt{S2D}(1\!-\!T)$ yields the {outwards distance-map} (see Fig.~\ref{fig:main1}) which is inline with Eq.~\ref{eq_1}. The above loss formulation effectively addresses all the design and gradient related difficulties thereby facilitating a seamless morphological silhouette alignment.
\vspace{-1mm}
\subsection{Adaptation training}
\vspace{-2mm}
We posit the training as a self-supervised domain adaptation problem.
To this end, we first train the CNN regressor by supervising on labeled source samples, $(I,\Theta_\textit{reg})\!\in\!\mathcal{D}_\textit{src}$ which is typically a graphics-based synthetic domain or an in-studio dataset. Upon deployment to an unknown target environment, we gather the image and silhouette pairs, $(I,S)\!\in\!\mathcal{D}_\textit{tgt}$ where $S$ is
obtained via classical FG-mask estimation techniques. Alongside the topology alignment objective (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot minimizing $\mathcal{L}_\textit{T}^{(sp)}$), we adopt the following to regularize the adaptation process.
\textbf{a) Enforcing natural priors}. Most
works
rely on adversarial prior-enforcing losses to
restrict the CNN predictions within the realm of natural pose and shape distributions. To simplify adaptation process,
we train an
adversarial auto-encoder~\cite{makhzani2015adversarial,kundu2020kinematic,kundu2020unsupervised,SMPL-X:2019} to learn a latent pose space $\phi\!\in\![-1,1]^{32}$ whose decoder, $\Psi$ generates plausible 3D pose vectors
$\theta\!\in\!\mathbb{R}^{3k}\!=\!\Psi(\phi)$. Thus, just integrating this frozen decoder into our framework constrains the pose to vary within the plausibility limits. We denote $\hat{\Theta}_\textit{reg}\!=\!\{\beta, \phi, R_c\}$ and regularize
$\beta$
to remain close to the mean shape, inline with~\cite{kolotouros2019learning}.
\begin{wrapfigure}[10]{r}{6cm}
\begin{minipage}{0.5\textwidth}
\vspace{-5mm}
\begin{algorithm}[H
{\small
$\Theta_\textit{CNN}$: Source training initialized CNN weights \\
$\Theta_\textit{reg}^{(\textit{opt})}$:
SMPL parameters initialized from $\hat{\Theta}_\textit{reg}$ \\
\For {$\textit{iter} < \textit{MaxIter}$}{
\If{$\textit{iter}\pmod{K}\neq 0$}{
Update $\Theta_\textit{CNN}$ by optimizing $\mathcal{L}_\textit{S}^{(sp)}$, $\mathcal{L}_\textit{T}^{(sp)}$.
}
\Else{
\For{$\textit{iter}_\textit{opt} < \textit{MaxIter}_\textit{opt}$}{
Fit $\Theta_\textit{reg}^\textit{(opt)}$ to minimize $\mathcal{L}_\textit{S}^{(sp)}$, $\mathcal{L}_\textit{T}^{(sp)}$
}
Update $\Theta_\textit{CNN}$ by optimizing $\mathcal{L}_{\Theta}$.
}}
}
\vspace{1mm}
\caption{\small Proposed adaptation procedure. $\;\;$}\label{algo:1}
\end{algorithm}
\end{minipage}
\end{wrapfigure}
\textbf{b) Optimization in the loop.} Motivated by the benefits of combining regression and optimization based training routines~\cite{kolotouros2019learning}, we propose a target adaptation procedure as shown in Algo.~\ref{algo:1}. We first initialize $\Theta_\textit{reg}^\textit{(opt)}$ by inferring $\hat{\Theta}_\textit{reg}$ from the current state of the source-trained CNN. These regression-initialized $\Theta_\textit{reg}^\textit{(opt)}$ parameters undergo an iterative fitting procedure which aims
to minimize $\mathcal{L}_\textit{T}^{(sp)}$
and $\mathcal{L}_\textit{S}^{(sp)}$ without updating the CNN parameters. Here, $\mathcal{L}_\textit{S}^{(sp)}$ represents the spatial chamfer-based loss directly applied on raw-silhouettes, obtained by replacing $T$ and $\hat{T}$ with $S$ and $\hat{S}$ in Eq.~\ref{eq_5}. This
loss further improves pose and shape fitting. Subsequently, the fitted $\Theta_\textit{reg}^\textit{(opt)}$ is used to form a supervised loss, $\mathcal{L}_\Theta\! = \!\Vert \hat{\Theta}_\textit{reg}\! - \!\Theta_\textit{reg}^\textit{(opt)} \Vert$ to update the CNN weights, $\Theta_\textit{CNN}$. Although, we primarily rely on the end-to-end training of the CNN regressor by directly minimizing the proposed
alignment objectives, alternating between
regression and fitting based routines allows the model to self-improve faster.
\begin{figure*}[!t
\begin{center}
\vspace{-6mm}
\includegraphics[width=0.99\linewidth]{figures_topo/Qualitative_results_pose_A_NeurIPS_mod.pdf}
\caption{\small
\textbf{A}, \textbf{B}, and \textbf{C.} Qualitative results of \textit{Ours(S$\rightarrow$R)} on the respective datasets. Failure cases are highlighted in magenta.
\textbf{D.} Progression of our articulation-centric adaptation across various domains (across columns).
}
\vspace{-6mm}
\label{fig:qualitative}
\end{center}
\end{figure*}
\begin{table}[t!]
\parbox{.42\linewidth}{
\footnotesize
\caption{
Comparison against prior-works on Human3.6M (P-2). Table is divided based on access to supervision. We beat prior arts at comparable supervision levels
\vspace{1.0mm}
}
\centering
\setlength\tabcolsep{3pt}
\resizebox{0.44\textwidth}{!}{
\begin{tabular}{llc}
\midrul
Sup. & Method & PA-MPJPE($\downarrow$) \\
\midrule\midrule
& Pavlakos~\emph{et al}\onedot~\cite{pavlakos2018learning} & 75.9 \\
Full & HMR~\cite{kanazawa2018end} & 56.8 \\
& SPIN~\cite{kolotouros2019learning} & 41.1 \\
\midrule
& HMR (unpaired)~\cite{kanazawa2018end} & 66.5 \\
Weak & SPIN (unpaired)~\cite{kolotouros2019learning} & 62.0 \\
& \textbf{\textit{Ours(S$\rightarrow$R,weak)}} & \textbf{58.1}
\\
\midrule
\multirow{2}{*}{\centering Unsup.} & Kundu~\emph{et al}\onedot(unsup)~\cite{kundu_human_mesh} & 90.5 \\
& \textbf{\textit{Ours(S$\rightarrow$R)}} & \textbf{81.3}
\\
\hline
\end{tabular}}
\vspace{-6mm}
\label{tab:h36-p2}
}
\hfill
\parbox{.54\linewidth}{
\footnotesize
\caption{ Evaluation on the
3DPW dataset, none of the works train on 3DPW. For fair comparison, we separate works based on access to target supervision.
\vspace{1mm}}
\centering
\setlength\tabcolsep{2pt}
\resizebox{0.53\textwidth}{!}{
\begin{tabular}{llcc}
\midrul
Sup. & Method & MPJPE($\downarrow$) & PA-MPJPE($\downarrow$)\\
\midrule\midrule
& HMR~\cite{kanazawa2018end} & 128.
& 81.3 \\
Full & Kanazawa \emph{et al}\onedot~\cite{kanazawa2019learning} & 116.5 & 72.6 \\
& SPIN~\cite{kolotouros2019learning} & 98.
& 59.2 \\
\midrule
\multirow{4}{*}{\centering Weak} & Martinez~\emph{et al}\onedot~\cite{martinez2017simple} & - & 157.0 \\
& SMPLify~\cite{bogo2016keep} & 199.2 & 106.1 \\
& Doersch~\emph{et al}\onedot(RGB+2D)~\cite{NEURIPS2019_d4a93297} & - & 82.4 \\
& \textbf{\textit{Ours(S$\rightarrow$R, weak)}} & \textbf{126.3}
& \textbf{79.1}
\\
\midrule
\multirow{4}{*}{\centering Unsup.} & Doersch~\emph{et al}\onedot(DANN)~\cite{NEURIPS2019_d4a93297} & - & 103.0 \\
& Kundu~\emph{et al}\onedot(unsup)~\cite{kundu_human_mesh} & 187.1 & 102.7 \\
& Doersch~\emph{et al}\onedot(Flow)~\cite{NEURIPS2019_d4a93297} & - & 100.1 \\
& \textbf{\textit{Ours(S$\rightarrow$R)}} & \textbf{159.0}
& \textbf{95.1}
\\
\hline
\end{tabular}}
\vspace{-6mm}
\label{tab:3dpwresults}
}
\end{table}
\section{Experiments}\label{sec:4}\vspace{-2mm}
We perform a thorough empirical analysis demonstrating our superiority for cross domain adaptation.
\noindent
\textbf{Implementation details.}
We use an ImageNet~\cite{russakovsky2015imagenet} pre-trained \textit{ResNet-50}~\cite{he2016identity} network as our
CNN backbone for the regressor. The final layer features are average pooled and subsequently passed through a series of fully-connected layers to regress the latent pose encoding $\phi$, shape and camera parameters. We use the Adam optimizer \cite{kingma2014adam} with a learning rate $1\mathrm{e}{-6}$ and batch size of 16, while setting
$\textit{MaxIter}_\textit{opt}$ to 10 and $K$=4. We use separate optimizers for $\mathcal{L}_\textit{T}^{(sp)}$ and $\mathcal{L}_\textit{S}^{(sp)}$.
A single
iteration of the iterative fitting procedure takes nearly 12ms on a Titan-RTX GPU
\noindent
\textbf{Datasets.}
Moshed~\cite{loper2014mosh} CMU-MoCap~\cite{cmumocap} and H3.6M training-set~\cite{ionescu2013human3} form our unpaired 3D pose data, which is used to train the 3D pose-prior.
Our adaptation settings use the following datasets.
\vspace{-1mm}
\textbf{a) Synthetic (S):} We use SURREAL~\cite{varol2017learning}
to train our synthetic source model.
It is a large-scale dataset with synthetically generated images of humans, rendered from 3D pose sequences of~\cite{cmumocap}.
\vspace{-1mm}
\textbf{b) Real (R):} We use a mixture of Human3.6M~\cite{ionescu2013human3} and
MPII~\cite{andriluka20142d} as our Real-domain dataset. This domain is used as both source and target in different adaptation settings.
\vspace{-1mm}
\textbf{c) {UAP-H3M} (UAP):} Universal Adversarial Perturbation (UAP) \cite{moosavi2017universal} is an instance-agnostic
perturbation
aimed at attacking a deployed model, thereby inflicting a drop in the task performance. Typically, the adversarial noise is added with the clean source domain images to construct the adversarially perturbed target domain samples. We craft a single UAP perturbation ($L_\infty$ and $\epsilon\!=\!8/255$) using an equivalent 3D pose estimation network while following the procedure from~\cite{gduap-mopuri-2018}. We aim to evaluate the effectiveness of the proposed self-supervised adaptation as a defence mechanism against such perturbed domains. We perturb the clean Human3.6M~\cite{ionescu2013human3} samples to construct such a target domain,
denoted as \texttt{UAP-H3M}. Further, we experiment on three perturbation strengths while varying $\epsilon$ as $4/255$, $8/255$, and $16/255$ (see Table \ref{tab:comm_adapt}).
\vspace{-1mm}
\textbf{d) {LR-3DPW} (LR):} We use low-resolution (LR) variants of 3DPW~\cite{von2018recovering} dataset as another target domain, denoted as \texttt{LR-3DPW}. LR holds practical
significance as they are common in surveillance and real-time streaming applications. Several works~\cite{neumann2018tiny,xu20203d} acknowledge the performance drop in models trained on high-resolution (HR) data, while evaluating on LR
targets. Addressing this, Xu~\emph{et al}\onedot~\cite{xu20203d} construct an LR-specific architecture and assume simultaneous access to pose-labeled samples from both LR and HR domains, hence not comparable in our setting. We aim to evaluate
the proposed self-supervised adaptation technique as a remedy to such domain-shifts in the absence of target domain pose-labels. We construct three different LR domains by downsampling the full-resolution image to $w\!\times\!w$ where $w$ is set as 96, 52 and 32 (see Table \ref{tab:comm_adapt}). Each LR domain image is upsampled (bi-cubic) back to $224\!\times\!224$ before forwarding them through the CNN regressor.
\textbf{Adaptation settings.}
We evaluate our approach on the following adaptation settings. We use full supervision for source pre-training and use only silhouettes for target adaptation. In static-camera moving-object scenarios, silhouettes obtained via BG subtraction remain unaffected by domain shifts.
\vspace{-1mm}
\noindent
\textbf{a) \textit{Ours(S$\rightarrow$R)}} and \textbf{\textit{Ours(S$\rightarrow$R, weak)}.} In this setting, we initialize our Synthetic source model by accessing full supervision from SURREAL~\cite{varol2017learning} and undertake topological-alignment based adaptation on the \textit{Real} domain. \textit{Ours(S$\rightarrow$R)} does not access any \textit{Real} 2D/3D pose annotation. However, the weakly-supervised variant, \textit{Ours(S$\rightarrow$R, weak)}, utilizes
2D pose supervision.
\vspace{-1mm}
\noindent
\textbf{b) \textit{Ours(R$\rightarrow$UAP)}} and \textbf{\textit{Ours(R$\rightarrow$LR)}.} Here, we adapt the \textit{Real} domain source model, \textbf{\textit{Ours(R)}}, to the \texttt{UAP-H3M} and \texttt{LR-3DPW} domains respectively, by only accessing silhouettes from the target domains.
\begin{wraptable}[11]{r}{6cm}
\vspace{-5mm}
\footnotesize
\caption{\small Ablation experiments on 3DPW for S$\rightarrow$R.
\textit{B1}, \textit{B2}, and \textit{B3} are adaptation baselines that use different loss variants as listed below.
\vspace{-2mm}
}
\centering
\setlength\tabcolsep{5.0pt}
\resizebox{0.42\textwidth}{!}{
\begin{tabular}{lccc}
\midrule
Method &
\multirow{2}{*}{\makecell{Loss \\on $S$ }} &
\multirow{2}{*}{\makecell{Loss \\on $T$ }} &
\multirow{2}{*}{\makecell{PA-MPJPE \\($\downarrow$) }} \\ \\
\midrul
\textit{Ours(S)} & - & - & 134.7 \\
\textit{B1(S$\rightarrow$R)} & as $L2$ & - & 129.6 \\
\textit{B2(S$\rightarrow$R)} & $\mathcal{L}_{S}^{(sp)}$ & - & 106.2 \\
\textit{B3(S$\rightarrow$R)} & $\mathcal{L}_{S}^{(sp)}$ & as $L2$ & 105.5 \\
\textbf{\textit{Ours(S$\rightarrow$R)}} & $\mathcal{L}_{S}^{(sp)}$ & $\mathcal{L}_{T}^{(sp)}$ & \textbf{95.1} \\
\midrule
\end{tabular}}
\vspace{-1mm}
\label{tab:ablations}
\end{wraptable}
\vspace{-1mm}
\subsection{Ablative study}
\vspace{-2mm}
We define three {baseline} models that use various silhouette-based losses for self-supervised adaptation. In Table \ref{tab:ablations}, \textit{Ours(S)} reports the pre-adaptation performance in order to gauge the corresponding post-adaption improvements. \textit{B1}, a model adapted using only the pixel-level $L2$ loss (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\Vert S\!-\!\hat{S} \Vert_2^2$) shows very poor performance.
This shows that $L2$ based losses fall short of offering an independent articulation-centric objective. On the other hand, replacing it with our spatial chamfer-based alignment loss $\mathcal{L}_{S}^{(sp)}$ on raw-silhouettes, gives reasonable adaptation performance, shown as \textit{B2}. We observe that naively supplementing \textit{B2} with an $L2$ loss on topological-skeleton (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $\Vert \!T-\!\hat{T} \Vert_2^2$) yields very minimal benefit, despite $T$ containing useful articulation-centric information. Finally, we compare these baselines with our proposed model to clearly establish the effectiveness of our Chamfer inspired topological-alignment loss towards driving articulation-centric learning.
\begin{figure*}[!t
\begin{center}
\vspace{-4mm}
\includegraphics[width=1.0\linewidth]{figures_topo/Qualitative_results_shape_and_adaptation_NeurIPS.pdf} \vspace{-5mm}
\caption{ \small
\textbf{A.} Inference-time fitting on Thermal images.
\textbf{B.} Results on \texttt{LR-3DPW})
\textbf{C.} Results on \texttt{UAP-H3M}). Yellow ellipses highlight the region of articulation errors (refer Table~\ref{tab:comm_adapt} caption for comparison).
}
\vspace{-6mm}
\label{fig:adapt}
\end{center}
\end{figure*}
\vspace{-1mm}
\subsection{Comparison against prior works}\label{sec:4_2}
\vspace{-2mm}
Adhering to standard metrics,
we compute the mean per joint position error (MPJPE) \cite{ionescu2013human3} and Procrustes aligned \cite{gower1975generalized} mean per joint position error (PA-MPJPE) to evaluate the pose specific adaptation performance.
We evaluate on Human3.6M \cite{ionescu2013human3} following Protocol-2~\cite{kanazawa2018end}.
\begin{table}[t!]
\vspace{-4mm}
\footnotesize
\caption{\small
Evaluation
on \texttt{UAP-H3M} (P2)
and
\texttt{LR-3DPW}. We compare against two strong adaptation baselines: using 1) \textit{A}1: 2D pose
from an off-the-shelf network~\cite{cao2017realtime},
2) \textit{A}2: additional point-set based silhouette fitting-loss~\cite{huang2017towards}.
Comparing \textit{A}1, \textit{A}2, and \textit{Ours} on pre-to-post performance recovery,
we observe that the effectiveness of our approach increases with increasing domain shift.
\textit{+} represents adaptation using the specified loss.
}
\centering
\setlength\tabcolsep{3.0pt}
\resizebox{1.0\textwidth}{!}{
\begin{tabular}{ll|ccc|ccc||ccc|ccc}
\midrule
& \multirow{3}{*}{\centering Method} & \multicolumn{6}{c}{\makecell{ Adaptation from R to \texttt{UAP-H3M} \\ \midrule }} & \multicolumn{6}{c}{\makecell{ Adaptation from R to \texttt{LR-3DPW} \\ \midrule }} \\
& & \multicolumn{3}{c}{\makecell{MPJPE ($\downarrow$) \\ \midrule}} & \multicolumn{3}{c}{\makecell{PA-MPJPE ($\downarrow$) \\ \midrule}} & \multicolumn{3}{c}{\makecell{MPJPE ($\downarrow$) \\ \midrule}} & \multicolumn{3}{c}{\makecell{PA-MPJPE ($\downarrow$) \\ \midrule}} \\
& & $4/255$ & $8/255$ & $16/255$ & $4/255$ & $8/255$ & $16/255$ & $96$ & $52$ & $32$ & $96$ & $52$ & $32$ \\
\midrule\midrul
\multirow{2}{5em}{Pre-Adapt.} & SPIN~\cite{kolotouros2019learning} & 65.8 & 98.2 & 160.1 & 44.6 & 60.8 & 90.7 & 104.3 & 120.3 & 176.4 & 63.7 & 71.1 & 87.9 \\
& \textbf{\textit{Ours(R)}} & 67.7 & 103.9 & 161.8 & 46.9 & 63.6 & 91.2 & 110.8 & 127.5 & 178.1 & 68.6 & 76.3 & 88.2 \\
\midrule
\multirow{4}{5em}{Post-Adapt.} & \textit{A}1: SPIN+$\mathcal{L}_{2D}^{(p)}$ & 64.5 & 94.0 & 151.2 & 43.4 & 59.5 & 89.8 & 100.2 & 117.0 & 153.6 & 61.7 & 70.3 & 85.4 \\
& \textit{A}2: SPIN+$\mathcal{L}_{2D}^{(p)}$+$\mathcal{L}_{S}^{(p)}$ & 64.1 & 89.1 & 136.5 & 43.4 & 58.9 & 85.1 & 100.1 & 115.2 & 147.5 & 61.5 & 69.8 & 82.3 \\
& \textbf{\textit{Ours(R$\rightarrow$UAP)}} & \textbf{63.6} & \textbf{84.7} & \textbf{125.2} & \textbf{43.2} & \textbf{57.6} & \textbf{79.4} & - & - & - & - & - & - \\
& \textbf{\textit{Ours(R$\rightarrow$LR)}} & - & - & - & - & - & - & \textbf{99.8} & \textbf{114.7} & \textbf{134.2} & \textbf{61.3} & \textbf{68.7} & \textbf{78.9} \\
\hline
\end{tabular}}
\vspace{-6mm}
\label{tab:comm_adapt}
\end{table}
\vspace{-1mm}
\noindent
\textbf{a) Adaptation from Synthetic to Real.}
We evaluate
the proposed approach against prior human mesh recovery works on both
Human3.6M~\cite{ionescu2013human3} (Protocol-2)
and 3DPW~\cite{von2018recovering} datasets.
Our model achieves \textit{state-of-the-art} performance on self-supervised and weakly-supervised settings (see Table~\ref{tab:h36-p2} and Table~\ref{tab:3dpwresults}). Though we access synthetic domain data similar to Doersch~\emph{et al}\onedot~\cite{NEURIPS2019_d4a93297}, we deem ourselves comparable as other
listed works~\cite{kanazawa2018end,kolotouros2019learning,pavlakos2018learning,kanazawa2019learning} benefit from access to additional \textit{Real} domain datasets such as LSP, LSP-Extended~\cite{johnson2011learning} and COCO~\cite{lin2014microsoft}.
\vspace{-1mm}
\noindent
\textbf{b) Adaptation from Real to UAP-H3M.}
We obtain pre-adaptation performance of the source trained networks via direct inference on the shifted target (first 2 rows of Table~\ref{tab:comm_adapt}).
We report these in order to gauge the severity of the domain-gap. In \textit{A}1, SPIN~\cite{kolotouros2019learning} is finetuned on \texttt{UAP-H3M} targets by using 2D pose predictions from off-the-shelf OpenPose~\cite{8765346, cao2017realtime, wei2016cpm} network. Only the high-confidence 2D predictions
are used, inline with \cite{kolotouros2019learning}.
To create a strong baseline, in \textit{A}2, we supplement the \textit{A}1 setting with additional silhouette supervision, via the \textit{fitting-loss}~\cite{huang2017towards}.
As shown in last two rows of Table~\ref{tab:comm_adapt}, the proposed
adaptation method clearly imparts superior recovery of target performance, despite accessing only silhouettes.
The baselines \textit{A}1 and \textit{A}2, fail to recover from domain shift despite having access to stronger supervision. We observe that, unlike the silhouette extraction process, OpenPose itself suffers from domain shift issues, thus is ineffective towards improving target performance.
\vspace{-1mm}
\noindent
\textbf{c) Adaptation from Real to LR-3DPW.}
Here, we undertake adaptation to the \texttt{LR-3DPW} domain and evaluate
against the same baseline
settings as in R$\rightarrow$UAP
task. But unlike in UAP, the target silhouettes here are of lower resolution. Despite this, in Table~\ref{tab:comm_adapt}, we clearly see superior adaptation performance against the baselines ($A1$ and $A2$) which access additional 2D-pose predictions obtained via OpenPose~\cite{8765346, cao2017realtime, wei2016cpm}. These observations shine light on the ineffectiveness of network-based pseudo supervision (via OpenPose) towards recovering from domain-shift issues, and voices the need and usefulness of our framework in practical deployment scenarios.
\textbf{Qualitative results.}
Fig. \ref{fig:adapt} shows qualitative results
on \texttt{LR-3DPW} and \texttt{UAP-H3M}. Interestingly, our formulation of fitting in-the-loop with silhouette based objectives allows us to fit meshes using only silhouettes. Hence enabling us to perform fast inference-time fitting on diverse unseen domains such as thermal images, without requiring any 2D/3D pose annotations. We show a few such fittings in Fig.~\ref{fig:adapt}{\color{red}A}. We also observe that in contemporary CNN-based approaches, domain-shift primarily manifests articulation degradation, while person-orientation remains relatively robust.
Additionally, in Fig.~\ref{fig:qualitative}, we show a few pose recovery results of \textit{Ours(S$\rightarrow$R)}
on
Human3.6, 3DPW and other in-the-wild images. In Fig.~\ref{fig:qualitative}{\color{red}D}, we visualize the gradual improvement in the predicted mesh output across intermediate iterations of the training.
\textbf{Limitations.} Silhouettes offer significantly weaker supervision than 2D/3D pose and hence the model fails in presence of multi-level inter-part occlusions. Some complex poses also elicit limb-swapped articulations
(shown in magenta in Fig.~\ref{fig:qualitative}). Though, an only fitting based approach severely fails to solve such cases, optimization coupled with CNN training fairly recovers this deficiency as a result of the inductive bias instilled in the source trained model. Here, the inductive bias can be improved by training on multi-source data, such as combining $S$ and $R$ for adapting to extreme domain shifts like thermal or NIR.
In future, it remains interesting to explore ways to address such deficiencies via temporal and multi-frame articulation consistencies.
\textbf{Negative societal impacts.}
Although our approach in itself does not pose any direct negative societal impacts, we understand that the task at hand (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot 3D human pose recovery) may be directed towards unwarranted and intrusive human activity surveillance. We strongly urge
the practitioner to confine the applicability of the proposed method to strictly ethical and legal use-cases.
\vspace{-3mm}
\section{Conclusion}\vspace{-2mm}
In this paper, we introduce a self-supervised domain adaptation framework that relies only on silhouette supervision. The proposed convolution friendly spatial distance field computation helps us disentangle the topological-skeleton, thereby facilitating seamless morphological silhouette alignment. The empirical evaluations clearly highlight the usefulness of the framework for practical deployment scenarios, even as a remedy to adversarial attack. Integrating temporal aspects, such as optical-flow, in our framework remains an interesting future direction.
\begin{ack}
This work was supported by a Qualcomm Innovation Fellowship (Jogendra) and a project grant from MeitY (No.4(16)/2019-ITEA), Govt. of India. We would also like to thank the anonymous reviewers for their valuable suggestions.
\end{ack}
\small
\bibliographystyle{plainnat}
|
\section{Introduction}
The growing industrial interest in small-scale unmanned aerial vehicles (SUAVs) for sensing, reconnaissance, and parcel delivery continues to spur scientific interest into novel aerodynamic design solutions for low-speed flows, inspired by biological fliers \citep{Hassanalian2017}.
A special focus has been given to the membrane wings of bats, who possess impressive maneuvering and gliding abilities without relying upon high-frequency wing flapping for lift \citep{Hedenstrom2015}. These membrane wings are lightweight and are therefore appealing for SUAV applications. However, the compliance of membrane wings couples their geometrical shape and dynamics to the surrounding fluid dynamics and complicates the prediction of their aerodynamic performance.
Several computational and experimental studies have examined the aerodynamics of these flexible membrane wings in steady flow conditions over the last two decades \citep[e.g.,][]{Song2008,Gordnier2009,Rojratsirikul2009,Arbos-Torrent2013,Serrano-Galiano2018}. \cite{Gordnier2009} and \cite{Rojratsirikul2009} showed that membrane wings in low Reynolds number flows delay stall and enhance the mean lift coefficient due to the onset of flow-membrane oscillations. The oscillations of the compliant membrane are essential to these aerodynamic benefits, as \cite{Gordnier2009} found no significant advantage for a static membrane wing when compared to an equivalent (cambered) rigid aerofoil. Thus, the unsteady behaviour of membrane wings is a principal source of interest to discover mechanisms for aerodynamic performance improvement.
While many studies investigated the membrane wing response to steady flow conditions \citep[see][for a recent literature survey]{Tiomkin2021}, few have studied its response to unsteady flow conditions or prescribed motions, where aeroelastic membrane deformation may yield further aerodynamic benefits.
The pursuit of novel mechanisms to improve the unsteady aerodynamic performance of membrane wings and understand their associated fluid mechanics has led to a recent research focus on flapping membrane wings.
The combination of aerofoil flexibility with flapping motions can potentially eliminate flow separation along the aerofoil altogether and improve the aerodynamic maneuverability of the vehicle, as demonstrated in the context of bat flight \citep{Muijres2008,Chin2016}.
Several computational \citep{Gopalakrishnan2010,Jaworski2015} and experimental \citep{Tregidgo2013} studies investigated the membrane wing response to prescribed flapping motions, which are generally described as pitch or heave oscillations, or as a combination of the two.
\cite{Gopalakrishnan2010} used coupled large eddy simulations of a rectangular membrane to show that induced camber enhances both the lift and the thrust during a flapping pitching motions.
These simulations identified the movement of the leading-edge vortex along the membrane aerofoil surface to be the main source of the increased lift and thrust relative to a flapping rigid wing, for which the leading-edge vortex detaches and moves away of the wing which causes a drop in the lift coefficient \citep{Eldredge2019ARFM}. These results were supported by the computations of \cite{Jaworski2015} that focused on the role of prestress and elastic modulus in the propulsion of a flapping membrane aerofoil, which is enhanced by the interaction of the leading-edge vortex with the local elastic deformation. Experiments of \cite{Tregidgo2013} focused on the membrane dynamic response to a transient sinusoidal pitch maneuver of reduced frequency of $k=0.022$ and amplitude of $10^\circ$. Different vibration modes were identified that depended on the stationary angle of attack about which the unsteady maneuver was carried out. For small stationary angles of attack $(0^\circ\le\alpha\le4^\circ)$, first mode oscillations were observed with a small lag in the membrane dynamic response relative to the prescribed motion. This delay was more pronounced for a larger stationary angle of attack of $\alpha=10^\circ$, for which hysteresis was identified between the pitch-up and the pitch-down sections of the motion, which were accompanied by different vibrational modes.
The above studies collectively emphasize the complexity of the flapping membrane wing problem. However, due to their computational and experimental nature, their scope is limited to a few specific points in the parameter regime of flapping membrane wings, i.e., specific values of reduced frequency, mass ratio, and membrane elasticity. An analytical solution of a simplified model problem is therefore desired to shed light on the role of each dimensionless group in the wide parameter space of flapping membrane wings.
Physical insights from such an analytical solution are expected to inform future computational and experimental studies en route to obtaining a more complete understanding of the physics of flapping membrane wings.
Several analytical studies have been carried out that focus on propulsive thrust and efficiency predictions for flapping flexible wings.
\cite{Alben2008} presented an analytical solution for a flapping inextensible elastic sheet (with a free trailing edge), utilizing unsteady thin aerofoil theory coupled to a beam structural model. Their work identified an optimal thrust condition at the resonance peaks for small pitching amplitudes.
More recently, \citet{Alon2019} showed via analytical solution that a heaving membrane wing transitions between thrust and drag near the membrane resonance frequency, as the reverse von K\'{a}rm\'{a}n wake transitions to a traditional von K\'{a}rm\'{a}n wake.
In their reviews on flapping wing aerodynamics of biological and bio-inspired flyers, \cite{Shyy2013book,Shyy2016} highlighted the importance of using a time-domain approach to predict accurately the aerodynamic performance of flapping wings at the scale of bats and birds due to the inseparable flapping and body time-scales of these flyers, which is not the case for smaller insect-scale flyers.
Thus, while a quasi-steady model can make accurate predictions for insect-scale vehicles, this model assumption is not recommended for SUAV applications, where a time-dependent approach is essential to address vehicle stability and control.
Furthermore, whilst the studies of \cite{Alben2008} and \cite{Alon2019} elucidate the propulsive potential of flapping flexible wings, a theoretical basis to understand the membrane wing aerodynamic performance in prescribed flapping motions remains underdeveloped, specifically in terms of the ability to predict its unsteady lift and structural dynamic response.
In addition to the unsteady lift and thrust enhancement mechanisms engendered by membrane wings under prescribed flapping motions,
an understanding of the response of these wings to flow disturbances such as gusts is important to the design of membrane wing SUAVs.
Due to their small size and slow flight speed, SUAVs are especially susceptible to flight disruption from small gusts typical of urban environments \citep[][]{Watkins2006atmospheric,elbanhawi2017enabling,Jones2022}.
Classical linear unsteady aerodynamic theory \citep[see][]{vonKarman1938,Sears1940} predicts the transient lift response of a rigid aerofoil to transverse gusts of small gust ratios, where the gust ratio is the transverse gust amplitude divided by the freestream flow speed.
This theory has long been utilized to predict the unsteady load on rigid wings in terms of lift amplitude and phase lag.
However, when compliant membrane wings are considered, the lift response is composed of both the local change in angle of attack and the resulting deformation of the aerofoil. The membrane deformation couples aeroelastically to the aerodynamic load, which may amplify or attenuate the unsteady lift response.
Initial results by \cite{Berci2013} from a semi-analytical state-space model indicate the appearance of structural oscillations in the \emph{massless} membrane response to a sharp-edged transverse gust. However, these oscillations were described only in terms of the mid-chord membrane deformation and without consideration of the structural mode of oscillation and the lift response of the aerofoil.
An analytical model that is able to predict the membrane response to unsteady flow is currently lacking in the literature.
The current study aims to fill this knowledge gap by presenting an unsteady analytical model and its solution for a membrane wing in inviscid incompressible flow, under the unsteady conditions of prescribed motions or transverse gust profiles.
The transient membrane response is determined in the Laplace domain; steady-state harmonic oscillations of the membrane deformation and the unsteady lift response are investigated using a simplified solution in the frequency (Fourier) domain, which is convenient to compare against established rigid aerofoil theory.
These solutions yield novel extensions to the classical unsteady aerodynamic functions for flexible membrane wings.
The remainder of this paper is organised as follows.
Section~\ref{sec:Formulation} presents the mathematical problem for the generalised case of a membrane wing in arbitrary motion or gust, and for specific canonical unsteady flow scenarios.
In \S~\ref{sec:results}, the results of the theoretical model are presented in terms of membrane wing deformation and aerodynamic performance, as represented by extensions to the classical unsteady aerodynamic functions by Theodorsen, Wagner, Sears, and K\"{u}ssner.
Section~\ref{sec:conclusions} closes with concluding remarks.
\section{Formulation}\label{sec:Formulation}
\subsection{Membrane wing}\label{sec:Formulation-membrane}
Consider an \emph{extensible} membrane aerofoil of thickness $h$ and density $\rho_m$, which is held by simple supports at a distance $2b$ from one edge to the other. The membrane is initially still and taut, and is immersed in a uniform and inviscid incompressible freestream of density $\rho$ and speed $U$, aligned parallel to the membrane chord.
Assuming small deformations of the membrane, the membrane dynamic equation is
\begin{equation}
\label{eq:DynamicMem-linear}
\rho_m h \,\tilde{y}_{\tilde{t}\tilde{t}} = T \,\tilde{y}_{\tilde{x}\tilde{x}} + \Delta p ,
\end{equation}
where $\tilde{y}$ denotes the membrane profile, $\tilde{t}$ represents time, $\tilde{x}$ is a coordinate along the chord, and $T$ and $\Delta p$ are the tension and pressure difference along the membrane, respectively.
While the membrane is extensible, we note that \cite{Tiomkin2017} showed that the tension coefficient can be considered constant to leading order for the small angles of attack and deformations assumed in the current study.
The non-dimensional form of the dynamic equation is
\begin{equation}
\label{eq:DynamicSail-LinearNorm}
4\mu \,y_{tt}=2C_T\,y_{xx}+\Delta C_p ,
\end{equation}
in which $b, b/U, \rho, U, \frac{1}{2}\rho U^2,$ and $\rho U^2 b$ are used as the units of length, time, density, circulation (per unit length), pressure, and force (per unit span).
Note that $b$ is used as the unit of length throughout the dynamic equation, but the mass ratio is normalised with $c$ as the unit of length, namely $\mu=\rho_m h/\rho c$, following the convention in previous membrane wing studies \citep[][]{Jaworski2012,Alon2019}.
The mass ratio $\mu$ and tension coefficient $C_T$ are fixed parameters in the present analysis, and the unsteady membrane deformation and pressure coefficient profiles, $y$ and $\Delta C_p$, respectively, are part of the solution.
\subsection{Incompressible potential flow}
The extensible membrane aerofoil may encounter or produce an unsteady flow field that superposes on the uniform background flow. Inviscid, incompressible potential flow is considered with an initial angle of attack of $\alpha=0^\circ$, which isolates the effects of unsteady angle of attack variations or transient gusts on the membrane dynamic response; this approach is similar to the traditional formulation available for the arbitrary motion of rigid aerofoils \citep[e.g.,][]{Bisplinghoff_book1996}.
For completeness, this section outlines the formulation of \cite{Tiomkin2017} for a membrane aerofoil in steady flow and extends it to include the dynamic membrane response to an unsteady flow.
The standard coordinate transformation
\begin{equation}
\label{eq:Tet-transformation}
x=-\cos\theta
\end{equation}
places the profile leading edge at $x=-1$ ($\theta=0$) and the trailing edge at $x=1$ ($\theta=\upi$).
Let us represent the membrane slope, $y_x$, by a Fourier series expansion per \cite{Nielsen1963}, which is augmented by permitting the Fourier coefficients to be time-dependent:
\begin{equation}
\label{eq:SailSlope-Fourier}
y_x(t,\theta)=\frac{1}{2}F_0(t)+\sum_{n=1}^\infty F_n(t)\cos n\theta .
\end{equation}
We proceed with expressing the membrane dynamic equation \eqref{eq:DynamicSail-LinearNorm} in terms of the new coordinate $\theta$, which yields a system of differential equations for the Fourier coefficients.
Integration of \eqref{eq:SailSlope-Fourier} along the horizontal coordinate, from the leading edge to a point $x$ along the chord, yields the membrane profile:
\begin{eqnarray}
\label{eq:SailShape_Fourier}
y\left(t,\theta\right)&=& \frac{1}{2}F_0(t)\left(1-\cos\theta\right)+\frac{1}{2}F_1(t)\sin^2\theta \nonumber\\
&& \mbox{}- \frac{1}{2}\sum_{n=2}^\infty F_n(t) \left(\frac{1}{n^2-1}\right)
\,\left[2+(n-1)\cos \left(n+1\right)\theta - (n+1)\cos \left(n-1\right)\theta\right] , \quad\qquad
\end{eqnarray}
which has to sustain the fixed boundary conditions of the membrane edges. The leading-edge boundary condition is automatically satisfied by \eqref{eq:SailShape_Fourier}. However, the fixed trailing-edge boundary condition imposes the constraint
\begin{equation}
\label{eq:F_0_Time}
F_0(t)=2\sum_{\substack{m=1}}^\infty \frac{F_{2m}(t)}{(2m)^2-1} .
\end{equation}
In addition, the assumption of an initially still and taut membrane yields zero initial conditions for the Fourier coefficients and their first time derivative.
The pressure difference across the membrane in \eqref{eq:DynamicSail-LinearNorm} is obtained by using the unsteady vortex sheet method \cite[][p. 274]{Bisplinghoff_book1996},
\begin{equation}
\label{eq:DP-def}
\Delta C_p(t,x)=2 \gamma(t,x) + 2\frac{\partial}{\partial t} \int_{-1}^x \gamma \left(t,\zeta\right) \mathrm{d} \zeta ,
\end{equation}
where $\gamma$ is the normalised vortex sheet strength per unit length along the profile.
The vorticity distribution along the aerofoil is determined by the fundamental equation of thin aerofoil theory,
\begin{equation}
\label{eq:ThinAirfoil-SmallCamber_norm}
\frac{1}{2\upi}\Xint-_{-1}^1\frac{\gamma(t,\xi)}{x-\xi}\mathrm{d}\xi= w_a(t,x) - \frac{1}{2\upi}\int_{1}^{1+t}\frac{\gamma_w\left(t,\eta\right)}{x-\eta}\mathrm{d}\eta, \qquad x\in \left(-1,1\right),
\end{equation}
where the dashed integral denotes the Cauchy principal value.
The normal velocity on the membrane surface (normalised by $U$) is
\begin{equation}
\label{eq:wa_def}
w_a(t,x)= w_{a_d}(t,x)+w_{a_f}(t,x),
\end{equation}
where $w_{a_\mathit{f}}(t,x)$ is the contribution of the unsteady flow (i.e., prescribed chord motion or a traveling gust) to the normal flow velocity on the membrane, and $w_{a_{d}}(t,x)$ is the respective contribution of the membrane deformation,
\begin{equation}
\label{eq:wa_s_def}
w_{a_d}(t,x)= -y_x(t,x)-y_t(t,x) .
\end{equation}
$\gamma_w (t,\eta)$ describes the normalised vorticity per unit length at location $\eta$ along the wake, $\eta\in(1,\infty)$, at time $t$.
Wake vortices are assumed to be continuously shed from the trailing edge at the freestream velocity into a flat wake and have fixed strength, which asserts that the wake vorticity distribution, $\gamma_w (t,\eta)$, is equivalent to the vorticity at the trailing edge at time $t-\eta+1$:
\begin{equation}
\gamma_w (t,\eta)=\gamma_w (t-\eta+1,1)\triangleq \gamma_{_{T\!E}}(t-\eta+1).
\end{equation}
Application of S\"{o}hngen's inversion formula to \eqref{eq:ThinAirfoil-SmallCamber_norm} and enforcement of Kelvin's theorem \citep[cf.,][p. 289]{Sohngen1939,Bisplinghoff_book1996} leads to
\begin{equation}
\label{eq:GammaInt-step5}
2\int_{-1}^{1} \sqrt{\frac{1+\xi}{1-\xi}}\,w_a(t,\xi)\mathrm{d}\xi = -\int_{1}^{1+t}\sqrt{\frac{\eta+1}{\eta-1}}\gamma_{_{T\!E}} (t-\eta+1) \mathrm{d}\eta .
\end{equation}
\cite{Tiomkin2017} showed that the application of the Laplace transform to \eqref{eq:GammaInt-step5} yields a closed-form expression for the wake vorticity distribution in the Laplace plane.
Provided that a solution for $\gamma_{_{T\!E}}$ is obtainable in the time domain, the method of \cite{Schwarz1940} \citep[see also][]{Iosilevskii2007} produces a general expression for the pressure difference along the aerofoil:
\begin{eqnarray}
\label{eq:DP-GeneralSol}
\Delta C_p(t,x)&=& -\frac{4}{\upi}\sqrt{\frac{1-x}{1+x}}\Xint-_{-1}^{1} \sqrt{\frac{1+\xi}{1-\xi}}\,\frac{w_a(t,\xi)}{x-\xi}\mathrm{d}\xi + \frac{4}{\upi}\Xint-_{-1}^{1}\Lambda_1 (x,\xi) w_{a_t}(t,\xi)\mathrm{d}\xi
\nonumber\\
&&\mbox{} + \frac{2}{\upi}\sqrt{\frac{1-x}{1+x}} \int_{1}^{1+t} \frac{\gamma_{_{T\!E}} (t-\eta+1)}{\sqrt{\eta^2-1}}\mathrm{d}\eta ,
\end{eqnarray}
where $\Lambda_1$ is an auxiliary function expressed in \eqref{eq:a4_Lam1} of appendix~\ref{a4:Math}.
The first term in the pressure difference expression describes the quasi-steady pressure difference, the second term is the apparent mass contribution (non-circulatory term), and the third term describes the contribution of the wake.
Substitution of \eqref{eq:wa_def} into \eqref{eq:DP-GeneralSol} permits a separation of the effect of the membrane deformation, $w_{a_d}$, from the effect of the unsteady flow, $w_{a_f}$, on the aerodynamic load, namely
\begin{equation}
\label{eq:Dcp_tot}
\Delta {C_p} (t,\theta) = \Delta C_{p_d} (t,\theta)+\Delta C_{p_f} (t,\theta) ,
\end{equation}
where the subscripts $d$ and $f$ denote terms due to membrane deformation and unsteady flow, respectively.
Details of the analytical expressions for $\Delta C_{p_d}$ (in the Laplace plane) are available in appendix A of \cite{Tiomkin2017}. We develop in \S~\ref{sec:UnsteadyFlow} the closed-form expressions for $\Delta C_{p_f}$ that are necessary to complete the description of the aerodynamic load on a membrane undergoing prescribed chord motion or encountering a gust.
Note that the membrane Fourier coefficients appear only in the expression for $\Delta C_{p_d}$, whilst $\Delta C_{p_f}$ depends only on the prescribed motion or gust.
The next section combines the terms obtained for the membrane deformation and the resulting aerodynamic load to obtain a set of equations for the coupled aeroelastic problem for any arbitrary prescribed chord motion or gust.
\subsection{Aeroelastic coupling and methods of solution}
The coupled aeroelastic equation that describes the membrane response to unsteady flow conditions is obtained by substituting \eqref{eq:SailSlope-Fourier}, \eqref{eq:SailShape_Fourier}, and \eqref{eq:Dcp_tot} into \eqref{eq:DynamicSail-LinearNorm}.
This procedure yields a matrix equation in which the unknowns are the Fourier coefficients that describe the membrane deformation.
The aeroelastic equation is described and solved in the Laplace domain for generalised time-dependent cases, or in the frequency domain for harmonic motions or gusts. Details of these two methods are given next in \S\S~\ref{sec:Laplace} and \ref{sc:FD_solution}, respectively.
Note that throughout the manuscript overbar and hat symbols are used to denote variables in the Laplace and frequency domains, respectively.
\subsubsection{Laplace domain}\label{sec:Laplace}
The membrane dynamic equation \eqref{eq:DynamicSail-LinearNorm} is expressed in the Laplace domain by applying the Laplace transform to \eqref{eq:SailSlope-Fourier}, \eqref{eq:SailShape_Fourier}, and \eqref{eq:Dcp_tot} and substituting the resulting expressions into the Laplace transform of \eqref{eq:DynamicSail-LinearNorm}.
We then multiply the resulting equation by $\sin\theta$ and use the mathematical relations \eqref{eq:a4_trigo3} and \eqref{eq:a4_trigo4} to construct a matrix system of equations,
\begin{equation}
\label{eq:MatrixEqUS-Lap}
\left\{\ssA s^2 + \ssB s + \ssE\right\}\bar{\boldsymbol{F}}= \boldsymbol{H} ,
\end{equation}
where $\bar{\boldsymbol{F}}$ is the vector of Fourier coefficients $\bar{F}_n (s) , n=1\ldots N$, and $N$ is the number of coefficients chosen to represent the membrane-profile slope in \eqref{eq:SailSlope-Fourier}, taken here as $N=24$ following the numerical convergence studies of \cite{Nielsen1963} and \cite{Tiomkin2017}. Here, the overbar denotes the Laplace transform of the variable, $\bar{F}_n(s) = \Lapl \left\{F_n(t);s\right\}$. The matrices $\ssA, \ssB, \ssE$, and the vector $\boldsymbol{H}$ are obtained by matching the coefficients of the harmonics of $\sin\theta$ in the dynamic equation \eqref{eq:DynamicSail-LinearNorm}. It is noted here that matrices $\ssB$ and $\ssE$ and vector $\boldsymbol{H}$ depend on the Laplace variable $s$, while $\ssA$ is constant.
We further note that the matrices $\ssA,\ssB,$ and $\ssE$ are obtained from the steady flow solution under zero angle of attack, i.e., by applying $\Delta {C_p} = \Delta C_{p_d}$ to the dynamic equation; these matrices are detailed in \cite{Tiomkin2017}.
The effect of the unsteady flow appears only in $\boldsymbol{H}$, on the right hand side of the resulting dynamic equation, and is determined by $\Delta C_{p_f}$. Thus, $\Delta C_{p_f}$ acts as an excitation force that is applied to the membrane.
The Fourier coefficients, $\bar{\boldsymbol{F}}$, can now be computed from \eqref{eq:MatrixEqUS-Lap}, and their substitution into the Laplace transform of \eqref{eq:SailShape_Fourier} produces the membrane dynamic solution in the Laplace domain.
This approach predicts the membrane dynamic response to any arbitrary motion or gust.
However, a numerical Laplace inversion is required to obtain a solution in the time domain, as no analytical expression is available for the inverse Laplace transform of our problem. We apply the numerical scheme of \cite{Valsa1998} to carry out this inversion, which is robust and reliable for both oscillatory and non-oscillatory functions.
A solution can alternatively be determined
in the frequency (Fourier) domain by setting $s=\mathrm{i} k$, where $k$ is the reduced frequency \citep[][p. 292]{Bisplinghoff_book1996}. This approach computes readily the steady-state response of the membrane wing to harmonic gusts or motions. However, this method cannot obtain the transient response of the membrane and will therefore only be used here for the harmonic cases and as a means of verification of the indicial lift responses obtained in the Laplace domain.
Details of the application of the frequency-domain method are presented in the next section.
\subsubsection{Frequency domain}\label{sc:FD_solution}
The assumption of harmonic motion for all variables converts the membrane dynamic solution to the frequency domain, where, for example $y(t,x)=\hat{y}(k,x)\, \mathrm{e}^{\mathrm{i} kt}$, and the hat denotes a complex-valued amplitude. Assignment of $s=\mathrm{i} k$ into \eqref{eq:MatrixEqUS-Lap} yields this equation in the frequency domain,
\begin{equation}
\label{eq:MatrixEqUS-Freq}
\left\{-\ssA\, k^2 + \hat{\ssB}\, \mathrm{i} k + \hat{\ssE}\right\}\hat{\boldsymbol{F}}= \hat{\boldsymbol{H}} ,
\end{equation}
where $\hat{\ssB}=\ssB(s=\mathrm{i} k), \hat{\ssE}=\ssE(s=\mathrm{i} k), \hat{\boldsymbol{H}}=\boldsymbol{H}(s=\mathrm{i} k)$, and $\hat{\boldsymbol{F}}$ is the vector of complex amplitudes of the Fourier coefficients $\hat{F}_n (s) , n=1\ldots N$.
Once determined by \eqref{eq:MatrixEqUS-Freq}, these Fourier coefficients produce the resulting membrane deformation through \eqref{eq:SailShape_Fourier}.
Note that the constant matrix $\ssA$ is unaffected by the shift from the Laplace \eqref{eq:MatrixEqUS-Lap} to the frequency \eqref{eq:MatrixEqUS-Freq} domain.
The frequency domain analysis is a natural approach to study the canonical unsteady aerodynamic scenarios of Theodorsen (harmonic oscillations) and Sears (sinusoidal gust). For these two scenarios, the frequency-domain solution yields the membrane dynamic response and the aerodynamic lift response; these results are compared against the Laplace domain results for verification.
Integration of the Theodorsen and Sears harmonic functions over the entire frequency domain yields the indicial lift responses to a step change in angle of attack (Wagner's function) and to a sharp-edged gust (K\"{u}ssner's function), respectively (\citeauthor{Bisplinghoff_book1996}, \citeyear{Bisplinghoff_book1996}, pp.~284-287; \citeauthor{Baddoo2021unsteady}, \citeyear{Baddoo2021unsteady}).
Thus, the frequency domain solution can generate all four canonical functions for the membrane aerofoil, whilst the transient dynamic response of the membrane can only be studied through the Laplace domain solution.
Case-specific technical details for this approach are further discussed in \S\S~\ref{sc:Formulation-EquivTheodorsen} and \ref{sc:Formulation-EquivSears} for harmonic heave motions and sinusoidal gusts, respectively.
\subsection{Unsteady flow conditions}\label{sec:UnsteadyFlow}
We next describe the unsteady flow conditions that the membrane aerofoil encounters for two cases of prescribed motion: generalised and harmonic heave motions, and a step change in angle of attack.
We describe the generalised problem of a membrane aerofoil that encounters a small-amplitude transverse gust, and then focus on two canonical gust profiles of sinusoidal and sharp-edged geometry.
For each of these cases, the aerodynamic load due to the unsteady flow, $\Delta C_{p_f}$, is derived by applying $w_a=w_{a_f}$ to \eqref{eq:DP-GeneralSol}, where the normal velocity on the aerofoil, $w_{a_f}$, is defined according to \cite{vonKarman1938} and \cite{wagner1925}. $\Delta C_{p_f}$ forms the $\boldsymbol{H}$ vector in \eqref{eq:MatrixEqUS-Lap} for the Laplace-domain solution, or the $\hat{\boldsymbol{H}}$ vector in \eqref{eq:MatrixEqUS-Freq} for the frequency-domain solution, which concludes the formulation of our problem.
\subsubsection{Prescribed heave motion}\label{sec:Heave_def}
Consider a membrane aerofoil that performs a prescribed translatory motion, $h(t)$, that is normal to the flight direction, where $h$ is normalised by $b$ and is positive downwards.
Under these conditions, the aerodynamic load on the membrane due to heave motion is obtained by substituting $w_a(t,x)=\dot{h}(t)$ into \eqref{eq:GammaInt-step5}, where the overdot denotes a time derivative.
The solution of \eqref{eq:GammaInt-step5} in the Laplace domain yields a closed-form expression for the wake vorticity distribution, which is substituted into the Laplace transform of \eqref{eq:DP-GeneralSol}. Subsequent application of the coordinate transformation \eqref{eq:Tet-transformation} yields the Laplace transform of the aerodynamic load due to harmonic heave oscillations:
\begin{equation}
\label{eq:DCp_h-Laplace}
\Delta \bar{C}_{p_h} (s,\theta ) = 4 s^2 \bar{h}(s)\left[\bar{\Phi}(s) \cot{\frac{\theta}{2}}+\sin{\theta}\right] ,
\end{equation}
where $\bar{h}(s)$ is the Laplace transform of the prescribed heave motion, and $\bar{\Phi}(s)$ is the Laplace transform of Wagner's function \citep[e.g.,][]{Sears1940},
\begin{equation}
\label{eq:Wagner_Lap}
\bar{\Phi}(s)=\frac{{C}(s)}{s}.
\end{equation}
Here ${C}(s)$ is the generalised Theodorsen function \citep[e.g.,][]{Edwards1979},
\begin{equation}
\label{eq:C_Th-Laplace}
C(s) = \frac{K_1(s)}{K_0(s)+K_1(s)},
\end{equation}
and $K_0$ and $K_1$ are modified Bessel functions of the second kind.
In the frequency domain, assuming $h(t) = h_0\, \mathrm{e}^{\mathrm{i} kt}$, the amplitude of the effective angle of attack is $\alpha_0=\mathrm{i} k h_0$, and the aerodynamic load due to harmonic heave oscillations is
\begin{equation}
\label{eq:DCp_h-Freq}
\hat{\Delta {C_p}_h} (k,\theta) = 4 (\mathrm{i} k)h_0 C(k)\, \cot{\frac{\theta}{2}} -4 k^2 h_0 \sin{\theta} ,
\end{equation}
where $C(k)$ is the frequency-domain Theodorsen's function,
\begin{equation}
\label{eq:C_Th-FreqDomain}
C(k) = \frac{H_1^{(2)}(k)}{H_1^{(2)}(k)+\mathrm{i} H_0^{(2)}(k)},
\end{equation}
and $H_0^{(2)}$ and $H_1^{(2)}$ are Hankel functions of the second kind.
The corresponding lift response is
\begin{equation}
\label{eq:CL_h-Freq}
{C_l}_{_h}(t) = 2\pi C(k)\,\dot{h}(t) + \pi\ddot{h}(t) ,
\end{equation}
which is in fact the rigid plate response to heave oscillations \citep[][p. 272]{Bisplinghoff_book1996}.
The aerodynamic load expressions \eqref{eq:DCp_h-Laplace} and \eqref{eq:DCp_h-Freq} form the term $\Delta C_{p_f}$ in \eqref{eq:Dcp_tot} for prescribed heave motions in the Laplace and frequency domains, respectively.
\subsubsection{Step angle of attack}
The canonical unsteady aerodynamics problem for the indicial lift response of an aerofoil to a step in angle of attack was originally solved by \citet{wagner1925} for a rigid flat plate.
The aerodynamic load on a membrane due to a step change in angle of attack may be computed using $w_a(t,x)=~\alpha(t)$, where
\begin{equation}
\label{eq:alpha_t-AoA_step}
\alpha(t) = \alpha_0 \,\mathcal{H}(t)\; \Rightarrow \; \bar{\alpha}(s) = \frac{\alpha_0}{s} ,
\end{equation}
and $\mathcal{H}(t)$ is the Heaviside function.
The procedure detailed in \S~\ref{sec:Heave_def} is repeated to derive an expression for the pressure difference coefficient,
\begin{equation}
\label{eq:DCp_a_AoA_step-Laplace}
\Delta \bar{C}_{p_{\alpha_0}}\left(s,\theta\right) = 4 \alpha_0\left[\bar{\Phi}(s)\,\cot{\frac{\theta}{2}}+\sin{\theta}\right].
\end{equation}
The corresponding lift coefficient due to a step change in angle of attack is
\begin{equation}
\label{eq:CL_a_AoA_step-Laplace}
\bar{C}_{l_{_{\alpha_0}}}(s) = 2\pi\alpha_0\left[\bar{\Phi}(s)+\frac{1}{2}\right],
\end{equation}
which recovers the indicial lift response found by \cite{wagner1925}.
\subsubsection{Generalised transverse gust}\label{sc:Formulation-Gust}
We consider a membrane aerofoil that encounters a vertically-oriented gust with an arbitrary profile. The leading edge of the aerofoil encounters the gust front at time $t=0$ (figure~\ref{fig:MemGustSkatch_t}), and the gust amplitude is assumed to be small with respect to the freestream velocity.
The aerodynamic load due to the imposed unsteady flow, $\Delta C_{p_f}$, is briefly detailed here to complete the formulation of the problem, as it is in fact the rigid aerofoil response to a transverse gust presented in appendix B of \cite{Iosilevskii2007}.
Note that $\Delta C_{p_f}$ in \eqref{eq:Dcp_tot} is denoted $\Delta C_{p_g}$ in the present context of gusts.
The gust model assumptions permit the substitution of $w_a(t,x)=\alpha_g(t-x-1)$ into \eqref{eq:GammaInt-step5} to produce the gust effect on the vorticity distribution along the wake, which leads to a closed-form expression for the aerodynamic load along the aerofoil using \eqref{eq:DP-GeneralSol}:
\begin{equation}
\label{eq:DCp_g-Laplace}
\Delta \bar{C}_{p_g}\left(s,\theta\right) = 4s\,\bar{\Psi}(s)\, \bar{\alpha}_g(s) \,\cot{\frac{\theta}{2}},
\end{equation}
where
\begin{equation}
\label{eq:Kussner-Laplace}
\bar{\Psi}(s)=\frac{\mathrm{e}^{-s}}{s^2}\frac{1}{K_0(s)+K_1(s)}
\end{equation}
is the Laplace transform of K\"{u}ssner's function, $\Psi(t)$ \citep[e.g.,][]{Sears1940}. Note that the aerodynamic load due to an arbitrary transverse gust is obtained by a convolution of K\"{u}ssner's function and the time derivative of the gust angle of attack,
\begin{equation}
\label{eq:DCp_g-Time}
\Delta C_{p_g}\left(t,\theta\right) = 4\cot{\frac{\theta}{2}}\int_0^t \Psi\left(t-\tau\right)\frac{\mathrm{d}\alpha_g(\tau)}{\mathrm{d} \tau}\mathrm{d}\tau .
\end{equation}
Equation \eqref{eq:DCp_g-Laplace} forms an expression for $\Delta C_{p_f}$ in the Laplace domain, for an arbitrary, small-amplitude gust profile.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Membrane_gust_t0.eps}
\caption{$t=0$}
\label{fig:MemGustSkatch_t0}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/MemGust_tg0_noleg_b.eps}
\caption{$t>0$}
\label{fig:MemGustSkatch_tg0}
\end{subfigure}
\caption{Sketch of the membrane aerofoil gust problem: (a) initial time $t=0$; (b) later time $t>0$. Note that the membrane is initially taut, at zero angle of attack, and deforms under transient (gust) fluid loads.}
\label{fig:MemGustSkatch_t}
\end{center}
\end{figure}
\subsubsection{Sinusoidal gust}\label{sc:Formulation-Sinus_gust}
A sinusoidal gust encountered at the leading edge at time $t=0$ can be expressed as
\begin{equation}
\label{eq:alpha_g-sinus}
\alpha_g\left(t-x-1\right) = \alpha_0\sin\left(k\left(t-x-1\right)\right) \, \mathcal{H}\left(t-x-1\right).
\end{equation}
The substitution of the Laplace transform of \eqref{eq:alpha_g-sinus} into \eqref{eq:DCp_g-Laplace} yields the aerodynamic load along the aerofoil due to a sinusoidal gust,
\begin{equation}
\label{eq:DCp_g-sinus}
\Delta \bar{C}_{p_g}\left(s,\theta\right) = 4k\alpha_0 \frac{s}{s^2+k^2}\,\bar{\Psi}(s)\, \cot{\frac{\theta}{2}}.
\end{equation}
It is natural to solve the steady-state problem for harmonic gusts in the frequency domain by assuming
\begin{equation}
\label{eq:alpha_g-exp_ikt}
\alpha_g\left(t-x-1\right) = \alpha_0\, \mathrm{e}^{\mathrm{i} k\left(t-x-1\right)},
\end{equation}
which yields
\begin{equation}
\label{eq:DCpg_exp_ikt-FD}
\Delta \hat{C}_{p_g}\left(k,\theta\right) = 4\alpha_0\mathrm{i} k\,\hat{\Psi}(k)\cot{\frac{\theta}{2}} .
\end{equation}
Here
\begin{equation}
\label{eq:Kussner-FD}
\hat{\Psi}(k)=\frac{2}{\pi}\frac{\mathrm{e}^{-\mathrm{i} k}}{k^2}\left(\frac{1}{H_1^{(2)}(k)+\mathrm{i} H_0^{(2)}(k)}\right)
\end{equation}
is the Fourier transform of the K\"{u}ssner function, which is obtained by assigning $s=\mathrm{i} k$ in \eqref{eq:Kussner-Laplace}.
The lift response to sinusoidal gusts in the frequency domain is
\begin{equation}
\label{eq:L_g-Time-sinus}
\hat{C}_{l_g}(k) = 2\pi\alpha_0\mathrm{i} k\, \hat{\Psi}(k) = 2\pi \alpha_0\, S(k),
\end{equation}
where the modified Sears function, $S(k)$, is given by \citep[][p. 287]{Bisplinghoff_book1996}
\begin{equation}
\label{eq:Sears-FD}
S(k) = \mathrm{i} k\,\hat{\Psi}(k) = \left\{C(k)\left[J_0(k)-\mathrm{i} J_1(k)\right]+\mathrm{i} J_1(k)\right\}\mathrm{e}^{-\mathrm{i} k} = \tilde{S}(k)\,\mathrm{e}^{-\mathrm{i} k} .
\end{equation}
$\tilde{S}(k)$ is the classical Sears function whose gust front is at the mid-chord location at time $t=0$, and $J_0$ and $J_1$ are Bessel functions of the first kind.
\subsubsection{Sharp-edged gust}
A sharp-edged gust is similarly described by
\begin{equation}
\label{eq:alpha_g-SEG}
\alpha_g\left(t-x-1\right) = \alpha_0 \,\mathcal{H}\left(t-x-1\right).
\end{equation}
The substitution of \eqref{eq:alpha_g-SEG} into \eqref{eq:DCp_g-Laplace} yields the aerodynamic load along the aerofoil due to a sharp-edged gust,
\begin{equation}
\label{eq:DCp_g-SEG}
\Delta \bar{C}_{p_g}\left(s,\theta\right) = 4\alpha_0\,\cot{\frac{\theta}{2}}\,\bar{\Psi}(s).
\end{equation}
The resulting lift response is
\begin{equation}
\label{eq:L_g-SEG}
\bar{C}_{l_g} (s)=2\pi\alpha_0\,\bar{\Psi}(s),
\end{equation}
which is the expected classical indicial lift of a rigid plate due to a sharp-edged gust.
The distributed aerodynamic load \eqref{eq:DCp_g-SEG} is in fact the external force applied on the membrane through the term $\boldsymbol{H}$ in \eqref{eq:MatrixEqUS-Lap} in the case of a sharp-edged gust.
This applied force initiates a membrane deformation which brings about a change in the aerodynamic load through aeroelastic coupling.
\subsection{Unsteady lift response functions}\label{sec:LiftFunctions}
The membrane unsteady lift coefficient is derived by integration of the aerodynamic load \eqref{eq:Dcp_tot} along the membrane chord-line, which leads to
\begin{equation}
\label{eq:CL_tot}
C_{l_m}(t) = C_{l_d}(t)+C_{l_f}(t) ,
\end{equation}
where the normalised lift due to membrane deformation is
\begin{equation}
\label{eq:Cl_d}
\frac{C_{l_d}(t)}{2\pi\alpha_0} = \int_0^t \Phi(t-\tau)\dot{f}(\tau)\mathrm{d}\tau +g(t) .
\end{equation}
Here, $\Phi(t)$ is the time-domain Wagner function, and $f(t)$ and $g(t)$ are functions of the Fourier coefficients given by
\begin{eqnarray}
\label{eq:f_t}
f(t) &=& \frac{1}{2}\mathcal{F}_{1}(t) - \frac{1}{2}\mathcal{F}_{0}(t) -\frac{1}{4} \dot{\mathcal{F}}_0(t) -\frac{1}{4} \dot{\mathcal{F}}_1(t) + \frac{1}{4} \dot{\mathcal{F}}_2(t) + \sum_{m=2}^{N/2} \frac{\dot{\mathcal{F}}_{2m-1}(t)}{(2m-1)^2-1},\\
\label{eq:g_t}
g(t) &=& -\frac{1}{4}\dot{\mathcal{F}}_0(t) + \frac{1}{4}\dot{\mathcal{F}}_2(t) - \frac{3}{16} \ddot{\mathcal{F}}_1(t) + \frac{1}{8} \ddot{\mathcal{F}}_3(t) +\frac{1}{2} \sum_{m=3}^{N/2} \frac{\ddot{\mathcal{F}}_{2m-1}(t)}{(2m-1)^2-1} ,
\end{eqnarray}
where $\mathcal{F}_n=F_n/\alpha_0$.
The term ${\alpha}_0$ is the unsteady angle of attack amplitude in the harmonic cases, or the steady angle of attack in the indicial cases; the reader may consult \S~\ref{sec:UnsteadyFlow} for details on the definition of ${\alpha}_0$ and the lift due to the unsteady flow, $C_{l_f}$, for each case of prescribed chord motion or gust encounter considered here.
The Laplace transform of $C_{l_d}(t)$,
\begin{equation}
\label{eq:Cl_d-Lap}
\frac{\bar{C}_{l_d}(s)}{2\pi\alpha_0 } = C(s)\bar{f}(s) +\bar{g}(s),
\end{equation}
obtains the lift coefficient due to membrane deformation in the Laplace domain for indicial scenarios.
Subsequent numerical Laplace inversion yields the membrane indicial lift response functions in the time domain.
For cases of harmonic oscillations, in which the membrane solution is obtained in the frequency domain, the lift coefficient due to membrane deformation is expressed in the frequency domain
\begin{equation}
\label{eq:Cl_d-Freq}
\frac{\hat{C}_{l_d}(k)}{2\pi\alpha_0 } = C(k)\hat{f}(k) +\hat{g}(k),
\end{equation}
where $\hat{f}(k)$ and $\hat{g}(k)$ are found by substituting $\mathcal{F}_n(t)=\hat{\mathcal{F}}_n(k)\, \mathrm{e}^{\mathrm{i} kt}$ into \eqref{eq:f_t} and \eqref{eq:g_t}, respectively, and the auxiliary functions in the time domain become {$f(t)=\hat{f}(k)\, \mathrm{e}^{\mathrm{i} kt}$} and {$g(t)=\hat{g}(k)\, \mathrm{e}^{\mathrm{i} kt}$}.
Substitution of $C_{l_d}$ (\eqref{eq:Cl_d-Lap} or \eqref{eq:Cl_d-Freq} for the indicial or harmonic scenarios, respectively) and the case-specific $C_{l_f}$ into \eqref{eq:CL_tot} yields a closed-form expression for the total membrane lift coefficient, $C_{l_m}$, from which extensions to the classical unsteady aerodynamic functions are derived after a solution for the Fourier coefficients is obtained.
\subsubsection{Equivalent Theodorsen function}\label{sc:Formulation-EquivTheodorsen}
An equivalent Theodorsen function is constructed for a flexible membrane wing following the classical approach presented in \citet[p.~279]{Bisplinghoff_book1996}. An extension for Theodorsen's function is obtained by computing the membrane response to prescribed heave oscillations in the frequency domain:
\begin{equation}
\label{eq:TheodorsenMem_def}
C_m(k) = \frac{\hat{L}^C_h(k)}{\mathrm{i} k h_0\, L_{s_\alpha}} ,
\end{equation}
where $L_h^C$ is the circulatory lift due to heave oscillations, and $L_{s_\alpha}$ is the static (aeroelastic) membrane lift-curve slope.
The membrane circulatory lift is obtained by superposition between the circulatory lift due to membrane deformation (first term in \eqref{eq:Cl_d-Freq}) and the circulatory lift due to the unsteady flow (first term in \eqref{eq:CL_h-Freq}).
Substitution of the membrane circulatory lift expression into \eqref{eq:TheodorsenMem_def} yields a closed-form expression for the membrane Theodorsen function in the frequency domain,
\begin{equation}
\label{eq:TheodorsenMem_expFn2}
C_m(k) = \frac{2\pi}{C_{l_{s\alpha}}}\,C(k)\left[1+\hat{f}(k) \right].
\end{equation}
The static membrane lift slope,
\begin{equation}
\label{eq:CLa_mem-static}
C_{l_{s\alpha}} = 2\pi \left[1+ \frac{1}{2}\mathcal{F}_{s_1} - \frac{1}{2}\mathcal{F}_{s_0} \right] ,
\end{equation}
is established by direct integration of the static pressure load given by \cite{Nielsen1963}.
Here, $\mathcal{F}_{s_n}$ are the static membrane Fourier coefficients, normalised by the static angle of attack, which depend solely on the tension coefficient. Note that as $k \rightarrow 0$ the Fourier coefficients in the unsteady membrane solution converge to the static solution, ${\hat{\mathcal{F}}_n \xrightarrow[]{} \mathcal{F}_{s_n}}$, where $\hat{\mathcal{F}}_n = F_n/\alpha_0$ and $\alpha_0=\mathrm{i} k h_0$. Application of this limit to \eqref{eq:TheodorsenMem_expFn2} after substituting the leading term in the asymptotic expansion of $\hat{f}(k)$ in \eqref{eq:a2_1-fk_lim_k0_1} recovers
\begin{equation}
\label{eq:TheodorsenMem_lim_k0}
C_m(k) = C(k) + \textit{O}\left(k\right) \quad \mbox{as}\quad k\ttz .
\end{equation}
In other words, in the limit of small reduced frequencies the equivalent Theodorsen function converges to the standard Theodorsen function, as expected.
Finally, we note that the Fourier coefficients of the membrane solution converge to zero for a very stiff membrane as $C_T\rightarrow\infty$ at any reduced frequency other than the fluid-loaded resonance frequencies. Under these conditions, the auxiliary function $f(t)$ \eqref{eq:f_t} goes to zero, and the static membrane lift-curve slope converges on $2\pi$. Therefore, the equivalent Theodorsen function recovers the rigid plate function for $C_T\rightarrow\infty$, as expected. Appendix~\ref{a2_1:approx_k0} reports further details on this limit.
\subsubsection{Equivalent Wagner function}\label{sc:Formulation-EquivWagner}
We next derive the equivalent Wagner function for a flexible membrane wing following \citet[pp.~284-287]{Bisplinghoff_book1996}.
Note that the equivalent Theodorsen function derived in \S~\ref{sc:Formulation-EquivTheodorsen} enables the computation of the equivalent Wagner function in the time domain through \citep[][p. 285]{Bisplinghoff_book1996}
\begin{equation}
\label{eq:WagnerMem_def}
\Phi_m(t) = \frac{L_{\alpha_0}^C}{\alpha_0\, L_{s_\alpha}} = 1+\frac{2}{\pi}\int_0^\infty \frac{\Im\left\{C_m(k)\right\}}{k}\cos{kt}\,\mathrm{d} k , \quad t>0,
\end{equation}
where $L_{\alpha_0}^C$ is the circulatory lift of the membrane due to a step change in angle of attack, expressed in the time domain.
The above equation allows for the computation of the equivalent Wagner function from both the Laplace-domain solution (first equality) or the frequency-domain solution (second equality).
However, whilst the frequency-domain solution is more efficient when computing the Theodorsen function, the application of this solution to \eqref{eq:WagnerMem_def} requires a broad range of frequencies to obtain an accurate initial (high-frequency) response of the membrane lift. Therefore, the equivalent Wagner function is presented using the Laplace-domain solution.
Following the procedure presented in \S~\ref{sc:Formulation-EquivTheodorsen} for the equivalent Theodorsen function, a closed-form expression is derived for the equivalent Wagner function in the Laplace domain,
\begin{equation}
\label{eq:WagnerMem_expFn-LD}
\bar{\Phi}_m(s) = \frac{2\pi}{C_{l_{s\alpha}}}\, \bar{\Phi}(s)\left[1+s\bar{f}(s)\right] ,
\end{equation}
where $\bar{f}(s)$ is the Laplace transform of $f(t)$ from \eqref{eq:f_t}. In the time domain, the equivalent Wagner function becomes
\begin{equation}
\label{eq:WagnerMem_expFn-TD}
\Phi_m(t) = \frac{2\pi}{C_{l_{s\alpha}}}\left\{ \Phi(t)+\int_0^t \Phi(t-\tau) \dot{f}(\tau) \mathrm{d}\tau \right\}.
\end{equation}
We note that $\Phi_m(0) = {\pi}/{C_{l_{s\alpha}}}$ at time $t=0$, and the static membrane lift slope, $C_{l_{s\alpha}}$, is higher than the rigid plate lift slope due to the static membrane camber. Therefore, the initial value of the equivalent Wagner function is smaller than the classical value of $1/2$ predicted by the standard Wagner function for a rigid flat plate.
At long times $(t\rightarrow\infty)$, we can use the fact that $\dot{f}(\tau)$ rapidly converges to zero as the membrane profile converges to the appropriate static profile. Thus,
\begin{eqnarray}
\nonumber
\lim_{t\rightarrow\infty} \left\{\int_0^t \Phi(t-\tau) \dot{f}(\tau) \mathrm{d}\tau\right\} \cong&& \lim_{t\rightarrow\infty} \left\{\Phi(t)\int_0^t \dot{f}(\tau) \mathrm{d}\tau\right\} \\
\label{eq:WagnerMem_int_t_inf}
\cong&& \lim_{t\rightarrow\infty} \left\{\Phi(t)\, f(t)\right\} \cong \left(\frac{C_{l_{s\alpha}}}{2\pi}-1\right) \Phi(t)
\end{eqnarray}
and the equivalent Wagner function converges to the standard Wagner function, ${\Phi_m(t\rightarrow\infty)\cong\Phi(t\rightarrow\infty)=1}$, as expected.
For a very stiff membrane $(C_T\rightarrow\infty)$, the static membrane lift slope becomes $2\pi$, and the equivalent Wagner function converges to the rigid-plate Wagner function during the entire time response.
\subsubsection{Equivalent Sears function}\label{sc:Formulation-EquivSears}
Extension of the canonical modified Sears function is next derived for a flexible membrane wing in the frequency domain, following the classical approach presented in \citet[pp. 286-287]{Bisplinghoff_book1996}.
The equivalent Sears function is obtained by calculating the membrane response to oscillating gusts (\S~\ref{sc:Formulation-Sinus_gust}) and normalising the expression by the membrane static lift:
\begin{equation}
\label{eq:SearsMem_def}
S_m(k) = \frac{\hat{L}_{gs}(k)}{L_s},
\end{equation}
where $\hat{L}_{gs}(k)$ is the (complex) amplitude of the lift due to harmonic gust with reduced frequency $k$.
Substitution of the membrane lift expression \eqref{eq:CL_tot} into \eqref{eq:SearsMem_def}, superposing the lift due to membrane deformation \eqref{eq:Cl_d-Freq} with the lift due to the sinusoidal gust \eqref{eq:L_g-Time-sinus} in the frequency domain, leads to a closed-form expression for the membrane equivalent Sears function,
\begin{equation}
\label{eq:SearsMemExpr}
S_m (k) = \frac{2\pi}{C_{l_{s\alpha}}} \left\{ S(k) + C(k)\hat{f}(k) + \hat{g}(k) \right\} .
\end{equation}
The first term in \eqref{eq:SearsMemExpr} describes the rigid aerofoil lift response, the second term is the circulatory lift response due to membrane deformation, and the third term represents the non-circulatory lift response due to membrane deformation.
Note that the unsteady membrane solution to an encounter with sinusoidal gusts converges to the static membrane solution as $k\rightarrow 0$, namely $\hat{\mathcal{F}}_n \xrightarrow[k \rightarrow 0]{} \mathcal{F}_{s_n}$, where $\hat{\mathcal{F}}_n = \hat{F}_n/\alpha_0$.
Thus, for very low reduced frequencies the equivalent Sears function converges to
\begin{eqnarray}
\nonumber
S_m(k) \cong&& \frac{2\pi}{C_{l_{s\alpha}}} \left\{ S(k) + \left(\frac{C_{l_{s\alpha}}}{2\pi} -1\right) \left[ 1 + k\left[\mathrm{i}\ln{\frac{k}{2}} -\frac{\pi}{2}\right]\right] \right\}\\
\label{eq:SearsMem_lim_k0}
&&\mbox{} + \textit{O}\left( k\hat{\mathcal{F}}_3 ; k\hat{\mathcal{F}}_4 ; k^2\ln{k} ; k^2 \right) \quad \mbox{as\ }\quad k\ttz .
\end{eqnarray}
We note that for the limiting case of $C_T\rightarrow\infty$ the static membrane lift slope approaches $2\pi$ and the lift due to the membrane deformation converges to zero for $k\neq\omega_{r_n}$. Therefore, the equivalent Sears function converges to the standard modified Sears function for ${C_T\rightarrow\infty}$, as expected. Appendix~\ref{a2_1:approx_k0} reports further details on these low-frequency limits.
\subsubsection{Equivalent K\"{u}ssner function}\label{sc:Formulation-EquivKussner}
The aeroelastic membrane K\"{u}ssner function in the time domain,
\begin{equation}
\label{eq:KussnerMem_def}
\Psi_m(t) = \frac{L_{\mathit{seg}}}{L_s} = 1+\frac{2}{\pi}\int_0^\infty \frac{\Im\left\{S_m(k)\right\}}{k}\cos{kt}\,\mathrm{d} k , \quad t>0,
\end{equation}
follows from the equivalent Sears function determined in \S~\ref{sc:Formulation-EquivSears} using the procedure outlined in \S~\ref{sc:Formulation-EquivWagner} for the Wagner function \citep[e.g.,][p. 287]{Bisplinghoff_book1996}.
Here $L_{\mathit{seg}}$ is the time-domain membrane lift response to a sharp-edged gust. The above equation enables the computation of the equivalent K\"{u}ssner function from both the Laplace-domain solution (first equality) or the frequency-domain solution (second equality).
A closed-form expression for the equivalent K\"{u}ssner function is determined in the Laplace domain by superposing the rigid aerofoil indicial lift \eqref{eq:L_g-SEG} and the lift due to membrane deformation \eqref{eq:Cl_d-Lap}, with substitution of the resultant unsteady lift into the Laplace transform of \eqref{eq:KussnerMem_def}:
\begin{equation}
\label{eq:KussnerMem_exp-LD}
\bar{\Psi}_m(s) = \frac{2\pi}{C_{l_{s\alpha}}}\, \left\{\bar{g}(s) + \bar{\Psi}(s)+\bar{C}(s)\bar{f}(s)\right\} .
\end{equation}
The equivalent K\"{u}ssner function in the time domain is clearly
\begin{equation}
\label{eq:KussnerMem_exp-TD}
{\Psi}_m(t) = \frac{2\pi}{C_{l_{s\alpha}}}\, \left\{g(t)+\Psi(t)+\int_0^t \Phi(t-\tau) \dot{f}(\tau) \mathrm{d}\tau\right\} .
\end{equation}
We note that the initial value of the equivalent K\"{u}ssner function is ${\Psi_m(0) = 0}$. For $t\rightarrow\infty$ the equivalent K\"{u}ssner function asymptotically converges to unity, and for a very stiff membrane of $C_T\rightarrow\infty$ the rigid-plate K\"{u}ssner function is recovered.
\section{Results and discussion}\label{sec:results}
The membrane response to prescribed chord motion or an incoming gust is derived in both the time domain (via inverse Laplace transform) and the frequency domain, the latter of which is used to study the steady-state response to harmonic motions or sinusoidal gusts.
Results are shown for four canonical cases: harmonic heave oscillations, step change in angle of attack, sinusoidal gust and sharp-edged gust, for which extensions of the four respective classical unsteady lift functions are presented for a flexible membrane wing.
We begin by studying the lift and dynamic response of a nominal membrane of $\mu=1$ and $C_T=2.5$, followed by analysis of the role of each of the membrane parameters.
\subsection{Prescribed motion}\label{sec:results-pres}
The membrane response to prescribed chord motion is derived for two canonical problems: harmonic heave oscillations, from which Theodorsen's function is derived, and step angle of attack (also known as Wagner's problem).
In both cases the membrane is free to deform around the chord-line, which adheres to the prescribed motion.
Extensions of the classical Theodorsen and Wagner functions are presented for flexible membrane wings, along with a discussion on the membrane dynamic response to these unsteady flow conditions and the role of the membrane parameters $\left(\mu, C_T\right)$ in its aerodynamic performance.
\subsubsection{Harmonic heave oscillations}\label{sec:results-pres-heave}
To assess the membrane wing response to prescribed oscillations in heave, we compute first the membrane amplitude at various reduced frequencies of oscillation, $k$, for various tension coefficients and two mass ratios (figure~\ref{fig:HeaveSSMem_CT_vs_k}). In addition, the resonance frequencies of the fluid-loaded membrane system ($\omega_{r_1}, \omega_{r_2},$ etc.) are computed from the homogeneous system of \eqref{eq:MatrixEqUS-Freq}, following the method of \cite{Kornecki1976}. The parametric dependence of the fluid-loaded resonance frequencies on $C_T$ and $k$ is illustrated with dashed red lines in figure~\ref{fig:HeaveSSMem_CT_vs_k}.
The left column of figure~\ref{fig:HeaveSSMem_CT_vs_k} presents maps of the resulting maximal membrane amplitude, obtained for $\mu=1$ (upper row) and $\mu=18$ (lower row) for varying tension coefficient and reduced frequency.
As expected, significant amplitudes of oscillation occur for frequencies near the resonance frequencies of the fluid-loaded membrane. In the heavy membrane case, $\mu=18$, where the mass ratio is encroaching upon the flutter instability threshold, predicted by \cite{Tiomkin2017} at $\mu\ge18.8$ for $C_T=2$, the peaks in the maximal amplitude map are more concentrated, with a significantly increased amplitude along the second fluid-loaded resonance frequency. Note that these narrow peaks in figure~\ref{fig:HeaveSSMem_CT_vs_k-mu18_max_amp_map} prevented the addition of the resonance frequencies to this plot, as these lines cover the peaks entirely; the relevant fluid-loaded resonance frequencies are plotted in figure~\ref{fig:HeaveSSMem_CT_vs_k-mu18} for reference.
The right column of figure~\ref{fig:HeaveSSMem_CT_vs_k} shows the membrane amplitude profiles obtained along the fluid-loaded resonance frequencies, with background colour used to indicate the maximum amplitude value. Dotted black lines indicate the in vacuo natural frequencies, $k_1, k_2,$ etc., that are compared against the fluid-loaded resonance frequencies.
Significant differences between the in vacuo and the fluid-loaded membrane resonance frequencies are obtained for the lowest mass ratio, $\mu=1$, in figure~\ref{fig:HeaveSSMem_CT_vs_k-mu1}.
This difference in frequencies evokes oscillations with membrane amplitude profiles that are noticeably different from the membrane in vacuo modes.
In addition, the membrane amplitude increases with reduced frequency, following the behaviour of the excitation term in \eqref{eq:DCp_h-Freq} whose amplitude increases monotonically with $k$.
As the mass ratio increases in figure~\ref{fig:HeaveSSMem_CT_vs_k-mu18}, the gap between the resonance and the structural frequencies diminishes, and second-mode oscillations become dominant; this is the first dynamically unstable membrane mode \citep[][]{Nielsen1963,Tiomkin2017}. In this case the membrane amplitude increases as the tension decreases (although the resonance frequency also decreases), which is typical of the membrane-wing behaviour on the verge of instability \citep{Tiomkin2017,Mavroyiakoumou2020}.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_max_amp_heave_ss_mu_1_CT_vs_k_fine_100_b_w_wr.eps}
\caption{}
\label{fig:HeaveSSMem_CT_vs_k-mu1_max_amp_map}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_shape_heave_ss_mu_1_CT_vs_k_with_wn_white_d.eps}
\caption{}
\label{fig:HeaveSSMem_CT_vs_k-mu1}
\end{subfigure}
\\
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_max_amp_heave_ss_mu_18_CT_vs_k_fine_100.eps}
\caption{}
\label{fig:HeaveSSMem_CT_vs_k-mu18_max_amp_map}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_shape_heave_ss_mu_18_CT_vs_k_with_wn_white_d.eps}
\caption{}
\label{fig:HeaveSSMem_CT_vs_k-mu18}
\end{subfigure}
\caption{Membrane amplitude in response to heave oscillations: (a), (c) maximum amplitude maps obtained for various values of $C_T$ and $k$ for $\mu=1$ and $\mu=18$, respectively. (b), (d) present the membrane amplitude profiles obtained at the fluid-loaded resonance frequencies for $\mu=1$ and $\mu=18$, respectively. Background colour depicts maximum amplitude of membrane deformation, normalised by the heave motion amplitude, $h_0$. Red dashed lines describe the first, second and third resonance frequencies of the fluid-loaded membrane, and black dotted lines present membrane in vacuo natural frequencies. A large value of the mass ratio, $\mu=18$, is chosen to assess the membrane response near the onset of flutter, which \cite{Tiomkin2017} predicts to occur for $\mu\ge18.8$ when $C_T=2$.}
\label{fig:HeaveSSMem_CT_vs_k}
\end{center}
\end{figure}
The membrane lift response to heave oscillations is next evaluated by comparing the membrane equivalent Theodorsen function \eqref{eq:TheodorsenMem_expFn2} to the standard Theodorsen function of a rigid flat plate.
Figure~\ref{fig:HeaveMem_mu1CT2_5-Lift} illustrates this comparison for the nominal membrane of $\mu=1$ and $C_T=2.5$ as an Argand diagram (figure~\ref{fig:HeaveMem_mu1CT2_5-Theodorsen}) and in terms of modulus and phase (figure~\ref{fig:HeaveMem_mu1CT2_5-Theodorsen_abs_phase}).
For low reduced frequencies, the membrane lift response follows the general behaviour of the rigid plate response, with reduced amplitude and increased phase lag.
As the reduced frequency of heave oscillations increases,
at some point (typically for $k$ smaller than the first resonance frequency) the membrane equivalent Theodorsen function changes its direction abruptly, where the unsteady lift response amplitude increases, rather than converging to zero as it would for rigid aerofoils. We mark this inflection point by reduced frequency $k_{{inv}_1}$. With a further increase in $k$ beyond $k_{{inv}_1}$, a circular path is obtained in the complex plain plot until the next inflection point is reached at $k=k_{inv_2}$, and so on. In figure~\ref{fig:HeaveMem_mu1CT2_5-Theodorsen}, we present results for reduced frequencies up to the second in-vacuo natural frequency, $k_2\cong3.5$, for the sake of clarity. Each of these circular arcs contains one of the system's resonance frequencies for which a local maximum is observed in the lift response amplitude (figure~\ref{fig:HeaveMem_mu1CT2_5-Theodorsen_abs_phase}).
A region of special interest is revealed around the first resonance frequency, for $0.65\le k\le 0.96$, where the membrane aerofoil demonstrates substantial benefits in lift magnitude over a rigid aerofoil. Interestingly, oscillations with lower or higher reduced frequency (in the examined range of $k\le 3.5$) result in substantial deficit in unsteady lift amplitude.
In terms of Strouhal number, the beneficial reduced frequency regime is translated to $0.2<\mathit{St}<0.3$, which agrees well with the flapping frequency of birds and bats, known to correspond to a Strouhal number between $0.2$ and $0.4$, as reported by \cite{Shyy2016}.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Theodorsen_CT2_5_mu1_w_wr_marks_black_k_arrow.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-Theodorsen}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Theodorsen_abs_phase_vs_k_Mu1_CT2_5_with_wr_marks_black_a.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-Theodorsen_abs_phase}
\end{subfigure}
\caption{Membrane lift response to heave oscillations, obtained for $C_T=2.5, \mu=1$ in terms of the membrane equivalent Theodorsen function: (a) Argand diagram; (b) modulus and phase. Frequencies of inflection points are denoted with black circles ($k_{{inv}_1}$) and crosses ($k_{{inv}_1}$), and resonance frequencies are denoted with red pluses ($\omega_{{r}_1}$) and magenta crosses ($\omega_{{r}_2}$).}
\label{fig:HeaveMem_mu1CT2_5-Lift}
\end{center}
\end{figure}
To further examine the origin of the circular paths in the membrane Theodorsen function, we recall that the equivalent Theodorsen function \eqref{eq:TheodorsenMem_expFn2} is in fact a product of the standard Theodorsen function and a function of the membrane Fourier coefficients. Figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik} presents the evolution of the first two Fourier coefficients (normalised by $\alpha_0=\mathrm{i} k h_0$) with varying reduced frequency, obtained for the nominal membrane case. These normalised Fourier coefficients are the most dominant coefficients in the Fourier series used to describe the membrane slope \eqref{eq:SailSlope-Fourier}, and are plotted as an Argand diagram (figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik_cmplx}) and in terms of its modulus and phase (figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik_abs_phase}).
Note that the unsteady solution recovers the static aeroelastic membrane results for $k\rightarrow 0$, as expected, where the static results are marked by pentagrams in the Argand diagram.
As the reduced frequency is increased from the static limit, the amplitudes of both of the normalised Fourier coefficients decrease at first, yielding a smaller amplitude of the oscillating membrane shape, and then increase as the reduced frequency approaches the first fluid-loaded resonance frequency. The first inflection point in the equivalent Theodorsen function corresponds to the first local minimum of $|\hat{\mathcal{F}}_1|$, which is followed by a circular path in the complex plane plots of all coefficients (figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik_cmplx}).
This entire frequency regime, in which the first circle appears in the Fourier coefficients, is dominated by the membrane's first~mode, as is evident by the dominance of the first Fourier coefficient in figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik_abs_phase}.
This dominance is maximal near the first resonance frequency and diminishes as $k$ increases, which continues until the second resonance frequency is approached and the second mode coefficient becomes dominant. The frequency at which $|\hat{\mathcal{F}}_2|$ crosses $|\hat{\mathcal{F}}_1|$ is in fact $k_{{inv}_2}$, which marks the transition from the first circle to the second circle in the equivalent Theodorsen function in figure~\ref{fig:HeaveMem_mu1CT2_5-Theodorsen}. This behaviour, in which the inflection points are identified by a switch of dominance between the membrane modes, also continues to higher modes as the frequency is further increased, but is not shown here for the sake of clarity and brevity.
Thus, it can be concluded that the circular arcs in the equivalent Theodorsen function are due to the membrane dynamic response, where each circle is related to a different dominant mode, and the inflection points between circles occur at the intersection between the modulus functions of two consecutive normalised Fourier coefficients.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Fn_ik_separate_complex_Heave_FD_nom_black_a_latex_arrow_w_zoom_c_axes.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-Fn_ik_cmplx}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Fn_ik_abs_phase_Heave_FD_nom_black_a_latex_b.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-Fn_ik_abs_phase}
\end{subfigure}
\caption{The first two (most dominant) complex-valued Fourier coefficients, normalised by $\alpha_0=\mathrm{i} k h_0$, for a nominal membrane of $C_T=2.5, \mu=1$ undergoing harmonic heave oscillations of reduced frequency $k$: (a) Argand diagram; (b) modulus and phase.
Static solutions are denoted by pentagram markers and recovered by the unsteady results as $k\rightarrow 0$. First inflection point is marked with black circles and identified by the first local minimum of $|\hat{\mathcal{F}}_1|$. Second inflection point is denoted by black crosses, identified by an intersection between $|\hat{\mathcal{F}}_1|$ and $|\hat{\mathcal{F}}_2|$. The insets in (a) introduce a zoom-in on the first inflection point region for illuminating the trend of both functions as $k$ passes $k_{\mathit{inv}_1}$.}
\label{fig:HeaveMem_mu1CT2_5-Fn_ik}
\end{center}
\end{figure}
Figure~\ref{fig:HeaveMem_mu1CT2_5-Mem} presents the membrane amplitude profiles computed for a nominal membrane undergoing heave oscillations of varying reduced frequency, where the profile represents the amplitude of oscillation at every point along the membrane chord.
A contour plot of the amplitude profiles is presented in figure~\ref{fig:HeaveMem_mu1CT2_5-MemContourf} for varying reduced frequency, $k$, with black dashed and dotted lines denoting the resonance and inflection point frequencies, respectively.
Small deformations relative to the heave amplitude are obtained along the entire frequency range, except in the vicinity of the system's (fluid-loaded) resonance frequencies, for which large-amplitude profiles are observed with a shape similar to the membrane structural modes.
This observation suggests a close coupling between the unsteady lift amplitude and the membrane amplitude in response to harmonic heave oscillations. Furthermore, a favorable lift is clearly achievable only for odd resonance frequencies ($\omega_{r_1}, \omega_{r_3}, ... $), which correspond to shapes that are symmetric around the mid-chord point, rather than anti-symmetric in the even-mode cases, as illustrated by comparing the lift response in figure~\ref{fig:HeaveMem_mu1CT2_5-Theodorsen_abs_phase} with the membrane amplitude profiles in figure~\ref{fig:HeaveMem_mu1CT2_5-MemContourf}.
Figures~\ref{fig:HeaveMem_mu1CT2_5-Mem_kinv1} and \ref{fig:HeaveMem_mu1CT2_5-Mem_kinv2} provides a more detailed view of the membrane amplitude profiles obtained for reduced frequencies near $k_{{inv}_1}$ and $k_{{inv}_2}$, respectively.
In accordance with the normalised Fourier coefficients (figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik}), we note that small-amplitude profiles are obtained for small values of reduced frequency (figure~\ref{fig:HeaveMem_mu1CT2_5-Mem_kinv1}). These membrane shapes are convex, with a maximum camber point at the fore section of the aerofoil, in accordance with the static membrane solution \citep{Nielsen1963}. As the reduced frequency increases, the amplitude of the profile decreases until for $k_{{inv}_1}$ an inflection point appears in the membrane profile and drastic changes in the profile shape are evoked with further increase in reduced frequency.
These deformations indicate the excitation of the membrane structural modes as the reduced frequency approaches the system's first resonance frequency.
For larger reduced frequencies near the second inflection point, $k_{{inv}_2}$, figure~\ref{fig:HeaveMem_mu1CT2_5-Mem_kinv2} shows that significant membrane oscillations are evoked, in which the inflection point ($k_{{inv}_2}$) represents the shift in the membrane amplitude profile, from the fully convex shape obtained for the first resonance frequency to the second mode shape obtained for the second resonance frequency. This shift is identified by an inflection point that appears near the leading edge of the membrane profile for $k\ge k_{{inv}_2}$, after which the second mode of the membrane becomes most dominant.
This result is in accordance with the behaviour of the Fourier coefficients, presented in figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik}, affirming the conclusion that the inflection points in the complex plane plot of the equivalent Theodorsen function mark the shift in dominance between consecutive membrane modes.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_deformation_contourf_Mu1_CT2_5_with_wr_kinv_w_leg_a.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-MemContourf}
\end{subfigure}
\\
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_deformation_around_k_inv1_Mu1_CT2_5_a.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-Mem_kinv1}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_deformation_around_k_inv2_Mu1_CT2_5_a.eps}
\caption{}
\label{fig:HeaveMem_mu1CT2_5-Mem_kinv2}
\end{subfigure}
\caption{Membrane dynamic response to harmonic heave oscillations, obtained for a nominal membrane of $C_T=2.5, \mu=1$: (a) amplitude profiles computed for varying values of reduced frequency. Resonance frequencies are denoted with dashed black lines, and inflection point frequencies are marked with dotted black lines. (b) and (c) present the membrane amplitude profiles obtained for reduced frequencies around the first and second inflection points, respectively.}
\label{fig:HeaveMem_mu1CT2_5-Mem}
\end{center}
\end{figure}
Figures~\ref{fig:TheodorsenMemLight-CTeffect} and \ref{fig:TheodorsenMemLight-Mueffect} illustrate the separate effects of the tension coefficient and the membrane mass ratio on the membrane Theodorsen function. Tension coefficients between $2.5$ and $4$, and mass ratios between $0.5$ and $2.5$ are chosen to represent realistic membrane wings \citep[e.g.,][]{Rojratsirikul2010,Tiomkin2021}, while still remaining in the membrane stable regime, as the membrane loses stability via divergence for {$C_T<1.73$} and loses stability via flutter only in the case of heavy membranes of {$\mu>18.8$} \citep[see][for details]{Tiomkin2017}.
Results are presented for reduced frequencies up to the second in vacuo frequency $(k=k_2)$ for the sake of clarity.
The membrane stiffness is strongly influenced by the tension coefficient, and figure~\ref{fig:TheodorsenMemLight-CTeffect_amp_phase} shows that the membrane lift response to low-frequency oscillations approaches the rigid plate response with increasing tension coefficient, as expected.
This result is further validated by examining an extreme case of $C_T=50$, presented in figure~\ref{fig:TheodorsenMemLight-CTeffect_amp_phase} with a dashed red line.
For this high tension coefficient the membrane is practically rigid, and indeed the resulting equivalent Theodorsen function follows closely the standard Theodorsen function for a wide range of frequencies up to about $k=2$, where differences in the amplitudes arise due to parametric proximity to the system's first resonance frequency. Because the system's resonance frequencies increase with $C_T$, the range of frequencies for which the equivalent Theodorsen function follows the standard Theodorsen function increases with $C_T$ as well, and the first resonance circle occurs at a larger reduced frequency (i.e., $k_{{inv}_1}$ increases).
In addition, the circle diameter increases with $C_T$ for $\mu=1$, indicating that an increase in tension coefficient leads locally to an increased amplitude of the unsteady lift response near the resonance frequency, as the membrane oscillation amplitude also increases (figure~\ref{fig:HeaveSSMem_CT_vs_k-mu1}).
Thus, the beneficial region, where substantial increase in unsteady lift amplitude is obtained, is clearly controlled by the tension coefficient, suggesting the possibility of optimizing the flapping wing performance by controlling the tension along the membrane.
\begin{figure}
\begin{center}
\begin{subfigure}[h!]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Theodorsen_Mu1_CT2to4_upto_k2_b_k_arrow.eps}
\caption{}
\label{fig:TheodorsenMemLight-CTeffect_complx}
\end{subfigure}
\quad
\begin{subfigure}[h!]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Theodorsen_abs_phase_vs_k_Mu1_CT2to4_b_CT50.eps}
\caption{}
\label{fig:TheodorsenMemLight-CTeffect_amp_phase}
\end{subfigure}
\caption{Effect of membrane tension coefficient on the membrane equivalent Theodorsen function for $\mu=1$: (a) Argand diagram; (b) modulus and phase. The rigid-plate Theodorsen function is presented with a black line for comparison with the membrane equivalent Theodorsen function. The solution for a very large tension coefficient of $C_T=50$ is presented with a red dashed line in (b) and indicates asymptotic convergence of the unsteady solution to the rigid plate solution as $C_T \rightarrow\infty$.}
\label{fig:TheodorsenMemLight-CTeffect}
\end{center}
\end{figure}
Analysis of the mass ratio effect on the membrane response to prescribed heave oscillations (figure~\ref{fig:TheodorsenMemLight-Mueffect}) shows that at small reduced frequencies the mass ratio has practically no effect on the lift response, as expected.
For higher reduced frequencies the first resonance circle appears earlier (at lower $k$) as the mass ratio is increased, in accordance with the decrease in resonance frequency, while the amplitude of the lift response is practically unaffected.
The beneficial region is also controlled by the mass ratio, as the peak in the lift amplitude follows the movement of the resonance frequency. However, in practical applications, this variable is often harder to control than the tension coefficient.
\begin{figure}
\begin{center}
\begin{subfigure}[h!]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Theodorsen_CT2_5_mu0_5to2_5_a_k_arrow.eps}
\caption{}
\label{fig:TheodorsenMemLight-Mueffect_complx}
\end{subfigure}
\quad
\begin{subfigure}[h!]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Theodorsen_abs_phase_vs_k_Mu0_5to2_5_CT2_5_a.eps}
\caption{}
\label{fig:TheodorsenMemLight-Mueffect_amp_phase}
\end{subfigure}
\caption{Effect of membrane mass ratio on the membrane equivalent Theodorsen function for $C_T=2.5$: (a) Argand diagram; (b)modulus and phase (b). The rigid-plate Theodorsen function is presented with a black line for comparison with the membrane equivalent Theodorsen function.}
\label{fig:TheodorsenMemLight-Mueffect}
\end{center}
\end{figure}
\subsubsection{Step angle of attack}\label{sec:results-pres-AoA}
We next examine the response of the nominal membrane to a step in angle of attack, in terms of its dynamic and aerodynamic response (figure~\ref{fig:MemAoAstepRes_mu1_CT2_5}). The membrane dynamic response begins with an initially taut profile, followed by membrane oscillations as a result of the abrupt change in angle of attack. These oscillations decrease in amplitude with time until a steady-state profile is obtained that is identical to the respective static solution (figure~\ref{fig:MemAoAstepRes_mu1_CT2_5-Mem}). The resulting lift coefficient history (figure~\ref{fig:MemAoAstepRes_mu1_CT2_5-CL}) presents a similar trend to the membrane deformation history and suggests a close coupling between the two.
In addition, when comparing the nominal membrane lift history in response to a step in angle of attack, $C_{l_m}$, with the rigid plate response, $C_{l_f}$, we see that the membrane wing achieves a larger lift across almost the entire response, converging to a value more than double that of the rigid-plate lift, due to membrane camber.
Interestingly, during the initial transient stage of the response (for $t<1.4$) the membrane lift is lower than the rigid plate lift.
Namely, the membrane deformation due to the abrupt change in angle of attack produces negative lift, as evident by the plot of $C_{l_d}^C$ and $C_{l_d}^{\mathit{NC}}$ in figure~\ref{fig:MemAoAstepRes_mu1_CT2_5-CL}, which describe the circulatory and non-circulatory terms in $C_{l_d}$ \eqref{eq:Cl_d}.
As the membrane is initially still and taut, the membrane surface accelerates in response to the sudden change in flow conditions, leading to a negative apparent mass lift at $t=0$ and zero circulatory lift.
The membrane inflates with time and its acceleration reduces, which yields a decrease in the non-circulatory lift magnitude.
The initial acceleration subsequently yields a negative circulatory lift, but it also causes an increase in the membrane velocity, which increases the circulatory lift.
This trend continues until at time $t=1.4$ the circulatory lift is able to compensate the lift deficit due to $C_{l_d}^{\mathit{NC}}$ and the membrane lift exceeds the rigid plate lift. For longer times the membrane deformation yields a higher circulatory lift that contributes to the further increase in the total membrane lift, as it proceeds to converge to the static membrane lift coefficient, $C_{l_s}$.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_AoA_step_re_surf_CT_2_5_mu_1.eps}
\caption{}
\label{fig:MemAoAstepRes_mu1_CT2_5-Mem}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/CL_vs_t_AoA_step_mem_nom_compare_rigid_w_terms_a_zoom_latex_b.eps}
\caption{}
\label{fig:MemAoAstepRes_mu1_CT2_5-CL}
\end{subfigure}
\caption{Membrane dynamic response (a) and lift response (b) to a step in angle of attack, obtained for a nominal membrane of $C_T=2.5$ and $\mu=1$. Black line in (a) and dashed black line in (b) denote the static solution \citep[][]{Nielsen1963}. The membrane unsteady lift coefficient, $C_{l_m}$, is computed by superposition between the rigid plate indicial lift, $C_{l_f}$, and the lift due to membrane deformation, $C_{l_d}$, which is composed of the circulatory and non-circulatory terms, $C_{l_d}^C$ and $C_{l_d}^\mathit{NC}$, respectively.}
\label{fig:MemAoAstepRes_mu1_CT2_5}
\end{center}
\end{figure}
Figure~\ref{fig:WagnerMem} illustrates the separate effects of the membrane tension coefficient and mass ratio on the membrane lift response to a step in angle of attack in terms of the equivalent Wagner function, and compares it to the standard Wagner function for rigid aerofoils.
The equivalent Wagner functions were computed in the Laplace domain \eqref{eq:WagnerMem_expFn-LD} and transformed to the time domain via numerical Laplace inversion.
In general, for all of the examined cases, the membrane lift response is slower than the rigid plate response.
However, we recall that the static membrane lift-curve slope is determined by the tension coefficient \eqref{eq:CLa_mem-static}, and can be significantly larger than the rigid-plate lift slope due to aeroelastic camber.
Thus, for example, while for $C_T=2,\mu=1$, at time $t=100$ the equivalent Wagner function reaches only $93.8\%$ of its steady-state solution, compared to $99\%$ for a rigid flat plate, the steady-state lift in this case is substantially higher for the membrane wing ($C_{l_{s\alpha}}\cong28$), yielding a lift that is more than $4$ times larger than the rigid plate lift.
As the tension along the membrane is increased, the membrane Wagner function approaches the classical solution for a rigid plate, as expected (figure~\ref{fig:WagnerMem-Mu1}).
In addition, the initial value of the equivalent Wagner function is significantly lower than its rigid value, as it recovers the result predicted in \S~\ref{sc:Formulation-EquivWagner}, $\Phi_m(0)=\pi/C_{l_{s\alpha}}$, marked by pentagrams in figure~\ref{fig:WagnerMem}.
The effect of the mass ratio on the equivalent Wagner function is much less pronounced than the tension coefficient effect (figure~\ref{fig:WagnerMem-CT2_5}) and is discernible only for short time periods (i.e., high frequencies) when inertial effects are important. For long time durations, the lift responses are equivalent for any practical use.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Wagner_LD_CT2to4_Mu1_w_limt0_zoom.eps}
\caption{$\mu=1$}
\label{fig:WagnerMem-Mu1}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_Wagner_LD_CT2_5_Mu0_5to2_5_w_zoom.eps}
\caption{$C_T=2.5$}
\label{fig:WagnerMem-CT2_5}
\end{subfigure}
\caption{Effect of the tension coefficient (a) and mass ratio (b) on the membrane equivalent Wagner function, as compared to the standard Wagner function (black line). Results are obtained via Laplace-domain solution and are verified against the expected initial values marked with pentagrams.}
\label{fig:WagnerMem}
\end{center}
\end{figure}
\subsection{Gust response}\label{sec:results-gust}
The membrane response to encounters with transverse gusts is now studied for two canonical cases: (i) a sinusoidal gust, and (ii) a sharp-edged gust. The sinusoidal and sharp-edged gusts produce unsteady lift responses described by the equivalent Sears and K\"{u}ssner functions, respectively. These extensions of the classical Sears and K\"{u}ssner functions are presented for flexible membrane wings, along with discussion on the membrane dynamic response to these unsteady flow conditions and the role of the membrane parameters $\left(\mu, C_T\right)$ in its aerodynamic performance. We note that while the sharp-edged gust has no physical meaning by itself, it is a very useful tool when predicting the aerofoil's response to an arbitrary (small amplitude) transverse gust by appeal to convolution theory \citep[][p. 288]{Bisplinghoff_book1996}.
\subsubsection{Sinusoidal gust}\label{sec:results-gust-sin}
The membrane lift and dynamic response to an encounter with a sinusoidal gust is controlled by the membrane tension coefficient, mass ratio, and the gust reduced frequency.
The response of a nominal membrane wing to sinusoidal gusts of varying reduced frequency is analyzed first, followed by a separate analysis of the effect of each of the membrane parameters on the resulting unsteady lift and dynamic response of the membrane.
Figure~\ref{fig:SearsCompMu1CT2_5} presents the lift response of the nominal membrane to sinusoidal gusts in terms of the equivalent Sears function \eqref{eq:SearsMemExpr}.
This equivalent Sears function is compared against the classical modified Sears function for a rigid flat plate in figure~\ref{fig:SearsCompMu1CT2_5-complx} using an Argand diagram and in figure~\ref{fig:SearsCompMu1CT2_5-amp2_phase} in terms of the squared magnitude and the phase.
The choice of a squared amplitude plot rather than a modulus plot follows the convention originated by \cite{Drischler1956} for harmonic gusts.
For low reduced frequencies of $k<k_{{inv}_1}\cong0.41$, the membrane equivalent Sears function closely follows the classical modified Sears function, with a slightly decreased amplitude and increased phase lag. However, at the point of inflection ($k=k_{{inv}_1}$) the lift amplitude begins to increase significantly with reduced frequency, creating a circular path in the complex plane, in a manner similar to the equivalent Theodorsen function (cf. figure~\ref{fig:HeaveMem_mu1CT2_5-Lift}). The lift amplitude increases beyond the rigid plate response in the vicinity of the first fluid-loaded resonance frequency ($k=\omega_{r_1}$), which is followed by a sharp decrease in amplitude at higher reduced frequencies. This behaviour suggests the existence of a reduced frequency region for which gust mitigation is achievable using flexible membrane wings, while also revealing a range of frequencies (near the first resonance frequency) for which membrane flexibility could have adverse results.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Sears_fun_mem_nom_complx_w_wr_kinv_from_Fn_k_short_w_leg_k_arrow.eps}
\caption{}
\label{fig:SearsCompMu1CT2_5-complx}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Sears_fun_mem_nom_abs2_phase_w_wr_kinv_from_Fn_k_short_a.eps}
\caption{}
\label{fig:SearsCompMu1CT2_5-amp2_phase}
\end{subfigure}
\caption{Membrane lift response to sinusoidal gusts of various frequencies, obtained for a nominal membrane of $C_T=2.5, \mu=1$ in terms of the membrane equivalent Sears function: (a) Argand diagram, and (b) squared modulus and phase. Frequencies of the inflection points are denoted with black circles ($k_{{inv}_1}$) and crosses ($k_{{inv}_2}$), and resonance frequencies are denoted with red pluses ($\omega_{{r}_1}$) and magenta cross signs ($\omega_{{r}_2}$).}
\label{fig:SearsCompMu1CT2_5}
\end{center}
\end{figure}
To further study the origin of the inflection points in the complex plane plot of the equivalent Sears function, we recall that the equivalent Sears function \eqref{eq:SearsMemExpr} depends on the standard modified Sears function, the standard Theodorsen function, and the Fourier coefficients used to describe the membrane deformation.
Figure~\ref{fig:SinGustMem_mu1CT2_5-Fn} presents the behaviour of the first two normalized Fourier coefficients obtained for the nominal membrane in response to sinusoidal gusts of varying reduced frequency. These Fourier coefficients are the most dominant coefficients in the membrane dynamic response for the range of reduced frequencies examined.
For $k\rightarrow 0$, the Fourier coefficients converge to the appropriate static solution, marked by pentagrams in figure~\ref{fig:SinGustMem_mu1CT2_5-Fn_cmplx}.
As the reduced frequency increases, the changes in both Fourier coefficients resemble the behaviours of the normalised coefficients in the heaving membrane case (cf. figure~\ref{fig:HeaveMem_mu1CT2_5-Fn_ik}), while the amplification at the resonance frequency is less pronounced for the sinusoidal gust response.
The first and second inflection points in the equivalent Sears function are identified here, similarly to the heaving membrane case, by locating the first local minimum in $|\hat{\mathcal{F}}_1|$, and the crossing between $|\hat{\mathcal{F}}_1|$ and $|\hat{\mathcal{F}}_2|$ as the dominance transfers from the first membrane mode to the second mode around $k_{{inv}_2}$. Namely, the circles in the complex plane plot of the equivalent Sears function are due to the membrane dynamic response, just like in the equivalent Theodorsen function, where each circle corresponds to a different dominant mode in the membrane oscillations. However, both the location of these circles in the frequency domain and the magnitude of the lift amplification differ significantly from the heaving membrane case, as the gust encounter introduces different flow mechanisms due to shedding of the unsteady flow conditions along the aerofoil.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/F1_F2_complx_mem_nom_SinGust_w_kinv_a_static_latex_k_arrow_b.eps}
\caption{}
\label{fig:SinGustMem_mu1CT2_5-Fn_cmplx}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/F1_F2_abs_phase_mem_nom_SinGust_w_kinv_a_latex_b.eps}
\caption{}
\label{fig:SinGustMem_mu1CT2_5-Fn_abs_phase}
\end{subfigure}
\caption{The first two (most dominant) complex-valued normalized Fourier coefficients obtained for a nominal membrane of $C_T=2.5, \mu=1$ that encounters a sinusoidal gust of reduced frequency $k$: (a) Argand diagram; (b) modulus and phase.
Static solutions are denoted by pentagram markers and recovered by the unsteady results as $k\rightarrow 0$. First inflection point is marked with black circles and identified by the first local minimum of $|\hat{\mathcal{F}}_1|$. Second inflection point is denoted by black crosses, identified by an intersection between $|\hat{\mathcal{F}}_1|$ and $|\hat{\mathcal{F}}_2|$.}
\label{fig:SinGustMem_mu1CT2_5-Fn}
\end{center}
\end{figure}
Figure~\ref{fig:MemSinGust_mu1CT2_5-MemContourf} presents the membrane amplitude profiles computed during steady state oscillations of a nominal membrane that encounters sinusoidal gusts of various reduced frequencies. For $k\to0$ a convex amplitude profile is obtained, in accordance with the static membrane solution. Then, as the reduced frequency is increased, the membrane amplitude profile is flattened until $k$ approaches the first resonance frequency, for which a large maximum amplitude is obtained. Figure~\ref{fig:MemSinGust_mu1CT2_5-Mem_kinv1} illustrates the membrane amplitude profiles computed for sinusoidal gusts at reduced frequencies near the first inflection point frequency, $k_{{inv}_1}$. For reduced frequencies smaller than $k_{{inv}_1}$ a convex shape with a maximum camber point at the fore part of the aerofoil is obtained. As the reduced frequency increases to $k_{{inv}_1}$ the maximum amplitude of the membrane decreases, and the maximum camber point slowly approaches the mid-chord location. A further increase in the reduced frequency beyond $k_{{inv}_1}$ yields a sudden shift of the maximum camber point downstream, as the membrane profile bears a close resemblance to the first unstable eigenshape of the membrane in cases of divergence instability, as reported by \cite{Sygulski2007} and \cite{Tiomkin2017}. This shift in the membrane amplitude profile signals the excitation of the membrane structural modes as the first fluid-loaded resonance frequency is approached.
As we further increase the reduced frequency to the vicinity of the second inflection point frequency, $k_{{inv}_2}$, in figure~\ref{fig:MemSinGust_mu1CT2_5-Mem_kinv2} we see a clear change in the membrane amplitude profile from a shape that is dominated by the first structural mode to a shape in which the second structural mode is most dominant, as supported by the Fourier coefficients in figure~\ref{fig:SinGustMem_mu1CT2_5-Fn}.
Thus, any inflection point in the complex plane plot of the equivalent Sears function is related to a shift in dominance between two consecutive membrane mode shapes. As the gust frequency is increased, higher membrane modes become dominant.
We further note that for the frequency regime in which membrane oscillations amplify the lift response (around $k=\omega_{r_1}$), large amplitude deformations are obtained with a convex amplitude profile. For higher reduced frequencies, the membrane amplitude profile is no longer convex due to the appearance of additional nodal points along the profile, and smaller maximum camber is obtained. Membrane oscillations with these amplitude profiles attenuate the aerofoil's lift response, presenting a reduced lift amplitude relative to the rigid plate lift (figure~\ref{fig:SearsCompMu1CT2_5-amp2_phase}).
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_mode_vs_k_sin_gust_FD_mu1_CT_2_5_w_wr_k_inv_w_leg_a.eps}
\caption{}
\label{fig:MemSinGust_mu1CT2_5-MemContourf}
\end{subfigure}
\\
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_mode_around_kinv1_sin_gust_FD_mu1_CT_2_5_a.eps}
\caption{}
\label{fig:MemSinGust_mu1CT2_5-Mem_kinv1}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_mode_around_kinv2_sin_gust_FD_mu1_CT_2_5_a.eps}
\caption{}
\label{fig:MemSinGust_mu1CT2_5-Mem_kinv2}
\end{subfigure}
\caption{Membrane dynamic response to an encounter with a sinusoidal gust, obtained for a nominal membrane of ${C_T=2.5}, {\mu=1}$. (a) amplitude profile computed for varying values of reduced frequency. Resonance frequencies, $\omega_r$, are denoted with dashed black lines and frequencies of inflection points, $k_{inv}$, are denoted by dotted black lines. Membrane amplitude profiles near the (b) first and (c) second inflection points.}
\label{fig:MemSinGust_mu1CT2_5-Mem}
\end{center}
\end{figure}
Figure~\ref{fig:SearsCompMu1CTvar} illustrates the effect of the tension coefficient on the equivalent Sears function, which is presented as a curve in the complex plane and in terms of its squared amplitude and phase, as compared to the classical modified Sears function. These results are obtained using the frequency-domain solution, which are verified by comparison with results of the Laplace domain solution (see dashed red line in figure~\ref{fig:SearsCompMu1CTvar-complx}).
The equivalent Sears function approaches the classical rigid aerofoil function uniformly at low reduced frequencies as the tension coefficient, $C_T$, increases.
An extreme case of $C_T=50$, in which the membrane is practically rigid, is presented in figure~\ref{fig:SearsCompMu1CTvar-amp2_phase} to confirm the approach of the analytical solution to that of the rigid aerofoil at large $C_T$. The equivalent Sears function of this extreme case follows closely the rigid-plate modified Sears function for a wide range of reduced frequencies up to about $k=2$ and validates our solution for large tension coefficients.
As the tension coefficient is increased from $2$ to $4$ in figure~\ref{fig:SearsCompMu1CTvar}, the first inflection point is delayed to a higher reduced frequency while also producing a larger unsteady lift amplification near the first resonance frequency.
The beneficial region, where significant gust mitigation is obtained, is also shifted to higher values of $k$ with an increase in the tension coefficient.
This unique characteristic of the membrane wing presents a special opportunity to calibrate the pre-tension on the membrane to successfully alleviate gusts of specific frequency regime, which is known to be one of the biggest challenges in SUAV design \citep[][]{Mohamed2014}.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Sears_mem_complex_FD_and_LD_mu1_CT_2to4_4latex_k_arrow.eps}
\caption{}
\label{fig:SearsCompMu1CTvar-complx}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Sears_mem_amp2_a_phase_FD_mu1_CT_2to4_w_CT50_4latex.eps}
\caption{}
\label{fig:SearsCompMu1CTvar-amp2_phase}
\end{subfigure}
\caption{Effect of tension coefficient on the membrane equivalent Sears function, for $\mu=1$: (a) Argand diagram; (b) squared amplitude and phase, as compared to the rigid-plate modified Sears function (black line).
An additional solution for a very large tension coefficient of $C_T=50$ is presented with a magenta dashed line in (b), to validate the unsteady solution by convergence to the rigid plate solution.
All results are obtained with the frequency-domain solution, presented for frequencies up to $k=k_2$ for clarity. Direct comparison with the Laplace-domain solution is made for $C_T=2$ (red dashed line in (a)), where the strong agreement indicates the equivalence of the methods.
}
\label{fig:SearsCompMu1CTvar}
\end{center}
\end{figure}
Looking into the effect of the mass ratio on the membrane lift response to a sinusoidal gust, figure~\ref{fig:SearsCompCT2Muvar} presents the equivalent Sears functions which are computed for varying values of the mass ratio, $\mu$, and are compared to the classical modified Sears function.
For very low frequencies, up to the point of inflection $k_{{inv}_1}$ (whose value decreases with increase in $\mu$), the lift response is practically unaffected by changes in the mass ratio, as predicted by the asymptotic analysis \eqref{eq:SearsMem_lim_k0}. Thus, the response can be considered quasi-steady for this frequency regime.
As the gust frequency is increased beyond the inflection point, a circular path appears at the complex plane plot with a radius (and amplitude) that slightly increases with $\mu$.
The membrane oscillates with the gust frequency so that an increase in $\mu$ leads to increase in the membrane inertia, which is proportional to $\mu k^2$, resulting with a larger amplitude of oscillation for a given gust frequency. However, as the system's resonance frequencies decrease with increase in $\mu$ (c.f., figure~\ref{fig:HeaveSSMem_CT_vs_k}), the lift amplification at the first resonance frequency increased only slightly with an increase in $\mu$.
We note that the in vacuo natural frequencies of the membrane are proportional to $\mu^{-\frac{1}{2}}$, indicating that for these frequencies the mass ratio has no effect on the membrane amplitude. Thus, the small effect of the mass-ratio on the lift amplitude, illustrated in figure~\ref{fig:SearsCompMu1CTvar}, is attributed to the added mass of the fluid-loaded membrane.
For frequencies beyond the first resonance frequency, the gust mitigation region is controlled by the membrane mass ratio in a manner that is similar to the effect of the tension coefficient (cf., figure~\ref{fig:SearsCompMu1CTvar-amp2_phase}), where a decrease in $\mu$ delays the region to higher values of $k$. However, in contrast to the tension coefficient, this variable is typically fixed in membrane wing applications, and thus is expected to be less useful for aerodynamic design purposes.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Sears_mem_complex_FD_and_LD_mu0_5to2_5_CT_2_5_4latex_k_arrow.eps}
\caption{}
\label{fig:SearsCompCT2Muvar-complx}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Sears_mem_amp2_a_phase_FD_mu0_5to2_5_CT_2_5_4latex.eps}
\caption{}
\label{fig:SearsCompCT2Muvar-amp2_phase}
\end{subfigure}
\caption{Effect of membrane mass ratio on the membrane equivalent Sears function plotted, $C_T=2.5$: (a) Argand diagram; (b) squared amplitude and phase, as compared to the rigid-plate modified Sears function (black line).
All results are obtained with the frequency-domain solution, presented for frequencies up to $k=k_2$ for clarity. Results from the Laplace-domain solution are shown in (a) with a red dashed line, presenting good agreement with the frequency-domain solution.}
\label{fig:SearsCompCT2Muvar}
\end{center}
\end{figure}
\subsubsection{Sharp-edged gust}\label{sec:results-gust-SEG}
The dynamic response of a nominal membrane wing to a sharp-edged gust is presented in figure~\ref{fig:SEG_mem_nom} in terms of its deformation in time and its lift response. Here we choose to present the actual lift response first (instead of the normalised response represented by the K\"{u}ssner function) to obtain quantitative conclusions on the difference in the lift produced by a flexible membrane wing (blue line) and a rigid flat plate (black line) during sharp-edged gust encounter. While the transient response up to $t\cong30$ introduces oscillations in the membrane deformation, resulting in oscillations in the lift response, at a later time a fully convex shape is obtained converging to the appropriate static solution \citep[][]{Nielsen1963}.
The lift response in figure~\ref{fig:SEG_mem_nom-CL} shows that, as expected, the membrane achieves a much higher lift coefficient than a rigid flat plate, converging to more than double the lift coefficient of the flat plate, as in the case of a step change in angle of attack (cf. figure~\ref{fig:MemAoAstepRes_mu1_CT2_5}). However, zooming in onto the transient lift response during gust penetration reveals that at initial stage, for $t<1.7$, the membrane presents a lower lift coefficient than the rigid plate.
The inset in figure~\ref{fig:SEG_mem_nom-CL} shows that the transient membrane deformation leads to a negative contribution to the lift coefficient for $t<1.7$, which results in a reduced lift coefficient compared to the rigid plate indicial lift. As the gust front approaches the trailing edge, the circulatory lift due to membrane deformation, $C_{l_d}^C$, increases first until it overcomes the non-circulatory lift, $C_{l_d}^{\mathit{NC}}$, for $t=1.7$. From this moment on, the membrane lift coefficient surpasses the rigid-plate lift and slowly converges to the static solution.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Mem_surf_vs_t_SEG_CT2_5_mu1_w_ss.eps}
\caption{}
\label{fig:SEG_mem_nom-dynamic}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/CL_vs_t_mem_nom_compare_rigid_and_012terms_w_zoom_latex_a.eps}
\caption{}
\label{fig:SEG_mem_nom-CL}
\end{subfigure}
\caption{Membrane dynamic response (a) and lift response (b) to a sharp edged gust, obtained for a nominal membrane of $C_T=2.5$ and $\mu=1$. Black line in (a) and dashed black line in (b) denote the static solution \citep[][]{Nielsen1963}. The membrane unsteady lift coefficient, $C_{l_m}$, is computed by superposition between the rigid plate indicial lift, $C_{l_g}$, and the lift due to membrane deformation, $C_{l_d}$, which is composed of the circulatory and non-circulatory terms, $C_{l_d}^C$ and $C_{l_d}^\mathit{NC}$, respectively. For $t<1.7$ the membrane deformation yields negative lift that reduces the total membrane lift compared to the rigid plate indicial lift. After time $t=1.7$ the membrane lift surpasses the rigid plate lift due to induced membrane camber and converges to the static solution.}
\label{fig:SEG_mem_nom}
\end{center}
\end{figure}
Figure~\ref{fig:SEG_mem_nom2-transient} presents the membrane profiles as obtained for $t\le2$, during gust penetration when the gust front travels along the chord.
The membrane encounters the gust at the leading edge at time $t=0$, in a taut initial position. As the gust front moves downstream, a small hump appears in the membrane profile near the leading edge, which increases in size and moves downstream with the advancement of the gust.
At time $t=1.7$, for which the membrane lift surpasses the rigid plate lift, the gust front has not yet reached the trailing edge, but a fully convex shape has developed with a maximum camber point at the aft part of the aerofoil. Note that the membrane profile obtains a positive camber during gust penetration, which would result with increased static lift compared to a rigid flat plate.
However, the unsteady response of the membrane, and in particular its acceleration, produces a negative non-circulatory lift at the initial response of the membrane (figure~\ref{fig:SEG_mem_nom-CL}), highlighting the importance of a full unsteady aerodynamic model for predicting the unsteady lift response of the membrane.
\begin{figure}
\begin{center}
\includegraphics[width=0.46\textwidth]{./Pictures/Mem_nom_profile_t0to2_CT2_5_mu1_w_max_camber.eps}
\caption{Membrane deformation in response to an encounter with a sharp-edged gust, obtained during gust penetration (when the gust front travels along the chord during ${0\le t\le 2}$) for a nominal membrane of $C_T=2.5, \mu=1$. Red points denote maximum camber point at each time step, and a dashed magenta line is used to identify the membrane profile at time $t=1.7$ (also denoted in figure~\ref{fig:SEG_mem_nom-CL} for the lift response).}
\label{fig:SEG_mem_nom2-transient}
\end{center}
\end{figure}
The membrane equivalent K\"{u}ssner function is presented in figure~\ref{fig:Kussner_mem} to study separately the effect of the tension coefficient and mass ratio. These plots were computed using the Laplace-domain solution \eqref{eq:KussnerMem_exp-LD} and were verified by comparison to the frequency-domain solution \eqref{eq:KussnerMem_def}. In general, the membrane response to a sharp-edged gust is slower than the rigid plate response, similarly to the Wagner function case. As the tension coefficient is reduced, a slower response is obtained since a larger camber profile is achieved at steady state, which takes a longer time to attain. Contrarily, the membrane mass ratio appears to affect only the initial oscillatory stage, while the rest of the response is practically independent of the mass ratio, as expected.
\begin{figure}
\begin{center}
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Kussner_LD_CT_2to4_mu1_w_zoom_W_FD_CT2.eps}
\caption{$\mu=1$}
\label{fig:Kussner_mem-CT_effect}
\end{subfigure}
\quad
\begin{subfigure}[]{0.46\textwidth}
\includegraphics[width=\textwidth]{./Pictures/Kussner_LD_CT_2_5_mu0_5to2_5_w_FD_mu1.eps}
\caption{$C_T=2.5$}
\label{fig:Kussner_mem-mu_effect}
\end{subfigure}
\caption{Effect of tension coefficient (a) and mass ratio (b) on the membrane equivalent K\"{u}ssner function, compared to the standard K\"{u}ssner function for a rigid aerofoil (black line). Results are obtained via Laplace domain solution and are validated against the frequency domain solution (red dashed line).}
\label{fig:Kussner_mem}
\end{center}
\end{figure}
\section{Concluding remarks}\label{sec:conclusions}
Unsteady aerodynamic theory is extended to include the aeroelastic deformations of flexible membrane wings in response to unsteady flow conditions. The pressure loads and membrane deformations due to dynamic fluid-membrane coupling are determined generally and exactly in the Laplace domain for small-amplitude prescribed chord motions or transverse gust encounters, which are also evaluated in the time domain using the numerical inversion scheme of \citet{Valsa1998}. Lift responses computed for the canonical unsteady aerodynamic scenarios of harmonic aerofoil motions or gusts, as well as of a step change in the angle of attack or gust profile, constitute aeroelastic extensions to the classical Theodorsen, Sears, Wagner, and K\"{u}ssner functions, respectively, for a membrane aerofoil.
The membrane structural response to harmonic heave motions and sinusoidal transverse gusts are verified against a separated analysis in the frequency domain. In each scenario, the membrane-equivalent Theodorsen and Sears functions follow the parametric trends of their classical, rigid aerofoil counterparts at low reduced frequency, $k$, albeit with a reduced unsteady lift amplitude and an increased phase lag. As the reduced frequency increases and approaches the first resonance of the fluid-loaded membrane, the membrane-equivalent Theodorsen and Sears functions introduce distinct circular paths in the complex plane, which initiate at different values of $k$ for each function. Closed-form expressions for both functions reveal that these circular paths are related to the membrane dynamic response during oscillations, where each circle corresponds to a different dominant mode, and the inflection points that connect the circles represent the shift in dominance between two consecutive membrane modes.
The present theoretical model results for harmonic motions or flow disturbances suggest parametric regions where the aeroelastic response of the membrane could enable performance improvements for flapping flight and gust resilience. In the case of harmonic heave motions, the unsteady lift amplitude of the membrane is superior to that of the rigid flat plate for a range of reduced frequencies centred about the first resonance. This parametric region is controlled by the tension coefficient, whereby the flapping aerodynamic performance may be enhanced through either passive or modest active control of the membrane pretension. The passive shape adaptation of the membrane may instead be used to weaken the unsteady lift response to a sinusoidal gust in a desired manner. Specifically, the present model predicts that the membrane aerofoil diminishes the unsteady lift response to sinusoidal gusts at reduced frequencies in a regime above the first resonance frequency. In addition, by increasing the membrane pretension, the beneficial region of reduced frequency shifts to higher frequencies, which could be exploited as a gust mitigation strategy in practice.
Closed-form expressions derived in the Laplace domain for the membrane equivalent Wagner and K\"{u}ssner functions reveal the direct dependence of these functions on the original rigid-plate functions and on the Fourier coefficients that describe the membrane profile. Each indicial lift response of the membrane aerofoil is slower relative to the rigid plate response but results in a significantly higher lift in the steady state due to aeroelastic membrane deformation. The membrane initial response in short times to an abrupt change in angle of attack or to an encounter with a sharp-edged gust produces a negative non-circulatory lift due to the acceleration of the membrane profile from a still and taut position. Therefore, the overall initial lift response of the membrane is smaller than the rigid plate response due to the gradual elastic reaction of the membrane to the changes in the fluid flow. However, at later times the non-circulatory lift due to the abrupt change in the flow field weakens and the circulatory lift increases, such that the membrane lift quickly overcomes the rigid plate lift response and converges to the static solution.
Results from the present theoretical effort invite computational and experimental companion efforts to elucidate the practical range of validity of the model, including the influence of nonlinear flow effects that occur at large gust ratios \citep[e.g.,][]{andreu2020effect,Jones2020}, which are expected to inform future improvements to the predictive aeroelastic framework.
\section*{Acknowledgement}
This work was supported by the Zuckerman-CHE STEM Leadership Program, with partial support from the National Science Foundation under award 1846852.
\section*{Declaration of interests}
The authors report no conflicts of interest.
|
\section{Introduction}
In this paper we study a local measure for singularity which
is based on a natural generalization of smoothness, the Gorenstein property.
The combinatorial properties of a polyhedral cone are deeply linked to the
singularities of the associated toric variety.
There are several results in this direction based on the following
theorem proved in [HHS19]:
the non-Gorenstein locus of a graded ring $R$ with canonical module $\omega$ is the the set of
prime ideals containing the trace ideal
\[
\operatorname{tr}(\omega) = \sum_{\varphi \in \operatorname{Hom}_R(\omega, R) } \varphi(\omega).
\]
This makes the ideal $\operatorname{tr}(\omega)$ a measure for singularity.
The aim of this paper is to provide a clear relation between the combinatorics
of the polyhedral cone and the geometry of the toric variety.
In particular, Theorem \ref{thm: charac} in Chapter 2 describes the non-Gorenstein loci of
toric varieties.
This is a vast generalization of results in \cite{Measuring} (Theorem 4.9), where
Herzog, Mohammadi and Page give criteria for certain simplicial toric varieties to be
Gorenstein on the punctured spectrum.
As an application of the toric framework
developed in section 2, we proceed with an investigation of
non-Gorenstein loci of Hibi rings in section 3 and 4, and
as a non-toric application we study secants of Segre varieties in the last section.
The separate treatment of the toric case enables us to give particularly
conceptual results and comparatively short and clear proofs.
A Hibi ring $k[P]$
is a toric algebra associated
to a finite, partially ordered set $P$ whose combinatorial properties
determine the geometry of the associated toric variety $\operatorname{Spec}(k[P])$.
They were introduced 1987
by Takayuki Hibi in \cite{Hibi_rings}. Initially, Hibi rings were studied because of their
appearance as flat degenerations of the
coordinate rings of Grassmannians and, more generally, flag varieties.
For a construction of these deformations consider \cite{Hodge}, and for a more modern construction using Sagbi bases consider \cite{Hibi_as_degen}.
Nowadays Hibi rings are objects of interest in their own, and appear in various geometric,
algebraic and combinatorial contexts. Consider for example \cite{Alexdual}, \cite{Fatemeh}, \cite{Howe2005WeylCA}, \cite{Sangjib}.
The non-Gorenstein loci of Hibi rings have been subject to extensive study:
a Hibi ring
$k[P]$
is Gorenstein if and only if $P$ is a pure poset, that is, all maximal chains of $P$ have the
same length ([Hib87], Corollary 3.d).
In \cite{Measuring} it is shown that the non-Gorenstein locus is zero-dimensional if and only if each connected component of $P$ is pure.
In an attempt to find a good regularity condition weaker than the Gorenstein property, but stronger than the Cohen-Macaulay property, in \cite{Herzog}, Herzog, Hibi, and Stamate define $k[P]$ to be nearly Gorenstein if
$\operatorname{tr}(\omega)$ is the graded maximal ideal.
They prove that $k[P]$ is almost Gorenstein if and only if
all connected components $P_i$ of $ P$ are pure, and the difference in their ranks is bounded by one.
In
\cite{Gor_loci_Janet} (Theorem 4.5), Miyazaki and Page give a description of the radical of $\operatorname{tr}(\omega)$
in combinatorial terms.
By proving Theorem \ref{theorem: max_comp} and Theorem \ref{theorem: graded_fibres}, we
give a new characterization of $\sqrt{\operatorname{tr}(\omega)}$ in a more conceptual, simple, and natural manner that unifies the above results.
Ultimately, we measure the deviation of $\operatorname{Spec}(k[P])$ from being Gorenstein, by how much $P$ deviates from being graded, in a precise sense, relating a geometric and a combinatorial notion of irregularity. These results take advantage of the discussion in section 2.
In Chapter five we proceed with an application of
Theorem \ref{thm: charac} to the non-toric case:
the first secant variety of the Segre variety
\[\mathbb{P}(V_1) \times \cdots \times \mathbb{P}(V_n) \xhookrightarrow{} \mathbb{P}(V_1 \otimes \cdots \otimes V_n) \]
is the closure of all tensors of rank $\leq 2$.
Secants of Segre varieties are objects in classical algebraic geometry that
are connected to the border rank of tensors, and to the computational complexity of matrix multiplication.
In \cite{Mateusz} the first secant variety, and in particular its singularities, are studied
using methods from statistics.
The most important tool are
secant-cumulant coordinates.
They are a special case of $L$-cumulants introduced in \cite{L-cumul}, and the induced coordinate change identifies affine, open patches
of the secant variety with certain toric varieties.
We extend the results of \cite{Mateusz} in the last chapter, the main result being Theorem \ref{thm: dim_Secants}.
\section{The toric trace ideal}
Our notation is close to \cite{cox2011toric}, which is the source that we refer to for information about toric varieties.
Let $N$ be a free abelian group of rank $n$ with dual $M = \operatorname{Hom}(N, \mathbb{Z})$,
and let
$\sigma \subseteq N_\mathbb{R} = N \otimes \mathbb{R}$ be a rational, pointed cone.
We denote its dual cone
$\sigma^\vee = \{l \in M_\mathbb{R}| \ l(\sigma) \subseteq \mathbb{R}_{\geq 0} \}$
and consider the normal toric ring $R = k[\sigma^\vee \cap M]$, where $k$ is an algebraically closed field in characteristic 0.
To a ray $\rho$ of $\sigma$ with primitive ray generator $u_\rho \in N$, we associate the torus-invariant prime-divisor
$D_\rho$. It is defined by the ideal
$I_ \rho$, spanned by all monomials away from the facet defined by $\rho$:
\[
I_\rho := \langle \chi^m | \ m \in \sigma^\vee \cap M, \ \langle u_\rho, m \rangle > 0 \rangle.
\]
The negative sum $K = \sum_{ \rho } - D_\rho$ of all torus-invariant, prime Weyl-divisors
is referred to as the canonical divisor of the toric variety $X = \operatorname{Spec}(R)$.
The module $\omega = \Gamma(\mathcal{O}_X(K)) $ is called the canonical module of $R$.
In \cite{Herzog} the non-Gorenstein locus of $X$ is described as the vanishing locus of its trace ideal:
\begin{definition}
The trace ideal of a module $M$ of $R$ is generated by the images of all $R$-module morphism $\varphi: M \longrightarrow R$.
\[
\operatorname{tr}(M) := \sum_{\varphi \in \operatorname{Hom}_R(M, R) } \varphi(M).
\]
\end{definition}
\begin{lemma} [\cite{Herzog}, Lemma 2.1]
\label{lemma:Herzog}
For every prime $\mathfrak{p} \in$ Spec$(R)$, $R_{\mathfrak{p}}$ is not Gorenstein if and only if
$\operatorname{tr}(\omega) \subseteq \mathfrak{p}$.
\end{lemma}
In other words,
the non-Gorenstein locus of $X$ is the vanishing locus of
the radical $\sqrt{\operatorname{tr}(\omega)}$.
In the remainder of this section we determine the minimal primes lying over $\operatorname{tr}(\omega)$, i.e. the irreducible components of the
non-Gorenstein locus.
The canonical module $\omega$ is spanned by all monomials
in the polyhedron
\[
P_K = \{ x \in M_\mathbb{R} | \ \forall \rho : \langle u_\rho, x \rangle \geq 1 \},
\]
which is obtained by translating all facets of $\sigma^\vee$ into the interior by one lattice-length.
Similarly, its dual $\omega^\vee= \operatorname{Hom}_R(\omega, R)$ is the module generated by the monomials in
the polyhedron
\[
P_{-K} = \{ x \in M_\mathbb{R} | \ \forall \rho : \langle u_\rho, x \rangle \geq -1 \}.
\]
We now proceed with a polyhedral description of the homogeneous ideal tr($\omega$).
\begin{proposition}
\label{prop:gens}
The trace ideal $\operatorname{tr}(\omega)$ is generated by all monomials
$
\chi^{m + m'} , $ where $ m \in P_K \cap M,$ and $ m'\in P_{-K} \cap M$. In particular, tr$(\omega)$ is $M-$graded.
\end{proposition}
\begin{proof}
By definition
$\omega$ is the module of global sections of the sheaf $ \mathcal{O}(K) $
associated to the canonical divisor.
Its dual module
$\omega^\vee$ is the module of global sections $\Gamma( \mathcal{O}(-K) )$:
\[
\omega^\vee = \bigoplus_{ \operatorname{div}(\chi^m) - K \geq 0 } k \cdot \chi^m
= \bigoplus_{ m \in P_{-K} \cap M } k \cdot \chi^m.
\]
Both $\omega$ and $\omega^\vee$ are submodules of the function field $K(X)$.
Evaluation of an element of $\sigma^\vee$ at an element of $\sigma$ is multiplication within $K(X)$. So
$\operatorname{tr}(\omega)$ is the product $\omega^\vee \omega$ in $K(X)$,
and
the statement of the proposition follows.
\end{proof}
Let $\mathfrak{p}$ be a prime in $\operatorname{Spec}(R)$, minimal with the property that it contains $\operatorname{tr}(\omega)$.
Since $\operatorname{tr}(\omega)$ is graded, $\mathfrak{p}$ is as well, and by the orbit-cone correspondence
it is associated to a face $F$ of $\sigma^\vee$:
\[
\mathfrak{p} = \langle \chi^m | \ m \in (\sigma^\vee \cap M) \setminus F \rangle.
\]
To decide whether an arbitrary graded prime $\mathfrak{p}$, associated to a face $F$ of $\sigma^\vee$, contains $\operatorname{tr}(\omega)$, we make the following definition:
\begin{definition}
Let
\[
F[1] :=
\{
x \in M_\mathbb{R} | \ \forall \rho , \ u_\rho \in F^{\perp}:
\langle u_\rho, x \rangle = 1
\}
\]
be the affine-linear space that is the intersection of all
facet-defining hyperplanes containing $F$, translated by one lattice-length into the interior of
$\sigma^\vee$.
\begin{example}
In this example $\sigma$ is the planar cone given by the inequalities
$ 2y \geq x, \ 2y \geq - x$. The face $F$ is the point $(0,0)$, and $F[1]$ is the point $(0, 0.5)$.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\tkzInit[xmax=3,ymax=1.5,xmin=-3,ymin=-0.5]
\tkzGrid
\tkzAxeXY
\coordinate (label) at (-0.3 , 0.8);
\coordinate (O) at (0,0);
\coordinate (X1) at (3, 1.5) ;
\coordinate (X2) at (-3,1.5) ;
\coordinate (Ox) at (0,0.5);
\coordinate (X1x) at (2, 1.5) ;
\coordinate (X2x) at (-2,1.5) ;
\filldraw [blue, opacity = 0.2] (O)--(X1)--(X2) ;
\filldraw [red, very thick] (O)--(X1) (O) --(X2) ;
\filldraw [yellow, very thick] (Ox)--(X1x) (Ox) --(X2x) ;
\end{tikzpicture}
\end{figure}
\end{example}
\end{definition}
The following theorem characterizes the graded primes lying over tr$(\omega)$
in terms of $F[1]$:
\begin{theorem}
\label{theorem: min_primes}
Let $\mathfrak{p} \subseteq R$ be a graded prime defined by a face $F$ of $\sigma^\vee$. Then
tr$(\omega) \subseteq \mathfrak{p}$ holds if and only if
$F[1]$ does not contain a lattice point.
\end{theorem}
\begin{proof}
Towards a contradiction we assume tr$(\omega) \subseteq \mathfrak{p}$,
and that there is a lattice point $z$ in $F[1]$.
Let $w$ be a lattice point in the relative interior of $F$.
Then
for every primitive generator $u_\rho$ of $\sigma$ the following inequalities hold.
\begin{align*}
\langle u_\rho, w \rangle > 0, \text{ if } u_\rho \not \in F^\perp, \
\langle u_\rho, w \rangle = 0, \text{ if } u_\rho \in F^\perp.
\end{align*}
After replacing $w$
with a positive integer multiple, we may assume the inequality
$ \langle u_\rho , w \rangle > \langle u_\rho, z\rangle + 1 $
to hold for all $u_\rho$ not in $F^\perp$.
Then $w - z \in P_{-K} $
and $z +w \in P_K$.
We obtain
\[2w = (w -z) + (z+w) \in P_{-K} \cap M + P_K \cap M. \]
By Proposition \ref{prop:gens}
$\chi^{2w}$ lies in $\sqrt{\operatorname{tr}(\omega)}$.
$\chi^{2w}$ is not contained in $\mathfrak{p}$, and hence the radical of tr($\omega$)
is not contained in $\mathfrak{p}$. Thus, tr($\omega$) is not contained either.\\
For the other direction let
$\chi^w \in \operatorname{tr}(\omega) \setminus \mathfrak{p}$.
Then $w$ lies in $F$, and by
Proposition \ref{prop:gens}
there is an element $z \in P_K \cap M$
with $w - z \in P_{-K}$:
\[
-\langle u_\rho ,z \rangle= \langle u_\rho ,w - z \rangle \geq -1
\]
holds for all $u_\rho$ in $F^\perp$.
So $\langle u_\rho, z \rangle \leq 1$, and hence $\langle u_\rho ,z \rangle = 1$ since $z \in P_K$.
We obtain
$z \in F[1]$.
\end{proof}
\begin{remark}
Theorem \ref{theorem: min_primes} can easily
be generalized to normal toric varieties that are
not affine. We use the notation from \cite{cox2011toric}.
Let $X$ be a normal toric variety defined by a fan $\Sigma$ of rational,
pointed cones in $N_\mathbb{R}$,
$Y \subseteq X$ a non-empty torus-invariant subvariety, given by a
cone $\sigma \in \Sigma$.
\end{remark}
\begin{theorem}
\label{thm: charac}
The subvariety $Y$ is contained in the non-Gorenstein locus $Z$ if and only if
there is no element $m$ of $M$, such that
$\langle u_\rho, m \rangle = 1$
holds for every ray $\rho$ of $\sigma$.
\end{theorem}
\begin{proof}
Both $Z$ and $Y$ are closed.
Since $Y$ intersects the affine open scheme $\operatorname{Spec}( k[ \sigma^\vee \cap M] )$ in
$\operatorname{Spec}( k[ \sigma^\perp \cap M] )$,
we may replace $Y$ with $\operatorname{Spec}( k[ \sigma^\perp \cap M] )$,
$X$ with $\operatorname{Spec}( k[ \sigma^\vee \cap M] )$, and $Z$ with the non-Gorenstein locus of $\operatorname{Spec}( k[ \sigma^\vee \cap M] )$.
By Lemma \ref{lemma:Herzog} and Theorem \ref{theorem: min_primes},
$Y$ is contained in $Z$ if and only if for the choice
$F = \sigma^\perp$, $F[1]$ does not contain a lattice point. In other words,
if there is no element $m$ of $M$ with
\[
u_\rho \in (\sigma^\perp)^\perp \implies \langle u_\rho, m \rangle = 1 \]
for every ray $\rho$ of $\sigma$.
All ray generators of $\sigma$ are contained in
$
(\sigma^\perp)^\perp = \sigma - \sigma.
$
\end{proof}
\section{Non-Gorenstein loci of Hibi rings}
Hibi rings are certain toric rings associated to finite, partially ordered sets.
We recall a description of the the associated polyhedral cone, and apply
the results from the previous chapter.
Finally, Theorem \ref{thm: charac} describes the non-Gorenstein locus as a combinatorial
measure for how far posets deviate from being graded, relating a geometric and a combinatorial notion of irregularity.
We call a set $P$, together with a transitive, reflexive order $\leq$,
a partially ordered set, or poset.
To elements
$a \leq b$ of $P$ we associate the interval
$
[a,b] : = \{ x \in P | \ a \leq x \leq b\},
$
and call $P$ bounded if
$P$ is the interval
$
[a, b]
$
for some elements $a$ and $b$.
For different elements $a \leq b$,
the covering relation $a \lessdot b$ is defined to hold if
$
\#[a, b] = 2.
$
That is, no elements lie properly between $a$ and $b$.
We call a totally ordered poset
$a_1 \leq \cdots \leq a_r$
of cardinality $r$
a chain of length $r-1$.
$P$ is defined to be pure if all chains contained in $P$,
maximal with respect to inclusion,
have the same length.
The set $\mathcal{I}(P)$ of subsets $I \subseteq P$, that are closed from below,
is called the lattice of order ideals.
\begin{definition}
Let $P$ be a finite poset and let $k[t, x_p, \ p \in P]$
be the free $k$-algebra in the variables
$x_p$ for all $p$ in $P$, and the variable $t$.
For each order ideal $I$ we denote the monomial
\[
x^I := \prod_{p \in I} x_p.
\]
The $k$-algebra
$R(P)$
generated by all monomials $tx^I$, where $I$ runs over all poset-ideals,
is the Hibi-ring associated to $P$.
\end{definition}
In fact,
$R(P)$ is the normal toric ring
$
k[C(P) \cap M]
$
associated to a cone $C(P)$, and of Krull-dimension $\#P + 1$
(\cite{Hibi_rings}).
The cone
\[
C(P) = \{ \psi: \overline{P} \longrightarrow \mathbb{R} | \ \forall a, b \in \overline{P}:
a \leq b \implies \psi(a) \geq \psi(b) \ , \psi(\infty) = 0 \}
\]
consists of the order-reversing maps from
$
\overline{P} := P \dot{\cup} \{-\infty, \infty\}
$
to $\mathbb{R}$, taking $0$ as minimal value.
It is the cone over the order-polytope
\[
Q(P) := \{ \psi \in C(P)|\ \psi(- \infty) = 1 \} .
\]
The generators of $R(P)$, the order ideals, and the vertices of $Q(P)$,
are in natural bijection, as can be seen by associating to an order ideal $I$ the vertex $\psi_I$ of
the order- polytope $Q(P)$:
\begin{align*}
\psi_I : \overline{P} &\longrightarrow \mathbb{Z} \\
p & \longmapsto
\begin{cases}
1, & \text{for } p \in I \cup \{ -\infty\} \\
0, & \text{for } p \notin I \cup \{- \infty\} \\
\end{cases}.
\end{align*}
The faces of $C(P)$ have a combinatorial description.
The maximal proper faces, called facets, are in bijective correspondence to
the order relations:
\[
F_{a \lessdot b } := \{ \psi \in C(P) | \ \psi(a) = \psi(b) \}.
\]
More generally, faces of $C(P)$ are families of functions $\psi$ that are constant along certain equivalence relations on $P$:
\begin{definition}
Let $P$ be a finite poset. A quotient poset of $P$ is a poset $P'$ together with a surjective, order preserving map
$\phi : P \longrightarrow P'$ with connected fibres, such that the order relation on $P'$ is the transitive hull of
the relation $a' \leq' b' := \exists a \in \phi^{-1}(a'),\ b \in \phi^{-1}(b') \mid \ a \leq b$.
\end{definition}
\begin{definition}
\label{def: Fphi}
Let $\phi: \overline{P} \twoheadrightarrow P'$ be a quotient poset of $\overline{P}$.
We denote by $F_\phi$ the face
\[F_\phi := \underset{\substack{ a \lessdot b, \\ \phi(a) = \phi(b) }}{\bigcap} F_{a \lessdot b} \]
of $C(P)$, consisting of the functions $\psi$ that are constant on the fibres of $\phi$.
\end{definition}
This definition bijectively identifies quotient posets and (possibly empty) faces of $C(P)$ (\cite{Geissinger}).
The inclusion order of faces corresponds to the refinement order of those equivalence relations $\sim_\phi$, that identify the fibres of $\phi$.
\begin{example}
\label{exmp: poset}
Consider the partially-ordered set
$
P = \{ p_1, p_2, p_3 \}
$
with only the relation $p_1 \geq p_2$.
\begin{figure}[H]
\centering
\subfloat[][]{
\begin{tikzpicture}
\filldraw[black] (4.5,1)node[anchor=west] {$\overline{P}:$};
[scale=.9,auto=center,every node/.style={circle,fill=red!10}]
\node (a2) at (6,0) {$p_2$};
\node (a1) at (6,2) {$p_1$};
\node (a3) at (8,1) {$p_3$};
\node (a4) at (7,3) {$\infty$};
\node (a5) at (7,-1) {$-\infty$};
\draw (a1) -- (a2);
\draw (a1) -- (a4);
\draw (a3) -- (a4);
\draw (a2) -- (a5);
\draw (a3) -- (a5);
\end{tikzpicture}
}
\subfloat[][]{
\tdplotsetmaincoords{60}{130}
\begin{tikzpicture}[fill=lightgray, tdplot_main_coords, scale = 2]
\tdplotsetrotatedcoords{12}{10}{15}
\begin{scope}[draw=red, tdplot_rotated_coords, axis/.style={->,dashed}]
\coordinate (A1) at (0,0,0);
\coordinate (B1) at (0,0,1) ;
\coordinate (A2) at (0,1,0) ;
\coordinate (B2) at (0,1,1) ;
\coordinate (A3) at (1, 1, 0) ;
\coordinate (B3) at (1,1,1);
\draw[axis, black] (0, 0, 0) -- (2, 0, 0) node [right] {$\psi(p_1)$};
\draw[axis, black] (0, 0, 0) -- (0, 2, 0) node [above] {$\psi(p_2)$};
\draw[axis, black] (0, 0, 0) -- (0, 0, 2) node [above] {$\psi(p_3)$};
\draw[fill] (A1) -- (A2) -- (A3) -- cycle;
\draw[fill] (A1) -- (A2) -- (B2) -- (B1) -- cycle;
\draw[fill] (A1) -- (A3) -- (B3) -- (B1) -- cycle;
\draw[fill] (A2) -- (A3) -- (B3) -- (B2) -- cycle;
\draw[dotted, thick] (A1) -- (A2);
\end{scope}
\end{tikzpicture}
}
\end{figure}
There are nine faces of $Q(P)$ of dimension one, corresponding to order preserving, surjective maps
to the unique bounded, partially ordered set $ q_0 \leq q_1 \leq q_2 $ with three elements.
We give a list
that matches the one-dimensional faces $F$ to the fibres of $\phi$.
\begin{center}
\begin{tabular}{ |c|c|c|c|c|c }
\hline
Nr. & $F$ & $\phi^{-1}(q_2) $ & $\phi^{-1}(q_1) $ & $\phi^{-1}(q_0) $ \\
\hline
1 & conv($ \{ (0, 0, 0), \ (0,0, 1) \} $) & $\{ \infty, p_1, p_2 \}$ & $\{p_3\}$ & $\{-\infty\}$\\
2 & conv($ \{ (0, 0, 0), \ (0,1, 0) \} $) & $\{ \infty, p_1, p_3 \}$ & $ \{p_2\} $ & $\{-\infty\}$\\
3 & conv($ \{ (0, 0, 0), \ (1, 1, 0) \} $) & $ \{ \infty, p_3 \}$ & $\{p_1, p_2\}$& $\{-\infty\}$\\
4 & conv($ \{ (0, 0, 1), \ (0, 1, 1) \} $) & $\{ p_1, \infty \} $ & $\{p_2\}$ & $\{ p_3, -\infty \}$ \\
5 & conv($ \{ (0, 0, 1), \ (1, 1, 1) \} $) & $\{\infty\}$ & $\{ p_1, p_2 \}$& $\{ p_3, - \infty\}$ \\
6 & conv($ \{ (0, 1, 0), \ (0, 1, 1) \} $) & $ \{ p_1, \infty\}$ & $\{p_3\}$ & $\{ p_2, -\infty \}$\\
7 & conv($ \{ (0, 1, 0), \ (1, 1, 0) \} $) & $ \{ p_3, \infty\}$ & $\{p_1\}$& $\{ p_2, -\infty \}$\\
8 & conv($ \{ (0, 1, 1), \ (1, 1, 1) \} $) & $\{\infty\}$ & $\{p_1\}$& $\{p_3, p_2, -\infty \}$ \\
9 & conv($ \{ (1, 1, 0), \ (1, 1, 1) \} $) & $\{\infty\}$ & $\{p_3\}$ & $\{ p_1, p_2, -\infty \}$ \\
\hline
\end{tabular}
\end{center}
\end{example}
Applying the results from Chapter 3 now allows us to characterize the non-Gorenstein locus
in terms of non-graded subsets of $P$:
\begin{definition}
Let $P$ be a finite poset. We call $P$ graded if
there is an order-reversing map $ \psi: P \longrightarrow \mathbb{Z}$,
such that for every covering relation $a \lessdot b $ it holds $ \psi(a) = \psi(b) +1 $.
$\psi$ is called a grading of $P$.
\end{definition}
Let the face $F_\phi$ of $C(P)$ be defined by a quotient poset $\phi: \overline{P} \twoheadrightarrow P'$, as in
definition \ref{def: Fphi}.
\begin{theorem}
\label{theorem: graded_fibres}
$F_\phi[1]$ contains a lattice point if and only if every fibre of $ \phi $, equipped with the restricted order relation, is graded.
\end{theorem}
\begin{proof}
By definition, $F_\phi$ is the intersection
\[
F_\phi = \bigcap_{ \substack{ a \lessdot b, \\ \phi(a) = \phi(b) } } F_{a \lessdot b}
\]
of all facets $F_{a \lessdot b}$, where $a$ and $b$ are elements in the same fibre.
Each such facet is supporting, so we obtain
\[
F_\phi[1] =
\{ \psi \in M_\mathbb{R} | \ \forall a \lessdot b, \ \phi(a) = \phi(b) : \
\psi(a) = \psi(b) + 1 \},
\]
and consequently the lattice points of $F_\phi[1]$
are functions $\psi$
that define a grading of each fibre of $\phi$, showing the implication from left to right.
Conversely, a separate choice of gradings on each fibre together form an integral element of $F[1]$.
\end{proof}
The remainder of this chapter is devoted to characterizing the maximal components of the non-Gorenstein locus.
\begin{lemma}
The subsets $A \subseteq \overline{P}$
that appear as fibres
of quotient poset maps $\phi: \overline{P} \longrightarrow P'$
are determined by the property that they be connected and
\[a \leq b \leq c, \ a, c \in A \implies b \in A. \]
\end{lemma}
\begin{proof}
To see that every such subset $A$ appears as a fibre, define
$P' := P \setminus A \cup\{ \star \}$ with the natural quotient map $\phi_A: \overline{P} \rightarrow P'$, and equip it with
the transitive closure of the relation
$a' \leq' b' := \exists a \in \phi_A^{-1}(a'),\ b \in \phi_A^{-1}(b') \mid \ a \leq b$.
The other direction is clear.
\end{proof}
We call subsets $A \subseteq \overline{P}$, appearing as fibres, complete.
For any quotient poset morphism $\phi$, having a non-graded fibre $A$,
consider the face
$F_{\phi_A}$ defined in the proof above.
It determines a subvariety of the non-Gorenstein locus
that is a superset of the variety determined by
$F_\phi$.
We obtain:
\begin{theorem}
\label{theorem: max_comp}
The map
\[
A \longmapsto
F_{\phi_A}
\]
bijectively identifies
the maximal components of the non-Gorenstein locus with subsets $A$ of $\overline{P}$, that are minimal with the property that
they be complete and not graded.
\end{theorem}
\begin{corollary}
\label{cor: dim_locus}
The dimension of the non-Gorenstein locus is
$ \max \{ \#P - \#A + 2 \},
$
where $A$ runs over all non-graded, complete subsets of $\overline{P}$.
\end{corollary}
\begin{proof}
The cone $C(P)$ has dimension
$\#P + 1$ and the cone $C(P \setminus A \cup\{ \star \})$
is of dimension $\#P - \#A + 2$.
\end{proof}
\begin{corollary}
The codimension of the non-Gorenstein locus is at least 4.
\end{corollary}
\begin{proof}
Every non-graded sub-poset $A \subseteq \overline{P}$ has at least 5 elements.
\end{proof}
\begin{remark}
Corollary \ref{cor: dim_locus}, generalizes the already known characterizations of
Gorenstein Hibi rings and Hibi rings that are Gorenstein on the pointed spectrum.
By the corollary, the non-Gorenstein locus is empty if and only if $\overline{P}$ is graded.
Since $\overline{P}$ is a bounded poset, it is graded if and only if it
is pure, which holds if and only if $P$ is pure.
Similarly, by the corollary, the locus is zero-dimensional if and only if $P \dot{\cup} \{ \infty \}$
and $P \dot{\cup} \{- \infty \}$ are graded.
It is easy to show that this is equivalent to every connected component of $P$ being pure
(for a proof consider Lemma 5.2 in \cite{Herzog}).
\end{remark}
\section{Comparison to \cite{Gor_loci_Janet}}
In the paper \cite{Gor_loci_Janet},
non-Gorenstein loci of Hibi rings are studied.
In particular, a family of graded ideals is
described in Theorem 4.5, that intersect in the radical ideal
$\sqrt{\operatorname{tr}(\omega)}$.
In this chapter we compare results and deduce Theorem 4.5 from the discussion in Chapter 3.\\
To state Theorem 4.5 we need the definitions of rank and distance of elements $a \leq b$ in a poset $P$:
the rank rank$(a, b)$ is defined to be the maximal
length $r-1$ of a chain $a = a_1 \lessdot a_2 < \cdots \lessdot a_r = b.$
Similarly, the distance dist$(a, b)$ is defined to be the minimal length $r-1$ of an inclusion-maximal chain
$a = a_1 \lessdot a_2 \lessdot \cdots \lessdot a_r = b. $
\begin{definition}
\label{def: a_i_b_j_vs_p}
Let $u$ be a natural number and
\[
a_1 < b_1 > a_2 < \cdots > a_u < b_u > a_1
\]
be elements of $\overline{P}$,
satisfying the inequality
\begin{equation}
\label{eq: inequality}
\sum_{i = 1}^u \operatorname{rank}(a_i, b_i) > \operatorname{dist}(a_2, b_1) + \cdots + \operatorname{dist}(a_u,b_{u-1}) +\operatorname{dist}(a_1,b_{u}) .
\end{equation}
We define the graded prime ideal
\[
\mathfrak{p}_{(a_1, \dots, a_u, b_1, \dots, b_u)} := \langle \chi^\psi| \psi \in C(P), \ \psi \text{ nonconstant on } \{a_1, \dots, a_u, b_1, \dots, b_u \} \rangle.
\]
\end{definition}
\begin{theorem}{(4.5, \cite{Gor_loci_Janet})}
The ideal $\sqrt{\operatorname{tr}(\omega)}$ is the intersecion of primes
\[
\sqrt{\operatorname{tr}(\omega)} = \bigcap_{(a_1, \dots, a_u, b_1, \dots, b_u)} \mathfrak{p}_{(a_1, \dots, a_u, b_1, \dots, b_u)}.
\]
\end{theorem}
By our description of the minimal primes lying over $\operatorname{tr}(\omega)$,
the monomials $\chi^\psi$ that lie in the radical $\sqrt{\operatorname{tr}(\omega)}$ are characterized by the property
that $\psi$ be non-constant on all non-graded, complete subsets of $\overline{P}$.
The observation to make is that,
given a complete subset $A$ of $\overline{P}$,
there exist elements $(a_1, \dots, a_u, b_1, \dots, b_u) $ in $A$ satisfying inequality \eqref{eq: inequality}, if and only if $A$ is not graded.
This proves the inclusion from right to left.
Conversely,
the union of all intervals $[a_i, b_j]$ forms a non-graded, complete subset $A$ of $\overline{P}$, showing the other inclusion.
\section{Secants of Segre varieties}
We start by introducing notation: let $k_1 \leq \dots \leq k_n$ be natural numbers, $n \geq 2$, and for each index $i$ let
$V_i$ denote a linear space of dimension $k_i$.
The image of the Segre map
\begin{align*}
\mathbb{P}(V_1) \times \cdots \times \mathbb{P}(V_n) & \longrightarrow \mathbb{P}(V_1 \otimes \cdots \otimes V_n) \\
(v_1, \dots, v_n) & \longmapsto v_1 \otimes \cdots \otimes v_n
\end{align*}
consists of all rank one tensors.
Its first secant variety $\operatorname{Sec}(k_1, \dots, k_n)$
is the Zariski-closure of the set of all tensors of
rank $\leq 2$.
In \cite{Mateusz}, an affine open covering of
$\operatorname{Sec}(k_1, \dots, k_n)$ is constructed, such that
each patch is isomorphic to the trivial vector bundle of rank $k_1 + \cdots + k_n$ over a certain toric variety $X$.
$X$ is the spectrum of the toric $k$-algebra
$k[\sigma^\vee \cap \mathbb{Z}^{1 + k_1 + \cdots + k_n} ]$, where $\sigma^\vee$ is the polyhedral cone
\begin{align*}
\sigma^\vee =
\{
&\left(q_0, q^1_1, \dots, q^1_{k_1}, \dots, q^n_1, \dots, q^n_{k_n} \right) \in \mathbb{R}^{1 + (k_1 + \cdots + k_n)}
\mid \\
&q^i_j \geq 0 \ \forall 1 \leq i \leq n, 1\leq j \leq k_i,\\
& q_0 - \sum_{j= 1}^{k_i} q^i_j \geq 0 \ \text{for } 1 \leq i \leq n,\\
&\sum_{i = 1}^n \sum_{j= 1}^{k_i} q^i_j -2q_0 \geq 0 \}.
\end{align*}
$\sigma^\vee$ is the cone over the polytope $Q = C \cap \{ q_0 = 1 \}$, which is the product of $n$ simplices $ \Delta_{k_i }$
of respective dimension $k_i$, intersected with the halfspace
$\{ \sum_{i,j} q^i_j \geq 2 \} $.
Using Theorem \ref{theorem: min_primes}, one can show that the non-Gorenstein locus of the secant variety
is the trivial vector bundle of rank $k_1 + \cdots + k_n$ over the locus of
$X$. This is done by expressing the faces of the cone $\sigma^\vee \times \mathbb{R}_{\geq 0}$, that contribute to the non-Gorenstein locus, as products
$F \times \mathbb{R}_{\geq 0}$.
We from now on assume that the non-Gorenstein locus is not empty.
According to \cite{Mateusz} (Theorem 7.18), this happens in all but the following cases.
\begin{itemize}
\item $n =5,: \ k_5 =1$,
\item $n =3: \ (k_1, k_2, k_3) \in \{ (1,1,1), (1,1,3), (1,3,3), (3,3,3) \},$
\item $n =2: \ k_2 = k_1$, or $k_1 = 1$.
\end{itemize}
In the the remainder of this section we investigate which positive-dimensional faces $F$ of $\sigma^\vee$ contribute to non-Gorenstein locus
of $X$, by applying Theorem \ref{theorem: min_primes}.
We show that any maximal such face is a cone over a product
$
\Delta_{k_l } \times \Delta_{k_m }
$
for distinct indices $l$ and $m$, and decide the existence of such faces.
Ultimately we obtain
\begin{theorem}
\label{thm: dim_Secants}
If $X$ is not Gorenstein,
the non-Gorenstein locus of $X$ is of dimension
\begin{align*}
&\bullet \emph{max}
\{
k_{l} +k_{m} +1, \text{ where } l \neq m , \ \sum _{i \neq l, m} k_i \neq 3 \}
& \text{if } n \geq 4 \text{ or } n = 3, k_1 > 1,\\
& \bullet k_{2} +k_{3} +1 & \text{ if } n = 3, k_2 \neq 1 ,\\
& \bullet 0 & \text{ else}.
\end{align*}
\end{theorem}
\begin{proof}
By construction of the cone $\sigma^\vee$, its dual cone $\sigma$ is generated by the vectors
\begin{itemize}
\item $R^i_j := e^i_j, \ \ \forall 1 \leq i \leq n, 1\leq j \leq k_i$,
\item $L_i := e_0 - \sum_{j= 1}^{k_i} e^i_j , \ \ \forall 1 \leq i \leq n$,
\item $S:= \sum_{i = 1}^n \sum_{j= 1}^{k_i} e^i_j -2e_0 $.
\end{itemize}
Here the vectors $e^i_j$ denote the standard basis of $\mathbb{R}^{k_1 + \cdots + k_n}$.
A list of the primitive ray generators of $\sigma$ is given in the proof of Theorem
7.18 in \cite{Mateusz}:
in the case $n=4$, all generators of $\sigma$ are ray generators.
If $n = 3$, all $R^i_1$ with $k_i = 1$ are omitted.
We now proceed with a proof of the theorem for the case
$n \geq 3$, and omit the case $n =2$. It is analogous, but easier, except
that for $n = 2$ the lineality space of $\sigma$ is spanned by $S, L_1, L_2$,
and the ray generators of
$\sigma / \langle S, L_1, L_2 \rangle$ are the vectors $[ R^i_j \operatorname ]$, so we need to replace
$\sigma$ with $\sigma / \langle S, L_1, L_2 \rangle$, and
$\sigma^\vee$ with $ \sigma^\vee \cap \langle S, L_1, L_2 \rangle^\perp$.\\
In the case $n=3$, $\sigma$ is a pointed cone,
and
we may apply Theorem \ref{theorem: min_primes}.
Let $F \subseteq \sigma^\vee$ be a positive-dimensional face.
An integral element of $F[1]$ is an intregral solution $(q^0,q^i_j) \in \mathbb{R}^{1 +k_1 + \cdots + k_n}$ to all affine-linear equations
\begin{itemize}
\item $ q^i_j = 1 $, if $F \subseteq (R^i_j)^\perp $,
\item $ q_0 - \sum_{j= 1}^{k_i} q^i_j = 1 $, if $F \subseteq L_i^\perp $,
\item $ \sum_{i = 1}^n \sum_{j= 1}^{k_i} q^i_j -2q_0 = 1 $, if $F \subseteq S^\perp $.
\end{itemize}
As can be seen by direct computation, such a solution always exists if $F \subseteq S^\perp $ does not hold, or if only for one ray generator $L_i$ it holds $F \subseteq L_i$.
Furthermore, $F$ is of dimension zero if $F \subseteq S^\perp $, and in addition there are at least three inclusions of the form $F \subseteq L_i$.
Let from now on $F \subseteq S^\perp $ and let $l$ and $m$ be the only two distinct indices such that $L_l$ and $L_m$ satisfy $F \subseteq L_l$, $F \subseteq L_m$.
We now investigate when $F[1]$ does not contain an integral element.
Observe that for each vector $R^i_j$ with $i \neq l, m$ it holds $F \subseteq (R^i_j)^\perp $.
If every such vector is a ray generator, by adding the three
three equations
\begin{gather*}
q_0 - \sum_{j= 1}^{k_{l}} q^{l}_j = 1, \ q_0 - \sum_{j= 1}^{k_{m}} q^{m}_j = 1, \
\sum_{i = 1}^n \sum_{j= 1}^{k_i} q^i_j -2q_0 = 1,
\end{gather*}
we obtain
\begin{equation}
\label{eq: 3 =}
3 = \sum_{i \neq l, m } \sum_{j= 1}^{k_i} q^i_j = \sum_{i \neq l, m } k_i.
\end{equation}
Consequently, $F[1]$ does not contain an integral element if equation \eqref{eq: 3 =} fails.
In the case $n = 3$ this is only true if $k_i \neq 1$ holds for every $i \neq l, m$, since our argument uses that all $R^i_j$, $i \neq l, m$ be ray generators.
If on the other hand equation \eqref{eq: 3 =} holds,
the choice
$q_0 =k_{l} +k_{m} $ and
$ q^i_j = 1 $ for all $i$ and $j$ except $q^{l}_1 = k_{m}, \ q^{m}_1 = k_{l}$
defines an integral element of $F[1]$.
The maximal faces $F$ contained in $S^\perp \cap L_l^\perp \cap L_m^\perp$ are of the form $F = C \cap L_l^\perp \cap L_m^\perp \cap S^\perp$, a cone over the product of simplices $F \cap Q = \Delta_{k_l } \times \Delta_{k_m } $, and of dimension $k_l + k_m +1$.
This proves the desired statement.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction}
\label{sec:intro}
In earth vision, change detection is a task to detect semantic changes in two high spatial resolution (HSR) images from different times and the same area (i.e. bi-temporal images).
Change detection is a very important task in the field of earth vision used for urban expansion, urban planning, environmental monitoring, and disaster assessment~\cite{hussain2013change,zhang2017separate}.
However, the manual comparison and change detection between two HSR images is a very labor-intensive and costly job.
To solve this problem, recent deep learning-based change detection methods~\cite{chen2021remote,fang2021snunet} have been proposed and the results are promising.
Due to the data-driven nature of deep learning methods, a large-scale training dataset of bi-temporal images and corresponding change labels is essential for supervised approaches~\cite{chen2021remote,fang2021snunet}.
The challenge lies in the expensive dataset: Collecting the correctly registered bi-temporal HSR images is expensive, and annotating the changes between them is more costly than general semantic segmentation~\cite{waqas2019isaid} or object detection datasets~\cite{xia2018dota}.
Another challenge is the imbalanced dataset: The change detection dataset requires two images taken at different times in the same area, and in real-world scenarios, changes are rare, so it is more difficult to collect a change detection dataset in which changes exist (e.g. class balanced).
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{images/fig1.pdf}
\caption{Qualitative comparison of unchanged bi-temporal pair images and changed bi-temporal pair images and pseudo-unchanged pair images. (a) and (b) are actual bi-temporal pair images, (c) is a pseudo-unchanged pair created by photometric transform.}
\label{fig:fig1}
\end{figure}
To solve this data collection problem, various unsupervised change detection (UCD) methods~\cite{he2021unsupervised, de2019unsupervised, ren2020unsupervised, li2019unsupervised} have been proposed.
UCD approaches effectively solve the problem of expensive annotations in change detection, but they still require correctly registered bi-temporal HSR images, or the performance was low compared to supervised learning methods.
In the existing UCD setting, because it does not explicitly train change and unchange, the prediction result is noisy both when the change occurs and when the change does not occur.
To solve this problem, most UCD methods use post-processing. However, they do not disclose how to post-processing, or they are overfitting for specific situations. (e.g. pixels with small area are removed)
Inspired by the study of unsupervised anomaly detection~\cite{li2021cutpaste,gong2019memorizing}, we rethink the UCD setup.
In both change detection and anomaly detection, change/anomaly situations are rare in the real world.
The unsupervised anomaly detection methods~\cite{li2021cutpaste,gong2019memorizing} train the image reconstruction model with only normal data, and the model is fitted to the normal distribution. During inference, normal inputs will be well reconstructed, as they fall in the normal distribution; on the other hand, anomaly inputs will have high reconstruction error, as they fall outside the normal distribution.
\textit{Can UCD use reconstruction errors like unsupervised anomaly detection?}
Since unchanged pairs can be generated synthetically, we can train an image reconstruction model that trains normal distributions.
For example, if the change detector is trained on the unchanged area by pairing $X^{t1}$ with itself, it can be trained on the unchanged area without $X^{t2}$ images or changed labels.
In this paper, we propose unsupervised change detection based on image reconstruction loss (CDRL) using only unlabeled single-temporal single source images.
The proposed method explicitly solves the challenges of data collection in change detection, as it does not require expensive bi-temporal HSR images, expensive annotations, nor balanced datasets with sufficient changes.
CDRL is trained to reconstruct the original source image by receiving the source image and the photometrically transformed source image as a pair.
The purpose of photometric transformations is to create pseudo-unchanged pairs that mimic unchanged pairs, as shown in \figref{fig:fig1}-(a) and \figref{fig:fig1}-(c). In the unchanged pairs, there are no structural changes by definition, but only style changes or photometric changes. The pseudo-changed pairs can be used to train CDRL instead of changed pairs.
Similar to unsupervised anomaly detection, CDRL receives only (pseudo-)unchanged pair images during training and is trained to reconstruct the original source image, so if untrained cases (changed pair~\figref{fig:fig1}-(b)) are input during inference, the reconstruction loss is high for that area.
However, unlike existing unsupervised anomaly detection studies, change detection usually receives two images, so there are two major problems.
First, the image reconstruction models should be able to reconstruct the original source image by receiving two pair images.
Second, the reconstruction models should focus more on the structure information of the photometrically transformed source image.
To solve this problem, we propose an image reconstruction model using generative adversarial networks based on encoder-decoder.
CDRL consists of a shared encoder to extract features from each image, and a decoder to fuse the features from the two images for image reconstruction.
To pay attention to the structure information of the photometrically transformed source image, spatial attention was performed only on the photometrically transformed source image.
To validate the efficacy of our proposed CDRL, we evaluated it on LEVIR-CD~\cite{Chen2020} and WHU-CD~\cite{liu2020novel}.
Even though CDRL does not use bi-temporal pairs or pre-trained weights, CDRL outperforms the existing UCD method using bi-temporal pairs and the UCD method using pre-trained weights by a large margin.
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth]{images/fig2_2.pdf}
\caption{Overview of the overall framework of \textit{CDRL}. CDRL is trained to reconstruct $X^{t1}$ by receiving a pseudo-unchanged pair during training, and when a changed bi-temporal pair that is not learned during training is input during inference, the reconstruction loss is large in the region with large structure change.}
\label{fig:fig2}
\end{figure*}
To sum up, our major contributions are as follows:
\begin{itemize}
\item We propose CDRL, a method to train a change detector on a single-temporal single source image in UCD. To the best of our knowledge, this is the first time to use a single-temporal single source image in UCD.
\item We propose an encoder-decoder-based generative adversarial network that receives paired images as input.
\item We evaluate CDRL on various change detection datasets, and CDRL outperforms previous UCD methods by a large margin.
\end{itemize}
\section{Related Work}
\label{sec:related}
The problem we are trying to solve is \textcircled{1} obtaining a matched bi-temporal image including a changed area is more difficult than obtaining a general single-temporal image \textcircled{2} pairwise annotation is very expensive and time consuming.
Therefore, this section focuses on the problems encountered by existing change detections, and finally briefly introduces the field of anomaly detection that we have been inspired by.
\subsection{Supervised Change Detection (SCD)}
Supervised change detection is largely divided into a method that uses only single-temporal information and an approach that performs temporal information modeling or different modeling~\cite{zheng2021change}.
A change detector that uses only single-temporal information, called post-classification comparison (PCC), trains a semantic segmentation model during training~\cite{zheng2021change, zheng2020foreground}.
After that, the semantic segmentation model predicts the change area through the xor operation of the results obtained by predicting images from two different times during inference.
PCC has the great advantage of not requiring coregistrated pair images, but this method only simply treats the change detection task as the semantic segmentation task and ignores the temporal information modeling, thus significantly decreasing the performance.
To solve this problem, change detection methods~\cite{chen2021remote,fang2021snunet} were proposed for temporal information modeling between pair images taken at different times in the same area. All of these methods achieved high performance, but the generalization performance of these models was not guaranteed because of the small size of the change detection benchmark datasets~\cite{lebedev2018change, tian2020hi, ji2018fully, chen2020spatial, daudt2019multitask, daudt2018urban, fujita2017damage, bourdis2011constrained, benedek2009change}.
The reason why the change detection benchmark datasets are small is that collecting bi-temporal pair images is much more difficult than collecting single-temporal images, and pairwise annotation is very expensive and time-consuming.
Since our proposed CDRL performs UCD using only unlabeled single-temporal single source images, it can alleviate the problem of collecting bi-temporal pair images and the cost and time-consuming problems of labeling.
\subsection{Unsupervised Change Detection (UCD)}
UCD is usually divided into a method~\cite{wu2013slow,thonfeld2016robust,blaschke2001object,im2008object} based on the concept of Change Vector Analysis~\cite{malila1980change} (CVA) or a method~\cite{ren2020unsupervised} based on a Generative Adversarial Network (GAN) using an unlabeled bi-temporal pair image.
However, because they use pre-trained weights without direct training on the dataset, the performance is low, or large-scale unlabeled bi-temporal pair images are required to train the GAN model.
Our proposed CDRL can be explicitly trained on an unchanged area and can be trained without bi-temporal pair images.
\subsection{Unsupervised Anomaly Detection}
The anomaly detection study we are inspired by is a reconstruction-based method~\cite{bergmann2019mvtec, perera2019ocgan}.
Reconstruction-based methods typically utilize generative models like auto-encoders or generative adversarial networks to encode and reconstruct the normal data. These methods hold the insights that the anomalies can not be reconstructed since they do not exist at the training samples.
These unsupervised anomaly detection methods achieved AUROC performance of over 95 in various benchmark datasets~\cite{bergmann2019mvtec}, even without explicitly training the anomaly data.
We also applied the fact that only unchanged pair (normal) images are trained during training like this reconstruction-based anomaly detection, and that when a changed pair (anomaly) is input during inference, the reconstruction loss is high.
\section{Method}
\label{sec:method}
This section describes the components of the CDRL in detail.
First, the training pipeline will be briefly described in ~\sref{sec:pipe}, and then, a method of performing photometric transform based on a single-temporal single source image will be described in ~\sref{sec:photo}.
~\sref{sec:model} describes the reconstructor that receives pair images and is trained as an objective to reconstruct the original source image.
Finally, ~\sref{sec:gan} describes the entire objective function of the CDRL including the GAN model.
\subsection{Overall Pipeline}
\label{sec:pipe}
CDRL performs photometric transform to create a pair image as a single-temporal single source image.
Photometric augmentation of simple rules such as brightness control and channel shuffling does not sufficiently express the style change of the corresponding bi-temporal pair image in the real-world unchanged area.
Therefore, in order to express the style change of the corresponding bi-temporal pair image of the unchanged area of the real-world, we perform photometric transform by style transfer using CycleGAN~\cite{zhu2017unpaired}.
After that, the generated pair image as shown in ~\figref{fig:fig2} is input to the U-Net-based original source image reconstructor during training.
For our purpose, to train the original source image reconstructor with high reconstruction loss for the region where the change has occurred, we need to pay attention to the channel information in the original source image and pay attention to the spatial information in the photometric transformed image.
To achieve this purpose, we applied spatial attention to the photometric transform image and channel attention to the original source image using the CBAM~\cite{woo2018cbam}.
Despite these efforts, the original source image reconstructor has a problem of overfitting the original source image too easily during training.
Therefore, to prevent overfitting, we made a discriminator and conducted adversarial training with the image reconstructor.
\subsection{Photometric Transformation}
\label{sec:photo}
The purpose of photometric transform for training CDRL is to create a natural style change while maintaining the structure like an actual unchanged bi-temporal pair image as a single-temporal single source image.
In order to achieve this purpose, we adopted CycleGan~\cite{zhu2017unpaired}, which receives unpaired images and changes the style while maintaining the structure.
In the existing CycleGan, when there are two domains $\{ x_{1}, x_{2}, ..., x_{n}\} \in X$ and $\{ y_{1}, y_{2}, ..., y_{n}\} \in Y$, it receives two samples $x_{i}$ and $y_{j}$ and is trained to optimize the parameters of two mapping functions $G: X \rightarrow Y$, $F: Y \rightarrow X.$
However, since we need to perform unpaired style transfer in one domain, we train a function that maps two randomly selected samples $x^{t1}_{i} \in X^{t1}$ and $x^{t2}_{i} \in X^{t2}$ in one domain $X$.
Therefore, when there is discriminator $D_{t2}$ for mapping function $G: X^{t1} \rightarrow X^{t2}$ and discriminator $D_{t1}$ for $F: X^{t2} \rightarrow X^{t1}$, our objective function is as follows:
\begin{equation}
\begin{aligned}
L(G,F,D_{t1},D_{t2}) = L_{GAN}(G,D_{t2}, X^{t1}, X^{t2}) \\
+ L_{GAN}(F,D_{t1}, X^{t2}, X^{t1}) + \lambda L_{cyc}(G,F),
\label{eq:cycle}
\end{aligned}
\end{equation}
where $\lambda$ controls the relative importance of the two objectives.
\subsection{Pair Image-based Source Image Reconstructor}
\label{sec:model}
The pair image-based Source Image reconstructor $R(.)$ is trained to reconstruct $X^{t1}$ by receiving the pseudo unchanged pair image $X^{t1}$, $X^{t2}$ previously created in ~\sref{sec:photo} as an input.
To achieve this purpose, the pair image-based source image reconstructor consists of a shared encoder and a decoder that concats and fuses each feature map of the pair image output from the encoder.
$R$ is trained to optimize the objective function as follows:
\begin{equation}
\begin{aligned}
L_{mae}(R) = MAE(R(X^{t1}, X^{t2}),X^{t1}),
\label{eq:recon}
\end{aligned}
\end{equation}
where ${MAE}$ is the mean absolute error between the reconstructed image and the source image.
A source image reconstructor trained only on pseudo unchanged pair images during training should have a high reconstruction loss when a changed pair image is received during inference.
However, if the source image reconstructor reconstructs by relying only on the structure information of the source image regardless of the photometrically transformed image, the reconstruction loss is low even when a changed pair image is input.
To alleviate this problem, we modified the CBAM structure to perform spatial attention on the photometric transformed image and channel attention on the original source image.
Through this process, the source image reconstructor is trained by paying attention to the structure information of the photometric transformed image and paying attention to the style information in the source image.
~\figureref{fig:fig9} shows our attention structure modified from the CBAM structure. As shown in the figure, channel attention is performed on the $X^{t1}$ image and spatial attention is performed on the $X^{t2}$ image, which is then added and concatenated to train.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{images/fig9.pdf}
\caption{Our proposed pseudo-pair CBAM structure. Channels are applied to the $X^{t1}$ image, and spatial attention is applied to the $X^{t2}$ image}
\label{fig:fig9}
\end{figure}
\subsection{GAN for detailed structure reconstruction}
\label{sec:gan}
As in the ~\cite{isola2017image} study, if only the $MLE$ loss is used, the reconstruction image does not reconstruct the structure well and is blurry. If the reconstruction result is blurry, the performance of the CDRL is degraded because it is insensitive to structure changes. Therefore, we use GAN as in \cite{isola2017image} to solve this problem. So, given the discriminator $D_{r}$, the objective function of the discriminator is:
\begin{equation}
\begin{aligned}
L_{gan}(R, D_{r}, X^{t1}, X^{t2}) = log(D_{r}(X^{t1})) \\
+ log(1-D_{r}(R(X^{t1}, X^{t2}))),
\label{eq:dis}
\end{aligned}
\end{equation}
where $R$ tries to reconstruct images $R(X^{t1}, X^{t2})$ that look similar to images from $X^{t1}$ , while $D_{r}$ aims to distinguish between translated samples $R(X^{t1}, X^{t2})$ and original source image $X^{t1}_{i}$. $R$ aims to minimize this objective against an adversary $D_{r}$ that tries to maximize it, i.e., $min_{R} max_{D_{r}} L(R, D_{r} , X^{t1}, X^{t2} )$.
The final objective function of the source image reconstructor that combines the GAN loss and the MAE loss is
\begin{equation}
\begin{aligned}
L_{total} = L_{gan} + \lambda L_{mae}
\label{eq:dis}
\end{aligned}
\end{equation}
where $\lambda$ controls the relative importance of the two objectives. We use $\lambda$ as 100 in all experiments.
\section{Experiments}
In this section, CDRL is evaluated on two HSR remote sensing change detection datasets.
Section ~\sref{sec:setting} describes the experimental setting in detail, and section ~\sref{sec:toy} describes the loss analysis experimental results in detail.
Also, in ~\sref{sec:pixel}, pixel level change detection, which is the same as the existing change detection experimental settings, and in ~\sref{sec:patch}, we describe the patch level change detection results suitable for our proposed change detector application situation.
Finally, in ~\sref{sec:qual}, qualitative results, and in ~\sref{sec:able}, the results of the ablation study are described in detail.
\subsection{Experimental Setting}
\label{sec:setting}
\paragraph{Datasets} Two HSR remote sensing change detection datasets were used to train and evaluate the performance of object change detection.
\begin{itemize}
\item \textbf{LEVIR-CD}~\cite{chen2020spatial}\textbf{.} The LEVIR-CD dataset contains 637 bi-temporal pairs of HSR images and 31,333 change labels on building instances. Each image has a spatial size of 1,024 $\times$ 1,024 pixels with a spatial resolution of 0.5 m. The change labels provide information about the construction of new buildings and the disappearance of existing buildings. This dataset provides an official split of 445 training, 128 validation, and 64 test pairs. The evaluation results are computed in the test pair set.
\end{itemize}
\begin{itemize}
\item \textbf{WHU building change detection}~\cite{ji2018fully}\textbf{.} The WHU dataset has one pair of aerial images of size 15,354 $\times$ 32,507 pixels obtained in 2012 and 2016 at the same area. It provides 12,796 and 16,077 building instances labels, respectively, and changed labels across the pair. We will use the change labels later to evaluate change detectors. Training, validation, and test sets are given specific areas containing 4,736, 1,036, and 2,416 tiles respectively.
\end{itemize}
\paragraph{Implementation details}
CycleGAN was used to generate pseudo unchanged pairs for training of CDRL. The two datasets $X^{t1}$ and $X^{t2}$ of CycleGAN are randomly divided into datasets $X$. All implementation details strictly follow the official CycleGAN code\footnote{https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix}.
For the data augmentation, RandomRotate90, HorizontalFlip, VerticalFlip, Transpose, RandomBrightnessContrast, and Sharpen of albumentations~\footnote{https://github.com/albumentations-team/albumentations} were used, and the probability $p$ of almost all applications is $0.3$.
We trained the source image reconstructor using the Adam optimizer with beta values equal to (0.5, 0.999). The learning rate set to 0.0002 and batch size of 1 to train the model. evaluation is using both the LEVIR-CD and WHU dataset validation (or test) sets.
Our all models are implemented on PyTorch and trained using a single NVIDIA Quadro RTX 8000 GPU.
The detailed structure of the network will be available soon on our project page.
\paragraph{Evaluation Metrics}
\begin{itemize}
\item \textbf{Pixel Level Change Detection.} We use the common metrics in pixel-by-pixel binary classification tasks and object change detection tasks: intersection over union (IoU), recall, precision score. Because our goal is also to classify whether it has changed or not at the pixel level.
\end{itemize}
\begin{itemize}
\item \textbf{Patch Level Change Detection.} We used classification AUC for patch level change detection. Note that if all of the output mask values of the change detector are 0, it is set to unchanged, if at least one is 1, it is set to change
\end{itemize}
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{images/fig6.pdf}
\caption{Sample images divided into unchange, small change, and large change for reconstruction loss analysis.}
\label{fig:fig6}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[width=\linewidth]{images/fig3.pdf}
\caption{Qualitative analysis of CDRL. The top 3 lines are the qualitative results of CDRL in the region where the change has occurred, and the bottom line is the qualitative result of CDRL in the unchanged pair. It can be seen that CDRL localizes the area where the change occurred.}
\label{fig:fig3}
\end{figure*}
\subsection{Loss Analysis Results}
\label{sec:toy}
We planned a loss analysis experiment to check whether the source image reconstructor has a high reconstruction loss in the part where the structure is changed a lot.
We divided the dataset into unchange, small change (the changed part is less than 30\% of the total image), and large change as shown in ~\figref{fig:fig6}.
When the dataset was split based on these criteria, the LEVIR-CD dataset was split into 8 unchanged pairs, 35 small change pairs, and 21 large change pairs.
The WHU dataset was split into 377 unchange pairs, 145 small change pairs, and 138 large change pairs.
\begin{table}
\centering
\resizebox{0.47\textwidth}{!}{
\begin{tabular}{c | c | c c c}
\toprule
Method & Dataset & \multicolumn{3}{c}{Reconstruction Loss} \\
& & Un & Small & Large \\
\midrule
CDRL (Aug) & LEVIR-CD& 24.79 & 36.81 & 44.91 \\
CDRL (Pseudo Unchange Pair) & LEVIR-CD & 14.45 & 34.74 & 43.49 \\
CDRL (Aug+Pseudo Unchange Pair) & LEVIR-CD& 10.15 & 31.41 & 38.53 \\
\hline
CDRL (Aug) & WHU& 35.95 & 41.11 & 50.89 \\
CDRL (Pseudo Unchange Pair) & WHU & 22.89 & 39.37 & 49.46 \\
CDRL (Aug+Pseudo Unchange Pair) & WHU& 17.20 & 38.03 & 47.65 \\
\bottomrule
\end{tabular}}
\caption{Loss analysis results of CDRL in LEVIR-CD dataset and WHU dataset.}
\label{tab:tab1}
\end{table}
~\tableref{tab:tab1} shows the results of loss analysis of CDRL in the LEVIR-CD test dataset and the WHU test dataset.
As shown in the table, the loss of the unchanged pair was the lowest in both the LEVIR-CD dataset and the WHU dataset, and the loss of the large change pair was the highest.
These experimental results indicate that the source image reconstructor is not good at reconstructing the source image when a pair with a large change in structure is input during the test because only pseudo unchanged pairs were input during training, as we intended.
Also, the fact that the loss was low in the unchanged pair indicates that our pseudo unchanged pair was generated at a level similar to that of the actual unchanged pair.
However, if the source image reconstructor works perfectly as we intended, the reconstruction loss should be close to 0 when an unchanged pair is input.
As shown in ~\tableref{tab:tab1}, the unchanged pair showed the lowest loss, but the value was not small.
The reason for this is that the LEVIR-CD dataset and the WHU dataset are labeled as only changes in building, and in fact, the unchanged pair includes many structural changes such as land becoming lakes, roads that did not exist, and cars.
A more detailed analysis result is described with an example in ~\sref{sec:qual}.
\begin{figure*}[t!]
\centering
\includegraphics[width=\linewidth]{images/fig7_2.pdf}
\caption{Qualitative analysis result of unchanged bi-temporal images for which CDRL showed high reconstruction loss. CDRL predicted that when roads, lakes, and cars were created, it was all change.
}
\label{fig:fig7}
\end{figure*}
\subsection{Pixel Level Change Detection Results}
\label{sec:pixel}
We compared and analyzed the performance of CDRL with other UCD and SCD methods.
In order to compare our CDRL with the existing UCD methods, we reproduced all ~\cite{li2019unsupervised, de2019unsupervised, li2019unsupervised} methods and tested them in LEVIR-CD and WHU.
As shown in ~\tableref{tab:ex2}, the existing UCD methods have high recall values and low precision because prediction results are very noisy and vulnerable to small structural changes.
These experimental results indicate that our CDRL is robust to small structural changes and also to style changes.
However, when compared with BIT, which is a state-of-the-art supervised change detection, the performance of CDRL was low due to a large gap. This reason is mainly analyzed for two reasons.
First, unlike supervised change detections, CDRL does not learn the pixel-level change area explicitly, so it can localize only the approximate location. Therefore, compared to BIT, our CDRL has similar recall value but clearly lower precision.
Second, supervised change detection can be explicitly trained on information about change objects of interest, so it can explicitly learn that a car is created or a lake is changed to be unchanged, but our CDRL predicts that they are all changed.
\begin{table}[t!]
\centering
\resizebox{0.47\textwidth}{!}{
\begin{tabular}{c| c | c | c c c}
\toprule
Method & Dataset & Supervision & Precision & Recall & IoU\\
\midrule
NMCD~\cite{li2019unsupervised} & LEVIR-CD & Unsup & 0.13 & 0.71 & 0.07\\
UCNN~\cite{de2019unsupervised} & LEVIR-CD & Unsup & 0.16 & 0.79 & 0.09\\
UCDGAN~\cite{li2019unsupervised} & LEVIR-CD & Unsup & 0.20 & 0.66 & 0.15\\
\textbf{CDRL} & LEVIR-CD & Unsup & \textbf{0.63} & \textbf{0.92} & \textbf{0.59}\\
BIT~\cite{chen2021remote} & LEVIR-CD & Sup & 0.89 & 0.89 & 0.80\\
\hline
NMCD~\cite{li2019unsupervised} & WHU & Unsup & 0.07 & 0.96 & 0.03\\
UCNN~\cite{de2019unsupervised} & WHU & Unsup & 0.07 & 0.95 & 0.03 \\
UCDGAN~\cite{li2019unsupervised} & WHU & Unsup & 0.09 & 0.93 & 0.08\\
\textbf{CDRL} & WHU & Unsup & \textbf{0.52} & \textbf{0.93} & \textbf{0.50}\\
BIT~\cite{chen2021remote} & WHU & Sup & 0.86 & 0.81 & 0.72 \\
\bottomrule
\end{tabular}}
\caption{Quantitative comparison results of CDRL and UCD, SCD methods. Note that, since there are no post-processing implementation details of the existing UCD methods, it was not applied.}
\label{tab:ex2}
\end{table}
\subsection{Patch Level Change Detection Results}
\label{sec:patch}
Many real-world applications that use change detectors do not rely solely on change detectors.
In these situations, the role of change detectors is to reduce human labor intensity by providing information on the patch or area where the change occurred among hundreds of patches.
Considering this application situation, we try to solve change detection with patch level classification.
\begin{table}[h!]
\centering
\begin{tabular}{c | c}
\toprule
Dataset & AUC \\
\midrule
LEVIR-CD & 83.52 \\
WHU & 87.18 \\
\bottomrule
\end{tabular}
\caption{Patch level binary classification results in the LEVIR-CD dataset and the WHU dataset.}
\label{tab:auc}
\end{table}
~\tableref{tab:auc} shows the patch level change detection results for our CDRL in the LEVIR-CD dataset and the WHU dataset.
As shown in the table, high AUC was achieved in both datasets despite using only single-temporal single source images.
Note that CVA-based methods predict that there is a change in all patches because the output result is noisy.
\subsection{Qualitative Results}
\label{sec:qual}
CDRL was qualitatively analyzed on the LEVIR-CD dataset.
As shown in ~\figref{fig:fig3}, it can be seen that the CDRL detects the changed part well because the reconstruction loss is high in the part where the structure change is large.
Also, for the unchanged pair, even if the style change is large, since there is no structure change, it can be seen that the reconstruction loss is low and no change is predicted
Therefore, it is thought that CDRL will be useful in applications where it is important not to localize the exact location, but to know the approximate location or whether or not changes have occurred in units of patches.
~\figref{fig:fig7} shows the qualitative analysis results for samples with poor CDRL performance
As shown in the figure, since CDRL cannot designate a specific change object of interest, it predicts that a change has occurred when a car is created, the ground becomes a lake, or a lake becomes the ground.
Therefore, the reason for the low IoU in all our experiments is dominant for the above reasons.
\subsection{Ablation Study}
\label{sec:able}
In order to compare and analyze the effect of the attention module and adversarial training constituting the CDRL, we conducted an ablation study on the LEVIR-CD dataset.
\paragraph{Attention Modules }
~\tableref{tab:abl2} shows the patch level classification results of CDRL in the LEVIR-CD dataset according to the existence of channels attention and spatial attention.
As shown in the table, performance was higher with CBAM than without CBAM. Also, in our proposed CDRL, the pseudo-pair CBAM, which provides channels attention to the $X^{t1}$ image and spatial attention to the $X^{t2}$ image, has the highest performance.
\begin{table}[h!]
\centering
\begin{tabular}{c| c}
\toprule
CDRL & AUC \\
\midrule
w/o Attention & 77.38 \\
w/ CBAM & 80.90 \\
w/ Pseudo-pair CBAM & 83.52 \\
\bottomrule
\end{tabular}
\caption{Patch level binary cleavage results of CDRL with or without attention module.}
\label{tab:abl2}
\end{table}
\paragraph{Adversarial Training}
We designed an experiment to check whether adversarial training solves the blurry problem of reconstruction images like ~\cite{isola2017image} in CDRL.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{images/fig11.pdf}
\caption{Qualitative comparison of reconstructed images with and without discriminator.}
\label{fig:fig10}
\end{figure}
~\figref{fig:fig10} is a sample of the reconstructed image with and without discriminator.
As shown in the figure, it can be seen that the reconstructed image is blurred when there is no discriminator. On the other hand, if there is a discriminator, it can be seen that the boundary is reconstructed more clearly.
These results indicate that adversarial training is also effective in CDRL.
~\tableref{tab:abl1} shows the results of measuring the AUC of CDRL with and without discriminator on the LEVIR-CD dataset. As shown in the table, it can be seen that higher AUC was achieved with the discriminator.
These experimental results show that the sensitivity of the CDRL to the structure of the pair image, like our design, helps the performance.
\section{Discussion and Future work}
In our work, we performed change detection with source image reconstruction loss using only unlabeled single-temporal single source images.
However, the semantic change we are interested in can exist in a variety of ways, such as buildings, seasons, cars, and trees.
The prerequisite for CDRL is to make the reconstruction loss appear high in the part where the structure change occurs regardless of the style change.
However, the semantic change that we are interested in can be diverse, such as natural scenery, artificial objects, weather, and environmental changes.
Therefore, in future work, based on the fact that CDRL has a significant performance improvement in UCD, we plan to study semi-supervised change detection to efficiently detect changes of interest (change of specific object)
\begin{table}
\centering
\begin{tabular}{c| c}
\toprule
CDRL & AUC \\
\midrule
w/o discriminator & 69.09 \\
w/ discriminator & 83.52 \\
\bottomrule
\end{tabular}
\caption{Patch level binary cleavage results of CDRL with or without discriminator.}
\label{tab:abl1}
\end{table}
\section{Conclusion}
In this paper, to solve the problem that it is difficult to construct a bi-temporal pair dataset containing semantic changes, we propose a CDRL that performs unsupervised change detection using only a single-temporal single source image.
In order to solve the unsupervised change detection problem as a reconstruction-based unsupervised anomaly detection problem, CDRL defined normal data as unchanged pairs and anomaly data as changed pairs.
After that, a change detector (reconstructor) that receives pair images was proposed.
We verified the CDRL on the WHU and LEVIR-CD datasets, and achieved significant performance despite unsupervised change detection using single-temporal single source images.
We hope that CDRL will be widely used in real-world scenarios where it is difficult to obtain labeled bi-temporal pair images.
\label{sec:con}
{\small
\bibliographystyle{ieee_fullname}
\section{Introduction}
After receiving paper reviews, authors may optionally submit a rebuttal to address the reviewers' comments, which will be limited to a {\bf one page} PDF file.
Please follow the steps and style guidelines outlined below for submitting your author response.
The author rebuttal is optional and, following similar guidelines to previous CVPR conferences, is meant to provide you with an opportunity to rebut factual errors or to supply additional information requested by the reviewers.
It is NOT intended to add new contributions (theorems, algorithms, experiments) that were absent in the original submission and NOT specifically requested by the reviewers.
You may optionally add a figure, graph, or proof to your rebuttal to better illustrate your answer to the reviewers' comments.
Per a passed 2018 PAMI-TC motion, reviewers should refrain from requesting significant additional experiments for the rebuttal or penalize for lack of additional experiments.
Authors should refrain from including new experimental results in the rebuttal, especially when not specifically requested to do so by the reviewers.
Authors may include figures with illustrations or comparison tables of results reported in the submission/supplemental material or in other papers.
Just like the original submission, the rebuttal must maintain anonymity and cannot include external links that reveal the author identity or circumvent the length restriction.
The rebuttal must comply with this template (the use of sections is not required, though it is recommended to structure the rebuttal for ease of reading).
\subsection{Response length}
Author responses must be no longer than 1 page in length including any references and figures.
Overlength responses will simply not be reviewed.
This includes responses where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide.
Note that this \LaTeX\ guide already sets figure captions and references in a smaller font.
\section{Formatting your Response}
{\bf Make sure to update the paper title and paper ID in the appropriate place in the tex file.}
All text must be in a two-column format.
The total allowable size of the text area is $6\frac78$ inches (17.46 cm) wide by $8\frac78$ inches (22.54 cm) high.
Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them.
The top margin should begin 1 inch (2.54 cm) from the top edge of the page.
The bottom margin should be $1\frac{1}{8}$ inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper;
for A4 paper, approximately $1\frac{5}{8}$ inches (4.13 cm) from the bottom edge of the page.
Please number any displayed equations.
It is important for readers to be able to refer to any particular equation.
Wherever Times is specified, Times Roman may also be used.
Main text should be in 10-point Times, single-spaced.
Section headings should be in 10 or 12 point Times.
All paragraphs should be indented 1 pica (approx.~$\frac{1}{6}$ inch or 0.422 cm).
Figure and table captions should be 9-point Roman type as in \cref{fig:onecol}.
List and number all bibliographical references in 9-point Times, single-spaced,
at the end of your response.
When referenced in the text, enclose the citation number in square brackets, for example~\cite{Alpher05}.
Where appropriate, include the name(s) of editors of referenced books.
\begin{figure}[t]
\centering
\fbox{\rule{0pt}{0.5in} \rule{0.9\linewidth}{0pt}}
\caption{Example of caption. It is set in Roman so that mathematics
(always set in Roman: $B \sin A = A \sin B$) may be included without an
ugly clash.}
\label{fig:onecol}
\end{figure}
To avoid ambiguities, it is best if the numbering for equations, figures, tables, and references in the author response does not overlap with that in the main paper (the reviewer may wonder if you talk about \cref{fig:onecol} in the author response or in the paper).
See \LaTeX\ template for a workaround.
\subsection{Illustrations, graphs, and photographs}
All graphics should be centered.
Please ensure that any point you wish to make is resolvable in a printed copy of the response.
Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print.
Readers (and reviewers), even of an electronic copy, may choose to print your response in order to read it.
You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic.
When placing figures in \LaTeX, it is almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below
{\small\begin{verbatim}
\usepackage{graphicx} ...
\includegraphics[width=0.8\linewidth]
{myfile.pdf}
\end{verbatim}
}
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
The concept of asymptotic expansion dates back to H. Poincar\'e in 1886, who gave an analytical meaning to the formal (generally divergent) power series solutions to some classical problems in optics or astronomy. Since then, its use has pervaded the theory of formal solutions to general classes of ordinary and partial differential equations, difference and $q$-difference equations, and so on.
For simplicity, we consider asymptotic expansions $\sum_{p=0}^\infty a_pz^p$ of functions $f$ analytic in unbounded sectors $S$ of the Riemann surface of the logarithm with vertex at 0. Although in its original formulation the remainders $f(z)-\sum_{p=0}^{n-1}a_pz^p$ were assumed to be just bounded for all $z\in S$ by a suitable constant (depending on $n$) times $|z|^n$, the asymptotic techniques have shown its real power when a precise rate of growth for the constants, in the form $CA^nM_n$ for $A,C>0$ and a sequence $(M_p)_{p=0}^\infty$ of positive real numbers, is either obtained from the concrete problem under study, or specified for theoretical considerations. In this way, ultraholomorphic classes of functions associated with the weight sequence $\mathbb{M}=(M_p)_{p=0}^\infty$ and the sector $S$ have been introduced, and it is important to decide about the injectivity and surjectivity of the (asymptotic) Borel map, sending a function in one of such classes into its series of asymptotic expansion. We refer the reader to the works~\cite{Mandelbrojt,Salinas,JimenezSanzSchindlInjectSurject} for
a complete solution of the injectivity problem. In the last of these references, the seminal results of V. Thilliez~\cite{Thilliez03}, proving surjectivity of the Borel map in a constructive way for strongly regular sequences $\mathbb{M}$ in sectors of opening less than $\pi\gamma(\mathbb{M})$, where $\gamma(\mathbb{M})$ is a growth index associated with $\mathbb{M}$, were shown to be optimal except possibly for the critical opening: if surjectivity occurs, the opening will be at most $\pi\gamma(\mathbb{M})$ (see Subsection~\ref{subsec.WeightSeq} for the notions related to weight sequences). Very recently, G. Schindl and the first and third authors~\cite{JimenezSanzSchindlSurjectDc} have obtained a similar result
for the more general case that $\widehat{\mathbb{M}}=(p!M_p)_{p=0}^\infty$ is a regular weight sequence in the sense of E.~M.~Dyn'kin~\cite{Dynkin80}, but now the proof is not constructive. Indeed, the key fact is due to A. Debrouwere~\cite{momentsdebrouwere}, who characterized the surjectivity of the Stieltjes moment problem in Gelfand-Shilov spaces defined by regular sequences by abstract functional-analytic techniques. The Fourier transform allows to transfer this information into the asymptotic framework in a halfplane, and in~\cite{JimenezSanzSchindlSurjectDc} Laplace and Borel transforms of arbitrary order extend the procedure to general sectors. In particular, the condition $\gamma(\mathbb{M})=\infty$ (we say the sequences satisfying this condition are of rapid growth, see Remark~\ref{rem.RapidGrowth}) is shown to amount to the surjectivity of the Borel map, and the existence of both local and global linear continuous right inverses for it, for unbounded sectors of arbitrary (finite) opening in the Riemann surface of the logarithm.
The main aim of this paper is to give a constructive proof of the surjectivity of the Borel map in sectors of the complex plane for the ultraholomorphic class associated with some specific sequences, $\mathbb{M}^{\tau,\sigma}=(p^{\tau p^{\sigma}})_{p=0}^\infty$, which are not strongly regular, but for which the sequence $\widehat{\mathbb{M}}$ is regular as long as $1<\sigma<2$. These sequences have been considered in a series of papers by S. Pilipovi{\'c}, N. Teofanov and F. Tomi{\'c}~\cite{ptt15,ptt16,ptt,ptt21}, inducing ultradifferentiable spaces of so-called extended Gevrey regularity. However, it is important to note that the estimates they impose on the elements of those classes are different, as their derivatives are controlled by expressions of the form $CA^{p^\sigma}p^{\tau p^{\sigma}}$ instead. This change of the geometric factor $A^{p}$ into the faster $A^{p^\sigma}$ entails some enhancements in the stability of the classes, but it is not well suited for our purposes, so we attach to the standard framework in ultraholomorphic classes. Our technique is similar to the one employed by S. Malek and the second and third authors~\cite{lastramaleksanzJMAA12} for reproving Thilliez's results, and by the third author in~\cite[Theorem\ 6.1]{SanzFlatProxOrder} for closing the surjectivity problem for strongly regular sequences admitting a nonzero proximate order. It rests on the construction of suitable kernel functions allowing us to define formal Borel- and truncated Laplace-like transforms, in terms of which the solution is provided. These kernels are constructed using the Lambert function and it is possible to prove that they have good behavior thanks to its properties. Regrettably, the multivalued character of the Lambert function, that appears in our procedure in a crucial way, causes some difficulties in order to reason in sectors of opening greater than $2\pi$; for such openings, although surjectivity of the Borel map is known to hold, our constructive proof is not available.
\section{Preliminaries}
\subsection{Notation} Throughout the paper, we use the following notation: $\mathbb{N}:=\{1,2,\dots \}$, $\mathbb{N}_0=\mathbb{N}\cup\{0\}$, $\mathbb{C}[[z]]$ is the space of formal power series in $z$ with complex coefficients.
We set $\mathcal{R}$ for the Riemann surface of the logarithm, and the notation $z=|z|e^{i\theta}$ refers to the element $(|z|,\theta)\in(0,\infty)\times\mathbb{R}$. For $\gamma>0$ we consider the unbounded sectors
$$S_{\gamma}:=\{z\in \mathcal{R} : |\arg(z)|<\frac{\gamma\pi}{2}\}.$$
For every $0<\gamma<2$ we identify $S_\gamma$ with the corresponding sector in $\mathbb{C}$. For $R>0$, $\gamma>0$, we write
$$L_{R,\gamma}:=\{z\in \mathcal{R} :|z|>R,|\text{arg} (z)|<\frac{\gamma\pi}{2}\}.$$
Given two unbounded sectors $T,S$, we write $T\prec S$ in the case that $\overline{T}\subseteq S\cup\{0\}$. For every $R>0$ and $z_0\in\mathbb{C}$, $D(z_0,R)$ stands for the open disc in $\mathbb{C}$ centered at $z_0$ and radius$~R$.
$\mathop{\rm Log}\nolimits(\cdot)$ stands for the principal branch of the complex logarithm, i.e., $\hbox{arg}(\mathop{\rm Log}\nolimits(z))\in(-\pi,\pi)$ for every $z\in\mathbb{C}\setminus(-\infty,0]$. We write $\ln(\cdot)$ for the natural logarithm, and for $x>0$, $\ln^+(x)$ denotes the value $\max(\ln(x),0)$.
\subsection{Weight sequences}\label{subsec.WeightSeq} In what follows, $\mathbb{M}=(M_p)_{p\in\mathbb{N}_0}$ will stand for a sequence of positive real numbers with $M_0=1$. The following properties will play a role in this paper:
\begin{itemize}
\item[(i)] $\mathbb{M}$ is \emph{logarithmically convex} (for short, (lc)) if
$$M_{p}^{2}\le M_{p-1}M_{p+1},\qquad p\in\mathbb{N}.$$%
\item[(ii)] $\mathbb{M}$ is \emph{stable under differential operators} or satisfies the \emph{derivation closedness condition} (briefly, (dc)) if there exists $D>0$ such that $$M_{p+1}\leq D^{p+1} M_{p}, \qquad p\in\mathbb{N}_{0}. $$
\item[(iii)] $\mathbb{M}$ is of, or has, \emph{moderate growth} (briefly, (mg)) whenever there exists $A>0$ such that
$$M_{p+q}\le A^{p+q}M_{p}M_{q},\qquad p,q\in\mathbb{N}_0.$$
\item[(iv)] $\mathbb{M}$ satisfies the condition (snq) if there exists $B>0$ such that
$$
\sum^\infty_{q= p}\frac{M_{q}}{(q+1)M_{q+1}}\le B\frac{M_{p}}{M_{p+1}},\qquad p\in\mathbb{N}_0.$$
\end{itemize}
In the classical work of H.~Komatsu~\cite{komatsu}, the properties (lc), (dc) and (mg) are denoted by $(M.1)$, $(M.2)'$ and $(M.2)$, respectively. The property (snq) for $\mathbb{M}$ is precisely the property $(M.3)$ when imposed on the sequence $\widehat{\mathbb{M}}:=(p!M_p)_{p=0}^\infty$.
Obviously, (mg) implies (dc).
If $\mathbb{M}$ is (lc), it is well-known that $((M_p)^{1/p})_{p\in\mathbb{N}}$ is nondecreasing; if moreover $(M_p)^{1/p}\uparrow\infty$, we say $\mathbb{M}$ is a \emph{weight sequence}.
Following E.~M.~Dyn'kin~\cite{Dynkin80}, if $\mathbb{M}$ is a weight sequence and satisfies (dc), we say $\widehat{\mathbb{M}}$ is \emph{regular}. According to V.~Thilliez~\cite{Thilliez03}, if $\mathbb{M}$ satisfies (lc), (mg) and (snq), we say $\mathbb{M}$ is \emph{strongly regular}; in this case $\mathbb{M}$ is a weight sequence, and the corresponding $\widehat{\mathbb{M}}$ is regular.
The index $\gamma(\mathbb{M})$ was introduced by V.~Thilliez~\cite[Sect.\ 1.3]{Thilliez03} for strongly regular sequences $\mathbb{M}$, in which case it is a positive real number, but his definition makes sense for (lc) sequences.
A sequence $(c_p)_{p\in\mathbb{N}_0}$ is \emph{almost increasing} if there exists $a>0$ such that for every $p\in\mathbb{N}_0$ we have that $c_p\leq a c_q $ for every $ q\geq p$; then, it was proved in~\cite{JimenezSanzSRSPO,JimenezSanzSchindlIndex} that for any weight sequence $\mathbb{M}$ one has
\begin{equation}\label{equa.indice.gammaM.casicrec}
\gamma(\mathbb{M})=\sup\{\gamma>0: \big(\frac{M_{p+1}}{M_p(p+1)^\gamma}\big)_{p\in\mathbb{N}_0}\hbox{ is almost increasing} \}\in[0,\infty].
\end{equation}
Whenever $\widehat{\mathbb{M}}$ is (lc) we have (see~\cite[Ch.~2]{PhDJimenez} and \cite[Cor.~3.13]{JimenezSanzSchindlIndex}) that
$\gamma(\mathbb{M})>0$ if and only if $\mathbb{M}$ is (snq).
Two sequences $\mathbb{M}=(M_{p})_{p\in\mathbb{N}_0}$ and $\mathbb{L}=(L_{p})_{p\in\mathbb{N}_0}$ of positive real numbers are said to be \emph{equivalent}, and we write $\mathbb{M}\approx\mathbb{L}$, if there exist positive constants $A,B$ such that
$$A^pM_p\le L_p\le B^pM_p,\qquad p\in\mathbb{N}_0.$$
\subsection{Ultraholomorphic classes}\label{subsec.UltraHolClasses}
In this section $S$ is a sector and $\mathbb{M}$ a sequence.
We say a holomorphic function $f:S\to\mathbb{C}$ admits the formal power series $\widehat{f}=\sum_{p=0}^{\infty} c_{p}z^{p}\in\mathbb{C}[[z]]$ as its \emph{uniform $\{\mathbb{M}\}$-asymptotic expansion in $S$ (of type $1/A$ for some $A>0$)} if there exists $C>0$ such that for every $p\in\mathbb{N}_0$, one has
\begin{equation}\Big|f(z)-\sum_{p=0}^{n-1}c_pz^p \Big|\le CA^nM_{n}|z|^n,\qquad z\in S.\label{desarasintunifo}
\end{equation}
In this case we write $f\sim_{\{\mathbb{M}\},A}^u\widehat{f}$ in $S$, and $\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\},A}(S)$ denotes the space of functions admitting uniform $\{\mathbb{M}\}$-asymptotic expansion of type $1/A$ in $S$, endowed with the norm
$$
\left\|f\right\|_{\mathbb{M},A,\overset{\sim}{u}}:=\sup_{z\in S,n\in\mathbb{N}_{0}}\frac{|f(z)-\sum_{p=0}^{n-1}c_pz^p|}{A^{n}M_{n}|z|^n},
$$
which makes it a Banach space. We set $\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\}}(S)$ for the $(LB)$ space of functions admitting uniform $\{\mathbb{M}\}$-asymptotic expansion in $S$, obtained as the union of the previous classes when $A$ runs over $(0,\infty)$.
When the type needs not be specified, we simply write $f\sim_{\{\mathbb{M}\}}^u\widehat{f}$ in $S$.
Note that, taking $n=0$ in~\eqref{desarasintunifo}, we deduce that every function in $\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\}}(S)$ is a bounded function.
One may accordingly define classes of formal power series
$$\mathbb{C}[[z]]_{\{\mathbb{M}\},A}=\Big\{\widehat{f}=\sum_{p=0}^\infty c_pz^p\in\mathbb{C}[[z]]:\, \left|\,\widehat{f} \,\right|_{\mathbb{M},A}:=\sup_{p\in\mathbb{N}_{0}}\displaystyle \frac{|c_{p}|}{A^{p}M_{p}}<\infty\Big\}.$$
$(\mathbb{C}[[z]]_{\{\mathbb{M}\},A},\left| \ \right|_{\mathbb{M},A})$ is a Banach space and we put $\mathbb{C}[[z]]_{\{\mathbb{M}\}}:=\cup_{A>0}\mathbb{C}[[z]]_{\{\mathbb{M}\},A}$, again an $(LB)$ space.
Given $f\in\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\}}(S)$ with $f\sim_{\{\mathbb{M}\}}^u\widehat{f}$, Cauchy's integral formula for the derivatives implies that for every $T\prec S$ there exists $A_T>0$ such that
\begin{equation}\label{eq.BoundsDerivUltraholClass}
\sup_{z\in T,p\in\mathbb{N}_{0}}\frac{|f^{(p)}(z)|}{A_T^{p}p!M_{p}}<\infty,
\end{equation}
and then, by Taylor's formula, for every such $T$ and every $p\in\mathbb{N}_0$ one has
\begin{equation*}
c_p=\lim_{ \genfrac{}{}{0pt}{}{z\to0}{z\in T}} \frac{f^{(p)}(z)}{p!},
\end{equation*}
So, we can set ${f^{(p)}(0)}:=p!c_p$, and it is straightforward that $\widehat{f}\in\mathbb{C}[[z]]_{\{\mathbb{M}\}}$, what makes natural to consider the \textit{asymptotic Borel map}
$$
\widetilde{\mathcal{B}}:\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\}}(S)\longrightarrow \mathbb{C}[[z]]_{\{\mathbb{M}\}}$$
sending a function $f\in\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\}}(S)$ into its uniform $\{\mathbb{M}\}$-asymptotic expansion $\widehat{f}$ in $S$.
$\widetilde{\mathcal{B}}$ is continuous when considered between these $(LB)$ spaces, and also when restricted to the Banach spaces with fixed type.
Finally, note that if $\mathbb{M}\approx\mathbb{L}$, then $\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\}}(S)=\widetilde{\mathcal{A}}^u_{\{\mathbb{L}\}}(S)$ and $\mathbb{C}[[z]]_{\{\mathbb{M}\}}=\mathbb{C}[[z]]_{\{\mathbb{L}\}}$, so the corresponding Borel maps are in all cases identical.
\begin{rem}
The inequalities mentioned in~\eqref{eq.BoundsDerivUltraholClass} suggest considering ultraholomorphic classes defined by imposing such estimates on the derivatives of its elements $f$, and the corresponding Borel map sending $f$ into $(f^{(p)}(0))_{p\in\mathbb{N}_0}$, where $f^{(p)}(0)$ is defined as above. Another possibility consists in regarding non-uniform asymptotics, with bounds $C_TA_T^pM_p|z|^p$ for the remainders on every bounded and proper subsector $T$ of $S$. Our forthcoming results can be stated in these frameworks, but we will not enter into details. The interested reader may consult~\cite{JimenezSanzSchindlInjectSurject,JimenezSanzSchindlSurjectDc}.
\end{rem}
\subsection{A family of sequences of rapid growth}
Let $\sigma>1$ and $\tau>0$ be real numbers. We consider the sequence of positive real numbers $\mathbb{M}^{\tau,\sigma}=(M_p^{\tau,\sigma})_{p\in\mathbb{N}_0}$ with $M_p^{\tau,\sigma}:=p^{\tau p^{\sigma}}$ for $p\ge1$ and $M^{\tau,\sigma}_0=1$.
First we recall the essential properties of the sequences $\mathbb{M}^{\tau,\sigma}$.
We refer to \cite{ptt15} for the proof of the next Lemma.
\begin{lemma} Let $\sigma>1$ and $\tau>0$ be real numbers. Then
the following properties hold:
\begin{enumerate}
\item[(i)] $\mathbb{M}^{\tau,\sigma}$ is (lc).
\item[(ii)] $\overline{(M.2)'}$ For all $q\in\mathbb{N}_{0}$ there exists $C_q\geq 1$ such that
$$M^{\tau,\sigma}_{p+q}\leq C_q^{p^\sigma} M^{\tau,\sigma}_{p}, \qquad p,q\in\mathbb{N}_{0}. $$
\item[(iii)] $\overline{(M.2)}$ There exists $A>0$ such that
$$M^{\tau,\sigma}_{p+q}\le A^{p^\sigma +q^\sigma}M^{\tau 2^{\sigma-1},\sigma}_{p}M^{\tau 2^{\sigma-1},\sigma}_{q},\qquad p,q\in\mathbb{N}_0.$$
\end{enumerate}
\end{lemma}
We note that the notation $\overline{(M.2)'}$, resp. $\overline{(M.2)}$, comes from~\cite{ptt}, while in~\cite{ptt15,ptt16,ptt21} $\widetilde{(M.2)'}$, resp. $\widetilde{(M.2)}$, is used.
These conditions $\overline{(M.2)'}$ and $\overline{(M.2)}$ are different from the classical ones, namely (dc) and (mg), appearing in the literature when dealing with Carleman-like classes (see~\cite{komatsu}). They play a prominent role in the study of the corresponding ultradifferentiable and ultradistributional classes carried out by S. Pilipovi{\'c}, N. Teofanov and F. Tomi{\'c} in~\cite{ptt15,ptt16,ptt,ptt21}, as they allow for a precise control of the flexibility obtained by introducing the two parameter dependence.
Accordingly, they introduce a convenient modification of the classical associated function by setting (see~\cite[Definition~2.1]{ptt})
$$T_{\tau,\sigma,h}(t):= \sup_{p\in\mathbb{N}} \ln^+\left(\frac{h^{p^{\sigma}}t^{p}}{M_{p}^{\tau,\sigma}}\right),\qquad t>0,$$
and $T_{\tau,\sigma,h}(0)=0$. We write $T_{\tau,\sigma}$ for the function $T_{\tau,\sigma,1}$, which is the standard function associated with the sequence $\mathbb{M}^{\tau,\sigma}$ in the literature (see~\cite[Definition~3.1]{komatsu}).
\begin{rem}\label{rem.RapidGrowth}
We note that, from the expression~\eqref{equa.indice.gammaM.casicrec}, it is easy to check that $\gamma(\mathbb{M}^{\tau,\sigma})=+\infty$ for every $\sigma>1$, so we say these sequences are of rapid growth. The terminology comes from the theory of regular variation, see~\cite{bgt} for the case of functions and~\cite{DjurcicKocinacZizovic07,JimenezSanzSchindlIndex} for some extensions of the theory for the case of sequences. Although there are different definitions of the notion of rapid variation, they all coincide when we restrict ourselves to nondecreasing sequences, see \cite{DjurcicElezKocinac15}. Then one can say that a nondecreasing sequence is of rapid variation if its lower Matuszewska index is $+\infty$. It turns out~\cite[Section~3.3]{JimenezSanzSchindlIndex} that, for a general sequence of positive real numbers $\mathbb{M}=(M_p)_{p\in\mathbb{N}_0}$, the index $\gamma(\mathbb{M})$ equals the lower Matuszewska index of the sequence of quotients $(M_{p+1}/M_p)_{p\in\mathbb{N}_0}$, which is nondecreasing if $\mathbb{M}$ is (lc); so, the sequence $(M^{\tau,\sigma}_{p+1}/M^{\tau,\sigma}_p)_{p\in\mathbb{N}_0}$ is of rapid variation, what justifies the previous nomenclature.
As mentioned before, the rapid growth of $\mathbb{M}^{\tau,\sigma}$ implies (snq) is satisfied, while (mg) is not, hence $\mathbb{M}^{\tau,\sigma}$ is not strongly regular for any $\sigma>1$. However, whenever $\sigma$ belongs to the interval $(1,2)$, the sequence $(M_p^{\tau,\sigma})_{p\in\mathbb{N}_0}$ not only satisfies $\overline{(M.2)'}$ but also the property (dc), i.e., there exists $C>0$ such that $M_{p+1}^{\tau,\sigma}\le C^{p+1}M_{p}^{\tau,\sigma}$. Indeed, for logarithmically convex sequences $(M_p)_{p\in\mathbb{N}_0}$, (dc) is equivalent to the condition $\log(M_p)=O(p^2)$, $p\to\infty$, which clearly holds for $\mathbb{M}^{\tau,\sigma}$ if, and only if, $\sigma\in(1,2)$. Hence, for such values $\widehat{\mathbb{M}}^{\tau,\sigma}$ is regular. This fact will make a difference with respect to the general case. Moreover, it implies the following well-known estimate: if one defines the auxiliary function
$$
h_{\tau,\sigma}(t):=e^{-T_{\tau,\sigma}(1/t)}=\inf_{p\ge 0}M_{p}^{\tau,\sigma}t^p,\quad t>0;\ h_{\tau,\sigma}(0)=0,
$$
for every $t>0$ one has
\begin{equation}
h_{\tau,\sigma}(t)\le
\inf_{p\ge 1}M_{p}^{\tau,\sigma}t^p=\inf_{p\ge 0}M_{p+1}^{\tau,\sigma}t^{p+1}\le Ct\inf_{p\ge 0}M_{p}^{\tau,\sigma}(Ct)^p=Ct h_{\tau,\sigma}(Ct).\label{eq.Estimate_h_tausigma_dc}
\end{equation}
\end{rem}
\section{On the Lambert function $W$}
In this section, and for the sake of completeness, we give some information about the Lambert function $W$ and its main properties to be considered in the present study. We refer to~\cite{corless} for further details.
The Lambert function $W$ is defined as the complex function which satisfies
\begin{equation}\label{eW}
W(z)e^{W(z)}=z.
\end{equation}
It holds that the Lambert function $W$ is a multivalued function which splits the $w=W(z)$ plane into an infinite number of regions. The principal branch of the Lambert function $W$, usually denoted by $W_0$, is defined in $\mathbb{C}\setminus (-\infty,-e^{-1}]$, and the curve in $\mathbb{R}^2$ defining the boundary of its image set $W_0(\mathbb{C}\setminus (-\infty,-e^{-1}])$ is given by
$$\{(-t\cot(t),t)\in\mathbb{R}^2:-\pi<t<\pi\}.$$
$W_0$ is a holomorphic and bijective map when restricted to the previous domains, with $W_0(0)=0$, being the origin the only value of its domain in which $W$ vanishes. Hereinafter, we will restrict our reasonings to the function $W_0$, so we will write $W$ instead of $W_0$ from now on.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{fig1.png}
\caption{Principal branch of the Lambert function $W$}
\end{figure}
\begin{prop}\label{propW}
The following properties hold for the principal branch of the Lambert function $W$:
\begin{enumerate}
\item[(a)] $W(0)=0$. More precisely, $z=0$ is a zero of order 1, and indeed $\displaystyle\lim_{z\to 0}W(z)/z=1$.
\item[(b)] $W(x)\in\mathbb{R}$ for $x>-e^{-1}$, and $\lim_{x\to\infty}W(x)=+\infty$. In addition to this, one has that $W(x)>0$ for $x>0$.
\item[(c)] For every $z\in\mathbb{C}\setminus(-\infty,-e^{-1}]$ one has
$$W'(z)=e^{-W(z)}\frac{1}{1+W(z)}.$$
\item[(d)] For every $z\in\mathbb{C}\setminus(-\infty,-e^{-1}]$ it holds that
$$\Re(z)=e^{\beta_1(z)}(\beta_1(z)\cos(\beta_2(z))-\beta_2(z)\sin(\beta_2(z))),$$
$$\mathop{\rm Im}\nolimits(z)=e^{\beta_1(z)}(\beta_2(z)\cos(\beta_2(z))+\beta_1(z)\sin(\beta_2(z))),$$
where $\beta_1(z)=\Re(W(z))$ and $\beta_2(z)=\mathop{\rm Im}\nolimits(W(z))$.
\item[(e)] Let $R>0$ and $\alpha\in (0,2)$. For $L_{R,\alpha}=S_{\alpha}\cap (\mathbb{C}\setminus D(0,R))$ one has
\begin{equation}\label{eq.imageunderW_kofasector_bis}
W(L_{R,\alpha})\subseteq\{\xi+i\eta \in\mathbb{C}: \exp(\xi)\sqrt{\xi^2+\eta^2}\ge R\}.
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof1}
(a) is a direct consequence of (\ref{eW}), from where the Taylor expansion of the Lambert function $W$ at the origin of the complex plane can be obtained, see~\cite{corless}. (b) also stems from (\ref{eW}), taking into account that $W$, when restricted to the interval $(-1,\infty)$, is the inverse function of $x\mapsto xe^x$. Differentiating (\ref{eW}) one arrives at an expression equivalent to (c). Observe that $W(z)\neq-1$ for $z\in\mathbb{C}\setminus(-\infty,-e^{-1}]$. (d) and (e) are obtained by splitting (\ref{eW}) into its real and imaginary parts.
\end{proof1}
\begin{rem}\label{remWslow}
It is worth remarking that property (e) in the previous proposition guarantees that for every $R>0$ and $\alpha\in(0,2)$,
\begin{equation}\label{e119}
\lim_{|z|\to\infty}\frac{zW'(z)}{W(z)}=\lim_{|z|\to\infty}\frac{1}{1+W(z)}=0,
\end{equation}
uniformly in $L_{R,\alpha}$. Indeed, given $M>0$ we may choose $R_0\ge Me^M$. For every $z\in L_{R_0,\alpha}$
put $W(z)=\xi +i\eta$. There are two possibilities: if $\xi\ge M$ then obviously $|W(z)|\ge M$; on the contrary, \eqref{eq.imageunderW_kofasector_bis} gives $|W(z)|=(\xi^2+\eta^2)^{1/2}\ge R_0e^{-\xi}>Me^Me^{-M}=M$, as desired.
Property (\ref{e119}) states that $W$ is slowly varying (see~\cite[Theorem~$A.1.2.(a)$]{bgt}) in $L_{R,\alpha}$. This statement is also true for any other choice of the branch of the Lambert function.
\end{rem}
\section{Kernel functions for sequences of rapid growth}
In this section, we construct a kernel function associated with the sequence $\mathbb{M}^{\tau,\sigma}$ which will be useful for our purposes.
\begin{defi}\label{defib}
For $\sigma>1$ and $\tau>0$, we define
\begin{equation}\label{e177}
a_{\tau,\sigma}:= \left(\frac{\sigma-1}{\tau\sigma}\right)^{\frac{1}{\sigma-1}}, \qquad b_{\tau,\sigma}:=e^{\frac{\sigma-1}{\sigma}} \frac{\sigma-1}{\tau\sigma}.
\end{equation}
and
$$
e_{\tau,\sigma}(z):=z\exp\left( -a_{\tau,\sigma} W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1)) \mathop{\rm Log}\nolimits^{\frac{\sigma}{\sigma-1}}(z+1)\right)
$$
for every $z\in\mathbb{C}\setminus(-\infty,0]$.
\end{defi}
For the sake of brevity, we write
\begin{equation}\label{e351}
g(z):=W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z)) \mathop{\rm Log}\nolimits^{\frac{\sigma}{\sigma-1}}(z).
\end{equation}
so $e_{\tau,\sigma}(z)=z\exp(-a_{\tau,\sigma}g(z+1))$.
The powers of the logarithm and of the Lambert function appearing in $e_{\tau,\sigma}$ and $g$ have their principal values, i.e., they are the ones associated with the principal branch of the logarithm.
\begin{prop}\label{prop347}
$e_{\tau,\sigma}(z)$ is a holomorphic function in $\mathbb{C}\setminus(-\infty,0]$.
\end{prop}
\begin{proof1}
On the one hand, if $z\in\mathbb{C}\setminus(-\infty,0]$, then $z+1\in\mathbb{C}\setminus(-\infty,1]$, and so
$\mathop{\rm Log}\nolimits(z+1)=\ln(|z+1|)+i\arg(z+1)\notin(-\infty,0]$. Since $b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1)$ differs from $\mathop{\rm Log}\nolimits(z+1)$ by a positive factor, this fact ensures that the maps
\begin{align*}
z&\mapsto \mathop{\rm Log}\nolimits^{\frac{\sigma}{\sigma-1}}(z+1)= \exp\Big(\frac{\sigma}{\sigma-1}\mathop{\rm Log}\nolimits\big(\mathop{\rm Log}\nolimits(z+1)\big)\Big),\\
z&\mapsto W(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1)),
\end{align*}
are both holomorphic in $\mathbb{C}\setminus(-\infty,0]$. Finally, observe that the principal branch of the Lambert function $W(w)$ only takes values in $(-\infty,0]$ when $w$ runs over the real interval $(-1/e,0]$. Since $w=b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1)\notin(-\infty,0]$ whenever $z\in\mathbb{C}\setminus(-\infty,0]$, the composition of $z\mapsto W(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1))$ with the principal branch of the logarithm is again holomorphic, and so is
$$
z\mapsto W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1)),
$$
what leads to the conclusion.
\end{proof1}
\begin{rem}
There is some freedom in the choice of $e_{\tau,\sigma}$. The factor $z$ may be changed into any $z^{\alpha}$ for some positive real number $\alpha$ (so that the assertion $(i)$ in Lemma~\ref{lema.Properties.e_h} holds true), where the principal branch of the power is considered. Our choice tries to make the following computations simpler.
\end{rem}
\begin{rem}\label{rem.e_h_boundedat0}
Taking into account property (a) of Proposition~\ref{propW}, one has that
\begin{align*}
\lim_{z\to 0}\frac{e_{\tau,\sigma}(z)}{z}&=\lim_{z\to 0}\exp\left(-\left(\frac{\sigma-1}{\tau\sigma}\right)^{\frac{1}{\sigma-1}} W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1))\mathop{\rm Log}\nolimits^{\frac{\sigma}{\sigma-1}}(z+1)\right)\\
&=\lim_{z\to 0}\exp\left(-\left(\frac{\sigma-1}{\tau\sigma}\right)^{\frac{1}{\sigma-1}} \frac{\mathop{\rm Log}\nolimits^{\frac{\sigma}{\sigma-1}}(z+1)}{(b_{\tau,\sigma}\mathop{\rm Log}\nolimits(z+1))^{\frac{1}{\sigma-1}}}\right)\\
&=\lim_{z\to 0}\exp\left(-\left(\frac{\sigma-1}{\tau\sigma}\right)^{\frac{1}{\sigma-1}} \frac{\mathop{\rm Log}\nolimits^{\frac{\sigma}{\sigma-1}}(z+1)}{e^{\frac{1}{\sigma}} \left(\frac{\sigma-1}{\tau\sigma}\right)^{\frac{1}{\sigma-1}} \mathop{\rm Log}\nolimits^{\frac{1}{\sigma-1}}(z+1)}\right)\\
&=\lim_{z\to 0}\exp\left(-e^{-\frac{1}{\sigma}}
\mathop{\rm Log}\nolimits(z+1)\right)=1.
\end{align*}
Since the function $e_{\tau,\sigma}(z)/z$ never vanishes, we deduce that it will remain bounded, and bounded away from 0, in bounded proper sectors of $\mathbb{C}\setminus(-\infty,0]$ with vertex at 0.
\end{rem}
\begin{lemma}\label{lema.g_hincreasing}
The restriction of the function $g$ in \eqref{e351} to $(1,\infty)$, given by
\begin{equation}\label{eq.defi.g_h_reals}
g(x)=W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\ln(x)) \ln^{\frac{\sigma}{\sigma-1}}(x),
\end{equation}
is positive and strictly increasing. Moreover, $\lim_{x\to\infty}g'(x)=0$.
\end{lemma}
\begin{proof1}
We clearly have $b_{\tau,\sigma}\ln(x)>0$ for every $x>1$, and $W$ is positive for positive values of its argument (see (b) in Proposition~\ref{propW}), hence $g(x)>0$ for every $x>1$. Property (c) of Proposition~\ref{propW} yields
\begin{equation}\label{eq.derivW}
W'(x)=\frac{1}{x+e^{W(x)}}=\frac{W(x)}{x(W(x)+1)}.
\end{equation}
So, applying~\eqref{eq.derivW} we can write
\begin{align*}
g'(x)&=-\frac{1}{\sigma-1}W^{-\frac{\sigma}{\sigma-1}}(b_{\tau,\sigma}\ln(x)) W'(b_{\tau,\sigma}\ln(x))b_{\tau,\sigma}\frac{1}{x}\ln^{\frac{\sigma}{\sigma-1}}(x)\\
&+W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\ln(x))\frac{\sigma}{\sigma-1} \ln^{\frac{1}{\sigma-1}}(x)\frac{1}{x}\\
&=-\frac{1}{\sigma-1}W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\ln(x)) \frac{1}{W(b_{\tau,\sigma}\ln(x))+1}\frac{1}{x}\ln^{\frac{1}{\sigma-1}}(x)\\
&+W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\ln(x))\frac{\sigma}{\sigma-1} \ln^{\frac{1}{\sigma-1}}(x)\frac{1}{x}\\
&=\frac{1}{\sigma-1}W^{-\frac{1}{\sigma-1}}(b_{\tau,\sigma}\ln(x)) \ln^{\frac{1}{\sigma-1}}(x)\frac{1}{x} \left(\sigma-\frac{1}{W(b_{\tau,\sigma}\ln(x))+1}\right)>0,
\end{align*}
since $x>1$ and $\sigma>1$. So, $g$ is strictly increasing in $(1,\infty)$. If we observe now that $\lim_{x\to\infty}b_{\tau,\sigma}\ln(x)=+\infty$ and (b) in Proposition~\ref{propW}, it is clear from the last expression that $\lim_{x\to\infty}g'(x)=0$.
\end{proof1}
The next technical result provides information on the function $e_{\tau,\sigma}$ with respect to the standard properties satisfied by kernel functions for summability, as described in~\cite{lastramaleksanzJMAA15}, and which were inspired by the work of W. Balser~\cite[Section~5.5]{balser} in the context of Gevrey classes.
\begin{lemma}\label{lema.Properties.e_h}
The function $e_{\tau,\sigma}(\cdot)$ enjoys the following properties:
\begin{itemize}
\item[1.] $z^{-1}e_{\tau,\sigma}(z)$ is integrable at the origin, it is to say, for every $t_0>0$ and $\theta\in(-\pi,\pi)$, the integral
$$\int_0^{t_0}t^{-1}|e_{\tau,\sigma}(te^{i\theta})|dt<\infty.$$
\item[2.] For every $x>0$ the value of $e_{\tau,\sigma}(x)$ is positive real.
\item[3.] For every $T\prec S_{2}$ there exist $C_1,K_1,C_2,K_2>0$ (depending on $\tau,\sigma$) such that
\begin{equation}\label{eq.Bounds_e_h_subsectors_real_axis}
C_1e_{\tau,\sigma}(K_1|z|)\le |e_{\tau,\sigma}(z)|\le C_2e_{\tau,\sigma}(K_2|z|),\quad z\in T.
\end{equation}
\end{itemize}
\end{lemma}
\begin{proof1}
\noindent 1. By the very definition of the function $e_{\tau,\sigma}$, it is evident that the function $z^{-1}e_{\tau,\sigma}(z)$ is measurable and bounded in every segment towards the origin, and so it is also integrable.
\noindent 2. Recall that the function $g$ in (\ref{eq.defi.g_h_reals}) is the restriction of the function defined in (\ref{e351}) to $(1,\infty)$. Then it suffices to observe that the value $g(x+1)$ is real for every $x>0$, see Lemma~\ref{lema.g_hincreasing}.
\noindent 3. Throughout this part of the proof, we make use of property (e) in Proposition~\ref{propW} to take control of the image of the elements contained in a sector of the complex plane which stay away from the origin via Lambert function.
The constants $a_{\tau,\sigma}$ and $b_{\tau,\sigma}$, defined in (\ref{e177}), will be named $a$ and $b$ throughout this proof for brevity.
For $\omega\neq 0$ one can apply (\ref{eW}) to $z=b\omega$ from where we deduce
$$
\frac{\omega^{\frac{\sigma}{\sigma-1}}}{W^{\frac{1}{\sigma-1}}(b\omega)}=
\frac{1}{b^{1/(\sigma-1)}}\omega e^{\frac{1}{\sigma-1}W(b\omega)}.
$$
So, putting $\omega=\mathop{\rm Log}\nolimits(z+1)$ we obtain
$$
g(z+1)=\frac{1}{b^{1/(\sigma-1)}}\mathop{\rm Log}\nolimits(z+1) e^{\frac{1}{\sigma-1}W(b\mathop{\rm Log}\nolimits(z+1))}.
$$
According to this expression, we will start by showing that given $\beta\in(0,2)$, there exist $R,K>0$ such that for every $z\in S_{\beta}$ with $|z|\ge R$ one has
$$
\Re\Big(\mathop{\rm Log}\nolimits(z+1) e^{\frac{1}{\sigma-1}W(b\mathop{\rm Log}\nolimits(z+1))}\Big)\le
\ln(K|z|+1) e^{\frac{1}{\sigma-1}W(b\ln(K|z|+1))};
$$
equivalently, if we write
\begin{equation}\label{eq.RealImaginaryPartsLambert}
W(b\mathop{\rm Log}\nolimits(z+1))=\xi(z)+i\eta(z),\quad \xi(z),\eta(z)\in\mathbb{R},
\end{equation}
one has
\begin{multline}\label{eq.IneqFineAdjustComplexLambert}
\ln(|z+1|) e^{\frac{1}{\sigma-1}\xi(z)} \cos\big(\frac{1}{\sigma-1}\eta(z)\big)-
\arg(z+1)e^{\frac{1}{\sigma-1}\xi(z)} \sin\big(\frac{1}{\sigma-1}\eta(z)\big)\\
\le
\ln(K|z|+1) e^{\frac{1}{\sigma-1}W(b\ln(K|z|+1))},
\end{multline}
where $W$ is the principal branch for the Lambert function. By taking exponentials in the last inequality, we obtain the first inequality in~\eqref{eq.Bounds_e_h_subsectors_real_axis}, if we take into account that the function
$$
\frac{|e_{\tau,\sigma}(z)|}{e_{\tau,\sigma}(K|z|)}= \frac{\exp(-a\Re(g(z+1)))}{K\exp(-ag(K|z|+1))}
$$
is bounded above, and bounded away from zero, in bounded proper subsectors of $\mathbb{C}\setminus(-\infty,0]$ (see Remark~\ref{rem.e_h_boundedat0}).
Observe that if $R>2$ and $z\in S_{\beta}$ with $|z|\ge R$, it is clear that $z+1\in S_{\beta}$ and $|z+1|\ge R-1>1$, so that we clearly have
\begin{multline}\label{eq.FirstIneqWlogzWlogKz}
\ln(|z+1|) e^{\frac{1}{\sigma-1}\xi(z)} \cos\big(\frac{1}{\sigma-1}\eta(z)\big)-
\arg(z+1)e^{\frac{1}{\sigma-1}\xi(z)} \sin\big(\frac{1}{\sigma-1}\eta(z)\big)\\
\le e^{\frac{1}{\sigma-1}\xi(z)}\Big(\ln(|z+1|)+\frac{\pi\beta}{2}\Big).
\end{multline}
Since $\lim_{x\to +\infty}(\ln(Kx+1)-\ln(x+1))=\ln(K)$, if we choose $K>e^{\pi\beta/2}$ there exists $R>2$ such that for every $z\in S_{\beta}$ with $|z|\ge R$ one has
\begin{equation}\label{eq.IneqLogarithmModulez}
\ln(|z+1|)+\frac{\pi\beta}{2}\le\ln(|z|+1)+\frac{\pi\beta}{2}\le\ln(K|z|+1).
\end{equation}
From this fact we also deduce that for such $z$,
\begin{equation*
\ln^2(K|z|+1)-\ln^2(|z|+1)= \ln\big((K|z|+1)(|z|+1)\big)\ln\Big(\frac{K|z|+1}{|z|+1}\Big)
\ge\Big(\frac{\pi\beta}{2}\Big)^2.
\end{equation*}
Hence, if we write $W(b\ln(K|z|+1))=\xi_1(z)\in\mathbb{R}$, and we take into account~\eqref{eq.RealImaginaryPartsLambert} and~\eqref{eW}, we can obtain that
\begin{align*}
\xi(z)e^{\xi(z)}&\le e^{\xi(z)}(\xi^2(z)+\eta^2(z))^{1/2}=|b\mathop{\rm Log}\nolimits(z+1)|\\
&=b\big(\ln^2(|z+1|)+\arg^2(z+1)\big)^{1/2}\le b\Big(\ln^2(|z|+1)+\Big(\frac{\pi\beta}{2}\Big)^2\Big)^{1/2}\\
&\le b\ln(K|z|+1)=\xi_1(z)e^{\xi_1(z)}.
\end{align*}
Since the map $\xi\in(-1,\infty)\mapsto\xi e^{\xi}$ is strictly increasing, we have
\begin{equation}\label{eq.IneqRealPartLambertandLambert}
\xi(z)\le\xi_1(z)=W(b\ln(K|z|+1)).
\end{equation}
Gathering \eqref{eq.FirstIneqWlogzWlogKz},~\eqref{eq.IneqLogarithmModulez} and~\eqref{eq.IneqRealPartLambertandLambert}, we conclude that~\eqref{eq.IneqFineAdjustComplexLambert} is satisfied, and so we get the first inequality in~\eqref{eq.Bounds_e_h_subsectors_real_axis} for a suitably enlarged constant $C_2\ge 1$.
We turn now to the second inequality in~\eqref{eq.Bounds_e_h_subsectors_real_axis}. Reasoning as before, it suffices to prove that given $\beta\in(0,2)$, there exist $R,H>0$ such that for every $z\in S_{\beta}$ with $|z|\ge R$ one has
\begin{multline}\label{eq.IneqFineAdjustBelowComplexLambert}
e^{\frac{1}{\sigma-1}\xi(z)}\Big(\ln(|z+1|) \cos\big(\frac{1}{\sigma-1}\eta(z)\big)-
\arg(z+1)\sin\big(\frac{1}{\sigma-1}\eta(z)\big)\Big)\\
\ge
\ln\big(\frac{|z|}{H}+1\big) e^{\frac{1}{\sigma-1}W\big(b\ln\big(\frac{|z|}{H}+1\big)\big)}.
\end{multline}
In view of Remark~\ref{remWslow} we derive that $W(b\mathop{\rm Log}\nolimits(z+1))$ tends to infinity as $z$ does so in $S_{\beta}$. Since $|\eta(z)|\le\pi$ for the principal branch of the Lambert function, necessarily $\xi(z)=\Re(W(b\mathop{\rm Log}\nolimits(z+1)))$ tends to $+\infty$ as $z\to\infty$ in $S_{\beta}$. Now, in view of property (d) in Proposition~\ref{propW} it is straightforward that
\begin{equation}\label{eq.EqualityImaginaryPartsLambert}
\mathop{\rm Im}\nolimits(b\mathop{\rm Log}\nolimits(z+1))=e^{\xi(z)}\big(\eta(z)\cos(\eta(z))+ \xi(z)\sin(\eta(z))\big).
\end{equation}
As the left-hand side is bounded by $b\beta\pi/2$ for $z$ in $S_{\beta}$, we deduce that
$$
\lim_{z\to\infty,\ z\in S_{\beta}}\big(\eta(z)\cos(\eta(z))+\xi(z)\sin(\eta(z))\big)=0.
$$
Again, $\eta(z)\cos(\eta(z))$ is bounded, so the only possibility is that $\lim_{z\to\infty,\ z\in S_{\beta}}\sin(\eta(z))=0$, and hence $\lim_{z\to\infty,\ z\in S_{\beta}}\eta(z)=0$. Here we have used the fact that, due to a connectedness argument, $W(b\mathop{\rm Log}\nolimits(z+1))$ cannot approach the boundary of the image set of the principal branch of the Lambert function, so that $\eta(z)$ cannot tend to either $\pi$ or $-\pi$. Going back to~\eqref{eq.EqualityImaginaryPartsLambert}, we get that, as $z$ tends to $\infty$, $\eta(z)$ is equivalent to
$$\sin(\eta(z))=\frac{b\arg(z+1)e^{-\xi(z)}-\eta(z)\cos(\eta(z))}{\xi(z)}.$$
This means there exists a function $\varepsilon(z)$, tending to 1 as $z\to\infty$, such that
$$
\varepsilon(z)\eta(z)= \frac{b\arg(z+1)e^{-\xi(z)}-\eta(z)\cos(\eta(z))}{\xi(z)},
$$
and so
\begin{align}\label{eq.AsymptoticsEtaz}
\eta(z)&= \frac{b\arg(z+1)e^{-\xi(z)}}{\varepsilon(z)\xi(z)+\cos(\eta(z))}
\sim \frac{b\arg(z+1)}{\xi(z)e^{\xi(z)}},\quad z\to\infty,\\
e^{\xi(z)}\eta(z)&\sim \frac{b\arg(z+1)}{\xi(z)}, \quad z\to\infty.
\label{eq.AsymptoticsEtazExpEtaz}
\end{align}
Again from the definition of the Lambert function, it is straightforward that
$$b\ln(|z+1|)=\Re(b\mathop{\rm Log}\nolimits(z+1))=e^{\xi(z)}\big(\xi(z)\cos(\eta(z))- \eta(z)\sin(\eta(z))\big),$$
from where we obtain that
\begin{equation}\label{eq.AsymptoticsbLnz+1}
b\ln(|z+1|)\sim \xi(z)e^{\xi(z)},\quad z\to\infty,\ z\in S_{\beta}.
\end{equation}
Observe that
\begin{equation}\label{eq.RewriteLnz+1Cos}
\ln(|z+1|) \cos\big(\frac{1}{\sigma-1}\eta(z)\big)=
\ln(|z+1|) -\ln(|z+1|) \frac{\sin^2\big(\frac{1}{\sigma-1}\eta(z)\big)}{1+ \cos\big(\frac{1}{\sigma-1}\eta(z)\big)}
\end{equation}
and then, because of~\eqref{eq.AsymptoticsbLnz+1} and~\eqref{eq.AsymptoticsEtaz},
\begin{equation}\label{eq.AsymptoticsLnz+1sinSquare}
\ln(|z+1|)\sin^2\big(\frac{1}{\sigma-1}\eta(z)\big)\sim \frac{\xi(z)e^{\xi(z)}b^2\arg^2(z+1)}{b(\sigma-1)^2 \big(\xi(z)e^{\xi(z)}\big)^2}\to 0 \ \text{as }z\to\infty.
\end{equation}
Hence, going back to~\eqref{eq.IneqFineAdjustBelowComplexLambert}, on the one hand there exists $R_1>2$ such that for every $z\in S_{\beta}$ with $|z|\ge R_1$ we have, by virtue of~\eqref{eq.RewriteLnz+1Cos} and~\eqref{eq.AsymptoticsLnz+1sinSquare},
$$
\ln(|z+1|) \cos\big(\frac{1}{\sigma-1}\eta(z)\big)-
\arg(z+1)\sin\big(\frac{1}{\sigma-1}\eta(z)\big)\ge
\ln(|z+1|)-\frac{1}{b}\ge \ln(|z|-1)-\frac{1}{b}.
$$
Since, for $H>0$, $\lim_{x\to +\infty}(\ln(x-1)-\ln(x/H+1))=\ln(H)$, if we choose $H>e^{1/b}$ there exists $R_2>R_1$ such that for $x\ge R_2$ one has
\begin{equation}\label{eq.Lnx-1LnxH+1}
\ln(x-1)-\ln\big(\frac{x}{H}+1\big)>\frac{1}{b},
\end{equation}
and so for every $z\in S_{\beta}$ with $|z|\ge R_2$ one gets
\begin{equation}\label{eq.IneqLogarithmModulez_bis}
\ln(|z+1|) \cos\big(\frac{1}{\sigma-1}\eta(z)\big)-
\arg(z+1)\sin\big(\frac{1}{\sigma-1}\eta(z)\big)>
\ln\big(\frac{|z|}{H}+1\big).
\end{equation}
On the other hand, by~\eqref{eq.AsymptoticsEtazExpEtaz} there exists
$R_3>2$ such that $e^{\xi(z)}|\eta(z)|<1$. If we take $z\in S_{\beta}$ with $|z|\ge R_4:=\max(R_3,R_2)$, we deduce that
\begin{align*}
\xi(z)e^{\xi(z)}&=e^{\xi(z)}\Re(W(b\mathop{\rm Log}\nolimits(z+1)))\\
&\ge e^{\xi(z)}\big(|W(b\mathop{\rm Log}\nolimits(z+1))|-|\mathop{\rm Im}\nolimits(W(b\mathop{\rm Log}\nolimits(z+1)))|\big)\\
&= b|\mathop{\rm Log}\nolimits(z+1)|-e^{\xi(z)}|\eta(z)|\ge b\ln(|z+1|)-1\\
&\ge b\ln(|z|-1)-1\ge b\ln\big(\frac{|z|}{H}+1\big),
\end{align*}
where the last inequality comes from~\eqref{eq.Lnx-1LnxH+1}. Due again to the fact that $\xi\in(-1,\infty)\mapsto\xi e^{\xi}$ is strictly increasing, we have that $\xi(z)\ge W(b\ln(|z|/H+1))$. Together with~\eqref{eq.IneqLogarithmModulez_bis}, this proves~\eqref{eq.IneqFineAdjustBelowComplexLambert}.
\end{proof1}
For every $p\in\mathbb{N}_0$ we define the $p-$th moment associated with the kernel function $e_{\tau,\sigma}$ by
$$m_{\tau,\sigma}(p):=\int_0^\infty t^{p-1}e_{\tau,\sigma}(t)\,dt=
\int_0^\infty t^p\exp\left(-a_{\tau,\sigma} g(t+1)\right)\,dt.$$
In order to compare the sequence of moments and the original sequence $\mathbb{M}^{\tau,\sigma}$, the following result~\cite[Theorem~2.1]{ptt} is important. It is particularized for the case $1<\sigma<2$, which we are considering here; a more general result is available which, however, does not meet our needs, see Remark~\ref{rem.ChangeBoundsSigmaGreater2}.
\begin{lemma}[\cite{ptt}, Theorem 2.1, Remark 2.2]\label{lema.doubleEstimatesSerbian}
Let $1<\sigma<2$.
There exist $A_{\tau,\sigma},\tilde{A}_{\tau,\sigma}>0$ such that
\begin{equation}\label{eq.doubleEstimatesSerbian}
A_{\tau,\sigma}\exp(a_{\tau,\sigma}g(x))\le e^{T_{\tau,\sigma}(x)}\le \tilde{A}_{\tau,\sigma}\exp(a_{\tau,\sigma}g(x)),
\end{equation}
for every $x>1$.
\end{lemma}
\begin{prop}\label{prop.boundsMomentsSigmaEntre1y2}
Suppose $1<\sigma<2$. Consider the sequence of moments $(m_{\tau,\sigma}(p))_{n\ge0}$ associated with the kernel function $e_{\tau,\sigma}(z)$. Then, there exist $B_1,B_2>0$ such that
\begin{equation}\label{eq.boundsMmomentsdc}
B_1^{p}M_p^{\tau,\sigma}\le m_{\tau,\sigma}(p)\le B_2^{p}M_p^{\tau,\sigma},\quad p\in\mathbb{N}_0.
\end{equation}
\end{prop}
\begin{proof1}
On the one hand, for every $s>0$ we may write
\begin{align}
m_{\tau,\sigma}(p)&=\int_0^s t^p\exp\left(-a_{\tau,\sigma}g(t+1)\right)\,dt
+\int_s^\infty \frac{1}{t^2}t^{p+2}\exp\left(-a_{\tau,\sigma}g(t+1)\right)\,dt\nonumber\\
&\le \int_0^s t^p\,dt
+\frac{1}{s}\sup_{t>0}(t+1)^{p+2}\exp\left(-a_{\tau,\sigma}g(t+1)\right)\nonumber\\
&\le \frac{s^{p+1}}{p+1}+ \frac{1}{s}\sup_{t>1}t^{p+2}\exp\left(-a_{\tau,\sigma}g(t)\right). \label{eq.firstBoundsMomentsAbovedc}
\end{align}
We use now an inequality essentially given in~\cite[$(2.3)$]{ptt}:
\begin{equation*
\sup_{t>1}t^{p}\exp\left(-a_{\tau,\sigma}g(t)\right)
\le A_{1}\sup_{t>1}t^pe^{-T_{\tau,\sigma}(t)}\le A_1\sup_{t>0}t^pe^{-T_{\tau,\sigma}(t)}= A_1M_p^{\tau,\sigma},
\end{equation*}
valid for some $A_1>0$ and every $p\in\mathbb{N}_0$.
With these estimates and property (dc), see~\eqref{eq.Estimate_h_tausigma_dc}, we obtain, continuing with~\eqref{eq.firstBoundsMomentsAbovedc}, that
$$m_{\tau,\sigma}(p)\le \frac{s^{p+1}}{p+1}+ \frac{1}{s}A_1M_{p+2}^{\tau,\sigma}\le \frac{s^{p+1}}{p+1}+ \frac{1}{s}A_1C^{p+2}M_{p}^{\tau,\sigma}.$$
Since $s>0$ was arbitrary, and it is immediate that for any $b>0$ one has
$$
\inf_{s>0}\left(\frac{s^{p+1}}{p+1}+\frac{b}{s}\right)= \frac{p+2}{p+1}b^{(p+1)/(p+2)},
$$
we finally get
$$m_{\tau,\sigma}(p)\le \frac{p+2}{p+1}A_1^{(p+1)/(p+2)} C^{p+1}(M_{p}^{\tau,\sigma})^{(p+1)/(p+2)}\le A_2D_1^{p}M_{p}^{\tau,\sigma}$$
for suitable positive constants $A_2,D_1>0$.
On the other hand, since $g$ is increasing in $(1,\infty)$ by Lemma~\ref{lema.g_hincreasing}, for every $s>0$ we may estimate
$$
m_{\tau,\sigma}(p)\ge \int_0^s t^p\exp\left(-a_{\tau,\sigma}g(t+1)\right)\,dt\ge
\exp\left(-a_{\tau,\sigma}g(s+1)\right)\frac{s^{p+1}}{p+1}.
$$
In particular, we deduce that
$$m_{\tau,\sigma}(p)\ge\sup_{s>1}\exp\left(-a_{\tau,\sigma}g(s+1)\right) \frac{s^{p+1}}{p+1}\ge \frac{1}{2^{2p+1}}\sup_{s>1}(s+1)^{p+1} \exp\left(-a_{\tau,\sigma}g(s+1)\right).$$
We apply now the left inequality in~\eqref{eq.doubleEstimatesSerbian} and obtain
$$
m_{\tau,\sigma}(p)\ge A_{\tau,\sigma}\sup_{s>1}(s+1)^{p+1}e^{-T_{\tau,\sigma}(s+1)}.
$$
Since the sequence $(M_p^{\tau,\sigma})_{p\in\mathbb{N}_0}$ is logarithmically convex, and $T_{\tau,\sigma}(s)$ is precisely its classical associated function (see~\cite{komatsu}), it is well known that there exists $p_0\in\mathbb{N}$ such that for every $p\ge p_0$ the function $s\in(0,\infty)\mapsto s^{p+1}e^{-T_{\tau,\sigma}(s)}$ reaches its supremum at a point greater than $2$, and the value of such supremum is $M_{p+1}^{\tau,\sigma}$.
Hence, we may deduce the existence of positive constants $E_1,F_1$ such that for every $p\in\mathbb{N}$ one has $m_{\tau,\sigma}(p)\ge E_1F_1^{p}M_{p}^{\tau,\sigma}$.
\end{proof1}
\begin{rem}\label{rem.ChangeBoundsSigmaGreater2}
A version of the previous result can be obtained also for $\sigma\ge 2$, by following an analogous reasoning, but there occur some changes in the form of the estimates. Besides the loss of property (dc), in this case the left-hand side inequality in~\eqref{eq.doubleEstimatesSerbian} suffers from a scaling in the constant $a_{\tau,\sigma}$ appearing in the exponent. As a consequence, we can only prove that there exist $\tilde{K}_1,\tilde{K}_2>0$ and $\tilde{\tau}>\tau$ such that $\tilde{K}_{1}^{p^{\sigma}}M_{p}^{\tau,\sigma}\le m_{\tau,\sigma}(p)\le \tilde{K}_{2}^{p^{\sigma}}M_p^{\tilde{\tau},\sigma}$, for every $p\in\mathbb{N}_0$.
It is also important to remark that the previous result guarantees that the ultraholomorphic classes associated with the sequences $(m_{\tau,\sigma}(p))_{p\in\mathbb{N}_0}$ and $(M^{\tau,\sigma}_{p})_{p\in\mathbb{N}_0}$ coincide.
\end{rem}
\begin{prop}\label{prop.eOptimalFlatatInfinity}
Let $1<\sigma<2$, $\tau>0$. For every $T\prec S_{2}$ there exist $C_3,K_3,C_4,K_4>0$ (depending on $\tau,\sigma$) such that
\begin{equation*
C_3\exp(-T_{\tau,\sigma}(|z|/K_3))\le |e_{\tau,\sigma}(z)|\le C_4\exp(-T_{\tau,\sigma}(|z|/K_4)),\quad z\in T.
\end{equation*}
\end{prop}
\begin{proof1}
From Lemma~\ref{lema.Properties.e_h}, there exist $C_2,K_2>0$ such that for every $z\in T$ one has $|e_{\tau,\sigma}(z)|\le C_2e_{\tau,\sigma}(K_2|z|)$. Regarding the values of $e_{\tau,\sigma}(t)$ for positive real $t$, we can apply Lemma~\ref{lema.g_hincreasing}, Lemma~\ref{lema.doubleEstimatesSerbian} and~\eqref{eq.Estimate_h_tausigma_dc} in order to obtain
constants $\tilde{A}_{\tau,\sigma},C>0$ such that
\begin{align*}
e_{\tau,\sigma}(t)&=t\exp(-ag(t+1))<t\exp(-ag(t))\le \tilde{A}_{\tau,\sigma}t\exp(-T_{\tau,\sigma}(t))\\
&=\tilde{A}_{\tau,\sigma}th_{\tau,\sigma}(1/t)\le
\tilde{A}_{\tau,\sigma}t\frac{C}{t}h_{\tau,\sigma}\big(\frac{C}{t}\big)=
C\tilde{A}_{\tau,\sigma}\exp(-T_{\tau,\sigma}(t/C)).
\end{align*}
The inequality on the right is so proved.
Regarding the inequality on the left, again from Lemma~\ref{lema.Properties.e_h} there exist $C_1,K_1>0$ such that for every $z\in T$ one has
$$
|e_{\tau,\sigma}(z)|\ge C_1e_{\tau,\sigma}(K_1|z|)=C_1K_1|z|\exp(-ag(K_1|z|+1)).
$$
Lemma~\ref{lema.g_hincreasing} guarantees that $\lim_{t\to\infty}g'(t)=0$, and the Mean Value Theorem implies then that there exists $c_0>0$ such that for every $t\ge K_1$ one has $g(t+1)-g(t)\le c_0$. So, for every $z\in T$ with $|z|\ge 1$ we have
\begin{align*}
|e_{\tau,\sigma}(z)|&\ge C_1K_1\exp(-ag(K_1|z|))\exp\big(-a(g(K_1|z|+1)-g(K_1|z|))\big)\\
&\ge C_1K_1e^{-ac_0}\exp(-ag(K_1|z|))\ge
C_1K_1e^{-ac_0}A_{\tau,\sigma}\exp(-T_{\tau,\sigma}(K_1|z|)),
\end{align*}
where in the last inequality we have used again Lemma~\ref{lema.doubleEstimatesSerbian}. By suitably enlarging the constants we can make the inequality valid for every $z\in T$.
\end{proof1}
According to~\cite[Theorem~2.3.1]{Thilliez03}, the estimates in Proposition~\ref{prop.eOptimalFlatatInfinity} amount to the fact that the function $e_{\tau,\sigma}(1/z)$ is an optimal flat function in the ultraholomorphic class of functions with a uniform asymptotic expansion associated with the sequence $\mathbb{M}^{\tau,\sigma}$ in any unbounded proper subsector of $\mathbb{C}\setminus(-\infty,0]$.
\section{Right inverses for the asymptotic Borel map in ultraholomorphic classes in sectors}\label{sectRightInver1var}
In this section we construct right inverses for the asymptotic Borel map in ultraholomorphic classes associated with the sequence $\mathbb{M}^{\tau,\sigma}$ in sectors of arbitrary opening of the complex plane.
\begin{theo}\label{tpral}
Let $\tau>0$ and $1<\sigma<2$ be given. For every $\hat{f}=\sum_{p=0}^{\infty}c_p z^p\in \mathbb{C}[[z]]_{\{\mathbb{M}^{\tau,\sigma}\}}$ there exists a function $f$, holomorphic in $\mathbb{C}\setminus(-\infty,0]$, such that for every unbounded proper subsector $S$ of $\mathbb{C}\setminus(-\infty,0]$, one has that $f$ admits $\hat{f}$ as its uniform asymptotic expansion in $S$.
\end{theo}
\begin{proof1} Let $(m_{\tau,\sigma}(p))_{p\in\mathbb{N}_0}$ be the sequence of moments associated with the function $e_{\tau,\sigma}(z)$.
Since $\hat{f} \in \mathbb{C}[[z]]_{\{\mathbb{M}^{\tau,\sigma}\}}$, there exist positive constants $C_1,D_1$ such that
$$|c_p|\le C_1D_1^{p} M^{\tau,\sigma}_p,\quad p\in\mathbb{N}_0.$$
From~\eqref{eq.boundsMmomentsdc}, we deduce that the series
$$
\hat{g}=\sum_{p=0}^{\infty}\frac{c_p}{m_{\tau,\sigma}(p)}z^p
$$
is convergent in a disc $D(0,R)$ for some $R>0$, and it defines a holomorphic function $g$ there. Let $0<R_0<R$.
We define
\begin{equation}\label{intope}
f(z):=\int_{0}^{R_0}e_{\tau,\sigma}\left(\frac{u}{z}\right)g(u)\frac{du}{u},\qquad z\in \mathbb{C}\setminus(-\infty,0],
\end{equation}
which may be called a truncated Laplace-like transform of the function $g$ with kernel $e_{\tau,\sigma}$.
By virtue of Leibniz's theorem on analyticity of parametric integrals and the definition of $e_{\tau,\sigma}$, $f$ turns out to be a holomorphic function in $\mathbb{C}\setminus(-\infty,0]$. In order to obtain our result, it suffices to prove that
$f\sim_{\{\mathbb{M}^{\tau,\sigma}\}}\hat{f}$ uniformly in $S_{\delta}$, for every $0<\delta<2$.
Let $N\in\mathbb{N}$ and $z\in S_{\delta}$. We have
\begin{align*}
f(z)-\sum_{p=0}^{N-1} c_p z^p &= f(z)-\sum_{p=0}^{N-1}\frac{c_p}{m_{\tau,\sigma}(p)} m_{\tau,\sigma}(p)z^p\\
&= \int_{0}^{R_0}e_{\tau,\sigma}\left(\frac{u}{z}\right) \sum_{k=0}^{\infty}\frac{c_{k}}{m_{\tau,\sigma}(k)} u^k \frac{du}{u} -\sum_{p=0}^{N-1}\frac{c_p}{m_{\tau,\sigma}(p)}\int_{0}^{\infty}u^{p-1}e_{\tau,\sigma}(u)du z^p.
\end{align*}
After a change of variable $v=zu$ in the second integral, by virtue of the estimate~(\ref{eq.Bounds_e_h_subsectors_real_axis}) one may use Cauchy's residue theorem in order to check that
$$
z^p\int_{0}^{\infty}u^{p-1}e_{\tau,\sigma}(u)du= \int_{0}^{\infty}v^{p}e_{\tau,\sigma}\left(\frac{v}{z}\right)\frac{dv}{v},
$$
which allows us to write the preceding difference as
\begin{multline*}
\int_{0}^{R_0}e_{\tau,\sigma}\left(\frac{u}{z}\right) \sum_{k=0}^{\infty}\frac{c_{k}}{m_{\tau,\sigma}(k)} u^k \frac{du}{u} -\sum_{p=0}^{N-1}\frac{c_p}{m_{\tau,\sigma}(p)} \int_{0}^{\infty}u^{p}e_{\tau,\sigma}\left(\frac{u}{z}\right) \frac{du}{u} \\
=\int_{0}^{R_0}e_{\tau,\sigma}\left(\frac{u}{z}\right) \sum_{k=N}^{\infty}\frac{c_{k}}{m_{\tau,\sigma}(k)} u^k\frac{du}{u} -\int_{R_0}^{\infty}e_{\tau,\sigma}\left(\frac{u}{z}\right) \sum_{p=0}^{N-1}\frac{c_p}{m_{\tau,\sigma}(p)} u^{p}\frac{du}{u}.
\end{multline*}
Then, we have
$$\left|f(z)-\sum_{p=0}^{N-1} c_p z^p\right|\le f_{1}(z)+f_2(z),$$
where
$$f_{1}(z)=\left|\int_{0}^{R_0}e_{\tau,\sigma}\left(\frac{u}{z}\right) \sum_{k=N}^{\infty}\frac{c_{k}}{m_{\tau,\sigma}(k)} u^k \frac{du}{u}\right|,$$
$$f_{2}(z)=\left|\int_{R_0}^{\infty}e_{\tau,\sigma}\left(\frac{u}{z}\right) \sum_{p=0}^{N-1}\frac{c_p}{m_{\tau,\sigma}(p)} u^{p}\frac{du}{u}\right|.$$
We now give suitable estimates for $f_1(z)$ and $f_2(z)$.
From~\eqref{eq.boundsMmomentsdc} there exist $C_2,D_2>0$ (not depending on $z$) such that
\begin{equation}\label{e327}
\frac{c_{k}}{m_{\tau,\sigma}(k)}\le \frac{C_1D_1^{k} M^{\tau,\sigma}_k}{m_{\tau,\sigma}(k)}\le C_2D_2^k,
\end{equation}
for all $k\in\mathbb{N}_0$. This yields
$$f_{1}(z)\le C_2\int_{0}^{R_0}\left|e_{\tau,\sigma}\left(\frac{u}{z}\right)\right| \sum_{k=N}^{\infty}(D_2u)^{k}\frac{du}{u}.$$
Taking $R_0\le(1-\epsilon)/D_2$ for some $\epsilon>0$ if necessary, we get
$$f_1(z)\le\frac{C_2}{\epsilon}D_2^{N}\int_{0}^{R_0} \left|e_{\tau,\sigma}\left(\frac{u}{z}\right)\right|u^{N-1}du.$$
Let us turn our attention to $f_{2}(z)$. We have $u^p\le R_0^pu^N/R_0^N$ for $u\ge R_0$ and $0\le p\le N-1$. So, according to~\eqref{e327}, we may write
\begin{equation} \label{e1211}
\sum_{p=0}^{N-1}\frac{c_pu^p}{m_{\tau,\sigma}(p)} \leq
\sum_{p=0}^{N-1} \frac{C_1D_1^pM^{\tau,\sigma}_pu^p}{m_{\tau,\sigma}(p)}
\leq C_2 \sum_{p=0}^{N-1}D_2^pu^p\le\frac{u^N}{R_0^N} C_2\sum_{p=0}^{N-1}D_2^pR_0^p\le C_5D_5^Nu^N,
\end{equation}
for some positive constants $C_5,D_5$. Then, we conclude
$$f_2(z)\le C_5D^N_5\int_{R_0}^{\infty}\left|e_{\tau,\sigma}\left(\frac{u}{z}\right) \right|u^{N-1}du.$$
So, both $f_1$ and $f_2$ can be estimated above essentially by
$$
\int_{0}^{\infty}\left|e_{\tau,\sigma}\left(\frac{u}{z}\right) \right|u^{N-1}du.
$$
The second inequality in~\eqref{eq.Bounds_e_h_subsectors_real_axis}, a simple change of variable and the estimates in~\eqref{eq.boundsMmomentsdc} imply that, for suitable constants $C_6,K_6,B>0$, one has
\begin{align*}
\int_{0}^{\infty}\left|e_{\tau,\sigma}\left(\frac{u}{z}\right) \right|u^{N-1}du
&\le
\int_{0}^{\infty}C_6e_{\tau,\sigma}\left(K_6\frac{u}{|z|}\right)u^{N-1}du
=C_6K_6^{-N}m_{\tau,\sigma}(N)|z|^N\\
&\le C_6(B/K_6)^{N}M_N^{\tau,\sigma}|z|^N,
\end{align*}
as desired.
\end{proof1}
As we have mentioned in Subsection~\ref{subsec.UltraHolClasses}, for each $A>0$ we know that $(\widetilde{\mathcal{A}}^u_{\{\mathbb{M}\},A}(S), \left\|\,\,\, \right\|_{\mathbb{M},A,\overset{\sim}{u}})$ and $(\mathbb{C}[[z]]_{\{\mathbb{M}\},A},\left| \ \right|_{\mathbb{M},A})$ are Banach spaces. Then, there is a topological consequence of the previous result: for each $A>0$ we have a linear and continuous right inverse for the asymptotic Borel map from $\mathbb{C}[[z]]_{\{\mathbb{M}\},A}$ to $\tilde{\mathcal{A}}^u_{\{\mathbb{M}\},dA}(S)$ for some $d>0$. More precisely, we observe from the proof of Theorem~\ref{tpral} that $D_2=d_1D_1$ for some $d_1\ge1$, in (\ref{e327}). In addition to this, the choice of $R_0$ guarantees that $D_5$ in (\ref{e1211}) is of the form $d_2D_1$ for some $d_2\ge1$. Therefore, one can choose $d:=\frac{2B}{K_6}\max\{d_1,d_2\}$. This entails that if $\hat{f}\in \mathbb{C}[[z]]_{\{\mathbb{M}\},A} $ there exist constants $c,d>0$, not depending neither on $\hat{f}$ nor on $A$, such that for every $N\in\mathbb{N}_0$ one has
$$\left|f(z)-\sum_{p=0}^{N-1} c_p z^p \right| \leq (cC_1)(d A)^N M^{\tau,\sigma}_{N}=
c|\hat{f}|_{\mathbb{M},A}(dA)^NM^{\tau,\sigma}_N,\qquad z\in S.$$
Then $\left\| f \right\|_{\mathbb{M}^{\tau,\sigma},dA,\overset{\sim}{u}} \leq c |\hat{f}|_{\mathbb{M}^{\tau,\sigma},A}$, where
the scaling factor $d$ only depends on the unbounded subsector $S$ of $\mathbb{C}\backslash(-\infty,0]$, that is, $d$ does not depend on $A$.
Consequently, we have the following result which is analogous to a result by V. Thilliez \cite[Thm. 3.2.1]{Thilliez03} for classes associated with strongly regular sequences.
\begin{theo}\label{corolpral}
Let $\tau>0$ and $1<\sigma<2$ be given. For each $\delta\in (0,2)$ there exists a positive constant $d\ge1$ such that for any $A>0$, the integral operator
$$ T_{\mathbb{M},A}: \mathbb{C}[[z]]_{\{\mathbb{M}\},A} \longrightarrow \widetilde{\mathcal{A}}^u_{\{\mathbb{M}\},dA}(S_\delta),$$
defined for $\hat{f}= \sum_{p=0}^{\infty}c_p z^p $ in \eqref{intope} by
$$T_{\mathbb{M},A}(\hat{f}):=\int_{0}^{R_0}e_{\tau,\sigma}(u/z) \Big(\sum_{p=0}^{\infty}\frac{c_{p}}{m_{\tau,\sigma}(p)} u^p\Big) \frac{du}{u}$$
is a linear and continuous right inverse for the asymptotic Borel map $\mathcal{B}$.
\end{theo}
|
\section{Introduction}
The Stirling number of the second $S_{2}(n,k)$ is the number of ways to partition a set of $n$ objects into $k$ nonempty subsets (see \eqref{6}). The (signed) Stirling number of the first kind $S_{1}(n,k)$ is defined in such a way that the number of permutations of $n$ elements having exactly $k$ cycles is the nonnegative integer $(-1)^{n-k}S_{1}(n,k)=|S_{1}(n,k)|$ (see \eqref{5}). The degenerate Stirling numbers of the second kind $S_{2,\lambda}(n,k)$ (see \eqref{9}) and of the first kind $S_{1,\lambda}(n,k)$ (see \eqref{7}) appear most naturally when we replace the powers of $x$ by the generalized falling factorial polynomials $(x)_{k,\lambda}$ in the defining equations (see \eqref{5}, \eqref{6}, \eqref{7}, \eqref{9}).\par
Carlitz initiated a study of degenerate versions of some special numbers and polynomials in [2], where the degenerate Bernoulli and Euler numbers were investigated. It is remarkable that in recent years intensive studies have been done for degenerate versions of quite a few special polynomials and numbers and have yielded many interesting results (see [6,7] and the references therein). They have been explored by various methods, including mathematical physics, combinatorial methods, generating functions, umbral calculus techniques, $p$-adic analysis, differential equations, special functions, probability theory and analytic number theory.
It turns out that the degenerate Stirling numbers play an important role in this exploration for degenerate versions of many special numbers and polynomials. The normal ordering of an integral power of the number operator $a^{+}a$ in terms of boson operators $a$ and $a^{+}$ can be written in the form
\begin{equation*}
(a^{+}a)^{k}=\sum_{l=0}^{k}S_{2}(k,l)(a^{+})^{l}a^{l}.
\end{equation*}
The normal ordering of the degenerate $k$th power of the number operator $a^{+}a$, namely $(a^{+}a)_{k,\lambda}$, in terms of boson operators $a, a^{+}$ can be written in the form
\begin{equation}
(a^{+}a)_{k,\lambda}=\sum_{l=0}^{k}S_{2,\lambda}(k,l)(a^{+})^{l}a^{l},\label{1}
\end{equation}
where the generalized falling factorials $(x)_{n,\lambda}$ are given by \eqref{1}.\par
By inversion, from \eqref{1} we obtain
\begin{equation}
(a^{+})^{k}a^{k}=\sum_{l=0}^{k}S_{1,\lambda}(k,l)(a^{+}a)_{l,\lambda}.\label{2}
\end{equation}
The aim of this paper is to derive some identities and recurrence relations for the degenerate Stirling numbers of the first kind and of the second kind by using the normal ordering in \eqref{1} and its inversion in \eqref{2}, certain relations of boson operators and the recurrence relations of the Stirling numbers themselves.
In more detail, our main results are as follows. Firstly, we derive some recurrence relation and identity involving the degenerate Stirling numbers of the second kind in Theorems 3 and 5 by using \eqref{1} and certain relations for boson operators. Secondly, we find some recurrence relation and identity involving the degenerate Stirling numbers of the first kind in Theorems 4 and 6 by exploiting \eqref{2} and certain relations for boson operators.
Thirdly, we investigate some recurrence relations for the degenerate Stirling numbers of the second kind in Theorems 7 and 9 by using the recurrence relation in
\eqref{15}, and those for the degenerate Stirling numbers of the first kind in Theorems 8 and 11 by using the recurrence relation in \eqref{8}. For the rest of this section, we recall the facts that are needed throughout this paper.
For any $\lambda\in\mathbb{R}$, the degenerate exponential functions are defined by
\begin{equation}
e_{\lambda}^{x}(t)=\sum_{k=0}^{\infty}(x)_{n,\lambda}\frac{t^{k}}{k!},\quad (\mathrm{see}\ [6,7]), \label{3}
\end{equation}
where
\begin{equation}
(x)_{0,\lambda}=1,\quad (x)_{k,\lambda}=x(x-\lambda)\cdots\big(x-(k-1)\lambda\big),\quad (k \ge 1).\label{4}
\end{equation}
When $x=1$, we use the notation $e_{\lambda}(t)=e_{\lambda}^{1}(t)$. It is well known that the Stirling numbers of the first kind are defined by
\begin{equation}
(x)_{n}=\sum_{k=0}^{n}S_{1}(n,k)x^{k},\quad (n\ge 0),\quad (\mathrm{see}\ [1,2,3,9,10]),\label{5}
\end{equation}
where $(x)_{0}=1,\ (x)_{n}=x(x-1)\cdots(x-n+1),\quad (n\ge 1)$. \par
As the inversion formula of \eqref{5}, the Stirling numbers of the second kind are defined by
\begin{equation}
x^{n}=\sum_{k=0}^{n}S_{2}(n,k)(x)_{k},\quad (n\ge 0).\label{6}
\end{equation}
In [6], the degenerate Stirling numbers of the first kind are defined by
\begin{equation}
(x)_{n}=\sum_{k=0}^{n}S_{1,\lambda}(n,k)(x)_{k,\lambda},\quad (n\ge 0).\label{7}
\end{equation}
From \eqref{7}, we note that
\begin{equation}
S_{1,\lambda}(k+1,l)=S_{1,\lambda}(k,l-1)-(k-l\lambda)S_{1,\lambda}(k,l),\label{8}
\end{equation}
where $k,l\in\mathbb{N}$ with $k\ge l$. \par
As the inverse formula of \eqref{7}, the degenerate Stirling numbers of the second kind are defined by
\begin{equation}
(x)_{n,\lambda}=\sum_{k=0}^{n}S_{2,\lambda}(n,k)(x)_{k},\quad (n\ge 0),\quad (\mathrm{see}\ [6]).\label{9}
\end{equation}
Note that $\displaystyle\lim_{\lambda\rightarrow 0}S_{2,\lambda}(n,k)=S_{2}(n,k)\displaystyle$ and $\displaystyle\lim_{\lambda\rightarrow 0}S_{1,\lambda}(n,k)=S_{1}(n,k)\displaystyle $.\par
By \eqref{7} and \eqref{9}, we easily obtain the next proposition.
\begin{proposition}
The following orthogonality and inverse relations hold true.
\begin{align*}
&\sum_{k=l}^{n}S_{1,\lambda}(n,k)S_{2,\lambda}(k,l)=\delta_{n,l},
\,\, \sum_{k=l}^{n}S_{2,\lambda}(n,k)S_{1,\lambda}(k,l)=\delta_{n,l},\,\,(0 \le l \le n), \\
&a_{n}=\sum_{k=0}^{n}S_{2,\lambda}(n,k)b_{k} \Longleftrightarrow b_{n}=\sum_{k=0}^{n}S_{1,\lambda}(n,k)a_{k},\\
&a_{n}=\sum_{k=n}^{m}S_{2,\lambda}(k,n)b_{k} \Longleftrightarrow b_{n}=\sum_{k=n}^{m}S_{1,\lambda}(k,n)a_{k},
\end{align*}
where $\delta_{n,l}$ is the Kronecker's symbol. \\
\end{proposition}
Let $a$ and $a^{+}$ be the boson annihilation and creation operators satisfying the commutation relation
\begin{equation}
[a,a^{+}]=aa^{+}-a^{+}a=1,\quad (\mathrm{see}\ [4,5,8]).\label{10}
\end{equation}
It has been known for some time that the normal ordering of $(a^{+}a)^{n}$ has a close relation to the Stirling numbers of the second kind ([4,5,8]).
Indeed, the normal ordering of an integral power of the number operator $a^{+}a$ in terms of boson operators $a$ and $a^{+}$, that satisfy the commutation $[a,a^{+}]=aa^{+}-a^{+}a=1$, can be written in the form
\begin{equation}
(a^{+}a)^{k}=\sum_{l=0}^{k}S_{2}(k,l)(a^{+})^{l}a^{l},\quad (\mathrm{see}\ [4,5,8]). \label{11}
\end{equation}
The number states $|m\rangle,\ m=1,2,\dots$, are defined as
\begin{equation}
a|m\rangle=\sqrt{m}|m-1\rangle,\quad a^{+}|m\rangle=\sqrt{m+1}|m+1\rangle.\label{12}
\end{equation}
By \eqref{12}, we get $a^{+}a|m\rangle=m|m\rangle$, (see [4,5,8]). The coherent states $|z\rangle$, where $z$ is a complex number, satisfy $a|z\rangle=z|z\rangle,$ $z|z\rangle=1$. To show a connection to coherent states, we recall that the harmonic oscillator has Hamiltonian $\hat{n}=a^{+}a$ (neglecting the zero point energy) and the usual eigenstates $|n\rangle$ (for $n\in\mathbb{N}$) satisfying $\hat{n}|n\rangle=n|n\rangle$ and $\langle m|n\rangle=\delta_{m,n}$, where $\delta_{m,n}$ is the Kronecker's symbol. In this paper, we derive some identities involving the degenerate Stirling numbers arising from the normal ordering of degenerate integral powers of the number operator $a^{+}a$ in terms of boson operators $a$ and $a^{+}$,
which are given by
\begin{displaymath}
(a^{+}a)_{k,\lambda}=\sum_{l=0}^{k}S_{2,\lambda}(k,l)(a^{+})^{l}a^{l}.
\end{displaymath}
In this paper, we study the degenerate Stirling numbers associated with the number operator $a^{+},a$ that satisfy $[a,a^{+}]=aa^{+}-a^{+}a=1$. Indeed, we give some new formulae of the degenerate Stirling numbers which are derived from an integral power of number operator $a^{+}a$ in terms of boson operators $a$ and $a^{+}$ with $$(a^{+}a)_{k,\lambda}=\sum_{l=0}^{k}S_{2,\lambda}(k,l)(a^{+})^{l}a^{l}.$$
\section{Some identities involving degenerate Stirling numbers}
We recall that the bosonic commutation relation
$[a,a^{+}]=aa^{+}-a^{+}a=1$ can be realized formally in a suitable space of functions by letting $a=\frac{d}{dx}$ and $a^{+}=x$ (the operator of multiplication by $x$). It is known that
\begin{equation}
\bigg(x\frac{d}{dx}\bigg)_{n,\lambda}f(x)=\sum_{k=0}^{n}S_{2,\lambda}(n,k)x^{k}\bigg(\frac{d}{dx}\bigg)^{k}f(x),\quad (\mathrm{see}\ [7]),\label{13}
\end{equation}
where $n$ is a nonnegative integer. \par
The equation \eqref{13} can be written as
\begin{equation}
(a^{+}a)_{n,\lambda}=\sum_{k=0}^{n}S_{2,\lambda}(n,k)(a^{+})^{k}a^{k}.\label{14}
\end{equation}
From \eqref{9}, we note that
\begin{equation}
S_{2,\lambda}(k+1,l)=S_{2,\lambda}(k,l-1)+(l-k\lambda)S_{2,\lambda}(k,l),\label{15}
\end{equation}
where $k,l\in\mathbb{N}$ with $k\ge l$. It is easy to show that
\begin{displaymath}
[a,\hat{n}]=a,\quad [\hat{n},a^{+}]=a^{+}.
\end{displaymath}
Inverting \eqref{13} by using Proposition 1, we have
\begin{equation}
x^{k}\bigg(\frac{d}{dx}\bigg)^{k}f(x)=\sum_{m=0}^{k}S_{1,\lambda}(k,m)\bigg(x\frac{d}{dx}\bigg)_{m,\lambda}f(x).\label{16}
\end{equation}
In view of \eqref{16}, the normal ordering of a degenerate integral power of the number operator $a^{+}a$ in terms of boson operators $a,a^{+}$ can be rewritten in the form
\begin{align}
(a^{+})^{k}a^{k}&=\sum_{m=0}^{k}S_{1,\lambda}(k,m)(a^{+}a)_{m,\lambda}\label{17} \\
&=\sum_{m=0}^{k}S_{1,\lambda}(k,m)(\hat{n})_{m,\lambda}=(\hat{n})_{k},\nonumber
\end{align}
where $k$ is a positive integer.
\begin{proposition}
For $k\in\mathbb{N}$, we have
\begin{align*}
(\hat{n})_{k}=(a^{+})^{k}a^{k}&=\sum_{m=0}^{k}S_{1,\lambda}(k,m)(a^{+}a)_{m,\lambda}
=\sum_{m=0}^{k}S_{1,\lambda}(k,m)(\hat{n})_{m,\lambda}.\end{align*}
\end{proposition}
We note that
\begin{align}
a^{+}(\hat{n}+1-\lambda)_{k,\lambda}a &=a^{+}\sum_{l=0}^{k}\binom{k}{l}(\hat{n})_{l,\lambda}(1-\lambda)_{k-l,\lambda}a \label{18}\\
&=\sum_{l=0}^{k}\binom{k}{l}(1-\lambda)_{k-l,\lambda}\sum_{m=0}^{l}S_{2,\lambda}(l,m)(a^{+})^{m+1}a^{m+1}\nonumber \\
&=\sum_{m=0}^{k}\bigg(\sum_{l=m}^{k}\binom{k}{l}(1-\lambda)_{k-l,\lambda}S_{2,\lambda}(l,m)\bigg)(a^{+})^{m+1}a^{m+1}. \nonumber
\end{align}
By \eqref{10}, we also have
\begin{align}
(\hat{n})_{k+1,\lambda}&=(\hat{n}-\lambda)_{k,\lambda}\hat{n}=a^{+}\big((\hat{n}+1-\lambda)\cdots(\hat{n}+1-k\lambda)\big)a\label{19}\\
&=a^{+}(\hat{n}+1-\lambda)_{k,\lambda}a.\nonumber
\end{align}
By \eqref{19} and \eqref{14}, we get
\begin{align}
a^{+}(\hat{n}+1-\lambda)_{k,\lambda}a&=(\hat{n})_{k+1,\lambda}=(a^{+}a)_{k+1,\lambda}\nonumber \\
&=\sum_{m=0}^{k+1}S_{2,\lambda}(k+1,m)(a^{+})^{m}a^{m}\label{20} \\
&=\sum_{m=1}^{k+1}S_{2,\lambda}(k+1,m)(a^{+})^{m}a^{m}\nonumber \\
&=\sum_{m=0}^{k}S_{2,\lambda}(k+1,m+1)(a^{+})^{m+1}a^{m+1}.\nonumber
\end{align}
Therefore, by \eqref{18} and \eqref{20}, we obtain the following theorem.
\begin{theorem}
For $m,k\in\mathbb{Z}$ with $k\ge m \ge 0$, we have
\begin{displaymath}
\sum_{l=m}^{k}\binom{k}{l}(1-\lambda)_{k-l,\lambda}S_{2,\lambda}(l,m)=S_{2,\lambda}(k+1,m+1).
\end{displaymath}
\end{theorem}
From Proposition 2, we note that
\begin{align}
(a^{+})^{k+1}a^{k+1}&=\sum_{m=0}^{k+1}S_{1,\lambda}(k+1,m)(a^{+}a)_{m,\lambda} \label{21}\\
&=\sum_{m=1}^{k+1}S_{1,\lambda}(k+1,m)(\hat{n})_{m,\lambda} \nonumber \\
&=\sum_{m=0}^{k}S_{1,\lambda}(k+1,m+1)(\hat{n})_{m+1,\lambda},\quad (k \ge 0).\nonumber
\end{align}
The degenerate rising factorial sequence is defined by
\begin{equation}
\langle x\rangle_{0,\lambda}=1,\quad \langle x\rangle_{n,\lambda}=x(x+\lambda)\cdots(x+(n-1)\lambda),\quad (n\ge 1).\label{22}
\end{equation}
Now, from \eqref{10} we note that
\begin{equation}
a^{+}(\hat{n})_{k,\lambda}a=(\hat{n}-1)_{k,\lambda}\hat{n}=\hat{n}(\hat{n}-1)_{k,\lambda},\quad(k\in\mathbb{N}). \label{23}
\end{equation}
On the other hand, by \eqref{17} and \eqref{23}, we get
\begin{align}
&(a^{+})^{k+1}a^{k+1}=\sum_{l=0}^{k}S_{1,\lambda}(k,l)a^{+}(\hat{n})_{l,\lambda}a=\sum_{l=0}^{k}S_{1,\lambda}(k,l)(\hat{n}-1)_{l,\lambda}\hat{n} \label{24}\\
&=\sum_{l=0}^{k}S_{1,\lambda}(k,l)\sum_{m=0}^{l}\binom{l}{m}(-1)^{l-m}\langle 1\rangle_{l-m,\lambda}(\hat{n})_{m,\lambda}\big(\hat{n}-m\lambda+m\lambda) \nonumber \\
&=\sum_{l=0}^{k}S_{1,\lambda}(k,l)\sum_{m=0}^{l}\binom{l}{m}(-1)^{l-m}\langle 1\rangle_{l-m,\lambda}(\hat{n})_{m+1,\lambda}\nonumber \\
&\quad\quad +\lambda\sum_{l=0}^{k}S_{1,\lambda}(k,l)\sum_{m=1}^{l+1}\binom{l}{m}(-1)^{l-m}\langle 1\rangle_{l-m,\lambda}(\hat{n})_{m,\lambda}m\nonumber\\
&=\sum_{m=0}^{k}\sum_{l=m}^{k}S_{1,\lambda}(k,l)\binom{l}{m}(-1)^{l-m}\langle 1\rangle_{l-m,\lambda}(\hat{n})_{m+1,\lambda}\nonumber \\
&\quad\quad +\sum_{m=0}^{k}\sum_{l=m}^{k}S_{1,\lambda}(k,l)(m+1)\lambda\binom{l}{m+1}(-1)^{l-m-1}\langle 1\rangle_{l-m-1,\lambda}(\hat{n})_{m+1,\lambda}.\nonumber
\end{align}
Therefore, by \eqref{21} and \eqref{24}, we obtain the following theorem.
\begin{theorem}
For $m,k\in\mathbb{Z}$ with $k \ge m \ge 0$, we have
\begin{align*}
&S_{1,\lambda}(k+1,m+1) \\
&\quad =\sum_{l=m}^{k}S_{1,\lambda}(k,l)\bigg\{\binom{l}{m}\langle 1\rangle_{l-m,\lambda}(-1)
^{l-m}+l\lambda\binom{l-1}{m}(-1)^{l-m-1}\langle 1\rangle_{l-m-1,\lambda}\bigg\}.
\end{align*}
\end{theorem}
From \eqref{23}, we get
\begin{align}
&a^{+}(\hat{n})_{k,\lambda}a=\hat{n}(\hat{n}-1)_{k,\lambda}=\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}\hat{n}(\hat{n})_{m,\lambda}\label{25} \\
&=\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}(\hat{n}-m\lambda+m\lambda)(\hat{n})_{m,\lambda}\nonumber \\
&=\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}\big((\hat{n})_{m+1,\lambda}+m\lambda(\hat{n})_{m,\lambda}\big) \nonumber \\
&= \sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}(\hat{n})_{m+1,\lambda}\nonumber\\
& \quad +\sum_{m=0}^{k}\binom{k}{m+1}(-1)^{k-m-1}\langle 1\rangle_{k-m-1,\lambda}(m+1)\lambda (\hat{n})_{m+1,\lambda} \nonumber\\
&=\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}\sum_{p=1}^{m+1}S_{2,\lambda}(m+1,p)(a^{+})^{p}a^{p} \nonumber \\
&\quad +\lambda\sum_{m=0}^{k}\binom{k}{m+1}(-1)^{k-m-1}\langle 1\rangle_{k-m-1,\lambda}(m+1)\sum_{p=1}^{m+1}S_{2,\lambda}(m+1,p)(a^{+})^{p}a^{p}. \nonumber \\
&=\sum_{p=1}^{k+1}\bigg(\sum_{m=p-1}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}S_{2,\lambda}(m+1,p)\bigg)(a^{+})^{p}a^{p} \nonumber \\
&\quad +\lambda\sum_{p=1}^{k+1}\sum_{m=p-1}^{k}\binom{k}{m+1}(-1)^{k-m-1}\langle 1\rangle_{k-m-1,\lambda}(m+1)S_{2,\lambda}(m+1,p)(a^{+})^{p}a^{p}\nonumber \\
&=\sum_{p=0}^{k}\bigg\{\sum_{m=p}^{k}\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}S_{2,\lambda}(m+1,p+1)\bigg\}(a^{+})^{p+1}a^{p+1} \nonumber \\
&\quad +\lambda\sum_{p=0}^{k}\sum_{m=p}^{k}k\binom{k-1}{m}(-1)^{k-m-1}\langle 1\rangle_{k-m-1,\lambda}S_{2,\lambda}(m+1,p+1)(a^{+})^{p+1}a^{p+1}. \nonumber
\end{align}
We observe that
\begin{align}
a^{+}(\hat{n})_{k,\lambda}a &= a^{+}\bigg(\sum_{p=0}^{k}S_{2,\lambda}(k,p)(a^{+})^{p}a^{p}\bigg)a \label{26}\\
&=\sum_{p=0}^{k}S_{2,\lambda}(k,p)(a^{+})^{p+1}a^{p+1}.\nonumber
\end{align}
Therefore, by \eqref{25} and \eqref{26}, we obtain the following theorem.
\begin{theorem}
For $p,k\in\mathbb{Z}$ with $0\le p \le k$, we have
\begin{align*}
S_{2,\lambda}(k,p)=&\sum_{m=p}^{k}\bigg\{\binom{k}{m}(-1)^{k-m}\langle 1\rangle_{k-m,\lambda}S_{2,\lambda}(m+1,p+1)\\
&\quad +\lambda k\binom{k-1}{m}(-1)^{k-m-1}\langle 1\rangle_{k-m-1,\lambda}S_{2,\lambda}(m+1,p+1)\bigg\}.
\end{align*}
\end{theorem}
By \eqref{17}, we obtain
\begin{align}
&(a^{+})^{k+1}a^{k+1}=\sum_{p=0}^{k+1}S_{1,\lambda}(k+1,p)(a^{+}a)_{p,\lambda}=\sum_{p=1}^{k+1}S_{1,\lambda}(k+1,p)(\hat{n})_{p,\lambda} \label{27} \\
&=\sum_{p=1}^{k+1}S_{1,\lambda}(k+1,p)\hat{n}(\hat{n}-1+1-\lambda)_{p-1,\lambda}\nonumber\\
&=\sum_{p=1}^{k+1}S_{1,\lambda}(k+1,p)\hat{n}\sum_{l=0}^{p-1}\binom{p-1}{l}(1-\lambda)_{p-1-l,\lambda}(\hat{n}-1)_{l,\lambda}\nonumber\\
&=\sum_{l=0}^{k}\bigg(\sum_{p=l+1}^{k+1}S_{1,\lambda}(k+1,p)\binom{p-1}{l}(1-\lambda)_{p-1-l,\lambda}\bigg)\hat{n}(\hat{n}-1)_{l,\lambda}\nonumber\\
&=\sum_{l=0}^{k}\bigg(\sum_{p=l}^{k}S_{1,\lambda}(k+1,p+1)\binom{p}{l}(1-\lambda)_{p-l,\lambda}\bigg)\hat{n}(\hat{n}-1)_{l,\lambda}.\nonumber
\end{align}
Therefore, from the first line of \eqref{24} and \eqref{27}, we obtain the following theorem.
\begin{theorem}
For $k,l\in\mathbb{Z}$ with $k\ge l \ge 0$, we have
\begin{displaymath}
S_{1,\lambda}(k,l)=\sum_{p=l}^{k}S_{1,\lambda}(k+1,p+1)\binom{p}{l}(1-\lambda)_{p-l,\lambda}.
\end{displaymath}
\end{theorem}
From \eqref{9}, we note that
\begin{equation}
\frac{1}{k}\big(e_{\lambda}(t)-1\big)^{k}=\sum_{n=k}^{\infty}S_{2,\lambda}(n,k)\frac{t^{n}}{n!},\quad (k\ge 0).\label{28}
\end{equation}
Thus, by \eqref{28}, we get
\begin{align}
\sum_{n=k}^{\infty}S_{2,\lambda}(n,k)\frac{t^{n}}{n!}&=\frac{1}{k!}\big(e_{\lambda}(t)-1\big)^{k}=\frac{1}{k!}\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}e_{\lambda}^{m}(t)\label{29} \\
&=\sum_{n=0}^{\infty}\bigg(\frac{1}{k!}\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}(m)_{n,\lambda}\bigg)\frac{t^{n}}{n!}.\nonumber
\end{align}
Comparing the coefficients on both sides of \eqref{29}, we have
\begin{displaymath}
\sum_{m=0}^{k}\binom{k}{m}(-1)^{k-m}(m)_{n,\lambda}=\left\{\begin{array}{ccc}
k!S_{2,\lambda}(n,k), & \textrm{if $n\ge k,$}\\
0, & \textrm{if $0 \le n<k$}.
\end{array}\right.
\end{displaymath}
By \eqref{15}, we get
\begin{align}
&S_{2,\lambda}(k+1,l+1)=S_{2,\lambda}(k,l)+(l+1-k\lambda)S_{2,\lambda}(k,l+1) \label{30}\\
&=S_{2,\lambda}(k,l)+(l+1-k\lambda)\nonumber \\
&\quad\quad \times \big\{S_{2,\lambda}(k-1,l)+(l+1-(k-1)\lambda)S_{2,\lambda}(k-1,l+1)\big\}\nonumber \\
&=S_{2,\lambda}(k,l)+(l+1-k\lambda)S_{2,\lambda}(k-1,l) \nonumber \\
&\quad\quad +\langle l+1-k\lambda \rangle_{2,\lambda}S_{2,\lambda}(k-1,l+1) \nonumber \\
&=S_{2,\lambda}(k,l)+(l+1-k\lambda)S_{2,\lambda}(k-1,l)+\langle l+1-k\lambda \rangle_{2,\lambda}S_{2,\lambda}(k-2,l)\nonumber\\
&\quad+\cdots + \langle l+1-k\lambda \rangle_{k-l,\lambda}S_{2,\lambda}(k,l) \nonumber \\
&=\sum_{m=l}^{k} \langle l+1-k\lambda \rangle_{k-m,\lambda}S_{2,\lambda}(l,l).\nonumber
\end{align}
Therefore, by \eqref{30}, we obtain the following theorem.
\begin{theorem}
For $l,k\in\mathbb{Z}$ with $0\le l\le k$, we have
\begin{displaymath}
S_{2,\lambda}(k+1,l+1)=\sum_{m=l}^{k}\langle l+1-k\lambda \rangle_{k-m,\lambda}S_{2,\lambda}(m,l).
\end{displaymath}
\end{theorem}
From \eqref{8}, we note that
\begin{align}
&S_{1,\lambda}(k+1,m+1)=S_{1,\lambda}(k,m)-\big(k-(m+1)\lambda\big)S_{1,\lambda}(k,m+1)\label{31}\\
&=S_{1,\lambda}(k,m)-\big(k-(m+1)\lambda\big)\nonumber\\
&\quad\quad \times\Big(S_{1,\lambda}(k-1,m)-((k-1)-(m+1)\lambda)S_{1,\lambda}(k-1,m+1)\Big)\nonumber \\
&=S_{1,\lambda}(k,m)-(k-(m+1)\lambda)S_{1,\lambda}(k-1,m)\nonumber \\
&\quad\quad +(-1)^{2}(k-(m+1)\lambda)_{2} S_{1,\lambda}(k-1,m+1)\nonumber \\
&=S_{1,\lambda}(k,m)-(k-(m+1)\lambda)S_{1,\lambda}(k-1,m) \nonumber \\
&\quad\quad +(-1)^{2}(k-(m+1)\lambda)_{2} S_{1,\lambda}(k-1,m+1)\nonumber \\
&\quad\quad +\cdots+(-1)^{k-m}(k-(m+1)\lambda)_{k-m}S_{1,\lambda}(m,m)\nonumber \\
&=\sum_{l=m}^{k}(-1)^{k-l}\big(k-(m+1)\lambda)_{k-l}S_{1,\lambda}(l,m).\nonumber
\end{align}
Therefore, by \eqref{31}, we obtain the following theorem.
\begin{theorem}
For $m,k\in\mathbb{Z}$ with $k \ge m \ge 0$, we have
\begin{displaymath}
S_{1,\lambda}(k+1,m+1)=\sum_{l=m}^{k}(-1)^{k-l}\big(k-(m+1)\lambda\big)_{k-l}S_{1,\lambda}(l,m).
\end{displaymath}
\end{theorem}
By \eqref{15}, we get
\begin{align}
&S_{2,\lambda}(m+k+1,m)=S_{2,\lambda}(m+k,m-1)+\big(m-(m+k)\lambda\big)S_{2,\lambda}(m+k,m)\label{32}\\
&=S_{2,\lambda}(m+k-1,m-2)+\big((m-1)-(m+k-1)\lambda\big)S_{2,\lambda}(m+k-1,m-1) \nonumber \\
&\quad\quad +\big(m-(m+k)\lambda\big)S_{2,\lambda}(m+k,m)\nonumber \\
&=\big(m-(m+k)\lambda\big)S_{2,\lambda}(m+k,m)+\big((m-1)-(m+k-1)\lambda\big)\nonumber \\
&\quad\quad \times S_{2,\lambda}(m+k-1,m-1)+\cdots+(0-k\lambda)S_{2,\lambda}(k,0) \nonumber \\
&=\sum_{l=0}^{m}\big(l-(k+l)\lambda\big)S_{2,\lambda}(k+l,l).\nonumber
\end{align}
Therefore, by \eqref{32}, we obtain the following theorem.
\begin{theorem}
For $k,m\in\mathbb{Z}$ with $k,m\ge 0$, we have
\begin{displaymath}
S_{2,\lambda}(m+k+1,m)=\sum_{l=0}^{m}\big(l-(k+l)\lambda\big)S_{2,\lambda}(k+l,l).
\end{displaymath}
\end{theorem}
From Theorem 3, we note that
\begin{displaymath}
S_{2,\lambda}(k+1,m+1)=\sum_{l=m}^{k}\binom{k}{l}(1-\lambda)_{k-l,\lambda}S_{2,\lambda}(l,m).
\end{displaymath}
Now, by using Propositon 1 we have
\begin{align}
\sum_{l=m}^{k}S_{2,\lambda}(k+1,l+1)S_{1,\lambda}(l,m)&=\sum_{l=m}^{k}\sum_{p=l}^{k}\binom{k}{p}(1-\lambda)_{k-p,\lambda}S_{2,\lambda}(p,l)S_{1,\lambda}(l,m).\label{33}\\
&=\sum_{p=m}^{k}\binom{k}{p}(1-\lambda)_{k-p,\lambda}\sum_{l=m}^{p}S_{2,\lambda}(p,l)S_{1,\lambda}(l,m)\nonumber \\
&=\binom{k}{m}(1-\lambda)_{k-m,\lambda}.\nonumber
\end{align}
Therefore, by \eqref{33}, we obtain the following theorem.
\begin{theorem}
For $m,k\in\mathbb{Z}$ with $k \ge m \ge 0$, we have
\begin{displaymath}
\binom{k}{m}=\frac{1}{(1-\lambda)_{k-m,\lambda}}\sum_{l=m}^{k}S_{2,\lambda}(k+1,l+1)S_{1,\lambda}(l,m).
\end{displaymath}
\end{theorem}
By \eqref{8}, we get
\begin{align}
&S_{1,\lambda}(m+k+1,m)=S_{1,\lambda}(m+k,m-1)-(m+k-m\lambda)S_{1,\lambda}(m+k,m)\label{34} \\
&=S_{1,\lambda}(m+k-1,m-2)-(m+k-1-(m-1)\lambda))S_{1,\lambda}(m+k-1,m-1)\nonumber\\
&\quad\quad -(m+k-m\lambda)S_{1,\lambda}(m+k,m)\nonumber \\
&=-(m+k-m\lambda)S_{1,\lambda}(m+k,m)-(m+k-1-(m-1)\lambda)\nonumber \\
&\quad\quad \times S_{1,\lambda}(m+k-1,m-1)- \cdots-k S_{1,\lambda}(k,0)\nonumber\\
&=-\sum_{l=0}^{m}(k+l-l\lambda)S_{1,\lambda}(k+l,l).\nonumber
\end{align}
Therefore, by \eqref{34}, we obtain the following theorem.
\begin{theorem}
For $m,k\in\mathbb{Z}$ with $m,k\ge 0$, we have
\begin{displaymath}
S_{1,\lambda}(m+k+1,m)=-\sum_{l=0}^{m}(k+l-l\lambda)S_{1,\lambda}(k+l,l).
\end{displaymath}
\end{theorem}
\section{Conclusion}
In recent years, studying degenerate versions of some special numbers and polynomials have drawn the attention of many mathematicians with their regained interests not only in combinatorial and arithmetical properties but also in applications to differential equations, identities of symmetry and probability theory. These degenerate versions include the degenerate Stirling numbers of the first and second kinds, degenerate Bernoulli numbers of the second kind and degenerate Bell numbers and polynomials. \par
As a degenerate version of the well known normal ordering of an integral power of the number operator, we considered the normal ordering of a degenerate integral power of the number operator in terms of boson operators and its inversion as well.
We derived some identities and recurrence relations for the degenerate Stirling numbers of the first kind and of the second kind by using the normal ordering in \eqref{1} and its inversion in \eqref{2}, certain relations of boson operators and the recurrence relations of the Stirling numbers themselves. \par
It is one of our future projects to continue to explore various degenerate versions of many special polynomials and numbers by using various methods mentioned in the Introduction.
\vspace{ 0.5cm}
{ \bf Conflict of Interest}
We have no conflicts of interest to disclose.
\vspace{ 0.5cm}
{\bf DATA AVAILABILITY}
Data sharing is not applicable to this article as no new data were created or analyzed in this study.
|
\section{Introduction}
Recently, Graph Neural Networks (GNNs) are widely applied in IR applications ranging from recommender system~\cite{he2020lightgcn, liu2021interest, wu2021self, jin2020multi, wang2019neural}, question answering~\cite{hu2020residual,zhang2020answer} to web search~\cite{mao2020item, lin2021graph}.
Specifically, graph convolution-based methods~\cite{zhao2021wgcn,xu2020label} incorporate rich attributes of nodes along with the structural information of graphs by recursively aggregating the neighborhood information.
Despite the success, the performance of GNNs on node classification significantly degrades when only few labeled nodes are given as GNNs tend to overfit to the few labeled nodes. To make the matter worse, GNNs suffer from ineffective propagation of supervisory signals due to the well-known over-smoothing issue~\cite{deeper}, which makes GNNs not being able to fully benefit from the given labeled nodes. Fig.~\ref{fig:teaser} demonstrates a severe performance degradation of GCN as the rate of the labeled nodes decreases.
\begin{figure*}[t]
\includegraphics[width=0.7\linewidth]{Img/Figure2.pdf}
\caption{
The overall model architecture of~\textsf{GraFN}. Given a graph $G$, we generate two differently augmented views $\tilde{G}$ and $\tilde{G}^{'}$ both of which are fed into a shared encoder $f_{\theta}$ to obtain node-level representation $\tilde{\mathbf{Z}}$ and $\tilde{\mathbf{Z}}^{'}$, respectively. Then,~\textsf{GraFN}~not only minimizes the difference between these two representations obtained from differently augmented graphs, i.e. node-wise consistency ($\mathcal{L}_\text{NC}$), but also minimizes the difference between the predicted class distributions computed in a non-parametrically manner by using the similarity between the anchor node and support nodes, which are randomly sampled from labeled nodes, i.e., label-guided consistency ($\mathcal{L}_\text{LC}$).
}
\label{fig:overall}
\end{figure*}
To train GNNs given limited labeled information, recent studies mainly focus on leveraging pseudo-labeling techniques. Specifically, co-training~\cite{deeper}~uses Parwalks~\cite{parwalks}~to provide confident pseudo-labels to help train GNNs, and self-training~\cite{deeper} expands the label set by obtaining pseudo-labels provided by GNNs trained in advance. Moreover, M3S~\cite{M3S} leverages a clustering technique to filter out pseudo-labels that do not align with the clustering assignments for improving the pseudo-labeling accuracy.
However, pseudo-labeling-based methods suffer from an inherent limitation originated from the incorrect pseudo-labels, which eventually incur confirmation bias~\cite{confirmation}. To alleviate this issue, it is crucial to fully benefit from the given label information.
On the other hand, self-supervised methods for graphs~\cite{GRACE,GCA,BGRL} learn node representations without any requirements of labeled nodes. In particular, based on graph augmentations, contrastive learning-based methods pull positive pairs of nodes together while pushing negative ones apart, whereas consistency regularization-based methods impose GNNs to consistently output the same node representations over various perturbations of the given graph. Although these methods have achieved the state-of-the-art results in node classification even outperforming the supervised counterparts, they fall short of learning class discriminative node representations since no labeled information is utilized during training. As shown in Fig.~\ref{fig:teaser}, although recent self-supervised methods for graphs, i.e., BGRL~\cite{BGRL} and GRACE~\cite{GRACE},
outperform GCN~\cite{GCN} over various rates of labeled nodes, the performance degradation is still severe as the rate decreases.
To this end, we propose a simple yet effective semi-supervised method for graphs that fully leverages a small amount of labeled nodes to learn class discriminative node representations, called \textsf{GraFN}. The main idea is to consistently impose the representations of nodes that belong to the same class to be grouped together on differently augmented graphs. Specifically, we randomly sample support nodes from the labeled nodes and anchor nodes from the entire graph, and non-parametrically compute two predicted class distributions from two augmented graphs based on the anchor-supports similarity. By minimizing the difference between the two class distributions,~\textsf{GraFN}~not only learns augmentation invariant parameters, but also enforces the representations of nodes that belong to the same class to be grouped together. As shown in Fig.~\ref{fig:teaser},~\textsf{GraFN}~consistently outperforms both the semi-supervised and self-supervised baselines over various rates of labeled nodes, especially outperforming when the number of labeled nodes is smaller, which demonstrates the robustness of~\textsf{GraFN}.
\smallskip
\noindent\textbf{Notations}. Let $G=(V, E)$ denote a graph, where $V$ is the set of $|V|=N$ nodes and $E$ is the set of edges between the nodes. The adjacency matrix is defined by $\mathbf{A} \in \mathbb{R}^{N \times N}$ with each element $\mathbf{A}_{ij}=1$ indicating the existence of an edge between nodes $v_i$ and $v_j$, otherwise $\mathbf{A}_{ij}=0$.
The node attributes are denoted by $\mathbf{X} \in \mathbb{R}^{N \times F}$, where $F$ is the number of features of each node. Additionally, the label matrix is denoted by $\mathbf{Y}\in\mathbb{R}^{N\times C}$, where $C$ is the number of classes, and each row, i.e., $\mathbf{Y}_i\in\mathbb{R}^C$, is the one-hot label vector for node $v_i$.
We denote $V_L$ and $V_U$ as the set of labeled and unlabeled nodes, respectively.
Our goal is to accurately predict the labels of nodes that belong to $V_U$ given few labeled nodes, i.e., $|V_{L}| \ll |V_{U}|$.
\vspace{-1ex}
\begin{table*}[t]
\centering
\caption{Test Accuracy on semi-supervised node classification.}
\vspace{-2ex}
\renewcommand{\arraystretch}{0.95}
\begin{tabular}{l|ccc|ccc|ccc|ccc|ccc}
\noalign{\smallskip}\noalign{\smallskip}
\toprule
{Methods} & \multicolumn{3}{c}{Cora} & \multicolumn{3}{c}{Citeseer} & \multicolumn{3}{c}{Pubmed} & \multicolumn{3}{c}{Am. Comp} & \multicolumn{3}{c}{Am. Photos} \\
\hline
{Label Rate} & 0.5\% & 1\% & 2\% & 0.5\% & 1\% & 2\% & 0.03\% & 0.06\% & 0.1\% & 0.15\% & 0.2\% & 0.25\% & 0.15\% & 0.2\% & 0.25\% \\
\midrule
MLP & 31.24 & 37.74 & 44.53 & 32.07 & 43.07 & 46.11 & 52.50 & 55.80 & 61.22 & 40.30 & 42.22 & 49.98 & 29.76 & 31.64 & 38.55 \\
LP & 50.77 & 58.28 & 64.43 & 31.15 & 37.95 & 41.71 & 50.93 & 55.83 & 62.14 & 60.46 & 65.90 & 68.79 & 63.67 & 66.38 & 70.40 \\
GCN & 56.00 & 66.36 & 72.35 & 44.67 & 54.61 & 60.59 & 59.28 & 64.00 & 73.74 & 62.71 & 66.81 & 71.75 & 66.70 & 70.72 & 75.74\\
\hline
GAT & 58.57 & 67.75 & 72.74 & 48.70 & 58.73 & 62.71 & 63.15 & 64.11 & 73.19 & 66.17 & 70.18 & 72.82 & 73.29 & 74.46 & 80.12 \\
SGC & 49.19 & 63.60 & 69.56 & 44.02 & 55.89 & 63.61 & 58.58 & 62.50 & 71.90 & 59.69 & 64.24 & 68.29 & 55.96 & 61.64 & 69.69 \\
APPNP & 62.02 & 71.45 & 76.89 & 41.79 & 54.70 & 62.86 & 63.15 & 64.11 & 73.19 & 68.53 & 72.47 & 74.27 & 75.54 & 78.49 & 82.75 \\
GRAND & 54.51 & 70.92 & 74.90 & 46.76 & 58.40 & 65.31 & 55.87 & 61.25 & 72.42 & 68.00 & 72.71 & 75.77 & 73.80 & 75.83 & 82.33 \\
\hline
GLP & 56.94 & 68.28 & 72.97 & 41.53 & 54.84 & 63.08 & 56.70 & 60.83 & 73.46 & 62.97 & 68.56 & 70.70 & 63.18 & 67.96 & 75.19 \\
IGCN & 58.81 & 70.10 & 74.34 & 43.28 & 57.00 & 64.62 & 57.50 & 62.06 & 73.13 & 65.48 & 70.05 & 71.03 & 71.27 & 73.28 & 77.93 \\
CGPN & 64.21 & 70.54 & 72.97 & 53.90 & 63.70 & 65.15 & 64.55 & 67.58 & 71.42 & 65.37 & 67.98 & 70.77 & 74.14 & 76.89 & 81.57 \\
\hline
GRACE & 60.95 & 68.69 & 74.68 & 52.01 & 58.00 & 63.76 & 64.86 & 68.35 & \textbf{75.92} & 65.25 & 67.79 & 71.79 & 70.19 & 71.89 & 77.32 \\
BGRL & 61.74 & 68.74 & 73.65 & 54.69 & 63.75 & 67.75 & 65.77 & \textbf{68.86} & 75.91 & 68.80 & 73.04 & 75.11 & 74.27 & 78.25 & 83.12 \\
\hline
Co-training & 62.75 & 68.72 & 74.05 & 43.76 & 54.75 & 61.13 & 63.01 & 68.15 & 74.24 & 67.06 & 71.62 & 71.34 & 72.85 & 74.65 & 79.92 \\
Self-training & 57.28 & 70.73 & 75.40 & 46.26 & 60.36 & 66.47 & 57.34 & 65.13 & 72.86 & 61.32 & 65.95 & 68.66 & 61.92 & 65.24 & 71.34 \\
M3S & 64.46 & \textbf{72.93} & 76.41 & 55.07 & 65.74 & 67.64 & 61.53 & 64.60 & 73.18 & 61.51 & 66.30 & 68.10 & 63.93 & 67.62 & 73.39 \\
\hline
\hline
\textbf{\textsf{GraFN}} & \textbf{66.73} & 72.50 & \textbf{77.20} & \textbf{57.48} & \textbf{66.47} & \textbf{69.89} & \textbf{65.91} & 68.41 & 75.74 & \textbf{71.73} & \textbf{74.26} & \textbf{77.37} & \textbf{79.25} & \textbf{80.87} & \textbf{85.36} \\
\bottomrule
\end{tabular}
\vspace{-1ex}
\label{tab:main_table}
\end{table*}
\section{Proposed Method : \textsf{G\lowercase{ra}FN}}
\label{sec:method}
\smallskip
\noindent\textbf{1) Graph Augmentations and Encoding.}
Given a graph, we first generate two graph views by applying stochastic graph augmentations, which randomly mask node features and drop partial edges. Two differently augmented views are denoted by
$\tilde{G}=(\tilde{\mathbf{A}},
\tilde{\mathbf{X}})$ and $\tilde{G}^{'}=(\tilde{\mathbf{A}}^{'}, \tilde{\mathbf{X}}^{'})$.
Each augmented view is fed into a \textit{shared} GNN encoder, $f_{\theta}: {\mathbb{R}^{N \times N}} \times \mathbb{R}^{N \times F} \rightarrow \mathbb{R}^{N \times D}$, to obtain low dimensional node-level representations
$f_\theta(\tilde{\mathbf{A}}, \tilde{\mathbf{X}})=\tilde{\mathbf{Z}}\in\mathbb{R}^{N\times D}$, and $f_\theta(\tilde{\mathbf{A}}^{'}, \tilde{\mathbf{X}}^{'})=\tilde{\mathbf{Z}}^{'}\in\mathbb{R}^{N\times D}$. Note that we adopt GCN as the backbone of the GNN encoder.
\smallskip
\noindent{\textbf{2) Node-wise Consistency Regularization.}}
Then, to learn augmentation invariant node representations, we minimize the difference, i.e., cosine distance, between the representations obtained from the two differently augmented graphs in a node-wise manner:
\begin{equation}
\label{eqn:nc}
\mathcal{L}_{\text{NC}} = -\frac{1}{N}\sum_{i=1}^{N}\frac{\mathbf{\tilde{Z}}_{i}\cdot\mathbf{\tilde{Z}}^{'}_{i}}{\| \mathbf{\tilde{Z}}_{i}\| \|\mathbf{\tilde{Z}}^{'}_{i} \|}
\vspace{-1ex}
\end{equation}
Note that the above loss can be considered as a simplified version of the self-supervised loss proposed in BGRL.
The major difference is that BGRL involves two separate encoders, where one encoder is updated by minimizing the distance between the node representations obtained from the two views, while the other one is updated by the exponential moving average of the parameters of the other encoder to prevent the collapsing of node representations to trivial solutions.
On the other hand,~\textsf{GraFN}~trains only one \textit{shared} encoder, and we find that the supervisory signals incorporated in the next step help avoid the collapse of representations.
\smallskip
\noindent{\textbf{3) Label-guided Consistency Regularization.}}
Although the above self-supervised loss has been shown to be effective, the learned node representations are not class discriminative because the node label information is not involved in the training process.
We argue that unlabeled nodes can be grouped together according to their classes by enforcing them to be consistently close with a certain class of labeled nodes on differently augmented graphs.
Hence, we compute the similarity between few labeled nodes and all the nodes in the two augmented graphs, and maximize the consistency between the two similarity-based class assignments, expecting that this would help nodes that belong to the same class to be grouped together.
More precisely, we first randomly sample $b$ labeled nodes per class to construct the support set $\mathcal{S}$, and let $\mathbf{Z}^\mathcal{S}\in\mathbb{R}^{(b\times C)\times D}$ denote ($b\times C$) support node representations\footnote{To make the best use of labeled nodes in the support set $S$, we fix $b$ to the number of labeled nodes that belong to the class with the fewest nodes, i.e., minority class.}.
Then, for each anchor node $v_i\in V$, we compute the similarity distribution, i.e., predicted class distribution, $\mathbf{p}_i \in \mathbb{R}^{C}$ using anchor-supports similarity in a non-parametric manner by applying Soft Nearest Neighbors(SNN) strategy~\cite{snn, paws} as follows:
\begin{equation}
\label{eqn:snn}
\mathbf{p}_i =\sum_{(\mathbf{Z}^\mathcal{S}_{j}, \mathbf{Y}^\mathcal{S}_{j}) \in \mathbf{Z}^\mathcal{S}}{\frac{\exp{(\textsf{sim}(\mathbf{Z}_{i}, \mathbf{Z}^\mathcal{S}_{j})/\tau)}}{\sum_{\mathbf{Z}^\mathcal{S}_{k}\in\mathbf{Z}^\mathcal{S}} \exp{({\textsf{sim}(\mathbf{Z}_{i}, \mathbf{Z}^\mathcal{S}_{k})}}/\tau)}}\cdot \mathbf{Y}^\mathcal{S}_j
\end{equation}
where $\textsf{sim}(\cdot, \cdot)$ computes the cosine similarity between two vectors, and $\tau$ is a temperature hyperparameter.
$\mathbf{p}_i$ can be considered as the soft pseudo-label vector for node $v_i$ since it is derived based on the labeled nodes in $\mathcal{S}$.
Having defined the predicted class distribution as in Eqn.~\ref{eqn:snn}, we compute the predicted class distributions for each node $v_i\in V$, i.e., $\mathbf{\tilde{p}}_i$ and $\mathbf{\tilde{p}^{'}}_i$ each of which is obtained from $\tilde{G}$ and $\tilde{G}^{'}$, respectively, where the former is considered as the prediction distribution, and the latter is considered as the target distribution.
Then,
we minimize the cross-entropy between them:
\begin{equation}
\label{eqn:lc_1}
\frac{1}{|V_U|}\sum_{v_i\in V_U}H(\tilde{\mathbf{p}}_{i}^{'},\tilde{\mathbf{p}}_{i}) + \frac{1}{|V_L|}\sum_{v_i\in V_L}H(\mathbf{Y}_i,\tilde{\mathbf{p}}_{i})
\end{equation}
where $H(\mathbf{y},\mathbf{\hat{y}})$ is the cross-entropy between the target $\mathbf{y}$ and the prediction $\mathbf{\hat{y}}$.
Note that for each labeled node, instead of computing the predicted class distribution as in Eqn.~\ref{eqn:snn}, we simply assign its one-hot label vector, i.e., $\mathbf{\tilde{p}^{'}}_i=\mathbf{Y}_i, \,\forall v_i\in V_L$, to fully leverage the label information.
{However, naively minimizing the above loss would incur confirmation bias~\cite{confirmation} due to inaccurate $\mathbf{\tilde{p}^{'}}_i$ computed for unlabeled nodes (i.e., $V_U)$, which is detrimental to the performance of pseudo-labeling-based semi-supervised methods.
To this end, we introduce a confidence-based label-guided consistency regularization:}
\begin{equation}
\label{eqn:lc}
\mathcal{L}_{\text{LC}} = \frac{1}{|V_\text{conf}|}\sum_{v_i\in V_\text{conf}}H(\tilde{\mathbf{p}}_{i}^{'},\tilde{\mathbf{p}}_{i})
+ \frac{1}{|V_L|}\sum_{v_i\in V_L}H(\mathbf{Y}_i,\tilde{\mathbf{p}}_{i})
\end{equation}
where $V_\text{conf} =\{v_i|\mathbbm{1}(\max(\tilde{\mathbf{p}}_{i}^{'})>\nu) = 1, \forall v_i\in V_U\}$ is the set of nodes with confident predictions, $\nu$ is the threshold for determining whether a node has confident prediction, and $\mathbbm{1}(\cdot)$ is an indicator function.
$\tilde{\mathbf{p}}_{i}^{'}$ is considered to be confident if its maximum element is greater than $\nu$.
We argue that setting $\nu$ to a high value helps alleviate confirmation bias~\cite{confirmation, fixmatch} by enforcing only high-quality target distribution $\tilde{\mathbf{p}}_{i}^{'}$ to be able to contribute to Eqn.~\ref{eqn:lc}.
It is important to note that we apply a relatively weak augmentation for graph $\tilde{G}^{'}=(\tilde{\mathbf{A}}^{'}, \tilde{\mathbf{X}}^{'})$ that is used to compute $\tilde{\mathbf{p}}_{i}^{'}$, because aggressive augmentations (e.g., dropping more than half of edges) drastically change the semantics of a graph, which may eventually incur inaccurate $\tilde{\mathbf{p}}_{i}^{'}$.
A further benefit of the label-guided consistency regularization defined in Eqn.~\ref{eqn:lc} is that since the class distributions are computed regardless of the structural information of graphs, the label information can be effectively propagated to distant nodes, whereas existing GNNs suffer from ineffective propagation incurred by the over-smoothing issue~\cite{deeper}.
Moreover, to break the symmetry of the model architecture thereby preventing the collapsing of node representations to trivial solutions, we stop gradient for the target distribution (i.e., $\tilde{\mathbf{p}}^{'}$), and only update the parameters associated with the prediction distribution (i.e., $\tilde{\mathbf{p}}$).
\smallskip
\noindent\textbf{4) Final Objective.}
Finally, we combine $\mathcal{L}_{\text{NC}}$ and $\mathcal{L}_{\text{LC}}$ with coefficients ${\lambda_{1}}$ and ${\lambda_{2}}$ to compute the final objective function as follows:
\begin{equation}
\mathcal{L}_{\text{Training}} = \lambda_{1}\mathcal{L}_{\text{NC}} + \lambda_{2}\mathcal{L}_{\text{LC}} + \mathcal{L}_{\text{sup}}
\end{equation}
We add the cross-entropy loss, i.e., $\mathcal{L}_\text{sup}$, defined over a set of labeled nodes $V_L$.
The overall pipeline of~\textsf{GraFN}~is shown in Fig.~\ref{fig:overall}.
\begin{table}[t]
\centering
\caption{Statistics for datasets used for experiments.}
\vspace{-2ex}
\renewcommand{\arraystretch}{0.95}
\begin{tabular}{c|ccccc}
\noalign{\smallskip}\noalign{\smallskip}\hline
& \# Nodes & \# Edges & \# Features & \# Classes \\
\hline
Cora & 2,708 & 5,429 & 1,433 & 7 \\
Citeseer & 3,327 & 4,732 & 3,703 & 6 \\
Pubmed & 19,717 & 44,338 & 500 & 3 \\
Am. Comp. & 13,752 & 245,861 & 767 & 10 \\
Am. Photos & 7,650 & 119,081 & 745 & 8 \\
\hline
\end{tabular}
\label{tab:dataset}
\vspace{-1ex}
\end{table}
\section{EXPERIMENTS}
\noindent{\textbf{Datasets.}}
To verify the effectiveness of~\textsf{GraFN},
we conduct extensive experiments on five widely used datasets (Table~\ref{tab:dataset}) including three citation networks (Cora, Citeseer, Pubmed)~\cite{cora_cite_pub} and two co-purchase networks (Amazon Computers, Amazon Photos)~\cite{amazon}.
\smallskip
\noindent{\textbf{Baselines.}}
We compare~\textsf{GraFN}~with fifteen baseline methods. (i) \emph{Two conventional methods}: \textbf{MLP} and \textbf{LP}~\cite{LPA}. (ii) \emph{Five graph convolution-based methods}: \textbf{GCN}~\cite{GCN} and \textbf{GAT}~\cite{GAT}.
\textbf{SGC}~\cite{SGC} simplifies GCN by removing repeated feature transformations and nonlinearities. \textbf{APPNP}~\cite{APPNP} and \textbf{GRAND}~\cite{GRAND} alleviate the limited receptive field issue of existing message passing models. (iii) \emph{Two self-supervised methods}: \textbf{GRACE}~\cite{GRACE} and \textbf{BGRL}~\cite{BGRL} maximize the agreement of the representations of the same nodes from two augmented views\footnote{For fair comparisons with~\textsf{GraFN}, we extend them to semi-supervised setting by adding the conventional supervised loss, i.e., cross-entropy loss.}.
(iv) \emph{Three label-efficient methods}: \textbf{GLP}~\cite{GLP_IGCN} and \textbf{IGCN}~\cite{GLP_IGCN} apply a low-pass graph filter to message propagation to achieve label efficiency. \textbf{CGPN}~\cite{cgpn} leverages Graph Poisson Network to effectively spread limited labels to the whole graph, and also utilizes contrastive loss to leverage information on unlabeled nodes. (v) \emph{Three pseudo-labeling-based methods}: \textbf{Co-training}~\cite{deeper}, \textbf{Self-training}~\cite{deeper}
and \textbf{M3S}~\cite{M3S}.
\smallskip
\noindent{\textbf{Evaluation Protocol.}}
We randomly create 20 splits to evaluate the effectiveness of~\textsf{GraFN}~on practical few-label setting. For citation networks, we closely follow the evaluation protocol of~\cite{deeper, investigating}.
For co-purchase networks, we evaluate models with \{0.15\%, 0.2\%, 0.25\%\} training label rates
to match with the average number of labeled nodes per class of the citation networks.
The remaining nodes are split into 1:9, and each split is used for validation and test, respectively. We report the averaged test accuracy when the validation accuracy is the highest.
For baseline methods including~\textsf{GraFN}, we search hidden dimension, learning rate, weight decay, and dropout ratio in \{32, 64, 128\}, \{0.1, 0.01, 0.05, 0.001, 0.005\}, \{1e-2, 1e-3, 1e-4, 5e-4\}, and \{0, 0.3, 0.5, 0.8\}, respectively, while other hyperparameter configurations are taken from each work.
Additionally, we conduct an extensive grid search for unreported hyperparameters for fair comparisons.
As mentioned in Sec.~\ref{sec:method}, for~\textsf{GraFN}, we apply relatively weaker augmentations on $\tilde{G}^{'}$ compared with $\tilde{G}$, and thus we search the augmentation hyperparameters of node feature masking and partial edge dropping in \{0.2, 0.3, 0.4\} for $\tilde{G}^{'}$, and \{0.5, 0.6\} for $\tilde{G}$. {Moreover, the temperature $\tau$ is fixed to 0.1}, the threshold $\nu$ is searched in \{0.8, 0.9\}, and the balance hyperparameters, i.e., $\lambda_{1}$ and $\lambda_{2}$, are searched in \{0.5, 1.0, 2.0\}.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{0.99}
\caption{Performance on similarity search. (Sim@$K$: Average ratio among $K$-NNs sharing the same label as the query node.)}
\vspace{-1ex}
\begin{tabular}{p{1.3cm}|p{1.0cm}|ccc}
\multicolumn{2}{c|}{} & GRACE & BGRL &~\textsf{GraFN}~\\ \hline \hline
\multirow{2}{*}{Cora} & Sim@5 & 0.8146 & 0.8047 & \textbf{0.8222} \\
& Sim@10 & 0.7947 & 0.7823 & \textbf{0.7984} \\ \hline
\multirow{2}{*}{Citeseer} & Sim@5 & 0.6407 & 0.6623 & \textbf{0.6810} \\
& Sim@10 & 0.6147 & 0.6396 & \textbf{0.6621} \\ \hline
\multirow{2}{*}{Pubmed} & Sim@5 & 0.7571 & \textbf{0.7815} & 0.7110 \\
& Sim@10 & 0.7493 & \textbf{0.7733} & 0.7015 \\ \hline
\multirow{2}{*}{Am.Comp} & Sim@5 & 0.8091 & 0.8335 & \textbf{0.8351} \\
& Sim@10 & 0.7965 & 0.8211 & \textbf{0.8246} \\ \hline
\multirow{2}{*}{Am.Photos} & Sim@5 & 0.8831 & 0.8886 & \textbf{0.9004} \\
& Sim@10 & 0.8761 & 0.8881 & \textbf{0.8937} \\ \hline
\end{tabular}
\vspace{-3ex}
\label{tab:similarity_search}
\end{table}
\subsection{Performance Analysis}
Table ~\ref{tab:main_table} shows the performance of various methods in terms of node classification over various label rates. We have the following observations:
\textbf{1)}~\textsf{GraFN}~outperforms methods that use advanced GNN encoders (i.e., APPNP and GRAND), and label efficient GNNs that effectively propagate label information to distant nodes (e.g., GLP and IGCN) and leverage unlabeled nodes with contrastive loss (i.e., CGPN).
We argue that the label-guided consistency regularization helps ~\textsf{GraFN}~ to effectively propagate the label information to distant nodes, thereby learning class discriminative representations. Note that~\textsf{GraFN}~especially outperforms other methods when the label rate is small demonstrating the robustness of~\textsf{GraFN}~(also refer to Fig.~\ref{fig:teaser}).
\textbf{2)} Even though~\textsf{GraFN}~adopts a simple self-supervised loss with a single shared GNN-based encoder, i.e. node-wise consistency regularization,~\textsf{GraFN}~outperforms advanced self-supervised methods, i.e. GRACE and BGRL. We attribute this to the label-guided consistency loss that groups nodes that belong to the same class by leveraging the given few labeled nodes.
\textbf{3)} To empirically verify the benefit of the label-consistency regularization of~\textsf{GraFN}, we compare the similarity search results of GRACE, BGRL and~\textsf{GraFN}~in Table ~\ref{tab:similarity_search}. Since GRACE and BGRL consider node labels through conventional supervised loss with advanced self-supervised loss, whereas~\textsf{GraFN}~does so through the label-guided consistency regularization with a simple self-supervised loss, the superior performance of~\textsf{GraFN}~implies the benefit of the label-guided consistency regularization of~\textsf{GraFN}~despite its simple self-supervised loss. We indeed observe that~\textsf{GraFN}~generally outperforms GRACE and BGRL, which corroborates the benefit of the label-consistency regularization for learning class discriminative node representations.
\textbf{4)} We observe that the pseudo-labeling-based methods, i.e., Co-training, Self-training and M3S, generally outperform vanilla GCN, which solely relies on the given label information. On the other hand,~\textsf{GraFN}~outperforms these methods without relying on the pseudo-labeling techniques. This implies that the pseudo-labeling techniques should be adopted with particular care as they may introduce incorrect labels.
\textbf{5)}~\textsf{GraFN}~shows relatively low performance on Pubmed dataset, which contains a small number of classes, i.e., 3 classes.
Since~\textsf{GraFN}~assigns class distributions based on the similarity with labeled nodes in different classes, having more classes leads to more class discriminative node representations.
\begin{figure}[t]
\includegraphics[width=0.99\linewidth]{Img/Figure3.pdf}
\vspace{-3ex}
\caption{Accuracy of pseudo-labeling and node classification.}
\label{fig:pseudo}
\vspace{-3ex}
\end{figure}
\smallskip
\noindent{\textbf{Adopting Pseudo-Labeling to~\textsf{GraFN}.}}
Since~\textsf{GraFN}~learns class discriminative node representations with few labels, we hypothesized that the confirmation bias~\cite{confirmation} of the pseudo-labeling technique suffered by existing methods would be alleviated when the pseudo-labeling technique is adopted to~\textsf{GraFN}.
Fig.~\ref{fig:pseudo}(a) indeed demonstrates that adopting the pseudo-labeling technique to~\textsf{GraFN}~gives the best pseudo-labeling accuracy, which in turn results in further improvements of~\textsf{GraFN}~in terms of node classification (Fig.~\ref{fig:pseudo}(b)).
\smallskip
\noindent{\textbf{Performance Comparison on Different Node Degrees. }}
In most real-world graphs, the node degrees follow a power-law distribution, which means that the majority of nodes are of low-degree.
Since the training of GNNs is based on the neighborhood aggregation scheme, high-degree nodes receive more information than low-degree nodes, which eventually leads to the models overfit to high-degree nodes and underfit to low-degree nodes~\cite{investigating}.
Fig.~\ref{fig:Degree} demonstrates that the node classification accuracy is indeed highly correlated with the node degree, i.e., high-degree nodes tend to result in better classification performance.
Furthermore, we observe that~\textsf{GraFN}~greatly outperforms other methods for low-degree nodes, while showing a comparable performance for high-degree nodes (i.e. degree $\geq$ 7).
We attribute the superior performance of~\textsf{GraFN}~on low-degree nodes to the label-guided consistency regularization that
promotes the supervision information to be evenly spread over the unlabeled nodes regardless of their node degree.
\begin{figure}[t]
\includegraphics[width=0.99\linewidth]{Img/Figure4.pdf}
\vspace{-1ex}
\caption{Node classification results on various node degrees.}
\label{fig:Degree}
\vspace{-1ex}
\end{figure}
\begin{figure}[h]
\vspace{-1ex}
\includegraphics[width=0.99\linewidth]{Img/Figure5.pdf}
\vspace{-1ex}
\caption{Ablation study on~\textsf{GraFN}.}
\vspace{-3ex}
\label{fig:ablation}
\end{figure}
\subsection{Ablation Studies}
To evaluate each component of \textsf{GraFN}, we conduct ablation studies on Cora and Computers datasets on the lowest label rate, i.e. 0.5\%, 0.15\%, respectively. In Fig~\ref{fig:ablation}, we have the following observations:
\textbf{1)} Only using the supervised cross-entropy loss shows poor performance because it suffers from overfitting and ineffective propagation of supervisory signal.
\textbf{2)} Using both node-wise and label-guided consistency regularization (i.e.,~\textsf{GraFN}) is more beneficial than using either one of them. We argue that this is because these two losses are complementary. More precisely, using only the node-wise loss cannot fully leverage the given label information, whereas using only the label-guided regularization loss can suffer from inaccurate target distribution, which incurs confirmation bias.
\section{CONCLUSIONS}
In this paper, we present a novel semi-supervised method for graphs that learns class discriminative node representations with only few labeled nodes.~\textsf{GraFN}~ not only exploits the self-supervised loss, i.e. node-wise consistency regularization, but also ensures nodes that belong to the same class to be grouped together by enforcing unlabeled nodes to be consistently close with a certain class of labeled nodes on differently augmented graphs. Through extensive experiments on real-world graphs, we show that~\textsf{GraFN}~outperforms existing state-of-the art methods given few labeled nodes. Moreover, it is worth noting that~\textsf{GraFN}~1) alleviates underfitting problem of low-degree nodes by propagating label information to distant nodes, and 2) enjoys further improvements by adopting the pseudo-labeling technique.
\smallskip
\noindent\textbf{Acknowledgements.}
This work was supported by the NRF grant funded by
the MSIT (No.2021R1C1C1009081), and the IITP grant
funded by the MSIT (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)).
\bibliographystyle{ACM-Reference-Format}
\balance
|
\section{Introduction}
Objective of the paper of to study the behavior of free boundary near the fixed boundary of domain, for minimizers of Bernoulli type functionals with Hölder continuous coefficients.
\begin{equation}\label{functional}
J(v;A,\lambda_+,\lambda_-,\Omega)=\int_{\Omega} \Big ( \< A(x)\nabla v,\nabla v \>+ \Lambda(v)\Big )\,dx
\end{equation}
$A$ is an elliptic matrix with Hölder continuous entries, and $\Lambda(v)=\lambda_+ \chi_{\{v>0\}}+\lambda_- \chi_{\{v\le0\}}$. We prove that if the value of boundary data and its derivative at a point are equal to zero (i.e. it satisfies the \eqref{DPT} condition mentioned below), then the contact of free boundary and the fixed boundary is tangential.
Boundary interactions of free boundaries have gained significant attention in recent years. Whenever there are two medias involved, interactions of their respective diffusions can be modeled by free boundary problems. Often, free boundary of solution and fixed boundary of set come in contact. In applications, the Dam problem \cite{AG82} and Jets, Wakes and Cavities \cite{ZG57} model phenomenas which involve understanding of free boundary and fixed boundary.
Very recently, works of Indrei \cite{I00}, \cite{I19} study interactions of free boundaries and fixed boundaries for fully non-linear obstacle problems. We refer to \cite{GL19} where authors shed more light into angle of contact between fixed boundary and free boundary for one phase Bernoulli problem.
As it is common by now, our strategy in this article is to classify blow up of minimizers by using the ideas from \cite{KKS06}. We prove that the blow up and also their positivity sets converge to a global solution in $\mathcal{P}_{\infty}$ as defined in \cite{KKS06}. In Section \ref{setting up problem}, we list the assumptions and set some notations and then in Section \ref{blow up analysis}, we prove that blow ups of minimizers converge to that of global solutions (c.f. Definition 2.6). In the last section we prove our main result.
\section{Setting up the problem}\label{setting up problem}
We consider the following class of function which we denote as $\mathcal{P}_r(\alpha, M,\lambda, \mathcal{D},\mu)$. Before definition, we set the following notation
\[
\begin{split}
B_R^+ := \Big \{ x\in B_R\mbox{ such that } x_N>0 \Big \}\\
B_R' := \Big \{ x\in B_R\mbox{ such that } x_N=0 \Big \}.
\end{split}
\]
For $x\in \mathbb{R}^N$ we denote $x'\in \mathbb{R}^{N-1}$ as the projection of $x$ on the plane $\{x_N=0\}$, we denote the tangential gradient $\nabla'$ as follows
$$
\nabla' u := \Big ( \frac{\partial u}{\partial x_1}, ...\,\frac{\partial u}{\partial x_{N-1}} \Big ).
$$
We define the affine space set $H^1_\phi(B_R^+)$ as follows,
\begin{equation}\label{Kf}
H^1_\phi(B_R^+) = \left \{ v\in H^1(B_R^+) \,:\, v-\phi\in H_0^1(B_R^+) \right \}.
\end{equation}
For a given function $v\in H^1(B_2^+)$, we denote $F(v)$ as $F(v):=\partial \{v>0\}$ and $Id$ is the notation for $N\times N$ identity matrix.
\begin{definition}
A function $u\in H^1(B_{2/r}^+)$ is said to belong to the class $\mathcal{P}_r(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$ if there exists $A\in C^{\alpha}(B_{2/r}^+)^{N\times N}$, $\phi \in C^{1,\alpha }(B_{2/r}^+)$, $ \lambda_{\pm}>0$, $0<\mu<1$ and $\mathcal{D}>0$ such that
\begin{enumerate}[label=\textbf{$($P\arabic*$)$}]
\item \label{P1} $\Linfty{A}{B_{2/r}^+}\le M$, $\Linfty{\nabla \phi}{B_{2/r}^+}\le M$, $[A]_{C^{\alpha}(B_{2/r}^+)},[\nabla \phi]_{C^{1,\alpha}(B_{2/r}^+)}\le r^{\alpha}M$ \newline and $|\phi(x')| \le Mr^{1+\alpha}|x'|^{1+\alpha}$ $($$x'\in B_{2/r}'$$)$. $\phi$ satisfies the following Degenerate Phase Transition condition (\ref{DPT}) mentioned below.
\begin{equation}\label{DPT}\tag{DPT}
\mbox{$\forall x'\in B_{2/r}'$ such that $\phi (x')=0$, then $|\nabla' \phi(x')|=0$}.
\end{equation}
\item \label{P2} $A(0)= Id$, $\mu|\xi|^2 \le \< A(x)\xi,\xi\> \le \frac{1}{\mu}|\xi|^2$ for all $x\in B_{2/r}^+$ and $\xi \in \mathbb{R}^N$.
\item \label{P3} $0<\lambda_-<\lambda_+$.
\item \label{P4} $u$ minimizes $J(\cdot; A,\lambda_+,\lambda_-,B_{2/r}^+)$ $($c.f. \eqref{functional}$)$ that is for every $u-v\in H_0^1(B_{2/r}^+)$
$$
\int_{B_{2/r}^+} \Big ( \< A(x) \nabla u,\nabla u \>+ \Lambda (u) \Big ) \,dx \le \int_{ B_{2/r}^+} \Big ( \< A(x) \nabla v,\nabla v \>+ \Lambda (v) \Big )\,dx \;\;
$$
$ (\Lambda(s)=\lambda_+ \chi_{\{s>0\}}+\lambda_- \chi_{\{s\le0\}} )$ and $0\in F(u)\cap \overline{B_{2/r}^+}$.
\item \label{P5} $u \in H_{\phi}^1(B_{2/r}^+)$.
\item \label{P6} There exists $0<r_0$ such that for all $0<\rho\le r_0$ we have
\begin{equation}\label{density}
\frac{|B_\rho^+(0)\cap \{u>0\}|}{|B_\rho^+(0)|}>\mathcal{D}.
\end{equation}
\end{enumerate}
\end{definition}
\begin{remark}
In fact, the functions $u\in \mathcal{P}_r(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$ carry more regularity than being only a Sobolev function. They are Hölder continuous in $ (\overline{B_{2/r}^+})$ (c.f. Lemma \ref{lipschitz}).
\end{remark}
In the absence of ambiguity on values of $\alpha, M,\lambda_{\pm}, \mathcal{D},\mu$ we use the notation $\mathcal{P}_r$ in place of $\mathcal{P}_r(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$. If $\phi\in C^{1,\alpha}(B_2^+)$ satisfies (\ref{DPT}), from \cite[Lemma 10.1]{BM21}, we know that $\phi^{\pm} \Big |_{B_2'} \in C^{1,\alpha}(B_2')$ and also
$$
\| \phi^{\pm} \|_{C^{1,\alpha}(B_2')} \le \| \phi \|_{C^{1,\alpha}(B_2')}.
$$
Given $v\in H^1(B_R^+)$ and $r>0$, we define the blow-up $v_r\in H^1(B_{R/r}^+)$ as follows
\begin{equation}\label{blowup}
v_r(x):=\frac{1}{r}v(rx).
\end{equation}
For the coefficient matrix $A$, $A^r(x)$ is defined as follows
\begin{equation}\label{Ar}
A^r(x):=A(rx).
\end{equation}
One can check that if $u\in \mathcal{P}_1(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$, then $u_r \in \mathcal{P}_r(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$. Indeed if $u\in \mathcal{P}_{1}$ and $u$ minimizer the functional $J$ (c.f. \ref{P4})
$$
J(v;A,\lambda_+,\lambda_-,B_2^+):= \int_{B_2^+}\Big ( \< A(x)\nabla v,\nabla v \> + \Lambda(v)\Big )\,dx,\qquad (\Lambda(s)=\lambda_+ \chi_{\{s>0\}}+\lambda_- \chi_{\{s\le0\}} )
$$
with boundary data $\phi\in C^{1,\alpha}(B_2^+)$ (i.e. $u\in H_{\phi}^1(B_2^+)$). Then by simple change of variables we can check that $u_r\in H^1_{\phi_r}(B_{2/r}^+)$ (this verifies \ref{P5}) and $u_r$ minimizes
$$
J(v;A^r,\lambda_+,\lambda_-,B_{2/r}^+):= \int_{B_{2/r}^+}\Big ( \< A^r(x)\nabla v,\nabla v \> + \Lambda(v)\Big )\,dx,\qquad (\Lambda(s)=\lambda_+ \chi_{\{s>0\}}+\lambda_- \chi_{\{s\le0\}} ).
$$
Moreover, if $A$ and $\phi$ satisfy the conditions \ref{P1}, \ref{P2} for $r=1$, then $A^r$ and $\phi_r$ satisfy \ref{P1}, \ref{P2} for $r$. \ref{P3} and \ref{P6} remains invariant under the change variables. Therefore $u_r\in \mathcal{P}_r$.
In order to study the blow-up limits ($\lim_{r\to 0} u_r$) of functions $u\in \mathcal{P}_1(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$, we define a class of global solutions $\mathcal{P}_{\infty}(C,\lambda_{\pm})$. Let us set the following notation before giving the definition
$$
\Pi:= \{x \in \mathbb{R}^N \,:\, x_N=0\}.
$$
\begin{definition}[Global solution]\label{global solution}
We say that $u\in H^1(\mathbb{R}^N_+)$ belongs to the class $\mathcal{P}_{\infty}(C, \lambda_{\pm})$, that is, $u$ is a global solution if there exists $C>0$ and $0<\lambda_+<\lambda_-$ such that
\begin{enumerate}[label=\textbf{(G\arabic*)}]
\item \label{G1} $|u(x)|\le C|x|$, for all $x\in \mathbb{R}^N_+$,
\item \label{G2}$u$ is continuous up to the boundary $\Pi$,
\item \label{G3} $u=0$ on $\Pi$,
\item \label{G4} and for every ball $B_r(x_0)$, $u$ is a minimizer of $J(\cdot; {Id,\lambda_+,\lambda_-,B_r(x_0)\cap \mathbb{R}^N_+})$ $($c.f. \eqref{functional}$)$, that is
$$
\int_{B_r(x_0)\cap \mathbb{R}^N_+} \Big ( |\nabla u|^2 + \Lambda(u)\Big )\,dx \le \int_{B_r(x_0)\cap \mathbb{R}^N_+} \Big ( |\nabla v|^2 +\Lambda(v) \Big )\,dx.
$$
Here $(\Lambda(s)=\lambda_+ \chi_{\{s>0\}}+\lambda_- \chi_{\{s\le0\}} )$ and for every $v\in H^1(B_r(x_0)\cap \mathbb{R}^N_+)$ such that $u-v\in H_0^1(B_r(x_0)\cap \mathbb{R}^N_+)$.
\end{enumerate}
\end{definition}
Our main result intends to show that for a minimizer $u$ of $J(\cdot; A,\lambda_+,\lambda_-,B_2^+)$ with $A,\lambda_{\pm}$ and $u$ satisfying the properties \ref{P1}-\ref{P6}, the free boundary of every such minimizer touches the flat part of fixed boundary tangentially at the origin. For this, we prove that as we approach closer and closer to the origin, the free boundary points cannot lie outside any cone which is perpendicular to the flat boundary and has its tip at the origin. The main result in this paper is stated below.
\begin{theorem}\label{main result}
There exists a constant $\rho_0$ and a modulus of continuity $\sigma $ such that if
$$u\in\mathcal{P}_1(\alpha, M,\lambda_{\pm}, \mathcal{D},\mu)$$
then
$$
F(u)\cap B_{\rho_0}^+ \subset \{ x\,:\, x_N\le \sigma (|x|)|x| \}
$$
Here $\sigma$ depends only on $\alpha, M,\lambda_{\pm}, \mathcal{D},\mu$.
\end{theorem}
\section{Blow-up analysis}\label{blow up analysis}
The following is a classical result (c.f. \cite[Remark 4.2]{AC81}) , we present the proof for the case of variable coefficients.
\begin{lemma}\label{subharmonic}
Given a strictly elliptic matrix and bounded $A(x)$ and a non-negative continuous function $w$ such that $\dive(A(x)\nabla w)=0$ in $\{w>0\}\cap B_2^+$, then $w\in H^1_{loc}(B_2^+)$ and $\dive(A(x)\nabla w)\ge 0$ in weak sense in $B_2^+$.
\end{lemma}
\begin{proof}
Let $D\Subset B_2^+$ and $\eta\in C_c^{\infty}(B_2^+)$ be cutoff function for $D$. That is $\eta\in C_c^{\infty}(B_2^+)$ be such that
$$
\eta(x) = \begin{cases}
1 \;\; \mbox{in $D$}\\
0 \;\; \mbox {on $\partial B_2^+$}.
\end{cases}
$$
Since $\dive(A(x)\nabla w)=0$ in $\{w>0\}$, we have
\[
\begin{split}
0&=\int_{B_2^+}\< A(x)\nabla w, \nabla \big ( (w-{\varepsilon})^+\eta^2 \big )\>\,dx\\
&=\int_{B_2^+\cap \{w>{\varepsilon}\}} \eta^2 \< A(x) \nabla w, \nabla w \>\,dx+ \int_{B_2^+\cap \{w>{\varepsilon}\}} w \< A(x) \nabla w, \nabla \eta^2 \>\,dx+{\varepsilon}\int_{B_2^+ \cap \{w>{\varepsilon}\}} \< A(x)\nabla w, \nabla \eta^2 \>\,dx
\end{split}
\]
which implies
\[
\begin{split}
\int_{B_2^+ \cap \{w>{\varepsilon}\}} \< A(x)\nabla w, \nabla w\> \eta^2 \,dx \le \int_{B_2^+\cap \{w>{\varepsilon}\}} & \Big | w \< A(x) \nabla w, \nabla \eta^2 \>\Big | \,dx \\&+{\varepsilon}\int_{B_2^+ \cap \{w>{\varepsilon}\}} \Big | \< A(x)\nabla w, \nabla \eta^2 \>\Big | \,dx.
\end{split}
\]
By the choice of $\eta$ and ellipticity of the matrix $A$, we obtain using Young's inequality
\[
\begin{split}
\mu \int_{B_2^+ \cap \{w>{\varepsilon}\}} |\nabla w|^2 \eta^2\,dx &\le \mu \Big | \int_{B_2^+ \cap \{w>{\varepsilon}\}} \eta w |\nabla w| |\nabla \eta| \,dx \Big | + {\varepsilon} \mu \Big | \int_{B_2^+ \cap \{w>{\varepsilon}\}} \eta |\nabla w| |\nabla \eta|\,dx \Big | \\
&\le C_1 (\mu) \Bigg [ \frac{1}{\delta}\int_{B_2^+ \cap \{w>{\varepsilon}\}} w^2 |\nabla \eta|^2 \,dx + \delta \int_{B_2^+ \cap \{w>{\varepsilon}\}} \eta^2 |\nabla w|^2 \,dx\\
&\qquad \qquad \qquad \qquad + \delta {\varepsilon} \int_{B_2^+ \cap \{w>{\varepsilon}\}} \eta^2 |\nabla w|^2\,dx +\frac{{\varepsilon}}{\delta} \int_{B_2^+ \cap \{w>{\varepsilon}\}} |\nabla \eta|^2\,dx \Bigg]
\end{split}
\]
after choosing of $\delta>0$ very small and rearranging the terms in the equation above, since $\eta=1$ in $D$ we finally get
\[
\begin{split}
\int_{D\cap \{w>{\varepsilon}\}} |\nabla w|^2\,dx & \le \int_{B_2^+ \cap \{w>{\varepsilon}\}} |\nabla w|^2 \eta^2\,dx\\
& \le C(\mu) \Bigg [ \int_{B_2^+ \cap \{w>{\varepsilon}\}} w^2 |\nabla n|^2 \,dx + \int_{B_2^+ \cap \{w>{\varepsilon}\}} |\nabla \eta|^2\,dx \Bigg ]
\end{split}
\]
As ${\varepsilon}\to 0$, we obtain
\[
\begin{split}
\int_{D}|\nabla w|^2\,dx = \int_{\{w>0\} \cap D}|\nabla w|^2\,dx &= \lim_{{\varepsilon}\to 0} \int_{\{w>{\varepsilon}\}\cap D} |\nabla w|^2\,\\
&\le C(\mu) \lim_{{\varepsilon}\to 0} \Bigg [ \int_{B_2^+ \cap \{w>{\varepsilon}\}} w^2 |\nabla n|^2 \,dx + \int_{B_2^+ \cap \{w>{\varepsilon}\}} |\nabla \eta|^2\,dx \Bigg ]\\
&\le C(\mu,D) \Bigg [ \int_{B_2^+\cap \supp(\eta)}w^2 \,dx + 1 \Bigg ].
\end{split}
\]
Since $w\in C(B_2^+)$ therefore, $w$ is uniformly bounded in $\supp(\eta)$ and therefore
$$
\int_{D}\Big ( |\nabla w|^2 + |w^2| \Big )\,dx \le C(\mu,D) \Bigg [ \int_{B_2^+\cap \supp(\eta)}w^2 \,dx + 1 \Bigg ] \le C(\mu, D, \| w \|_{L^{\infty}(\supp(\eta))}).
$$
Now, for $0\le \varphi \in C_c^{\infty}(B_2^+)$ consider the test function
$$
v = \varphi \Big ( 1 - \big (\min (2-\frac{w}{{\varepsilon}} , 1) \big ) ^+ \Big ).
$$
$$
\int_{B_2^+}\< A(x) \nabla w, \nabla \varphi \>\,dx=\int_{B_2^+}\< A(x)\nabla w, \nabla \big ( \varphi ((2-\frac{w}{{\varepsilon}})\wedge 1 )^+ )\big )\>\,dx
$$
We can easily check that $v\ge 0$ in $B_2^+$ and $v\in H_0^1(B_2^+)$, in particular
$$
\varphi \Big ((2-\frac{w}{{\varepsilon}})\wedge 1 \Big )^+ =
\begin{cases}
\varphi \;\;\;\qquad \qquad \;\; \qquad \qquad x\in \{w\le {\varepsilon}\},\\
\varphi \cdot \Big ( 2-\frac{w}{{\varepsilon}} \Big )\; \qquad\;\; \;\qquad x\in \{{\varepsilon}<w\le 2{\varepsilon}\},\\
0\;\;\;\; \qquad \qquad \qquad \qquad \;\; x\in \{w>2{\varepsilon}\}.
\end{cases}
$$
Therefore, we have
\[
\begin{split}
\int_{B_2^+}\< A(x) \nabla w, \nabla \varphi \>\,dx&=\int_{B_2^+}\< A(x)\nabla w, \nabla \big ( \varphi ((2-\frac{w}{{\varepsilon}})\wedge 1 )^+ )\big )\>\,dx\\
&= \int_{B_2^+\cap \{w\le {\varepsilon}\}} \< A(x)\nabla w,\nabla \varphi \>\,dx + \int_{B_2^+\cap \{{\varepsilon}<w\le 2{\varepsilon}\}} \< A(x)\nabla w,\nabla \Big ( \varphi \big ( 2- \frac{w}{{\varepsilon}}\big ) \Big ) \>\,dx\\
&= \int_{B_2^+\cap \{w\le {\varepsilon}\}} \< A(x)\nabla w,\nabla \varphi \>\,dx +2\int_{B_2^+\cap \{{\varepsilon}<w\le 2{\varepsilon}\}} \< A(x)\nabla w,\nabla \varphi \>\,dx\\
&\qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad-\frac{2}{{\varepsilon}} \int_{B_2^+\cap \{{\varepsilon} <w\le 2{\varepsilon}\}} \< A(x)\nabla w,\nabla (w\varphi) \>\,dx\\
&\le C(\mu) \int_{B_2^+ \cap \{{\varepsilon} <u \le 2{\varepsilon}\}} |\nabla w||\nabla \varphi|\,dx+ \frac{2}{{\varepsilon}} \int_{B_2^+ \cap \{{\varepsilon} <w\le 2{\varepsilon}\}}w\< A(x)\nabla w,\nabla \varphi \>\,dx\\
&\qquad \qquad \qquad \qquad\qquad \qquad -\frac{2}{{\varepsilon}} \int_{B_2^+ \cap \{ {\varepsilon} <w\le 2{\varepsilon}\}} \varphi \< A(x)\nabla w,\nabla w\> \,dx\\
\le & C(\mu) \int_{B_2^+ \cap \{{\varepsilon}<w\le 2{\varepsilon}\}} |\nabla w||\nabla \varphi|\,dx
\end{split}
\]
The last term goes to zero as ${\varepsilon}\to 0$. Therefore, we can say that
$$
\int_{B_2^+}\< A(x) \nabla w, \nabla \varphi \>\,dx\le 0.
$$
This concludes the proof.
\end{proof}
\begin{lemma}[Hölder continuity]\label{lipschitz}
If $u\in \mathcal{P}_1$ then $u\in C^{0,\alpha_0}(\overline {B_2^+})$ for some $0<\alpha_0<1$. In fact,
$$
\|u\|_{C^{\alpha_0}(B_2^+)} \le C(\mu, \lambda_{\pm})\Linfty{u}{B_2^+}.
$$
\end{lemma}
\begin{proof}
The functional $J(\cdot;A, \lambda_+,\lambda_-,B_2^+)$ satisfy the hypothesis of \cite[Theorem 7.3]{eg05} and $\phi\in C^{1,\alpha}(\overline{B_2^+})$, therefore Lemma \ref{lipschitz} follows from the arguments in \cite[Section 7.8]{eg05}.
\end{proof}
\begin{remark}
Since every function $u\in \mathcal{P}_1$ is continuous. Therefore, the positivity set $\{u>0\}$ is an open set.
\end{remark}
\begin{corollary}\label{linear growth}
If $u\in \mathcal{P}_1$, then $u^{\pm}$ are $A$-subharmonic.
\end{corollary}
\begin{proof}
The claim follows directly from Lemma \ref{lipschitz} and Lemma \ref{subharmonic}.
\end{proof}
\begin{lemma}\label{lemma 3.5}
If $u \in \mathcal{P}_1$ then
\begin{equation}\label{lg}
|u(x)| \le C(\mu)M \,|x| \mbox{ in $B_1^+$.}
\end{equation}.
\end{lemma}
\begin{proof}
Let $w$ be such that
$$
\begin{cases}
\dive(A(x)w) = 0,\qquad \mbox{in $B_2^+$}\\
w = \phi^+,\qquad \qquad\qquad\mbox{in $\partial B_2^+$}.
\end{cases}
$$
Since $u$ is $A$-subharmonic in $B_2^+$ (c.f. Corollary \ref{linear growth}), by maximum principle, if $x\in B_1^+$ we have
\begin{equation}\label{borsuk 2}
\begin{split}
u^+(x)\le w(x)&\le w(x)-w(x')+w(x')\\
&\le \|\nabla w\|_{L^{\infty}(B_1^+)}|x-x'|+|\phi^+(x')|\\
&=\|\nabla w\|_{L^{\infty}(B_1^+)}x_N+|\phi^+(x')|\\
&\le (\|\nabla w\|_{L^{\infty}(B_1^+)} + M)|x|\;\forall x\in B_1^+
\end{split}
\end{equation}
in the last inequality, we have used \ref{P1}. Now, we prove that the term $\|\nabla w\|_{L^{\infty}(B_1^+)}$ is uniformly bounded.
From \cite[Theorem 2]{B98}, we have
\begin{equation} \label{borsuk 1}
\|\nabla w\|_{L^{\infty}(B_1^+)}\le {C(\mu)} \Big [ \|w\|_{L^{\infty}(B_2^+)} + \|\phi^+ \|_{C^{1,\alpha}(B_2')} \Big ].
\end{equation}
By comparison principle, $\| w\|_{L^{\infty}(B_2^+)} = \| w\|_{L^{\infty}(\partial B_2^+)}= \| \phi^+ \|_{L^{\infty}(\partial B_2^+)}\le M$. Plugging this information in \eqref{borsuk 1}
$$
\|\nabla w\|_{L^{\infty}(B_1^+)}\le C(\mu)M
$$
and then using \eqref{borsuk 2}, we obtain
\begin{equation}\label{u+}
u^+(x) \le C(\mu) M|x|\;\forall \;x\in B_1^+.
\end{equation}
And analogously,
\begin{equation}\label{u-}
u^-(x) \le C(\mu) M|x|\;\forall \;x\in B_1^+.
\end{equation}
By adding \eqref{u+} and \eqref{u-}, we prove \eqref{lg}.
\end{proof}
\begin{remark}\label{linear growth for blowup}
We can check that for every $u\in \mathcal{P}_1$, $u_r\in \mathcal{P}_r$ and $u_r$ is $A^r$-subharmonic and satisfies \eqref{lg} in $B_{1/r}^+$. That is
$$
|u_r(x)| \le C(\mu)M\,|x|, \,\,x\in B_{1/r}^+.
$$
\end{remark}
\begin{lemma}[Uniform bounds in $H^1(B_1^+)$ norm]\label{global H1}\label{L2 bounds}
Let $u\in \mathcal{P}_1$. Then for $R>0$ such that $2R\le \frac{2}{r}$, we have
$$
\int_{B_{R}^+}|\nabla u_r|^2\,dx\le C (N, \lambda, \mu, R, M).
$$
\end{lemma}
\begin{proof}
Since $u_r\in \mathcal{P}_r$, from \ref{P4} we can say that $u_r$ is a minimizer of $J(\cdot; A^r, \lambda_{\pm}, B_{2R}^+)$ with boundary data $\phi_r$. Here $A^r$ and $\phi_r$ satisfy the conditions \ref{P1} and \ref{P2}. Precisely speaking, $u_r$ is minimizer of the following functional
$$
J(v; A^r, \lambda_{\pm}, B_{2R}^+) : =\int_{B_{2R}^+} \Big ( \< A^r(x) \nabla v,\nabla v \> + \Lambda (v) \Big )\,dx
$$
here $\big ( \Lambda(v) = \lambda_{+} \chi_{\{v>0\}} + \lambda_{-} \chi_{\{v\le0\}} \big )$. Consider $h\in H^1(B_{2R}^+))$ be a harmonic replacement
$$
\begin{cases}
\dive(A^r(x)\nabla h) = 0\qquad \mbox{in $B_{2R}^+$}\\
h-u_r\in H_0^1(B_{2R}^+).
\end{cases}
$$
in other words, $h$ is the minimizer of $\int_{B_{2R}^+}\< A^r(x)\nabla h,\nabla h \> \,dx $, in the set $H_{u_r}^1(B_{2R}^+)$.
From minimality of $u_r$ and the choice of $h$, we have
\[
\begin{split}
\int_{B_{2R}^+}\langle A^r(x) \nabla (u_r-h),\nabla &(u_r-h) \rangle\,dx = \int_{B_{2R}^+}\langle A^r(x) \nabla (u_r-h),\nabla (u_r+h-2h) \rangle\,dx\\
&= \int_{B_{2R}^+}\langle A^r(x) \nabla (u_r-h),\nabla (u_r+h) \rangle\,dx - 2\int_{B_{2R}^+}\langle A^r(x) \nabla (u_r-h),\nabla h \rangle\,dx\\
&= \int_{B_{2R}^+} \Big ( \langle A^r(x) \nabla u_r,\nabla u_r \rangle- \langle A^r(x) \nabla h,\nabla h \rangle \Big )\,dx\;\;\; \mbox{(since $h$ is $A^r$-harmonic in $B_R^+$)}\\
& \le \int_{B_{2R}^+} \Big ( \Lambda(h)-\Lambda(u_r) \Big )\,dx \le C(N,\lambda,R).
\end{split}
\]
We use ellipticity of $A$, we get
$$
\int_{B_{ R}^+}|\nabla (u_r-h)|^2\,dx \le \int_{B_{2R}^+}|\nabla (u_r-h)|^2\,dx\le C(N, \lambda, \mu, R)
$$
expanding the left hand side, we get
\[
\begin{split}
\int_{B_{R}^+}|\nabla u_r|^2\,dx\le \int_{B_{R}^+}|\nabla u_r|^2\,dx+\int_{B_{R}^+}|\nabla h|^2\,dx&\le C(N, \lambda, \mu, R)+2\int_{B_{R}^+}\nabla u_r\cdot \nabla h\,dx\\
&\le C(N, \lambda, \mu, R) + {\varepsilon} \int_{B_{R}^+}|\nabla u_r|^2\,dx +\frac{1}{{\varepsilon}} \int_{B_{R}^+}|\nabla h|^2\,dx
\end{split}
\]
by choosing ${\varepsilon}=\frac{1}{8}$ we are left with the following,
$$
\int_{B_{R}^+}|\nabla u_r|^2\,dx\le C(N, \lambda, \mu, R) \Big ( 1 + \int_{B_{R}^+}|\nabla h|^2\,dx \Big ).
$$
From \cite[Theorem 2]{B98}, $\Linfty{\nabla h}{B_{R}^+}\le C(\mu,M)$. Thus we obtain a uniform bound on $\int_{B_{R}^+}|\nabla h|^2\,dx$.
\end{proof}
\begin{lemma}[Compactness]\label{convergence}
Let $r_j\to 0^+$, and a sequence $\{v_j\} \in \mathcal{P}_1$. Then the blow-ups $u_j : = (v_j)_{r_j}$ $($as defined in \eqref{blowup}$)$ converges $($up to subsequece$)$ uniformly in $B_R^+$ and weakly in $H^1(B_R^+)$ to some limit for any $R>0$. Moreover, if $u_0$ is such a limit of $u_j$ in the above mentioned topologies, then $u_0$ belongs to $\mathcal{P}_{\infty}$.
\end{lemma}
\begin{proof}
We fix $R>0$, since $v_j\in \mathcal{P}_1$, therefore $u_j\in \mathcal{P}_{r_j}$ and as argued in the proof of previous Lemma, the functions $u_j$ are minimizers of the functional $J (\cdot; A_j, \lambda_{\pm}, B_{R}^+)$ for $j$ sufficiently large that $R<\frac{1}{r_j}$. We set the notation for the functional $J_j$ as
\begin{equation}\label{Jj}
J_j(v):=J (v; A_j, \lambda_{\pm}, B_{R}^+) = \int_{B_{R}^+} \Big ( \langle A_{j}(x) \nabla v,\nabla v\rangle+\Lambda(v) \Big )\,dx,\qquad v\in H_{u_j}^1(B_{R}^+).
\end{equation}
We also denote the boundary values for $u_j \in \mathcal{P}_{r_j}$ as $\phi_j$. Here the sequences $A_j\in {C^{\alpha}(B_{2/r_j}^+)}^{N\times N}$ and $\phi_j\in C^{1,\alpha}(B_{2/r_j})$ satisfy the condition \ref{P1}, \ref{P2} with $r=r_j$, $\Lambda(v) := \lambda_{+} \chi_{\{v>0\}} + \lambda_{-} \chi_{\{v\le0\}} $. We set the following notation for the functional $J_0$
\begin{equation}\label{J0}
J_0(v;B_R^+):=\int_{B_{R}^+}|\nabla v|^2+\Lambda(v)\,dx.
\end{equation}
From Lemma \ref{lipschitz}, we know that $u_j\in C^{\alpha_0}(\overline{B_{2/r_j}^+)}$ which implies $C^{\alpha_0}(\overline{B_{R}^+)}$. In particular $\|u_j\|_{C^{0,\alpha_0}(B_R^+)} \le C(\mu,\lambda_{\pm})$. Hence, $u_j$ is a uniformly bounded and equicontinuous sequence in $\overline{B_R^+}$, we can apply Arzela Ascoli theorem to show that $u_j $ uniformly converges to a function $u_0 \in C^{0,\alpha_0}(\overline{B_R^+})$.
Since $u_j=\phi_j$ on $B_{R}'$, from \ref{P1} we have $|\phi_j(x)| \le M r_j^{1+\alpha} |x|^{1+\alpha}$ for $x\in B_{R}'$, therefore \newline $|\phi_j(x)| \le C(M,\alpha) r_j^{1+\alpha} R^{1+\alpha}$. Hence $\phi_j\to 0$ uniformly on $B_R'$. We have
$$ u_0 =\lim_{j\to \infty} u_j = \lim_{j\to \infty} \phi_j =0 \mbox{ on $B_R'$.}$$
Thus $u_0$ satisfies \ref{G2} and \ref{G3} inside the domain $\overline{B_R^+}$. Also, from Lemma \ref{L2 bounds} we have
\begin{equation}\label{bound}
\int_{B_R^+} |\nabla u_j|^2\,dx \le C (N, \lambda_{\pm}, \mu, R, M, \alpha)
\end{equation}
then, by the linear growth condition (c.f. Remark \ref{linear growth for blowup}), $u_j$ also satisfies
\begin{equation}
|u_j(x)| \le C(\mu,\alpha)M|x|\;\;x\in B_R^+.
\end{equation}
Hence, passing to the limit, we have $|u_0(x)| \le C(\mu,\alpha)M|x|,\;\forall x\in B_R^+$. In other words $u_0$, satisfies \ref{G1} in $\overline{B_R^+}$. Moreover, we have
\begin{equation}\label{bound L2}
\int_{B_R^+}|u_j|^2\,dx\le C(\mu,\alpha,M) \int_{B_R^+} |x|^2\,dx \le C(\mu, \alpha, M, N, R).
\end{equation}
Thus \eqref{bound} and \eqref{bound L2} imply that $u_j$ is a bounded sequence in $H^1(B_R^+)$. Hence, up to a subsequence, $u_j \rightharpoonup u_0$ weakly in $H^1(B_R^+)$. We rename the subsequence again as $u_j$.
We have found a blow-up limit up to a subsequence $u_0$ and have shown that $u_0$ satisfies \ref{G1}, \ref{G2} and \ref{G3} in $\overline{B_R^+}$. In order to show that $u_0\in \mathcal{P}_{\infty}$, it only remains to verify that $u_0$ satisfies \ref{G4}, i.e. $u_0$ is a local minimizer of $J_{0}(\cdot; B_R^+)$ for all $R>0$ (c.f. \eqref{J0}). For that, we first claim that
\begin{equation}\label{gamma convergence}
\int_{B_R^+} \Big ( |\nabla u_0|^2+\Lambda(u_0) \Big )\,dx \le \liminf_{j\to \infty} \int_{B_R^+} \Big ( \langle A_{j}(x) \nabla u_j,\nabla u_j\rangle+\Lambda(u_j) \Big )\,dx.
\end{equation}
Indeed, let us look separately at the term $J_j(u_j)$ on the right hand side of the above equation
$$
J_{j}(u_j)= \int_{B_R^+} \Big ( \< A_j(x) \nabla u_j,\nabla u_j \> + \lambda_+ \chi_{\{u_j>0\}} + \lambda_- \chi_{\{u_j\le 0\}}\Big ) \,dx.
$$
We rewrite the first term as follows
\begin{equation}\label{difference}
\int_{B_R^+} \< A_j(x) \nabla u_j,\nabla u_j \> \,dx = \int_{B_R^+} \< (A_j(x) -Id )\nabla u_j,\nabla u_j \> \,dx+\int_{B_R^+}|\nabla u_j|^2\,dx.
\end{equation}
From \ref{P1} and \ref{P2} we have for all $x\in B_R^+$
$$
\|A_j(x) - Id\|_{L^{\infty}(B_R^+)} \le M r_j^{\alpha} |x|^{\alpha}\le C(M,R,\alpha)r_j^{\alpha} \to 0\mbox{ as $j\to \infty$}.
$$
Therefore $A_j\to Id$ uniformly and $\|\nabla u_j\|_{L^2(B_2^+)}$ is bounded (c.f. \eqref{bound}). Hence, the first term on the right hand side of \eqref{difference} tends to zero as $j\to \infty$. Thus, from \eqref{difference} and by weak lower semi-continuity of $H^1$ norm, we have
\begin{equation}\label{eq10}
\int_{B_R^+}|\nabla u_0|^2\,dx
\le \liminf_{j\to \infty} \int_{B_R^+} |\nabla u_j|^2\,dx = \liminf_{j\to \infty} \int_{B_R^+} \< A_j(x) \nabla u_j,\nabla u_j \> \,dx .
\end{equation}
For the second term, we claim that
\begin{equation}\label{last term}
\begin{split}
\int_{B_R^+}\lambda_+\chi_{\{u_0>0\}}+ \lambda_-\chi_{\{u_0\le0\}} \,dx &\le \liminf_{j\to \infty} \int_{B_R^+} \lambda_+\chi_{\{u_j>0\}}+\lambda_-\chi_{\{u_j\le0\}}\,dx.\\
\end{split}
\end{equation}
To see this, we first show that for almost every $x\in B_R^+$, we have
\begin{equation}\label{char}
\lambda_{+}\chi_{\{u_0>0\}}(x)+ \lambda_- \chi_{\{u_0\le0\}}(x) \le \liminf_{j\to \infty} \left ( \lambda_+ \chi_{\{u_j>0\}}(x)+\lambda_-\chi_{\{u_j\le0\}}(x)\right ).
\end{equation}
Indeed, let $x_0\in B_R^+\cap \big ( \{u_0>0\}\cup \{u_0<0\} \big )$. Then by the uniform convergence of $u_j$ to $ u_0$, we can easily see that $u_j(x_0)$ attains the sign of $u_0(x_0)$ for sufficiently large value of $j$. Hence, \eqref{char} holds in $\{u_0>0 \} \cup \{u_0<0\}$.
Now, assume $x_0\in B_R^+\cap \{u_0=0\}$. Then left hand side of \eqref{char} is equal to
$$
\lambda_{+}\chi_{\{u_0>0\}}(x_0)+ \lambda_- \chi_{\{u_0\le0\}}(x_0) = \lambda_{-}.
$$
Regarding RHS of \eqref{char}, we see that
$$
\lambda_+ \chi_{\{u_j>0\}}(x_0)+\lambda_-\chi_{\{u_j\le0\}}(x_0) =
\begin{cases}
\lambda_+, \qquad \mbox{if $u_j(x_0)>0$}\\
\lambda_-, \qquad \mbox{if $u_j(x_0)\le 0$}.
\end{cases}
$$
Since $\lambda_- <\lambda_+$ (c.f. \ref{P3}), the right hand side in the equation above is always greater than or equal to $ \lambda_- $. Then
\[
\begin{split}
\lambda_{+}\chi_{\{u_0>0\}}(x_0)+ \lambda_- \chi_{\{u_0\le0\}}(x_0) = \lambda_{-} \le \liminf_{j\to \infty} \left ( \lambda_+ \chi_{\{u_j>0\}}(x_0)+\lambda_-\chi_{\{u_j\le0\}}(x_0) \right ).
\end{split}
\]
Thus, \eqref{char} is proven for all $x\in B_R^+$ and hence \eqref{last term} holds by Fotou's lemma.
By adding \eqref{eq10} and \eqref{last term} and \cite[Theorem 3.127]{D17} we obtain \eqref{gamma convergence}. Now we will use \eqref{gamma convergence} prove the minimality of $u_0$ for the functional $J_0(\cdot;B_R^+)$ (c.f. \ref{J0}).
Pick any $w\in H^1(B_R^+)$ such that, $u_0-w\in H_0^1(B_R^+)$. We construct an admissible competitor $w_j^{\delta}$ to compare the minimality of $u_j$ for the functional $J_j(\cdot;B_R^+)$. Then we intend to use \eqref{gamma convergence}.
In this direction, we define two cutoff functions $\eta_{\delta}: \mathbb{R}^N\to \mathbb{R}$ and $\theta :\mathbb{R} \to \mathbb{R}$ as follows,
$$
\eta_{\delta}(x):=
\begin{cases}
1,\; x\in B_{R-\delta}\\
0, \;x \in \mathbb{R}^N\setminus B_{R}
\end{cases},
\theta(t):=
\begin{cases}
1,\;|t|\le 1/2\\
0,\; |t| \ge 1.
\end{cases}
$$
we can take $|\nabla \eta_{\delta}|\le \frac{C(N)}{\delta}$. We define $\theta_j(x)= \theta(\frac{x_N}{d_j})$, for a sequence $d_j\to 0$, which we be suitably chosen in later steps of the proof. Let $w_j^{\delta}$ be a test function defined as
\begin{equation}\label{test function}
w_j^{\delta} := w + (1-\eta_{\delta}) (u_j-u_0) +\eta_{\delta} \theta_{j} \phi_j.
\end{equation}
Since, the function $w_j^{\delta} -w = (1-\eta_{\delta}) (u_j-u_0)+\eta_{\delta} \theta_{j} \phi_j$ is continuous in $\overline{B_R^+}$ and is pointwise equal to zero on $\partial B_R^+$, which is a Lipschitz surface in $\mathbb{R}^N$. Therefore, $u_j-w_j^{\delta} \in H_0^1(B_R^+)$. For further steps, the reader can refer to the Figure 1.
\begin{figure}[h]\label{fig1}
\centering
\includegraphics[width=0.7\textwidth]{figure2}
\caption{(curvy line represents the free boundary of $w$)}
\end{figure}
Let $\Omega_{\delta,j}= B_R^+ \cap \{ \theta_j=0 \} \cap \{\eta_{\delta}=1\}$, and $\mathcal{R}_{\delta,j} = B_R^+ \setminus \Omega_{\delta,j}$ by observing $w_j^{\delta} = w$ on $\Omega_{\delta,j}$ we see that
\[
\begin{split}
|\{w_j^{\delta}>0 \} \cap B_R^+| &= |\{w_j^{\delta}>0 \}\cap \Omega_{\delta,j} |+ |\{w_j^{\delta}>0 \}\cap \mathcal{R}_{\delta,j}|\\
&= |\{w>0 \}\cap \Omega_{\delta,j} |+ |\{w_j^{\delta}>0 \}\cap \mathcal{R}_{\delta,j}|\\
&= |\{w>0 \}\cap (B_R^+\setminus \mathcal{R}_{\delta,j} )|+ |\{w_j^{\delta}>0 \}\cap \mathcal{R}_{\delta,j}|\\
&= |\{w>0 \}\cap B_R^+| - |\{w>0 \} \cap \mathcal{R}_{\delta,j}| + |\{w_j^{\delta}>0 \}\cap \mathcal{R}_{\delta,j}|.
\end{split}
\]
From the above discussions, we have
\[
\begin{split}
|\{w>0\}\cap B_R^+| - |\mathcal{R}_{\delta,j}| \le |\{w_j^{\delta}>0\}\cap B_R^+| \le |\{w>0\}\cap B_R^+| + |\mathcal{R}_{\delta,j}|.
\end{split}
\]
Since we know that $\lim_{\delta\to 0}\left ( \lim_{j\to 0} |\mathcal{R}_{\delta,j}| \right )=0$, we have
\begin{equation}\label{eq15}
\lim_{\delta \to 0}\left ( \lim_{j\to \infty} |\{ w_{j}^{\delta}>0 \} \cap B_R^+| \right )= |\{w>0\} \cap B_R^+|
\end{equation}
and similarly
\begin{equation}\label{eq16}
\lim_{\delta \to 0}\left ( \lim_{j\to \infty} |\{ w_{j}^{\delta}\le 0 \} \cap B_R^+| \right )= |\{w\le 0\} \cap B_R^+|.
\end{equation}
Given $u_j\in \mathcal{P}_{r_j}$ and $w_j^{\delta}-u_j \in H_0^1(B_R^+)$, from the minimility of $u_j$ for the functional $J_j$ we have
$$
\int_{B_R^+} \Big ( \< A_j(x) \nabla u_j, \nabla u_j \> +\lambda_+ \chi_{\{u_j>0\}} + \lambda_- \chi_{\{u_j\le 0\}} \Big ) \,dx \le \int_{B_R^+} \Big ( \< A_j(x) \nabla w_j^{\delta}, \nabla w_j^{\delta} \> +\lambda_+ \chi_{\{w_j^{\delta}>0\}} + \lambda_- \chi_{\{w_j^{\delta}\le 0\}} \Big ) \,dx
$$
and from \eqref{gamma convergence} we obtain
\begin{equation}\label{eq12}
\begin{split}
\int_{B_R^+} \Big ( |\nabla u_0|^2 + \lambda_+ \chi_{\{u_0>0\}}+\lambda_- \chi_{\{u_0\le 0\}}\Big ) \,dx & \le \liminf_{j\to \infty} \int_{B_R^+}\Big ( \< A_j(x) \nabla w_j^{\delta}, \nabla w_j^{\delta} \> +\lambda_+ \chi_{\{w_j^{\delta}>0\}} + \lambda_- \chi_{\{w_j^{\delta}\le 0\}}\Big ) \,dx\\
&\le \limsup_{j\to \infty} \int_{B_R^+}\Big ( \< A_j(x) \nabla w_j^{\delta}, \nabla w_j^{\delta} \> +\lambda_+ \chi_{\{w_j^{\delta}>0\}} + \lambda_- \chi_{\{w_j^{\delta}\le 0\}}\Big ) \,dx
\end{split}
\end{equation}
from the same reasoning as for the justification of \eqref{eq10}, we have
\begin{equation}\label{eq13.0}
\limsup _{j\to \infty} \int_{B_R^+} \< A_j(x) \nabla w_j^{\delta}, \nabla w_j^{\delta} \> \,dx = \limsup_{j\to \infty} \int_{B_R^+}|\nabla w_j^{\delta}|^2\,dx
\end{equation}
therefore rewriting \eqref{eq12}
\begin{equation}\label{eq13}
\int_{B_R^+} \Big ( |\nabla u_0|^2 + \lambda_+ \chi_{\{u_0>0\}}+\lambda_- \chi_{\{u_0\le 0\}} \Big ) \,dx \le \limsup_{j\to \infty} \int_{B_R^+} \Big ( |\nabla w_j^{\delta}|^2\,dx+\lambda_+ \chi_{\{w_j^{\delta}>0\}} + \lambda_- \chi_{\{w_j^{\delta}\le 0\}} \Big ) \,dx.
\end{equation}
We claim that
\begin{equation}\label{eq14}
\lim_{\delta \to 0}\left ( \limsup_{j\to \infty} \int_{B_R^+} | \nabla w_j^{\delta}|^2 \,dx \right )= \int_{B_R^+} |\nabla w|^2\,dx
\end{equation}
To obtain the claim above, we prove that
$$
\lim_{\delta \to 0}\left ( \limsup_{j\to \infty} \int_{B_R^+} |\nabla (w_j^{\delta}-w)|^2\,dx \right ) = 0.
$$
From the definition of $w_j^{\delta}$, we know that
$$
w^{\delta}_j-w = (1-\eta_{\delta}) (u_j-u_0) +\eta_{\delta} \theta_{j} \phi_j
$$
therefore we have
\begin{equation}\label{all terms}
\begin{split}
\int_{B_R^+} |\nabla (w_j^{\delta}-w)|^2\,dx & \le C \Bigg ( \int_{B_R^+} |\nabla \big ( (1-\eta_{\delta})(u_j-u_0) \big )|^2 \,dx + \int_{B_R^+} |\nabla (\theta_j \eta_{\delta} \phi_j)|^2\,dx \Bigg )\\
&\le C(N) \Bigg ( \int_{B_R^+} (1-\eta_{\delta})^2 |\nabla (u_j-u_0)|^2\,dx+\frac{1}{\delta^2}\int_{B_R^+}|u_j-u_0|^2\,dx\\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad+ \int_{B_R^+} |\nabla (\theta_j \eta_{\delta} \phi_j)|^2\,dx \Bigg ).
\end{split}
\end{equation}
Let us consider the first term on the right hand side. We know that $\int_{B_R^+} |\nabla (u_j-u_0)|^2\,dx$ is uniformly bounded in $j\in \mathbb{N}$ (c.f. \eqref{bound}). Therefore
\begin{equation}\label{t1}
\begin{split}
\lim_{\delta\to 0 } \Big ( \limsup_{j\to \infty} \int_{B_R^+} (1-\eta_{\delta})^2 |\nabla (u_j-u_0)|^2\,dx \Big ) &= \Big (\lim_{\delta\to 0}\Linfty{1-\eta_{\delta}}{B_R^+}^2 \Big ) \Big ( \limsup_{j\to \infty} \int_{B_R^+} |\nabla (u_j-u_0)|^2\,dx \Big )\\
&\le C(N,\lambda,\mu,R,M,\alpha)\lim_{\delta\to 0}\Linfty{1-\eta_{\delta}}{B_R^+} =0
\end{split}
\end{equation}
Regarding the second term, since $|u_j-u_0|$ tends to zero in $L^2(B_R^+)$ as $j\to \infty$, therefore the second term also tends to zero as $j\to \infty$. We write
\begin{equation}\label{t2}
\lim_{\delta\to 0 } \Big ( \lim_{j\to \infty}\frac{1}{\delta^2}\int_{B_R^+}|u_j-u_0|^2\,dx \Big ) =0.
\end{equation}
Lastly, we claim that
\begin{equation}\label{t3}
\lim_{j\to \infty}\int_{B_R^+} |\nabla (\theta_j \eta_{\delta} \phi_j)|^2\,dx =0.
\end{equation}
Indeed, we have
\begin{equation}\label{final part}
\int_{B_R^+} |\nabla (\theta_j \eta_{\delta} \phi_j)|^2\,dx\le C \left (\int_{B_R^+} |\nabla \eta_{\delta}|^2 (\theta_j \phi_j)^2\,dx +\int_{B_R^+}|\nabla \phi_j|^2 (\eta_{\delta} \theta_j)^2 \,dx+\int_{B_R^+} |\nabla \theta_j |^2(\eta_{\delta} \phi_j)^2\,dx \right ).
\end{equation}
Since $\eta_{\delta}, \theta_j\le 1$, $|\nabla \eta_{\delta}| \le \frac{C(N)}{\delta}$ and $\Linfty{\nabla \phi_j}{B_R^+}\le M$ (c.f. \ref{P1}), we obtain
\begin{equation}\label{dterm}
\begin{split}
\int_{B_R^+} |\nabla \eta_{\delta}|^2 (\theta_j \phi_j)^2\,dx +\int_{B_R^+}|\nabla \phi_j|^2 (\eta_{\delta} \theta_j)^2 \,dx &\le C(N) |\{ \theta_j \neq 0 \}\cap B_R^+| \Big ( \frac{1}{\delta^2} + M^2 \Big )\\
\end{split}
\end{equation}
We know that $|\{ \theta_j \neq 0\}\cap B_R^+|\to 0$ as $j\to \infty$, hence from \eqref{dterm}, the first and second term in \eqref{final part} tend to zero as $j\to \infty$. The last term in \eqref{all terms} also tends to zero as $j\to \infty$, indeed from \ref{P1} we have $[\nabla \phi_j]_{C^{\alpha}(B_R^+)} \le r_j^\alpha M$. Since $\phi_j(0) = 0$, therefore we have $|\phi_j|\le R^{\alpha} [\nabla \phi_j]_{B_R^+}\le R^{\alpha} r_j^{\alpha} M$ in $B_R^+$. Also, observing that $|\nabla \theta_j| \le \frac{1}{d_j}$, $\eta_{\delta}\le 1$ in $B_{R}^+$ we have
$$
\int_{B_R^+} |\nabla \theta_j |^2 (\eta_{\delta} \phi_j)^2\,dx\le M R^{2\alpha}\frac{r_j^{2 \alpha}}{d_j^2} |B_R^+|
$$
if we choose a sequence $d_j\to 0^+$ such that we also have $\frac{r_j^{\alpha}}{d_j} \to 0$, the third term in \eqref{final part} tends to zero as $j\to \infty$. Plugging in the estimates above \eqref{t1}, \eqref{t2}, \eqref{t3} in \eqref{all terms}, we obtain the claim (\ref{eq14}).
From the equations (\ref{eq15}), (\ref{eq16}), and (\ref{eq14}) we obtain that the right hand side of \eqref{eq13} is equal to $J_0(w;B_R^+)$, therefore $u_0$ is a minimizer of $J_0(\cdot;B_R^+)$. That is
$$
J_0(u_0;B_R^+) \le J_0(w;B_R^+)\;
$$
for every $w \in H^1(B_R^+)$ such that $u_0-w\in H_0^1(B_R^+)$. Since the inequality above (which corresponds to \ref{G4}) and other verified properties of $u_0$ (i.e. \ref{G1}, \ref{G2} and \ref{G3} in $B_R^+$) hold for every $R>0$, therefore $u_0 \in H_{loc}^1(\mathbb{R}^N_+)$ satisfies all the properties in the Definition \ref{global solution}. Hence $u_0\in \mathcal{P}_{\infty}$.
\end{proof}
After proving that the (subsequential) limits of blow-up are global solutions, we proceed to show that the positivity sets (and hence the free boundaries) of blow-ups converge in certain sense to that of blow-up limit. For this we will need to establish that the minimizers $u\in \mathcal{P}_r$ are non-degenerate near the free boundary. In the proof below, we adapt the ideas from \cite{ACF84}.
\begin{comment}
\begin{theorem}\label{lower bound}
Let $u\in \mathcal{P}_1$, then, for every compact subset $K\subset B_{2}^+$, there exists a constant $C(K)$ such that for every $x\in K\cap \{u>0\}$.
$$
u(x)\ge C(K)\dist(x, F(u)\cap B_2^+).
$$
\end{theorem}
\begin{proof}
We fix $x_0\in K \cap B_{2}^+$ and define $d=\dist(x_0,\partial F(u) \cap B_{2}^+) $. Let $w:B_1(0)\to \mathbb{R}$ be a rescaled function defined as
$$
w(y)=\frac{1}{d}u(x_0+dy)
$$
note that $w(0)=\frac{u(x_0)}{d}$, and we will denote $\alpha = w(0)$. Our aim in the proof is to show that $\alpha$ is bounded from below.
Observe that,
$$
-\dive(A_d(x)\nabla w)=0\; \; \mbox{in $B_1(0)$}.
$$
Since for all $x\in B_1(0)$, $A_d(x) = A(dx)$ is an elliptic matrix, we have by Harnack inequality
$$
\frac{1}{c}\alpha\le w\le c\alpha \;\; \mbox{in $B_{1/2}(0)$}
$$
(here, the constant $c$ does not depend on $d$).
We make use of minimality of $w$ in order to prove our claim, for this we construct a test function $v$ as follows,
$$
v(x)=\begin{cases}
\min (w,c \alpha \psi)\;\;\mbox{in $B_{1/2}(0)$}\\
w \;\;\;\;\;\;\;\;\;\;\;\; \;\;\; \;\;\; \mbox{in $B_1(0)\setminus B_{1/2}(0)$}
\end{cases}
$$
here, $\psi \in C^{\infty}(B_1(0))$ is a cut-off function defined as,
$$
\psi=\begin{cases}
0\;\;\;\mbox{in $B_{1/4}(0)$}\\
1\;\;\;\mbox{in $B_1\setminus B_{1/2}(0)$}.
\end{cases}
$$
Since $w>0$ in $B_1(0)$, therefore $|\{w>0\}|=|B_1(0)|$ Now, we observe that
\[
\begin{split}
\int_{B_1(0)} \< A_d(x) \nabla w, \nabla w \>+\Lambda (w)\,dx&\le \int_{B_1(0)}\< A_d(x) \nabla v, \nabla v \>+\Lambda (v)\,dx\\
\int_{B_1(0)}\< A_d(x) \nabla w, \nabla w \>\,dx+\lambda_{+}|B_1(0)|&\le \int_{B_1(0)}\< A_d(x) \nabla v, \nabla v \>+\lambda_{+} \chi_{\{v>0\}}+\lambda_-\chi_{\{v\le 0\}}\,dx\\
\end{split}
\]
rearranging the terms we get,
\[
\begin{split}
(\lambda_+-\lambda_-)|B_{1/4}(0)|&=\lambda_{+}|B_1(0)|-\lambda_+|\{v>0\}|-\lambda_-|\{v\le 0\}| \\
&\le \int_{B_1(0)}\< A_d(x)\nabla v,\nabla v \> - \< A_d(x) \nabla w, \nabla w\> \\
&=\int_{B_1(0)\cap \{ w>c\alpha \psi \}} (c\alpha)^2 \< A_d(x) \nabla \psi,\nabla \psi \>- \< A_d(x)\nabla w,\nabla w \>\,dx\\
&\le c^2|B_1(0)|\alpha^2
\end{split}
\]
Therefore, $\alpha\ge C(A) \sqrt{\lambda_+-\lambda_-}$. Here, $C(A)$ depends solely on ellipticity of $A$ and dimension $N$.
\end{proof}
\end{comment}
\begin{proposition}[Non-degeneracy near the free boundary] \label{non degeneracy}
Let $u\in \mathcal{P}_{r_0}$ for some $r_0>0$ and $x_0\in B_{2/r_0}^+$. Then, for every $0<\kappa<1$ there exists a constant $c(\mu,N,\kappa,\lambda_{\pm})>0$ such that for all $B_{r}(x_0)\subset B_{\frac{2}{r_0}}^+$, we have
\begin{equation}
\frac{1}{r} \fint _{\partial B_r(x_0)} u^+\,d\mathcal{H}^{N-1}(x)\,<\,c(\mu,N,\kappa,\lambda_{\pm}) \implies \mbox{$u^+=0$ in $B_{\kappa r}(x_0)$}.
\end{equation}
\end{proposition}
\begin{proof}
We fix $x_0\in \{u>0\}\cap B_2^+$ and $r>0$ such that $B_r(x_0) \subset B_2^+$. We denote \newline $\gamma := \frac{1}{r} \fint _{B_r(x_0)} u^+\,dx$. We know from Lemma \ref{lipschitz} that the set $\{u>0\}$ is open. Also, since $u\in \mathcal{P}_{r_0}$, there exists $A\in C^{\alpha}(B_{2/r_0}^+)^{N\times N}$, $\varphi \in C^{1,\alpha}(B_{2/r_0}^+)$, $\lambda_{\pm}$ satisfying \ref{P1}-\ref{P6}. Therefore $u$ solves the PDE $\dive(A(x)\nabla u) = 0$ in $\{u>0\}\cap B_{2/r_0}^+$. By elliptic regularity theory, $u$ is locally $C_{loc}^{1,\alpha}(\{u>0\}\cap B_{2/r_0}^+)$. Then, for almost every $\epsilon>0$, $B_r\cap \partial \{u>{\varepsilon}\}$ is a $C^{1,\alpha}$ surface. Pick one such small ${\varepsilon}>0$ and we consider the test function $v_{{\varepsilon}}$ given by
$$
\begin{cases}
\dive(A(x)\nabla v_{{\varepsilon}}) =0 \qquad \mbox{in $(B_r(x_0)\setminus B_{\kappa r}(x_0))\cap \{u>{\varepsilon}\}$}\\
v_{{\varepsilon}} = u\qquad \qquad \;\; \qquad \qquad \mbox{in $B_{r}(x_0)\cap \{u\le {\varepsilon} \}$}\\
v_{{\varepsilon}}={\varepsilon} \qquad \;\; \qquad \qquad \qquad\mbox{in $B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}$}\\
v_{{\varepsilon}}=u \qquad \; \; \qquad \qquad \qquad \mbox{on $\partial B_r(x_0)$.}
\end{cases}
$$
The function $v_{{\varepsilon}}$ defined above belongs to $H^1(B_r(x_0))$, thanks to \cite[Theorem 3.44]{DD12} $\big ($\cite[Theorem 3.44]{DD12} is proven for $C^1$ domains, but the proof can also be adapted for Lipschitz domains \cite[Theorem 4.6]{GE15}$\big )$.
\begin{figure}[h]\label{fig1}
\centering
\includegraphics[width=0.7\textwidth]{figure}
\caption{Graph of $v_{{\varepsilon}}$.}
\end{figure}
We intend to show that $ v_{{\varepsilon}}$ is bounded in $H^{1}(B_r(x_0))$. This ensures the existence of limit $\lim_{{\varepsilon}\to 0}v_{{\varepsilon}}$ exists in weak sense in $H^1(B_r(x_0))$ and strong sense in $L^2(B_r(x_0))$. Let $G$ be the Green function for $L(v)=\dive(A(x)\nabla v)$ in the ring $B_r(x_0)\setminus B_{\kappa r}(x_0)$. Then if there is a function $w$ such that
\begin{equation}\label{PDE w}
\begin{cases}
\dive(A(x)\nabla w)=0 \; &\mbox{in $B_r(x_0)\setminus B_{\kappa r}(x_0)$}\\
w=u \; &\mbox{on $\partial B_r(x_0)\cap \{u>{\varepsilon}\}$}\\
w={\varepsilon} \; &\mbox{elsewhere on $\partial (B_r(x_0)\setminus B_{\kappa r}(x_0)).$}
\end{cases}
\end{equation}
We can also write that $w-{\varepsilon} = (u-{\varepsilon})^+$ on $\partial B_r(x_0)$ and $w-{\varepsilon} =0$ on $\partial B_{\kappa r}(x_0)$. Consider any sequence $\{x_k\} \subset B_r(x_0)\setminus B_{\kappa r}(x_0)$ such that $x_k \to \bar x \in \partial B_{\kappa r}(x_0)$.
By Green representation formulae for $w-{\varepsilon}$ in \eqref{PDE w}, we have $w(x_k) \to w(\bar x)$, indeed since
\[
\begin{split}
0=(w-{\varepsilon})(\bar x)&=\lim_{k\to \infty}(w-{\varepsilon})(x_k)\\
&=\lim_{k\to \infty}\int_{\partial (B_r(x_0)\setminus B_{\kappa r}(x_0))}\left(u-{\varepsilon})^+(y) (A(y)\nabla_yG(x_k,y)\right )\cdot \nu_y\,d \sigma(y)\\
&= \lim_{k\to \infty}\int_{\partial B_r(x_0) \cap \{u>{\varepsilon}\}}\left(u-{\varepsilon})^+(y) (A(y)\nabla_yG(x_k,y)\right )\cdot \nu_y\,d \sigma(y) ,\qquad \forall x\in \partial B_{\kappa r}(x_0)
\end{split}
\]
where $\nu_y$ is the unit outer normal vector at a point $y$ on the boundary.
We apply same arguments as above to $\nabla w(\bar x)$ and from \cite[Theorem 3.3 (vi)]{GW82} on $G(x,y)$ and therefore for $\bar x\in \partial B_{\kappa r}(x_0)$
\begin{equation}\label{green estimates}
\begin{split}
\big | \nabla w(\bar x)\big | &\le C(\mu) \lim_{k\to \infty} \int_{\partial B_r(x_0)\cap \{u>{\varepsilon}\}}\Big | \nabla_x \left ( \frac{\partial}{\partial _{\nu_y}}G(x_k,y) \right ) (u-{\varepsilon})^+ \Big |\,dx\\
& \le C(\mu) \lim_{k\to \infty} \int_{\partial B_r(x_0)} \frac{1}{|x_k-y|^N}(u-{\varepsilon})^+\,d\mathcal{H}^{N-1}(y)\\
&\le \frac{C(\mu,N)}{(1-\kappa)^N}\frac{1}{r}\fint _{\partial B_r} (u-{\varepsilon})^+\,d\mathcal{H}^{N-1}(y)\le C(\mu,N,\kappa) \gamma\;\mbox{on $\partial B_{\kappa r}(x_0)$}.
\end{split}
\end{equation}
We can easily check by respective definitions that $w\ge v_{{\varepsilon}}$ on $\partial (B_r(x_0)\setminus B_{\kappa r}(x_0))$, moreover, by maximum principle, since $\dive(A(x)\nabla w) =0$ in $B_r(x_0)\setminus B_{\kappa r}(x_0)$ and $w\ge {\varepsilon}$ on $\partial B_r(x_0)\setminus B_{\kappa r}(x_0)$, we have $w>{\varepsilon}$ in $B_r(x_0)\setminus B_{\kappa r}(x_0)$. In particular $w\ge v_{{\varepsilon}}$ on $\partial D_{{\varepsilon}}$ where
$$D_{{\varepsilon}}:= (B_r(x_0)\setminus B_{\kappa r}(x_0))\cap \{u>{\varepsilon}\}.$$
By comparison principle, we know $w\ge v_{{\varepsilon}}$ in $D_{{\varepsilon}}$ and since $w=v_{{\varepsilon}} = {\varepsilon}$ on $\partial B_{\kappa r}(x_0) \cap \{u>{\varepsilon}\}$, hence from \eqref{green estimates}
\begin{equation}\label{eq 2.6}
|\nabla v_{{\varepsilon}}|\le |\nabla w|\le C(\mu,N,\kappa) \gamma\;\;\mbox{on $\partial B_{\kappa r}(x_0) \cap \{u>{\varepsilon}\}$}.
\end{equation}
Given that $\dive (A(x)\nabla v_{{\varepsilon}})=0$ in $D_{{\varepsilon}}$, we have by divergence theorem and \eqref{eq 2.6}
\[
\begin{split}
\int_{D_{{\varepsilon}}} (A(x)\nabla v_{{\varepsilon}})\cdot \nabla (v_{\varepsilon}-u)\,dx&= \int_{\partial B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}} (u-v_{{\varepsilon}})(A(x)\nabla v_{{\varepsilon}})\cdot \nu(y)\, d\mathcal{H}^{N-1}(y)\\
&\le {C(\mu)} \int_{\partial B_{\kappa r}(x_0) \cap \{ u>{\varepsilon} \}} |u-{\varepsilon}||\nabla v_{{\varepsilon}}|\,d\mathcal{H}^{N-1}(y)\\
&\le {C(\mu,N,\kappa) \gamma}\int_{\partial B_{\kappa r}(x_0) \cap \{ u>{\varepsilon} \}} |u-{\varepsilon}|\,d\mathcal{H}^{N-1}(y)=: M_0(u)
\end{split}
\]
justification of use of divergence theorem in $D_{{\varepsilon}}$ can be found in \cite[equation (3.4)]{ACF84}. From the calculations above, we can write
\begin{align*}
\;\; \;\;& \int_{D_{{\varepsilon}}} (A(x)\nabla v_{{\varepsilon}})\cdot \nabla (v_{\varepsilon}-u)\,dx \le M_0\\
\Rightarrow& \int_{D_{{\varepsilon}}} (A(x)\nabla v_{{\varepsilon}})\cdot \nabla v_{\varepsilon}\,dx \le M_0+\int_{D_{{\varepsilon}}} (A(x)\nabla v_{{\varepsilon}})\cdot \nabla u\,dx\\
\Rightarrow &\mu \int_{D_{{\varepsilon}}}|\nabla v_{{\varepsilon}}|^2\,dx \le M_0 + \frac{1}{\mu} \int_{D_{{\varepsilon}}} |\nabla v_{{\varepsilon}}||\nabla u|\,dx\\
\Rightarrow & \mu \int_{D_{{\varepsilon}}}|\nabla v_{{\varepsilon}}|^2\,dx\le M_0+\frac{{\varepsilon}_0}{2\mu} \int_{D_{{\varepsilon}}} |\nabla v_{{\varepsilon}}|^2\,dx+\frac{1}{2{\varepsilon}_0 \mu}\int_{D_{\varepsilon}} |\nabla u|^2\,dx
\end{align*}
putting very small ${\varepsilon}_0>0$ in the last inequality, we have
$$
\int_{D_{{\varepsilon}}}|\nabla v_{{\varepsilon}}|^2\,dx \le M_0+C(\mu)\int_{D_{\varepsilon}} |\nabla u|^2\,dx=:M_1(u).
$$
Since $v_{{\varepsilon}}={\varepsilon} \mbox{ in $B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}$}$ which implies $\nabla v_{{\varepsilon}}=0$ in $B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}$ and $v_{\varepsilon} = u $ in $B_{r}(x_0) \setminus D_{{\varepsilon}}$,
\begin{equation}\label{b1}
\int_{B_r(x_0)}|\nabla v_{\varepsilon}|^2\,dx = \int_{D_{{\varepsilon}}}|\nabla v_{{\varepsilon}}|^2\,dx+\int_{B_r(x_0)\setminus D_{{\varepsilon}}}|\nabla u|^2\,dx=:M_2(u).
\end{equation}
By the definition of $v_{{\varepsilon}}$, $0< v_{{\varepsilon}}\le u$ on $\partial D_{{\varepsilon}}$ and $\dive(A(x)\nabla v_{{\varepsilon}}) = \dive(A(x)\nabla u) = 0$ in $D_{{\varepsilon}}$, therefore by comparison principle $0< v_{{\varepsilon}} < u $ in $D_{{\varepsilon}}$. In the set $B_{\kappa r}(x_0) \cap \{u>{\varepsilon}\} $ we have $v_{{\varepsilon}} ={\varepsilon}<u$ and $v_{{\varepsilon}} = u$ in $B_r(x_0)\cap \{u\le {\varepsilon}\}$. Overall we have $0< v_{{\varepsilon}} \le u$ in $B_r(x_0)\cap \{u>{\varepsilon}\}$ Therefore
\begin{equation}\label{b2}
\int_{B_r(x_0)}| v_{\varepsilon}|^2\,dx \le \int_{B_{r}(x_0) \cap \{u\le {\varepsilon}\}} |u|^2\,dx + \int_{B_{r}(x_0) \cap \{u> {\varepsilon}\}} |u|^2\,dx = \int_{B_{r}(x_0)} | u|^2\,dx.
\end{equation}
Hence, from \eqref{b1} and \eqref{b2}, $v_{{\varepsilon}}$ is bounded in $H^1(B_r(x_0))$. Therefore, up to a subsequence, there exists a limit $v = \lim_{{\varepsilon} \to 0} v_{{\varepsilon}} $ in weak $H^1$ sense, such that $v$ satisfies the following
\begin{equation}\label{prop of v}
\begin{cases}
\dive(A(x)\nabla v) =0 \qquad \mbox{in $(B_r(x_0)\setminus B_{\kappa r}(x_0))\cap \{u>0\}$}\\
v = u\qquad \qquad \;\; \qquad \qquad \mbox{in $B_{r}(x_0)\cap \{u\le 0 \}$}\\
v =0 \qquad \;\; \qquad \qquad \qquad\mbox{in $B_{\kappa r}(x_0)\cap \{u>0\}$}\\
v=u \qquad \; \; \qquad \qquad \qquad \mbox{on $\partial B_r(x_0)$}.
\end{cases}
\end{equation}
We verify the above properties \eqref{prop of v} of $v$ at the end of this proof.
Let us use the function $v$ as a test function with respect to minimality condition on $u$ in $B_r(x_0)$, we have
\[
\begin{split}
\int_{B_r(x_0)} \Big ( \< A(x)\nabla u,\nabla u\>+ \lambda (u)\Big ) \,dx& \le \int_{B_r(x_0)}\Big ( \< A(x)\nabla v,\nabla v\> + \lambda (v) \Big )\,dx\\
\end{split}
\]
since, $v=u$ in $\{u\le 0\}$ and $\{ v>0 \} \subset \{u>0\}$, the integration in the set $\{u\le 0\}$ gets cancelled from both sides and we are left with the terms mentioned below. \newline Set $D_0:= (B_r(x_0)\setminus B_{\kappa r}(x_0))\cap \{u>0\}$, we have
\[
\begin{split}
\int_{B_r(x_0)\cap \{u> 0\}} \Big ( \< A(x) \nabla u,\nabla u \> - \< A(x) \nabla v,\nabla v \> \Big )\,dx &\le \int_{B_{ r}(x_0)\cap \{u>0\}} (\Lambda(v)-\Lambda(u))\,dx\\
&= \int_{B_{\kappa r(x_0)}\cap \{u>0\}} (\Lambda(v)-\Lambda(u))\,dx\\
&= \lambda_0 |B_{\kappa r}(x_0)\cap \{u>0\}| . \qquad(\lambda_0: = -(\lambda_ + - \lambda_-)). \\
\end{split}
\]
We have second equality above because $\chi_{\{v>0\}}= \chi_{\{u>0\}}$ in $D_0$. Since $v=0$ in $D_0$, we have
\[
\begin{split}
\int_{B_{\kappa r}(x_0)\cap \{u>0\}} \< A(x) \nabla u,\nabla u \>\,dx + \int_{D_0} \Big ( \< A(x) \nabla u,\nabla u \> - \< A(x) \nabla v,\nabla v \> \Big )\,dx\le \lambda_0 |B_{\kappa r}(x_0)\cap \{u>0\}|.
\end{split}
\]
Using the ellipticity of $A$ and shuffling the terms in the above equation, we obtain
\begin{equation}\label{eq2.5}
\begin{split}
\int_{B_{\kappa r}(x_0)\cap \{u>0\}} \Big (\mu |\nabla u|^2-\lambda_0 \Big ) \,dx &\le \int_{D_0} \Big ( \< A(x) \nabla v,\nabla v \>- \< A(x) \nabla u,\nabla u \> \Big ) \,dx\\
&= \int_{D_0} \Big ( \< A(x) \nabla (v-u), \nabla (v+u) \> \Big )\,dx\\
&= \int_{D_0} \Big (\< A(x) \nabla (v-u), \nabla (u-v+2v) \> \Big ) \,dx\\
& \le 2\int_{D_0} \< A(x)\nabla v, \nabla (v-u) \>\,dx \\
& \le \liminf_{{\varepsilon}\to 0} 2\int_{D_0} \< A(x) \nabla v_{{\varepsilon}}, \nabla (v_{{\varepsilon}}-u) \>\,\\
&= \liminf_{{\varepsilon}\to 0} 2 \int_{D_{{\varepsilon}}} \< A(x) \nabla v_{{\varepsilon}}, \nabla (v_{{\varepsilon}}-u) \>\,dx\qquad \mbox{(since $v_{{\varepsilon}}=u$ in $D_{{\varepsilon}}\setminus D_0$)}\\
&=\liminf_{{\varepsilon}\to 0} 2\int_{\partial B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}} (u-{\varepsilon}) ( A(x)\nabla v_{{\varepsilon}})\cdot \nu \,dx\\
&\le \liminf_{{\varepsilon}\to 0} \frac{2}{\mu} \int_{\partial B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}}(u-{\varepsilon})\big | \nu\cdot \nabla v_{{\varepsilon}} \big |\,dx\,:=\,M.
\end{split}
\end{equation}
The second to last equality in above calculation is obtained from integration by parts, its justification can be found in \cite[equation (3.4)]{ACF84}. From \eqref{eq2.5} and \eqref{eq 2.6}, and using the trace inequality in $H^1(B_{\kappa r})$ we have (for some different constant $C(\kappa)$),
\begin{equation}\label{eq2.8}
\begin{split}
M &\le C(\mu,N,\kappa) \gamma \int_{\partial B_{\kappa r(x_0)}} u^+\,d\mathcal{H}^{N-1}(x)\\
&\le C(\mu,N, \kappa) \gamma \int_{B_{\kappa r(x_0)}} \Big ( |\nabla u^+|+\frac{1}{r} u^+ \Big )\,dx \\
& \le C(\mu,N, \kappa) \gamma \Bigg [ |B_{\kappa r(x_0)}\cap \{u>0\}|^{1/2}\left ( \int_{B_{\kappa r(x_0)}}|\nabla u^+|^2\,dx \right )^{1/2}+ \frac{1}{r}\sup_{B_{\kappa r}(x_0)} (u^+)\big | \{B_{\kappa r}(x_0)\cap \{u>0\}\}\big | \Bigg ]\\
& \le C(\mu,N, \kappa) \gamma \Bigg [ \frac{1}{2 \sqrt{-\lambda_0} } \int_{B_{\kappa r}(x_0)\cap \{u>0\}} |\nabla u^+|^2\,dx + 2 \sqrt{-\lambda_0} |B_{\kappa r(x_0)}\cap \{u>0\}| \\
&\qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad + \frac{1}{r} \sup_{B_{\kappa r}(x_0)}(u^+) \int_{B_{\kappa r}(x_0)\cap \{u>0\}}1\,dx \Bigg ] \\
&= \frac{C(\mu,N, \kappa)\gamma }{2 \sqrt{-\lambda_0} }\left ( \int_{B_{\kappa r}(x_0)\cap \{u>0\}} |\nabla u^+|^2 - \lambda_0 \,dx \right )+\frac{C(\mu,N,\kappa)\gamma}{\lambda_0 r}\sup_{B_{\kappa r}(x_0)}(u^+) \int_{B_{\kappa r}(x_0)\cap \{u>0\}}\lambda_0 \,dx
\end{split}
\end{equation}
we have used Hölder's inequality and then Young's inequality above. From Lemma \ref{subharmonic}, $u^+$ is $A-$subharmonic in $B_r(x_0)$. If $G'$ is the Green's function for $L'(v)=\dive(A(x)\nabla v)$ in $B_r(x_0)$, then by comparison principle and Green's representation
$$
u^+(x)\le \int_{\partial B_r(x_0)} u^+(y) \left ( A(y)\nabla_y G'(x,y)\right )\cdot \nu_y \,d\mathcal{H}^{N-1}(y)\;\; \forall x\in B_{\kappa r}(x_0).
$$
Since for all $y\in \partial B_r(x_0)$ and $x\in B_{\kappa r}(x_0)$, we have $\frac{1}{|x-y|^{N-1}}\le \frac{C(\kappa)}{r^{N-1}}$, then using the Green's function estimates c.f. \cite[Theorem 3.3 (v)]{GW82} we get
\begin{equation}\label{eq2.9}
\begin{split}
\sup_{B_{\kappa r}(x_0)}u^+&\le C(\mu) \int_{\partial B_r(x_0)} \frac{u^+(y)}{|x-y|^{N-1}}\,d\mathcal{H}^{N-1}(y)\\
&\le C(\mu,\kappa,N) \fint_{\partial B_r(x_0)} u^+\,d\mathcal{H}^{N-1}(y)= C(\mu, \kappa,N)\gamma r.
\end{split}
\end{equation}
Use \eqref{eq2.5} and \eqref{eq2.9} in \eqref{eq2.8} and we have
\[
\begin{split}
\mu \int_{B_{\kappa r}(x_0)\cap \{u>0\}} \Big ( |\nabla u|^2-\lambda_0 \Big )\,dx & \le
\frac{C(\mu,\kappa,N)\gamma }{2 \sqrt{-\lambda_0} }\int_{B_{\kappa r}(x_0)\cap \{u>0\}}\Big ( |\nabla u^+|^2 -\lambda_0 \Big )\,dx\\
&\qquad+\frac{C(\mu,\kappa,N)\gamma}{\lambda_0 r}\sup_{B_{\kappa r}(x_0)}(u^+) \int_{B_{\kappa r}(x_0)\cap \{u>0\}}\lambda_0 \,dx\\
&\le \frac{C(\mu,\kappa,N) \gamma}{\mu \sqrt{-\lambda_0} } \left ( 1+ \frac{C(\kappa)\gamma}{\sqrt{-\lambda_0}} \right )\mu\int_{B_{\kappa r}(x_0)\cap \{u>0\}} \Big ( |\nabla u|^2-\lambda_0 \Big ) \,dx.
\end{split}
\]
If $\gamma$ is small enough, then
$$
\int_{B_{\kappa r}(x_0)\cap \{u>0\}} \Big ( |\nabla u|^2-\lambda_0 \Big ) \,dx=0
$$
in particular $|\{u>0\}\cap B_{\kappa r(x_0)}|=0$, that is $u^+=0$ almost everywhere in $B_{\kappa r}(x_0)$.
It remains to verify the properties of $v$ in \eqref{prop of v}. Before looking at the proof, we observe that for a given $\varphi \in C_c^{\infty}(D_0)$, then there exists ${\varepsilon}_0>0$ such that $\varphi \in C_c^{\infty}(D_{{\varepsilon}})$ for all ${\varepsilon}<{\varepsilon}_0$. Indeed, since $\supp(\varphi)$ is a compact set, and $\bigcup_{{\varepsilon}>0} D_{{\varepsilon}}$ is a cover of $\supp(\varphi)$, then for a finite set $\{{\varepsilon}_1, ...,{\varepsilon}_n\}$ we have $\supp (\varphi) \subset \bigcup_{i=1}^{n} D_{{\varepsilon}_i} \subset D_{{\varepsilon}_{max}}$ where ${\varepsilon}_{max} = \max ({\varepsilon}_1,...,{\varepsilon}_n)$. Therefore, $\varphi \in C_c^{\infty}(D_{{\varepsilon}})$ for all ${\varepsilon}<{\varepsilon}_{max}$.
Let us first verify that $\dive(A(x)\nabla v)=0 $ in $D_0$. For this let $\varphi \in C_c^{\infty}(D_0)$, then from continuity of $u$, there exists a ${\varepsilon}_0>0$ such that $\supp (\varphi) \subset D_{\varepsilon}$ for all ${\varepsilon}<{\varepsilon}_0$, also we have
\begin{equation}\label{weak form}
\int_{D_0}\langle A \nabla v, \nabla \varphi \rangle \,dx = \int_{\supp (\varphi)} \langle A \nabla v, \nabla \varphi \rangle \,dx
\end{equation}
since $\supp (\varphi) \subset D_{\varepsilon}$, from the definition of $v_{{\varepsilon}}$ we have
$$
\int_{\supp (\varphi)} \langle A \nabla v_{{\varepsilon}}, \nabla \varphi \rangle \,dx=0
$$
and we know that $v$ is a weak limit of $v_{{\varepsilon}}$ in $H^1({B_r(x_0)})$, therefore from \eqref{weak form} we have
$$
\int_{D_0}\langle A \nabla v, \nabla \varphi \rangle \,dx = \int_{\supp (\varphi)} \langle A \nabla v, \nabla \varphi \rangle \,dx= \lim_{{\varepsilon}\to 0} \int_{\supp (\varphi)} \langle \nabla v_{{\varepsilon}}, \nabla \varphi \rangle \,dx=0.
$$
Hence we show that $\dive(A(x)\nabla v)=0 $ in $D_0$. To show that $v=0$ in $B_{\kappa r}(x_0)\cap \{u>0\}$, we now take the function $\varphi \in C_c^{\infty}(B_{\kappa r}(x_0)\cap \{u>0\})$. From the same reasoning as above we know that there exists an ${\varepsilon}_0>0$ such that $\supp(\varphi)\subset B_{\kappa r}(x_0)\cap \{u>{\varepsilon}\}$ for all ${\varepsilon}<{\varepsilon}_0$. From the definition of $v_{{\varepsilon}}$, we have
$$
\int_{\supp (\varphi)} v_{{\varepsilon}} \varphi\,dx=\int_{\{u>{\varepsilon}\} \cap B_{\kappa r}(x_0)} v_{{\varepsilon}} \varphi\,dx = {\varepsilon} \int_{\{u>{\varepsilon}\} \cap B_{\kappa r}(x_0)}\varphi\,dx
$$
and in limit ${\varepsilon}\to 0$, from the above equation we have
$$
\int_{\supp(\varphi)} v\varphi\,dx= \lim_{{\varepsilon}\to 0} \int_{\supp (\varphi)} v_{{\varepsilon}} \varphi\,dx=0
$$
and therefore $v=0$ a.e. in $B_{\kappa r}(x_0)\cap \{u>0\}$. To prove that $v=u$ in $\{u\le 0\}$, we observe that $\{u\le 0\}\subset \{ u\le {\varepsilon}\}$, hence from the definition of $v_{{\varepsilon}}$, have
$$
v_{{\varepsilon}} = u \mbox{ in $\{u\le 0\}$}.
$$
since the weak limits maintain the equality (c.f. \cite[Lemma 3.14]{MS211}) the claim follows in the limit ${\varepsilon} \to 0$. Apart from that, since $v_{{\varepsilon}}=u$ on $\partial B_r(x_0)$ therefore from conservation of traces in weak convergence, it follows that $v=u$ on $\partial B_r(x_0)$.
This completes the proof of Proposition \ref{non degeneracy}.
\end{proof}
\begin{remark}
In the proposition above the constant is local in nature, this means, the value of the constant depends on the choice of compact set $K\subset \subset B_2^+$ where $x_0\in K$.
\end{remark}
\begin{lemma}
Let $u_0$ and $u_k$ be as in Theorem \ref{convergence}. Then, for a subsequence of $u_k$, for any $R>0$ we have
\begin{equation}\label{pw convergence}
\chi_{\{u_k>0\} \cap B_R^+} \to \chi_{\{u_0>0\} \cap B_R^+}\qquad \mbox{ a.e. in $B_R^+$}.
\end{equation}
This in turn implies
\begin{equation}\label{L1 convergence}
\chi_{\{u_k>0\} \cap B_R^+} \to \chi_{\{u_0>0\} \cap B_R^+}\qquad \mbox{ in $L^1(B_R^+)$}.
\end{equation}
\end{lemma}
\begin{proof}
From Lemma \ref{convergence}, we can consider a subsequence of $u_k$ such that $u_k\to u_0$ in $L^{\infty}(B_R^+)$.
Let $x\in B_R^+$. If $x\in \{u_0>0\}\cap B_R^+$ (or $\chi_{\{u_0>0\}\cap B_R^+}(x)=1$), then $u_k(x) > \frac{u(x)}{2}>0$ (or $\chi_{\{u_k>0\}\cap B_R^+}(x)=1$) for $k$ sufficiently large. Thus we conclude that
$$
\mbox{ $\chi_{\{u_k>0\}\cap B_R^+} (x) \to \chi_{\{u_0>0\}\cap B_R^+} (x)$ as $k\to \infty$ for all $x\in \{u_0>0\}\cap B_R^+$}.
$$
If $x\in \{u_0\le 0\}^o \cap B_R^+$ (or $\chi_{\{u_0>0\}\cap B_R^+}(x)=0$), then there exists $\delta>0$ such that $B_{\delta}(x) \subset \{u_0\le 0\} \cap B_R^+$. Thus we have $\frac{1}{\delta} \fint_{\partial B_\delta(x)} u_0^+ \,d\mathcal{H}^{N-1} =0$. Again, by the uniform covergence of $u_k$ to $u_0$ in $B_R^+$ (c.f. Lemma \ref{convergence}) we obtain
\begin{equation}\label{c+}
\frac{1}{\delta} \fint_{\partial B_\delta(x)} u_k^+ \,d\mathcal{H}^{N-1} \le \frac{1}{2}c(\mu,N,\lambda_{\pm})\qquad\mbox{for $k$ sufficiently large.}
\end{equation}
Here $c(\mu,N,\lambda_{\pm})$ is as in Proposition \ref{non degeneracy}. This implies $u_k \le 0$ in $B_{\frac{\delta}{2}}(x)$ (c.f. Proposition \ref{non degeneracy}). In particular, $\chi_{\{u_k(x)\le 0\}}(x) =0$ for $k$ sufficiently large. This way, we obtain
\begin{equation}\label{c-}
\mbox{ $\chi_{\{u_k>0\}\cap B_R^+} (x) \to \chi_{\{u_0> 0\}\cap B_R^+} (x)$ as $k\to \infty$ for all $x\in \{u_0\le 0\}\cap B_R^+$}.
\end{equation}
From the representation theorem \cite[Theorem 7.3]{ACF84}, we know that $|\partial \{u_0>0\}\cap B_R^+| =0$. From \eqref{c+}, \eqref{c-} and the fact that $|\partial \{u_0>0\}\cap B_R^+| =0$, we obtain the claim \eqref{pw convergence}. Since $|\chi_{\{u_k>0\}\cap B_R^+}|\le 1$, the claim \eqref{L1 convergence} follows from Lebesgue's dominated convergence theorem.
\end{proof}
\begin{comment}
\begin{theorem}[Linear growth of minimizer near free bounadry]\label{nondeg}
Suppose $r>0$ and $R<\frac{1}{r}$, let $D\Subset B_R^+$. If $u\in \mathcal{P}_1$ and $u_r$ be the blow-up c.f. (\eqref{blowup}). Then for every $x\in (\{u_r>0\} \cup F(u_r))\cap D$, and for every $\rho>0$ such that $B_{\rho}(x)\Subset B_R^+$. We have
\begin{equation}\label{non degeneracy}
\sup_{B_{\rho}(x)} u_r \,\ge\, C \rho
\end{equation}
where $C$ depends only on $D$, the dimension $N$ and ellipticity of coefficient matrix $A$ in \eqref{P1}.
\end{theorem}
\begin{proof}
Note that $-\dive(A_j(x)\nabla u_j)=0$ in $\{u_j >0\}$. Then \eqref{non degeneracy} follows from Theorem \ref{lower bound} and \cite[Theorem 4.1, Corollary 4.5]{ MW14}.
\end{proof}
\end{comment}
\begin{comment}
\begin{lemma}\label{L1 convergence}
Let $u_0$ and $u_j$ be as in Theorem \ref{convergence}. Then, for any $R>0$ and $K\subset \subset B_R^+$, the sets $\{u_{j}>0\} \cap K$ converge upto subsequence to $\{u_0>0\}\cap K$ in Hausdorff sense for some $u_0\in \mathcal{P}_{\infty}$.
In particular, $\{u_k>0\} \to \{u_0>0\}$ in $L^1(B_R^+)$.
\end{lemma}
\begin{proof}
We begin by showing the Hausdorff convergence claimed in the statement. We fix some $\delta>0$ and $x_0\in K$ be such that $x_0\in \big [ \mathcal{N}_{\delta}(\{u_0>0\}) \big ]^c \cap K$. This means $u_0 \le 0$ in $B_{\delta}(x_0)$. Since $u_k\to u_0$ uniformly in $B_R^+$ (c.f. Lemma \ref{convergence}),
$$
\frac{1}{\delta} \fint_{\partial B_{\delta}(x_0)}u_k^+\,dx \le \frac{C(K, \mu, N,\lambda_{\pm})}{2}\,\,\mbox{for $k\in \mathbb{N}$ large enough (say for $k\ge k_{K,\delta}$)}
$$
therefore by non-degeneracy (c.f. Proposition \ref{non degeneracy}) we have $u_k\le 0$ in $B_{\frac{\delta}{2}}(x_0)$, in particular $x \notin \{u_k>0\} \cap K$ for $k\ge k_{K,\delta}$. Hence we get
\begin{equation}\label{containment 1}
\{u_k>0\} \cap K \subset \mathcal{N}_{\delta} \{u_0>0\}\cap K \;\;\forall\; k\ge k_{K,\delta}.
\end{equation}
Next, we claim that for a subsequence of $u_k$, there exists $k_{K,\delta}$ such that
\begin{equation}\label{containment 2}
\{u_0>0\} \cap K \subset \mathcal{N}_{\frac{\delta}{2}} \{u_k>0\}\cap K \subset \mathcal{N}_{\delta} \{u_k>0\}\cap K \;\mbox{for $k\ge k_{K,\delta}$}.
\end{equation}
Consider the following set $\mathcal{A} \subset K$
$$
\mathcal{A}:= \Big \{y\in K \Big | \mbox { $y\notin \mathcal{N}_{\frac{\delta}{2}}\{u_k>0\}\cap K$ for infinitely many $k\in\mathbb{N}$} \Big \}
$$
If the set $\mathcal{A}$ is empty. Thus for every $x\in K$, there exists $k_x\in \mathbb{N}$ such that $x\in \mathcal{N}_{\frac{\delta}{2}}\{u_k>0\}$ for all $k\ge k_x$. In this case, since $K$ is compact, there exists a finite set $\{x_1,..,x_n\} \subset \mathcal{N}_{\frac{\delta}{2}}\{u_k>0\}\cap K$ for $k\ge k_0({K,\delta}):=\max(k_{x_1},...,k_{x_n})$ and such that $K \subset \bigcup_{i=1}^n B_{\frac{\delta}{2}}(x_i)$. Therefore, for the subsequence $\{u_k\}_{k\ge k_0}$ we have $K\subset \mathcal{N}_{\delta}(\{u_k>0\})$. This way, the conclusion \eqref{containment 2} trivially holds when $\mathcal{A}=\{\phi\}$.
If $\mathcal{A}$ is not empty, we consider a subsequence of $u_k$ such that
$$\mathcal{A}': =\bigcap\limits_{k=1}^{\infty} \Big (\big [ \mathcal{N}_{\delta}(u_k>0)\big ]^c \cap K \Big ) \neq \{\phi\}. $$
For that subsequence, if $y_0\in \mathcal{A}'$ then $y_0\notin \mathcal{N}_{{\delta}}\{u_k>0\}\cap B_R^+$ for all $k\in \mathbb{N}$. Hence, $u_k(y_0)\le 0$ for all $k\in \mathbb{N}$. From Theorem \ref{convergence} we have $u_0(y_0) = \lim_{k\to \infty} u_k(y_0) \le 0$. Hence, $y_0\notin \{u_0>0\}$. Therefore, \eqref{containment 2} also holds when $\mathcal{A}\neq \{\phi\}$. From \eqref{containment 1} and \eqref{containment 2} we conclude that
$$
\{u_k>0\}\cap K \to \{u_0>0\}\cap K\;\mbox{in Hausdorff metric upto a subsequence which depends on $K$}.
$$
In particular, we have
$$
\{u_k>0\} \to \{u_0>0\}\;\mbox{in $L^1(K)$ upto a subsequence which depends on $K$}
$$
or
$$
|(\{u_k>0\} \Delta \{u_0>0\} ) \cap K| \to 0\;\mbox{in $L^1(K)$ upto a subsequence which depends on $K$}.
$$
In order to prove the last part of Lemma \ref{L1 convergence}, we consider a sequence $D_j$ of compact subsets of $B_R^+$ such that $|B_R^+\setminus D_j|\le\frac{1}{2j}$ and we extract a subsequence out of $u_{j}$ (and renaming it again as $u_j$) such that
$$
|(\{u_{j}>0\} \Delta \{u_0>0\})\cap D_j|\le \frac{1}{2j}.
$$
Hence
$$
|(\{u_{j}>0\} \Delta \{u_0>0\}) \cap B_R^+|\le |\{u_{j}>0\} \Delta \{u_0>0\}\cap D_j|+|B_R^+\setminus D_j|\le \frac{1}{j}
$$
and therefore we have,
$$
\lim_{j\to \infty}|(\{u_{j}>0\} \Delta \{u_0>0\}) \cap B_R^+|=0.
$$
This proves that $\{u_k>0\} \cap B_R^+ \to \{u_0>0\} \cap B_R^+$ in $L^1(B_R^+)$ as $k\to\infty$ (up-to a subsequence).
\end{proof}
\end{comment}
\begin{comment}
\begin{proof}
Let $\delta>0$ and $x\in \{u_0>\delta\}\cap B_R^+$, since $u_{j}\to u_0$ in $L^{\infty}(B_R^+)$ (from Theorem \ref{convergence}), $u_{j}(x)>\frac{\delta}{2}>0$ for large value of $j$. Thus $x\in \{u_{j}>0\}$ for $j$ large. We can write that
$$
\{u_0>\delta\} \subset \{u_{j}>0\}\qquad \mbox{for $j\ge j_\delta$}
$$
we take limit $\delta \to 0$ and get
\begin{equation}\label{1.1}
\lim_{j\to \infty}|(\{u_0>0\} \setminus \{u_{j}>0\}) \cap B_R^+|=0
\end{equation}
Now, we claim that for any given $\delta>0$ there exist $j$ large such that for every $D\Subset B_R^+$,
\begin{equation}\label{1.5}
\{u_{j}>0\} \cap S\subset \mathcal{N}_{\delta}(\{u_0>0\})\cap D
\end{equation}
(we define $\mathcal{N}_{\delta}(K)$ as $\delta-$neighbourhood of a set $K$). Before proving the above claim, let us show that the claim (\ref{1.5}) implies \eqref{3}.
Indeed, observe that the sets $(\{u_j>0\}\setminus \{u_0>0\}) \cap D$ and $((\mathcal{N}_{\delta}\{u_0>0\})\setminus \{u_0>0\}) \cap D$ are sets of finite measure. Also for a given $\delta>0$ there exists $j_{D,\delta}$ such that for $j\ge j_{D,\delta}$, \eqref{1.5} holds. We have the following
\[
\begin{split}
|(\{u_j>0\}\setminus \{u_0>0\}) \cap D| &\le |((\mathcal{N}_{\delta}\{u_0>0\})\setminus \{u_0>0\}) \cap D|:= m_{D}(\delta).
\end{split}
\]
Where $m_{D}(\delta)$ tends to zero as $\delta \to 0$. That is, for any given $\delta'>0$ we can choose $\delta>0$ such that $m_D(\delta) < \delta'$. Therefore from the above equation
$$
|(\{u_j>0\}\setminus \{u_0>0\}) \cap D| \le m_D(\delta) < \delta', \qquad j \ge j_{D, \delta}.
$$
Since the choice of $\delta$ depends on $\delta'$ we can replace $j_{D, \delta}$ in the above equation by $j_{D, \delta'}$. And therefore we can say that in the limit $j\to \infty$
\begin{equation}\label{other part}
\lim_{j\to \infty}|(\{u_{j}>0\} \setminus \{u_0>0 \})\cap D |=0.
\end{equation}
Now let us prove the claim (\ref{1.5}). Assuming the contrary that the claim (\ref{1.5}) is not true. Then we have a sequence $x_j\in \{u_{j}>0\}$ such that $\dist(x_j, \{u_0>0\})>\delta$. That is,
\begin{equation}\label{assumption}
B_{\delta/2}(x_j)\Subset \{ u_0 \le 0\}.
\end{equation}
Let $y_j\in B_{\delta/2}(x_j)$ be such that $u_{j}(y_j)=\sup_{B_{\delta/2}(x_j)}u_{j}$. From \eqref{non degeneracy}, we have
\begin{equation}\label{2}
u_{j}(y_j)=\sup_{B_{\delta/2}(x_j)} u_{j}\ge C\frac{\delta}{2}.
\end{equation}
Let $y_j\to y^*$ as $j\to \infty$, from \eqref{2} we have $\lim_{j\to \infty} u_{j}(y_j)=u_0(y^*)\ge C \frac{\delta}{2}>0$. Since $y_j\in B_{\delta/2}(x_j)$ thus from \eqref{assumption} we have $u_0(y_j)\le 0$, and therefore $\lim_{j\to \infty}u_0(y_j)=u_0(y^*)\le 0$. Which is a contradiction.
Thus we have \eqref{1.5} and hence
\begin{equation}\label{3}
\lim_{j\to \infty} |(\{u_0>0\}\setminus \{u_{r_j}>0\})\cap D| = 0,\qquad \forall D\Subset B_R^+
\end{equation}
combining \eqref{1.1} and \eqref{3} we have
$$
\lim_{j\to \infty}|\{u_{r_j}>0\} \Delta \{v>0\}\cap D|=0\qquad \forall D\Subset B_R^+
$$
($\Delta$ refers to the symmetric difference operator).
We can now take a sequence $D_j$ for compact subsets of $B_R^+$ such that $|B_R^+\setminus D_j|\le\frac{1}{2j}$ and we extract a subsequence out of $u_{j}$ (and renaming it again as $u_j$) such that
$$
|(\{u_{j}>0\} \Delta \{u_0>0\})\cap D_j|\le \frac{1}{2j}.
$$
Hence
$$
|(\{u_{j}>0\} \Delta \{u_0>0\}) \cap B_R^+|\le |\{u_{j}>0\} \Delta \{u_0>0\}\cap D_j|+|B_R^+\setminus D_j|\le \frac{1}{j}
$$
and therefore we have,
$$
\lim_{j\to \infty}|(\{u_{j}>0\} \Delta \{u_0>0\}) \cap B_R^+|=0.
$$
\end{proof}
\end{comment}
\begin{comment}
We will see in the last section of the article, how the $L^1$ convergence of positivity sets of blow-ups implies that free boundary touches fixed boundary tangentially at the origin. Readers who have watched the steps of above proofs closely, must have realised that the rate of convergence of blow-ups and also convergence of their positivity sets depends on the minimizer. But in fact, we can show that it is not the case. We can establish uniformity in rate of convergence by showing that any minimizer of $J$ in \eqref{P} is actually uniformly bounded in $H^1(B_1^+)$ norm. This uniform bound on $H^1(B_1^+)$ norm of $u\in\mathcal{P}_1$ will also imply uniformity in rate at which free boundary of $u\in \mathcal{P}_1$ approaches tangentially towards the fixed boundary.
\begin{remark}
The above result also hold true for $u\in \mathcal{P}_r$, and the steps of proof are exactly the same as above.
\end{remark}
\subsection{Blow-up analysis for almost minimizers}
We will prove the parallel results as in the previous section for function $u$ belonging to the family of almost minimizers, that is $u\in \mathcal{Q}_1$.
\begin{lemma}\label{continuity}
Let $u\in \mathcal{Q}_1$, then there exists $0<\alpha_0<1$, such that
$u\in C^{0,\alpha_0}(\overline{B_2^+})$.
\end{lemma}
\begin{proof}
Note that the almost minimizer $u\in \mathcal{Q}_1$ is same as $\omega$-minimizer defined in \cite[Section 7.7]{eg05}. Moreover, since the boundary data $\phi\in C^{1,\beta_0}(\overline{B_2^+})$, we can apply the discussion in \cite[Section 7.8]{eg05} on $u\in \mathcal{Q}_1$. This proves Lemma \ref{continuity}.
\end{proof}
\begin{remark}[local Lipschitz continuity]
Moreover, since $\Lambda(0)=0$, apart from being Hölder continuous till the boundary, the function $u\in \mathcal{Q}_1$ is locally Lipschitz in $B_2^+$. Precisely speaking, given $D\Subset B_2^+$, there exists $C>0$ depending on $D$ such that
$$
\|\nabla u\|_{L^{\infty}(D)}\le C(D)
$$
the above claim is proven in \cite[Theorem 8.1]{DEGT19}.
\end{remark}
\begin{theorem}\label{convergence 2}
Suppose $u\in \mathcal{Q}_1$. Let $r_j \to 0$ and $R>0$ such that $R<1/2r_j$. We define the blow-up sequence $u_j$ as $u_j=u_{r_j}$ (c.f. \eqref{blowup}), then there exists $u_0\in \mathcal{P}_{\infty}$ such that up to a subsequence we have $u_j\rightharpoonup u_0$ weakly in $H^1(B_R^+)$ and $u_j \to u_0$ uniformly in $ \overline {B_R^+}$.
\end{theorem}
\begin{proof}
From the same arguments as in the proof of Theorem \ref{convergence}, we know by Arzela Ascoli theorem that for some $u_0\in C^{\alpha_0}(\overline {B_R^+})$ we have $u_j\to u_0$ uniformly in $B_2^+$ and therefore $u_0=0$ on $\Pi$. Also, with just like we did in the proof of Theorem \ref{convergence}, we have $u_j \rightharpoonup u_0$ weakly in $H^1(B_R^+)$ up to a subsequence and then we obtain
\begin{equation}\label{gamma convergence 2}
\int_{B_R^+}|\nabla u_0|^2+\Lambda(u_0)\,dx \le \liminf_{j\to \infty} \int_{B_R^+}|\nabla u_j|^2+\Lambda(u_j)\,dx
\end{equation}
Notice that \eqref{gamma convergence 2} is same as \eqref{gamma convergence} with $A_j=Id$ for all $j\in \mathbb{N}$ and $\Lambda(0)=0$. Since $u\in \mathcal{Q}_1$, then we have $u_j \in \mathcal{Q}_{r_j}$. Let $r>0$, from the definition of $\mathcal{Q}_{r_j}$ (replacing $s$ by $\frac{r}{r_j}$ and $x_0$ by $0$ in \ref{Q1}) we have
$$
\int_{B_{\frac{2}{r_j}}^+\cap B_{\frac{r}{r_j}}(0)} |\nabla u_j|^2+\Lambda(u_j)\,dx\,\le \, (1+\kappa r^{\alpha}) \int_{B_{\frac{2}{r_j}}^+\cap B_{\frac{r}{r_j}}(0)}|\nabla v|^2+\Lambda(v)\,dx
$$
for all $v$ such that $v-u_j \in H_0^1(B_{\frac{2}{r_j}}^+\cap B_{\frac{r}{r_j}}(0))$. We choose $r=Rr_j$ in the above equation. Then for large value of $j$ such that $R<2/r_j$, we have
$$
\int_{B_R^+}|\nabla u_j|^2+\Lambda(u_j)\,dx \,\le \, (1+k R^{\alpha} r_j^{\alpha}) \int_{B_R^+}|\nabla v|^2+\Lambda (v)\,dx
$$
for all $v \in H^1(B_R^+)$ such that $v-u_j\in H_0^1(B_R^+)$. Let $w\in H^1(\mathbb{R}^N_+)$ such that $u_0-w\in H_0^1(B_R^+)$, we consider the function $w^{\delta}_{j} \in H^1(B_R^+)$ as constructed in \eqref{test function}. Since $w_j^{\delta}-u_j\in H_0^1(B_R^+)$ we have
$$
\int_{B_R^+}|\nabla u_j|^2+\Lambda(u_j)\,dx \,\le \, (1+k R^{\alpha} r_j^{\alpha}) \int_{B_R^+}|\nabla w^{\delta}_j|^2+\Lambda (w^{\delta}_j)\,dx.
$$
From \eqref{gamma convergence 2} and using the fact that $r_j\to 0$, we get
\[
\begin{split}
\int_{B_R^+}|\nabla u_0|^2+\Lambda(u_0)\,dx &\le \liminf_{j\to \infty} \int_{B_R^+}|\nabla u_j|^2+\Lambda(u_j)\,dx\\
&= \liminf_{j\to \infty} \int_{B_R^+}|\nabla w^{\delta}_j|^2+\Lambda (w^{\delta}_j)\,dx.
\end{split}
\]
Now, after proceeding exactly as in the proof of Theorem \ref{convergence}, we obtain
\begin{equation}\label{L2}
\begin{split}
\lim_{\delta\to 0} \left ( \liminf_{j\to \infty} \int_{B_R^+}|\nabla w^{\delta}_j|^2\,dx\right) = \int_{B_R^+}|\nabla w|^2\,dx
\end{split}
\end{equation}
and we have also shown in Theorem \ref{convergence} that
\begin{equation}\label{measure}
\begin{split}
&\lim_{\delta\to 0} \left ( \liminf_{j\to \infty} |\{w_j^{\delta}>0\}| \right) = |\{w>0\} |\\ \mbox{and }
&\lim_{\delta\to 0} \left ( \liminf_{j\to \infty} |\{w_j^{\delta}\le 0\}| \right) = |\{w \le 0\} |.
\end{split}
\end{equation}
Combining equations (\ref{L2}) and (\ref{measure}) we obtain the following
$$
\int_{B_R^+}|\nabla u_0|^2+\Lambda(u_0)\,dx\,\le \, \int_{B_R^+}|\nabla w|^2+\Lambda (w)\,dx.
$$
Hence $u_0$ satisfies \ref{G4}. We have already assume that $u_0$ satisfies \ref{G1} and we can easily verify that $u_0$ also satisfies \ref{G2} and \ref{G3} just like we did in the proof of Theorem \ref{convergence}. Hence we prove that $u_0\in \mathcal{P}_{\infty}$.
\end{proof}
\begin{theorem}\label{non degeneracy 2}
Suppose $r>0$, then for every $D\Subset B_2^+$ there exists $\eta_0\equiv \eta_0(D)>0$ such that of for every $x\in D$ and $B_s(x)\Subset B_2^+$, if we have
$$
\frac{1}{s} \fint _{\partial B_s(x)} u_r^+\,dx\,\le \, \eta_0
$$
then $u_r\le 0$ in $B_{s/4}(x)$.
\end{theorem}
\begin{proof}
Theorem \ref{non degeneracy 2} is an immediate corollary of \cite[Theorem 10.1]{DT15}
\end{proof}
\begin{theorem}\label{L1 convergence 2}
Let $u_j$, $u_0$ be same as in Theorem \ref{convergence 2}. For all $R>0$, the sequence $\{u_{j}>0\}\cap B_R^+$ converge in $L^1(B_R^+)$ to $\{u_0>0\}\cap B_R^+$ upto a subsequence.
\end{theorem}
\begin{proof}
As we proved in Theorem \ref{L1 convergence}, we will prove Theorem \ref{L1 convergence} by proving the following claim
\begin{equation}\label{claim L1}
\lim_{j\to \infty}|\{u_{r_j}>0\} \Delta \{v>0\}\cap D|=0\qquad \forall D\Subset B_R^+.
\end{equation}
The following convergence \eqref{eq3.23} can be proven by following the same steps as in the proof of Theorem \ref{L1 convergence}
\begin{equation}\label{eq3.23}
\lim_{j\to 0}|\{u_0>0\} \setminus \{u_{j}>0\} |=0.
\end{equation}
Fix $D\Subset B_R^+$. We now claim that given $\delta>0$ there exists $j_{\delta, D}\in \mathbb{N}$ such that
\begin{equation}\label{containment in nbd}
\{u_{j}>0\}\cap D \subset \mathcal{N}_{\delta}(\{u_0>0\}\cap D)\qquad\forall j\ge j_{\delta,D}
\end{equation}
($\mathcal{N}_{\delta}(K)$ is $\delta-$ neighbourhood of a set $K$). In the proof of Theorem \ref{L1 convergence}, we have shown that the above claim implies (\ref{other part 2}).
To prove the claim (\ref{containment in nbd}), we assume on the contrary that (\ref{containment in nbd}) is not true. Then there exists a subsequence of $u_j$ (which we again rename it $u_{j}$) and a corresponding sequence of points $\{x_j\}$ such that
$$
x_j \in \{u_{j}>0\}\cap D,\qquad \mbox{and}\qquad \dist(x_j,\{u_0>0\})>\delta
$$
for all $j\in \mathbb{N}$.
That is $u_0\le 0$ in $B_{\delta/2}(x_j)$. Since $x_j \in \{u_{j}>0\}$, from Theorem \ref{non degeneracy 2} we can write
$$
\fint_{\partial B_{\delta/2}(x_j)}u_{j}^+\,dx > \eta_0 \frac{\delta}{2}.
$$
On the other hand, since $u_{j}\to u_0$ uniformly in $D$ and let $x_j\to x^*$ we have
\begin{equation}\label{1.2}
\fint_{\partial B_{\delta/2}(x_j)}u_{j}^+\,dx\to \fint_{\partial B_{\delta/2}(x^*)}u_0^+\,dx
\end{equation}
this implies that $\fint_{\partial B_{\delta/2}(x^*)}u_0^+\,dx\ge \eta_0\frac{\delta}{2}>0$
and also we have,
$$
\fint_{\partial B_{\delta/2}(x_j)}u_0^+\,dx\to \fint_{\partial B_{\delta/2}(x^*)}u_0^+\,dx
$$
since $B_{\delta/2}(x_j) \subset \{u_0\le 0\}$, thus $ \fint_{\partial B_{\delta/2}(x_j)}u_0^+\,dx\to \fint_{\partial B_{\delta/2}(x^*)}u_0^+\,dx= 0$. Which is a contradiction. Thus \eqref{containment in nbd} holds for all $\delta>0$ and therefore
\begin{equation}\label{other part 2}
\lim_{j\to 0}| (\{u_{j}>0\} \setminus \{u_0>0\})\cap D|=0.
\end{equation}
By following same arguments as in proof of Theorem \ref{L1 convergence}, we can prove Theorem \ref{L1 convergence 2}.
\end{proof}
\end{comment}
\begin{comment}
\begin{lemma}[Boundary estimates for $\|\nabla v_j\|_{L^2(B_R^+)}$]
Provided $j$ is large enough that $2S\le \frac{R}{r_j}$, we have
$$
\|\nabla v_j\|_{L^2(B_R^+)}\le C(R)
$$
$C(R)$ depends solely on $R$, dimension, and data of original problem \eqref{P1}.
\end{lemma}
\begin{proof}
We proceed in the proof in the same way as we did in the proof Lemma \ref{L2 bounds}.
Suppose $w_j\in H^1(B_{2R}^+)$ is a function such that $-\Delta w_j=0$ in $B_{2R}^+$ and $w_j-v_j\in H_0^1(B_{2R}^+)$. Since $v_j\in \mathcal{Q}_j$, we have for all $x\in \overline{ B_{2R}^+}$ and $r>0$
$$
\int_{ B_{2R}^+} |\nabla v_j|^2+\Lambda (v_j)\,dx\le (1+kR^{\alpha})\int_{B_{2R}^+}|\nabla w_j|^2+\Lambda (w_j)\,dx.
$$
Thus we can write
$$
\int_{ B_{2R}^+} |\nabla v_j|^2\,dx \le C_1(R) \int_{ B_{2R}^+} |\nabla w_j|^2\,dx + C_2(R)
$$
from this point, we can proceed same as proof of Lemma \ref{L2 bounds} and prove uniform bound on $\int_{ B_{2R}^+} |\nabla w_j|^2\,dx$.
\end{proof}
\end{comment}
\section{The main result}
We rephrase the notion of the tangential touch of the free boundary to the fixed boundary, which is equivalent to the tangential touch condition mentioned in statement of Theorem \ref{main result}.
In the proof of our main result, we will show that given $u\in \mathcal{P}_1$, for every ${\varepsilon}>0$ there exists $\rho_{{\varepsilon}}>0$ such that
\begin{equation}
\partial \{u>0\}\cap B_{\rho}^+ \subset B_{\rho}^+\setminus K_{{\varepsilon}},\qquad \forall\; 0< \rho\le \rho_{{\varepsilon}}
\end{equation}
where
$$
K_{{\varepsilon}}:= \big \{x \in \mathbb{R}^N_+ \,:\, x_N \ge {\varepsilon} \sqrt{x_1^2+...+x_{N-1}^2} \big \}.
$$
\begin{proof}[Proof of Theorem \ref{main result}]
We assume, by contradiction that the free boundaries of functions in $\mathcal{P}_1$ do not touch the origin in a tangential fashion to the plane. Then, there exists ${\varepsilon}>0$ and sequences $v_j \in \mathcal{P}_1$ and $x_j\in F(v_j)\cap K_{{\varepsilon}}$ such that
$ |x_j|\to 0$ as $j\to \infty$. Let $r_j=|x_j|$ and we consider the blowups $u_j := (v_j)_{r_j}$.
Let $u_0 := \lim_{j\to \infty}u_j$ as in Lemma \ref{convergence}. Also, let $x_0\in \partial B_1^+ \cap K_{{\varepsilon}}$ be a limit up to a subsequence (still called $x_j$) such that $x_0= \lim_{j\to \infty}\frac{x_j}{|x_j|}$. Since $x_j\in F(v_j)$, we have $v_j(x_j)=0$. Therefore on rescaling, $u_j(\frac{x_j}{r_j}) =\frac{1}{r_j}v_j(x_j)=0$. In the limit as $j\to \infty$ we have
$$
u_0(x_0)=\lim_{j\to \infty} u_j \left (\frac{x_j}{r_j} \right )=0.
$$
From the density assumption that $u_j$ satisfy condition \eqref{density} and Lemma \ref{L1 convergence} we have for any given $R>0$
\begin{equation}\label{density on rescaling}
\begin{split}
\frac{|\{u_0>0\}\cap B_R^+|}{|B_R^+|}= \fint_{B_R^+}\chi_{\{u_0>0\}}\,dx &=\lim_{j\to \infty}\fint_{B_R^+}\chi_{\{u_j>0\}}\,dx\\
&= \lim_{j\to \infty} \frac{1}{|B_{R r_j}^+|} \int_{B_{Rr_j}^+}\chi_{\{v_j>0\}}\,dx\\
& = \lim_{j\to \infty}\frac{|\{v_j>0\}\cap B_{R r_j}^+|}{|B_{R r_j}^+|} >\mathcal{D}.
\end{split}
\end{equation}
We can see that the computations done in \eqref{density on rescaling}, in fact shows that the density property remains invariant under blowup of any function $v$,. This way, we conclude that the function $(u_0)_0$ which is the blowup limit of $(u_0)$ (in particular $(u_0)_0:= \lim_{r\to 0} (u_0)_r$) also satisfies
\begin{equation}\label{density 2}
\frac{ |\{ (u_0)_0>0 \}\cap B_R^+ |}{|B_R^+ |}>\mathcal{D},\qquad \forall R>0.
\end{equation}
Now, we note that from Lemma \ref{convergence} $u_0\in \mathcal{P}_{\infty}$. Moreover, from \eqref{density on rescaling} $u_0\not\equiv 0$ and from \cite[Theorem 4.2, Lemma 4.3]{KKS06} we have $u_0 \ge 0$ also, from \eqref{density 2}, we conclude $(u_0)_0 \not \equiv 0$. This way, again by \cite[Theorem 4.9]{KKS06}, we have $u_0(x)=c\, x_N^+$ for all $x\in \mathbb{R}^N_+$ for some constant $c>0$.
Hence the function $u_0$ cannot be equal to zero at any point in $\mathbb{R}^N_+$. But we have $x_0\in \partial B_1^+\cap K_{{\varepsilon}}$ and $u_0(x_0)=0$. This leads to a contradiction.
\end{proof}
\begin{remark}
Interested readers may check that the modulus of continuity $\sigma$ mentioned in the statement of Theorem \ref{main result} can be written as
$$
\sigma(r) = \sup \left ( \frac{x_n}{\sqrt {x_1^2+...+x_{N-1}^2}} \,:\,x\in (B_{\rho}^+\setminus \{0\}) \cap F(u), \rho \le r \right ).
$$
\end{remark}
\newpage
\newcommand{\noop}[1]{}
|
\section{Introduction}
\IEEEPARstart{H}{uman} motion prediction is a challenging task that involves modeling dynamic spatiotemporal correlations between body joints. This task is becoming increasingly important in autonomous driving \cite{paden2016survey}, human-machine interaction \cite{kong2018human,unhelkar2018human}, and healthcare \cite{troje2002decomposing}.
Traditional methods can handle simple and periodic scenarios with classical time series processing approaches \cite{taylor2010dynamical,lehrmann2014efficient}, but make unrealistic predictions when the human motion becomes more complex and erratic. Recently, researchers have adopted different deep learning techniques to human motion prediction, such as Convolution Neural Network (CNN) \cite{li2018convolutional,liu2020trajectorycnn}, Recurrent Neural Network (RNN) \cite{al2020attention,guo2019human,martinez2017human,fragkiadaki2015recurrent,martinez2017human,li2018convolutional,tang2018long,liu2019towards,pavllo2020modeling,shu2021spatiotemporal,wang2021pvred}, Generative Adversarial Network (GAN) \cite{gui2018adversarial,hernandez2019human,cui2021efficient,ke2019learning,kundu2019bihmp} and Transformer \cite{cai2020learning,aksan2021spatio}. While these methods have achieved significant advancements over traditional approaches, they are unable to explicitly represent body connections and rely on manual filter size selection \cite{dang2021msr}.
As a human motion sequence is naturally represented as a spatiotemporal graph and can be decomposed into a combination of spatial graphs and temporal graphs (see Figure \ref{fig:motivation} (a)), Graph Convolutions Network (GCN) have been applied by many researchers in motion prediction \cite{mao2019learning,mao2020history,li2020dynamic,cui2020learning,dang2021msr,liu2021motion,sofianos2021space}. Earlier GCN-based methods utilized graph convolutions {GCs} to model shared spatial correlations between joints \cite{mao2019learning, li2020dynamic, dang2021msr, cui2020learning, liu2021motion}. Sofianos et al. \cite{sofianos2021space} first introduced a GC with unshared spatial and temporal correlations, which outperforms correlation-shared GCNs with far fewer parameters.
Despite the promising results of GCN-based models, there are two challenges for realizing high predictability of human motions: (1) \emph{Lightweight spatiotemporal correlation representation}: Although Sofianos \textit{et al.} \cite{sofianos2021space} directly parameterized the spatiotemporal correlations as unshared spatial and temporal adjacency matrices, this method still introduce redundant parameters to represent shared prior knowledge like body connections and temporal context.
There is still room to explore more parameter-saving decomposition representations of spatiotemporal graph.
(2) \emph{Sample-specific spatiotemporal correlations}: as shown in Figure \ref{fig:motivation} (b), spatiotemporal correlations are dynamic and vary between input samples. Graph correlations in all previous works \cite{mao2019learning,mao2020history,li2020dynamic,cui2020learning,liu2021motion,dang2021msr,sofianos2021space} adopt sample-shared graph correlations. Although the sample-shared correlations are learned through back-propagation and optimal for all data, they may disturb individual prediction by mistakenly enhancing unrelated correlations or reducing critical connection strengths. This leads to amplification of slight movement or the inability to infer future motion from tiny cues.
To handle the aforementioned challenges, we look into \emph{constrained dynamic nature of spatiotemporal correlations}, a key characteristic which all previous GCN-based methods ignore. On the one hand, the spatiotemporal correlations are subject to shared constraints like joint connections and temporal context. On the other hand, the correlation strength between certain joints is dynamically enhanced or reduced according to different samples.
Inspired by this, we introduce Dynamic SpatioTemporal Decompose Graph Convolutions (DSTD-GC), which contains Dynamic Spatial Graph Convolution (DS-GC) and Dynamic Temporal Graph Convolution (DT-GC) for constrained dynamic spatial and temporal correlation modeling. Within the constrained correlation modeling, the shared constrained correlation is parameterized as a vanilla spatial/temporal adjacency matrix, which is shared by all time/joints and all samples. This shared correlation serves as motion prior and indicates general connections between vertices. On the other hand, the dynamic refinements are extracted by an adjustment function according to different samples. The sample-specific temporal/spatial-wise adjustments complete sample-shared temporal/spatial-shared constraints. With this formulation, we manage to represent unshared spatial/temporal correlations in quarter space complexity, while the previous representation \cite{sofianos2021space} is of cubic space complexity. Besides, DSTD-GC captures dynamic motion patterns between different samples and adjust correlations between joints, which precisely enhances correlations for mutual movement and reduce correlations for unrelated motion. Furthermore, we analyze different kinds of spatiotemporal graph convolutions in a unified form and theoretically prove that DSTD-GC enhances feature representation by relaxing certain restrictions of spatiotemporal graph convolutions. Combing DSTD-GC and prior knowledge like body connections and temporal context, we propose DCTD-GCN which consistently outperforms other state-of-the-art methods in prediction accuracy with the fewest parameters.
In brief, our contributions are summarized as follow:
\begin{itemize}
\item We propose Dynamic SpatioTemporal Decompose Graph Convolution, which utilizes Constrained Dynamic Correlation Modeling to represent sample-specific unshared correlations in spatiotemporal graph, leading to effective spatiotemporal correlation representation and accurate motion prediction.
\item We mathematically unify all forms of graph convolutions in human motion prediction and show that DSTD-GCN improves representative ability by relaxing limitations of other graph convolutions.
\item We show with extensive experiments that our proposed DSTD-GCN outperforms other state-of-the-art methods in terms of prediction errors and parameter numbers on two benchmark datasets.
\end{itemize}
\section{Related Work}
\subsection{Human Motion Prediction}
Researchers have applied various deep learning techniques in human motion prediction. CNN-based methods mainly treat a human pose as a pseudo-image. \cite{li2018convolutional,liu2020trajectorycnn}. Liu \textit{it al.} \cite{liu2020trajectorycnn} stacked pose sequence along the channel dimension and extract multi-level motion features in different CNN layers. RNN-based methods shows their power in modeling temporal motion features and make consistent predictions \cite{fragkiadaki2015recurrent,martinez2017human,li2018convolutional,tang2018long,liu2019towards,li2020dynamic,pavllo2020modeling,liu2021motion,su2021motion,shu2021spatiotemporal,wang2021pvred,dong2022skeleton}. Martinez \textit{it al.} proposed a simple yet efficient Seq2Seq model and make prediction based on residual velocities. Wang \textit{it al.} \cite{wang2021pvred} extends this framework by introducing joint velocity and frame position.
GAN-based methods make multiple future predictions based on data pattern similarities and generate realistic results \cite{gui2018adversarial,cui2020learning,liu2021aggregated,lyu2021learning}. Lyu \textit{it al.} \cite{lyu2021learning} utilized stochastic differential equations and path integrals.
Transformer-based methods directly model long-range spatial and temporal dependencies \cite{cai2020learning,aksan2021spatio}. Aksan \textit{it al.} designed spatial and temporal transformer to simultaneously update spatiotemporal representation. Although these methods make remarkable advancements compared with traditional methods, they can't directly model natural connectivity between body joints. The natural connectivity is crucial for human motion prediction as human motion follows kinetic chains and trajectory path, which are constrained by body connections and temporal context, respectively. In order to explicitly model body connections, researchers have looked into graph convolution networks (GCNs). Graph Convolutions (GCs) are suitable for non-grid and graph-structural data, which has been successfully applied to social network \cite{derr2018signed}, point cloud \cite{qian2021pu} and traffic prediction \cite{yu2017spatio}. With joints as vertices, bone and trajectory connections as edges, a human pose can be naturally represented as a spatiotemporal graph. Recently, many researchers has applied GCs in motion prediction \cite{mao2019learning,cui2020learning,mao2020history,dang2021msr,sofianos2021space,liu2021motion,li2021multiscale}.
\subsection{Graph Convolution Networks in Motion Prediction}
Nearly all GCN-based prediction approaches were developed based on Kipt \text{et al.} \cite{kipf2016semi}, where features are updated in a two-step manner: (1) Feature transformation with a simple linear transformation or a multi-layer perceptron; and (2) Feature aggregation with graph correlations. The graph correlation is the key component that set GCN apart from other deep learning techniques, which explicitly depicts body connections flexibly. For the spatiotemporal graph correlation, all human prediction methods decomposed it into spatial correlations and temporal correlations. Based on decomposed graph correlations, there are two classifications: (1) \textit{Correlation-shared/Correlation-unshared}: this is according to whether graph correlations are shared by joints in temporal-wise spatial graphs or spatial-wise temporal graphs. (2) \textit{Sample-shared/Sample-specific}: this is based on whether graph correlations are adjusted dynamically to different samples.
\subsubsection{Correlation-shared/Correlation-unshared Methods} For correlation-shared methods, the spatial/temporal correlations are shared across time/joints. Most of previous works \cite{mao2019learning,mao2020history,cui2020learning,liu2021motion,dang2021msr,li2021multiscale} belong to this type. A challenge for these methods is that they can't explicitly model changing spatial relationships in different stages of the motion and diverse temporal patterns in different joints. Thus, they generally need to stack many GC layers to model complex spatiotemporal correlations in human motion. For \textit{correlation-unshared methods}, the spatial correlations vary from time and the temporal correlations vary from joints. Thus, these methods can explicitly depict joint relationships in spatiotemporal graphs and thus explicitly model complex spatiotemporal correlations in human motion with far fewer parameters \cite{sofianos2021space}.
\subsubsection{Sample-shared/Sample-specific Methods} For sample-shared methods, graph correlations keep unchanged for all samples. Nearly all previous works applied this strategy. Mao \textit{et al.} \cite{mao2019learning} and Dang \textit{et al.} \cite{dang2021msr} directly learned graph correlations from data, while Li \textit{et al.} \cite{li2020dynamic} initialize the trainable correlation as a predefined graph. Cui \textit{et al.} \cite{cui2020learning} and Liu \textit{el al.} \cite{liu2021motion} combined previous two correlations and proposed a semi-constrained graph correlation. For \textit{sample-specific methods}, graph correlations can be adjusted to each sample. To our best knowledge, sample-specific GCs are hardly studied in motion prediction. We are the first work to model \emph{sample-specific correlation-unshared GC} in motion prediction.
\begin{figure*}[htbp]
\centering
\includegraphics[width=1\textwidth]{figures/dstdgc_nochannel.pdf}
\caption{The framework of Dynamic Spatial Graph Convolution. Constrained feature update aggregates motion representation with temporal-shared
correlations. Temporal-wise Feature Adjustment obtains motion representations with temporal-specific correlations, where the correlations are obtained with
temporal-wise correlation modeling. Eventually, the output feature is obtained by constrained correlations with temporal-wise adjustment. The introduction of DT-GC is omitted due the spatiotemporal-equivalence between DS-GC and DT-GC.}
\label{fig:framework}
\end{figure*}
\section{Methodology}
\subsection{Preliminaries}
\subsubsection{Problem Definition} Human motion prediction is to predict $L$ future human pose frames with $K$ historical observations. As a human pose is represented by $J$ joints with $D-$dimensional spatial information, we denote the human pose at time $t$ as $X_t \in \mathbb{R}^{J \times D}$. The historical observations are formulated as $X_{1:K} = \{X_1,\cdots,X_K\}$. Our goal is to predict the future human motion $\tilde{X}_{K+1:K+L} = \{\tilde{X}_{K+1}, \cdots,\tilde{X}_{K+L}\}$ where the corresponding ground truth is denoted as $X_{K+1:K+L} = \{X_{K+1},\cdots,X_{K+L}\}$.
\subsubsection{Notations}\label{sec:notations} We formally define spatiotemporal graph and its decomposition used in this paper.
\emph{Spatiotemporal Graph Representation.} As shown in Figure \ref{fig:motivation} (a), a human motion sequence with $T$ frames is presented as a spatiotemporal graph $\mathcal{G}^{st}=(\mathcal{V}^{st}, \mathcal{E}^{st})$, where $\mathcal{V}^{st} \in \mathbb{R}^{JT}$ is the set of all joints across time and $\mathcal{E}^{st}$ is the spatiotemporal edge set. $\boldsymbol{A}^{st} \in \mathbb{R}^{JT \times JT} $ is the spatiotemporal adjacency matrix, which represent correlations between vertices in the graph. However, it cost a lot of parameters to directly store and learn spatiotemporal correlations with $O((JT)^2)$ space complexity. Therefore, we further decompose a spatiotemporal graph into a unique pair of $T$ spatial graphs across time and $J$ temporal graphs across joints. The spatial graphs are denoted as $\mathcal{G}^{s} = (\mathcal{V}^{s}, \mathcal{E}^{s})$ where $\mathcal{V}^{s} \in \mathbb{R}^{T \times J}$ is the joint vertex set. $\mathcal{E}^{s}$ is the spatial edge set and is formulated as a temporal-unshared spatial adjacency matrix\footnote{We denote both second and third order tensor as matrix in this paper for clarity.} $\boldsymbol{A}^{s} \in \mathbb{R}^{T \times J \times J}$. When spatial correlations are shared across time, the spatial adjacency matrix is degraded into vanilla form and is represented by a matrix from $\mathbb{R}^{J \times J}$. Similarly, the temporal graphs are denoted as $\mathcal{G}^{t} = (\mathcal{V}^{t}, \mathcal{E}^{t})$ where $\mathcal{V}^{t} \in \mathbb{R}^{J \times T}$ is the trajectory vertex set. $\mathcal{E}^{t}$ is the temporal edge set and is formulated as a spatial-unshared temporal adjacency matrix $\boldsymbol{A}^{t} \in \mathbb{R}^{J \times T \times T}$.
\emph{Spatiotemporal-equivalence.} Spatiotemporal-equivalence is defined between two functions on spatial graphs and corresponding temporal graphs. Formally, function $\mathcal{F}^{s}$ on $\mathcal{G}^{s}$ and function $\mathcal{F}^{t}$ on corresponding $\mathcal{G}^{t}$ are spatiotemporal-equivalent if and only if $\mathcal{F}^{s}$ is equivalent to $\mathcal{F}^{t}$ after switching all operations for time and space dimension and vice versa. GCs with spatiotemporal-equivalent operations are called spatiotemporal-equivalent GC.
\subsubsection{Spatiotemporal Decompose Graph Convolutions} We introduce different GCs with the example of the feature updating process for the joint $q$ of time $n$. The neighbor indices of the selected joint in the spatial graph and the temporal graph are $p$ and $m$, respectively. For a typical spatiotemporal graph convolution (ST-GC), the output feature $\mathbf{y}_{qn}$ is obtained by:
\begin{equation}
\mathbf{y}_{qn} = \sum_{p}^{J} \sum_{m}^{T} a^{st}_{(pm)(qn)} \mathbf{x}_{pm} \boldsymbol{W},
\label{eq:st-gc}
\end{equation}
where $\mathbf{x}_{pm} \in \mathbb{R}^{C}$ is the input feature, $\mathbf{y}_{qn} \in \mathbb{R}^{C'}$ is the output feature, $a^{st}_{(pm)(qn)}$ is a correlation strength in the spatiotemporal adjacency matrix $\boldsymbol{A}^{st}$ and $\boldsymbol{W} \in \mathbb{R}^{C \times C'}$ is a trainable parameter for feature transformation.
With the graph decomposition introduced above, we define two graph GCs: spatial graph convolution (S-GC) and temporal graph convolution (T-GC), which are described as:
\begin{equation}
\mathbf{x}^{s}_{qn}= \sum_{p}^{J} a^{s}_{npq} \mathbf{x}_{pn} \boldsymbol{W},
\label{eq:sgcn}
\end{equation}
\begin{equation}
\mathbf{x}^{t}_{qn} = \sum_{m}^{T} a^{t}_{qmn} \mathbf{x}_{qm} \boldsymbol{W}.
\label{eq:tgcn}
\end{equation}
Specifically, the S-GC models temporal-wise spatial correlations, while T-GC models spatial-wise temporal correlations. We then propose \textit{SpatioTemporal Decompose Graph Convolution (STD-GC)} by alternatively stacking a S-GC (Equation \ref{eq:sgcn}) and a T-GC (Equation \ref{eq:tgcn}):
\begin{equation}
\mathbf{y}_{qn} = \sum_{m}^{T} a^{t}_{qmn} (\sum_{p}^{J} a^{s}_{npq} \mathbf{x}_{pm} \boldsymbol{W}_{1}) \boldsymbol{W}_{2},
\label{eq:stdgcn}
\end{equation}
where $a^{s}_{npq}$ and $a^{t}_{qmn}$ are from spatial adjacency matrix $\boldsymbol{A}^{s}$ and temporal adjacency matrix $\boldsymbol{A}^{t}$. STD-GC decomposes ST-GC on a spatiotemporal graph into S-GC on spatial graphs and T-GC on temporal graphs, spatiotemporal feature modeling is accomplished by stacking these two GCs. The spatiotemporal modeling capability is invariant to the stacking order of S-GC and T-GC. Besides, STD-GC is equivalent to STS-GC \cite{sofianos2021space}. We demonstrate all of these with experiments in Section \ref{sec:gc_compare}. Furthermore, STD-GC becomes vanilla STD-GC (VSTD-GC ) when the spatial and temporal correlations are in vanilla form.
In sample-shared methods, $a^{s}_{npq}$ and $a^{t}_{qmn}$ are set based on prior knowledge or defined as trainable parameters. In sample-specific methods, $a^{s}_{npq}$ and $a^{t}_{qmn}$ are generated by model according to input samples.
\begin{comment}
STS-GC \cite{sofianos2021space} alleviates this problem by discomposing the spatiotemporal adjacency matrix into a spatial adjacency matrix and a temporal adjacency matrix. The feature updating process is defined as:
\begin{equation}
\mathbf{y}_{qn} = \sum_{p}^{J} \sum_{m}^{T} a^{s}_{npq} a^{t}_{qmn} \mathbf{x}_{pm} \boldsymbol{W}.
\label{eq:sts-gc}
\end{equation}
For one thing, STD-GC is a more general model and mathematically equivalent to STS-GC when $W_{1}W_{2} = W$. For another, STD-GC is also a decomposition form of ST-GC, the spatiotemporal feature modeling is accomplished by consecutive stacking of the two GCs.
Besides, STD-GC is invariant to the stacking order of S-GC and T-GC. We will illustrate this in the experiment.
\end{comment}
\subsection{Constrained Dynamic Correlation Modeling}
The correlations between joints are both constrained and dynamic. For one thing, the correlations between joints are constrained by the inherent human body structure and learned motion prior. For another, these constrained correlations are varied from different motion sequences. To show these two aspects, we utilize correlations of the elbow joint in the walking and directing traffic scenarios as an example. On the one hand, the elbow movement generally follows the motion chain of the human body and shared trajectory patterns. These constraints are represented by shared spatiotemporal correlations. On the other hand, the spatial correlations between the elbow joint and the knee joint are generally stronger than the correlation in directing traffic scenarios, and temporal correlations in the knee joint are stronger than the ones in directing traffic.
Inspired by this, we can represent spatiotemporal graph correlations with lightweight vanilla spatial and temporal adjacency matrices and a function to extract dynamic unshared correlations. Therefore, we introduce \textit{Dynamic SpatioTemporal Decompose Graph Convolution (DSTD-GC)}. DSTD-GC extends STD-GC with Dynamic Spatial Graph Convolution and Dynamic Temporal Graph Convolution. Specifically, DS-GC update motion feature with constrained dynamic spatial correlations, while DT-GC with constrained dynamic temporal correlations. By combing these two GCs, DSTD-GC can model constrained dynamic spatiotemporal correlations. These two GCs are spatialtemporal-equivalent. For clarity, We only introduce DS-GC, and its spatiotemporal-equivalent counterpart DT-GC can be deduced naturally. The general framework of DS-GC is illustrated in Figure \ref{fig:framework}. Specifically, DS-GC consists of two parts: (1) Constrained feature update with shared vanilla correlations; (2) Temporal-wise feature adjustment. The first part updates motion features with shared constraints, while the last part completes feature representations by modeling sample-specific temporal-wise spatial correlation adjustments. Specifically, A DS-GC layer receives motion feature $\boldsymbol{X} \in \mathbb{R}^{J \times T \times C}$ and shared constrained spatial correlation $\boldsymbol{A}^{s} \in \mathbb{R}^{J \times J}$ as inputs, and then outputs $\boldsymbol{Y} \in \mathbb{R}^{J \times T \times C}$.
\subsubsection{Constrained Feature Update} As shown in the blue block of Figure \ref{fig:framework}, constrained feature update is identical to vanilla S-GC where spatial correlations are shared across time. Specifically, feature transformation is accomplished by transformation function $\mathcal{T}$. Here we set the function as a linear transformation function for clarity while other functions like a multi-layer perceptron can also be used. The function is defined as:
\begin{equation}
\boldsymbol{F}^{s} = \mathcal{T}(\boldsymbol{X}) = \boldsymbol{X} \boldsymbol{W},
\end{equation}
where $\boldsymbol{F}^{(s)} \in \mathbb{R}^{J \times T \times C^{'}}$ is the high-level representations. After feature transformation, features are aggregated across time with aggregation function $\mathcal{A}^{s}_{t}$ to obtain updated constrained feature $\boldsymbol{C}^{s}$. The whole process to obtain in constrained feature update is formulated as:
\begin{equation}
\boldsymbol{C}^{s} = \mathcal{A}^{s}_{t}(\boldsymbol{F}^{s}, \boldsymbol{A}^{s}) = [\boldsymbol{F}^{s}_{:, 1} \boldsymbol{A}^{s} ||_{t} \cdots ||_{t} \boldsymbol{F}^{s}_{:, T} \boldsymbol{A}^{s}],
\end{equation}
where $||_{t}$ is concatenation function along the time dimension. $\boldsymbol{A}^{s}$ is vanilla spatial correlations and guide feature aggregation with the broadcast mechanism.
\subsubsection{Temporal-wise Feature Adjustment} The temporal-wise feature adjustment is shown in the green part of Figure \ref{fig:framework}. Here temporal-specific spatial correlations $\boldsymbol{M}^{s} \in \mathbb{R}^{T \times J \times J}$ is inferred from the input sample to capture dynamic temporal-wise spatial correlations. Specifically, $\boldsymbol{M}^{s}$ is extracted by the temporal-wise spatial correlation modeling function $\mathcal{M}^{s}_{t}$, which extract correlations between all pairs of joints. This function for a joint pair $(j_{p}, j_{q})$ is formulated as:
\begin{equation}
\boldsymbol{M}^{s}= \mathcal{M}^{s}_{t}(\mathbf{x}_{p}, \mathbf{x}_{q}) = \operatorname{MLP}_{1}(\theta(\boldsymbol{x}_{p}) ||^{p}_{s} \phi(\mathbf{x}_{q})),
\label{eq:dsgc-dc}
\end{equation}
where $(\mathbf{x}_{p},\mathbf{x}_{q})$ are from the input feature $\boldsymbol{X}$ of the joint pair $(j_{p}, j_{q})$ and $||^{p}_{s}$ is pair-wise feature concatenation along the spatial dimension. $\theta$ and $\phi$ are two linear transformation functions, which project the joint features into low-dimensional representations to reduce computational cost. $\operatorname{MLP}_{1}$ is a multi-layer perceptron, we utilize $\operatorname{MLP}$ because it can model complex spatial correlations. Note that $\boldsymbol{M}^{s}$ is not required to be symmetric, this benefits correlation modeling by increasing analysis flexibility and enhancing representative ability. Besides, $\boldsymbol{M}^{s}$ is also varied from different input samples.
Given the temporal-specific spatial correlations $\boldsymbol{M}^{s}$ from input samples, we aggregate features with temporal-wise aggregation function $\mathcal{A}^{s}_{t}$, where aggregated feature $\hat{\boldsymbol{F}^{s}} \in \mathbb{R}^{J \times T \times C^{'}}$ is obtained by:
\begin{equation}
\hat{\boldsymbol{F}}^{s} = \mathcal{A}^{s}_{t}(\boldsymbol{F}^{s}, \boldsymbol{M}^{s}) = [\boldsymbol{F}^{s}_{:, 1} \boldsymbol{M}^{s} ||_{t} \cdots ||_{t} \boldsymbol{F}^{s}_{:, T} \boldsymbol{M}^{s}].
\end{equation}
Finally, the output representation with dynamic temporal-wise adjustment, $\boldsymbol{H}^{(l+1)}$, is obtained by adjusting constrained representation $\boldsymbol{C}^{s}$ with temporal-specific representation $\hat{\boldsymbol{F}}^{s}$:
\begin{equation}
\boldsymbol{E}^{s} = \mathcal{U}^{s}_{t}(\boldsymbol{C}^{s},\hat{\boldsymbol{F}}^{s}) = \boldsymbol{C}^{s} + \alpha \cdot \hat{\boldsymbol{F}}^{s},
\label{eq:dsgc}
\end{equation}
where $\alpha$ is a learnable parameter to control the adjustment intensity. Therefore, DS-GC can be formulated as:
\begin{equation}
\boldsymbol{Y} = \mathcal{U}^{s}(\mathcal{A}^{s}_{c}(\mathcal{T}(\boldsymbol{X}), \boldsymbol{A}^{s}), \mathcal{A}^{s}_{t}(\mathcal{T}(\boldsymbol{X}), \mathcal{M}^{s}_{t}(\boldsymbol{X}))).
\end{equation}
Since DT-GC is spatiotemporal-equivalent to DS-GC, DT-GC can be directly obtained by replacing all operations in DS-GC with their spatiotemporal-equivalent counterparts.
\begin{comment}
\subsubsection{Channel-wise Feature Adjustment} Similar to the temporal-wise feature adjustment, the channel-wise feature modeling first models channel-wise spatial correlations $\boldsymbol{R}^{s}$ then aggregates and adjusts features on the channel dimension. We omit detail introduction and provide formulation for clarification:
\begin{equation}
\boldsymbol{R}^{s}= \mathcal{M}^{s}_{c}(\boldsymbol{f}^{s}_{x}, \boldsymbol{f}^{s}_{y}) = \operation{MLP}_2(\gamma(\boldsymbol{f}^{s}_{x}) - \delta(\boldsymbol{f}^{s}_{y})),
\end{equation}
\begin{equation}
\hat{\boldsymbol{E}}^{s} = \mathcal{A}^{s}_{c}(\boldsymbol{E}^{s}, \boldsymbol{R}^{s}) = [\boldsymbol{E}^{s}_{:,:, 1} \boldsymbol{R}^{s} ||_{c} \cdots ||_{c} \boldsymbol{E}^{s}_{:, :, C^{'}} \boldsymbol{R}^{s}],
\end{equation}
\begin{equation}
\boldsymbol{H}^{(l+1)} = \mathcal{U}^{s}_{c}(\boldsymbol{E}^{s},\hat{\boldsymbol{F}}^{s}) = \boldsymbol{C}^{s} + \beta \cdot \hat{\boldsymbol{E}}^{s}.
\end{equation}
\end{comment}
\subsubsection{Discussion} Compared with STS-GC/STD-GC, DSTD-GC contains three advantages: (1) \textit{Parameter-saving}: STS-GC learn a spatial adjacency matrix $\boldsymbol{A}^{s}$ and a temporal adjacency matrix $\boldsymbol{A}^{t}$, where the space complexity is $O(JT^{2} + J^{2}T)$. Conversely, DSTD-GC saves a great many parameters by learning the previous two adjacency matrices in vanilla form with few extra operations, where the space complexity is $O(T^{2} + J^{2})$. As the joint number is always proportional to the sequence length, The space complexity for STS-GC is $O(T^{3})$ while the one DSTD-GC is only with $O(T^{2})$. (2) \textit{Joint optimization}: DSTD-GC learns shared graph correlations and dynamic correlation variations between different times or different joints. This mechanism has less optimization difficulty than learning optimal spatial correlation at different times or temporal correlations in different joints separately. (3) \textit{Dynamic}: DSTD-GC adopts an adjustment strategy that can better model dynamic spatiotemporal correlations in different motion inputs.
\begin{figure*}[!ht]
\centering
\includegraphics[width=1\textwidth]{figures/details_nochannel.pdf}
\caption{(a) Problem reformulation and prediction framework. Human poses in the blue dashed box are duplicate inputs, while poses in the red box are target predictions. (b) The basic block of DSTD-GCN. The spatial modeling aims to adjust natural connection constraint and semantic connection constraint separately with two DS-GCs, while the temporal modeling refines self-loop temporal adjacency constraint with a DT-GC. (c) Implementation of DS-GC. Operations with different colors indicate different functions. The introduction of DT-GC is omitted due to the spatiotemporal-equivalence between DS-GC and DT-GC.}
\label{fig:details}
\end{figure*}
\subsection{Analysis of Graph Convolutions on Spatiotemporal Graphs}
By reformulating several graph convolutions on spatiotemporal graphs into an uniform form, we assess their representation capabilities in human motion prediction. We begin by introducing the spatiotemporal graph decomposition constraint, and then evaluate several GCs on the basis of their spatial and temporal modeling capability. We focus exclusively on the correlation between distinct GCs, as correlation dominates feature aggregation and is hence critical for obtaining representative spatiotemporal feature. Based on previous formulation, we add $(i)$ in the superscript to indicate different samples.
By comparing ST-GC (Equation \ref{eq:st-gc}) and STD-GC (Equation \ref{eq:stdgcn}), the spatiotemporal correlations can be represented by a combination of spatial correlations and temporal correlations with the following constraints: \\
\textit{Constraint 1: $a^{st(i)}_{(pm)(qn)}$ equals to the product of $a^{s(i)}_{npq}$ and $a^{t(i)}_{qmn}$}. \\
Note that here $a^{st(i)}_{(pm)(qn)}$ is the spatiotemporal correlation strength, $a^{s(i)}_{npq}$ is corresponding spatial correlation strength, and $a^{t(i)}_{qmn}$ is corresponding temporal correlation strength.
We further reformulate decomposed ST-GC from the spatial and temporal graph perspectives using the decomposition constraint. These two views can be defined as follows:
\begin{equation}
\begin{aligned}
\mathbf{y}^{(i)}_{qn} &= \sum_{p}^{J} a^{s(i)}_{npq} \sum_{m}^{T} \mathbf{x}^{(i)}_{pm} (a^{t(i)}_{qmn} \boldsymbol{W})\\ &= \sum_{p}^{J} a^{s(i)}_{npq} \sum_{m}^{T} \mathbf{x}^{(i)}_{pm} \boldsymbol{R}^{t(i)}_{qmn},
\end{aligned}
\label{eq:gc-s}
\end{equation}
\begin{equation}
\begin{aligned}
\mathbf{y}^{(i)}_{qn} &= \sum_{m}^{T} a^{t(i)}_{qmn} \sum_{p}^{J} \mathbf{x}^{(i)}_{pm} (a^{s(i)}_{npq} \boldsymbol{W}) \\ &= \sum_{m}^{T} a^{t(i)}_{qmn} \sum_{p}^{J} \mathbf{x}^{(i)}_{pm} \boldsymbol{R}^{s(i)}_{npq},
\end{aligned}
\label{eq:gc-t}
\end{equation}
where $a^{s(i)}_{npq}$ and $a^{t(i)}_{qmn}$ are the spatial correlation weight and temporal correlation weight for the input sample. Furthermore, we combine correlation strength and feature transformation weights into a generalized weight matrix $\boldsymbol{R}^{s(i)}_{npq}$ or $\boldsymbol{R}^{t(i)}_{qmn}$. Therefore, the feature updating process consists of feature transformation with the generalized weight matrix and feature aggregation with temporal-wise spatial or spatial-wise temporal correlations. With these two formulations, we then analyze different graph GCs based on the spatial and temporal correlations for feature aggregation.
\subsubsection{Sample-shared Correlation-shared GCs} Most recent works \cite{yan2018spatial,mao2019learning,mao2020history,cui2020learning,dang2021msr,li2021multiscale} adopted correlation-shared GC, where joint vertices from all times share one spatial correlation and trajectory vertices for all joints share one temporal correlation.
The whole GC process can be formulated as:
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{p}^{J} \sum_{m}^{T} a^{s}_{pq} a^{t}_{mn} \mathbf{x}^{(i)}_{pm} \boldsymbol{W},
\label{eq:s/tgc}
\end{equation}
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{p}^{J} a^{s}_{pq} \sum_{m}^{T} \mathbf{x}^{(i)}_{pm} (a^{t}_{mn} \boldsymbol{W}),
\label{eq:s/tgc-s}
\end{equation}
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{m}^{T} a^{t}_{mn} \sum_{p}^{J} \mathbf{x}^{(i)}_{pm} (a^{s}_{pq} \boldsymbol{W}).
\label{eq:s/tgc-t}
\end{equation}
Note that $a^{s}_{pq}$ is temporal-shared spatial correlation strength and $a^{t}_{mn}$ is spatial-shared spatial correlation strength, they are respectively from the vanilla form of the spatial and the temporal adjacency matrix $\boldsymbol{A}^{s}$ and $\boldsymbol{A}^{t}$, respectively. Equation \ref{eq:s/tgc-s} and \ref{eq:s/tgc-t} are the reformulations of Equation \ref{eq:s/tgc} in spatial graph's view and temporal graph's view respectively. Besides, these two correlation strengths are stay unchanged across input samples Therefore, the sample-shared correlation-shared GCs are subject to following constraints: \\
\textit{Constraint 2: $a^{s(i)}_{m_{1}pq}$ and $a^{s(i)}_{m_{2}pq}$ are forced to be the same, $a^{t(i)}_{q_{1}mn}$ and $a^{t(i)}_{q_{2}mn}$ are forced to be the same.} \\
\textit{Constraint 3: $a^{s(i_1)}_{mpq}$ and $a^{s(i_2)}_{mpq}$ are forced to be the same, $a^{t(i_1)}_{qmn}$ and $a^{t(i_2)}_{qmn}$ are forced to be the same.} \\
Note that $i_{1},i_{2}$ are different input sample indices, $m_{1},m_{2}$ are indices of different joints and $q_{1},q_{2}$ indicate indices of different time.
\subsubsection{Sample-shared Correlation-unshared Graph GCs} The only difference between sample-shared correlation-shared GCs and sample-shared correlation-unshared GCs is that the sample-shared spatiotemporal ones adopt unshared spatial and temporal correlations. Sample-shared correlation-unshared GCs can be formulated as:
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{p}^{J} \sum_{m}^{T} a^{s}_{npq} a^{t}_{qmn} \mathbf{x}^{(i)}_{pm} \boldsymbol{W},
\label{eq:stgc-static}
\end{equation}
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{p}^{J} a^{s}_{npq} \sum_{m}^{T} \mathbf{x}^{(i)}_{pm} (a^{t}_{qmn} \boldsymbol{W}),
\label{eq:stgc-s-static}
\end{equation}
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{m}^{T} a^{t}_{qmn} \sum_{p}^{J} \mathbf{x}^{(i)}_{pm} (a^{s}_{npq} \boldsymbol{W}),
\label{eq:stgc-t-static}
\end{equation}
where $a^{s}_{npq}$ is temporal-wise spatial correlation strength and $a^{t}_{qmn}$ is spatial-shared temporal correlation strength. They are respectively from the spatial adjacency matrix and the temporal adjacency matrix. With the two unshared correlations, sample-shared correlation-unshared GCs capture evolving spatial relationships between joints at different action stages and diverse temporal relationships for different body parts. Thus, they generally outperform correlation-shared GCs. However, these correlations are optimized across all data and may not be optimized for individual action sequences. From the formula's view, sample-shared correlation-unshared GCs still suffer from \textit{Constraint 2} but relax \textit{Constraint 1} into the following constraint: \\
\textit{Constraint 4: $a^{s(i)}_{m_{1}pq}$ and $a^{s(i)}_{m_{2}pq}$ differ by a scaling factor, $a^{t(i)}_{q_{1}mn}$ and $a^{t(i)}_{q_{2}mn}$ differ by a scaling factor.}
\subsubsection{Sample-specific Correlation-unshared Graph GCs} In comparison to sample-shared correlation-unshared GCns, the dynamic ones infer sample-specific correlations between vertices and thus have a higher capacity for representation. The formulation for this GCs is:
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{p}^{J} \sum_{m}^{T} a^{s(i)}_{npq} a^{t(i)}_{qmn} \mathbf{x}^{(i)}_{pm} \boldsymbol{W},
\label{eq:stgc-dynamic}
\end{equation}
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{p}^{J} a^{s(i)}_{npq} \sum_{m}^{T} \mathbf{x}^{(i)}_{pm} (a^{t(i)}_{qmn} \boldsymbol{W}),
\label{eq:stgc-s-dynamic}
\end{equation}
\begin{equation}
\mathbf{y}^{(i)}_{qn} = \sum_{m}^{T} a^{t(i)}_{qmn} \sum_{p}^{J} \mathbf{x}^{(i)}_{pm} (a^{s(i)}_{npq} \boldsymbol{W}).
\label{eq:stgc-t-dynamic}
\end{equation}
Note that both $a^{s(i)}_{npq}$ and $a^{t(i)}_{qmn}$ are adjustable to the $i$-th input sample. Based on these formulations, sample-specific correlation-unshared GCs relax both \textit{Constraint 2} and \textit{3}. Concretely, they relax \textit{Constraint 2} into \text{Constraint 4} and relax \textit{Constraint 3} into the following constraint:\\
\textit{Constraint 5: $a^{s(i_1)}_{npq}$ and $a^{s(i_2)}_{npq}$ differ by a scaling factor, $a^{t(i_1)}_{qmn}$ and $a^{t(i_2)}_{qmn}$ differ by a scaling factor.}
We summarize different types of GCs on spatiotemporal graphs in Table \ref{tab:graph modeling}. It can be seen that sample-specific correlation-unshared GC is the least constrained spatiotemporal decompose GC. Our DSTD-GC belongs to sample-specific correlation-unshared GC. Thus, DSTD-GC has a stronger representation capability than previous graph convolutions \cite{mao2019learning,sofianos2021space}. We can reformulate DSTD-GC into Equation \ref{eq:stgc-dynamic}, Equation \ref{eq:stgc-s-dynamic} and Equation \ref{eq:stgc-t-dynamic} respectively, as we will shown in the appendix.
\subsection{Model Architecture}
Based on DSTD-GC, we build a light-weight and powerful model DSTD-GCN, which builds upon the prediction framework proposed by \cite{mao2019learning}. We first introduce the prediction framework and then the implementation of DSTD-GCN.
\subsubsection{Prediction Framework} We duplicate the last input human pose for $L$ times and formulate a new input motion sequence $X_{1:K+L}$. Then the original prediction task is reformulated into predicting the residual motion sequence between $\tilde{X}_{1:K+L}$ and corresponding ground truth $X_{1:K+L}$. Based on the problem reformulation, the prediction framework is proposed as shown in Figure \ref{fig:details} (a). The framework contains a encoding GC layer to project motion input to high-level representation, several GC basic blocks for spatiotemporal representation modeling and a decoding GC to make predictions from motion feature. Besides, the global residual connection enforces the residual learning, while the local residual connection stabilizes the training process. The effectiveness of this framework is illustrated in recent studies \cite{mao2019learning,cui2020learning,dang2021msr}. With this framework, we can model the input and output human poses as a whole sequence and model associated spatiotemporal correlations, while the previous framework \cite{li2020dynamic,sofianos2021space} only models spatiotemporal correlations for the input frames. We will use this framework to build our model and compare different GCs.
\subsubsection{Model Implementation} The entire network adopts the prediction framework and stacks five basic blocks. The basic block of DSTD-GCN is shown in Figure \ref{fig:details} (b), which contains two parallel DS-GCs for spatial modeling and a DT-GC for temporal modeling. Due to the spatiotemporal-equivalence between DS-GC and DS-GC, we only introduce the detailed implementation of DS-GC. Specifically, the modeling function $\mathcal{M}^{s}_{t}$ receives inputs $X \in \mathbb{R}^{J \times C}$. We squeeze the temporal dimension into the channel dimension and direct extract temporal-wise spatial correlations with a linear transformation. Since the computation for direct transformation from $T \times C$ to $T$ is high. We utilize two linear transformation functions $\theta$ and $\phi$ to transform $\boldsymbol{X}$ into a neatly compact representation, then preforms a temporal squeeze operation. After, a pair-wise concatenation and $MLP$ are adapted to extract temporal-wise spatial correlations. The final result is a combination of constrained and dynamic correlations.
We further consider prior knowledge connections and initialize constrained spatial correlations as natural connections and semantic connections such as relationships between two hands, respectively. The initial constrained temporal correlations are set as connections to oneself and its context frames.
\section{Experiments}
To evaluate the effectiveness of our proposed model, we run experiments on two standard benchmark motion capture datasets, which include Human3.6M \cite{ionescu2013human3} and the CMU Mocap dataset. Experiments are designed to answer the following research questions: \textit{(1) what is the performance of DSTD-GC as compared to other GCs? (2) what is the performance of our proposed DSTD-GCN as compared to other state-of-art motion prediction methods? (3) what is the influence of different components and what insights and findings can we learn from the experimental results?}
\subsection{Datasets Settings}
\subsubsection{Human3.6M} Human3.6M \cite{ionescu2013human3} is considered to be a wide-spread benchmark dataset for evaluating motion prediction. It contains 15 actions performed by 7 actors. Following the data preprocessing procedure of \cite{martinez2017human,mao2019learning}, the original data in exponential mapping format is transformed into 3D joint coordinate space and a single pose is represented by 22 body joints. Besides, each data sequence is down-sampled to 25 FPS. The data from S5 and S11 are used as test and validation datasets, while the remaining 5 subjects are for training.
\subsubsection{CMU Mocap} CMU Mocap dataset\footnote{http://mocap.cs.cmu.edu/} is another well-known benchmark dataset. Each pose is represented by 25 body joints. The rest pre-processing procedure is the same as Human3.6M. We adopt the same train/test split strategy as \cite{mao2019learning,dang2021msr}.
\subsection{Comparison Settings}
\subsubsection{Metrics} The Mean Per Joint Position Error (MPJPE) is a standard evaluation metric used in previous studies \cite{martinez2017human,li2018convolutional,dang2021msr}. With the predicted motion sequence $\tilde{X}_{K+1:K+L}$ and corresponding ground truth $X_{K+1:K+L}$, the MPJPE loss is defined as
\begin{equation}
\mathcal{L}_{\text {MPJPE}}=\frac{1}{J \times L} \sum_{t=K+1}^{K+L} \sum_{j=1}^{J}\left\|\tilde{p}_{j, t}-p_{j, t}\right\|_{2},
\end{equation}
where $\tilde{p}_{j,t} \in \mathbb{R}^{D}$ is the predicted $j$th joint position in frame $t$, while $p_{i, t} \in \mathbb{R}^{D}$ is the corresponding ground truth.
\subsubsection{Implementation Details} We conduct all experiments on one RTX 3080 Ti GPU with the Pytorch framework \cite{paszke2019pytorch}. For the comparison baseline and DSTD-GCN, we set the dropout rate as 0.1 and chose PReLU \cite{he2015delving} as the activation function. Besides, all the models were trained with Adam optimizer \cite{kingma2014adam} with an initial learning rate of $3e-3$, which decays by 0.9 for every 5 epochs. We also use MPJPE as the loss function and set the batch size as 32. For the test set and sequence length settings, we follow the experiment setting of Dang \textit{et al.} \cite{dang2021msr}. We use the full test set and set the input length as 10, output length as 25.
\begin{comment}
\begin{table*}[htbp]
\setlength\tabcolsep{1pt}
\begin{center}
\caption{Comparison of different graph convolutions.}
\label{tab:graph modeling}
\resizebox{\textwidth}{30mm}{
\begin{tabular}{cc|ccccc|ccc|ccccccc}
\hline
\multicolumn{2}{c|}{Correlations} &
\multicolumn{5}{c|}{Constraints} &
\multirow{2}{*}{GCs} &
\multirow{2}{*}{Params.} &
\multirow{2}{*}{FLOPs} &
\multicolumn{7}{|c}{\textit{3D Error}} \\
Unshared &
sample-specific &
1 &
2 &
3 &
4 &
5 &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c|}{} &
80 &
160 &
320 &
400 &
560 &
1000 &
Average \\
\hline
\Checkmark & \XSolidBrush & & \Checkmark & \Checkmark & & & ST-GC (\ref{sec:notations}) & 5.44M & 63.63M & 8.70 & 16.16 & 31.34 & 39.16 & 53.91 & 85.30 & 39.09 \\
\hline
\XSolidBrush & \XSolidBrush & \Checkmark & \Checkmark & \Checkmark & & & VSTD-GC (\ref{sec:notations}) & 0.10M & 63.63M & 8.48 & 16.38 & 33.63 & 42.33 & 58.37 & 93.05 & 40.79 \\
\hline
\XSolidBrush & \XSolidBrush & \Checkmark & \Checkmark & \Checkmark & & & FC-GC \cite{mao2019learning} & 0.29M & 9.12M & 13.11 & 24.87 & 46.47 & 56.16 & 73.75 & 106.75 & 50.84 \\
\hline
\Checkmark & \XSolidBrush & \Checkmark & & \Checkmark & \Checkmark & & STS-GC \cite{sofianos2021space} & 0.45M & 63.63M & 8.33 & 15.62 & 30.97 & 38.77 & \underline{53.70} & \underline{86.34} & 38.29 \\
\hline
\Checkmark & \XSolidBrush & \Checkmark & & \Checkmark & \Checkmark & & STD-GC & 0.46M & 63.63M & \underline{8.78} & \underline{16.25} & \underline{31.12} & \underline{38.41} & 52.15 & 80.93 & \underline{37.85} \\
\hline
\Checkmark & \Checkmark & \Checkmark & & & \Checkmark & \Checkmark & DSTD-GC & 0.13M & 70.54M & \textbf{8.10} & \textbf{15.16} & \textbf{29.88} & \textbf{37.01} & \textbf{50.24} & \textbf{79.67} & \textbf{36.68} \\
\hline
\end{tabular}}
\end{center}
\end{table*}
\end{comment}
\begin{table}[!htp]
\setlength\tabcolsep{1.5pt}
\begin{center}
\caption{Comparison of different graph convolutions.}
\label{tab:graph modeling}
\resizebox{0.48 \textwidth}{16mm}{
\begin{tabular}{cc|cc|ccccc|ccccccc}
\hline
\multirow{2}{*}{GCs} &
\multirow{2}{*}{Params.} &
\multicolumn{2}{c|}{Correlations} &
\multicolumn{5}{c|}{Constraints} &
\multicolumn{7}{c}{\textit{MPJPE}} \\
\multicolumn{1}{c}{} &
\multicolumn{1}{c|}{} &
Unshared &
Sample-specific &
1 &
2 &
3 &
4 &
5 &
80 &
160 &
320 &
400 &
560 &
1000 &
Average \\
\hline
ST & 5.44M & - & - & & & & & & 8.70 & 16.16 & 31.34 & 39.16 & 53.91 & 85.30 & 39.09 \\
\hline
\hline
VSTD & 0.10M &\XSolidBrush & \XSolidBrush & \Checkmark & \Checkmark & \Checkmark & & & 8.48 & 16.38 & 33.63 & 42.33 & 58.37 & 93.05 & 42.04 \\
FC \cite{mao2019learning} & 0.29M &\XSolidBrush & \XSolidBrush & \Checkmark & \Checkmark & \Checkmark & & & 13.11 & 24.87 & 46.47 & 56.16 & 73.75 & 106.75 & 53.52 \\
\hline
STS \cite{sofianos2021space} & 0.45M &\Checkmark & \XSolidBrush & \Checkmark & & \Checkmark & \Checkmark & & 8.33 & 15.62 & 30.97 & 38.77 & 53.70 & 85.34 & 38.79 \\
TSD & 0.46M & \Checkmark & \XSolidBrush & \Checkmark & & \Checkmark & \Checkmark & & 7.85 & 14.97 & 30.61 & 38.94 & 54.90 & 87.51 & 39.13 \\
STD & 0.46M & \Checkmark & \XSolidBrush & \Checkmark & & \Checkmark & \Checkmark & & 8.14 & 15.45 & 30.71 & 38.51 & 53.75 & 86.10 & 38.78 \\
\hline
DTSD & 0.13M & \Checkmark & \Checkmark & \Checkmark & & & \Checkmark & \Checkmark & \underline{7.74} & \textbf{14.11} & \underline{29.72} & \underline{37.68} & \underline{52.47} & \underline{86.20} & \underline{37.99} \\
DSTD & 0.13M & \Checkmark & \Checkmark & \Checkmark & & & \Checkmark & \Checkmark & \textbf{7.36} & \underline{14.21} & \textbf{29.29} & \textbf{36.91} & \textbf{51.57} & \textbf{84.66} & \textbf{37.33} \\
\hline
\end{tabular}}
\end{center}
\end{table}
\vspace{-10pt}
\subsection{RQ1: Comparison with other GCs}\label{sec:gc_compare}
we compare the parameters and prediction error of DSTD-GC against other GCs in Table \ref{tab:graph modeling}. Specifically, we exam three key aspects of DSTD-GCN: unshared correlation, sample-specific correlation, and stacking order invariance. For a fair comparison, we keep the prediction framework unchanged and merely replace graph convolution in the encoder, decoder and basic blocks. Moreover,
we exchange linear transformation to a two-layer $MLP$ for ST-GC, VST-GC, FC-TC and STS-GC. We also introdce TSD-GC and DTSD-GC by switching order of spatial and temporal GC in STD-GC and DSTD-GC, respectively. Besides, we initialized adjacency matrices randomly for all GCs.
\subsubsection{Shared Correlation vs Unshared Correlation} As these two GCs are defined for decomposing spatiotemporal GC, we compare them with original ST-GC. We observe that correlation-shared GCs are more parameter-saving than correlation-unshared GCs but perform worse than ST-GC. On the other hand, the correlation-unshared GCs achieve similar errors as ST-GC, this suggests that unshared spatial and temporal correlations are more suitable forms for describing spatiotemporal correlations, because the unshared correlations can explicitly describe the varying spatiotemporal correlations in different body parts and different motion stages. Our DSTD-GC utilizes a shared correlation to store general correlation and a function to extract unshared correlation variances. Thus, our DSTD-GC is parameter-saving and explicitly models spatiotemporal correlations.
\subsubsection{Sample-shared Correlation vs Sample-specific Correlation} We compare the two sample-shared GCs (TSD-GCD and STD-GC) with their sample-specific counterparts (DTSD-GC and DSTD-GC). We find sample-specific GC make consistent improvement in accuracy. The reason is that the learnt sample-shared correlation can't precisely depict correlations in individual sequences, the correlation has to make a balance between various input samples. This global optimal strategy may introduce false correlation that disturb individual predictions. More studies about this is shown in Section \ref{sec:vis_cdc}.
\subsubsection{Spatial-temporal Order vs Temporal-spatial Order} As STD-GC and TSD-GC, DSTD-GC and DTSD-GC perform similarly, the modeling capability of spatiotemporal decomposed GC is invariant to the stacking order of spatial and temporal GCs.
\begin{comment}
we observe that (1) correlation-unshared methods outperform correlation-shared methods, and sample-specific methods outperform sample-shared methods, highlighting the importance of sample-specific unshared correlation modeling; (2) STD-GC, STS-GC and ST-GC perform similarly, but ST-GC has far more parameters than the other two GCs, demonstrating the critical role of spatiotemporal decomposition and the equivalence between STS-GC and STD-GC; (3) DSTD-GC outperforms STD-GC by 1.45 with 28.3\% parameters, proving that our constrained dynamic modeling strategy is effective to model spatiotemporal correlations in human motions. (4) STD-GC and TSD-GC, DSTD-GC and DTSD-GC perform similarly, indicating that the modeling capability of spatiotemporal decomposef GC is invariant to the stacking order of spatial and temporal GCs.
\end{comment}
\begin{table*}[!ht]
\linespread{1}
\setlength\tabcolsep{1.5pt}
\begin{center}
\caption{Comparisons of predictions on Human3.6M. The best results are highlighted in \textbf{bold}, while the second best results are shown in \underline{underline}.}
\label{tab:h36m_all}
\resizebox{\textwidth}{64mm}{
\begin{tabular}{c|cccccc|cccccc|cccccc|cccccc}
\hline
Action & \multicolumn{6}{c|}{Walking} & \multicolumn{6}{c|}{Eating} & \multicolumn{6}{c|}{Smoking} & \multicolumn{6}{c}{Discussion} \\
\hline
Millisecond & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 24.13 & 38.59 & 55.09 & 59.26 & 65.14 & 82.04 & 18.64 & 30.38 & 47.97 & 54.84 & 67.80 & 105.56 & 18.06 & 29.63 & 45.80 & 52.56 & 64.25 & 91.95 & 27.74 & 48.22 & 79.13 & 90.02 & 106.71 & 134.22 \\
DMGNN \cite{li2020dynamic} & 15.52 & 27.42 & 46.20 & 54.96 & 58.86 & 83.74 & 10.34 & 19.84 & 37.60 & 46.66 & 57.95 & 86.55 & 9.94 & 18.90 & 35.02 & 42.82 & 53.23 & 77.76 & 14.77 & 30.86 & 62.05 & 75.26 & 93.24 & 123.67 \\
FC-GCN \cite{mao2019learning} & 12.29 & 23.03 & 39.77 & 46.12 & 54.05 & \textbf{59.75} & 8.36 & 16.90 & 33.19 & 40.70 & 53.39 & 77.75 & 7.94 & 16.24 & 31.90 & 38.90 & 50.74 & 72.62 & 12.50 & 27.40 & 58.51 & 71.68 & 91.61 & 121.53 \\
Traj-CNN \cite{liu2020trajectorycnn} & \underline{11.91} & \underline{22.54} & 38.66 & 45.71 & 54.49 & 62.01 & 8.41 & \underline{16.56} & \underline{32.44} & \underline{39.82} & 53.47 & 78.40 & 8.41 & 16.17 & \underline{31.06} & \underline{37.58} & \underline{49.30} & 72.31 & 11.74 & \underline{26.30} & 57.29 & 70.36 & 91.45 & 122.66 \\
STS-GCN \cite{sofianos2021space} & 11.98 & 22.96 & 41.53 & 48.47 & 56.62 & 62.69 & \underline{7.90} & 16.79 & 33.38 & 40.74 & 53.10 & \underline{76.73} & \underline{7.48} & \underline{15.69} & 31.33 & 38.45 & 50.67 & 73.10 & \underline{11.39} & 26.41 & 57.78 & 71.08 & 91.16 & 120.79 \\
MSR-GCN \cite{dang2021msr} & 12.16 & 22.65 & \textbf{38.64} & \underline{45.24} & \underline{52.72} & 63.04 & 8.39 & 17.05 & 33.03 & 40.43 & \underline{52.54} & 77.11 & 8.02 & 16.27 & 31.32 & 38.15 & 49.45 & \underline{71.64} & 11.98 & 26.76 & \underline{57.08} & \underline{69.74} & \underline{88.59} & \underline{117.59} \\
\hline
Ours & \textbf{11.05} & \textbf{22.35} & \underline{38.81} & \textbf{45.19} & \textbf{52.70} & \underline{59.76} & \textbf{6.95} & \textbf{15.51} & \textbf{31.74} & \textbf{39.19} & \textbf{51.86} & \textbf{76.19} & \textbf{6.64} & \textbf{14.75} & \textbf{29.78} & \textbf{36.67} & \textbf{48.09} & \textbf{71.16} & \textbf{9.98} & \textbf{24.37} & \textbf{54.53} & \textbf{67.40} & \textbf{87.00} & \textbf{116.30} \\
\hline
\noalign{\smallskip}
\hline
Action & \multicolumn{6}{c|}{Directions} & \multicolumn{6}{c|}{Greetings} & \multicolumn{6}{c|}{Phoning} & \multicolumn{6}{c}{Posing} \\
\hline
Millisecond & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 21.93 & 37.38 & 61.51 & 71.86 & 88.09 & 122.08 & 35.28 & 61.98 & 99.06 & 111.04 & 127.44 & 160.59 & 21.05 & 35.49 & 57.89 & 67.26 & 83.74 & 130.34 & 32.99 & 59.71 & 104.87 & 123.67 & 158.74 & 300.18 \\
DMGNN \cite{li2020dynamic} & 10.77 & 22.81 & 46.70 & 57.35 & 73.51 & 104.37 & 20.45 & 41.13 & 78.99 & 94.34 & 114.68 & 148.75 & 12.35 & 24.25 & 47.22 & 58.27 & 73.62 & 113.05 & 15.40 & 32.07 & 67.36 & 84.19 & 113.60 & 171.56 \\
FC-GCN \cite{mao2019learning} & 8.97 & 19.87 & 43.35 & \underline{53.74} & 71.01 & 101.79 & 18.65 & 38.68 & 77.74 & 93.39 & 114.43 & 148.69 & 10.24 & 21.02 & 42.54 & 52.30 & 69.56 & 104.50 & 13.66 & 29.89 & 66.62 & 84.05 & 113.56 & 171.31 \\
Traj-CNN \cite{liu2020trajectorycnn} & 8.69 & 19.29 & 43.57 & 54.36 & 74.56 & 109.42 & 15.81 & 35.12 & 73.56 & \underline{88.89} & \underline{110.78} & 149.55 & 10.14 & 20.52 & 41.95 & 51.86 & 69.29 & 104.41 & 12.09 & \underline{26.94} & \underline{62.44} & \underline{79.33} & \underline{108.36} & 170.86 \\
STS-GCN \cite{sofianos2021space} & \underline{7.82} & \underline{18.72} & \underline{42.58} & 53.25 & 70.95 & 102.10 & \underline{15.33} & \underline{35.01} & \underline{73.44} & 89.08 & 112.19 & \underline{143.91} & \underline{9.54} & \underline{20.35} & 41.55 & \underline{51.07} & 68.32 & \underline{103.69} & \underline{11.61} & 27.60 & 63.85 & 81.23 & 111.68 & \underline{168.41} \\
MSR-GCN \cite{dang2021msr} & 8.61 & 19.65 & 43.28 & 53.82 & 71.18 & \underline{100.59} & 16.48 & 36.95 & 77.32 & 93.38 & 116.26 & 147.26 & 10.10 & 20.74 & \underline{41.51} & 51.26 & \underline{68.29} & 104.27 & 12.79 & 29.38 & 66.95 & 85.01 & 116.27 & 174.33 \\
\hline
Ours & \textbf{6.86} & \textbf{17.39} & \textbf{40.96} & \textbf{51.67} & \textbf{69.05} & \textbf{99.05} & \textbf{14.27} & \textbf{33.53} & \textbf{72.15} & \textbf{87.31} & \textbf{108.66} & \textbf{142.28} & \textbf{8.52} & \textbf{19.22} & \textbf{40.31} & \textbf{49.87} & \textbf{66.69} & \textbf{102.20} & \textbf{10.08} & \textbf{25.40} & \textbf{60.60} & \textbf{77.34} & \textbf{106.54} & \textbf{163.31} \\
\hline
\noalign{\smallskip}
\hline
Action & \multicolumn{6}{c|}{Purchases} & \multicolumn{6}{c|}{Sitting} & \multicolumn{6}{c|}{Sitting Down} & \multicolumn{6}{c}{Taking Photo} \\
\hline
Millisecond & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 30.54 & 52.89 & 86.31 & 99.07 & 121.00 & 169.66 & 22.08 & 38.58 & 64.59 & 75.94 & 96.33 & 145.90 & 27.73 & 47.24 & 79.55 & 93.33 & 117.92 & 170.19 & 21.53 & 37.30 & 63.83 & 75.78 & 95.70 & 145.82 \\
DMGNN \cite{li2020dynamic} & 17.60 & 36.24 & 69.43 & 83.20 & 104.09 & 145.26 & 11.77 & 23.42 & 48.33 & 60.69 & 80.75 & 123.31 & 16.93 & 32.04 & 62.19 & 76.52 & 101.16 & 153.00 & 11.49 & 23.07 & 46.89 & 58.46 & 78.20 & 121.69 \\
FC-GCN \cite{mao2019learning} & 15.60 & 32.78 & \underline{65.72} & \underline{79.25} & \underline{100.19} & 141.14 & 10.62 & 21.90 & 46.33 & 57.91 & 79.38 & 122.44 & 16.14 & 31.12 & 61.74 & 76.46 & 99.24 & 149.30 & 9.88 & 20.89 & 44.95 & 56.58 & 76.52 & 119.33 \\
Traj-CNN \cite{liu2020trajectorycnn} & 14.54 & 31.88 & 66.55 & 80.75 & 103.65 & 141.01 & 10.97 & 21.17 & 45.48 & 57.50 & 78.95 & 120.12 & 16.13 & \underline{29.56} & \underline{58.74} & \underline{72.59} & \underline{97.00} & \textbf{146.96} & 10.43 & 20.64 & 44.37 & 55.83 & 76.78 & 120.11 \\
STS-GCN \cite{sofianos2021space} & \underline{13.87} & \underline{31.66} & 66.00 & 80.04 & 102.46 & 142.46 & \underline{9.63} & \underline{20.65} & \underline{45.22} & \underline{57.26} & 78.96 & 122.03 & \underline{14.98} & 29.60 & 59.41 & 73.55 & 98.80 & 149.52 & \underline{9.15} & \underline{19.87} & \underline{43.42} & \underline{54.99} & \underline{76.15} & \underline{118.76} \\
MSR-GCN \cite{dang2021msr} & 14.75 & 32.39 & 66.13 & 79.64 & 101.63 & \underline{139.16} & 10.53 & 21.99 & 46.26 & 57.80 & \underline{78.20} & \underline{120.04} & 16.10 & 31.63 & 62.45 & 76.84 & 102.84 & 155.47 & 9.89 & 21.01 & 44.56 & 56.30 & 77.97 & 121.91\\
\hline
Ours & \textbf{12.68} & \textbf{29.65} & \textbf{62.29} & \textbf{75.79} & \textbf{97.54} & \textbf{137.76} & \textbf{8.78} & \textbf{19.32} & \textbf{42.88} & \textbf{54.33} & \textbf{74.94} & \textbf{117.75} & \textbf{14.10} & \textbf{28.03} & \textbf{57.33} & \textbf{71.18} & \textbf{96.08} & \underline{147.25} & \textbf{8.41} & \textbf{18.84} & \textbf{42.00} & \textbf{53.50} & \textbf{74.50} & \textbf{117.91} \\
\hline
\noalign{\smallskip}
\hline
Action & \multicolumn{6}{c|}{Waiting} & \multicolumn{6}{c|}{Walking Dog} & \multicolumn{6}{c|}{Walking Together} & \multicolumn{6}{c}{Average} \\
\hline
Millisecond & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 25.78 & 44.52 & 72.29 & 82.44 & 98.79 & 136.76 & 39.19 & 67.36 & 105.56 & 117.96 & 135.99 & 186.09 & 22.03 & 36.08 & 54.91 & 60.30 & 67.89 & 85.19 & 25.91 & 44.36 & 71.86 & 82.35 & 99.90 & 144.44 \\
DMGNN \cite{li2020dynamic} & 12.89 & 25.70 & 51.27 & 62.84 & 79.82 & 112.99 & 26.09 & 50.47 & 88.90 & 102.55 & 118.44 & 156.09 & 13.76 & 25.55 & 44.49 & 53.47 & 59.41 & 79.06 & 14.67 & 28.91 & 55.51 & 67.44 & 84.04 & 129.06 \\
FC-GCN \cite{mao2019learning} & 11.43 & 23.99 & 50.06 & 61.48 & 78.15 & 108.77 & 23.39 & 46.17 & 83.47 & 95.96 & \underline{110.98} & \textbf{146.24} & 10.47 & 21.04 & 38.47 & 45.19 & 54.71 & 66.96 & 12.68 & 26.06 & 52.29 & 63.58 & 81.17 & 114.14 \\
Traj-CNN \cite{liu2020trajectorycnn} & 10.51 & \underline{21.76} & \underline{45.79} & \underline{56.29} & \underline{73.36} & \underline{104.53} & 21.30 & 43.29 & 80.77 & 94.50 & 115.56 & 153.50 & 10.30 & 21.11 & 38.48 & 44.82 & 54.78 & 68.00 & 12.09 & 24.86 & 50.74 & 61.95 & 80.78 & 114.92 \\
STS-GCN \cite{sofianos2021space} & \underline{10.00} & 21.93 & 46.98 & 58.22 & 76.39 & 107.68 & 20.79 & 43.56 & 81.81 & 95.20 & 114.36 & 151.92 & \underline{10.06} & \underline{20.69} & 39.07 & 46.00 & 54.92 & \underline{62.91} & \underline{11.43} & \underline{24.77} & \underline{51.16} & \underline{62.58} & \underline{81.12} & \underline{113.78} \\
MSR-GCN \cite{dang2021msr} & 10.68 & 23.06 & 48.25 & 59.23 & 76.33 & 106.27 & 20.65 & 42.88 & \underline{80.35} & \underline{93.31} & 111.89 & 148.24 & 10.56 & 20.92 & \underline{37.40} & \underline{43.85} & \underline{52.94} & 65.94 & 12.11 & 25.56 & 51.64 & 62.93 & 81.14 & 114.19 \\
\hline
Ours & \textbf{8.71} & \textbf{20.15} & \textbf{44.28} & \textbf{55.25} & \textbf{73.19} & \textbf{105.66} & \textbf{19.64} & \textbf{41.82} & \textbf{77.61} & \textbf{90.24} & \textbf{109.84} & \underline{147.68} & \textbf{9.07} & \textbf{19.79} & \textbf{36.33} & \textbf{42.67} & \textbf{50.54} & \textbf{61.22} & \textbf{10.38} & \textbf{23.34} & \textbf{48.77} & \textbf{59.84} & \textbf{77.81} & \textbf{111.02} \\
\hline
\end{tabular}}
\end{center}
\end{table*}
\begin{comment}
\begin{table*}[!htp]
\linespread{1}
\setlength\tabcolsep{1pt}
\begin{center}
\caption{Comparisons of long-term predictions on Human3.6M. The best results are highlighted in \textbf{bold}, while the second best results are shown in \underline{underline}.}
\label{tab:cmu_long}
\resizebox{\textwidth}{10.5mm}{
\begin{tabular}{c|cc|cc|cc|cc|cc|cc}
\hline
Action & \multicolumn{2}{c|}{Walking} & \multicolumn{2}{c|}{Eating} & \multicolumn{2}{c|}{Smoking} & \multicolumn{2}{c|}{Discussion} & \multicolumn{2}{c|}{Directions} & \multicolumn{2}{c}{Average} \\
\hline
Millisecond (ms) & 560 & 1000 & 560 & 1000 & 560 & 1000 & 560 & 1000 & 560 & 1000 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 81.73 & 100.68 & 79.87 & 100.20 & 94.83 & 137.44 & 121.30 & 161.70 & 110.05 & 152.48 & 97.56 & 130.50 \\
DMGNN \cite{li2020dynamic} & 73.36 & 95.82 & 58.11 & 86.66 & 50.85 & 72.15 & 81.90 & 138.32 & 110.06 & 115.75 & 74.85 & 101.74 \\
FC-GCN \cite{mao2019learning} & 54.05 & 59.75 & 53.39 & 77.75 & 50.74 & 72.62 & 91.61 & 138.32 & 101.79 & 115.75 & 64.16 & 86.69 \\
Traj-CNN \cite{liu2020trajectorycnn} & 52.49 & 62.01 & 53.47 & 78.40 & 49.30 & 72.31 & 91.45 & 122.66 & 74.56 & 109.42 & 64.25 & 88.96\\
MSR-GCN \cite{dang2021msr} & 52.72 & 63.04 & 52.54 & 77.11 & 49.45 & 71.64 & 88.59 & 117.59 & 71.18 & 100.59 & 62.89 & 86.00 \\
STS-GCN \cite{sofianos2021space} & &&& &&&& &&& \\
Ours & 52.75 & 58.92 & 51.19 & 74.86 & 48.24 & 70.59 & 87.17 & 115.50 & 69.38 & 99.49 & 61.75 & 83.87 \\
\hline
\end{tabular}}
\end{center}
\end{table*}
\end{comment}
\begin{table*}[!htp]
\setlength\tabcolsep{1.5pt}
\centering
\caption{Comparisons of prediction results on CMU Mocap dataset. The best results are highlighted in \textbf{bold}, while the second best results are shown in \underline{underline}.}
\label{tab:cmu_all}
\resizebox{\textwidth}{32mm}{
\begin{tabular}{c|cccccc|cccccc|cccccc|cccccc}
\hline
Action & \multicolumn{6}{c|}{Basketball} & \multicolumn{6}{c|}{Basketball Signal} & \multicolumn{6}{c|}{Directing Traffic} & \multicolumn{6}{c}{Jumping} \\
\hline
Millisecond & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 29.50 & 53.05 & 91.22 & 106.03 & 128.74 & 157.38 & 14.63 & 22.07 & 39.07 & 46.56 & 59.98 & 89.93 & 21.77 & 38.78 & 70.45 & 85.3 & 110.29 & 165.13 & 30.18 & 53.02 & 89.35 & 103.9 & 125.55 & 160.49 \\
DMGNN \cite{li2020dynamic} & 14.97 & 27.07 & 49.36 & 61.45 & 84.83 & 145.18 & 4.91 & 8.78 & 15.93 & 19.60 & 27.43 & 47.28 & 10.62 & 19.63 & 36.03 & 44.34 & 62.87 & 115.71 & 19.47 & 36.77 & 67.42 & 81.28 & 105.62 & 148.53 \\
FC-GCN \cite{mao2019learning} & 11.67 & 21.09 & 40.70 & 50.58 & 68.03 & 95.66 & 3.35 & 6.23 & 13.48 & 17.87 & 27.34 & 51.88 & 6.78 & 13.36 & 29.57 & 39.06 & 59.64 & 112.83 & 17.10 & 32.06 & 59.82 & 72.51 & 94.33 & 127.20 \\
Traj-CNN \cite{liu2020trajectorycnn} & 11.84 & 19.12 & \underline{36.72} & 46.02 & 62.47 & 95.76 & 4.42 & 6.20 & 12.29 & 16.19 & 25.48 & 51.76 & 6.95 & \underline{11.03} & \underline{25.89} & 40.99 & 54.76 & 112.43 & \underline{14.88} & \underline{27.01} & \underline{55.31} & 71.72 & 94.23 & 126.97 \\
STS-GCN \cite{sofianos2021space} & \underline{10.23} & \underline{18.67} & 36.93 & \underline{45.98} & \underline{61.19} & 91.36 & \underline{2.96} & \underline{5.52} & \underline{12.12} & \underline{16.12} & \underline{25.15} & 50.88 & \underline{5.95} & 11.99 & 27.55 & \underline{36.75} & 57.05 & \underline{111.53} & 15.66 & 30.63 & 59.13 & 71.87 & 93.32 & \textbf{125.94}\\
MSR-GCN \cite{dang2021msr} & 10.28 & 18.94 & 37.68 & 47.03 & 62.01 & \textbf{86.27} & 3.04 & 5.63 & 12.51 & 16.61 & 25.46 & \underline{50.04} & 6.13 & 12.61 & 29.39 & 39.24 & \underline{50.49} & 114.58 & 15.19 & 28.86 & 55.98 & \underline{69.12} & \underline{92.40} & 126.18\\
\hline
Ours & \textbf{9.60} & \textbf{17.64} & \textbf{35.44} & \textbf{44.43} & \textbf{59.97} & \underline{88.44} & \textbf{2.57} & \textbf{4.72} & \textbf{10.37} & \textbf{13.86} & \textbf{21.85} & \textbf{46.17} & \textbf{5.02} & \textbf{10.01} & \textbf{23.35} & \textbf{31.40} & \textbf{49.28} & \textbf{99.57} & \textbf{12.81} & \textbf{26.05} & \textbf{54.62} & \textbf{68.47} & \textbf{91.83} & \underline{126.07} \\
\hline
\noalign{\smallskip}
\hline
Action & \multicolumn{6}{c|}{Soccer} & \multicolumn{6}{c|}{Walking} & \multicolumn{6}{c|}{Wash Window} & \multicolumn{6}{c}{Average} \\
\hline
Millisecond & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 & 80 & 160 & 320 & 400 & 560 & 1000 \\
\hline
Residual sup. \cite{martinez2017human} & 26.51 & 46.98 & 81.45 & 96.18 & 117.9 & 139.06 & 14.61 & 22.87 & 36.09 & 40.90 & 51.10 & 69.49 & 19.32 & 31.77 & 56.05 & 66.00 & 83.62 & 125.87 & 22.36 & 38.36 & 66.24 & 77.84 & 96.74 & 129.62 \\
DMGNN \cite{li2020dynamic} & 17.64 & 31.86 & 56.81 & 68.84 & 92.70 & 130.80 & 12.23 & 21.89 & 36.03 & 41.32 & 51.46 & 64.68 & 9.10 & 16.90 & 32.55 & 41.13 & 57.09 & 97.08 & 12.70 & 23.27 & 42.01 & 51.13 & 68.86 & 107.04 \\
FC-GCN \cite{mao2019learning} & 13.62 & 24.30 & 44.40 & 54.31 & 73.14 & 111.64 & 6.74 & 11.09 & 18.08 & 20.95 & 25.16 & \textbf{32.38} & 5.87 & 11.33 & 24.14 & 30.95 & 43.44 & \textbf{66.93} & 9.30 & 17.06 & 32.89 & 40.89 & 55.86 & 85.50 \\
Traj-CNN \cite{liu2020trajectorycnn} & 13.46 & 21.25 & 38.65 & 47.26 & \underline{62.66} & \underline{97.33} & 7.69 & 11.28 & 18.02 & 20.62 & 25.67 & 40.35 & 6.64 & 11.04 & 24.14 & 31.22 & 44.19 & 71.34 & 9.41 & 15.27 & 30.15 & 39.14 & 52.63 & 85.13 \\
STS-GCN \cite{sofianos2021space} & 11.30 & 20.45 & 39.04 & 48.88 & 69.12 & 102.54 & 6.87 & 11.29 & 18.13 & 21.06 & 26.12 & 37.86 & 5.44 & \underline{10.84} & 23.90 & \underline{30.72} & \underline{44.00} & 71.42 & 8.33 & 15.62 & 30.97 & 38.77 & 53.70 & 84.50 \\
MSR-GCN \cite{dang2021msr} & \underline{10.92} & \underline{19.39} & 37.41 & \underline{47.01} & 65.26 & 101.86 & \underline{6.39} & \textbf{10.25} & \underline{16.89} & \underline{20.05} & \underline{25.49} & 36.82 & \underline{5.41} & 10.94 & 24.51 & 31.80 & 45.14 & 70.19 & \underline{8.19} & \underline{15.20} & \underline{30.53} & \underline{38.65} & \underline{52.32} & \underline{83.70} \\
\hline
Ours & \textbf{10.25} & \textbf{18.96} & \textbf{36.79} & \textbf{45.65} & \textbf{62.29} & \textbf{96.93} & \textbf{6.34} & \underline{10.35} & \textbf{16.09} & \textbf{18.62} & \textbf{23.28} & \underline{33.56} & \textbf{4.75} & \textbf{9.53} & \textbf{21.98} & \textbf{28.99} & \textbf{42.48} & \underline{68.93} & \textbf{7.33} & \textbf{13.90} & \textbf{28.37} & \textbf{35.91} & \textbf{50.11} & \textbf{79.95} \\
\hline
\end{tabular}}
\end{table*}
\begin{comment}
\begin{table*}[!htp]
\setlength\tabcolsep{1pt}
\centering
\caption{Comparison summary of average MPJPE and parameter numbers}
\label{tab:summary_all}
\begin{tabular}{c|cccccc|c|cccccc|c}
\hline
\multirow{3}{*}{Model} & \multicolumn{7}{c|}{Human3.6M} & \multicolumn{7}{c}{CMU Mocap} \\
\cline{2-15}
\multicolumn{1}{c|}{} & \multicolumn{6}{c|}{Average MPJPE} & \multirow{2}{*}{Params.} & \multicolumn{6}{c|}{Average MPJPE} & \multirow{2}{*}{Params.} \\
\multicolumn{1}{c|}{} & 80 & 160 & 320 & 400 & 560 & 1000 & \multicolumn{1}{c|}{} & 80 & 160 & 320 & 400 & 560 & 1000 & \multicolumn{1}{c}{} \\
\hline
Residual sup. \cite{martinez2017human} & 25.91 & 44.36 & 71.86 & 82.35 & 99.90 & 144.44 & 3.42M & 22.50 & 38.56 & 65.74 & 76.54 & 93.80 & 123.19 & 3.46M \\
DMGNN \cite{li2020dynamic} & 14.67 & 28.91 & 55.51 & 67.44 & 84.04 & 129.06 & 46.93M & 13.31 & 24.12 & 42.48 & 50.91 & 66.96 & 103.86 & 46.94M \\
FC-GCN \cite{mao2019learning} & 12.68 & 26.06 & 52.29 & 63.58 & 81.17 & 114.14 & 2.55M & 10.02 & 18.08 & 33.66 & 41.07 & 54.29 & 81.51 & 2.70M \\
Traj-CNN \cite{liu2020trajectorycnn} & 12.09 & 24.86 & 50.74 & 61.95 & 80.78 & 114.92 & 2.42M & 10.03 & 16.29 & 31.07 & 39.19 & 51.14 & 78.69 & 2.42M \\
STS-GCN \cite{sofianos2021space} & \underline{11.43} & \underline{24.77} & \underline{51.16} & \underline{62.58} & \underline{81.12} & \underline{113.78} & \underline{0.39M} & 8.95 & 16.40 & 31.35 & 38.64 & 52.11 & 80.27 & \underline{0.46M} \\
MSR-GCN \cite{dang2021msr} & 12.11 & 25.56 & 51.64 & 62.93 & 81.14 & 114.19 & 6.30M & \underline{8.81} & \underline{15.91} & \underline{30.45} & \underline{37.99} & \underline{50.54} & \underline{76.69} & 6.37M \\
\hline
Ours & \textbf{10.38} & \textbf{23.34} & \textbf{48.77} & \textbf{59.84} & \textbf{77.81} & \textbf{111.02} & \textbf{0.18M} & \textbf{7.92} & \textbf{14.78} & \textbf{29.04} & \textbf{36.09} & \textbf{48.85} & \textbf{76.39} & \textbf{0.20M} \\
\hline
\end{tabular}
\end{table*}
\end{comment}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\textwidth]{figures/result_comparison_cmu.pdf}
\caption{Visualization of predicted results of state-of-the-art methods on two action examples on CMU-Mocap dataset. The blue and red poses indicate ground truths and predictions respectively. (a) Basketball. (b) Directing Traffic.}
\label{fig:skeleton_comparison}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[width=1\textwidth]{figures/heatmap.pdf}
\caption{Visualization of spatial and temporal correlation. Color brightness indicates correlation strength. (a) Spatial correlation from a predicted human pose frame. (b) Temporal correlation from a joint trajectory. (I) Constrained correlation. (II) and (III) are sample-specific correlation adjustments of a pose frame or a joint trajectory for two different action sequences.}
\label{fig:heatmap_comparison}
\end{figure*}
\subsection{RQ2: Comparison with the State-of-the-art}
To evaluate the performance of DSTD-GCN, we present the quantitative and qualitative results on Human3.6M and CMU Mocap datasets. We first introduce the compared baselines and then DSTD-GCN's performance. Following literature from previous works \cite{mao2019learning,dang2021msr}, we divide the results into short-term ($<500$ msecs) and long-term ($>500$ msecs) predictions.
\subsubsection{Baselines} We select six state-of-the-art baselines, including RNN-based model (Residual sup \cite{martinez2017human}), CNN-based model (Traj-CNN \cite{liu2020trajectorycnn}) and GCN-based model (FC-GCN \cite{mao2019learning}, DMGNN \cite{li2020dynamic} and MSR-GCN \cite{dang2021msr} and STS-GCN \cite{sofianos2021space}). For GCN-based models, FC-GCN, DMGNN, and MSR-GCN belong to the sample-shared correlation-shared method, and STS-GCN\footnote{We utilized the model from RQ1} is the only sample-shared correlation-unshared method. For the graph correlations, DMGNN adopts fixed correlations from body connections, while the other three GCN-based methods initialize the adjacency matrix randomly and then optimize it with back-propagation. Among all the baseline methods, FC-GCN, Traj-CNN, MSR-GCN, STS-GCN, and our approaches utilize the prediction framework depicted in Figure \ref{fig:details} (a).
\subsubsection{Results} The detail quantitative comparisons for both short-term and long-term prediction results are presented in Table \ref{tab:h36m_all} and Table \ref{tab:cmu_all}. Besides, we also present the average results of selected frames in Table \ref{tab:compare_summary}. Apparently, methods with the prediction framework make more accurate predictions with fewer parameters. Moreover, GCN-based methods generally outperform the RNN-based method (Residual Sup.) and CNN-based methods (Traj-CNN). For GCN-based methods, Correlation-unshared methods outperforms correlation-unshared methods in terms of prediction accuracy and parameter numbers. Note that MSR-GCN contains 65 GC layers and combines DCT temporal encoding \cite{mao2019learning} with multi-scale modeling, while STS-GCN and DSTD-GCN merely contain 7 GC layers with raw 3D inputs. With Constrained Dynamic Correlation Modeling, DSTD-GCN also models sample-specific correlations and achieves best prediction accuracy with fewest parameters. For qualitative comparison, we show two examples of the predicted poses for different methods in Figure \ref{fig:skeleton_comparison}. In the basketball scenario, correlation-shared methods generally work better than correlation-shared ones as correlation-unshared methods explicitly model individual joint's spatiotemporal correlations and thus can infer large movement from tiny motion cues. Here, both STS-GCN and our method capture the backward trend of the right feet. However, sample-shared modeling spatiotemporal correlations may disturb individual motion prediction, especially in some static action scenarios. As we see in directing traffic scenario, most joints stay still but STS-GCN amplifies the motion of the right arm and infers feet movement mistakenly. With the Constrained Dynamic Correlation Modeling, our method can dynamically model sample-specific motion pattern and thus make more accurate predictions under different scenarios.
\subsubsection{Visualization of Constrained Dynamic Correlations}\label{sec:vis_cdc} To better understand why our method takes effect, we visualize the sample spatial and temporal correlations from the two visualization sequences in Figure \ref{fig:heatmap_comparison}. We obtain two observations: (1) The constrained correlation strength is larger than the adjusted one, demonstrating that the basics of motion prediction come from shared constrained correlation. (2) Sample-specific correlation adaptively extracts adjustments which contribute to accurate motion prediction. As shown in Figure \ref{fig:heatmap_comparison} (a) (II) and (III), the adjustment strength from the right arm to the feet (green box) is enhanced in basketball and reduced in directing traffic. These adjustments are reasonable because arms and feet collaborate intensively in the basketball example but feet are still and unrelated to the movement of the right arm in the directing traffic case. Moreover, we also present temporal correlation adjustments from a right foot joint in Figure \ref{fig:heatmap_comparison} (b), we find the adjustment strengths are generally stronger in the basketball case. As the right foot moves backward in the basketball example, it correlates to other vertices of the trajectory more intensively. With this visualization, we observe that our proposed Constrained Dynamic Correlation Modeling can generate sample-specific correlation and thus make more accurate predictions.
\subsubsection{Effectiveness Analysis} We present model parameters and inference time of single epoch in Table \ref{tab:compare_summary}. Generally, correlation-unshared methods outperform other methods in prediction accuracy and parameters. Our method outperforms other methods in prediction accuracy with fewest parameters. The parameter number of our method is half of the most lightweight model STS-GCN and only $3\%$ of the best baseline MSR-GCN on the CMU Mocap dataset. Besides, our method is slightly slower than STS-GCN due to the additional computation of unshared correlation adjustments.
\begin{comment}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\textwidth]{figures/plt_compare.pdf}
\caption{Comparisons Summaries on the Human3.6M and the CMU-Mocap dataset. (a) and (b) show average prediction error at different forecast times. (c) and (d) show average prediction error and model parameters.}
\label{fig:plot_comparison}
\end{figure*}
\end{comment}
\begin{comment}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/plt_compare_small.pdf}
\caption{Average prediction loss and parameter number of different methods (a) Human3.6M. (b) CMU Mocap.}
\label{fig:plot_comparison}
\end{figure}
\end{comment}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.48\textwidth]{figures/cmu_ablation.png}
\caption{Comparison of Unseen Action Prediction. On each of the seven actions in the CMU Mocap dataset, we train STS-GCN \cite{sofianos2021space}, MSR-GCN \cite{dang2021msr}, and our model. Then, we test these models on all actions. Our method outperforms STS-GCN and MSR-GCN with lower MPJPE.}
\label{fig:single_ablation}
\end{figure}
\begin{table}[!ht]
\setlength\tabcolsep{1pt}
\centering
\caption{Comparison summary of average MPJPE, parameter numbers and inference time per epoch}
\label{tab:compare_summary}
\begin{tabular}{c|ccc|ccc}
\hline
\multirow{2}{*}{Model} & \multicolumn{3}{c}{Human3.6M} & \multicolumn{3}{|c}{CMU Mocap} \\
\cline{2-7}
& MPJPE & Params. & Inference & MPJPE & Params. & Inference \\
\hline
DMGNN \cite{li2020dynamic} & 63.27 & 46.90M & 422ms & 50.84 & 46.94M & 590ms \\
FC-GCN \cite{mao2019learning} & 58.32 & 2.55M & 46ms & 40.30 & 2.70M & 47ms \\
Traj-CNN \cite{liu2020trajectorycnn} & 57.56 & 1.20M & 95ms & 38.62 & 1.20M & 110ms \\
STS-GCN \cite{sofianos2021space} & 57.47 & 0.40M & \textbf{39ms} & 38.64 & 0.46M & \textbf{41ms} \\
MSR-GCN \cite{dang2021msr} & 57.88 & 6.30M & 70ms & 38.10 & 6.37M & 82ms \\
\hline
Ours & \textbf{55.19} & \textbf{0.18M} & 48ms & \textbf{35.93} & \textbf{0.20M} & 50ms \\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\setlength\tabcolsep{1pt}
\centering
\caption{Comparison of Ablation Variants}
\label{tab:ablation}
\begin{tabular}{c|ccccccc}
\hline
\multirow{2}{*}{Model} & \multicolumn{7}{c}{MPJPE} \\
& 80 & 160 & 320 & 400 & 560 & 1000 & Average \\
\hline
DSTD-C & 7.84 & 15.02 & 30.81 & 38.90 & 54.26 & 88.74 & 39.26 \\
DSTD-D & 10.68 & 19.34 & 35.69 & 43.53 & 57.56 & 89.81 & 42.77 \\
DSTD-R & 8.62 & 16.37 & 33.15 & 41.90 & 58.55 & 90.69 & 41.55 \\
DSTD-P & 7.58 & 14.25 & 29.02 & 36.73 & 51.35 & 81.81 & 36.79 \\
DSTD-F & 7.99 & 14.97 & 30.23 & 37.53 & 52.96 & 84.23 & 37.99 \\
DSTD-S & 8.89 & 16.24 & 32.48 & 40.87 & 56.85 & 91.38 & 41.12 \\
DSTD-T & 7.68 & 15.07 & 31.47 & 39.78 & 55.13 & 87.14 & 39.38 \\
DSTD & \textbf{7.33} & \textbf{13.90} & \textbf{28.37} & \textbf{35.91} & \textbf{50.11} & \textbf{79.95} & \textbf{35.93} \\
\hline
\end{tabular}
\end{table}
\begin{comment}
\begin{table}[!ht]
\setlength\tabcolsep{1pt}
\centering
\caption{Comparison of Ablation Variants}
\label{tab:ablation}
\begin{tabular}{c|c|c|c|c|c|c|c}
\hline
Model & DSTD-C & DSTD-D & DSTD-R & DSTD-P & DSTD-S & DSTD-T & DSTD \\
\hline
MPJPE & 40.07 & 43.48 & 47.78 & 36.26 & 49.91 & 47.61 & \textbf{35.93} \\
\hline
\end{tabular}
\end{table}
\end{comment}
\subsection{RQ3: Ablation Study}
We investigate some major components of our method in this section by comparing DSTD-GCN to its variants and other state-of-the-art methods. We propose eight ablation variants: (1) DSTD-C: DSTD-GCN only with constrained correlation update in DG-GC and DT-GC; (2) DSTD-D: DSTD-GCN only with dynamic correlation adjustments in DG-GC and DT-GC. (3) DSTD-R: constrained correlation is used to update dynamic correlation adjustments. (4) DSTD-P: DSTD-GCN without any predefined correlations. (6) DSTD-F: replacing DS-GCs and DT-GCs by S-GCs and T-GCs, respectively. (7) DSTD-S: replacing DT-GCs by DS-GCs. (8) DSTD-T: all DS-GCs are replaced by DT-GCs. The detail results are shown in Table \ref{tab:ablation}.
\subsubsection{Constrained Dynamic Correlation Modeling} We investigate our proposed constrained dynamic modeling strategy by comparing DSTD-C, DSTD-D, DSTD-R and DSTD. First, we study the relationships between the constrained correlation and the dynamic correlation. Comparing the prediction error of DSTD-C and DSTD-D: we observe that DSTD-C outperforms DSTD-D by 3.51, indicating that constrained correlations are more important for accurate motion prediction. Then, we study the updating strategy studying DSTD-R. DSTD-R performs worse than DSTD by 5.64, indicating that constrained correlation adjustment disturbs dynamic correlation. So dynamic correlations adjust constrained correlations but not vice versa.
\subsubsection{Prior Connections} We study the effects of prior connections in constrained correlations with DSTD-F, DSTD-P and DSTD. Comparing the improvement in average MPJPE for prior connection (DSTD-P and DSTD) and Constrained Dynamic Correlation Modeling (DSTD-F and DSTD), we find the reduction of prior connections (0.86) is much less than the reduction of Constrained Dynamic Correlation Modeling (2.06). Note that other state-of-the-art baselines don't get prior connections, DSTD-P still outperforms the best baseline MSR-GCN by 1.31 in average MPJPE, indicating the effectiveness of our proposed Constrained Dynamic Correlation Modeling.
\subsubsection{Spatial and Temporal GCs} To investigate the contribution of spatial and temporal GCs, we compare DSTD-S, DSTD-T, and DSTD. We observe that (1) DSTD outperforms DSTD-S and DSTD-T by 13.98 and 11.68, respectively, indicating that individual spatial or temporal GCs cannot make accurate prediction results and emphasizing the importance of spatiotemporal correlation modeling; and (2) DSTD-T outperforms DSTD-S by 2.30, demonstrating that temporal correlations are more important than spatial correlations for human motion prediction.
\subsubsection{Unseen Action Prediction} To further understand the effectiveness of our proposed model, we conduct experiments on unseen action predictions. Specifically, we train the model on one action and test on all actions. Figure \ref{fig:single_ablation} shows the comparison results. Our method outperforms two state-of-the-art models by a large margin. It illustrates that our DSTD-GCN is capable of generating sample-specific spatiotemporal correlations that may be used to improve human motion prediction in response to the individualized motions associated with various actions.
\subsection{Limitations} We only design a multi-layer perceptron based on pair-wise concatenation to extract temporal/spatial-wise spatial/temporal correlation adjustments. This exhaustive correlation modeling strategy calculates correlations between some unrelated vertices. There is space to design a more powerful and efficient correlation modeling strategy. Another limitation is the accuracy of unseen action prediction. The unseen action prediction error is 76.85, while the supervised prediction error is 35.93. There is space to explore methods for unseen action predictions. One possible method is to combine constrained dynamic modeling with few-shot learning \cite{gui2018few}.
\section{Conclusions}
In this work, we propose a novel Dynamic SpatioTemporal Decompose Graph Convolutions (DSTD-GC) for human motion prediction. DSTD-GC deploys the Constrained Dynamic Correlation Modeling, which extends conventional graph convolutions by combining constrained correlations from training or prior knowledge and dynamic correlations from input motion sequences. We make mathematical analysis and conduct extensive experiments to illustrate the power spatiotemporal modeling capability of DSTD-GC, which break certain constraints of state-of-the-art graph convolutions on spatiotemporal graphs. With this strategy, we propose DSTD-GCN which outperforms other state-of-the-art methods in prediction accuracy with the fewest parameters.
\section*{Acknowledgments}
This work was supported partly by the National Natural Science Foundation of China (Grant No. 62173045, 61673192), and partly by the Fundamental Research Funds for the Central Universities (Grant No. 2020XD-A04-2).
|
\section{Introduction}
Investigating Hausdorff dimensions of sets of points approximable at certain ``speed rate'' by a given sequence of points $(x_n)_{n\in\mathbb{N}}$ is an important topic in Diophantine approximation (see \cite{BV} and \cite{BS2} among other references), in dynamical systems \cite{HV,LS,PR} and in multifractal analysis \cite{Ja,BS3,BS4}. These studies consists in general, knowing that $\mu(\limsup_{n\rightarrow+\infty}B_n:=B(x_n ,r_n))=1$ for a certain measure $\mu$ and a sequence of radius $(r_n)_{n\in\mathbb{N}}$, in investigating the Hausdorff dimension of $\limsup_{n\rightarrow+\infty}U_n$ where $U_n\subset B_n $. Typically $U_n$ is a contracted ball inside $B_n$, but recently, general sets $U_n$ have been considered \cite{KR,rect,ED1}.The so-called ubiquity theorems or mass transference principles mainly focus on finding a lower bound, using an adequate measure $\mu$, for the Hausdorff dimension (or Hausdorff measure) of those sets. However it is key in many situations to understand whether this lower bound is optimal or not. This article is dedicated to this problem.
\sk
In order to do so, we introduce, given a probability measure $\mu$ on $\mathbb{R}^d$, the concept of $\mu$-asymptotically covering sequence of balls. This notion is a generalization of a covering property used in the $KGB$ Lemma stated in \cite{BV}.
For a sequence $(B_n)_{n\in\mathbb{N}}$, verifying this condition will be proved to be almost equivalent to verify $\mu(\limsup_{n \rightarrow+\infty}B_n)=1 $ (it is equivalent if the measure is doubling for instance, so that working under this settings is very reasonable).
As said above, given a sequence of balls $(B_n)_{n\in\mathbb{N}}$ and another $(U_n \subset B_n)_{n\in\mathbb{N}}$, ubiquity theorems or mass transference principles give lower bounds for the Hausdorff dimension of $\limsup_{n\rightarrow+\infty}U_n$ when, roughly speaking, some information is known about the geometry of $ \limsup_{n\rightarrow+\infty}B_n.$ Of course, there is no reason in general for a lower-bound for $\dim_H (\limsup_{n\rightarrow+\infty}U_n)$ obtained only knowing that $\mu(\limsup_{n\rightarrow+\infty}B_n )=1$, to be sharp (i.e $=\dim_H (\limsup_{n\rightarrow+\infty}U_n)$). If one hopes such a lower-bound to be accurate, the measure $\mu$ has to be particularly adapted, in some sense, to the pair of sequences $((B_n)_{n\in\mathbb{N}},(U_n)_{n\in\mathbb{N}}).$ The approach adopted in this article is to extract some sub-sequences from $(B_n)_{n\in\mathbb{N}}$ which are still $\mu$-a.c (or still verifies $\mu(\limsup_{n\rightarrow+\infty}B_n)=1$) but are adapted to the measure $\mu$. Applying mass transference principles (which are proved only for measures presenting enough self-similarity) to those sub-sequences, it will be proved that Hausdorff dimension of the limsup set associated with the corresponding $U_n$'s is given by the lower-bounds found in \cite{ED3} (so that it is also the case for the lower-bounds given in \cite{ED1, BS2, rect}).
This shows that those lower-bounds are sharp in a strong sens: for any self-similar measure $\mu$, any $\mu$-a.c sequence $(B_n)_{n\in\mathbb{N}}$, if one only considers the balls that are relevant for the measure $\mu$, the limsup set obtained by considering the sub-sequence of the corresponding $U_n$'s ($(U_n)_{n\in\mathbb{N}}$ being also the sequence of sets involved in the articles mentioned) has the expected dimension.
More precisely, it will be proven first that, under those very weak condition over a $\mu$-a.c sequence of balls $(B_n)_{n\in \mathbb{N}}$, it is always possible to extract a sub-sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$, still $\mu$-ac, "weakly redundant" (see Definition \ref{wr}) and such that the balls $(B_{\phi(n)})_{n\in\mathbb{N}}$ have prescribed behavior with respect to the measure $\mu$, roughly meaning (see Theorem \ref{theoremextra}) that the balls $(B_{\phi(n)})_{n\in\mathbb{N}}$ satisfies
\begin{equation}
\label{roughcondi}
\vert B_{\phi(n)} \vert^{\overline{\dim}_P (\mu)} \lessapprox \mu(B_{\phi(n)})\lessapprox\vert B_{\phi(n)} \vert^{\underline{\dim}_H (\mu)}.
\end{equation}
In a second time, it will be proved that, for weakly redundant sequences satisfying \eqref{roughcondi}, the Hausdorff dimension of $\limsup$ set associated with any sequence of shrunk balls or very thin rectangles $(R_n \subset B_n)_{n\in\mathbb{N}}$ (see Theorem \ref{rectssmajo}) can be bounded by above precisely by the lower-bound given in \cite{ED3}, which proves the optimality of those bounds.
\section{Notation and definition}
Let $d$ $\in\mathbb{N}$. \smallskip
For $n\in\mathbb{N}$, the set of dyadic cubes of generation $n$ of $\mathbb{R}^d$ is denoted $\mathcal{D}_n (\mathbb{R}^d)$ and defined as $\mathcal{D}_n (\mathbb{R}^d) = \left\{\prod_{i=1}^d [\frac{k_i}{2^{n}},\frac{k_{i}+1}{2^{n}})\right\}_{(k_i)_{i\in\mathbb{Z}}\in\mathbb{Z}^{\mathbb{Z}}}.$
For $x\in\mathbb{R}^{d}$, $r>0$, $B(x,r)$ stands for the closed ball of ($\mathbb{R}^{d}$,$\vert\vert \ \ \vert\vert_{\infty}$) of center $x$ and radius $r$.
Given a ball $B$, $\vert B\vert$ is the diameter of $B$.
For $t\geq 0$, $\delta\in\mathbb{R}$ and $B:=B(x,r)$, $t B$ stand for $B(x,t r)$, i.e. the ball with same center as $B$ and radius multiplied by $t$, and the $\delta$-contracted $B^{\delta}$ is defined by $B^{\delta}=B(x ,r^{\delta})$.
\smallskip
Given a set $E\subset \mathbb{R}^d$, $\overset{\circ}{E}$ stands for the interior of the $E$, $\overline{E}$ its closure and $\partial E$ is the boundary of $E$, i.e, $\partial E =\overline{E}\setminus \overset{\circ}{E}.$
\smallskip
The $\sigma$-algebra of Borel sets of $\mathbb{R}^d$ is denoted by $\mathcal{B}(\mathbb{R}^d)$,
$\mathcal{L}^d$ is the Lebesgue measure on $\mathcal{B}(\mathbb{R}^d)$ and
$\mathcal{M}(\mathbb{R}^d)$ stands for the set of Borel probability measure over $\mathbb{R}^d.$
\smallskip
For $\mu \in\mathcal{M}(\mathbb{R}^d)$, $\supp(\mu):=\left\{x: \ \forall r>0, \ \mu(B(x,r))>0\right\}$ is the topological support of $\mu.$
\smallskip
Given $E\in\mathcal{B}(\mathbb{R}^d)$, $\dim_{H}(E)$ and $\dim_{P}(E)$ denote respectively the Hausdorff and the packing dimension of $E$.
\smallskip
\subsection{Definition and recalls}
\begin{definition}
\label{hausgau}
Let $\zeta :\mathbb{R}^{+}\mapsto\mathbb{R}^+$ be an increasing mapping verifying $\zeta (0)=0$. The Hausdorff measure at scale $t\in(0,+\infty)$ associated with $\zeta$ of a set $E$ is defined by
\begin{equation}
\label{gaug}
\mathcal{H}^{\zeta}_t (E)=\inf \left\{\sum_{n\in\mathbb{N}}\zeta (\vert B_n\vert) : \ (B_n)_{n\in \N} \mbox{ closed balls, }\vert B_n \vert \leq t \text{ and } E\subset \bigcup_{n\in \mathbb{N}}B_n\right\}.
\end{equation}
The Hausdorff measure associated with $\zeta$ of a set $E$ is defined by
\begin{equation}
\mathcal{H}^{\zeta} (E)=\lim_{t\to 0^+}\mathcal{H}^{\zeta}_t (E).
\end{equation}
\end{definition}
For $t\in (0,+\infty)$, $s\geq 0$ and $\zeta:x\to x^s$, one simply uses the usual notation $\mathcal{H}^{\zeta}_t (E)=\mathcal{H}^{s}_t (E)$ and $\mathcal{H}^{\zeta} (E)=\mathcal{H}^{s} (E).$ In particular, the $s$-dimensional Hausdorff outer measure at scale $t\in(0,+\infty]$ of the set $E$ is defined by
\begin{equation}
\label{hcont}
\mathcal{H}^{s}_{t}(E)=\inf \left\{\sum_{n\in\mathbb{N}}\vert B_n\vert^s : \ (B_n)_{n\in \N} \mbox{ closed balls, }\vert B_n \vert \leq t \text{ and } E\subset \bigcup_{n\in \mathbb{N}}B_n\right\}.
\end{equation}
\begin{definition}
\label{dim}
Let $\mu\in\mathcal{M}(\mathbb{R}^d)$.
For $x\in \supp(\mu)$, the lower and upper local dimensions of $\mu$ at $x$ are
\begin{align*}
&\underline\dim (\mu,x)=\liminf_{r\rightarrow 0^{+}}\frac{\log\mu(B(x,r))}{\log r} \\
\mbox{ and } \ \ \ \ & \overline\dim (\mu,x)=\limsup_{r\rightarrow 0^{+}}\frac{\log \mu(B(x,r))}{\log r}.
\end{align*}
Then, the lower and upper dimensions of $\mu$ are defined by
\begin{equation}
\label{dimmu}
\dimm(\mu)=\mbox{infess}_{\mu}(\underline\dim (\mu,x)) \ \ \mbox{ and } \ \ \overline{\dim}_P (\mu)=\mbox{supess}_{\mu}(\overline\dim (\mu,x)).
\end{equation}
\end{definition}
It is known that (for more details see \cite{F})
$$\dimm(\mu)=\inf_{E\in\mathcal{B}(\mathbb{R}^d):\ \mu(E)>0}\dim_{H}(E) \ \ \ \mbox{ and } \ \ \overline{\dim}_P (\mu)=\inf_{E\in\mathcal{B}(\mathbb{R}^d):\ \mu(E)=1}\dim_{P}(E).$$
A measure verifying $\dimm(\mu)=\overline{\dim}_P (\mu):=\alpha$ will be called an $\alpha$ exact dimensional measure. From Definition \ref{dim}, such measures verify, for $\mu$-almost every $x\in \mathbb{R}^d$, $\lim_{r\rightarrow 0^{+}}\frac{\log \mu(B(x,r))}{\log r}=\alpha.$
\subsection{Main statements}
Before stating the Theorems proved in this article, one starts by recalling the following definition, introduced in \cite{BS}.
\begin{definition}
\label{wr}
Let $\mathcal{B}=( B _n =:B(x_n ,r_n))_{n\in\mathbb{N}}$ be a family of balls in $\mathbb{R}^d$. Denote by $\mathcal{T}_k(\mathcal{B})=\left\{B_n \ : \ 2^{-k-1}< r_{n}\leq 2^{-k}\right\}.$
The family $\mathcal{B}$ is said to be weakly redundant when for all $k$, there exists an integer $J_k$ and $\mathcal{T}_{k,1}(\mathcal{B}),..,\mathcal{T}_{k,J_k}(\mathcal{B})$ a partition of $\mathcal{T}_k(\mathcal{B})$ such that:\medskip
\begin{itemize}
\item[\textbf{$(C_1)$}] $\mathcal{T}_k(\mathcal{B})=\bigcup_{1\leq j\leq J_k} T_{k,j}(\mathcal{B}),$\medskip
\item[\textbf{$(C_2)$}] For every $1\leq j\leq J_k$ and every pair of balls $B\neq B^{\prime}\in \mathcal{T}_{k,j}(\mathcal{B})$,
$B\cap B^{\prime}=\emptyset,$\medskip
\item[\textbf{$(C_3 )$}] $\lim_{k\rightarrow +\infty}\frac{\log_{2}(J_k)}{k}=0.$
\end{itemize}
\end{definition}
So, a sequence of balls $(B_n)_{n\in\mathbb{N}}$ is weakly redundant when at each scale $2^{-k}$, the balls of the family $\left\{B_n \right\}_{n\in\mathbb{N}}$ that have radii $\approx 2^{-k}$ can be sorted in a relatively small number of families of pairwise disjoint balls.
The main property we introduce for a sequence of balls $\mathcal{B}=(B_n)_{n\in\mathbb{N}}$ is meant to ensure that any set can be covered efficiently by the limsup of the $B_n$'s, with respect to a measure $\mu$. This property is a general version of the KGB Lemma of Beresnevitch and Velani, stated in \cite{BV}, using a Borel probability measure $\mu$. Such properties (like the KGB Lemma) are usually key (cf \cite{Ja,BV,BS} for instance) to prove ubiquity or mass transference results.
\begin{definition}
\label{ac}
Let $\mu\in \mathcal{M}(\mathbb{R}^d)$. The sequence $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ of balls of $\mathbb{R}^d$ is said to be $\mu$-asymptotically covering (in short, $\mu$-a.c) when there exists a constant $C>0$ such that for every open set $\Omega\subset \mathbb{R}^d $ and $g\in\mathbb{N}$, there is an integer $N_\Omega \in\mathbb{N}$ as well as $g\leq n_1 \leq ...\leq n_{N_\Omega}$ such that:
\begin{itemize}
\item $\forall \, 1\leq i\leq N_\Omega$, $B_{n_i}\subset \Omega,$\medskip
\item $\forall \, 1\leq i\neq j\leq N_\Omega$, $B_{n_i}\cap B_{n_j}=\emptyset,$\medskip
\item one has
\begin{equation}
\label{majac}
\mu\left(\bigcup_{1\leq i\leq N_\Omega}B_{n_i} \right)\geq C\mu(\Omega).
\end{equation}
\end{itemize}
\end{definition}
In other words, for any open set $\Omega$ and any $g>0$, there exists a finite set of disjoint balls of $\left\{B_n\right\}_{n\geq g}$ covering a large part of $\Omega$ from the $\mu$-standpoint.
This notion of $\mu$-asymptotically covering is related to the way the balls of $\mathcal{B}$ are distributed according to the measure $\mu$. In particular, given a measure $\mu$, this property is slightly stronger than being of $\limsup$ of full $\mu$-measure, as illustrated by the following Theorem.
\begin{theoreme}
\label{equiac}
Let $\mu\in\mathcal{M}(\mathbb{R}^d)$ and $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ be a sequence of balls of $\mathbb{R}^d$ with $\lim_{n\to +\infty} r_{n}= 0$.
\begin{enumerate}
\smallskip
\item
If $\mathcal{B} $ is $\mu$-a.c, then $\mu(\limsup_{n\rightarrow+\infty}B_n)=1.$
\smallskip
\item
If there exists $v<1$ such that $ \mu \big(\limsup_{n\rightarrow+\infty}(v B_n) \big)=1$, then $\mathcal{B} $ is $\mu$-a.c.
\end{enumerate}
\end{theoreme}
Moreover, it results from the proof of the KGB-Lemma \cite{BV} that if the $\mu$ is doubling, $\mu\Big(\limsup_{n\rightarrow+\infty}B_n \Big)=1 \Leftrightarrow (B_n)_{n \in\mathbb{N}}$ is $\mu$-a.c.
\sk
One now states the main result about extraction of sub-sequences of balls of this article.
\begin{theoreme}
\label{theoremextra}
Let $\mu \in\mathcal{M}(\mathbb{R}^d)$ Let $(B_n)_{n\in\mathbb{N}}$ be a sequence of balls of $\mathbb{R}^d .$
\begin{enumerate}
\item If $(B_n)_{n\in\mathbb{N}}$ is $\mu$-a.c, then there exists a $\mu$-a.c sub-sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ which is weakly redundant.\medskip
\item If there exists $v<1$ such that $\mu(\limsup_{n\rightarrow+\infty}vB_n)=1$, then there exists a $\mu$-a.c sub-sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ verifying
\begin{equation}
\label{condimesu}
\underline{\dim}_H (\mu)\leq \liminf_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log\vert B_{\phi(n)}\vert}\leq \limsup_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log \vert B_{\phi(n)}\vert}\leq \overline{\dim}_P (\mu).
\end{equation}
\end{enumerate}
\end{theoreme}
\begin{remarque}
Theorem \ref{theoremextra} implies in particular that if the sequence of balls $(B_n)_{n \in\mathbb{N}}$ verifies $\mu( \limsup_{n\rightarrow+\infty}vB_n)=1$, for some $v<1$, it is possible to extract a $\mu$-a.c sub-sequence verifying items $(1)$ and $(2)$.
\end{remarque}
\bigskip
Section \ref{secequiac} and Section \ref{sec-extrac} are respectively dedicated to the proof of Theorem \ref{equiac} and Theorem \ref{theoremextra}.
Section \ref{sec-exam} provides some explicit applications of Theorem \ref{theoremextra}.
In the last section, Section \ref{sec-upper}, one proves Theorem \ref{majoss} and some final remark are given about Corollary \ref{equass1}.
\subsection{Application to the study of the optimality of lower-bounds in obtained via mass transference principles}
\subsubsection{An upper-bound for ubiquity Theorem in the self-similar case}
In this section, one shows how the previous extraction theorem can be used to investigate optimal bounds in inhomogeneous mass transference principles. Let us recall first the definition of the following geometric quantity, introduced in \cite{ED3}.
\begin{definition}
\label{mucont}
{Let $\mu \in\mathcal{M}(\R^d)$, and $s\geq 0$.
The $s$-dimensional $\mu$-essential Hausdorff content at scale $t\in(0,+\infty]$ of a set $A\subset \mathcal B(\R^d)$ is defined as}
{\begin{equation}
\label{eqmucont}
\mathcal{H}^{\mu,s}_{t}(A)=\inf\left\{\mathcal{H}^{s}_{t}(E): \ E\subset A , \ \mu(E)=\mu(A)\right\}.
\end{equation}}
\end{definition}
Let us also recall the definition of a self-similar measure.
\begin{definition}
\label{def-ssmu} A self-similar IFS is a family $S=\left\{f_i\right\}_{i=1}^m$ of $m\ge 2$ contracting similarities of $\mathbb{R}^d$.
Let $(p_i)_{i=1,...,m}\in (0,1)^m$ be a positive probability vector, i.e. $p_1+\cdots +p_m=1$.
The self-similar measure $\mu$ associated with $ \left\{f_i\right\}_{i=1}^m$ and $(p_i)_{i=1}^m$ is the unique probability measure such that
\begin{equation}
\label{equass}
\mu=\sum_{i=1}^m p_i \mu \circ f_i^{-1}.
\end{equation}
The topological support of $\mu$ is the attractor of $S$, that is the unique non-empty compact set $K\subset X$ such that $K=\bigcup_{i=1}^m f_i(K)$.
\end{definition}
The existence and uniqueness of $K$ and $\mu$ are standard results \cite{Hutchinson}. Recall that due to a result by Feng and Hu \cite{FH} any self-similar measure is exact dimensional.
Our goal is to investigate whether the lower-bound given by the following theorem, proved in \cite{ED3}, is sharp.
\begin{theoreme}
\label{zzani}
Let $\mu\in\mathcal{M}(\R^d)$ be a self-similar measure and $\mathcal{B}= (B_n)_{n\in\mathbb{N}}$ be a $\mu$-a.c. sequence of closed balls of~$\R^d$ centered in $\supp (\mu)$.
Let $\mathcal{U}=(U_n)_{n\in\mathbb{N}}$ be a sequence of open sets such that
$U_n \subset B_n$ for all $n\in \mathbb N$, and $0\leq s\leq \dim(\mu)$. If, for every $n\in\mathbb{N}$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\geq \mu(B_n) $, then
$$\dim_{H}(\limsup_{n\rightarrow +\infty}U_n)\geq s .
$$
\end{theoreme}
One now states the main result of this section.
\begin{theoreme}
\label{majoss}
Let $\mu \in \mathcal{M}(\R^d)$ be a self-similar measure, $K$ its support and $(B_n)_{n\rightarrow+\infty}$ be a weakly redundant sequence of balls of $\R^d$ verifying $\vert B_n \vert \to 0.$ Let $(U_n)_{n\in \mathbb{N}}$ be a sequence of open sets satisfying $U_n \subset B_n$. For any $0\leq s <\dim (\mu)$ such that, for all large enough $n\in\mathbb{N}$, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \mu(B_n)$,
\begin{equation}\label{uppernound18}
\dim_H (\limsup_{n\rightarrow+\infty}U_n \cap K)\leq s.
\end{equation}
\end{theoreme}
Note that if, for any $n\in\mathbb{N}$, the ball $B_n$ intersects $K$, $\limsup_{n\rightarrow+\infty}U_n \cap K=\limsup_{n\rightarrow+\infty}U_n$.
Theorem \ref{zzani} and Theorem \ref{majoss} yields the following useful corollary.
\begin{corollary}
\label{equass1}
Let $\mu \in\mathcal{M}(\mathbb{R}^d)$ be a self-similar measure. Let $(B_n)_{n\in\mathbb{N}}$ be a weakly redundant $\mu$-a.c sequence of balls satisfying $\vert B_n \vert \to 0$ and $B_n \cap K \neq \emptyset$ for any $n\in\mathbb{N}.$
Let $(U_n)_{n\in\mathbb{N}}$ be a sequence of open sets satisfying that, for any $n\in\mathbb{N}$, $U_n \subset B_n$.
Assume that there exists $s_0$ such that
\begin{itemize}
\item[•] for any $s<s_0$, for $n$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\geq \mu(B_n),$\mk
\item[•] for any $s>s_0$, for $n$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \mu(B_n).$\mk
\end{itemize}
Then by Theorem \ref{zzani} and Theorem \ref{majoss}, $$\dim_H (\limsup_{n\rightarrow+\infty}U_n)=s_0 .$$
\end{corollary}
\begin{remark}
\label{remajo}
It is easily seen from the proof that the condition $\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \mu(B_n)$ in Theorem \ref{majoss} can be weakened into $\liminf_{n\rightarrow+\infty}\frac{\log \mathcal{H}^{\mu,s}_{\infty}(U_n)}{\log \mu(B_n)}\geq 1$.
\end{remark}
\subsubsection{Application in the case of balls and rectangles}
We can now show in which sense, in view of Theorem~\ref{theoremextra}, Theorem~\ref{zzani} is sharp by applying Corollary \ref{equass1} to the specific cases where the sets $U_n$ are balls or rectangles.
\begin{corollary}\label{sharp}
Let $\mu \in\mathcal{M}(\R^d)$ be a self-similar measure of support $K$ and $\mathcal{B}=(B_n)_{n\in\mathbb{N}}$ be a sequence of balls centered in $K$ satisfying $\vert B_n\vert \to 0 $ and $\mu\Big( \limsup_{n\rightarrow+\infty}B_n\Big)=1.$ Then \cite{ED3},
$$\dim_H (\limsup_{n\rightarrow+\infty}B_n ^{\delta})\geq \frac{\dim(\mu)}{\delta}.
$$
Assume furthermore that $\mathcal{B} $ is weakly redundant and $\limsup_{n\rightarrow+\infty}\frac{\log \mu(B_n )}{\log(\vert B_n \vert)}=\dim (\mu)$, then for every $\delta\geq 1$,
$$\dim_H (\limsup_{n\rightarrow+\infty}B_n ^{\delta})=\frac{\dim(\mu)}{\delta}.
$$
\end{corollary}
\begin{corollary}
\label{rectssmajo}
Let $\mu$ be a self-similar measure verifying that its support, $K$, is the closure of its interior. Let $1\leq \tau_1\leq...\leq \tau_d$, $\tau=(\tau_1,..., \tau_d)$ and $(B_n :=B(x_n,r_n))_{n\in\mathbb{N}}$ be a sequence of balls of $\R^d$ satisfying $r_n \to 0$, $\mu(\limsup_{n\rightarrow+\infty}B_n)=1.$ Define $R_n =\widering{R}_{\tau}(x_n,r_n),$ where $
R_{\tau}(x_n,r_n)=x_n+\prod_{i=1}^d [-\frac{1}{2}r_n^{\tau_i},\frac{1}{2}r_n^{\tau_i}].$
Then \cite{ED3}
\begin{equation}
\dim_H (\limsup_{n\rightarrow+\infty}R_n)\geq \min_{1\leq i\leq d }\left\{\frac{\dim (\mu)+\sum_{1\leq j\leq i}\tau_i-\tau_j}{\tau_i}\right\}.
\end{equation}
Assume furthermore that $(B_n)_{n\in\mathbb{N}}$ is weakly redundant and $\lim_{n\rightarrow+\infty}\frac{\log \mu(B_n)}{\log \vert B_n \vert}=\dim (\mu)$, then
\begin{equation}
\dim_H (\limsup_{n\rightarrow+\infty}R_n)= \min_{1\leq i\leq d }\left\{\frac{\dim (\mu)+\sum_{1\leq j\leq i}\tau_i-\tau_j}{\tau_i}\right\}.
\end{equation}
\end{corollary}
\begin{remarque}
\begin{itemize}
\item[•] Corollaries \ref{sharp} and \ref{rectssmajo} are direct consequences of second item of Remark 5.1 and Remark 5.3 in \cite{ED3}, together with Corollary \ref{equass1} (applied to, respectively, $s_0 =\frac{\dim (\mu)}{\delta}$ and $s_0 =s(\mu,\tau)$).\mk
\item[•] Note that, by Theorem \ref{theoremextra} combined with Corollary \ref{sharp} and Corollary \ref{rectssmajo}, for any sequence of balls $(B_n)_{n\in\mathbb{N}}$ satisfying $\mu\Big(\limsup_{n\rightarrow+\infty}\frac{1}{2} B_n\Big)=1$ ($\mu$ a self-similar measure satisfying the of hypothesis of Corollaries \ref{sharp} or \ref{rectssmajo} for $\mu$), it is always possible to extract a $\mu$-a.c sub-sequence of balls so that the Hausdorff dimension of the limsup set associated with corresponding $U_n$'s is the bound stated in \cite{ED3} and recalled in those corollaries. This in particular proves that those bounds are sharp. \medskip
\item[•]In the case of the Lebesgue measure, it is always verified that $\lim_{n\rightarrow+\infty}\frac{\log \mu(B_n)}{\log \vert B_n \vert}=\dim (\mu).$ As a consequence, the lower-bound provided by Theorem \ref{zzani} (which is established in \cite{KR}) in the case of balls or rectangles is precisely the dimension of $\limsup_{n\rightarrow+\infty}U_n$ as soon as the sequence $(B_n)$ is weakly redundant. More explicitly, given a weakly redundant sequence of balls $(B_n)_{n\in\mathbb{N}}$ of $[0,1]^d$ satisfying $\vert B_n \vert \to 0$ and $ \mathcal{L}^d (\limsup_{n\rightarrow+\infty}B_n)=1$, for any sequence rectangles associated with a vector $\tau$ as in Theorem \ref{rectssmajo}, one has
$$\dim_H ( \limsup_{n\rightarrow+\infty}R_n)= \min_{1\leq i\leq d }\left\{\frac{d+\sum_{1\leq j\leq i}\tau_i-\tau_j}{\tau_i}\right\}.$$
\end{itemize}
\end{remarque}
\medskip
Section \ref{secpequiac} is dedicated to the proof of Theorem \ref{equiac}. In the next section, Section \ref{sec-extrac}, Theorem \ref{theoremextra} is established. Then some explicit examples of application of Theorem \ref{theoremextra} are given in Section \ref{sec-exam}.
In the penultimate section, Section \ref{sec-upper}, Theorem \ref{majoss} is proved.
The last section, Section \ref{conclu}, draws some conclusions and gives some perspectives about the results established in this article.
\section{Proof of Theorem \ref{equiac}}
\label{secpequiac}
\subsection{A useful modified version of Besicovitch covering Lemma}
\label{secbesi}
One focuses on a modified version of Besicovitch's covering Lemma.
\begin{proposition}
\label{besimodi}
For any $0<v\leq 1$, there exists $Q_{d,v} \in\mathbb{N}^{\star}$, a constant depending only on the dimension $d$ and $v$, such that for every $E\subset [0,1]^{d}$, for every set $\mathcal{F}=\left\{B(x, r_{(x)} ): x\in E, r_{(x)} >0 \right\}$, there exists $\mathcal{F}_1,...,\mathcal{F}_{Q_{d,v}}$ finite or countable sub-families of $\mathcal{F}$ such that:\medskip
\begin{itemize}
\item
$\forall 1\leq i\leq Q_{d,v}$, $\forall L\neq L'\in\mathcal{F}_i$, one has $\frac{1}{v}L \cap \frac{1}{v}L'=\emptyset.$\medskip
\item
$E$ is covered by the families $\mathcal{F}_i$, i.e.
\begin{equation}\label{besi}
E\subset \bigcup_{1\leq i\leq Q_{d,v}}\bigcup_{L\in \mathcal{F}_i}L.
\end{equation}
\end{itemize}
\end{proposition}
The case $v=1$ corresponds to the standard Besicovich's covering lemma (see \cite{Ma}, Chapter 2, pp. 28-34 for instance).
A first step toward Proposition \ref{besimodi} is the next lemma, that allows to split a given family of "weakly" overlapping balls into a finite number of families of disjoint balls.
\begin{lemme}
\label{fami}
Let $0<v<1$ and $\mathcal{B} =(B_n)_{n\in\mathbb{N}}$ be a countable family of balls such that $\lim_{n\to +\infty} |B_n|=0$, and for every $n\neq n' \in \N$, $vB_n \cap vB_n' =\emptyset$.
There exists $\gamma_{d,v}+1$ ($\gamma_{d,v}$ being the constant appearing in Lemma \ref{dimconst} below) sub-families of $\mathcal{B}$, $(\mathcal{F}_i)_{1\leq i\leq \gamma_{d,v}+1}$, such that:
\begin{itemize}
\item
$\mathcal{B}=\bigcup_{1\leq i\leq \gamma_{d,v}+1}\mathcal{F}_i$,
\item
$\forall \, 1\leq i\leq \gamma_{d,v}+1$, $\forall L\cap L' \in\mathcal{F}_i$, one has $L \cap L'=\emptyset.$
\end{itemize}
\end{lemme}
\begin{proof}
The proof is based on the following lemma, whose proof can be found in \cite{Ma}, Lemma 2.7, pp.30 - there, the result is obtained for $v= {1}/{2}$ but the proof remains valid for any $v<1$.
\begin{lemme} \label{dimconst}
For any $0<v\leq 1$ there exists a constant $\gamma_{v,d} >0 $ depending only on $v$ and the dimension $d$ only, satisfying the following: if a family of balls $\mathcal{B} =(B_n )_{n\in\mathbb{N}}$ and a ball $B$ are such that
\begin{itemize}
\item
$\forall \ n\geq 1$, $\vert B_n\vert \geq \frac{1}{2}\vert B\vert,$
\item
$\forall \ n_1 \neq n_2 \geq 1 $, $vB_{n_1}\cap vB_{n_2}=\emptyset,$
\end{itemize}
then $B$ intersects at most $\gamma_{v,d}$ balls of $\mathcal{B}$.
\end{lemme}
The families $\mathcal{F}_1 ,..., \mathcal{F}_{\gamma_{d,v}+1}$ are built recursively.
For $k\in\mathbb{N}$, call $\mathcal{G}^{(k)}=\left\{L\in\mathcal{F}: 2^{-k-1} < \vert L\vert \leq 2^{-k}\right\}$. Notice that, because $\lim_{n\to +\infty} |B_n|=0$, each $\mathcal{G}^{(k)}$ is empty or finite.
Observe first that for every $k\in\mathbb{N}$ and every ball $B\in\mathcal{G}^{(k)}$, and every pair of balls $B_1 \neq B_2\in\bigcup_{k'\leq k}\mathcal{G}^{(k')}\setminus\left\{B\right\}$, one has $vB_1 \cap v B_2=\emptyset$ and for $i=1,2$, $\vert B_i \vert \geq \frac{\vert B \vert }{2}$. By Lemma \ref{dimconst}, this implies that $B$ intersects at most $\gamma_{d,v}$ balls of $\bigcup_{k'\leq k}\mathcal{G}^{(k')}\setminus\left\{B\right\}$.
To get Lemma \ref{fami}, we are going to sort the balls of $\bigcup_{k '\leq k}\mathcal{G}^{(k')}$ recursively on $k$ into families $\mathcal{F}_1 ,..., \mathcal{F}_{\gamma_{d,v}+1}$ of pairwise disjoint balls. At each step, a new ball $B$ will be added to one of those families of balls $\mathcal{F}_{i}$ and the resulting family, $\mathcal{F}_i \bigcup \left\{B\right\}$ will be denoted again by $\mathcal{F}_i$.
\mk
Let $k_0$ be the smallest integer such that $\mathcal{G}^{(k_0)}$ is non-empty. Consider an arbitrary $L_0 \in\mathcal{G}^{(k_0)}$. By Lemma \ref{dimconst}, $L_0$ intersects $n_{0}\leq\gamma_{d,v}$ other balls of $\mathcal{G}^{(k_0)}$, that are denoted by $L_1 ,...,L_{n_0}$. The sets $\mathcal{F}_i$ are then set as follows:
\begin{itemize}
\item
$\forall \, 1\leq i\leq n_0$, $\mathcal{F}_i =\left\{L_i \right\}$,
\item
$\forall \ n_0 +1\leq i\leq \gamma_{d,v}$, $\mathcal{F}_i=\emptyset$,
\item
$\mathcal{F}_{\gamma_{d,v}+1}=\left\{L_0\right\}.$
\end{itemize}
Further, consider $\widetilde L \notin \bigcup_{0\leq i\leq n_0}\left\{L_i\right\}$ (whenever such an $\widetilde L$ exists). The same argument (Lemma \ref{dimconst}) ensures that $\widetilde L $ intersects at most $\gamma_{d,v}$ balls of $\mathcal{G}^{(k_0)}$.
In particular there must exists $1\leq i\leq \gamma_{d,v}+1$ such that for every $L\in\mathcal{F}_i$, $\widetilde L \cap L=\emptyset$. Choosing arbitrarily one of those indices $i$, one adds $\widetilde L$ to $\mathcal{F}_i :=\widetilde L\bigcup \mathcal{F}_i $ (we keep the same name for this new family).
\mk
The same argument remains valid for any other ball $L^{\prime \prime} \notin \bigcup_{1\leq j\leq \gamma_{d,v}+1}\bigcup_{L\in\mathcal{F}_j}\left\{L\right\}$. Hence, proceeding recursively on all balls of $\mathcal{G}^{(k_0)}$ allows to sort the balls of $\mathcal{G}^{(k_0)}$ into $ \gamma_{d,v}+1$ families $(\mathcal{F}_i)_{1 \leq i \leq \gamma_{d,v}+1}$ of pairwise disjoint balls.
\mk
Next, let $k_1$ be the smallest integer such that $k_1 >k_0$ and $\mathcal{G}^{(k_1 )}$ is non empty, take an arbitrary $L_{0}^{(1)}\in\mathcal{G}^{(k_1)}$. It is trivial to check that the family $\mathcal{G}^{(k_0)}\cup \mathcal{G}^{(k_1)}$ and the ball $L_{0}^{(1)}$ satisfy the conditions of Lemma \ref{dimconst}. Subsequently, $L_{0}^{(1)}$ intersects at most $\gamma_{d,v}$ balls of $\mathcal{G}^{(k_0)}\bigcup \mathcal{G}^{(k_1)}$, and there must exist an integer $1\leq i _0 \leq \gamma_{d,v}+1$ such that $L_{0}^{(1)}\cap \bigcup_{L\in\mathcal{F}_{i_0}}L=\emptyset$. As before, we add this ball $L_{0}^{(1)} $ to the family $\mathcal{F}_{i_0} $.
Consider $\widetilde {L } \in \mathcal{F}^{(k_1)}$ such that $\widetilde {L } \notin \bigcup_{1\leq i\leq \gamma_{d,v}+1}\mathcal{F}_i$ (whenever such a ball exists). The exact same argument shows the existence of an integer $1\leq \widetilde i \leq \gamma_{d,v}+1$ such that $\widetilde {L } $ intersects at most $ \gamma_{d,v}$ balls of $\mathcal{G}^{(k_0)}\bigcup \mathcal{G}^{(k_1)}$. One adds $\widetilde {L } $ to the family $\mathcal{F}_{\, \widetilde i}$, which remains composed only of pairwise disjoint balls.
One applies this argument to every ball of $\mathcal{F}^{(k_1)}$, hence finally sorting the balls of $\mathcal{F}^{(k_0)}\cup \mathcal{F}^{(k_1)}$ into $\gamma_{d,v}+1$ families of pairwise disjoint balls, as requested.
\mk
It is now easily seen that one can proceed recursively on $k\geq k_0$, ending up with the families $\mathcal{F}_1,...,\mathcal{F}_{\gamma_{d,v}+1}$ fulfilling the desired properties.
\end{proof}
We are now ready to prove Proposition \ref{besimodi}.
\begin{proof} Fix $E\subset [0,1]^{d}$ and $\mathcal{F}=\left\{B(x, r_{(x)} ): x\in E, r_{(x)} >0 \right\}$.
One applies Besicovich's theorem (i.e. Proposition \ref{besimodi} with $v=1$) to $\mathcal{F}=\left\{B(x, r_{(x)} ):x\in E \ r_{(x)} >0 \right\}$. This provides us with a finite set of families of balls $\mathcal{G}_1 ,...,\mathcal{G}_{\gamma_{d,1}+1}$ composed of pairwise disjoint balls satisfying \eqref{besi}, i.e. $
E\subset \bigcup_{1\leq i\leq Q_{\gamma_{d,1}+1}} \bigcup_{L\in \mathcal{G}_i}L.$
For every $1\leq i\leq Q_{\gamma_{d,1}+1}$, one sets $\mathcal{G}^{(v)}_i=\left\{\frac{1}{v}L: L\in\mathcal{G}_i\right\}$, i.e. the sets of balls with same centers as $\mathcal{G}_i$ but with radii multiplied by $v^{-1}>1$. Notice that by construction, $\forall \, 1\leq i\leq Q_{\gamma_{d,1}+1}$, $\forall \, L \neq L' \in\mathcal{G}_i ^{(v)}$, one has $vL \cap vL' =\emptyset$. Hence, Lemma \ref{fami} yields $\gamma_{d,v}+1$ sub-families $(\mathcal{G}^{(v)}_{i,j})_{1\leq j \leq \gamma_{d,v}+1}$ of $\mathcal{G}_{i}^{(v)}$ such that: \medskip
\begin{itemize}
\item
$\forall \ 1\leq j\leq \gamma_{d,v}+1$, $\forall \ L \neq L' \in \mathcal{G}^{(v)}_{i,j}$, one has $L \cap L' =\emptyset,$\medskip
\item
$ \mathcal{G}_{i}^{(v)}=\bigcup_{1\leq j\leq \gamma_{d,v}+1}\mathcal{G}_{i,j}^{(v)}.$\medskip
\end{itemize}
Finally, we set for every $ 1\leq i\leq Q_{d,1} $ and $1\leq j\leq \gamma_{d,v}+1$
$$\mathcal{F}_{i,j}=\left\{vL: L\in\mathcal{G}_{i,j}^{(v)}\right\} \ \ \mbox { and } \ \ \ \mathcal{F}_{i} = \bigcup_{1\leq j\leq \gamma_{d,v}+1}\mathcal{F}_{i,j} .$$
These sets verify that:
\begin{itemize}
\item
$\forall \ 1\leq i\leq Q_{d,1}$, $\forall \ 1\leq j\leq \gamma_{d,v}+1$, $\forall L \neq L' \in\mathcal{F}_{i,j}$, $\frac{1}{v}L \cap \frac{1}{v} L' = \emptyset$ (because the balls of $\mathcal{G}_{i,j}$ are pairwise disjoint),\medskip
\item
$E\subset \bigcup_{1\leq i\leq Q_{d,1}}\mathcal{G}_i =\bigcup_{1\leq i\leq Q_{d,1}} \ \bigcup_{1\leq j\leq \gamma_{d,v}+1}\mathcal{F}_{i,j}.$\medskip
\end{itemize}
This proves the statement and the fact that $Q_{d,v}=Q_{d,1}.(\gamma_{d,v}+1)$.
\end{proof}
\subsection{Consequences of the $\mu$-asymptotic covering property}
\label{seccons}
One first shows that the constant $C$ in Definition \ref{ac} can be replaced by 1 if infinite subsequences of balls are authorized. In fact Definition \ref{ac} ensures that any open set can be covered (with respect to the $\mu$-measure) by disjoint balls $B_n$ of arbitrary large indices.
\begin{lemme}
\label{covO}
Let $\mu\in\mathcal{M}(\mathbb{R}^d)$ and $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ be a $\mu$-a.c sequence of balls of $ \mathbb{R}^d$ with $\lim_{n\to +\infty} r_{n}= 0$.
Then for every open set $\Omega$ and every integer $g\in\mathbb{N}$, there exists a subsequence $(B_{(n)}^{(\Omega)})\subset \left\{B_n \right\}_{n\geq g} $ such that:
\begin{enumerate}
\item
$\forall \, n\in\mathbb{N}$, $B_{(n)}^{(\Omega)}\subset \Omega,$
\sk
\item
$\forall \, 1\leq n_1\neq n_2$, $B_{(n_1)}^{(\Omega)}\cap B_{(n_2)}^{(\Omega)}=\emptyset$,
\sk
\item
$\mu\left (\bigcup_{n\geq 1}B_{(n)}^{(\Omega)}\right)= \mu(\Omega).$
\end{enumerate}
In addition, there exists an integer $N_{\Omega}$ such that for the balls $ (B_{(n)}^{(\Omega)})_{n=1,...,N_\Omega}$, the conditions (1) and (2) are realized, and (3) is replaced by $\mu\left (\bigcup_{n=1}^{N_\Omega} B_{(n)}^{(\Omega)}\right)\geq \frac{3}{4} \mu(\Omega).$
\end{lemme}
The last part of Lemma \ref{covO} simply follows from item (3) and the $\sigma$-additivity of $\mu$.
\begin{proof}
The idea consists in covering $\Omega$ by pairwise disjoint balls amongst those balls of $\mathcal{B} $, such that their union has measure at least $C\mu(\Omega)$, then in covering the complementary of the union of those balls in $\Omega$ (that is still open) with at least a proportion $C$ of its measure, and so on.
More precisely, this is achieved as follows:
\mk$\bullet$
{\bf Step 1:} By application of Definition \ref{ac} to $\Omega_0 :=\Omega$ and $g\in \N$, there exists $C>0$ and some integers $g\leq n_1 \leq ...\leq n_{N_{0}}$ so that the family of balls $\mathcal{F}_{0}:=\left\{B_{n_i}:=B^{(0)}_i\right\}_{1\leq i\leq N_{0}}$ is pairwise disjoint and $\mu(\bigcup_{1\leq i\leq N_{0}}B_{n_i})\geq C\mu(\Omega).$
\mk$\bullet$ {\bf Step 2:} Setting $\Omega_1 =\Omega\setminus \bigcup_{ L\in\mathcal{F}_0}L$, applying Definition \ref{ac} to $\Omega_1$ with the integer $g$ provides us with a family $\mathcal{F}_1$ of pairwise disjoint balls $B_{1}^{(1)},...,B_{N_1}^{(1)}\in\left\{B_n \right\}_{n\geq g}$ such that $\forall \ 1\leq i\leq N_1$ $B^{(1)}_i \subset \Omega_1$ and $$ \mu(\bigcup_{1\leq i\leq N_2}B_{i}^{(1)})\geq C\mu(\Omega_1).$$
One sets $\mathcal{F}_1 =\mathcal{F}_0 \bigcup \mathcal{F}_1 $. One sees that
\begin{align*}
\mu \left(\bigcup_{L\in\mathcal{F}_1}L\right )&= \mu\left(\bigcup_{L\in\mathcal{F}_0}L\right)+\mu\left(\bigcup_{L\in\mathcal{G}_1}L\right)\\
& \geq \mu\left(\bigcup_{L\in\mathcal{F}_0}L\right)+C\left(\mu(\Omega)-\mu\left( \bigcup_{L\in\mathcal{F}_0}L\right)\right)\\
&\geq (1-C)\mu\left(\bigcup_{L\in\mathcal{F}_0}L\right)+C\mu\left(\Omega\right) \\
&\geq (C+C(1-C))\mu(\Omega) .
\end{align*}
Observe that the balls of $\mathcal{F}_0$ and $\mathcal{F}_1$ are disjoint by construction.
\mk$\bullet$ {\bf Following steps :} Proceeding recursively, and applying the exact same argument as above, one constructs an increasing sequence of families $(\mathcal{F}_{i})_{i\in\mathbb{N}}$ and a decreasing sequence of open sets $\Omega_i$ such that:
\begin{itemize}
\sk
\item
$\forall \, i\in\mathbb{N}$, $L\in \{B_n \}_{n\geq g} $ and $\forall L\in\mathcal{F}_i $, $L\subset \Omega_i\subset \Omega,$
\sk
\item
$\forall \,i\in\mathbb{N}$, $\forall L \neq L'\in\mathcal{F}_i $, $L \cap L' =\emptyset$,
\sk
\item
$\forall \, i\neq j\in\mathbb{N}$, $\forall L \in\mathcal{F}_i $ and $\forall L \in\mathcal{F}_j $, $L \cap L' =\emptyset$,
\sk
\item $\forall \, i\in\mathbb{N}$,
$\mu\left(\bigcup_{L\in\mathcal{F}_i}L\right)\geq \mu(\Omega) \sum_{1\leq k\leq i}C(1-C)^{k-1}.$
\end{itemize}
Finally, setting $\mathcal{F}=\bigcup_{i\in\mathbb{N}}\mathcal{F}_i$, one sees that $\mathcal{F}$ is constituted by pairwise disjoint balls chosen amongst $\left\{B_n\right\}_{n\geq g}$ satisfying
\begin{equation}
\label{usefulmath}
\mu(\Omega)\geq \mu\left(\bigcup_{L\in\mathcal{F}}L\right )\geq \mu(\Omega)\sum_{k\geq 1}C(1-C)^{k-1}=\mu(\Omega),
\end{equation}
so that $\mathcal{F}$ fulfills the conditions of Lemma \ref{covO}.
\end{proof}
An easy consequence is the following.
\begin{corollary}
\label{covBor}
Let $\mu \in\mathcal{M}(\mathbb{R}^d)$ and $(B_n )_{n\in\mathbb{N}}$ be a $\mu$-a.c sequence of balls. Then for any Borel set $E$, for any $g\in\mathbb{N}$, there exists a sub-sequence of balls $(B_{(n)}^{(E)})\subset \left\{B_n\right\}_{n\geq g}$ such that:
\begin{enumerate}
\item $\forall 1\leq n_1 \neq n_2$, $B_{(n_1)}^{(E)}\cap B_{(n_2)}^{(E)}=\emptyset,$\mk
\item $\mu\Big(\bigcup_{n\in\mathbb{N}}B_{(n)}^{(E)}\cap E\Big)=\mu(E),$\mk
\item $\mu\Big(\bigcup_{n\in\mathbb{N}}B_{(n)}^{(E)}\Big)\leq \mu(E)+\varepsilon,$
\end{enumerate}
\end{corollary}
\begin{proof}
By outer regularity, there exists an open set $\Omega$ such that $E\subset \Omega$ and $m(\Omega)\leq \mu(E)+\varepsilon.$ Applying Lemma \ref{covO} to $\Omega$, the sequence $(B_n)_{n\in\mathbb{N}}$ fulfills the condition of Corollary \ref{covBor}.
\end{proof}
\subsection{Proof of Theorem \ref{equiac}}
\label{secequiac}
{\bf (1)} Assume first that $\mathcal{B} = (B_n)_{n\in\mathbb{N}}$ is $\mu$-a.c, and let us prove that $\mu(\limsup_{n\rightarrow+\infty}B_n)=1$.
For every $g\in\mathbb{N}$, applying Lemma \ref{covO}, there exists a sub-family of balls, $\mathcal{F}_g \subset \left\{B_{n}\right\}_{n\geq g}$ such that $\mu(\bigcup_{L\in\mathcal{F}_g}L)=\mu(\mathbb{R}^d)=1.$ In particular, $\mu(\bigcup_{n\geq g}B_n)=1$ for every $g\geq 1$, and $\mu(\limsup_{n\rightarrow+\infty}B_n)=\mu(\bigcap_{g\geq 1}\bigcup_{n\geq g}B_n)=1$.
\medskip
{\bf (2)} Suppose next that there exists $v<1$ such that $\mu(\limsup_{n\rightarrow+\infty}v B_n)=1$, and let us show that $\mathcal{B} $ is $\mu$-a.c.
Let $\Omega$ be an open set in $\mathbb{R}^d$. Our goal is to find a constant $C$ such that the conditions of Definition \ref{ac} are realized.
\mk
Let $E= \Omega \cap \limsup_{n\rightarrow+\infty}vB_n$.
For every $y\in E$, consider an integer $n_y\geq g$ large enough so that $y\in vB_{n_y}$ and $B(y,2r_{n_y} )\subset \Omega.$ This is possible since $\lim_{n\to +\infty} r_n=0$.
Since $y\in vB_{n_y}$, one has
\begin{equation}
\label{eq11}
B(y,(1-v)r_{n_y}) \subset B_{n_y} \subset B(y, (1+v) r_{n_y}) \subset B(y,2r_{n_y}),
\end{equation}
and the family $\mathcal{F}= \left\{ B(y, (1-v) r_{n_y}): y\in E\right\}$ covers $E $ by balls centered on $E$.
Applying Proposition \ref{besimodi} with constant $v'= \frac{1-v}{2} <1$ allows to extract from $\mathcal{F}$ finite or countable sub-families $\mathcal{F}_1,...,\mathcal{F}_{Q_{d,v'}}$ such that:
\begin{itemize}
\item
$\forall 1\leq i\leq Q_{d,v'}$, $L\neq L'\in\mathcal{F}_i$, one has $ \frac{1}{v'} L \cap \frac{1}{v'} L'=\emptyset.$\medskip
\item
$E$ is covered by the families $\mathcal{F}_i$, i.e. \eqref{besi} holds true.
\end{itemize}
Now, $\mu(\Omega) =\mu(E) \leq \mu \left ( \bigcup_{i=1}^{ Q_{d,v'}} \bigcup_{L\in\mathcal{F}_{i}} L \right ) $.
There must exist $1\leq i_0 \leq Q_{d, v'}$ such that
$$
\mu \left (\bigcup_{L\in\mathcal{F}_{i_0}}L \right)\geq \frac{1}{Q_{d,v'}}\mu(E )= \frac{1}{Q_{d, v'}}\mu(\Omega).
$$
There exist $L_1$, $L_2$, ... $L_N$ balls of $\mathcal{F}_{i_0}$ such that
$$\mu\left (\bigcup_{1\leq k\leq N}L_{k} \right )\geq\frac{1}{2 Q_{d,v'}}\mu(\Omega),
$$
Notice the following facts:
\begin{itemize}
\item
$\forall 1\leq i\leq Q_{d,v'}$, every $L \in\mathcal{F}_i$ is naturally associated with some $y \in E$ and some ball $B_{n_y}$, with $L\subset B_{n_y} \subset \Omega$,
\item
$\forall 1\leq i\leq Q_{d,v'}$, if $L \in\mathcal{F}_i$ is associated with $y \in E$ and $B_{n_y}$ and $L' \in\mathcal{F}_i$ is associated with $y' \in E$ and $B_{n_{y'}}$, then $ \frac{1}{v'} L \cap \frac{1}{v'} L'=\emptyset$ implies by \eqref{eq11} that $B_{n_{y}} \cap B_{n_{y'}}=\emptyset$.
\end{itemize}
The first fact implies that there exist $N$ integers $n_1$, ..., $n_N$ such that $B_{n_k}\subset \Omega$ and
$$
\mu\left (\bigcup_{1\leq k\leq N} B_{n_k} \right )\geq\frac{1}{2 Q_{d,v'}}\mu(\Omega),
$$
The second fact implies that these balls $B_{n_k} $, $k=1,..., N$ are pairwise disjoint.
This exactly proves that $\mathcal{B}$ is $\mu$-a.c.
\section{proof of Theorem \ref{theoremextra}}
\label{sec-extrac}
The following section is dedicated to the study of the properties one can ask an $\mu$-a.c sequence $(B_n)_{n \in\mathbb{N}}$ to verify, up to an $\mu$-a.c extraction.
The concept of conditioned ubiquity was introduced by Barral and Seuret in \cite{BS3}. It consists in asking the balls of the sequence $(B_n)$ to verify some specific properties with respect to the measure $\mu$. When investigating the Hausdorff dimension of some sets $(U_n)_{n\in\mathbb{N}}$, where $U_n \subset B_n$, in practical cases (when the measure carries some self-similarity), it turns out that when a lower-bound is found for $(U_n)_{n\in\mathbb{N}}$ using the fact that the sequence $(B_n)_{n\in\mathbb{N}}$ is of limsup of full $\mu$-measure, it is often quite easy to prove that $\limsup_{n\rightarrow+\infty}U_n$ has precisely the expected measure provided that the sequence $(B_n)$ verifies some specific properties with respect to $\mu$.
Note that in full generality, understanding the optimality of a bound as mentioned above, means understanding very finely the behavior of the measure $\mu$ on the sets $U_n$ (the sequence $(B_n)$ being $\mu$-a.c). It will be proved in this article that, under mild conditions on the sequence $(B_n)_{n\in\mathbb{N}}$, it is always possible to give a natural upper-bound for $\dim_H (\limsup_{n\rightarrow+\infty}U_n)$. This upper-bound turns out to be optimal when the measure carries enough self-similarity (in particular it works for quasi-Bernoulli measures or fully supported self-similar measures).
\textbf{In this section, the balls $(B_n)_{n\in\mathbb{N}}$ are supposed to be pairwise distinct and such that $\vert B_n\vert \underset{n\rightarrow+\infty}{\rightarrow}0$.}
\subsection{Extraction of weakly redundant $\mu$-a.c subsequences}
\label{secnest1}
The main result of this section is stated here.
\begin{proposition}
\label{extrawr}
Let $\mu \in \mathcal{M}(\mathbb{R}^d)$ and $(B_n)_{n\in\mathbb{N}}$ be a $\mu$-a.c sequence of balls $\lim_{n\rightarrow+\infty}\vert B_n \vert=0.$ There exists a subsequence $(B_{\psi(n)})_{n\in\mathbb{N}}$ of $(B_n)_{n\in\mathbb{N}}$ which is weakly redundant and $\mu$-a.c.
\end{proposition}
\begin{proof}
Let $g_k \in\mathbb{N}$ be large enough so that $\forall n\geq g_k$, $\vert B_n \vert \leq 2^{-k}.$ By Proposition \ref{covO}, applied with the sequence $(B_n)_{n\in\mathbb{N}}$, $\Omega=\mathbb{R}^d$ for any $k \in \mathbb{N}$, there exists a sub-sequence $(B_{(n,k)})$ of $\left\{B_n\right\}_{n\geq g_k}^{\mathbb{N}}$ satisfying
\begin{enumerate}
\item $\forall 1\leq n_1 \neq n_2$, $B_{(n_1 ,k)}\cap B_{(n_2 ,k)}=\emptyset,$\mk
\item $\mu\Big(\bigcup_{n\in\mathbb{N}}B_{(n,k)}\Big)=1.$\mk
\end{enumerate}
Define $\mathcal{B}_{\psi}=(B_{\psi(n)})_{n\in\mathbb{N}}$ as the sub-sequence of balls corresponding to $\bigcup_{k\in\mathbb{N}}\left\{B_{(n,k)}\right\}_{n\in\mathbb{N}}.$
Since the following inclusion holds
\begin{equation}
\label{reuint}
\bigcap_{k\in\mathbb{N}}\bigcup_{n\in\mathbb{N}}B_{(n,k)}\subset \limsup_{n\to+\infty}B_{\psi(n)},
\end{equation}
by item $(2)$ one has
$$\mu(\limsup_{n\to+\infty}B_{\psi(n)})=1.$$
Note that, for all $k\in\mathbb{N}$, all $B\in \left\{B_{(n,k)}\right\}_{n\in\mathbb{N}}$, $\vert B\vert \leq 2^{-k}$. Following the notation of Definition \ref{wr}, for any $k\in\mathbb{N}$, $\mathcal{T}_k (\mathcal{B}_{\psi})$ can contain only balls of the sequence of the $k$ first families $\left\{B_{(n,k)}\right\}_{n\in\mathbb{N}}$, which are composed of pairwise disjoint balls. This proves that $\mathcal{T}_k (\mathcal{B}_{\psi})$ can be sorted in at most $k+1$ families of pairwise disjoint balls. In particular, $\mathcal{B}_{\psi}$ is weakly redundant.
It remains to show that $(B_{\psi(n)})_{n\in\mathbb{N}}$ is $\mu$-a.c.
Let $\Omega$ be an open set and $g\in\mathbb{N}$. One will extract from $\mathcal{B}_{\psi}$ a finite number of balls satisfying the condition of Definition \ref{ac}.
There exists $k_0 $ so large that, $\mu\big(\left\{x\in \Omega(x,2^{-k_0+1})\right\}\big)\geq \frac{3\mu(\Omega)}{4}$ for any $k\geq k_0$,
$\left\{B_{(n,k)}\right\}_{n\in\mathbb{N}}\subset \left\{B_{\psi(n)}\right\}_{n\geq g}$ and $\mu(\limsup_{n\rightarrow+\infty}B_{\psi(n)}\cap \Omega )\geq \frac{3\mu(\Omega)}{4}$. Setting
$$\widehat{E}=\left\{x\in\limsup_{n\rightarrow+\infty}B_{\psi(n)}\cap \Omega: B(x,2^{-k_0 +1})\subset \Omega\right\},$$
it holds that
$$\mu(\widehat{E})\geq \frac{1}{2}\mu(\Omega).$$
Recalling \eqref{reuint}, for every $x\in \widehat{E}$, consider $B_{x}$, the ball of $\left\{B_{(n,k_0)}\right\}_{n\in\mathbb{N}}$ containing $x$. Note that, because for $B\in\left\{B_{(n,k_0)}\right\}_{n\in\mathbb{N}}$, $\vert B\vert \leq 2^{-k_0}$, one has $B_x \subset B(x,2^{-k_0 +1})\subset \Omega .$ Set $\mathcal{F}_1 =\left\{B_x :x\in\widehat{E}\right\}.$ The set $\mathcal{F}_1$ is composed of pairwise disjoint balls (by item $(1)$ above) of $\left\{B_{\psi(n)}\right\}_{n\geq g} $ included in $\Omega$ and such that
\begin{equation}
\mu\Big(\bigcup_{L\in\mathcal{F}_1}L\Big)\geq \mu(\widehat{E})\geq \frac{1}{2}\mu(\Omega).
\end{equation}
Using the $\sigma$-additivity of $\mu$ concludes the proof.
\end{proof}
\medskip
\subsection{Extraction of sub-sequences of balls with conditioned measure}
\label{seccondi}
Let $\mu \in\mathcal{M}(\mathbb{R}^d)$ and $(B_n)_{n\in\mathbb{N}}$ be an $\mu$-a.c sequence of balls.
This part aims to understand what condition can be assumed about the measure of the ball of the sequence $(B_n)_{n\in\mathbb{N}}$ in general under the $\mu$-a.c condition.
More precisely, item $(2)$ of Theorem \ref{theoremextra} is proved.
\begin{proposition}
\label{mainextraprop}
Let $\mu\in\mathcal{M}(\mathbb{R}^d)$. For any sequence of balls $(B_n)_{n\in\mathbb{N}}$ satisfying $\mu(\limsup_{n\rightarrow+\infty}vB_n)=1$ for some $0<v<1$, there exists an $\mu$-a.c sub-sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ verifying
\begin{equation*}
\label{condimesu1}
\underline{\dim}_H (\mu)\leq \liminf_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log\vert B_{\phi(n)}\vert}\leq \limsup_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log \vert B_{\phi(n)}\vert}\leq \overline{\dim}_P (\mu).
\end{equation*}
\end{proposition}
\begin{remarque}
For the left part of \eqref{condimesu1}, the proof actually only uses the fact that $\mu (\limsup_{n\rightarrow+\infty}B_n)=1.$
\end{remarque}
Let us introduce some useful sets to prove Lemma \ref{gscainf} and Lemma \ref{gscaup}, which are key in order to prove \eqref{condimesu1}.
{
\begin{definition}
\label{emudef}
Let $0\leq \alpha\leq\gamma$ be real numbers, $\mu\in\mathcal{M}(\mathbb{R}^d)$, and $\varepsilon,\rho>0$ two positive real numbers. Then define
\begin{equation}
\label{emuhat}
{{E}_{\mu}^{[\alpha,\gamma],\rho,\varepsilon}=\left\{x\in\mathbb{R}^d : \ \underline\dim (\mu,x)\in[\alpha,\gamma] \text{ and }\forall r\leq \rho, \ \mu(B(x,r))\leq r^{\underline{\dim} (\mu,x)-\varepsilon}\right\}},
\end{equation}
\begin{equation}
\label{fmu}
F_{\mu}^{[\alpha,\beta],\rho,\varepsilon}=\left\{x\in \mathbb{R}^d : \overline{\dim} (\mu,x)\in [\alpha, \beta]\text{ and }\forall r<\rho, \ \mu(B(x,r))\geq r^{\overline{\dim} (\mu,x)+\varepsilon}\right\}.
\end{equation}
and
\begin{align}
\label{emut}
E_{\mu}^{[\alpha,\gamma],\varepsilon} =\bigcup_{n\geq 1}E_{\mu}^{[\alpha,\gamma],\frac{1}{n},\varepsilon} \ \ \ \mbox{ and }\ \ \ \ F_{\mu}^{[\alpha,\gamma],\varepsilon} =\bigcup_{n\geq 1}F_{\mu}^{[\alpha,\gamma],\frac{1}{n},\varepsilon}.
\end{align}
\end{definition}}
{The following statements are easily deduced from Definition \ref{dim}. }
\begin{proposition}
\label{résu}
For every $\mu\in\mathcal{M}(\mathbb{R}^d)$, $\rho>0$, every $0\leq \alpha \leq \gamma$ and $\varepsilon>0$,
\begin{align*}
\mu(E_{\mu}^{[\alpha,\gamma],\varepsilon}) & =\mu(\left\{x:\underline{\dim} (\mu,x)\in [\alpha,\gamma]\right\})\\
\mu(F_{\mu}^{[\alpha,\gamma],\varepsilon})& =\mu(\left\{x:\overline{\dim} (\mu,x)\in [\alpha,\gamma]\right\})
\end{align*}
and
\begin{align}
\label{mlecoro}
E_{\mu}^{[\alpha,\gamma],\rho,\varepsilon} & \subset \left\{x\in \mathbb{R}^d : \ \forall r\leq \rho, \ \mu(B(x,r))\leq r^{\alpha-\varepsilon}\right\}\\
\nonumber F_{\mu}^{[\alpha,\gamma],\rho,\varepsilon} & \subset \left\{x\in \mathbb{R}^d : \ \forall r\leq \rho, \ \mu(B(x,r))\geq r^{\gamma+\varepsilon}\right\}.
\end{align}
Furthermore, for $\alpha_1=\underline{\dim}_H (\mu)$ and $\gamma_1=\mbox{supess}_{\mu}(\underline{\dim} (\mu,x))$, one has
\begin{equation}
\label{mle}
\mu(E_{\mu}^{[\alpha_1,\gamma_1],\varepsilon})=1.
\end{equation}
Similarly, for $\alpha_2=\mbox{infess}_{\mu} (\overline{\dim} (\mu,x))$ and $\gamma_2=\overline{\dim}_P (\mu)$, one has
\begin{equation}
\label{mleup}
\mu\left(F_{\mu}^{[\alpha_2,\gamma_2],\varepsilon}\right)=1.
\end{equation}
\end{proposition}
\begin{proof}
{ For any $x \in \mathbb{R}^d$, for any $\varepsilon>0$, there exists $r_x >0$ such that, $\forall r\leq r_x $, $r^{\overline{\dim} (\mu,x)+\varepsilon}\leq \mu(B(x,r))\leq r^{\underline{\dim} (\mu,x)-\varepsilon}.$ This implies
\begin{equation*}
F_{\mu}^{[\alpha,\gamma],\rho,\varepsilon}\subset \left\{x\in \mathbb{R}^d : \ \forall r\leq \rho, \ \mu(B(x,r))\geq r^{\gamma+\varepsilon}\right\},
\end{equation*}
\begin{equation*}
E_{\mu}^{[\alpha,\gamma],\rho,\varepsilon}\subset \left\{x\in\mathbb{R}^d : \ \forall r\leq \rho, \ \mu(B(x,r))\leq r^{\alpha-\varepsilon}\right\},
\end{equation*}}
and
$$E_{\mu}^{[\alpha,\gamma],\varepsilon}=\left\{x:\underline{\dim} (\mu,x)\in [\alpha,\gamma]\right\} \text{ and } F_{\mu}^{[\alpha,\gamma],\varepsilon}=\left\{x:\overline{\dim} (\mu,x)\in [\alpha,\gamma]\right\}.$$
{
Since $$\mu([\mbox{infess}_{\mu} (\underline{\dim} (\mu,x)),\mbox{supess}_{\mu}(\underline{\dim} (\mu,x))])=1$$ and $$\mu([\mbox{infess}_{ \mu} (\overline{\dim} (\mu,x)),\mbox{supess}_{\mu}(\overline{\dim} (\mu,x))])=1,$$}
{ recalling Definition \ref{dim}, it holds that, following the notation of Proposition \ref{résu},
$$\mu\left(E_{\mu}^{[\alpha_1,\gamma_1],\varepsilon}\right)=\mu\left(F_{\mu}^{[\alpha_2,\gamma_2],\varepsilon}\right)=1.$$ }
\end{proof}
Before showing Proposition \ref{mainextraprop}, let us start by the two following Lemmas \ref{gscainf} and \ref{gscaup}. The first one will be used to prove the left part of the inequality \eqref{condimesu} while the second one will be useful to prove the right part.
\begin{lemme}
\label{gscainf}
Let $\mu \in\mathcal{M}(\mathbb{R}^d)$ and $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ be a $\mu$-a.c sequence of balls of $ \mathbb{R}^d$ with $\lim_{n\to +\infty} r_{n}= 0$.
For any $\varepsilon>0$, there exists a $\mu$-a.c subsequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ of $\mathcal{B} $ such that for every $n\in\mathbb{N}$, $\mu(B_{\phi(n)})\leq (r_{\phi(n)})^{ \underline{\dim}_H(\mu)-\varepsilon}.$
\end{lemme}
\begin{proof}
Set $\alpha= \underline{\dim}_H (\mu)$ and $\gamma=suppess_{\mu}(\underline{\dim} (\mu,x)).$
Let $\Omega$ be an open set and $\ep>0$. By \eqref{mle}, $\mu(E^{[\alpha,\gamma],\frac{\varepsilon }{2}}_{\mu })=1$ and $\mu ( \Omega\cap E^{[\alpha,\gamma],\frac{\varepsilon }{2}}_{\mu}) = \mu(\Omega)$.
For every $x\in \Omega\cap E^{[\alpha,\gamma],\frac{\varepsilon }{2}}_{\mu}$, there exists $r_x>0 $ such that $B(x,r_x) \subset \Omega$ and $x\in E_{\mu}^{[\alpha,\gamma],r_x ,\frac{\varepsilon}{2}}.$
Recall \eqref{emut} and that the sets $E_{\mu}^{[\alpha,\gamma], \rho ,\varepsilon}$ are non-increasing in $\rho$. In particular there exists $\rho_{\Omega}>0$ such that the set $E_\Omega:=\left\{x\in \Omega\cap E^{[\alpha,\gamma],\rho_{\Omega},\frac{\varepsilon }{2}}_{\mu} : \ r_x\leq \rho_\Omega \right\}$ verifies
\begin{equation}
\label{muomegainf}
\mu(E_\Omega ) \geq \frac{3\mu(\Omega)}{4}.
\end{equation}
Let $g\in\mathbb{N}$. Applying Lemma \ref{covO}
to $\Omega$, the sequence $(B_n)$ and the measure $m$, there exists $N_\Omega$ as well as $g\leq n_{1}\leq ... \leq n_{N_{\Omega}}$ verifying:
\begin{enumerate}
\item for every $ 1\leq i\neq j\leq N_\Omega,$ $B_{n_i}\cap B_{n_j}=\emptyset$,
\item for every $ 1\leq i\leq N_\Omega$, $ 2r_{n_i} \leq \rho_\Omega $ and $2^{\alpha -\frac{\varepsilon}{2}}\leq r_{n_i}^{-\frac{\varepsilon}{2}}$,
\item $\mu(\bigcup_{1\leq i\leq N_\Omega}B_{n_i})\geq \frac{\mu(\Omega)}{2}.$
\end{enumerate}
We may assume that $\mu(B_{n_i}) >0$ for every $i$, otherwise $B_{n_i}$ does not play any role.
Item $(3)$ together with \eqref{muomegainf} implies that
$$\mu\left (\bigcup_{1\leq i\leq N_\Omega}B_{n_i}\cap E_\Omega \right)\geq \frac{\mu(\Omega)}{4}.$$
Furthermore, for every $1\leq i\leq N_\Omega$ verifying $B_{n_i}\cap E_\Omega \neq \emptyset $, it holds that $0<\mu(B_{n_i}) \leq (r_{n_i})^{\alpha-\varepsilon }.$ Indeed, let $x\in B_{n_i}\cap E_\Omega$. By item (2), $B_{n_i} \subset B(x,2r_{n_i})$, and by \eqref{emuhat} , item $(2)$, and \eqref{mlecoro}, it holds that
$$\mu(B_{n_i})\leq \mu(B(x,2r_{n_i}))\leq (2r_{n_i})^{\alpha-\frac{\varepsilon}{2}}\leq (r_{n_i})^{\alpha-\varepsilon }.$$
Writing $\mathcal{B}'=\left\{B_n : \mu(B_{n})\leq r_n ^{\alpha-\varepsilon }\right\}$, the argument above shows that only balls of $\mathcal{B}'$ have been used to cover $\Omega$ . This is satisfied for every open set $\Omega$, so that $\mathcal{B}'$ is a sub-sequence of $\mathcal{B}$ satisfying the condition of Definition \ref{ac}, which concludes the proof of Lemma \ref{gscainf}.
\end{proof}
\begin{lemme}
\label{gscaup}
Let $\mu \in\mathcal{M}(\mathbb{R}^d)$, $v<1$ and $\mathcal{B} =(B_n :=B(x_n ,r_n))_{n\in\mathbb{N}}$ a sequence of balls of $ \mathbb{R}^d$ verifying $\mu(\limsup_{n\to +\infty}vB_n)=1$.
For all $\varepsilon>0$, there exists a sub-sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ of $\mathcal{B} $ as well as $0<v^{\prime}<1$ such that $\mu(\limsup_{n\rightarrow+\infty}v^{\prime}B_{\phi(n)})=1$ and for all $n\in\mathbb{N}$, one has $\mu(B_{\phi(n)})\geq (r_{\phi(n)})^{\overline{\dim}_H (\mu)+\varepsilon}$.
\end{lemme}
\begin{remarque}
The sequence $(B_{\phi(n)})_{n\in\mathbb{N}}$ found in Lemma \ref{gscaup} is in particular $\mu$-a.c by Theorem \ref{equiac}.
\end{remarque}
\begin{proof}
Let $\alpha =\mbox{infess}_{\mu} (\overline{\dim} (\mu,x))$ and $\gamma=\overline{\dim}_P (\mu).$
Let $\ep>0$ and $v< v^{\prime}<1$.
By \eqref{mleup} and Theorem \ref{equiac}, $\mu(\limsup_{n\rightarrow+\infty}vB_n \cap F_{\mu}^{[\alpha,\gamma],\frac{3\varepsilon}{2}})=1.$ For all $x\in\limsup_{n\rightarrow+\infty}vB_n \cap F^{[\alpha,\gamma],\frac{3\varepsilon}{2}}_{\mu}$, there exists $r_x>0$ small enough so that
\begin{equation}
\label{randcond}
r_x ^{\frac{\varepsilon}{2}}\leq {(v^{\prime}-v)}^{\gamma +\frac{3\varepsilon}{2}} \text{ and } \forall 0<r\leq r_x, \ \mu(B(x,r))\geq r^{\gamma +\frac{3\varepsilon}{2}}.
\end{equation}
Since $x\in \limsup_{n\rightarrow+\infty}vB_n$, for all $n\in\mathbb{N}$, there exists $n_x \geq n$ such that $x\in vB_{n_x}$ and $ (v^{\prime}-v)r_{n_x} \leq r_x .$ Note that $B(x,(v^{\prime}-v)r_{n_x})\subset v^{\prime}B_{n_x}.$ This implies the following inequalities:
$$\mu(B_{n_x})\geq \mu(v^{\prime}B_{n_x})\geq \mu(B(x,(v^{\prime}-v)r_{n_x})\geq ((v^{\prime}-v)r_{n_x})^{\gamma+ \frac{3\varepsilon}{2}}\geq r_{n_x}^{\gamma+2\varepsilon}.$$
Set $\mathcal{B}_{\gamma,2\varepsilon}=\left\{B_n : \mu(B_n)\geq r_n ^{\gamma+2\varepsilon}\right\}$.
One just showed that
$$\limsup_{n\rightarrow+\infty}vB_n \cap F^{[\alpha,\gamma],\frac{\varepsilon}{2}}\subset\limsup_{B\in\mathcal{B}_{\gamma,2\varepsilon}}v^{\prime}B.$$
This proves that $\mu(\limsup_{B\in\mathcal{B}_{\gamma,2\varepsilon}}v^{\prime}B)=1$.
Since $\varepsilon>0$ was arbitrary, the results also holds with $\frac{\varepsilon}{2}$, which proves Lemma \ref{gscaup}.
\end{proof}
We are now ready to prove Proposition \ref{mainextraprop}.
\begin{proof}
Set $\alpha=\underline{\dim}_H (\mu)$ and $\beta=\overline{\dim}_H (\mu).$
Let us fix $(\varepsilon_n)_{n\in\mathbb{N}}\in(\mathbb{R}^{*+})^{\mathbb{N}}$ verifying $\lim_{n\rightarrow+\infty}\varepsilon_n =0.$
The strategy of the proof consists in constructing recursively coverings of the cube $\mathbb{R}^d$ by using Lemma \ref{gscainf} and Lemma \ref{gscaup} and a diagonal argument (on the choice of $\varepsilon$) at each step.
More precisely, at step 1, one will build a family of balls $(\mathcal{F}_{1,i})_{i\in\mathbb{N}}$ verifying:
\begin{itemize}
\item[•] for all $i,j\geq 1$, $\forall L\in\mathcal{F}_{1,i}$, $\forall L^{\prime}\in\mathcal{F}_{1,j}$ such that $L\neq L^{\prime},$ one has $L\cap L^{\prime}=\emptyset,$\medskip
\item[•] for all $i\geq 1$, $\mathcal{F}_{1,i}$ is a finite sub-family of $\left\{B_n\right\}_{n\geq 1},$\mk
\item[•] for all $i \geq 1$, for all $L\in\mathcal{F}_{1,i}$, $\vert L\vert ^{\beta+\varepsilon_i}\leq \mu(L)\leq \vert L\vert^{\alpha-\varepsilon_i},$\mk
\item[•] $\mu\left(\bigcup_{i\in\mathbb{N}}\bigcup_{L\in\mathcal{F}_{1,i}}L\right)=1.$
\end{itemize}
Note that for each $i\in\mathbb{N}$, only a finite number of balls $L \in\mathcal{F}_1 :=\bigcup_{j\in\mathbb{N}} \mathcal{F}_{1,j}$ verifies (for that $\varepsilon_i$ naturally associated with those balls) $\vert L\vert ^{\beta+\varepsilon_i}\leq \mu(L)\leq \vert L\vert^{\alpha-\varepsilon_i}$.
At step 2, a family of balls $(\mathcal{F}_{2,i})_{i\in\mathbb{N}}$ will be constructed such that:
\begin{itemize}
\item[•] for all $i,j\geq 1$, $L\in\mathcal{F}_{2,i}$, $L^{\prime}\in\mathcal{F}_{2,j}$, $L\neq L^{\prime}=\emptyset,$\medskip
\item[•] for all $i\geq 1$, $\mathcal{F}_{2,i}$ is a finite sub-family of $\left\{B_n\right\}_{n\geq 2},$\mk
\item[•] for all $i \geq 1$, for all $L\in\mathcal{F}_{2,i}$, $\vert L\vert ^{\beta+\varepsilon_{i+1}}\leq \mu(L)\leq \vert L\vert^{\alpha-\varepsilon_{i+1}},$
\item[•] one has
\begin{equation}
\label{eqfulmes}
\mu\left(\bigcup_{i\in\mathbb{N}}\bigcup_{L\in\mathcal{F}_{2,i}}L\right)=1.
\end{equation}
\end{itemize}
Write $\mathcal{F}_2 =\bigcup_{i \geq 1}\mathcal{F}_{2,i}$. Note that the family of balls $\mathcal{F}_2$ verifies, by construction, that any $L\in\mathcal{F}_{2}$ the natural $\varepsilon_i$ associated with $L$ is never equal to $\varepsilon_1$, so that only some balls constructed in step 1 are associated with $\varepsilon_1$ .
The other steps are achieved following the same scheme.
The construction is detailed below:
\mk
\textbf{Step 1:}
\mk
Let $\Omega_{1,1}=\mathbb{R}^d .$
\mk
\textbf{Sub-step 1.1:}
\mk
By Lemma \ref{gscainf} and Lemma \ref{gscaup} applied to $\varepsilon=\varepsilon_1$, there exists a $\mu$-a.c sub-sequence $(B_{\psi_{1,1}(n)})_{n\in\mathbb{N}}$, satisfying, for every $n\in\mathbb{N}$,
$$\vert B_{\psi_{1,1}(n)}\vert^{\beta+\varepsilon_1}\leq \mu(B_{\psi_{1,1}(n)})\leq \vert B_{\psi_{1,1}(n)} \vert^{\alpha-\varepsilon_1}.$$
By Lemma \ref{covO} applied to $\Omega_{1,1}$, the sequence $(B_{\psi_{1,1}(n)})_{n\in\mathbb{N}}$ and $g=1$, there exists an integer $N_{1,1}$ as well as some balls $L_{1,1,1} ,...,L_{1,1,N_{1,1}}\in\left\{B_n\right\}_{n\geq 1}$ verifying:
\begin{itemize}
\item[•] for all $1\leq i<j\leq N_{1,1}$, $L_{1,1,i} \cap L_{1,1,j}=\emptyset,$\mk
\item[•] for all $1\leq i\leq N_{1,1}$, $\vert L_{1,1,i}\vert^{\beta+\varepsilon_1}\leq \mu(L_{1,1,i})\leq \vert L_{1,1,i}\vert^{\alpha-\varepsilon_1},$\mk
\item[•] $\mu(\bigcup_{1\leq i\leq N_{1,1}}L_{1,1,i})\geq \frac{1}{2}.$\mk
\end{itemize}
Set $\mathcal{F}_{1,1}=\left\{L_{1,1,i}\right\}_{1\leq i\leq N_{1,1}}.$
\mk
\textbf{Sub-step 1.2:}
\mk
Let $\Omega_{1,2}=\Omega_{1,1}\setminus \bigcup_{L\in\mathcal{F}_{1,1}}L.$
By Lemma \ref{gscainf} and Lemma \ref{gscaup} with $\varepsilon=\varepsilon_2$, there exists a $\mu$-a.c sub-sequence $(B_{\psi_{1,2}(n)})_{n\in\mathbb{N}}$ satisfying
$$\vert B_{\psi_{1,2}(n)}\vert^{\beta+\varepsilon_2}\leq \mu(B_{\psi_{1,2}(n)})\leq \vert B_{\psi_{1,2}(n)} \vert^{\alpha-\varepsilon_2}.$$
One applies Lemma \ref{covO} to the open set $\Omega_{1,2}$, the sub-sequence of balls $(B_{\psi_{1,2}(n)})_{n\in\mathbb{N}}$ and $g=1$. There exists $N_{1,2}\in\mathbb{N}$ such that $L_{1,2,1},...,L_{1,2,N_{1,2}}$ verifies:
\begin{itemize}
\item[•] for all $1\leq i<j\leq N_{1,2}$, $L_{1,2,i} \cap L_{1,2,j}=\emptyset,$\mk
\item[•] for all $1\leq i\leq N_{1,2}$, $\vert L_{1,2,i}\vert^{\beta+\varepsilon_2}\leq \mu(L_{1,2,i})\leq \vert L_{1,2,i}\vert^{\alpha-\varepsilon_2},$\mk
\item[•] $\mu(\bigcup_{1\leq i\leq N_{1,2}}L_{1,2,i})\geq \frac{1}{2}\mu(\Omega_{1,2}).$\mk
\end{itemize}
The family $\mathcal{F}_{1,2}$ is defined as
$\mathcal{F}_{1,2}=\left\{L_{1,2,i}\right\}_{1\leq i\leq N_{1,2}}.$
Proceeding iteratively as Sub-step $1.1$ and Sub-step $1.2$, for any $i\in\mathbb{N}$, at Sub-step $1.i$ a family of balls $(\mathcal{F}_{1,i})_{i\in\mathbb{N}}$ is constructed so that it verifies:
\begin{itemize}
\item[•] for all $i,j\geq 1$, $L\in\mathcal{F}_{1,i}$, $L^{\prime}\in\mathcal{F}_{1,j}$, if $L\neq L^{\prime},$ then $L\cap L^{\prime}=\emptyset,$\medskip
\item[•] for all $i\geq 1$, $\mathcal{F}_{1,i}$ is a finite subset of $\left\{B_n\right\}_{n\geq 1},$\mk
\item[•] for all $i \geq 1$, for all $L\in\mathcal{F}_{1,i}$, $\vert L\vert ^{\beta+\varepsilon_i}\leq \mu(L)\leq \vert L\vert^{\alpha-\varepsilon_i},$
\item[•] $\mu\left(\bigcup_{i\in\mathbb{N}}\bigcup_{L\in\mathcal{F}_{1,i}}L\right)=1.$
\end{itemize}
Recall that, to justify the last item, this recursive scheme allows to cover $\mathbb{R}^d$, up to a set of $\mu$-measure 0 (the argument is similar to the one developed at the end of the proof of Lemma \ref{covO} to obtain \eqref{usefulmath}).
Set $\mathcal{F}_1 =\bigcup_{i\geq 1}\mathcal{F}_{1,i}.$ With each ball $L\in\mathcal{F}$ is naturally associated a positive real number $\varepsilon(L)$, such that $\varepsilon(L)=\varepsilon_i $ if $L\in\mathcal{F}_{1,i}.$
Let us notice that the construction of the family $\mathcal{F}_2$ does not rely on the existence of the family $\mathcal{F}_1$, so that the families $\mathcal{F}_k$ can actually be built independently, following the same scheme, as described below.
\mk
\textbf{Step $k$:}
\mk
As in step 1, one constructs a family of balls $(\mathcal{F}_{k,i})_{i\geq 1}$ verifying:
\begin{itemize}
\item[•] for all $i,j\geq 1$, $L\in\mathcal{F}_{k,i}$, $L^{\prime}\in\mathcal{F}_{k,j}$, $L\neq L^{\prime},$ $L\cap L^{\prime}=\emptyset,$\medskip
\item[•] for all $i\geq 1$, $\mathcal{F}_{k,i}$ is a finite subset of $\left\{B_n\right\}_{n\geq k},$\mk
\item[•] for all $i \geq 1$, for all $L\in\mathcal{F}_{k,i}$, $\vert L\vert ^{\beta+\varepsilon_{i+k}}\leq \mu(L)\leq \vert L\vert^{\alpha-\varepsilon_{i+k}},$\mk
\item[•] one has
\begin{equation}
\label{eqfulmesbis}
\mu\left(\bigcup_{i\in\mathbb{N}}\bigcup_{L\in\mathcal{F}_{k,i}}L\right)=1.
\end{equation}
\end{itemize}
Set $\mathcal{F}_{k}=\bigcup_{i\geq 1}\mathcal{F}_{k,i}$ and
$\mathcal{F}=\bigcup_{k\geq 1}\mathcal{F}_k.$
Denote by $(B_{ \phi(n)})_{n\in\mathbb{N}}$ the sub-sequence of balls that constitutes the family $\mathcal{F}.$
By construction, for all $i\in\mathbb{N}$, only a finite number of balls $L\in\mathcal{F}$ verifies $\varepsilon(L)=\varepsilon_i$ (and $\vert L\vert ^{\beta+\varepsilon_i}\leq \mu(L)\leq \vert L\vert^{\alpha -\varepsilon_i}$). In particular, for all $\varepsilon>0$, there exists $N$ large enough so that, for every $n\geq N$, $\varepsilon_n \leq \varepsilon$. Similarly, there exists $N^{\prime}\in\mathbb{N}$ so large that for every $n^{\prime}\geq N^{\prime}$,
$$\vert B_{\phi(n^{\prime})}\vert^{\beta+\varepsilon}\leq \mu(B_{ \phi(n^{\prime})})\leq \vert B_{\phi(n^{\prime})} \vert^{\alpha-\varepsilon}.$$
It follows that
$$\alpha-\varepsilon\leq \liminf_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log\vert B_{\phi(n)}\vert}\leq \limsup_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log \vert B_{\phi(n)}\vert}\leq \beta +\varepsilon.$$
Letting $\varepsilon\to 0$ shows that
$$\underline{\dim}_H (\mu)\leq \liminf_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log\vert B_{\phi(n)}\vert}\leq \limsup_{n\rightarrow+\infty}\frac{\log \mu(B_{\phi(n)})}{\log \vert B_{\phi(n)}\vert}\leq \overline{\dim}_H (\mu).$$
It only remains to prove that $(B_{\phi(n)})_{n\in \mathbb{N}}$ is $\mu$-a.c.
Let $\Omega $ be an open set and $g\in\mathbb{N}$. We find a finite family of balls $\left\{L\right\}_{i\in\mathcal{I}}\subset \left\{B_{\phi(n)}\right\}_{n\geq g}$ satisfying the conditions of Definition \ref{ac}.
Note that, by \eqref{eqfulmes},
$$\mbox{setting } E=\bigcap_{k\geq 1}\bigcup_{L\in\mathcal{F}_k}L, \ \ \mbox{ then } \ \mu\left(E\right)=1.$$
Let $x\in\Omega \cap E$ and $r_x >0$ small enough so that $B(x,r_x)\subset \Omega .$ Consider $k_x \geq \phi(g)\geq g$ large enough so that, for all $n\geq k_x$, $\vert B_n \vert \leq 2 r_x$. Recall that $\mathcal{F}_{k_x}\subset \left\{B_n\right\}_{n\geq k_x}.$ Finally, let us fix $k$ large enough so that $\mu(\widehat{E})\geq \frac{\mu(\Omega)}{2},$ where $\widehat{E}=\left\{x\in E \ : \ k_x \leq k\right\}$. For $x\in\widehat{E}$, let $L_x \in\mathcal{F}_k$ be the ball that contains $x$ (the balls of $\mathcal{F}_k$ being pairwise disjoint, $L_x$ is well defined) and $\left\{L_i\right\}_{i\geq 1}=\left\{L_x : x\in\widehat{E}\right\}$. One has
\begin{itemize}
\item[•] for all $1\leq i<j$, $L_i \cap L_j =\emptyset,$
\item[•] for all $i\in\mathbb{N}$, $L_i \in\left\{B_{\phi(n)}\right\}_{n\geq g}$ and $L_i \subset \Omega,$
\item[•] $\mu(\bigcup_{i \geq 1}L_i)\geq \mu(\widehat{E})\geq \frac{\mu(\Omega)}{2}.$
\end{itemize}
By $\sigma$-additivity, there exists $N\in\mathbb{N}$ such that $\mu(\bigcup_{1\leq i\leq N}L_i)\geq \frac{\mu(\Omega)}{4}$, which proves that $(B_{\phi(n)})_{n\in\mathbb{N}}$ satisfies Definition \ref{ac} with $C=\frac{1}{4}$ and is indeed $\mu$-a.c.
\end{proof}
One finishes this section with the following proposition, which supports the idea that, roughly speaking, for an $\alpha$ exact-dimensional measure $\mu$ and a $\mu$-a.c sequence of balls $(B_n)$, considering balls $(B_n)_{n\in\mathbb{N}}$ which does not verify $\mu(B_n)\approx \vert B_n \vert^{\alpha}$ is not relevant from the $\mu$-standpoint.
\begin{proposition}
Let $\mu\in\mathcal{M}( \mathbb{R}^d)$ be an $\alpha$ exact-dimensional measure and $(B_n)_{n\in\mathbb{N}}$ a sequence of balls satisfying $\vert B_n \vert\to 0.$ Let $\varepsilon>0.$ Let is also define $\mathcal{B}_{>}^{\varepsilon}=\left\{B_n : \mu(B_n)\leq \vert B_n \vert ^{\alpha+\varepsilon}\right\}$ and $\mathcal{B}_{<}^{\varepsilon}=\left\{B_n : \mu(B_n)\geq \vert B_n \vert ^{\alpha-\varepsilon}\right\}.$ Then
\begin{enumerate}
\item for any $v<1$, $\mu(\limsup_{B\in\mathcal{B}_{>}^{\varepsilon}}vB)=0,$\mk
\item $\mu(\limsup_{B\in\mathcal{B}_{<}^{\varepsilon}}B)=0.$
\end{enumerate}
\end{proposition}
\begin{proof}
(1) Suppose that there exists $0<v<1$ such that $\mu(\limsup_{B\in\mathcal{B}_{>}^{\varepsilon}}vB)>0.$ Then, since $\mu$ is assumed to be exact-dimensional, there exists $x\in\limsup_{B\in\mathcal{B}_{>}^{\varepsilon}}vB$ such that $\lim_{r\to 0}\frac{\log \mu(B(x,r))}{\log r}=\alpha.$
Consider $r_x>0$ small enough so that, for any $0<r\leq r_x$, $\mu(B(x,r))\geq r^{\dim (\mu)+\frac{ \varepsilon}{2}}$ and $(\frac{1-v}{2})^{\dim (\mu)+\frac{\varepsilon}{2}}\geq r_x ^{\frac{\varepsilon}{4}}.$ Let also $n$ be large enough so that $x\in B_n$ and $\vert B_n \vert \leq r_x .$ Then $ B(x,\frac{(1-v)}{2} \vert B_n \vert)\subset B_n$, so that
\begin{align}
\label{cone}
\mu(B_n)\geq \mu(B(x,\frac{1-v}{2}\vert B_n \vert))\geq \vert B_n \vert^{\alpha+\frac{\varepsilon}{2}}(\frac{1-v}{2})^{\alpha+\frac{\varepsilon}{2}}\geq \vert B_n \vert^{\alpha+\frac{3\varepsilon}{4}}.
\end{align}
This contradicts the definition of $\mathcal{B}_{>}^{\varepsilon}.$
\medskip
\noindent(2) Assume that $\mu(\limsup_{B\in\mathcal{B}_{<}^{\varepsilon}}B)>0.$ Then, again, there exists $x\in\limsup_{B\in\mathcal{B}_{<}^{\varepsilon}}B$ so that $\lim_{r\to 0}\frac{\log \mu(B(x,r))}{\log r}=\alpha.$ Consider $r_x >0$ small enough so such that, for any $0<r\leq r_x$, $\mu(B(x,r))\leq r^{\alpha-\frac{ \varepsilon}{2}}.$ Consider $n\in\mathbb{N}$ large enough so that $x\in B_n$ and $\vert B_n \vert \leq r_x.$ One has $B_n \subset B(x,\vert B_n \vert)$, hence
\begin{align*}
\mu(B_n)\leq \mu(B(x,\vert B_n \vert))\leq \vert B_n \vert^{\alpha-\frac{\varepsilon}{2}}.
\end{align*}
This contradicts the definition of $\mathcal{B}_{<}^{\varepsilon}.$
\end{proof}
\begin{remarque}
For doubling measures, it is straightforward that item $(1)$ can be replaced by simply $\mu(\limsup_{B\in\mathcal{B}_{>}^{\varepsilon}}B)=0.$ It can be proved that this is also the case for 1-average d-1 unrectifiable measures (as a consequence of \cite[Theorem 2.11]{KS}). Some self-similar measures with open set condition satisfies this property (see \cite{KS} again for more details).
\end{remarque}
\
\section{Some explicit examples}
\label{sec-exam}
In this section, applications of Theorem \ref{theoremextra} are given.
\subsection{Rational approximation}
Let us recall the following result from Hurwitz see \cite{HW}, p 219, for more details.
\begin{theoreme}
\label{hure}
Let $x\in[0,1]\setminus \mathbb{Q}.$ There exists an infinite number of pairs $(p,q)\in\mathbb{N}\times \mathbb{N}^*$ with $p\wedge q=1$ and
\begin{equation}
\label{hur}
\Big\vert x -\frac{p}{q}\Big\vert<\frac{1}{\sqrt{5}q^2}.
\end{equation}
\end{theoreme}
An immediate corollary of Theorem \ref{equiac}, Theorem \ref{theoremextra} and Theorem \ref{hure} is the following:
\begin{corollary}
Let $\mu \in\mathcal{M}([0,1])$ be any diffuse measure. Then the sequence of balls $\Big(B(\frac{p}{q},\frac{1}{q^2})\Big)_{0\leq p\leq q , q\in\mathbb{N}^* ,p\wedge q =1}$ is weakly redundant (see \cite{BS2}) and $\mu$-a.c. In particular, if $\mu$ is $\alpha$-exact-dimensional, for $0\leq \alpha\leq d$, then there exists a sequence $(\varepsilon_n )_{n\in\mathbb{N}}\in(\mathbb{R}^* _+ )^{\mathbb{N}} $ with $ \lim_{n\rightarrow+\infty}\varepsilon_n =0$ and an infinite number of pairs $(p_n ,q_n)\in\Big(\mathbb{N}\times\left\{0,...,q_n\right\}\Big)^{\mathbb{N}} $ such that $p_n \wedge q_n =1$ and $$\Big(\frac{1}{q_n^{2}}\Big)^{\alpha+\varepsilon_n}\leq\mu\Big(B\Big(\frac{p_n}{q_n},\frac{1}{q_n ^2}\Big)\Big)\leq \Big(\frac{1}{q_n^{2}}\Big)^{\alpha-\varepsilon_n}.$$
Moreover, writing $$\mathcal{B}_{\mu}=\left\{ B(\frac{p_n}{q_n},\frac{1}{q_n ^2}):\Big(\frac{1}{q_n^{2}}\Big)^{\alpha+\varepsilon_n}\leq\mu\Big(B\Big(\frac{p_n}{q_n},\frac{1}{q_n ^2}\Big)\Big)\leq \Big(\frac{1}{q_n^{2}}\Big)^{\alpha-\varepsilon_n}\right\}_{n\in\mathbb{N}},$$ one has
$$\dim_H(\limsup_{B\in\mathcal{B}_{\mu}}B)=\alpha.$$
\end{corollary}
\subsection{Application to Random balls}
Let us recall Shepp's Theorem of Shepp \cite{shepp1972randomcovering}.
\begin{theoreme}
\label{she}
Let $(l_n)_{n\in\mathbb{N}}\in(\mathbb{R}_+ ^{*} )^{\mathbb{N}}$ and $(X_n )_{n\in\mathbb{N}}$ be a sequence of i.i.d uniformly distributed random variables on $[0,1]$. Then
$$
\limsup_{n\rightarrow+\infty}B(X_n, l_n)=[0,1] \Leftrightarrow \sum_{n\geq 0}\frac{1}{n^2}\exp(l_1 +...+l_n)=+\infty. $$
\end{theoreme}
From Theorem \ref{equiac}, Theorem \ref{theoremextra} and Theorem \ref{she}, the following corollary is deduced.
\begin{corollary}
For any $\alpha$ exact-dimensional measure $\mu \in\mathcal{M}([0,1])$, $0\leq\alpha\leq d$, for almost any i.i.d sequence of random variables uniformly distributed on $[0,1]$, $(X_n)_{n\in\mathbb{N}}$, there exists a sequence of positive real numbers $(\varepsilon_k )_{k\in\mathbb{N}}$ with $\varepsilon_k \to 0$ and a subsequence $(n_k)_{k\in\mathbb{N}}\to+\infty$ satisfying
$$\Big(\frac{2}{n_k} \Big)^{\alpha+\varepsilon_k}\leq \mu\Big(B\Big(X_{n_k} ,\frac{2}{n_k} \Big)\Big)\leq \Big(\frac{2}{n_k} \Big)^{\alpha-\varepsilon_k}.$$
Writing again $$\mathcal{B}_{\mu}=\left\{B\Big(X_{n_k} , \frac{2}{n_k}\Big):\Big(\frac{2}{n_k}\Big)^{\alpha+\varepsilon_k}\leq \mu\Big(B\Big(X_{n_k} ,\frac{2}{n_k} \Big)\Big)\leq \Big(\frac{2}{n_k} \Big)^{\alpha-\varepsilon_k}\right\}_{k\in\mathbb{N}},$$
one has $$\dim_H (\limsup_{B\in\mathcal{B}_{\mu}}B)=\alpha.$$
\end{corollary}
\subsection{Examples in dynamical systems}
Let us introduce some notation.
Let $m\geq 2$ and $S=\left\{f_1 ,...,f_m\right\}$ be a system of $m$ similarities of $\mathbb{R}^d \to \mathbb{R}^d$ of ratio of contraction $0<c_1 <1 ,...,0<c_m <1.$
Let us also write $\Lambda=\left\{1,...,n\right\}$, $\Lambda^{*}=\bigcup_{k\geq 0}\Lambda^k$ and for $\underline{i}=(i_1,...,i_k)\in\Lambda^k$,
\begin{itemize}
\item[•]$f_{\underline{i}}=f_{i_1}\circ ... \circ f_{i_k},$\mk
\item[•] $X_{\underline{i}}=f_{\underline{i}}([0,1]^d),$\mk
\item[•]$c_{\underline{i}}=c_{i_1}\times ... \times c_{i_k}.$
\end{itemize}
Let us fix also $(p_1,...,p_m)\in\mathbb{R}_+ ^{\mathbb{N}}$ a probability vector, i.e a vector verifying $\sum_{1\leq i\leq m}p_i =1$.
\begin{remarque}
\label{covks}
Let $\mu$ defined by \eqref{equass} and $x\in K_S$. Then, for any $k\in\mathbb{N}$, the balls $\left\{B(f_{\underline{i}}(x),2\vert K_S \vert c_{\underline{i}})\right\}_{\underline{i}\in\Lambda^k}$ covers $K_S$. In particular, by Theorem \ref{equiac} $\Big(B(f_{\underline{i}}(x),3\vert K_S \vert c_{\underline{i}})\Big)_{\underline{i}\in\Lambda^*}$ is $\mu$-a.c.
\end{remarque}
As a consequence of Theorem \ref{theoremextra} and Lemma \ref{covks}, one gets:
\begin{corollary}
Let $\mu\in\mathcal{M}(\mathbb{R}^d)$ be a measure defined by \eqref{equass} and $x\in K_S$. There exists a $\mu$-a.c weakly redundant sub-sequence of balls $(B_{n})_{n\in\mathbb{N}}$ extracted from $\Big(B(f_{\underline{i}}(x),3c_{\underline{i}})\Big)_{\underline{i}\in\Lambda^*}$ such that, for all $n\in\mathbb{N}$ and for some sequence $(\varepsilon_k)_{k\in\mathbb{N}}\in (\mathbb{R}^{*}_+ )^{\mathbb{N}}$ verifying $\varepsilon_k \to 0$,
$$\vert B_n \vert^{\dim(\mu)+\varepsilon_n}\leq \mu(B_n)\leq \vert B_n \vert^{\dim(\mu)-\varepsilon_n},$$
and $\mathcal{B}_{\mu}=\left\{B_n \right\}_{n\in\mathbb{N}}$ satisfies
$$ \dim_H (\limsup_{n\rightarrow+\infty}B_n)=\dim (\mu).$$
\end{corollary}
\
\section{Proofs of Theorem \ref{majoss}}
\label{sec-upper}
The proof strongly relies on the following result proved in \cite{ED3}.
\begin{theoreme}
\label{contss}
Let $S$ be a self-similar IFS of $\R^d$. Let $K$ be the attractor of $S$. Let~$\mu$ be a self-similar measure associated with $S$. For any $0\leq s<\dim(\mu)$, there exists a constant $c=c(d,\mu,s)>0 $ depending on the dimension $d$, $\mu$ and $s$ only, such that for any ball $B=B(x,r)$ centered on $K$ and $r\leq 1$, any open set $\Omega$, one has
\begin{align}
\label{genhcontss}
&c(d,\mu,s)\vert B\vert ^{s}\leq\mathcal{H}^{\mu, s }_{\infty}(\widering{B})\leq \mathcal{H}^{ \mu, s}_{\infty}(B)\leq\vert B\vert ^{s}\text{ and } \nonumber\\
&c(d,\mu,s)\mathcal{H}^{s}_{\infty}(\Omega \cap K)\leq \mathcal{H}^{\mu,s}_{\infty}(\Omega)\leq \mathcal{H}^{s}_{\infty}(\Omega \cap K).
\end{align}
For any $s>\dim(\mu)$, $ \mathcal{H}^{\mu,s}_{\infty}(\Omega)=0.$
\end{theoreme}
Recall that the sequence $\mathcal{B}=(B_n)_{n\in\mathbb{N}}$ is assumed to be weakly redundant. In such a case, for any $\varepsilon>0$, following the notation involved in Definition \ref{wr}, it holds that
\begin{align*}
\sum_{n\geq 0}\vert B_n \vert^{\varepsilon}\mu(B_n)=\sum_{k\geq 0}\sum_{B\in\mathcal{T}_k (\mathcal{B})}\vert B \vert^{\varepsilon}\mu(B)\leq \sum_{k\geq 0}\sum_{1\leq j\leq J_k}2^{-k\varepsilon}\sum_{B\in T_{k,j} (\mathcal{B})} \mu(B).
\end{align*}
Since for every $(k,j)$ the family $T_{k,j}$ is composed of pairwise disjoint balls, this yields
\begin{equation}
\label{majowr}
\sum_{n\geq 0}\vert B_n \vert^{\varepsilon}\mu(B_n)\leq \sum_{k\geq 0}\sum_{1\leq j\leq J_k}2^{-k\varepsilon}=\sum_{k\geq 0}J_k 2^{-k\varepsilon}<+\infty.
\end{equation}
Now, for $n \in\mathbb{N}$, consider a sequence of balls $(A_k ^{n})_{k\in\mathbb{N}}$, with $\vert A_k ^n \vert \leq \vert B_n \vert$ and such that $U_n \cap K\subset \bigcup_{k\geq 0}A_k ^n$. Recall Theorem \ref{contss} and its notations. One has.
\begin{equation}
\label{equapmaj1}
\mathcal{H}^s_{\infty}(U_n)\leq \sum_{k\geq 0}\vert A_{k}^n \vert^s \leq 2\mathcal{H}^s_{\infty}(U_n)\leq \frac{2}{c(d,\mu,s)}\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \frac{2}{c(d,\mu,s)}\mu(B_n).
\end{equation}
Since for each $n\in\mathbb{N}$, $U_n \cap K \subset \bigcup_{k\geq 0}A_k ^n$, it holds that $\limsup_{n\rightarrow+\infty}U_n \cap K\subset \limsup_{k,n \rightarrow+\infty}A_k ^n .$ For any $\varepsilon>0$, one gets
\begin{align*}
\sum_{n\geq 0}\sum_{k\geq 0}\vert A_k ^n\vert^{s+\varepsilon}\leq \sum_{n\geq 0}\vert B_n \vert^{\varepsilon}\frac{2}{c(d,\mu,s)}\mu(B_n).
\end{align*}
In particular, by \eqref{majowr},
\begin{equation}
\sum_{n\geq 0}\sum_{k\geq 0}\vert A_k ^n\vert^{s+\varepsilon}<+\infty.
\end{equation}
One concludes that
$$\mathcal{H}^{s+\varepsilon}(\limsup_{n\rightarrow+\infty}U_n \cap K)\leq \mathcal{H}^{s+\varepsilon}(\limsup_{k,n\rightarrow+\infty}A_k^n)<+\infty .$$
This implies that $\dim_H (\limsup_{n\rightarrow+\infty}U_n \cap K)\leq s+\varepsilon$ and $\varepsilon$ being arbitrary, $$\dim_H (\limsup_{n\rightarrow+\infty}U_n \cap K)\leq s .$$
\begin{remarque}
\label{remajo1}
\item[•] An important fact to underline here is that the convergences established in \eqref{majowr} and \eqref{equapmaj1} do not rely on the fact that the measure is self-similar, but hold for any measure $\mu$. One could state a comparable upper-bound Theorem for any measure $\mu$ by replacing $K$ by a $G_{\delta}$ set of full measure in \eqref{uppernound18}.
These computations also have the following straightforward consequence for a measure $\mu\in\mathcal{M}(\mathbb{R}^d)$ without the self-similarity assumption: Assume that, for $n$ large enough, $\mathcal{H}^{\mu,s}_{\infty}(U_n)\leq \mu(B_n).$ If the sequence $(U_n)_{n\in\mathbb{N}}$ verifies that for any ball $B_i \subset U_n$ one also has $B_i \subset \bigcup_{k\geq n}A_k^n$ (where the balls $(A_{k,n})_{k\in\mathbb{N}}$ are chosen as in the proof of Theorem \ref{majoss}), then $\dim_H (\limsup_{n\rightarrow+\infty}U_n)\leq s$. In particular if this holds for any $s>s(\mu,\mathcal{B},\mathcal{U}),$ then $\dim_H (\limsup_{n\rightarrow+\infty}U_n)=s(\mu,\mathcal{B},\mathcal{U}).$\medskip
\item[•] When the self-similar measure verifies $\supp (\mu)=[0,1]^d$, the existence of $s_0$ as in Corollary \ref{equass1} is ensured as soon as the shapes of the sets $U_n$ are ``uniform'' in $n$. For instance, consider the case where $\mu=\mathcal{L}^d$ and $(U_n =R_n )_{n\in\mathbb{N}}$, where $R_n$ is an open rectangle associated with some vector $\boldsymbol{\tau}=(\tau_1 ,...,\tau_d)$ defined as in Theorem \ref{rectssmajo}. Recall that by Theorem \ref{contss}, the Lebesgue essential Hausdorff content and the classical Hausdorff content are equivalent. It is easily verified that, for any $n\in\mathbb{N}$, $\mathcal{H}^{s}_{\infty}(R_n)=\vert B_n \vert^{g_{\boldsymbol{\tau}}(s)}$, for the mapping $g_{\boldsymbol{\tau}}:\mathbb{R}^+ \to \mathbb{R}^+$, defined as (see \cite{KR})
$$ g_{\boldsymbol{\tau}}(s)= \max_{1\leq k\leq d}\left\{s\tau_k -\sum_{1\leq i\leq k}\tau_k-\tau_i\right\}.$$
Note that $g_{\boldsymbol{\tau}}(s) $ does not depend on $n$. Corollary can therefore be applied with $s_0=\min_{s: g_{\boldsymbol{\tau}}(s)\geq d}\left\{s\right\}$. \mk
\item[•] Unfortunately, when such an $s_0$ does not exist, the Hausdorff dimension of $\limsup_{n\rightarrow+\infty}U_n$ has to depend on the structure of the sequence $(U_n)$ itself.
Consider $0<s_1 < s_2 \leq d$ and two vectors $\boldsymbol{\tau_1}$ and $\boldsymbol{\tau_2}$ such that $s_1=\min_{s: g_{\boldsymbol{\tau_1}}(s)\geq d}\left\{s\right\}$ and $s_2=\min_{s: g_{\boldsymbol{\tau_2}}(s)\geq d}\left\{s\right\}.$ Consider a weakly redundant sequence of balls $(B_n)_{n\in\mathbb
N}$ of $[0,1]^d$ and a sequence of open sets $(U_n)_{n\in\mathbb{N}}$, $U_n \subset B_n$ satisfying:\mk
\begin{itemize}
\item[•] $\vert B_n \vert \to 0,$\mk
\item[•] $\mathcal{L}^d (\limsup_{n\rightarrow+\infty}B_n)=1,$\mk
\item[•] for any $n\in\mathbb{N}$, $B_n \subset [0,\frac{1}{2})\times \prod_{i=2}^d [0,1]$ or $B_n \subset (\frac{1}{2},1]\times \prod_{i=2}^d [0,1],$\mk
\item[•] for any $n\in\mathbb{N}$ such that $B_n \subset [0,\frac{1}{2})\times \prod_{i=2}^d [0,1]$, $U_n =(R_n )$ with $R_n$ an open rectangle associated with $\boldsymbol{\tau_1}$ as in Theorem \ref{rectssmajo},\mk
\item[•] for any $n\in\mathbb{N}$ such that $B_n \subset (\frac{1}{2},1]\times \prod_{i=2}^d [0,1],$ $U_n =R_n $ with $R_n$ an open rectangle associated with $\boldsymbol{\tau_2}.$\mk
\end{itemize}
Then smallest real number such that the condition of Theorem \ref{majoss} holds is $s_2$, the largest real number such that the condition of Theorem \ref{zzani} holds is $s_1$ and $\dim_H (\limsup_{n\rightarrow+\infty}U_n)=s_2.$
On the other hand, following the scheme of example 3.5 in \cite{KR}, it is also possible to construct two weakly redundant sequences of balls $(B_{n,1})_{n\in \mathbb{N}}$ and $(B_{n,2})_{n\in\mathbb{N}}$ such that:
\begin{itemize}
\item[•] $\vert B_{n,1} \vert \to 0$ and $\vert B_{n,2}\vert \to 0,$\mk
\item[•] $0<\mathcal{L}^d (\limsup_{n\rightarrow+\infty}B_{n,2})<1,$\mk
\item[•] $\limsup_{n\rightarrow+\infty}R_{n,2}=\emptyset$, where $R_{n,2}\subset B_{n,2}$ is an open rectangle associated with $\boldsymbol{\tau_2}$,\mk
\item[•] $\limsup_{n\rightarrow+\infty}B_{n,1}\subset [0,1]^d \setminus \limsup_{n\rightarrow+\infty}B_{n,2}$ and $$\mathcal{L}^{d}(\limsup_{n\rightarrow+\infty}B_{n,1})=1-\mathcal{L}^d (\limsup_{n\rightarrow+\infty}B_{n,2}).$$ For any $n\in\mathbb{N}$, denote by $R_{n,1}\subset B_{n,1}$ an open rectangle associated with $\boldsymbol{\tau_1}.$
\end{itemize}
Those properties implies that that any sequence $(B_n)_{n\rightarrow+\infty}$ corresponding to the family $\left\{B_{n,i}\right\}_{n\in\mathbb{N}, i\in\left\{1,2\right\}}$ is weakly redundant and satisfies $\mathcal{L}^d (\limsup_{n\rightarrow+\infty}B_n)=1.$
Again, the smallest real number such that the condition of Theorem \ref{rectssmajo} holds is $s_2$, the largest real number such that the condition of Theorem \ref{zzani} holds is $s_1$ but this time, $\dim_H (\limsup_{n\rightarrow+\infty}U_n)=s_1.$
\end{remarque}
\section{Conclusion and perspectives}
\label{conclu}
The properties stated in Theorem \ref{theoremextra} (the prescribed measure and the weak redundancy) are of course non exhaustive and maybe more can be imposed to well chosen subsequences of $\mu$-.a.c sequences of balls. It turns out that in the quasi-Bernoulli case it is enough to get condition under which the lower-bound found in \cite{ED1} is also an upper-bound , but it is likely that in some other cases, one needs to ask the sequence to verify more properties to ensure the sharpness of a certain lower-bound. In particular, it can be proved that, under very weak hypothesis on a $\mu$-a.c sequence $(B_n)_{n\in\mathbb{N}}$, given a $G_{\delta}$ set of full measure $G$, it is possible to assume (up to an extraction) that the set $G^{\prime}=\limsup_{n\rightarrow+\infty}B_n$ is a $G_{\delta}$ set of full measure with $G^{\prime}\subset G$ (so that one can assume that $\limsup_{n\rightarrow+\infty}B_n$ is always included in any $G_{\delta}$ of full measure if needed).
\mk
\bibliographystyle{plain}
|
\section{Introduction}
In \cite{Serre}, Serre defined the concept of a $p$-adic modular form as the $p$-adic limit of a
$q$-expansion of modular forms with rational Fourier coefficients. The $p$-adic Eisenstein series
was introduced as a typical example of $p$-adic modular forms, and its relation with the
modular forms on $\Gamma_0(p)$ was also studied. Later, Serre's $p$-adic Eisenstein series was
extended to the case of Siegel modular groups, revealing various interesting properties. For example,
it was shown that some $p$-adic Siegel Eisenstein series becomes the usual Siegel modular form
of level $p$ (\cite{Nagaoka1}, \cite{Kat-Nag}).
Let $p$ be a prime number and $\{k_m\}$ the sequence defined by
$$
k_m:=2+(p-1)p^{m-1}.
$$
For sequence $\{k_m\}$, the $p$-adic Siegel Eisenstein series
$$
\widetilde{E}_2^{(n)}:=\lim_{m\to\infty}E_{k_m}^{(n)}
$$
is defined, where $E_k^{(n)}$ is the ordinary Siegel Eisenstein series of degree $n$ and weight $k$.
Let $S^{(p)}$ be a positive definite, half-integral symmetric matrix of degree $4$ with
level $p$ and determinant $p^2/16$. We denote by genus\,$\Theta^{(n)}(S^{(p)})$ the genus theta series of
degree $n$ associated with $S^{(p)}$ (for the precise definition, see $\S$ \ref{genusT}). Genus theta
series genus\,$\Theta^{(n)}(S^{(p)})$ is a Siegel modular form of weight 2 on the level $p$ modular
group $\Gamma_0^{(n)}(p)$.
In \cite{KN}, Kikuta and the second author showed the coincidence between the two objects
$\widetilde{E}_2^{(2)}$ and genus\,$\Theta^{(2)}(S^{(p)})$. This result asserts that there is a correspondence
between some $p$-adic Siegel Eisenstein series and some genus theta series when the degree
is 2. The main purpose of this paper is to show that this coincidence still exists for any $n$.
Namely we prove the following theorem.
\vspace{3mm}
\\
\textbf{Theorem}\quad {\it We assume that $p$ is an odd prime number.
Then the degree $n$ $p$-adic Siegel Eisenstein series $\widetilde{E}_2^{(n)}$ coincides with
the degree $n$ genus theta series {\rm genus}\,$\Theta^{(n)}(S^{(p)})$:
$$
\widetilde{E}_2^{(n)}={\rm genus}\,\Theta^{(n)}(S^{(p)}).
$$
In particular, the $p$-adic Siegel Eisenstein series $\widetilde{E}_2^{(n)}$ is a Siegel
modular form of weight 2 on $\Gamma_0^{(n)}(p)$.
}
\\
The equality of this theorem is proved by showing that the Fourier coefficients on both
sides are equal. In particular, as can be seen from the discussion in the text, the computation
of local densities for $n=3$ or $4$ is essential
(cf. \S\; \ref{LC34}).
By considering the $p$-adic first and second approximation of $\widetilde{E}_2^{(n)}$,
we obtain the following results.
\vspace{2mm}
\\
\textbf{Corollary}\quad (1)\;\; {\it Assume that $p>n$. Then the modular form
$\widetilde{E}_2^{(n)}$ of level $p$ and weight 2 is congruent to the weight $p+1$
Siegel Eisenstein series $E_{p+1}^{(n)}$ mod $p$\;{\rm :}
$$
\widetilde{E}_2^{(n)} \equiv E_{p+1}^{(n)} \pmod{p}.
$$
}
(2)\;\;{\it
Assume that $p\geq 3$. Then we have
$$
\varTheta (E_{p+1}^{(3)}) \equiv 0 \pmod{p}\quad and \quad \varTheta (E_{p^2-p+2}^{(4)}) \equiv 0 \pmod{p^2},
$$
where $\varTheta$ is the theta operator {\rm (cf. $\S$ \ref{ThetaOp})}.
}
\vspace{3mm}
\\
Statement (1) is motivated by Serre's
result in the case of elliptic modular
forms: For any modular form $f$ of weight 2 on $\Gamma_0(p)$, there is a modular
form $g$ of level one and weight $p+1$ satisfying
$$
f \equiv g \pmod{p}.
$$
The result described in (2) is related to the theory of the mod $p$ kernels of theta operators.
The second congruence provides an example of a Siegel modular form contained in the mod $p^2$ kernel
of a theta operator.
\vspace{4mm}
\\
{\sc Notation.}
Let $R$ be a commutative ring. We denote by $R^{\times}$ the unit group of $R$.
We denote by $M_{mn}(R)$ the set of
$m \times n$-matrices with entries in $R.$ In particular put $M_n(R)=M_{nn}(R).$ Put $GL_m(R) = \{A \in M_m(R) \ | \ \det A \in R^\times \},$ where $\det
A$ denotes the determinant of a square matrix $A$. For an $m \times n$-matrix $X$ and an $m \times m$-matrix
$A$, we write $A[X] = {}^t X A X,$ where $^t X$ denotes the
transpose of $X$. Let $\text{Sym}_n(R)$
denote
the set of symmetric matrices of degree $n$ with entries in
$R.$ Furthermore, if $R$ is an integral domain of characteristic different from $2,$ let $\calh_n(R)$ denote the set of half-integral matrices of degree $n$ over $R$, that is, $\calh_n(R)$ is the subset of symmetric
matrices of degree $n$ with entries in the field of fractions of $R$ whose $(i,j)$-component belongs to
$R$ or ${1 \over 2}R$ according as $i=j$ or not.
We say that an element $A$ of $M_n(R)$ is non-degenerate if $\det A \not=0$. For a subset $\cals$ of $M_n(R)$ we denote by $\cals^{\nd}$ the subset of $\cals$
consisting of non-degenerate matrices. If $\cals$ is a subset of $\Sym_n(\RR)$ with $\RR$ the field of real numbers, we denote by $\cals_{>0}$ (resp. $\cals_{\ge 0}$) the subset of $\cals$
consisting of positive definite (resp. semi-positive definite) matrices.
We sometimes write $\Lambda_n$ (resp. $\Lambda_n^+$) instead of $\calh_n(\ZZ)$
(resp. $\calh_n(\ZZ)_{>0}$).
The group
$GL_n(R)$ acts on the set $\Sym_n(R)$ by
$$
GL_n(R) \times \Sym_n(R) \ni (g,A) \longmapsto A[g] \in \Sym_n(R).
$$
\noindent
Let $G$ be a subgroup of $GL_n(R).$ For a $G$-stable subset ${\mathcal B}$ of $\Sym_n(R)$ we denote by $\calb/G$ the set of equivalence classes of $\calb$ under the action of $G.$ We sometimes use the same symbol $\calb/G$ to denote a complete set of representatives of $\calb/G.$ We abbreviate $\calb/GL_n(R)$ as $\calb/\!\!\sim$ if there is no fear of confusion. Let $G$ be a subgroup of $GL_n(R)$. Then two symmetric matrices $A$ and $A'$ with
entries in $R$ are said to be $G$-equivalent with each
other and write $A \sim_{G} A'$ if there is
an element $X$ of $G$ such that $A'=A[X].$ We also write $A \sim A'$ if there is no fear of confusion.
For square matrices $X$ and $Y$ we write $X \bot Y = \begin{pmatrix} X &O \\ O & Y \end{pmatrix}$.
We put ${\bf e}(x)=\exp(2 \pi \sqrt{-1} x)$ for $x \in \CC,$ and for a prime number $q$ we denote by ${\bf e}_q(*)$ the continuous additive character of $\QQ_q$ such that ${\bf e}_q(x)= {\bf e}(x)$ for $x \in \ZZ[q^{-1}].$
For a prime number $q$ we denote by $\ord_q(*)$ the additive valuation of $\QQ_q$ normalized so that $\ord_q(q)=1$.
Moreover for any element $a, b \in \ZZ_q$
we write $b \equiv a \pmod {q}$ if $\ord_q(a-b) >0$.
\section{Siegel Eisenstein series and genus theta series}
\subsection{Siegel modular forms}
Let $\mathbb{H}_n$ be the Siegel upper-half space of degree $n$; then the Siegel
modular group $\Gamma^{(n)}:=Sp_n(\mathbb{R})\cap M_{2n}(\mathbb{Z})$ acts
discontinuously on $\mathbb{H}_n$. For a congruence subgroup $\Gamma'$ of
$\Gamma^{(n)}$, we denote by $M_k(\Gamma')$ the corresponding space of Siegel
modular forms of weight $k$ for $\Gamma'$. Later we mainly deal with the case $\Gamma'=\Gamma^{(n)}$
or $\Gamma_0^{(n)}(N)$ where
$$
\Gamma_0^{(n)}(N)=\left\{ \binom{A\;B}{C\;D}\in\Gamma^{(n)}\;\big{|}\; C \equiv O_n \pmod{N}\;\right\}.
$$
In both cases, $F\in M_k(\Gamma')$ has a Fourier expansion of the form
$$
F(Z)=\sum_{0\leq T\in\Lambda_n}a(F,T)\,{\bf e}(\text{tr}(TZ)).
$$
Taking $q_{ij}:={\bf e}(z_{ij})$
with $Z=(z_{ij})\in\mathbb{H}_n$, we write
$$
q^T:={\bf e}(\text{tr}(TZ))
=\prod_{1\leq i<j\leq n}q_{ij}^{2t_{ij}}\prod_{i=1}^nq_i^{t_i},
$$
where $T=(t_{ij})$ and $q_i=q_{ii}$,\,$t_i=t_{ii}$\,$(i=1,\cdots,n)$. Using this notation, we obtain
\begin{align*}
F=\sum_{0\leq T\in\Lambda_n}a(F,T)\,q^T & =\sum_{t_i}\left(\sum_{t_{ij}}a(F,T)\prod_{i<j}q_{ij}^{2t_{ij}}\right)
\prod_{i=1}^nq_i^{t_{ij}} \\
& \in\mathbb{C}[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!].
\end{align*}
For a subring $R\subset\mathbb{C}$, we denote by $M_k(\Gamma')_R$ the set consisting of
modular forms $F$ all of whose Fourier coefficients $a(F,T)$ lie in $R$. Therefore, an element
$F\in M_k(\Gamma')_R$ may be regarded as an element of
$$
R[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!].
$$
\subsection{Siegel Eisenstein series}
Define
$$
\Gamma_\infty^{(n)}:=\left\{\;\binom{A\;B}{C\;D}\in\Gamma^{(n)}\;\big{|}\; C=O_n\; \right\}.
$$
For an even integer $k>n+1$, define a series by
$$
E_k^{(n)}(Z)=\sum_{\binom{*\;*}{C\;D}\in\Gamma_\infty^{(n)}\backslash \Gamma^{(n)}}
\text{det}(CZ+D)^{-k},\quad Z\in\mathbb{H}_n.
$$
This series is an element of $M_k(\Gamma^{(n)})_{\mathbb{Q}}$ called the {\it Siegel Eisenstein series} of
weight $k$ for $\Gamma^{(n)}$.
\subsection{Genus theta series}
\label{genusT}
Fix $S\in\Lambda_{m}^{+}$ and define
$$
\theta^{(n)}(S;Z):=\sum_{X\in M_{mn}(\mathbb{Z})}\,{\bf e}(\text{tr}(S[X]Z)),\quad Z\in\mathbb{H}_n.
$$
Let $\{S_1,\ldots, S_d\}$ be a set of representatives of $GL_m(\mathbb{Z})$-equivalence classes
in $\text{genus}\,(S)$. The {\it genus theta series} associated with $S$ is defined by
$$
\text{genus}\,\Theta^{(n)}(S)(Z):=
\left(\sum_{i=1}^d\frac{\theta^{(n)}(S_i;Z)}{a(S_i,S_i)}\right)\Big{/}
\left( \sum_{i=1}^d\frac{1}{a(S_i,S_i)}\right), \quad Z\in\mathbb{H}_n
$$
where
$$
a(S_i,S_i):=\sharp\{\;X\in M_m(\mathbb{Z})\;|\;S_i[X]=S_i\;\}\quad (\text{cf. \S\;\ref{LC34})}.
$$
\subsection{$p$-adic Siegel Eisenstein series}
Let $\{k_m\}_{m=1}^{\infty}$ be an increasing sequence of even positive integers which
is $p$-adically convergent.
If the corresponding sequence of Siegel Eisenstein series
$$
\{E_{k_m}^{(n)}\}\subset \mathbb{Q}[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!]
$$
converges $p$-adically to an element of $\mathbb{Q}_p[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!]$,
then we call the limit
$$
\lim_{m\to\infty}E_{k_m}^{(n)}\in \mathbb{Q}_p[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!]
$$
a {\it $p$-adic Siegel Eisenstein series}.
\section{Main result}
\subsection{Statement of the main theorem}
As stated in the Introduction, the main result of this paper asserts a coincidence between
some $p$-adic
Siegel Eisenstein series and some genus theta series.
First we consider an element $S^{(p)}$ of $\Lambda_4^+$ with level $p$ and
determinant $p^2/16$.
(The existence of such a matrix will be proved in Lemma \ref{lem.existence-of-S}.)
Next we consider a special $p$-adic Siegel Eisenstein series.
Let $\{k_m\}_{m=1}^\infty\subset \mathbb{Z}_{>0}$ be a sequence defined by
$$
k_m=k_m(p):=2+(p-1)p^{m-1}\qquad (p:\;\text{prime}).
$$
This sequence converges $p$-adically to 2.
We associate this sequence with
the sequence of Siegel Eisenstein series
$$
\{E_{k_m}^{(n)}\}_{m=1}^\infty\subset \mathbb{Q}[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!].
$$
As we prove in the following, this sequence defines a $p$-adic Siegel Eisenstein series.
We set
$$
\widetilde{E}_2^{(n)}:=\lim_{m\to\infty}E_{k_m}^{(n)}.
$$
Our main theorem can be stated as follows.
\begin{theorem}
\label{statementmain}
Let $p$ be an odd prime number and $S^{(p)}$ be as above.
Then the following identity holds:
$$
\widetilde{E}_2^{(n)}={\rm genus}\,\Theta(S^{(p)})
$$
\end{theorem}
If $n=2$, the above identity has already been proved in \cite{KN},
and the proof for a general $n$ is essentially the
case $n=3$ or $4$, which will be presented in the next section.
\subsection{Case $\boldsymbol{n=3}$ or $\boldsymbol{4}$}
\label{n=3-4}
We prove our main result in the case that $n$ is $3$ or $4$.
\begin{theorem} \label{th.main-result}
\begin{itemize}
\item[{\rm (1)}] For any $T \in \Lambda_3^+$, we have
\begin{align*}
a(\widetilde E_2^{(3)},T)=a(\mathrm{genus} \ \Theta^{(3)}(S^{(p)}),T).
\end{align*}
\item[{\rm (2)}] For any $T \in \Lambda_4^+$, we have
\begin{align*}
a(\widetilde E_2^{(4)},T)=a(\mathrm{genus} \ \Theta^{(4)} (S^{(p)}),T).
\end{align*}
\end{itemize}
\end{theorem}
By the above theorem and \cite{KN}. we have the following result.
\begin{corollary}
\label{cor.main-result2}
Let $n=3$ or $4$. Then
\begin{align*}
\widetilde E_2^{(n)}=\mathrm{genus} \ \Theta^{(n)}(S^{(p)}).
\end{align*}
\end{corollary}
To prove Theorem \ref{th.main-result}, we compute the both-hand
sides of the equality in the theorem explicitly.
We also use the notation in \cite{IK22}.
Let $q$ be a prime number.
Let $\langle \ {} \ , \ {} \ \rangle=\langle \ {} \ , \ {} \ \rangle_Q$ be the Hilbert symbol on $\QQ_q$.
Let $T$ be a non-degenerate symmetric matrix with entries in $\QQ_q$ of degree $n$. Then $T$ is $GL_n(\QQ_q)$-equivalent to $b_1 \bot \cdots \bot b_n$ with $b_1,\ldots,b_n \in \QQ_q^{\times}$.
Then we define the Hasse invariant $h_q(T)$ as
\[h_q(T)=\prod_{1 \le i \le j \le n} \langle b_i,b_j \rangle.\]
We also define $\varepsilon_q(T)$ as
\[\varepsilon_q(T)=\prod_{1 \le i < j \le n} \langle b_i,b_j \rangle.\]
These do not depend on the choice of $b_1,\ldots,b_n$. We also denote by $\eta_q(T)$ the Clifford invariant of $T$
(cf. \cite{IK22}).
Then we have
\[\eta_q(T)=
\begin{cases}
\langle -1,-1 \rangle^{m(m+1)/2} \langle (-1)^m,\det (T) \rangle \varepsilon_q(T) & \ \text{if $n=2m+1$,}\\
\langle -1,-1 \rangle^{m(m-1)/2} \langle (-1)^{m+1},\det (T) \rangle \varepsilon_q(T) & \ \text{if $n=2m$,}
\end{cases}\]
and hence,
\[\eta_q(T)=
\begin{cases}
\langle -1,-1 \rangle^{m(m+1)/2} \langle (-1)^m \det (T),\det (T) \rangle \varepsilon_q(T) & \ \text{if $n=2m+1$,}\\
\langle -1,-1 \rangle^{m(m-1)/2} \langle (-1)^{m+1} \det (T),\det (T) \rangle \varepsilon_q(T) & \ \text{if $n=2m$.}
\end{cases}\]
Let $T \in \Sym_n(\QQ_q)$. Then, by the product formula for the Hilbert symbol, we have
$\prod_q h_q(T)=1$ and
\[\prod_q \eta_q(T)=\begin{cases} (-1)^{(n^2-1)/8} & \text{ if } n \text{ is odd} ,\\
(-1)^{n(n-2)/8} & \text{ if } n \text{ is even}.
\end{cases}\]
For $a \in \ZZ_q, a \not=0$, define $\chi_q(a)$ as
\[
\chi_q(a)=
\begin{cases}
1 & \text{ if } \QQ_q(\sqrt{a})=\QQ_q,\\
-1 & \text{ if } \QQ_q(\sqrt{a})/\QQ_q \text { is unramified quadratic}, \\
0 & \text{ if } \QQ_q(\sqrt{a})/\QQ_q \text{ is ramified quadratic}.
\end{cases}
\]
We now prove the exisitence of the matrix $S^{(p)}$ that appears in the main theorem.
\vspace{2mm}
\\
Let $H_k = \overbrace{H \bot ...\bot H}^k$ with
$H =\left(\begin{matrix}
0 & 1/2 \\
1/2 & 0
\end{matrix}\right)$.
We take an element $\epsilon \in \ZZ_p^\times$ such that $\chi_p(-\epsilon)=-1$,
and put $U_0=1 \,\bot\, \epsilon$.
\begin{lemma}\label{lem.existence-of-S}
For each prime number $q$, let $S_q$ be an element of $\calh_4(\ZZ_q)$
such that
\begin{align*}
S_q =\begin{cases} U_0 \bot pU_0 & \text{ if } q=p, \\
H_2 & \text{ otherwise}.
\end{cases}
\end{align*}
Then there exists an element $S$ of $\Lambda_4^+$ with level $p$ and determinant $16^{-1}p^2$ such that
\begin{align*}
S \sim_{GL_4(\ZZ_q)} S_q \text{ for any } q. \tag{E}
\end{align*}
Conversely, if $S$ is an element of $\Lambda_4^+$ with level $p$ and determinant $16^{-1}p^2$ , then $S$ satisfies condition {\rm (E)}.
\end{lemma}
\begin{proof}
Let $S_q$ be as above. By assumption, we have
\[ (16^{-1}p^2)^{-1}\det (S_q) \in (\ZZ_q^\times)^2 \text{ for any } q,\]
and
\begin{align*} h_q(S_q)=\begin{cases} -1 & \text{ if } q=p \text{ or } 2, \\
1 & \text{ otherwise}.
\end{cases}
\end{align*}
Hence, by \cite[Theorem 4.1.2]{Ki2}, there exists an element $S$ of $\text{Sym}_4(\QQ)_{>0}$ satisfying condition (E). By construction, we easily see that $S$ belongs to $\calh_4(\ZZ)_{>0}$ with level $p$ and determinant $16^{-1}p^2$.
Conversely, let $S$ be an element of $\Lambda_4^+$ with level $p$ and determinant $16^{-1}p^2$.
Then, we have
\[\det (2S) \in (\ZZ_q^\times)^2 \text{ for any } q\not=p.\]
Hence, we have $S \sim_{GL_4(\ZZ_q)} H_2$,
so we have $h_q(S)=1$ if $q \not=p,\,2$ and $h_2(S)=-1$.
Thus, we have $h_p(S)=-1$. By definition, we have
$p^{-2}\det (S) \in (\ZZ_p^\times)^2$ and $pS^{-1} \in \calh_4(\ZZ_p)$. Then, we easily see that
$S \sim_{GL_4(\ZZ_p)} U_0 \bot pU_0$.
This proves the assertion.
\end{proof}
\subsubsection{Computation of $\boldsymbol{a(\widetilde{E}_2^{(n)},T)}$ for $\boldsymbol{n=3}$ or $\boldsymbol{4}$}
In the case $n$ is even, for $T \in \calh_n(\ZZ_q)^{\rm nd}$ we put
$\xi_q(T)=\chi_q((-1)^{n/2} \det (2T))$.
For $T \in \calh_n(\ZZ_q)^{\rm nd}$, let $b_q(T,s)$ be the Siegel series of $T$.
Then $b_q(T,s)$ is a polynomial in $q^{-s}$. More precisely, we define a polynomial $\gamma_q(T,X)$ in $X$ by
$$\gamma_q(T,X)=
\begin{cases}
(1-X)\prod_{i=1}^{n/2}(1-q^{2i}X^2)(1-q^{n/2}\xi_q(T) X)^{-1} & \text{ if $n$ is even,}
\vspace{2mm}
\\
(1-X)\prod_{i=1}^{(n-1)/2}(1-q^{2i}X^2) & \text{ if $n$ is odd.} \end{cases}$$
Then there exists a polynomial $F_q(B,X)$ in $X$ with coefficients in $\ZZ$ such that
$$F_q(T,q^{-s})={b_q(T,s) \over \gamma_q(T,q^{-s})}.$$
The properties of the polynomial $F_q(B,X)$ is studied in detail by the first author
in \cite{K99}. (In particular, see \cite[Theorem 3.2]{K99} for the properties used below.)
We put
\[\widetilde b_q(T,X)=\gamma_q(T,X)F_q(T,X).\]
\begin{proposition}\label{prop.FC-Siegel}
For any $T \in \Lambda_n^+$, we have
\begin{align*}
a(E_k^{(n)},T)&= (-1)^{[(n+1)/2]}2^{n-[n/2]} \frac{k}{B_k} \prod_{i=1}^{[n/2]}\frac{2k-2i}{B_{2k-2i}}
\prod_q F_q(T,q^{k-n-1})\\
&\times \begin{cases} \frac{B_{k-n/2,\chi_T}}{k-n/2} & \text{ if } n \text{ is even,} \\
1 & \text{ if } n \text{ is odd,}
\end{cases}
\end{align*}
where $\chi_T$ denotes the primitive Dirichlet character corresponding to the extension
$\mathbb{Q}(\sqrt{(-1)^{n/2}{\rm det}(2T)})/\mathbb{Q}$.
\end{proposition}
\begin{remark}\label{rem.comariosn-KN}
Let $F_q^{(n)}(T,X)$ be the polynomial in \cite[Theorem 2.2]{Kat-Nag}.
Then it coincides with $F_q(T,X)$ if $n$ is even. But
it is $\eta_q(T)F_q(T,X)$ if $n$ is odd.
\end{remark}
\begin{proposition} \label{prop.limit-of-Kitaoka-polynomial}
Let $T \in \calh_n(\ZZ_q)^{\rm nd}$.
\begin{itemize}
\item[{\rm (1)}] Let $q \not=p$. Then
we have
\begin{align*}
&\lim_{m\to\infty} F_q(T,q^{k_m-n-1})=F_q(T, q^{2-n-1}).\\
\end{align*}
\item[{\rm (2)}] We have
\begin{align*}
&\lim_{m \to\infty} F_p(T,p^{k_m-n-1})=1
\end{align*}
\end{itemize}
\end{proposition}
\begin{proof}
By construction, $F_q(T,X)$ belongs to $\ZZ[X]$ and its first coefficient is $1$.
Thus the assertion holds.
\end{proof}
\begin{lemma} \label{lem.vanishing-of-F}
Let $T \in \calh_3(\ZZ_q)^{\rm nd}$.
Suppose that $\eta_q(T)=-1$. Then
\[F_q(T,q^{-2})=0.\]
\end{lemma}
\begin{proof} By the functional equation of $F_q(T,X)$ (cf. \cite[Theorem 3.2]{K99}), we have
\begin{align*}
F_q(T,q^{-4}X^{-1}) & =\eta_q(T)(q^2X)^{\mathrm{ord}_q(2^{2}\det T)}F_q(T,X)\\
& =-(q^2X)^{\mathrm{ord}_q(2^{2}\det T)}F_q(T,X).
\end{align*}
Hence, we have
\begin{align*}
F_q(T,q^{-2})=-F_q(T,q^{-2}).
\end{align*}
Therefore, the assertion holds.
\end{proof}
Here we prepare some result on Bernoulli numbers.
\begin{lemma}\;{\rm (Carlitz \cite[{\sc Theorem 3}]{Carlitz1})}
\label{lem.limit-of-Bernoulli-number}
Assume that $p$ is an odd prime number.
For $t=rp^k(p-1)$\,{\rm (}$r\in\mathbb{Z}_{>0},\,k\in\mathbb{Z}_{\geq 0}${\rm )},
the numerator of
$\displaystyle B_t+\frac{1}{\,p\,}-1$ is divisible by $p^k$. In particular
$$
\lim_{k\to\infty}B_{r(p-1)p^k}=1-\frac{1}{\,p\,}.
$$
for any $r\in\mathbb{Z}_{>0}$.
\end{lemma}
Now we obtain the following theorem.
\begin{theorem} \label{th.limit-of-FC-of-Eisenstein}
\begin{itemize}
\item[{\rm (1)}] Let $T \in \Lambda_3^+$.
Then, we have
\begin{align*}
a(\widetilde E_2^{(3)},T)=\frac{576}{(1-p)^2} \prod_{q \not=p} F_q(T,q^{-2}),
\end{align*}
and in particular if $\eta_p(T)=1$, we have
\begin{align*}
a(\widetilde E_2^{(3)},T)=0.
\end{align*}
\item[{\rm (2)}] Let $T \in \Lambda_4^+$.
\begin{itemize}
\item[{\rm (2.1)}] Suppose that $\det (2T)$ is not square. Then
\[a(\widetilde E_2^{(4)},T)=0.\]
\item[{\rm (2.2)}] Suppose that $\det (2T)$ is square. Then
\begin{align*}
a(\widetilde E_2^{(4)},T)=\frac{1152}{(1-p)^2} \prod_{q \not=p} F_q(T,q^{-3}).
\end{align*}
In particular if $\eta_p(T)=1$ or $\mathrm{ord}_p(\det(2T)) =0$, then
\[a(\widetilde E_2^{(4)},T)=0.\]
\end{itemize}
\end{itemize}
\end{theorem}
\begin{proof}
(1) We have
\begin{align*} (\star)\qquad\qquad
\lim_{m\to\infty} \frac{B_{k_m}}{k_m}=
\lim_{m\to\infty} \frac{B_{2k_m-2}}{2k_m-2}=
\frac{(1-p)B_2}{2}=\frac{1-p}{12}.
\end{align*}
These identities can be proved by Kummer's congruence when $p>3$.
For $p=3$, these identities are shown using the extended Kummer's congruence
(cf. \cite[COROLLAIRE 3]{Fr}). This proves the first part of the assertion.
Suppose that $\eta_p(T)=1$.
We note that
\[\prod_q \eta_q(T)=-1.\]
Therefore, if $\eta_p(T)=1$, there is a prime number $q \not=p$ such that $\eta_q(T)=-1$. Therefore, by Lemma \ref{lem.vanishing-of-F}, we have $F_q(T,q^{-2})=0$. This proves the second part of the assertion.
(2) Put
\[\calb_{T,m}=\Big(\frac{2k_m-4}{B_{2k_m-4}} \Big)\Big(\frac{B_{k_m-2,\chi_T}}{k_m-2}\Bigr).\]
Then
\[\calb_{T,m} =\Big(\frac{2(p-1)p^{m-1}}{B_{2(p-1)p^{m-1}}} \Big)
\Big(\frac{B_{(p-1)p^{m-1},\chi_T}}{(p-1)p^{m-1}}\Bigr).\]
First suppose that $\det (2T)$ is not square.
Then, $\chi_T$ is non-trivial, and $\frac{B_{(p-1)p^{m-1},\chi_T}}{(p-1)p^{m-1}}$ is $p$-integral
for any $m \ge 1$ (cf. \cite{Carlitz2}). Hence, by the theorem of von Staudt-Clausen,
\[\lim_{m\to\infty} \calb_{T,m}=0.\]
This proves the assertion (2.1).
Next suppose that $\det (2T)$ is square. Then, $B_{(p-1)p^{m-1},\chi_T}=B_{(p-1)p^{m-1}}$, and
\[\calb_{T,m} =2\frac{B_{(p-1)p^{m-1}}}{B_{2(p-1)p^{m-1}}}.\]
By Lemma \ref{lem.limit-of-Bernoulli-number},
\[\lim_{m\to\infty} \calb_{T,m}=2\cdot\lim_{m\to\infty}\frac{B_{(p-1)p^{m-1}}}{B_{2(p-1)p^{m-1}}}=2.\]
This fact, together with $(\star)$, proves the first part of the assertion (2.2).
To prove the remaining part, first suppose that $\eta_p(T)=1$.
We note that
\[\prod_q \eta_q(T)=-1.\]
Therefore, there is a prime number $q \not=p$ such that $\eta_q(T)=-1$.
We have $F_q(T,q^{-3})=0$ as will be proved in Corollary \ref{add-cor2-2}.
Next suppose that $\mathrm{ord}_p(\det (2T)) =0$. Then, $\eta_p(T)=1$, and the assertion follows from the above.
\end{proof}
\subsubsection{Computation of
$\boldsymbol{a(\text{genus}\,\Theta^{(n)}(S^{(p)}),T)}$ for $\boldsymbol{n=3}$ or $\boldsymbol{4}$}
\label{LC34}
For $S \in \calh_m(\ZZ_q)^{\nd}$ and $T \in \calh_n(\ZZ_q)^{\nd}$ with $m \ge n$, we define the local density
$\alpha_q(S,T)$ as
\[\alpha_q(S,T)=2^{-\delta_{m,n}}\lim_{e\to\infty} q^{e(-mn+n(n+1)/2)}\cala_e(S,T),
\]
where
\begin{align*}
\cala_e(S,T)=\{X \in M_{mn}(\ZZ_q)/q^eM_{mn}(\ZZ_q) \ | \ S[X] \equiv T\!\pmod{ q^e \calh_n(\ZZ_q)}\,\},
\end{align*}
and $\delta_{m,n}$ is the Kronecker delta.
We say that an element $X$ of $M_{mn}(\ZZ_q)$ with $m \ge n$ is
{\it primitive} if $\mathrm{rank}_{\ZZ_q/q\ZZ_q} X=n$.
We also say that an element $X \mod q^e$ of $M_{mn}(\ZZ_q)/q^eM_{mn}(\ZZ_q)$ is {\it primitive} if $X$ is primitive.
This definition does not depend on the choice of $X$.
From now on, for $X \in M_{mn}(\ZZ_q)$, we often use the same symbol $X$ to denote the class of $X$ mod $q^e$.
For $S \in \calh_m(\ZZ_q)^{\nd}$ and $T \in \calh_n(\ZZ_q)$ with $m \ge n$, we define the primitive local density $\beta_q(S,T)$ as
\[\beta_q(S,T)=2^{-\delta_{m,n}}\lim_{e\to\infty} q^{e(-mn+n(n+1)/2)}\calb_e(S,T),\]
where
\[\calb_e(S,T)=\{X \in \cala_e(S,T) \ | \ X \text{ is primitive}\}.
\]
The following is due to \cite{Ki1}.
\begin{lemma} \label{lem.local-density-via-primitive-local-density}
Let $S \in \calh_m(\ZZ_q)^{\rm nd}$ and $T \in \calh_n(\ZZ_q)^{\rm nd}$ with $m \ge n$. Then
\[\alpha_q(S,T)=\sum_{g \in GL_n(\ZZ_q) \backslash M_n(\ZZ_q)^{\rm nd}} q^{(-m+n+1) \det g}\beta_q(S,T[g^{-1}]).\]
\end{lemma}
For $S \in \Sym_m(\ZZ_q)^{\nd}$ and $T \in \Sym_n(\ZZ_q)^{\nd}$ we also define another local density $\widetilde \alpha_q(T,S)$ as
\begin{align*}
\widetilde \alpha_q(T,S)=2^{-\delta_{m,n}}\lim_{e\to\infty} q^{e(-mn+n(n+1)/2)}\widetilde \cala_e(T,S),
\end{align*}
where
\begin{align*}
\widetilde \cala_e(T,S)=\{X \in M_{mn}(\ZZ_q)/q^eM_{mn}(\ZZ_q) \ |
\ S[X] \equiv T \!\pmod{q^e \mathrm{Sym}_n(\ZZ_q)}\,\}.
\end{align*}
\begin{remark} \label{rem.comparison-densities}
We note that
\begin{align*}
\widetilde \alpha_q(2T,2S)=2^{m\delta_{2,p}}\alpha_q(S,T)
\end{align*}
for $S \in \calh_m(\ZZ_q)$ and $T \in \calh_n(\ZZ_q)$.
\end{remark}
\vspace{3mm}
For $S \in \Lambda_m^+$ and $T \in \Lambda_n^+$, put
\[a(S,T)=\#\{X \in M_{mn}(\ZZ) \ | \ S[X]=T\}.\]
Moreover put
\[M(S)=\sum_{i=1}^{d} {1 \over a(S_i,S_i)} \]
and
\[\widetilde a(S,T)=M(S)^{-1}\sum_{i=1}^d {a(S_i,T) \over a(S_i,S_i)},\]
where $S_i$ runs over a complete set of $GL_m(\ZZ)$-equivalence classes in $\text{genus}(S)$.
Then we have the following formula.
\begin{proposition} \label{prop.Siegel-main-theorem}
Under the above notation, we have
\begin{align*}
&\widetilde a(S,T)=\\
&2^n \vep_{n,m}\pi^{n(2m-n+1)/4}\prod_{i=1}^{n-1} \Gamma((m-i)/2)^{-1}(\det (2S))^{-n/2}(\det (2T))^{(m-n-1)/2}\\
&\times \prod_q \alpha_q(S,T),
\end{align*}
where
\[\vep_{n,m}=\begin{cases} 1/2 & \text{ if either } m=n+1 \text{ or } m=n>1, \\
1 & \text{otherwise}. \end{cases}\]
\end{proposition}
\begin{proof}
We note that we have $\widetilde a(S,T)=\widetilde a(2S,2T)$. By \cite[Theorem 6.8.1]{Ki2} we have
\begin{align*}
&\widetilde a(2S,2T)\\
&=\vep_{n,m}\pi^{n(2m-n+1)/4}\prod_{i=1}^{n-1} \Gamma((m-i)/2)^{-1}(\det (2S))^{-n/2}(\det (2T))^{(m-n-1)/2}\\
&\times \prod_q \widetilde \alpha_q(2T,2S).
\end{align*}
Then, the assertion follows from Remark \ref{rem.comparison-densities}.
\end{proof}
\begin{proposition} \label{prop.mass-formula}
\begin{itemize}
\item[{\rm (1)}] Let $T \in \Lambda_3^+$. Then we have
\[a(\mathrm{genus} \ \Theta^{(3)}(S^{(p)}),T)=8p^{-3}\pi^{4} \alpha_p(U_0 \bot pU_0,T)\prod_{q \not=p} \alpha_q(H_2,T).\]
\item[{\rm (2)}] Let $T \in \Lambda_4^+$. Then we have
\[a(\mathrm{genus}\ \Theta^{(4)}(S^{(p)}),T)=16p^{-4}\pi^{4} \det(2T)^{-1/2}\alpha_p(U_0 \bot pU_0,T)\prod_{q \not=p} \alpha_q(H_2,T).\]
\end{itemize}
\end{proposition}
\begin{proof}
(1) By Proposition \ref{prop.Siegel-main-theorem}, we have
\[a(\mathrm{genus} \ \Theta^{(3)}(S^{(p)}),T)=8p^{-3}\pi^{4} \prod_{q } \alpha_q(S^{(p)},T).\]
By Lemma \ref{lem.existence-of-S}, we prove the assertion.
(2) Again by Proposition \ref{prop.Siegel-main-theorem}, we have
\[a(\mathrm{genus} \ \Theta^{(4)}(S^{(p)}),T)=16p^{-4}\pi^{4} \det(2T)^{-1/2}\prod_{q } \alpha_q(S^{(p)},T).\]
Again by Lemma \ref{lem.existence-of-S}, we prove the assertion.
\end{proof}
The following lemma has been essentially proved in \cite[Lemma 14.8]{Sh1}.
\begin{lemma} \label{lem.local-density-and-Siegel-series}
Let $T \in \calh_n(\ZZ_q)^{\rm nd}$ with $k \ge n/2$. Then, we have
\[\alpha_q(H_k,T)=2^{-\delta_{2k,n}}\widetilde b_q(T,q^{-k}).\]
\end{lemma}
\begin{proposition}\label{prop.local-density-at-q}
\label{prop2-5}
Let $q \not=p$.
\begin{itemize}
\item[{\rm (1)}] Let $T \in \calh_3(\ZZ_q)^{\rm nd}$. Then,
\[\alpha_q(H_2,T)=(1-q^{-2})^2 F_q(T,q^{-2}).\]
\item[{\rm (2)}]
Let $T \in \calh_4(\ZZ_q)^{\rm nd}$ and suppose that $\det(2T)$ is square in $\ZZ_q$. Then,
\[\alpha_q(H_2,T)=(1-q^{-2})^2 q^{\mathrm{ord}_q(\det (2T))/2}F_q(T,q^{-3}).\]
\end{itemize}
\end{proposition}
\begin{proof}
(1) By definition, $\gamma_q(T,X)=(1-X)(1-q^{2}X^2)$. Thus the assertion is easy to prove using Lemma \ref{lem.local-density-and-Siegel-series}.
(2) Again by definition, $\gamma_q(T,X)=(1-X)(1-q^{2}X^2)(1+q^2X)$. Hence, by Lemma \ref{lem.local-density-and-Siegel-series}, we have
\[\alpha_q(H_2,T)=(1-q^{-2})^2 F_q(T,q^{-2}).\]
Thus the assertion follows from the functional equation of $F_q(T,X)$.
\end{proof}
\begin{corollary}
\label{add-cor2-2} Assume that $\det (2T)$ is square in $\mathbb{Z}_q$.
Let $T \in \mathcal{H}_4(\mathbb{Z}_q)^{\mathrm{nd}}$ and suppose that $\eta_q(T)=-1$.
Then $F_q(T,q^{-3})=0$.
\end{corollary}
\begin{proof}
Since $\eta_q(T)=1$, we have $\alpha_q(H_2,T)=0$. Thus the assertion follows from
Proposition \ref{prop2-5} (2).
\end{proof}
The following proposition is one of key ingredients in the proof of our main result.
\begin{proposition}\label{prop.local-density-at-p}
\begin{itemize}
\item[{\rm (1)}] Let $T \in \calh_3(\ZZ_p)^{\nd}$. Then
\[\alpha_p(U_0 \bot pU_0,T)=(1+p)(1+p^{-1})(1-\eta_p(T)).\]
\item[{\rm (2)}] Let $T \in \calh_4(\ZZ_p)^{\nd}$.
If $\alpha_p(U_0 \bot pU_0,T) \not=0$, then $\det (2T) =p^{2m}\xi^2$ with $m \in \ZZ_{>0}, \xi \in \ZZ_p^\times$ and $\eta_p(T)=-1$.
Conversely, if $T$ satisfies this condition, then
\begin{align*}
\alpha_p(U_0 \bot pU_0, T)=2(1+p^{-1})^2 p^m.
\end{align*}
\end{itemize}
\end{proposition}
The above proposition may be proved by the explicit formula in \cite[THEOREM]{SH}. But to apply the formula, we have to compute too many quantities associated with $U_0 \bot pU_0$ and $T$.
Therefore, we here use another method.
We say that $S \in \calh_n(\ZZ_q)^{\rm nd}$ is {\it maximal} if
there is no element
$g \in M_n(\ZZ_q)^{\mathrm{nd}}$ such that $\det g \in q\ZZ_q$ and $S[g^{-1}] \in \calh_n(\ZZ_q)$.
First we prove Proposition \ref{prop.local-density-at-p} (1).
The following lemma is easy to prove.
\begin{lemma}
\label{lem.maximal-element3}
Let $q$ be an odd prime number.
Suppose that $T \in \calh_3(\ZZ_q)^{\rm nd}$ is maximal. Then
$T$ is one of the following matrices.
\begin{itemize}
\item[{\rm (1)}] $T \sim_{GL_3(\ZZ_q)} \epsilon_1 \bot \epsilon_2 \bot \epsilon_3$ with $\epsilon_i \in \ZZ_q^\times \ (i=1,2,3)$.
\item[{\rm (2)}] $T \sim_{GL_3(\ZZ_q)} \epsilon_1 \bot \epsilon_2 \bot q\epsilon_3$ with $\epsilon_i \in \ZZ_q^\times \ (i=1,2,3)$.
\item[{\rm (3)}] $T \sim_{GL_3(\ZZ_q)} \epsilon_1 \bot q\epsilon_2 \bot q\epsilon_3$ with $\epsilon_i \in \ZZ_q^\times \ (i=1,2,3)$ such that $\chi_q(-\epsilon_2\epsilon_3)=-1$.
\end{itemize}
Moreover, we have
\[ \eta_q(T)=\begin{cases}1 & \text{ in the case {\rm (1)},} \\
\chi_q(-\epsilon_1\epsilon_2) & \text{ in the case {\rm (2)},} \\
-1 & \text{ in the case {\rm (3)}}.
\end{cases}\]
\end{lemma}
\begin{lemma}
\label{lem.explicit-primitive-locla-density3}
Let $T \in \calh_3(\ZZ_p)^{\rm nd}$.
\begin{itemize}
\item[{\rm (1)}] Suppose that $T$ is non-maximal. Then $\beta_p(U_0 \bot pU_0,T)=0$.
\item[{\rm (2)}] Suppose that $T$ is maximal and $\eta_p(T)=1$. Then
$\beta_p(U_0 \bot pU_0,T)=0$.
\item[{\rm (3)}] Suppose that $T$ is maximal and $\eta_p(T)=-1$. Then \[\beta_p(U_0 \bot pU_0,T)=2(1+p)(1+p^{-1}).\]
\end{itemize}
\end{lemma}
\begin{proof}
We may suppose that $T=\epsilon_1 p^{a_1} \bot \epsilon_2 p^{a_2}
\bot \epsilon_3 p^{a_3}$ with $a_1 \le a_2 \le a_3$ and
and $\epsilon_i \in \ZZ_p^\times \ (i=1,2,3)$.
Suppose that $T$ is non-maximal. Then, by Lemma \ref{lem.maximal-element3}, we have $a_1 \ge 1$, or $a_3 \ge 2$, or $a_1=0,a_2=a_3=1$ and $\chi_p(-\epsilon_2\epsilon_3)=1$.
Suppose that $\beta_p(U_0 \bot pU_0,T) \not=0$. Then, there is a primitive matrix $X=(x_{ij})_{1 \le i \le 4, 1 \le j \le 3} \in M_{4,3}(\ZZ_p)$ such that
\begin{align*}
(U_0 \bot pU_0)[X] \equiv T \pmod{p^e \calh_3(\ZZ_p)}
\tag{a}
\end{align*}
for an integer $e \ge 2$.
In the case $a_1 \ge 1$, (a) implies that ${\bf x}_j=(x_{i,j})_{1 \le i \le 2}$ is primitive for some $j=1,2,3$ and that
\[U_0 [{\bf x}_j]\equiv 0 \pmod{p}.\]
This is impossible because $\chi_p(\det U_0)=-1$.
In the case $a_3 \ge 2$, (a) implies that
${\bf y}=(x_{i,3})_{1 \le i \le 4}$ is primitive, and that
\[(U_0 \bot pU_0) [{\bf y}]\equiv 0 \pmod{p^2}.\]
This is also impossible by the same reason as above.
In the case, $a_1=0,a_2=a_3=1$ and $\chi_p(-\epsilon_2\epsilon_3)=1$, (a) implies that ${\bf z}=(x_{ij})_{3 \le i \le 4, 2 \le j \le 3}$ belongs to $GL_2(\ZZ_p)$ and
\[pU_0[{\bf z}] \equiv \epsilon_2 p \bot \epsilon_3 p \pmod{p^2 \calh_2(\ZZ_p)}.\]
This is also impossible because $\chi_p(-\det (pU_0)) \not=\chi_p(-\det (\epsilon_2 p \bot \epsilon_3 p))$.
This proves the assertion (1).
Next suppose that $T$ is maximal and $\eta_p(T)=1$. Then, again by Lemma \ref{lem.maximal-element3}, $a_1=a_2=a_3=0$, or $a_1=a_2=0,a_3=1$ and $\chi_p(-\epsilon_1\epsilon_2)=1$.
In the first case, since the $\ZZ_p/p\ZZ_p$-rank of $U_0 \bot pU_0$ is smaller than that of $T$, clearly we have $\beta_p(U_0 \bot pU_0,T)=0$. In the second case, since $\chi_p(-\det U_0) \not=\chi_p(-\det (\epsilon_1 \bot \epsilon_2))$, we also have
$\beta_p(U_0 \bot pU_0,T)=0$. Finally suppose that
$T$ is maximal and $\eta_p(T)=-1$. Let $X$ be a primitive matrix satisfying the condition (a).
First let
$T=\epsilon_1 \bot \epsilon_2 \bot \epsilon_3 p$ with $\chi_p(-\epsilon_1\epsilon_2)=-1$.
Write $X=\begin{pmatrix} X_{11} & X_{12} \\ X_{21} & X_{22} \end{pmatrix}$ with $X_{11} , M_{21} \in M_{2}(\ZZ_p)$ and $X_{12},X_{22} \in M_{21}(\ZZ_p)$. Then we have
\begin{align*}
U_0[X_{11}] +pU_0[X_{21}] \equiv \epsilon_1 \bot \epsilon_2 \pmod{p^e}, \tag{b11}
\end{align*}
\begin{align*}
{}^tX_{11}U_0X_{12}+{}^tX_{21}pU_0X_{22} \equiv 0 \pmod{p^e}, \tag{b12}
\end{align*}
\begin{align*}
U_0[X_{12}]+pU_0[X_{22}] \equiv \epsilon_3 p \pmod{p^e}.\tag{b22}
\end{align*}
Since $U_0$ and $\epsilon_1 \bot \epsilon_2$ are invertible in $M_2(\ZZ_p)/p^eM_2(\ZZ_p)$, so is $X_{11}$ by (b11).
Hence, by (b12), we have
\[X_{12}\equiv -p({}^tX_{11}U_0)^* \ {}^t\!X_{21}U_0X_{22} \pmod{p^e},\]
and by (b22), we have
\begin{align*}
p(pU_0[({}^tX_{11}U_0)^*\ {}^t\!X_{21}U_0]+U_0)[X_{22}] \equiv \epsilon_3 p \pmod{p^e}, \tag{b$22'$}
\end{align*}
where $({}^tX_{11}U_0)^*$ is the inverse of ${}^tX_{11}U_0$ in $M_2(\ZZ_p)/p^eM_2(\ZZ_p)$.
For $X_{21} \in M_2(\ZZ_p)/p^eM_2(\ZZ_p)$ put
\[\calc_e(X_{21})=\calb_e(U_0,-pU_0[X_{21}]+(\epsilon_1 +\epsilon_2)),\]
and for $X_{21} \in M_2(\ZZ_p)/p^eM_2(\ZZ_p)$ and $X_{11} \in \calc_e(X_{21})$ put
\[\calc_e(X_{21},X_{11})=\calb_e(p(pU_0[({}^tX_{11}U_0)^*\ {}^t\!X_{21}U_0]+U_0),\epsilon_3 p).\]
Then, by (b11) and (b$22'$), we have
\begin{align*}
&\#\calb_e(U_0 \bot pU_0,T) \\
&=\sum_{X_{21} \in M_2(\ZZ_p)/p^eM_2(\ZZ_p)} \sum_{X_{11} \in \calc_e(X_{21})} \# \calc_e(X_{21},X_{11}).
\end{align*}
For non-negative integers $e' \le e$, let
$$\pi_{e'}^e:M_{rs}(\ZZ_p)/p^eM_{rs}(\ZZ_p) \longrightarrow
M_{rs}(\ZZ_p)/p^{e'}M_{rs}(\ZZ_p) $$
be the natural projection. For $Y \in M_{rs}(\ZZ_p)/p^eM_{rs}(\ZZ_p)$, we use the same symbol $Y$ to denote the element $\pi_{e'}^e(Y) \in M_{rs}(\ZZ_p)/p^{e'}M_{rs}(\ZZ_p)$. Then $\pi_{e'}^e$ induces a mapping from $\calc_e(X_{21})$ to $\calc_{e'}(X_{21})$, which will also be denoted by $\pi_{e'}^e$.
We have
\[\calc_{1}(X_{21})=\calb_1(U_0,\epsilon_1 \bot \epsilon_2)\]
for any integer $e \ge 1$ and $X_{21} \in M_2(\ZZ_p)/p^eM_2(\ZZ_p)$.
Hence, by the $p$-adic Newton approximation method, we easily see that
\[\#\calc_e(X_{21})=p^{e-1}\#\calc_1(X_{21})=p^{e-1}\#\calb_1(U_0,\epsilon_1 \bot \epsilon_2),\]
and in particular
\[\#\calb_e(U_0,\epsilon_1 \bot \epsilon_2)=p^{e-1}\#\calb_1(U_0,\epsilon_1 \bot \epsilon_2).\]
This implies that
\[p^{-1}\#\calb_1(U_0,\epsilon_1 \bot \epsilon_2)=\beta_p(U_0,\epsilon_1 \bot \epsilon_2),\]
and
\[\#\calc_e(X_{21})=p^e\beta_p(U_0,\epsilon_1 \bot \epsilon_2).\]
Moreover, we have
\[\calc_2(X_{21},X_{11})=\calb_2(pU_0,\epsilon_3 p)\]
for any integer $e \ge 2$, $X_{21} \in M_2(\ZZ_p)/p^eM_2(\ZZ_p)$ and $X_{11} \in \calc_e(X_{21})$. Therefore, in a way similar to above, we have
\[\#\calc_e(X_{21},X_{11})=p^e\beta_p(pU_0,\epsilon_3 p).\]
Hence, we have
\begin{align*}
&\#\calb_e(U_0 \bot pU_0,T)\\
& =p^{2e}\beta_p(U_0,\epsilon_1 \bot \epsilon_2)\beta_p(pU_0,\epsilon_3p)\sum_{X_{21} \in M_2(\ZZ_p)/p^eM_2(\ZZ_p)} 1\\
&=p^{6e} \beta_p(U_0,\epsilon_1 \bot \epsilon_2)\beta_p(pU_0,\epsilon_3 p).
\end{align*}
Therefore,
we have
\[\beta_p(U_0 \bot pU_0,T)=\beta_p(U_0,\epsilon_1 \bot \epsilon_2)\beta_p(pU_0,\epsilon_3 p).\]
We note that $\beta_p(pU_0,\epsilon_3 p)=p\beta_p(U_0,\epsilon_3 )
=p(1+p^{-1})$. Thus the assertion follows from \cite[Theorem 5.6.3]{Ki2}. Next let $T=\epsilon_1 \bot \epsilon_2p \bot \epsilon_3 p$ with $\chi_p(-\epsilon_2\epsilon_3)=-1$. Then, in the same way as above, we have
\[\beta_p(U_0 \bot pU_0,T)=\beta_p(U_0,\epsilon_1) \beta_p(pU_0,\epsilon_2 p \bot \epsilon_3 p).\]
Thus the assertion follows from \cite[Theorem 5.6.3]{Ki2}.
\end{proof}
A non-degenerate ${m \times m}$ matrix $D=(d_{ij})$ with entries in $\ZZ_q$ is said to be {\it reduced} if $D$ satisfies the following two conditions:\begin{enumerate}
\item[(R-1)] For
$i=j$, $d_{ii}=q^{e_{i}}$ with a non-negative integer $e_i$; \vspace*{1mm}
\item[(R-2)] For
$i\ne j$, $d_{ij}$ is a non-negative integer satisfying
$ d_{ij} \le q^{e_j}-1$ if $i <j$ and $d_{ij}=0$ if $i >j$.
\end{enumerate}
It is well known that we can take the set of all reduced matrices as a complete set of representatives of
$GL_m(\ZZ_q) \backslash M_m(\ZZ_q)^{\rm nd}.$
\begin{lemma}\label{lem.uniquenes-of-maximal-element3}
Let $q$ be an odd prime number.
Let $T \in \calh_3(\ZZ_q)^{\rm nd}$ and suppose that $\eta_q(T)=-1$.
Then there is a unique element $g \in GL_3(\ZZ_q) \backslash M_3(\ZZ_q)^{\rm nd}$
such that $T[g^{-1}]$ is maximal.
\end{lemma}
\begin{proof}
Since $\eta_q(T)=-1$, without loss of generality we assume
that
$T=\epsilon_1 q^{a_1} \bot \epsilon_2 q^{a_2} \bot \epsilon_3 q^{a_3}$
with $a_1$ even, and $a_2$ or $a_3$ is odd.
Let $g \in M_3(\ZZ_q)^{\rm nd}$ such that $T[g^{-1}]$ is maximal.
We may assume that
\[g=\begin{pmatrix} q^{e_1} & d_{12} & d_{13} \\
0 & q^{e_2} & d_{23} \\ 0 & 0 & q^{e_3}
\end{pmatrix}\]
satisfies the conditions (R-1) and (R-2).
Then we have
\begin{align*}
&T[g^{-1}]=\\
&{\footnotesize \begin{pmatrix}
\epsilon_1q^{a_1-2e_1} & q^{-e_1}\epsilon_1q^{a_1} d_{12}^*& q^{-e_1}\epsilon_1q^{a_1}d_{13}^*\\
q^{-e_1}\epsilon_1q^{a_1} d_{12}^* & \epsilon_1q^{a_1}(d_{12}^*)^2+ \epsilon_2q^{a_2-2e_2}& d_{12}^*\epsilon_1q^{a_1}d_{13}^*+q^{-e_2}\epsilon_2q^{a_2}d_{23}^*\\
d_{13}^*\epsilon_1q^{a_1-e_1} & d_{13}^*\epsilon_1q^{a_1}d_{12}^*+q^{-e_2}\epsilon_2q^{a_2}d_{23}^* &
\epsilon_1q^{a_1}(d_{13}^*)^2+\epsilon_2 q^{a_2}(d_{23}^*)^2+\epsilon_3 q^{a_3-2e_3}
\end{pmatrix}},
\end{align*}
where
\vspace{1mm}
\\
$d_{12}^*=-q^{-e_1-e_2}d_{12},d_{13}^*=q^{-e_1-e_2-e_3}(d_{12}d_{23}-q^{e_2}d_{13})$ and
$d_{23}^*=-q^{-e_2-e_3}d_{23}.$
\vspace{1mm}
\\
First assume $a_2$ is even and $a_3$ is odd.
Since $T[g^{-1}]$ is maximal, we have $e_1=a_1/2$, and hence
$q^{-e_1}\epsilon_1q^{a_1} d_{12}^*=-\epsilon_1q^{-e_2}d_{12}$, Then, by (R-2), we have $d_{12}=0$. Hence, we have $q^{-e_1}\epsilon_1q^{a_1}d_{13}^*=-\epsilon_1q^{-e_3}d_{13}$, and again by (R-2), we have $d_{13}=0$.
We also have $\epsilon_1q^{a_1}(d_{12}^*)^2+ \epsilon_2q^{a_2-2e_2}=\epsilon_2q^{a_2-2e_2}$, and by the maximality condition, we have $e_2=a_2/2$ and again by (R-2), we have $d_{23}=0$. This proves the uniqueness of $g$.
Next assume that $a_2$ and $a_3$ are odd. Since we have $\eta_q(T)=-1$, we have $\chi_q(-\epsilon_2\epsilon_3)=-1$.
In the same way as above, we can prove the uniqueness of $e_1=a_1/2$ and $d_{12}=d_{13}=0$. Then, $\epsilon_1q^{a_1}(d_{12}^*)^2+ \epsilon_2q^{a_2-2e_2}=\epsilon_2q^{a_2-2e_2}$, and by the maximality condition, we have $e_2=(a_2-1)/2$.
Then, we have $d_{12}^*\epsilon_1q^{a_1}d_{13}^*+q^{-e_2}\epsilon_2q^{a_2}d_{23}^*=-\epsilon_2q^{-e_3+1}d_{23}$, and
therefore, by (R-2), we have $-q^{-e_3+1}d_{23} \in \ZZ_p^\times$ if $d_{23} \not=0$.
Hence, we have
\[\epsilon_1q^{a_1}(d_{13}^*)^2+\epsilon_2 q^{a_2}(d_{23}^*)^2+\epsilon_3 q^{a_3-2e_3}=q^{-1}\epsilon_2(q^{-e_3+1}d_{23})^2+\epsilon_3q^{a_3-2e_3},\]
and it does not belong to $\ZZ_q$ because $\chi_q(-\epsilon_2\epsilon_3)=-1$. This implies that $d_{23}=0$ and the assertion holds.
\end{proof}
\noindent
{\bf Proof of Proposition \ref{prop.local-density-at-p} \,(1).}
The assertion follows from Lemmas
\ref{lem.local-density-via-primitive-local-density}, \ref{lem.explicit-primitive-locla-density3}, and
\ref{lem.uniquenes-of-maximal-element3}.
Next we prove Proposition \ref{prop.local-density-at-p} (2).
The following lemma is easy to prove.
\begin{lemma} \label{lem.maximal-element4}
Let $q$ be an odd prime number.
Suppose that $T \in \calh_4(\ZZ_q)^{\rm nd}$ is maximal. Then
$T$ is one of the following matrices.
\begin{itemize}
\item[{\rm (1)}] $T \sim_{GL_4(\ZZ_q)} \epsilon_1 \bot \epsilon_2 \bot \epsilon_3 \bot \epsilon_4$ with $\epsilon_i \in \ZZ_q^\times \ (i=1,2,3,4)$.
\item[{\rm (2)}] $T \sim_{GL_4(\ZZ_q)} \epsilon_1 \bot \epsilon_2 \bot \epsilon_3 \bot q\epsilon_4$ with $\epsilon_i \in \ZZ_q^\times \ (i=1,2,3,4)$.
\item[{\rm (3)}] $T \sim_{GL_4(\ZZ_q)} \epsilon_1 \bot \epsilon_2 \bot q\epsilon_3 \bot q\epsilon_4$ with $\epsilon_i \in \ZZ_q^\times \ (i=1,2,3,4)$ such that \\ $\chi_q(-\epsilon_3\epsilon_4)=-1$.
\end{itemize}
Moreover, in case {\rm (3)}, we have
$\eta_q(T)=-1 $.
\end{lemma}
\begin{lemma}\label{lem.explicit-primitive-locla-density4}
Let $T \in \calh_4(\ZZ_p)^{\rm nd}$ be maximal.
If $\alpha_p(U_0 \bot pU_0,T) \not=0$, then
$$T \sim_{GL_4(\ZZ_p)} \epsilon_1 \bot \epsilon_2 \bot p\epsilon_3 \bot p\epsilon_4$$
such that $\epsilon_1 \epsilon_2\epsilon_3 \epsilon_4 \in (\ZZ_p^\times)^2$ and $\chi_p(-\epsilon_1\epsilon_2)=\chi_p(-\epsilon_3\epsilon_4)=-1$.
Conversely, if $T$ satisfies these conditions, then
\begin{align*}
\alpha_p(U_0 \bot pU_0, T)=2(1+p^{-1})^2 p.
\end{align*}
\end{lemma}
\begin{proof}
By the assumption, $\det (2T)$ is divided by $p^2$, and $T$ must be the case (3) of Lemma \ref{lem.maximal-element4}, $T$.
Moreover, in this case we have $p^{-2}\det (2T)=\prod_{i=1}^4 \epsilon_i=\epsilon^2$ with $\epsilon \in \ZZ_p^\times$. Hence, we have $\chi_p(-\epsilon_1\epsilon_2)=\chi_p(-\epsilon_3\epsilon_4)=-1$.
Hence, $T \sim_{GL_4(\ZZ_p)} U_0 \bot pU_0$, and the second assertion follows from \cite[Theorem 6.8.1]{Ki2}.
\end{proof}
\begin{lemma}\label{lem.uniquenes-of-maximal-element4} Let $q$ be an odd prime number.
Let $T \in \calh_4(\ZZ_q)^{\rm nd}$ and suppose that $\eta_q(T)=-1$. Then there is a unique element $g \in GL_4(\ZZ_q) \backslash M_4(\ZZ_q)^{\rm nd}$ such that $T[g^{-1}]$ is maximal.
\end{lemma}
\begin{proof}
The assertion can be proved in the same manner as Lemma \ref{lem.uniquenes-of-maximal-element3}.
\end{proof}
\noindent
{\bf Proof of Proposition \ref{prop.local-density-at-p}\,(2).}
The assertion follows from
Lemmas \ref{lem.local-density-via-primitive-local-density}, \ref{lem.explicit-primitive-locla-density4},
and \ref{lem.uniquenes-of-maximal-element4}.
\vspace{3mm}
\\
{\bf Proof of Theorem \ref{th.main-result}.}
(1) If $\eta_p(T)=1$, by Theorem \ref{th.limit-of-FC-of-Eisenstein} (1) and Propositions \ref{prop.mass-formula} and
\ref{prop.local-density-at-p}, we have
\[a(\widetilde E_2^{(3)},T)=a(\mathrm{genus} \ \Theta^{(3)}(S^{(p)}),T)=0.\]
Suppose that $\eta_p(T)=-1$. Then, we have
\[a(\mathrm{genus} \ \Theta^{(3)}(S^{(p)}),T)=16 (1+p)^2 p^{-4} \pi^4 \prod_{q \not=p} (1-q^{-2})^2 \prod_{q \not=p} F_q(T,q^{-2}).\]
We have
\[\prod_{q \not=p} (1-q^{-2})^2=\zeta(2)^{-2}(1-p^{-2})^{-2}=\frac{36}{\pi^4(1-p^{-2})^2}.\]
Hence, we have
\[a(\mathrm{genus} \ \Theta^{(3)}(S^{(p)}),T)=\frac{576}{(p-1)^2} \prod_{q \not=p} F_q(T,q^{-2}).\]
This coincides with $a(\widetilde E_2^{(3)},T)$ in view of Theorem \ref{th.limit-of-FC-of-Eisenstein} (1).
(2) If $p^{-2}\det (2T)$ is not a square integer or $\eta_p(T)=1$, by Theorem \ref {th.limit-of-FC-of-Eisenstein} (2), and Propositions \ref{prop.mass-formula} and
\ref{prop.local-density-at-p}, we have
\[a(\widetilde E_2^{(4)},T)=a(\mathrm{genus} \ \Theta^{(4)}(S^{(p)}),T)=0.\]
Suppose that $p^{-2}\det (2T)$ is a square integer and $\eta_p(T)=-1$. Then, we have
\[a(\mathrm{genus} \ \Theta^{(4)}(S^{(p)}),T)=32 (1+p)^2 p^{-4} \pi^4 \prod_{q \not=p} (1-q^{-2})^2 \prod_{q \not=p} F_q(T,q^{-3}).\]
We have
\[\prod_{q \not=p} (1-q^{-2})^2=\zeta(2)^{-2}(1-p^{-2})^{-2}=\frac{36}{\pi^4(1-p^{-2})^2}.\]
Hence, we have
\[a(\mathrm{genus} \ \Theta^{(4)}(S^{(p)}),T)=\frac{1152}{(p-1)^2} \prod_{q \not=p} F_q(T,q^{-3}).\]
This coincides with $a(\widetilde E_2^{(4)},T)$ in view of Theorem \ref{th.limit-of-FC-of-Eisenstein} (2).
These complete the proof of Theorem \ref{th.main-result}.
\subsection{General case}
To prove the main theorem, we must consider the case $n\geq 5$. For the Siegel operator $\Phi$,
we have
$$
\begin{cases}
\Phi (\widetilde{E}_2^{(n)})=\widetilde{E}_2^{(n-1)},\\
\Phi (\text{genus}\,\Theta^{(n)}(S^{(p)}))=\text{genus}\,\Theta^{(n-1)}(S^{(p)}).
\end{cases}
$$
From this, it is sufficient to prove the following proposition.
\begin{proposition}
\label{n5}
\quad Assume that $n\geq 5$. For any $T\in\Lambda_n^+$, we have
$$
(*)\qquad\qquad \qquad a(\widetilde{E}_2^{(n)},T)=a({\rm genus}\,\Theta^{(n)}(S^{(p)}),T)=0.
$$
\end{proposition}
\begin{proof}
First we consider ${\rm genus}\,\Theta^{(n)}(S^{(p)})$.
Since $S^{(p)}\in\Lambda_4^+$ and $T\in\Lambda_n^+$\;$(n\geq 5)$,
$a(\theta^{(n)}(S_i;Z),T)=0$ holds for each theta series $\theta^{(n)}(S_i;Z)$. Hence, we obtain
$$
a({\rm genus}\,\Theta^{(n)}(S^{(p)}),T)=0.
$$
Next we investigate $\widetilde{E}_2^{(n)}$ and prove
$$
\label{vanishE2}
(**)\qquad\qquad\qquad \lim_{m\to\infty}a(E_{k_m}^{(n)},T)=a(\widetilde{E}_2^{(n)},T)=0
$$
for any $T\in\Lambda_n^+$.
We recall the formula for $a(E_k^{(n)},T)$
given in Proposition \ref{prop.FC-Siegel}.
We extract the following factor in the formula for $a(E_k^{(n)},T)$:
$$
A_{k,n}(T):=\frac{k}{B_k}\cdot\prod_{i=1}^{[n/2]}\frac{k-i}{B_{2k-2i}}\cdot
\begin{cases}
\frac{B_{k-n/2,\chi_T}}{k-n/2} & (\text{if $n$ is even}),\\
1 & (\text{if $n$ is odd}).
\end{cases}
$$
To prove $(**)$, it suffices to show that
$$
(\ddagger)\qquad\qquad\qquad \lim_{m\to\infty}A_{k_m,n}(T)=0\quad (\text{$p$-adically}).
$$
First we assume that $p>3$.
By Kummer's congruence for Bernoulli numbers, the factors
$$
\frac{k_m}{B_{k_m}}\;\;\text{and}\;\;\frac{k_m-i}{B_{2k_m-2i}}\;\;(1\leq i \leq [n/2]
\;\;\text{with}\;\;
i \not\equiv 2 \pmod{(p-1)})
$$
in $A_{k_m,n}(T)$ have $p$-adic limits when $m\longrightarrow \infty$.
\vspace{2mm}
\\
We forcus our attention on the factors
$$
\frac{k_m-i}{B_{2k_m-2i}}\;\;\text{for}\;\; i\;\;\text{with}\quad i \equiv 2 \pmod{(p-1)}.
$$
In these cases, by the von Staudt-Clausen theorem, we obtain
$$
\text{ord}_p\left(\frac{k_m-i}{B_{2k_m-2i}}\right)\geq 1.
$$
In particular, in the case of $i=2\leq [n/2]$, the following identity holds:
$$
\text{ord}_p\left(\frac{k_m-2}{B_{2k_m-4}}\right)=m.
$$
( Such $i$ also appears when $n=4$. See Remark \ref{cacellation}.) Consequently, there is a constant $C$ such that
$$
\text{ord}_p\left(\frac{k_m}{B_{k_m}}\cdot\prod_{i=1}^{[n/2]}\frac{k_m-i}{B_{2k_m-2i}}\right)\geq
C+m
$$
for sufficiently large $m$. Regarding the factor $B_{k_m-n/2,\chi_T}/(k_m-n/2)$, we use
Carlitz's result (\cite{Carlitz2}) for the generalized Bernoulli numbers in the case of quadratic
Dirichlet characters. We have
\begin{align*}
\text{ord}_p\left( \frac{B_{k_m-n/2,\chi_T}}{k_m-n/2}\right)
& =\text{ord}_p\left(B_{k_m-n/2,\chi_T}\right)
-\text{ord}_p\left(k_m-n/2\right)\\
& \geq -1-\text{ord}_p\left(2-n/2\right),
\end{align*}
for sufficiently large $m$. By assumption $n\geq 5$, the values
$$
\text{ord}_p\left( \frac{B_{k_m-n/2,\chi_T}}{k_m-n/2}\right)
$$
have a lower bound for sufficiently large $m$.
\vspace{1mm}
\\
Combining these results, we can prove $(\ddagger)$ when $p>3$.
Next we consider the case $p=3$. By von Staudt-Clausen theorem, we obtain
$$
\text{ord}_3\left(\frac{k_m}{B_{k_m}}\right)=1,\quad
\text{ord}_3\left(\frac{k_m-i}{B_{2k_m-2i}}\right)=\text{ord}_3(k_m-i)+1\geq 1
\quad (1\leq i\leq [n/2]).
$$
In particular, when $i=2\geq [n/2]$, the following identity holds:
$$
\text{ord}_3\left(\frac{k_m-2}{B_{2k_m-4}}\right)=m.
$$
Since we know that $\text{ord}_3(B_{k_m-n/2,\chi_T}/(k_m-n/2))$ has a lower bound as in the case
$p>3$, the statement $(\ddagger)$ is also proven in the case $p=3$.
This completes the proof of Proposition \ref{n5}.
\end{proof}
\noindent
\begin{remark}
\label{cacellation}
\quad An exceptional factor $(k_m-2)/\,B_{2k_m-4}$ in the product
\\
$\prod (k_m-i)/\,B_{2k_m-2i}$
also appears when $n=4$. However, in this case, cancellation occurs between
$$
\frac{k_m-2}{B_{2k_m-4}}\quad \text{and}\quad \frac{B_{k_m-2,\chi_T}}{k_m-2}.
$$
\vspace{2mm}
\end{remark}
By combining Corollary \ref{cor.main-result2} and Proposition \ref{n5}, we have
proved our main result, Theorem \ref{statementmain}.
\section{Applications}
\subsection{Modular forms on $\boldsymbol{\Gamma_0^{(n)}(p)}$}
In \cite{Serre}, Serre proved the following result.
\begin{theorem}
{\rm (Serre \cite{Serre})}\quad Let $p$ be an odd prime number and $\mathbb{Z}_{(p)}$ the local
ring consisting of $p$-integral rational numbers. For any $f\in M_2(\Gamma^{(1)}_0(p))_{\mathbb{Z}_{(p)}}$,
there is a modular form $g\in M_{p+1}(\Gamma^{(1)})_{\mathbb{Z}_{(p)}}$ satisfying
$$
f \equiv g \pmod{p}.
$$
\end{theorem}
An attempt to generalize this result to the case of Siegel modular forms can be found in \cite{BN2}.
\\
Here we consider the first $p$-adic approximation of $\widetilde{E}_2^{(n)}$, that is,
$$
E_{k_1}^{(n)}=E_{p+1}^{(n)}.
$$
\begin{theorem}
Let $p$ be a prime number such that $p>n$. The modular form
$\widetilde{E}_2^{(n)}\in M_2(\Gamma_0^{(n)}(p))_{\mathbb{Z}_{(p)}}$ is
congruent to $E_{p+1}^{(n)}\in M_{p+1}(\Gamma^{(n)})_{\mathbb{Z}_{(p)}}$ mod $p$:
$$
\widetilde{E}_2^{(n)} \equiv E_{p+1}^{(n)} \pmod{p}.
$$
\end{theorem}
The above result provides an example of Serre's type congruence in the case of Siegel
modular forms. ( For $n=2$, this theorem has already been proved in \cite[\sc Propositon 4]{KN}.
The $p$-integrality of $\widetilde{E}_2^{(n)}$ comes from the explicit
formula for the Fourier coefficients.)
\subsection{Theta operators}
\label{ThetaOp}
For a Siegel modular form $\displaystyle F=\sum a(F,T)q^T$, we define
$$
\varTheta (F):=\sum a(F,T)\cdot\text{det}(T)\,q^T\in\mathbb{C}[q_{ij}^{-1},q_{ij}][\![q_1,\ldots,q_n]\!].
$$
The operator $\varTheta$ is called the {\it theta operator}. This operator was first studied by
Ramanujan in the case of elliptic modular forms, and the generalization to the case of
Siegel modular forms can be found in \cite{BN1}.
If a Siegel modular form $F$ satisfies
$$
\varTheta (F) \equiv 0 \pmod{N},
$$
we call it an element of the space of the {\it mod $N$ kernel of the theta operator}. For example,
Igusa's cusp form $\chi_{35}\in M_{35}(\Gamma^{(2)})_{\mathbb{Z}}$ satisfies the congruence
relation
$$
\varTheta (\chi_{35}) \equiv 0 \pmod{23}\qquad\qquad (\text{cf. \cite{K-K-N}}),
$$
namely, $\chi_{35}$ is an element of the space of mod $23$ kernel of the theta operator.
\begin{theorem}
\label{modpaquare}
Assume that $p\geq 3$. Then we have
$$
\varTheta (E_{p+1}^{(3)}) \equiv 0 \pmod{p},\qquad
\varTheta (E_{p^2-p+2}^{(4)}) \equiv 0 \pmod{p^2}.
$$
The second congruence shows that the Siegel Eisenstein series $E_{p^2-p+2}^{(4)}$ is
an element of the mod $p^2$ kernel of the theta operator.
\end{theorem}
\begin{proof}
To prove the first congruence relation, we consider the first approximation of $\widetilde{E}_2^{(3)}$:
$$
\widetilde{E}_2^{(3)} \equiv E_{2+p-1}^{(4)}=E_{p+1}^{(3)} \pmod{p}.
$$
If $T\in\Lambda^+_3$ satisfies $a({\rm genus}\,\Theta^{(3)}(S^{(p)}),T)=a(\widetilde{E}_2^{(3)},T)\ne 0$,
then, by Lemma \ref{lem.explicit-primitive-locla-density3}, we have $\text{det}(2T) \equiv 0 \pmod{p}$.
This fact implies that
$$
\varTheta (E_{p+1}^{(3)}) \equiv \varTheta(\widetilde{E}_2^{(3)})
= \varTheta ({\rm genus}\,\Theta^{(3)}(S^{(p)}) )\equiv 0 \pmod{p}.
$$
We consider the second congruence relation.
Considering the second $p$-adic approximation of $\widetilde{E}_2^{(4)}$, we obtain
$$
\widetilde{E}_2^{(4)} \equiv E_{2+(p-1)p}^{(4)}=E_{p^2-p+2}^{(4)} \pmod{p^2}.
$$
Therefore, it is sufficient to prove that
$$
\varTheta (\widetilde{E}_2^{(4)}) \equiv 0 \pmod{p^2}.
$$
Assume that $a(\widetilde{E}_2^{(4)},T)\ne 0$ for $T\in\Lambda_4^+$. Then $\text{det}(2T)$
is square by Theorem \ref{th.limit-of-FC-of-Eisenstein}, (2). Under this condition,
if we further
assume that $\text{det}(2T)\not\equiv 0 \pmod{p}$, then we have
$a(\widetilde{E}_2^{(4)},T)=0$ by Theorem \ref{th.limit-of-FC-of-Eisenstein}, (2.2).
This is a contradiction. Therefore, we have $\text{det}(2T) \equiv 0 \pmod{p}$, equivalently,
$\text{det}(2T) \equiv 0 \pmod{p^2}$. This means that,
if $a(\widetilde{E}_2^{(4)},T)\ne 0$ for $T\in\Lambda_4^+$, then $T\in\Lambda_4^+$ satisfies
$\text{det}(2T) \equiv 0 \pmod{p^2}$. This implies
$$
\varTheta (\widetilde{E}_2^{(4)}) \equiv 0 \pmod{p^2}
$$
and completes the proof of Theorem \ref{modpaquare}.
\end{proof}
\noindent
\begin{remark}
The first congruence relation in Theorem \ref{modpaquare} can also be shown as a special
case of the main theorem in \cite[Theorem 2.4]{N-T}.
\end{remark}
\subsection{Numerical examples}
In this section, we provide examples of Fourier coefficients of $\widetilde{E}_2^{(n)}$
and \\
genus\,$\Theta^{(n)}(S^{(p)})$, which certify the validity of our identity in our main result.
\vspace{2mm}
\\
\fbox{\textbf{Case $\boldsymbol{n=3:}$}}
\vspace{2mm}
\\
We take
$$
p=11,\quad T={\scriptsize \begin{pmatrix} 1 & 0 & \tfrac{1}{2} \\
0 & 1 & 0 \\
\tfrac{1}{2} & 0 & 3 \end{pmatrix}}
\in \Lambda_3^+\quad\text{with}\quad\text{det}(T)=11/4,
$$
and calculate
$a(\widetilde{E}_2^{(3)},T)$ and $a(\text{genus}\,\Theta^{(3)}(S^{(p)}),T)$.
\vspace{2mm}
\\
\textbf{Calculation of} $\boldsymbol{a(\widetilde{E}_2^{(3)},T)}$:
\vspace{2mm}
\\
By Theorem \ref{th.limit-of-FC-of-Eisenstein}, (1),
$$
a(\widetilde{E}_2^{(3)},T)=\frac{576}{(1-11)^2}\cdot\lim_{m\to\infty}(1-11^{10\cdot 11^{m-1}})=
\frac{144}{25}.
$$
\textbf{Calculation of} $\boldsymbol{a({\rm genus}\,\Theta^{(3)}(S^{(11)}),T)}$:
\vspace{2mm}
\\
We can take three representatives of $GL_4(\mathbb{Z})$-equivalence classes in genus$(S^{(11)})$:
$$
{\scriptsize
S_1^{(11)}:=\begin{pmatrix} 1 & 0 & \tfrac{1}{2} & 0 \\ 0 & 1 & 0 & \tfrac{1}{2} \\
\tfrac{1}{2} & 0 & 3 & 0 \\ 0 & \tfrac{1}{2} & 0 & 3
\end{pmatrix},\;
S_2^{(11)}:=\begin{pmatrix} 1 & \tfrac{1}{2} & \tfrac{1}{2} & \tfrac{1}{2} \\ \tfrac{1}{2} & 1 & 0 & \tfrac{1}{2} \\
\tfrac{1}{2} & 0 & 4 & 2 \\ \tfrac{1}{2} & \tfrac{1}{2} & 2 & 4
\end{pmatrix},\;
S_3^{(11)}:=\begin{pmatrix} 2 & 1 & \tfrac{1}{2} & \tfrac{1}{2} \\ 1 & 2 & 0 & \tfrac{1}{2} \\
\tfrac{1}{2} & 0 & 2 & 1 \\ \tfrac{1}{2} & \tfrac{1}{2} & 1 & 2
\end{pmatrix}.
}
$$
Moreover we have
$$
a(S_1^{(11)},S_1^{(11)})=32,\quad a(S_2^{(11)},S_2^{(11)})=72,
\quad a(S_3^{(11)},S_3^{(11)})=24.\qquad (\text{cf.\;\cite{Nipp}}).
$$
Therefore,
\begin{align*}
& a(\text{genus}\,\Theta^{(3)}(S^{(11)}),T)=\\
&[a(\theta^{(3)}(S_1^{(11)};Z),T)/32+a(\theta^{(3)}(S_2^{(11)};Z),T)/72+a(\theta^{(3)}(S_3^{(11)};Z),T)/24]\\
& \cdot [(1/32)+(1/72)+(1/24)]^{-1}.
\end{align*}
Direct calculations show that
$$
a(\theta^{(3)}(S_1^{(11)};Z),T)=16,\qquad
a(\theta^{(3)}(S_2^{(11)};Z),T)=a(\theta^{(3)}(S_3^{(11)};Z),T)=0
$$
for the above $T\in\Lambda_3^+$.
Hence, we have
$$
a(\text{genus}\,\Theta^{(3)}(S^{(11)}),T)=16 \times\frac{9}{25}=\frac{144}{25}.
$$
Of course, this value is consistent with the value obtained using the equations
given in Propositions \ref{prop.mass-formula}\, (1) and \ref{prop.local-density-at-p}\, (1):
\begin{align*}
a(\text{genus}\,\Theta^{(3)}(S^{(11)}),T) & = \frac{8}{11^3}\pi^4
\cdot \alpha_{11}(U_0\,\bot\,11\cdot U_0,T)
\cdot \prod_{q\ne 11}\alpha_q(H_2,T) \\
& = \frac{8}{11^3}\pi^4
\cdot 2(1+11)(1+11^{-1})
\cdot \frac{36}{\pi^4(1-11^{-2})^2}\\
&=\frac{144}{25}.
\end{align*}
Further numerical examples for the case $n=3$ can be found in \cite{Okuma}.
\vspace{2mm}
\\
\fbox{\textbf{Case $\boldsymbol{n=4:}$}}
\vspace{2mm}
\\
The values $a(\text{genus}\,\Theta^{(4)}(S^{(p)}),T)$ at $T=S^{(p)}$ can be
calculated as
$$
a(\text{genus}\,\Theta^{(4)}(S^{(p)}),S^{(p})=\left(\sum_{i=1}^d\frac{1}{a(S_i,S_i)}\right)^{-1}=M(S^{(p)})^{-1}
$$
and the values $M(S^{(p)})$ for small $p$ can be found in \cite{Nipp}:
\begin{table}[hbtp]
\begin{center}
\begin{tabular}{c | cccccc}
$p$ & $3$ & $5$ & $7$ & $11$ & $13$ & $\cdots$
\\
\hline
$M(S^{(p)})^{-1}$ & $288$ & $72$ & $32$ & $\frac{288}{25}$& $8$ & $\cdots$
\end{tabular}
\end{center}
\end{table}
\vspace{1mm}
\\
These numerical data can be verified to be consistent with those obtained from the
the formula
$$
a(\widetilde{E}_2^{(4)},S^{(p)})=\frac{1152}{(p-1)^2},
$$
which is a result of Theorem \ref{th.limit-of-FC-of-Eisenstein} (2.2).
\section{Acknowledgements}
$\bullet$\quad We thank Prof. Fumihiro Sato for valuable discussion.
\vspace{1mm}
\\
$\bullet$\quad
This work was supported by
JSPS KAKENHI: first author, Grant-in-Aid (B) (No.16H03919) and Grant-in-Aid (C) (No. 21K03152);
second author, Grant-in-Aid (C) ( No. 20K03547).
|
\section{Introduction}
The numerical solution of fluid flow problems is an important topic in computational science and engineering, which has received much attention in the last few decades \cite{bathe2007benchmark,brandt1979multigrid,brandt2011multigrid,connor2013finite,vanka1986block,wesseling2001geometric}. Stokes-Darcy Brinkman problem is one of them used to model fluid motion in porous media with fractures. The discretization of the fluid flow problems often leads to a saddle-point system, which is ill-conditioned. Designing fast numerical solution of these problems is often challenging due to the small magnitude of physical parameters of the model.
We consider the multigrid numerical solution of the Stokes-Darcy Brinkman equations
\begin{subequations}
\label{eq:SDB}
\begin{align}
- \epsilon^2 \Delta\boldsymbol{u} +\boldsymbol{u} + \nabla p =&\boldsymbol{f} \qquad \text{in}\,\, \Omega \label{eq:stokes-DB-one}\\
\nabla \cdot \boldsymbol{u}=&g \qquad \text{in}\,\, \Omega \label{eq:stokes-DB-two} \\
\boldsymbol{u} =&0 \qquad \text{on}\,\, \partial \Omega \label{eq:stokes-DB-three},
\end{align}
\end{subequations}
where $\epsilon \in(0,1]$. The source term $g$ is assumed to satisfy the solvability condition
\begin{equation*}
\int_{\Omega} g\, d\Omega =0.
\end{equation*}
Then, equations \eqref{eq:stokes-DB-one}, \eqref{eq:stokes-DB-two}, and \eqref{eq:stokes-DB-three} have a unique solution.
A variety of discretization schemes are available for equations \eqref{eq:stokes-DB-one}, \eqref{eq:stokes-DB-two}, and \eqref{eq:stokes-DB-three}, including finite element methods \cite{gulbransen2010multiscale,vassilevski2014mixed,xie2008uniformly,xu2010new,zhai2016new,zhang2009low,zhao2020new}, finite difference techniques \cite{he2017finite,sun2019stability}, and divergence-conforming B-spline methods \cite{evans2013isogeometric}. When $\epsilon= 0$, the model problem is reduced to the Darcy problem \cite{arraras2021multigrid,harder2013family}. For $\epsilon \in(0,1]$, designing a robust discretization and numerical solver is challenging. The convergence rate deteriorates as the Stokes-Darcy Brinkman becomes Darcy-dominating when certain stable Stokes elements are used \cite{hannukainen2011computations}, for example, Taylor–Hood element. While, as the Stokes-Darcy Brinkman problem becomes Stokes-dominating when Darcy stable elements such as the lowest order Raviart-Thomas elements are used, degradation on convergence is observed \cite{mardal2002robust}.
Upon discretization, large-scale indefinite linear systems typically need to be solved, at times repeatedly. For saddle-point systems,
within the context of multigrid, there are several effective block-structured relaxation schemes for solving such linear systems, such as Braess-Sarazin smoother \cite{braess1997efficient, he2018local,MR1810326}, distributive smoother \cite{chen2015multigrid,he2018local}, Schwarz-type smoothers \cite{schoberl2003schwarz}, Vanka smoother \cite{claus2021nonoverlapping,MR3217219,manservisi2006numerical,molenaar1991two,wobker2009numerical}, and Uzawa-type relaxation \cite{MR1451114,MR3217219,MR1302679,luo2017uzawa}.
We note also that a number of effective preconditioning methods are available for the Stokes-Darcy Brinkmans problems, for example the scalable block diagonal preconditioner \cite{vassilevski2013block}, and Uzawa-type preconditioning \cite{kobelkov2000effective,sarin1998efficient}.
Multigrid methods are studied in depth \cite{butt2018multigrid,coley2018geometric,kanschat2017geometric,larin2008comparative,olshanskii2012multigrid}. Braess-Sarazin, Uzawa, and Vanka smoothers within multigrid with finite element discretization have been discussed \cite{larin2008comparative}. However, the convergence rate is highly dependent on physical parameters. A Gauss–Seidel smoother based on a Uzawa-type iteration is studied \cite{MR3217219}, where the authors provide an upper bound on the smoothing factor. Moreover, the performance of Uzawa with a Gauss–Seidel type coupled Vanka smoother \cite{vanka1986block} has been investigated \cite{MR3217219}, in which the pressure and the velocities in a grid cell, are updated simultaneously, showing that the actual convergence of the W-cycle of Uzawa is approximately the same as that obtained by the Vanka smoother.
Our interest is in the marker and cell scheme (MAC), a finite difference method on a staggered mesh. On a uniform mesh discretization, the method is second-order accurate for both velocity and pressure \cite{sun2019stability}. We propose a Vanka-type Braess-Sarazin relaxation (V-BSR) scheme for the Stokes-Darcy Brinkman equations discretized by the MAC scheme on staggered meshes. In contrast to the Vanka smoother \cite{MR3217219}, our work builds an algorithm that decouples velocity and pressure, which is often preferred considering the cost efficiency. Specifically, in our relaxation scheme, the shifted Laplacian operator, $ - \epsilon^2 \Delta\boldsymbol{u} +\boldsymbol{u}$, is solved by an additive Vanka-type smoother. Instead of solving many subproblems involved in Vanka setting, we derive the stencil of the Vanka smoother, which means that we can form the global matrix of the Vanka smoother. As a result, in our multigrid method we only have matrix-vector products. This represents significant savings compared to traditional methods that require computationally expensive exact solves; in V-BSR, we solve the Schur complement system by only two or three iterations of the Jacobi method, which achieves the same performance as that of exact solve. We apply local Fourier analysis (LFA) to select the multigrid damping parameter and predict the actual multigrid performance. From this analysis, we derive an optimal damping parameter and optimal smoothing factors. Those parameters are dependent on physical parameters and the meshsize, which means that we can propose adaptive damping parameter in each multigrid level. The optimal parameter turns out to be close to one and relatively insensitive to physical parameters and meshsize. This allows for an easy choice of an approximately optimal damping parameter. We quantitatively compare the results with optimal parameter and the value of one from LFA and present numerical results of two-grid and V-cycle multigrid to validate the high efficiency of our methods. Our V-cycle results outperform these of Uzawa and Vanka smoothers \cite{MR3217219}, especially for small $\epsilon$.
The rest of the work is organized as follows. In Section \ref{sec:Discretization-relaxation} we review the MAC scheme for our model problem and propose the afore-mentioned Vanka-based Braess-Sarazin relaxation. We apply LFA to study the smoothing process in Section \ref{sec:LFA}, where optimal LFA smoothing factor is derived. In Section \ref{sec:numerical} we present our LFA predictions for the two-grid method and actual multigrid performance. Finally, we draw conclusions in Section \ref{sec:con}.
\section{Discretization and relaxation}\label{sec:Discretization-relaxation}
As mentioned in the Introduction, we use throughout the well-known MAC scheme to solve~\eqref{eq:SDB}. For the discretization of \eqref{eq:SDB}, a staggered mesh is needed to guarantee numerical stability. The discrete unknowns $u,v,p,$ are placed in different locations; see Figure \ref{fig:MAC-stokes-brinkman}. The stability and convergence of the MAC scheme for this problem has been studied \cite{sun2019stability}.
\begin{figure}[H]
\centering
\includegraphics[width=0.5\textwidth]{./figures/figure_MAC_StokesBrinkman.pdf}
\caption{The location of the unknowns in the staggered grid: $\Box-u,\,\, \lozenge-v, \,\, \bigcirc-p$.}\label{fig:MAC-stokes-brinkman}
\end{figure}
The stencil representation of MAC for the Stokes-Darcy Brinkmann equations is
\begin{equation}\label{eq:Kh--stencil-operator}
\mathcal{K}_h =\begin{pmatrix}
- \epsilon^2\Delta_{h} + I & 0 & (\partial_{x})_{h/2}\\
0 & - \epsilon^2 \Delta_{h} + I & (\partial_{y})_{h/2} \\
- (\partial_{x})_{h/2} & - (\partial_{y})_{h/2} & 0
\end{pmatrix},
\end{equation}
where
\begin{equation*}
-\Delta_{h} =\frac{1}{h^2}\begin{bmatrix}
& -1 & \\
-1 & 4 & -1 \\
& -1 &
\end{bmatrix},\quad
(\partial_{x})_{h/2} =\frac{1}{h}\begin{bmatrix}
-1& 0 & 1 \\
\end{bmatrix},\quad
(\partial_{y})_{h/2} =\frac{1}{h}\begin{bmatrix}
1 \\
0 \\
-1
\end{bmatrix}.
\end{equation*}
After discretization, the corresponding linear system is
\begin{equation}\label{eq:linear-system}
\mathcal{K}_h \boldsymbol{x}_h=\begin{pmatrix}
\mathcal{A} & \mathcal{B}^T\\
\mathcal{B} & 0
\end{pmatrix}
\begin{pmatrix}
\boldsymbol{u}_h
\\
p_h
\end{pmatrix}
=\begin{pmatrix}
\boldsymbol{f}_h
\\
g_h
\end{pmatrix}=b_h,
\end{equation}
where $\mathcal{A}$ is the matrix corresponding to the discretization of $ - \epsilon^2 \Delta\boldsymbol{u} +\boldsymbol{u}$, and $\mathcal{B}^T$ is the discrete gradient.
In order to solve \eqref{eq:linear-system} efficiently by multigrid we use BSR, with the smoother
\begin{equation}\label{eq:Mh-form}
\mathcal{M}_h= \begin{pmatrix}
\mathcal{C} & \mathcal{B}^T\\
\mathcal{B} & 0
\end{pmatrix},
\end{equation}
where $\mathcal{C}$ is an approximation to $\mathcal{A}$. In the context of preconditioning, such an approach is known as {\em constraint preconditioning} \cite{keller2000constraint,chidyagwai2016constraint,rees2007preconditioner} and it has received quite a bit of attention due to its attactive property of computing interim approximate solutions that satisfy the constraints.
A number of studies \cite{he2018local,YH2021massStokes} have shown that the efficiency of solving the Laplacian will determine the convergence of BSR. To construct an efficient approximation $\mathcal{C}$, we first investigate the discrete operator $-\epsilon^2 \Delta u +u $ denoted by
\begin{equation}\label{eq:positive-shift-Laplace-system}
L= A+ I,
\end{equation}
where $A$ corresponds to the five-point discretization of operator $-\epsilon^2 \Delta u$. The stencil notation for the discrete operator $-\epsilon^2 \Delta u +u $ is
\begin{equation}\label{eq:shift-Laplace-stencil}
L= \frac{\epsilon^2}{h^2}
\begin{bmatrix}
& -1 & \\
-1 & 4+\frac{h^2}{\epsilon^2} & -1 \\
& -1 &
\end{bmatrix}=\frac{\epsilon^2}{h^2}
\begin{bmatrix}
& -1 & \\
-1 & 4+r & -1 \\
& -1 &
\end{bmatrix},
\end{equation}
where $r=\frac{h^2}{\epsilon^2}$. When $r=0$, \eqref{eq:shift-Laplace-stencil} reduces to the discretization of $-\epsilon^2 \Delta u$.
Recently, we proposed an additive element-wise Vanka smoother \cite{CH2021addVanka} for $\Delta u$. Our current goal is to extend our approach to \eqref{eq:positive-shift-Laplace-system}. An immediate challenge here contrary to \cite{CH2021addVanka} is the difference in scale between the discretized scaled Laplacian and the identity operator.
Denote the element-wise smoother as $M_e$, which has the form
\begin{equation}\label{eq:Vanka-operator}
M_e = \sum_{j=1}^{N} V_j^T D_j L_j^{-1} V_j,
\end{equation}
where $D_j=\frac{1}{4}I$ with $I$ be the $4\times 4$-identity matrix, $L_j$ is the coefficient matrix of $j$-th subproblem defined for one element, and $V_j$ is a restriction operator mapping the global vector to the $j$-th subproblem.
We consider
\begin{equation*}
\mathcal{C}^{-1} =
\begin{pmatrix}
M_e & 0\\
0 & M_e
\end{pmatrix}.
\end{equation*}
The relaxation scheme for \eqref{eq:linear-system} is
\begin{equation}\label{eq:BSR-relax-scheme}
\boldsymbol{x}^{k+1}_h = \boldsymbol{x}^k_h+\omega \mathcal{M}^{-1}_h(b_h- \mathcal{K}_h\boldsymbol{x}^k_h).
\end{equation}
We refer to the above relaxation as {\em Vanka-based Braess-Sarazin relaxation} (V-BSR).
Let $b_h- \mathcal{K}_h\boldsymbol{x}^k_h=( r_{ \boldsymbol{u}}, r_p)$. In \eqref{eq:BSR-relax-scheme}, we need to solve for
$(\delta \boldsymbol{u}, \delta p)=\mathcal{M}^{-1}_h(r_{\boldsymbol{u}}, r_p)$ by
\begin{eqnarray}
(\mathcal{B}\mathcal{C}^{-1}\mathcal{B}^{T})\delta p&=&\mathcal{B}\mathcal{C}^{-1} r_{ \boldsymbol{u}}- r_{p}, \label{eq:solution-schur-complement}\\
\delta \boldsymbol{u}&=& \mathcal{C}^{-1}(r_{ \boldsymbol{u}}-\mathcal{B}^{T}\delta p).\nonumber
\end{eqnarray}
Solving \eqref{eq:solution-schur-complement} exactly is prohibitive and impractical in the current context, and it has been shown in a few studies \cite{he2018local,MR1810326} that an inexact solve can be applied and perform well. In the sequel we will present a smoothing analysis for the exact solve, but in practice, for assessing the performance of the multigrid scheme we apply a few iterations of weighted Jacobi to \eqref{eq:solution-schur-complement}.
The relaxation error operator for \eqref{eq:BSR-relax-scheme} is given by
\begin{equation}\label{eq:relaxation-error-operator}
\mathcal{S}_h = I-\omega \mathcal{M}^{-1}_h\mathcal{K}_h,
\end{equation}
where $\omega$ is a damping parameter to be determined.
For a two-grid method, the error propagation operator is
\begin{equation}\label{eq:two-grid-error-operator}
E_h=S^{\nu_2}_h (I- P_h(L_{2h})^{-1} R_h L_h) S^{\nu_1}_h,
\end{equation}
where $L_{2h}$ is rediscretization for the coarse-grid operator and the integers $\nu_1$ and $\nu_2$ are the numbers of pre- and post-smoothing steps. For simplicity, we denote the overall number of those steps by $\nu=\nu_1+\nu_2$. We consider simple restriction operators using six points for the $u$ and $v$ components of the velocity, that is,
\begin{equation*}
R_{h,u} = \frac{1}{8}\begin{bmatrix}
1 & & 1 \\
2 & \star & 2 \\
1 & & 1
\end{bmatrix},
\quad
R_{h,v} = \frac{1}{8}\begin{bmatrix}
1 & 2 & 1 \\
& \star & \\
1 & 2 & 1
\end{bmatrix},
\end{equation*}
where the $\star$ denotes the position (on the coarse grid) at which the discrete operator is applied. For interpolation, we take $P_{h,u} =4 R^T_{h,u}$ and $P_{h,v} =4 R^T_{h,v}$.
For the restriction for the pressure, we use
\begin{equation*}
R_{h,p} = \frac{1}{4}\begin{bmatrix}
1 & & 1 \\
& \star & \\
1 & & 1
\end{bmatrix},
\end{equation*}
and $P_{h,p}=4R^T_{h,p}$. Consequently,
\begin{equation*}
R_h = \begin{pmatrix}
R_{h,u} & 0 & 0 \\
0 & R_{h,v} & 0 \\
0 & 0 & R_{h,p}
\end{pmatrix}, \quad P_h = 4 R^T_h.
\end{equation*}
\section{Local Fourier analysis}\label{sec:LFA}
To identify a proper parameter $\omega$ in \eqref{eq:BSR-relax-scheme} to construct fast multigrid methods, we apply LFA \cite{wienands2004practical,trottenberg2000multigrid} to examine the multigrid relaxation scheme.
The LFA smooting factor, see Definition \ref{def:LFA-mu}, often offers a sharp prediction of actual multigrid performance.
\begin{definition}
Let $L_h =[s_{\boldsymbol{\kappa}}]_{h}$ be a scalar stencil operator acting on grid ${G}_{h}$ as
\begin{equation*}
L_{h}w_{h}(\boldsymbol{x})=\sum_{\boldsymbol{\kappa}\in{V}}s_{\boldsymbol{\kappa}}w_{h}(\boldsymbol{x}+\boldsymbol{\kappa}h),
\end{equation*}
where $s_{\boldsymbol{\kappa}}\in \mathbb{R}$ is constant, $w_{h}(\boldsymbol{x}) \in l^{2} ({G}_{h})$, and ${V}$ is a finite index set.
Then, the symbol of $L_{h}$ is defined as:
\begin{equation}\label{eq:symbol-calculation-form}
\widetilde{L}_{h}(\boldsymbol{\theta})=\displaystyle\sum_{\boldsymbol{\kappa}\in{V}}s_{\boldsymbol{\kappa}}e^{i \boldsymbol{\theta}\cdot\boldsymbol{\kappa}},\qquad i^2=-1.
\end{equation}
\end{definition}
We consider standard coarsening. The low and high frequencies are given by $$\boldsymbol{\theta} \in T^{\rm{L}} =\left[-\frac{\pi}{2}, \frac{\pi}{2}\right)^d; \qquad \boldsymbol{\theta} \in T^{\rm{H}} =\left[-\frac{\pi}{2}, \frac{3\pi}{2}\right)^d \setminus T^{\rm{L}}.$$
\begin{definition}\label{def:LFA-mu}
We define the LFA smoothing factor for relaxation error operator $\mathcal{S}_h$ as
\begin{equation*}
\mu_{\rm loc}(\mathcal{S}_h) = \max_{\boldsymbol{\theta} \in T^{\rm{H}}}\{\rho(\widetilde{\mathcal{S}}_h(\boldsymbol{\theta}))\},
\end{equation*}
where $\rho(\widetilde{S}_h(\boldsymbol{\theta}))$ stands for the spectral radius of $\widetilde{\mathcal{S}}_h(\boldsymbol{\theta})$.
\end{definition}
The symbol of $\mathcal{S}_h$ defined in \eqref{eq:relaxation-error-operator} is a $3\times 3$ matrix since $\mathcal{K}_h$ is a $3\times 3$ block system; see \eqref{eq:Kh--stencil-operator}. The same holds for $\mathcal{M}_h$, see \eqref{eq:Mh-form}, and the symbol of each block is a scalar. For more details on how to compute the symbol of
$\mathcal{S}_h$, refer to other studies \cite{farrell2021local,he2018local}. Since $\mu_{\rm loc}(\mathcal{S}_h)$ is a function of the parameter $\omega$, we are interested in minimizing $\mu_{\rm loc}(\mathcal{S}_h)$ over $\omega$ to obtain a fast convergence speed. We define the optimal smoothing factor as
\begin{equation*}
\mu_{\rm opt} =\min_{\omega}\mu_{\rm loc}(\mathcal{S}_h).
\end{equation*}
For the two-grid error operator $E_h$ defined in \eqref{eq:two-grid-error-operator}, the two-grid LFA convergence factor is
\begin{equation}\label{eq:LFA-two-grid-convergence-factor}
\rho_h(\nu)= \max_{\boldsymbol{\theta}\in T^{\rm L}}\{\rho( \widetilde{\mathbf{E}}_h(\omega,\boldsymbol{\theta}))\},
\end{equation}
where $\widetilde{\mathbf{E}}_h$ is the two-grid error operator symbol and $\rho( \widetilde{\mathbf{E}}_h)$ stands for the spectral radius of matrix $\widetilde{\mathbf{E}}_h$. Since $E_h$ contains the coarse and fine grid operators, its symbol is a $12\times 12$ matrix, including four harmonic frequencies.
From this point onward, let us drop the subscript $h$, unless it is necessary.
The element-wise Vanka-type smoother has been successfully applied to complex-shifted Laplacian systems arising from optimal control problem \cite{HL2022shiftLaplacianPiTMG}. Here, we consider an element-wise additive Vanka smoother applied to \eqref{eq:shift-Laplace-stencil}. The subproblem coefficient matrix $L_j$ in \eqref{eq:Vanka-operator} has a symmetric structure
\begin{equation*}
L_j =\frac{\epsilon^2}{h^2}
\begin{pmatrix}
4+r & -1 & -1 &0 \\
-1 & 4+r& 0 &-1\\
-1 & 0 & 4+r &-1\\
0 & -1 & -1 &4+r
\end{pmatrix}.
\end{equation*}
It follows that
\begin{equation} \label{eq:inverse-Li-form}
L^{-1}_j =\frac{h^2}{\epsilon^2}
\begin{pmatrix}
a & b & b &c \\
b & a& c &b\\
b & c & a &b \\
c & b & b &a
\end{pmatrix},
\end{equation}
where
\begin{subequations}
\label{eq:Vanka}
\begin{eqnarray}
a&=& \frac{r^2+8r+14}{(2+r)(4+r)(6+r)}, \label{eq:Vanka-a}\\
b&=& \frac{1}{(2+r)(6+r)}, \label{eq:Vanka-b}\\
c&= & \frac{2}{(2+r)(4+r)(6+r)}.\label{eq:Vanka-c}
\end{eqnarray}
\end{subequations}
It is easy to show that $a>b>c$, which is useful for our analysis.
Based on \eqref{eq:Vanka-operator} and \eqref{eq:inverse-Li-form}, the stencil of the element-wise Vanka smoother $M_e$ is given by
\begin{equation*}
M_e = \frac{h^2}{4\epsilon^2}
\begin{bmatrix}
c & 2b &c\\
2b & 4a &2b\\
c & 2b &c
\end{bmatrix}.
\end{equation*}
Using \eqref{eq:symbol-calculation-form}, we have
\begin{eqnarray*}
\widetilde{L} &=&\frac{\epsilon^2}{h^2}(4+r-2\cos \theta_1 -2\cos \theta_2),\\
\widetilde{M}_e &=&\frac{h^2}{\epsilon^2} (a+b\cos \theta_1 +b\cos \theta_2 +c\cos\theta_1\cos \theta_2).
\end{eqnarray*}
Let $t= \epsilon^2 (4+r-2\cos \theta_1 -2\cos \theta_2)$ and $\hat{t}=\frac{\epsilon^2}{ a+b\cos \theta_1 +b\cos \theta_2 +c\cos\theta_1\cos \theta_2)}$. Then,
\begin{equation*}
\widetilde{\mathcal{K}}= \frac{1}{h^2}\begin{pmatrix}
t & 0 & i 2h \sin(\theta_1/2)\\
0 & t & i 2h \sin(\theta_2/2)\\
-i 2h \sin(\theta_1/2) & -i 2h \sin(\theta_2/2) &0
\end{pmatrix},
\end{equation*}
and
\begin{equation*}
\widetilde{\mathcal{M}}= \frac{1}{h^2} \begin{pmatrix}
\hat{t} & 0 &i 2h \sin(\theta_1/2)\\
0 & \hat{t} & i 2h \sin(\theta_2/2)\\
-i 2h \sin(\theta_1/2) & -i 2h \sin(\theta_2/2) &0
\end{pmatrix},
\end{equation*}
To identify the eigenvalues of $\widetilde{\mathcal{M}}^{-1} \widetilde{\mathcal{K}}$, we first compute the determinant of $\widetilde{\mathcal{K}}-\lambda \widetilde{\mathcal{M}}$:
\begin{eqnarray*}
| \widetilde{\mathcal{K}}-\lambda\widetilde{\mathcal{M}}| &= &
\frac{1}{h^2}\begin{vmatrix}
t-\lambda \hat{t} & 0 & (1-\lambda) i 2h \sin(\theta_1/2) \\
0 & t-\lambda \hat{t} & (1-\lambda) i 2h \sin(\theta_2/2) \\
-(1-\lambda) i 2h \sin(\theta_1/2) &(1-\lambda) i 2h \sin(\theta_2/2) &0
\end{vmatrix} \\
&=& \frac{1}{h^2}(t-\lambda \hat{t})(1-\lambda)^2\left((i 2h \sin(\theta_1/2))^2 + (i 2h \sin(\theta_2/2) )^2\right)\\
&=& 4\hat{t}\left((\sin(\theta_1/2))^2 + ( \sin(\theta_2/2) )^2\right)(1-\lambda)^2 (\lambda -\frac{t}{\hat{t}}).
\end{eqnarray*}
The three eigenvalues of $\widetilde{\mathcal{M}}^{-1} \widetilde{\mathcal{K}}$ are $1, 1$ and $\frac{t}{\hat{t}}=:\lambda^*$, where
\begin{equation}\label{eq:lambda-form-r}
\lambda^*(r;\cos \theta_1,\cos \theta_2) =(a+b\cos \theta_1 +b\cos \theta_2 +c\cos\theta_1\cos \theta_2)(4+r-2\cos \theta_1 -2\cos \theta_2).
\end{equation}
For $\boldsymbol{\theta} \in T^{\rm H}$, it is easy to show that
\begin{equation}\label{eq:cos-range}
(\cos\theta_1,\cos\theta_2) \in \mathcal{D}=[-1,1] \times [-1,0] \bigcup [-1,0]\times [0, 1].
\end{equation}
Next, we explore the range of $\lambda^*$ over $\boldsymbol{\theta}$ for high frequencies.
\begin{theorem}\label{thm:maxmin-and-theta}
For $\boldsymbol{\theta} \in T^{\rm H}$,
\begin{eqnarray*}
\max_{\boldsymbol{\theta} }\lambda^*(r;\cos \theta_1,\cos \theta_2) &=& \lambda^*(r;-1,-1) =(a-2b+c)(8+r)=:d_1(r),\\
\min_{\boldsymbol{\theta} }\lambda^*(r;\cos \theta_1,\cos \theta_2) &=& \lambda^*(r;1,0) =(a+b)(2+r)=:d_2(r).
\end{eqnarray*}
\end{theorem}
\begin{proof}
For simplicity, let $\eta_1=\cos\theta_1$ and $\eta_2=\cos \theta_2$. Then, we rewrite \eqref{eq:lambda-form-r} as
\begin{equation*}
\lambda^*=\psi(\eta_1,\eta_2) =(a+b\eta_1+b\eta_2 +c\eta_1\eta_2)(4+r-2\eta_1 -2\eta_2).
\end{equation*}
We first consider the critical point of $ \psi(\eta_1,\eta_2) $ in $\mathcal{D}$, see \eqref{eq:cos-range}, by computing the partial derivatives of $ \psi(\eta_1,\eta_2)$, which are given by
\begin{eqnarray}
\psi'_{\eta_1} (\eta_1,\eta_2)&=&rb+4b-2a-4b\eta_1+(4c+cr-4b)\eta_2-2c\eta_2^2-4c\eta_1\eta_2=0, \label{eq:partial-psi-eta1} \\
\psi'_{\eta_2}(\eta_1,\eta_2) &=& rb+4b-2a-4b\eta_2+(4c+cr-4b)\eta_1-2c\eta_1^2-4c\eta_1\eta_2=0. \label{eq:partial-psi-eta2}
\end{eqnarray}
Subtracting \eqref{eq:partial-psi-eta2} from \eqref{eq:partial-psi-eta1} gives
\begin{equation}\label{eq:critical-point-solu}
(\eta_1-\eta_2)\left(2(\eta_1+\eta_2)-4-r\right) =0.
\end{equation}
It follows that $\eta_1=\eta_2$ or $2(\eta_1+\eta_2)-4-r=0$. However, $\eta_1+\eta_2< 2$, so the latter does not have a real solution. For $\eta_1=\eta_2$, we
replace $\eta_2$ by $\eta_1$ in \eqref{eq:partial-psi-eta1}, leading to
\begin{equation}\label{eq:solve-eta1=0}
6c\eta_1^2-(4c+cr-8b)\eta_1-(rb+4b-2a)=0.
\end{equation}
We claim that there is no real solution for \eqref{eq:solve-eta1=0} for $r>0$. We will show that the discriminant is not positive.
We first simplify $rb+4b-2a$. Using \eqref{eq:Vanka-a} and \eqref{eq:Vanka-b} gives
\begin{eqnarray*}
rb+4b-2a &=&\frac{ 4+r}{(2+r)(6+r)}- \frac{2(r^2+8r+14)}{(2+r)(4+r)(6+r)}\\
&=&\frac{ (4+r)^2-2(r^2+8r+14)}{(2+r) (4+r)(6+r)}\\
&=& -\frac{1}{4+r}.
\end{eqnarray*}
Using \eqref{eq:Vanka-b} and \eqref{eq:Vanka-c}, the discriminant of \eqref{eq:solve-eta1=0} is
\begin{eqnarray*}
\Phi &=& (4c+cr-8b)^2+4 \cdot 6c(rb+4b-2a)\\
&=&\left(\frac{8+2r}{(2+r)(4+r)(6+r)} - \frac{8(4+r)}{(2+r)(4+r)(6+r)} \right)^2-\frac{48}{(2+r)(4+r)(6+r)} \frac{1}{4+r}\\
&=&\left (\frac{-6}{(2+r)(6+r)}\right)^2-\frac{48}{(2+r)(4+r)^2(6+r)}\\
&=& \frac{12}{(2+r)(6+r)}\left(\frac{3}{(2+r)(6+r)}-\frac{4}{(4+r)^2}\right)\\
&=& \frac{-12r(r+8)}{(2+r)^2(4+r)^2(6+r)^2}\leq 0.
\end{eqnarray*}
The case $r=0$ has been discussed \cite{CH2021addVanka}, where $\psi'_{\eta_1} (\eta_1,\eta_2)=\psi'_{\eta_2} (\eta_1,\eta_2)=0$ gives $(\eta_1,\eta_2)=(-1,-1)$ the boundary point of $\mathcal{D}$, and $\lambda^*_{\rm max}=\frac{4}{3}$. When $r>0$, \eqref{eq:critical-point-solu} has no real solution and $\psi(\eta_1,\eta_2)$ cannot have extreme values at interior of $\mathcal{D}$. This means that we only need to find the extreme values of $\psi(\eta_1,\eta_2)$ at the boundary of $\mathcal{D}$, see \eqref{eq:cos-range}. To do this, we split the boundary of $\mathcal{D}$ as follows:
\begin{eqnarray*}
\partial \mathcal{D}_1 &=&\{-1\} \times [-1,1], \\
\partial \mathcal{D}_2 &=& [-1,1] \times \{-1\}, \\
\partial \mathcal{D}_3 &=& \{1\} \times [-1,0], \\
\partial \mathcal{D}_4 &=& [0,1]\times \{0\}, \\
\partial \mathcal{D}_5 &=& \{0\} \times [0,1], \\
\partial \mathcal{D}_6&=& [-1,0] \times \{1\}.
\end{eqnarray*}
Due to the symmetry of $\psi(\eta_1,\eta_2)$, that is $\psi(\eta_1,\eta_2)=\psi(\eta_2,\eta_1)$, we only need to find the extreme values of $ \psi(\eta_1,\eta_2) $ at $\partial \mathcal{D}_1, \partial \mathcal{D}_3$ and $\partial \mathcal{D}_4$. We present below the results.
\begin{enumerate}
\item For $(\eta_1,\eta_2) \in \partial \mathcal{D}_1$,
\begin{equation}\label{eq:psi-case1}
\psi(\eta_1,\eta_2) =\psi(-1,\eta_2) =(a-b+b\eta_2-c\eta_2)(6+r-2\eta_2).
\end{equation}
Note that the two roots of the quadratic form \eqref{eq:psi-case1} are $\frac{6+r}{2}$ and $\frac{a-b}{c-b}$.
Using \eqref{eq:Vanka}, we have
\begin{equation*}
\frac{a-b}{c-b}= -(5+r).
\end{equation*}
Thus, the axis of symmetry is $\eta_2=\frac{(6+r)/2-(5+r)}{2}=-1-\frac{r}{4}\leq -1$. Using the fact that $a>b>c$, see \eqref{eq:Vanka-a}, \eqref{eq:Vanka-b}, and \eqref{eq:Vanka-c}, the quadratic function opens downward. Therefore,
the maximum and minimum of $\psi(-1,\eta_2) $ for $\eta_2 \in[-1,1]$ are
\begin{align}
\begin{aligned}
\psi(-1,\eta_2) _{\rm max} &=\psi(-1,-1) =(a-2b+c)(8+r),\label{eq:case1-max}\\
\psi(-1,\eta_2) _{\rm min} &=\psi(-1,1) = (a-c)(4+r).
\end{aligned}
\end{align}
\item For $(\eta_1,\eta_2) \in \partial \mathcal{D}_3$,
\begin{equation}\label{eq:psi-case2}
\psi(\eta_1,\eta_2) =\psi(1,\eta_2) =(a+b+b\eta_2+c\eta_2)(2+r-2\eta_2).
\end{equation}
The two roots of quadratic form \eqref{eq:psi-case2} are $\frac{2+r}{2}$ and $-\frac{a+b}{b+c}$.
Using \eqref{eq:Vanka}, we have
\begin{equation*}
-\frac{a+b}{b+c}= -(3+r).
\end{equation*}
Thus, the axis of symmetry is $\eta_2=\frac{(2+r)/2-(3+r)}{2}=-1-\frac{r}{4}\leq -1$. Using the fact that $a>b>c$, the quadratic function opens downward. It follows that
for $\eta_2 \in[-1,0]$, the maximum and minimum of $\psi(1,\eta_2) $ are given by
\begin{eqnarray*}
\psi(1,\eta_2) _{\rm max} &=&\psi(1,-1) =(a-c)(4+r),\\
\psi(1,\eta_2) _{\rm min} &=&\psi(1,0) = (a+b)(2+r).
\end{eqnarray*}
\item For $(\eta_1,\eta_2) \in \partial \mathcal{D}_4$,
\begin{equation*}
\psi(\eta_1,\eta_2) =\psi(\eta_1,0) =(a+b\eta_1)(4+r-2\eta_1).
\end{equation*}
Note that the two roots of quadratic form \eqref{eq:psi-case2} are $\frac{4+r}{2}$ and $-\frac{a}{b}$.
Using \eqref{eq:Vanka-a} and \eqref{eq:Vanka-b}, we have
\begin{equation*}
-\frac{a}{b}= -(4+r)+\frac{2}{4+r}.
\end{equation*}
Thus, the axis of symmetry is $\eta_1=\frac{(4+r)/2-(4+r)+\frac{2}{4+r}}{2}=-1-\frac{r}{4}+\frac{1}{4+r}< 0$. Thus, the maximum and minimum of $\psi(\eta_1,0) $ for $\eta_1 \in[0,1]$ are
\begin{eqnarray}
\psi(\eta_1,0) _{\rm max} &=&\psi(0,0) =a(4+r),\label{eq:case3-max}\\
\psi(\eta_1,0) _{\rm min} &=&\psi(0,1) = (a+b)(2+r). \nonumber
\end{eqnarray}
\end{enumerate}
Based on the above discussions, the minimum of $\psi(\eta_1,\eta_2) $ over $\partial \mathcal{D}$ is
\begin{equation*}
\psi(\eta_1,\eta_2) _{\rm min} =\psi(1,0)=\psi(0,1) = (a+b)(2+r).
\end{equation*}
Next, we compare $\psi(-1,-1)$ (see \eqref{eq:case1-max}) and $\psi(0,0)$ (see \eqref{eq:case3-max}) to determine the maximum.
Using \eqref{eq:Vanka}, we have
\begin{eqnarray*}
\psi(-1,-1)- \psi(0,0) &=&(a-2b+c)(8+r)-a(4+r)\\
&=& 4a+(c-2b)(8+r)\\
& =&\frac{4(r^2+8r+14)}{(2+r)(4+r)(6+r)} - \frac{2-2(4+r)}{(2+r)(4+r)(6+r)}(8+r)\\
&=& \frac{2r^2+10r+8}{(2+r)(4+r)(6+r)}>0.
\end{eqnarray*}
It follows that the maximum of $\psi(\eta_1,\eta_2)$ is given by
\begin{equation*}
\psi(\eta_1,\eta_2) _{\rm max} =\psi(-1,-1) =(a-2b+c)(8+r).
\end{equation*}
Thus, for $(\eta_1,\eta_2) \in \mathcal{D}$, the maximum and minimum of $\psi(\eta_1,\eta_2)$ are $\psi(-1,-1)$ and $\psi(1,0)=\psi(0,1)$, respectively.
\end{proof}
Based on the results in Theorem \ref{thm:maxmin-and-theta}, we can further estimate the range of extreme values of $\lambda^*$, which plays an important role in determining the optimal smoothing factor for V-BSR.
\begin{theorem}\label{thm:form-d1-d2}
Suppose $r\in[0,\infty)$. Then,
\begin{equation}\label{eq:explicit-form-low-up-bound}
d_1(r)=\frac{8+r}{6+r}, \quad d_2(r)=\frac{3+r}{4+r}.
\end{equation}
Furthermore,
\begin{eqnarray*}
1< &d_1(r)& \leq \frac{4}{3}, \\
\frac{3}{4}\leq &d_2(r)& <1.
\end{eqnarray*}
\end{theorem}
\begin{proof}
Using \eqref{eq:Vanka}, we simplify $d_1(r)$ as follows:
\begin{eqnarray*}
d_1(r) &=&(a-2b+c)(8+r) \\
&=& \frac{r^2+8r+14-2(4+r)+2}{(2+r)(4+r)(6+r)}(8+r)\\
&=& \frac{8+r}{6+r}.
\end{eqnarray*}
Since $d_1(r)$ is a decreasing function of $r$, $\max_r d_1(r)=d_1(0)=\frac{4}{3}$.
Using \eqref{eq:Vanka-a} and \eqref{eq:Vanka-b}, we have
\begin{eqnarray*}
d_2(r) &=&(a+b)(2+r) \\
&=& \frac{r^2+8r+14+4+r}{(2+r)(4+r)(6+r)} (2+r)\\
&=&\frac{r+3}{r+4}.
\end{eqnarray*}
Since $d_2(r)$ is an increasing function of $r$, $\min_r d_2(r)=d_2(0)=\frac{3}{4}$.
\end{proof}
Now, we are able to derive the optimal smoothing factor for V-BSR for the Stokes-Darcy Brinkman problems.
\begin{theorem}\label{thm:opt-mu-omega}
For the V-BSR relaxation scheme \eqref{eq:BSR-relax-scheme} the optimal smoothing factor is given by
\begin{equation}\label{eq:opt-mu-form-r}
\mu_{\rm opt}(r)=\min_{\omega} \max_{\boldsymbol{\theta} \in T^{\rm H}} \{|1-\omega|, |1-\omega \lambda^*|\} = \frac{3r+14}{2r^2+21r+50},
\end{equation}
provided that
\begin{equation}\label{eq:opt-omega}
\omega =\omega_{\rm opt}= \frac{2r^2+20r+48}{2r^2+21r+50}.
\end{equation}
Moreover,
\begin{equation*}
\mu_{\rm opt}(r)\leq \frac{7}{25}=0.28.
\end{equation*}
\end{theorem}
\begin{proof}
From Theorem \ref{thm:maxmin-and-theta} and \eqref{eq:explicit-form-low-up-bound}, we know that
\begin{equation}\label{eq:lambda*-two-roots}
\max_{\boldsymbol{\theta} \in T^{\rm H}} \{|1-\omega|, |1-\omega \lambda^*|\} =\max\{|1-\omega d_2(r)|, |1-\omega d_1(r)| \}.
\end{equation}
To minimize $\max_{\boldsymbol{\theta} \in T^{\rm H}} \{|1-\omega \lambda^*|\}$, we require
\begin{equation*}
|1-\omega d_2(r)|= |1-\omega d_1(r)|,
\end{equation*}
which gives $\omega_{\rm opt}(r) =\frac{2}{d_1(r)+d_2(r)}$. Using Theorem \ref{thm:form-d1-d2}, we obtain
\begin{equation*}
\omega_{\rm opt}(r) =\frac{2}{d_1(r)+d_2(r)} = \frac{2r^2+20r+48}{2r^2+21r+50}
\end{equation*}
and
\begin{equation*}
\mu_{\rm opt}(r) =\frac{d_1(r)-d_2(r)}{d_1(r)+d_2(r)} = \frac{3r+14}{2r^2+21r+50}\leq \frac{7}{25}.
\end{equation*}
\end{proof}
\begin{remark}
When $r=0$, Theorem \ref{thm:opt-mu-omega} is consistent with the existing results \cite{CH2021addVanka}, which amount to applying an element-wise Vanka smoother to the Poisson equation.
\end{remark}
\begin{proposition}\label{prop:optimal-mu-decrease}
$\mu_{\rm opt}(r)$ given by \eqref{eq:opt-mu-form-r} is a decreasing function of $r$.
\end{proposition}
\begin{proof}
The derivative of $\mu_{\rm opt}$ is given by
\begin{equation*}
\mu'_{\rm opt} = \frac{-2(r^2+28r+72)}{(2r^2+21r+50)^2}<0.
\end{equation*}
This suggests that when $r$ increases, the optimal smoothing factor decreases.
\end{proof}
Let us look at the optimal parameter, \eqref{eq:opt-omega}. It can be shown that
\begin{equation*}
\omega'_{\rm opt}(r) = \frac{2(r^2+4r-4)}{(2r^2+21r+50)^2}.
\end{equation*}
When $r \in[0, 2\sqrt{2}-2]$, $\omega_{\rm opt}(r)$ is decreasing and when $r \in[2\sqrt{2}-2, \infty)$, $\omega_{\rm opt}(r)$ is increasing. It follows that
\begin{equation}\label{eq:estimate-opt-omega}
(\omega_{\rm opt}(r))_{\rm min}=\omega_{\rm opt}(2\sqrt{2}-2) = \frac{(\sqrt{2}-1)(4\sqrt{2}+16)+24}{(\sqrt{2}-1)(4\sqrt{2}+17)+25}\approx 0.959 \leq \omega_{\rm opt}(r) <1.
\end{equation}
%
Thus, for simplicity, if we take $\omega=1$, then \eqref{eq:lambda*-two-roots} gives
\begin{equation}\label{eq:omega-one-mu}
\mu(\omega=1)=\max_{\omega=1}\{|1-\omega d_2(r)|, |1-\omega d_1(r)| \}=\frac{2}{6+r}\leq \frac{1}{3}.
\end{equation}
In practice, we can consider $\omega=1$. In multigrid, for fixed $\epsilon$, in each level, the relaxation schemes has a different convergence speed in each level, which can be computed from \eqref{eq:omega-one-mu}. However, note that \eqref{eq:omega-one-mu} is a decreasing function of $r=\frac{h^2}{\epsilon^2}$ or $h$. This means that at the coarse level, the relaxation scheme has a smaller convergence speed compared with that of the fine level.
\section{Numerical experiments}\label{sec:numerical}
In this section, we first compute LFA two-grid convergence factors using two choices of the damping parameter, that is, $\omega=1$ and $\omega=\omega_{\rm opt}$, then we report V-cycle multigrid results for different values of the physical parameter $\epsilon$.
\subsection{LFA prediction}
We compute the two-grid LFA convergence factor \eqref{eq:LFA-two-grid-convergence-factor}, using $h=1/64,\ \omega=1$ and optimal $\omega$, see \eqref{eq:opt-omega}, derived from optimizing LFA smoothing factors for different $\epsilon$. From Table \ref{tab:LFA-results-h64-omega} we see a strong agreement between two-grid convergence factors $\rho_h(1)$ and the LFA smoothing factors. Moreover, the convergence factors for optimal $\omega$ are slightly better than those for $\omega=1$, which is reasonable since the optimal $\omega$, see \eqref{eq:estimate-opt-omega}, is very close to $1$. From our smoothing analysis, we know that even though the smoothing factor is dependent on $h$ and $\alpha$, the upper bound on the smoothing factor is $\frac{1}{3}$. This is also confirmed by our two-grid LFA convergence factor $\rho_h(1)$ in Table \ref{tab:LFA-results-h64-omega}.
\begin{table}[H]
\caption{Two-grid LFA convergence factors, $\rho_h(\nu)$ with $h=1/64$ and different choices of $\omega$.}
\centering
\begin{tabular}{lccccc}
\hline
$\epsilon,\omega=1$ & $\mu_{\rm opt}$ &$\rho_h(1)$ &$\rho_h(2)$ &$\rho_h(3)$ & $\rho_h(4)$ \\ \hline
$1$ & 0.333 & 0.333 & 0.119 & 0.054 &0.043 \\
$ 2^{-2} $ &0.333 & 0.333 & 0.119 & 0.054 & 0.042 \\
$2^{-4}$ &0.330 & 0.330 &0.115 &0.052 &0.040 \\
$2^{-6}$ &0.286 & 0.286 & 0.082 & 0.023 & 0.012 \\
$2^{-8}$ & 0.091 & 0.091 & 0.008 & 0.001 &0.000 \\
\hline
$1, \omega_{\rm opt}$ &0.280 &0.280 & 0.096 & 0.056 & 0.044 \\
$2^{-2}$ & 0.280 &0.280 &0.096 & 0.056 &0.044 \\
$2^{-4}$ & 0.276 & 0.276 & 0.093 & 0.055 &0.042 \\
$2^{-6}$ & 0.233 &0.233 & 0.057 & 0.026 & 0.014 \\
$2^{-8}$ & 0.069 & 0.069 & 0.005 & 0.000 & 0.000 \\
\hline
\end{tabular}\label{tab:LFA-results-h64-omega}
\end{table}
To illustrate how the smoothing factor changes as a function of $r$, we plot $\mu_{\rm opt}$ defined in \eqref{eq:opt-mu-form-r} and $\mu(\omega=1)$ in \eqref{eq:omega-one-mu} as functions of $r$ in Figure \ref{fig: mu-vs-r}. It is evident that when $r$ increases, the smoothing factor decreases and approaches zero, and $\mu(\omega=1)$ tends towards $\mu_{\rm opt}$.
\begin{figure}[H]
\centering
\includegraphics[width=0.6\textwidth]{./figures/plot-mu-vs-r.pdf}
\caption{Smoothing factors with optimal $\omega$ and $\omega=1$.}\label{fig: mu-vs-r}
\end{figure}
\subsection{Multigrid performance}
Consider the model problems \eqref{eq:SDB} on the unit square domain $[0,1]\times[0,1]$ with an exact solution \cite[Section 5]{sun2019stability},
and given by
\begin{align*}
u(x,y) &=\pi \sin^2(\pi x) \sin(2\pi y),\\
v(x,y) &=-\pi \sin(2\pi x) \sin^2(\pi y),\\
p(x,y) &=\sin(\pi y) -\frac{2}{\pi},
\end{align*}
with $g=0$. The source term is computed via $\boldsymbol{f}=(f_1,f_2)=- \epsilon^2 \Delta\boldsymbol{u} +\boldsymbol{u} + \nabla p$, and it is
\begin{align*}
f_1 &=(4\pi^3\epsilon^2+\pi)\sin^2(\pi x) \sin(2\pi y)-2\pi^3\epsilon^2\cos(2\pi x)\sin(2\pi y), \\
f_2& = -(4\pi^3\epsilon^3+\pi)\sin(2\pi x)\sin^2(\pi y)+2\pi^3\epsilon^2\sin(2\pi x)\cos(2\pi y)+\pi \cos(\pi y).
\end{align*}
To validate our theoretical LFA predictions, we compute the actual multigrid convergence factors by
\begin{equation*}
\hat{\rho}^{(k)}_h=\left( \frac{||r_k||}{||r_0||}\right)^{1/k},
\end{equation*}
where $r_k=b_h-\mathcal{K}_h\boldsymbol{z}_k$ is the residual and $\boldsymbol{z}_k$ is the $k$-th multigrid iteration. The initial guess is chosen randomly. In our test, we report $\hat{\rho}^{(k)}_h=:\hat{\rho}_h $ with the smallest $k$ such that $||r_k||/|r_0|\leq 10^{-10}$.
As mentioned before, computing the exact solution of the Schur complement system \eqref{eq:solution-schur-complement} is expensive. For our multigrid tests, we apply a few weighted ($\omega_J$) Jacobi iterations to the Schur complement system. We choose $\omega_J=0.8$ which seems more robust to $\epsilon$. The number of Jacobi iterations is set as three.
\subsubsection{Two-grid results}
We first report actual two-grid convergence factor using $h=1/64$ and three Jacobi iterations for solving the Schur complement. Table \ref{tab:TG-omega1} shows that two-grid actual performance using $\omega=1$ matches with the LFA predictions in Table \ref{tab:LFA-results-h64-omega}, except for a small difference for a $\epsilon =2^{-8}$, which might suggest that more iterations are needed for the Schur complement system. However, due to the satisfactory convergence factor of the actual performance, we do not further explore this. Using optimal $\omega$, Table \ref{tab:TG-omega2} shows that the actual two-grid performance matches two-grid LFA predictions reported in Table \ref{tab:LFA-results-h64-omega}, except for $\epsilon=2^{-6}, 2^{-8}$. Again, the measured convergence factor is satisfactory, and there is no need to consider more Jacobi iterations for the Schur complement system.
\begin{table}[H]
\caption{Two-grid measured convergence factor, $\hat{\rho}_h(\nu)$, using three Jacobi iterations for solving the Schur complement, $h=1/64$ and $\omega=1$.}
\centering
\begin{tabular}{lcccc }
\hline
$ \epsilon, \omega=1 $ &$\hat{\rho}_h(1)$ &$\hat{\rho}_h(2)$ &$\hat{\rho}_h(3)$ & $\hat{\rho}_h(4)$ \\ \hline
$1$ & 0.319 &0.111 &0.033 & 0.023 \\
$2^{-2} $ &0.317 & 0.109 &0.033 & 0.023 \\
$2^{-4} $ & 0.300 & 0.094 & 0.029 & 0.021 \\
$2^{-6}$ & 0.209 &0.047 &0.023 & 0.015 \\
$2^{-8} $ &0.145 & 0.035 &0.020 & 0.015 \\
\hline
\end{tabular}\label{tab:TG-omega1}
\end{table}
\begin{table}[H]
\caption{Two-grid measured convergence facto, $\hat{\rho}_h(\nu)$, using three Jacobi iterations for solving the Schur complement, $h=1/64$ and $\omega_{\rm opt}$, see \eqref{eq:opt-omega}.}
\centering
\begin{tabular}{lcccc }
\hline
$\epsilon, \omega_{\rm opt}$ &$\hat{\rho}_h(1)$ &$\hat{\rho}_h(2)$ &$\hat{\rho}_h(3)$ & $\hat{\rho}_h(4)$ \\ \hline
$1$ & 0.266 & 0.082 &0.030 & 0.023 \\
$2^{-2} $ & 0.264 &0.080 & 0.030 & 0.024 \\
$2^{-4}$ & 0.248 &0.068 &0.029 & 0.023 \\
$2^{-6}$ &0.163 & 0.044 & 0.024 & 0.016 \\
$2^{-8}$ & 0.165 &0.040 & 0.019 &0.015 \\
\hline
\end{tabular}\label{tab:TG-omega2}
\end{table}
\subsubsection{V(1,1)-cycle results}
A two-grid method is computationally costly since we have to solve the coarse problem directly and if the initial mesh is fine, then the next coarser mesh may give rise to a large problem as well. In practice, deeply-nested W-cycle and V-cycle are preferred. We now explore the V(1,1)-cycle multigrid methods with two choices of $\omega$ and varying values of the physical parameter $\epsilon$. In order to study the sensitivity of solving the Schur complement system, we consider one, two, and three Jacobi iterations for Schur complement system. We consider different $n\times n$ finest meshgrids, where $n=32, 64,128, 256$.
{\bf One iteration for Schur complement system:} We first report the iteration counts for V(1,1)-cycle multigrid methods using one iteration of Jacobi relaxation for the Schur complement system in Table \ref{tab:Itn-Schur-Jacobi-number-one} to achieve the tolerance $||r_k||/|r_0|\leq 10^{-10}$. We see that using $\omega=1$ and optimal $\omega$ give similar performance. When $\epsilon =2^{-6}, 2^{-8}$, the iteration count increase dramatically.
To mitigate the effect of this degradation, we will consider two or three Jacobi iterations for the Schur complement system.
\begin{table}[H]
\caption{Iteration accounts for V(1,1)-cycle multigrid with one Jacobi iteration for solving the Schur complement.}
\centering
\begin{tabular}{lcccc }
\hline
$\epsilon, \omega=1$ &$n=32$ &$n=64$ &$n=128$ & $n=256$ \\
$1$ & 13 & 13 &13 & 15 \\
$2^{-2} $ &12 & 13 & 13 & 14 \\
$2^{-4}$ &11 & 11 & 12 & 12 \\
$2^{-6}$ &23 & 18 &13 & 11 \\
$2^{-8}$ & 50 & 50 &47 & 34 \\
\hline
$\epsilon, \omega_{\rm opt}$ &$n=32$ &$n=64$ &$n=128$ & $n=256$ \\
$1$ & 12 &12 & 12 & 15 \\
$2^{-2} $ & 12 & 12 & 12 & 14 \\
$2^{-4}$ & 11 & 11 &11 & 11 \\
$2^{-6}$ & 26 &19 & 14 & 12 \\
$2^{-8}$ &50 &50 &50 & 38 \\
\hline
\end{tabular}\label{tab:Itn-Schur-Jacobi-number-one}
\end{table}
{\bf Two iterations for Schur complement system:} We report the convergence history of the relative residual norm $\frac{||r_k||}{||r_0||}$ as a function of the number of V(1,1)-cycles using two Jacobi iterations for Schur complement system. Figure \ref{fig:V-vs-J2-eps0} reports the results for $\epsilon=1$. We see that using optimal $\omega$ takes 12 V(1,1)-cycle iterations to achieve the stopping tolerance and it takes 13 iterations for $\omega=1$. The convergence behavior is independent of meshsize $h$. A similar performance is seen for $\epsilon=2^{-2}, 2^{-4}, 2^{-6}, 2^{-8}$ in Figures \ref{fig:V-vs-J2-eps2}, \ref{fig:V-vs-J2-eps4}, \ref{fig:V-vs-J2-eps6} and \ref{fig:V-vs-J2-eps8}. Observe that for smaller values of $\epsilon$, the iteration count does not increase. Using optimal $\omega$ has one iteration number fewer than that of $\omega=1$. Thus, it is simple and reasonable to use $\omega=1$ in practice.
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J2-eps0.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J2-eps0.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=1$ and two Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-J2-eps0}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J2-eps2.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J2-eps2.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-2}$ and two Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-J2-eps2}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J2-eps4.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J2-eps4.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-4}$ and two Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-J2-eps4}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J2-eps6.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J2-eps6.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-6}$ and two Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-J2-eps6}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J2-eps8.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J2-eps8.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-8}$ and two Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-J2-eps8}
\end{figure}
{\bf Three iterations for Schur complement system:} We explore V(1,1)-cycle iterations with two choices of $\omega$ and a varying physical parameter $\epsilon$, using three Jacobi iterations for Schur complement system. We report the history of relative residual $\frac{||r_k||}{||r_0||}$ as a function of the V(1,1)-cycle iteration counts for $n\times n$ meshgrid ($n=32, 64,128, 256$). Figure \ref{fig:V-vs-eps0} shows the results for $\epsilon=1$. We see that using optimal $\omega$ takes 12 iterations of V(1,1)-cycle to achieve the stopping tolerance and it takes 13 iterations for $\omega=1$. We see that the convergence behavior is independent of meshsize $h$. A similar performance is seen for $\epsilon=2^{-2}, 2^{-4}, 2^{-6}, 2^{-8}$ in Figures \ref{fig:V-vs-eps2}, \ref{fig:V-vs-eps4}, \ref{fig:V-vs-eps6} and \ref{fig:V-vs-eps8}. Compared with two Jacobi iterations for solving the Schur complement system, three Jacobi iterations give a slightly better results for small $\epsilon=2^{-6}, 2^{-8}$. Again using optimal $\omega$ has one iteration number less than that of $\omega=1$. Thus, it is simple and reasonable to use $\omega=1$ in practice. Moreover, two Jacobi iterations are enough to achieve robustness V(1,1)-cycle multgrid with respect to meshgrid and physical parameter $\epsilon$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J3-eps0.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J3-eps0.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=1$ and three Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-eps0}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J3-eps2.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J3-eps2.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-2}$ and three Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-eps2}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J3-eps4.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J3-eps4.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-4}$ and three Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-eps4}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J3-eps6.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J3-eps6.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-6}$ and three Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-eps6}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.49\textwidth]{./figures/V-omega1J3-eps8.pdf}
\includegraphics[width=0.49\textwidth]{./figures/V-omega2J3-eps8.pdf}
\caption{Convergence history: Number of iterations versus relative residual of V(1,1)-cycle with $\epsilon=2^{-8}$ and three Jacobi iterations for Schur complement system (left $\omega=1$ and right optimal $\omega$).} \label{fig:V-vs-eps8}
\end{figure}
\section{Conclusions}\label{sec:con}
We propose a parameter-robust multigrid method for solving the discrete system of Stokes-Darcy problems, with the maker and cell scheme used for the discretization. The resulting linear system is a saddle-point system. In contrast to existing Vanka smoothers, where the velocities and pressure unknowns in a grid cell are updated simultaneously, we propose Vanka-based Braess-Sarazin relaxation scheme, where the Laplace-like term in the saddle-point system is solved by an additive Vanka algorithm. This approach decouples the velocities and pressure unknowns. Moreover, only matrix-vector products are needed in our proposed multigrid method. LFA is used to analyze the smoothing process and help choose the optimal parameter that minimizing LFA smoothing factor. From LFA, we derive the stencil of additive Vanka for the Laplace-like operator, which can help form the global iteration matrix, avoiding solving many subproblems in the classical additive Vanka setting.
Our main contribution is that we derive the optimal algorithmic parameter and optimal LFA smoothing factor for Vanka-based Braess-Sarazin relaxation scheme, and show that this scheme is highly efficient with respect to physical parameter. Our theoretical results reveal that although the optimal damping parameter is related to physical parameter and meshsize, it is very close to one. We also present the theoretical LFA smoothing factor with damping parameter one. In Vanka-based Braess-Sarazin relaxation, we have to solve a Schur complement system. Direct solver is often expensive. We propose an inexact version of Vanka-based Braess-Sarazin relaxation, where we apply only two or three iterations of Jacobi to the Schur complement system to achieve the same performance as that of an exact solve. We show that using damping parameter one can achieve almost the same performance as that of optimal result, and the results are close to exact version. Thus, using damping parameter one is recommended. Our V-cycle multigrid illustrates high efficiency of our relaxation scheme and robustness to physical parameter.
We comment that the proposed Vanka-based Braess-Sarazin multigrid method can be used
as a preconditioner for Krylov subspace methods. We have limited ourselves for the MAC scheme on uniform grids. However, it is possible to extend the Vanka-smoother to non-uniform grids.
\bibliographystyle{siam}
|
\section{Introduction}
A remarkable development of accelerator science from the beginning of the 20th century to the present
has enabled the use of various quantum beams, such as electron, proton, and ion beams, and has opened the way to greater human understanding,
including the discovery of the Higgs boson and quantum beam therapy etc.
The acceleration of muons using a radio-frequency accelerator was recently demonstrated for the first time~\cite{bib:Bae18},
and has opened a new era of accelerator science using accelerated muon beams.
\par
A muon is an elementary particle similar to an electron, with an electric charge of $-e$ and a spin of $\frac{1}{2}$,
but with a mass 200-times heavier.
After the successful generation of muons using a proton driver half a century after their discovery during the study of cosmic-rays~\cite{bib:And37, bib:Mic75},
muons are now widely used in various scientific fields.
In recent years, the demand for muon acceleration has been increasing in many different areas.
For example, a muon collider, where muons are accelerated to a high energy for colliding, is one of the future plans of particle physics~\cite{bib:Man19, bib:Bos18}.
In the material and life sciences, one promising
application of muon acceleration is in the construction of
a transmission muon microscope~\cite{bib:tmm}.
If the muons can be cooled to thermal temperature and subsequently re-accelerated,
transmission muon microscopes will be
realized.
\rev{Among the future programs, a new experiment (E34 experiment) in the Materials and Life Science Experimental Facility (MLF) of the Japan Proton Accelerator Research Complex (J-PARC)
is planning to measure the muon anomalous magnetic moment (\ensuremath{a_{\mu}}) and search for the electric dipole moment (EDM) as a pioneer in muon acceleration~\cite{bib:Abe20}.}
\par
Although the discovery of the Higgs boson using the Large Hadron Collider (LHC) has established the Standard Model (SM)
as a successful description of particle interactions,
we are still confronted with many problems that can be solved only through experimental clues.
One of the most interesting clues is the precise measurement of \ensuremath{a_{\mu}}, which has paved the way for understanding the nature of elementary particles through the quantum effects.
In a series of three experiments at \rev{European Organization for Nuclear Research (CERN)}~\cite{bib:Cha62, bib:Bai72, bib:Bai79} and an experiment at Brookhaven National Laboratory (BNL)~\cite{bib:Ben06},
successive improvements in the accuracy of the measurements had enabled a deeper understanding of the SM and
there is a large discrepancy between measurements and predictions of the SM.
After many years of scrutiny, challenging calculations and corrections were conducted by stimulated theorists
and still 3.7 sigma deviation remains~\cite{bib:Aoyama20}.
This discrepancy should be addressed by new measurements.
A new experiment at Fermi National Accelerator Laboratory (FNAL E989~\cite{bib:Gra15}) is currently being conducted, with the storage ring from the BNL experiment being reused.
\rev{They published their first result in early FY2021~\cite{bib:abi21}.
The result is consistent to that of the previous BNL experiment and
the the tension between experiment and the SM calculation becomes 4.2 sigma.
It strengthens the importance for confirming the tension with with independent measurements different from BNL and FNAL. }
\par
As described above, continuous studies have continued improving the accuracy of \ensuremath{a_{\mu}}.
In particular, the improvement of the beam has been one of the driving forces moving the measurements forward.
Our ancestors at CERN, BNL, and FNAL have been struggling with beam-related uncertainties in their measurements
because they use muons obtained directly from a pion decay with a large emittance.
The J-PARC E34 experiment~\cite{bib:Abe20} aims to measure \ensuremath{a_{\mu}}~ with a precision of 0.1 ppm using an unprecedentedly low-emittance muon beam realized by the acceleration of thermal muons.
\par
The reminder of this paper is organized as follows.
In \revr{Section}~\ref{sec:e34}, the J-PARC E34 experiment, particularly the radio-frequency linear accelerator (linac) dedicated to muons is described.
Section~\ref{sec:muacc} describes the first demonstration of muon acceleration.
A summary and outlook are shown in \revr{Section}~\ref{sec:sum}.
\section{Muon linac for the J-PARC E34 experiment}\label{sec:e34}
In this section, details of the muon linac will be described after successive descriptions explaining an overview of the J-PARC E34 experiment.
\par
The J-PARC E34 experiment aims to measure \ensuremath{a_{\mu}}~ with a precision of 0.1 ppm and search for the EDM with a sensitivity
of approximately $10^{-21}$~$e\cdot$cm using a low-emittance muon beam realized by an accelerated thermal muon beam.
\rev{The total emittance in the transverse direction is required to be 1.5$\pi$~mm~mrad to realize the measurement
with a 3 T compact MRI-type magnet with sufficient injection efficiency.
The cyclotron radius with this magnet is 333~mm that is about a factor of 20 smaller than that for the BNL and FNAL experiments.
Because of the high uniformity of the magnetic field in the muon storage region, the uncertainty due to the field uniformity is much smaller than in the BNL and FNAL experiments. }
Figure~\ref{fig:e34} shows the experimental setup.
The pulsed high-power primary proton beam generates secondary surface muons produced by $\pi^{+}$ decay
near the surface of the production target~\cite{bib:Kawamura18}.
The produced surface muons are extracted and thermalized to form muoniums \rev{(bound state made up of a positive muon and an electron)}, which are then emitted into vacuum region adjacent to the muonium production target~\cite{bib:Bak13, bib:Beer14, bib:Bea20}.
The paired electron in the muonium is knocked out by a laser, and thermal muon (3 keV/c) is generated.
After acceleration to 300~MeV/c, the muon beam has an extremely low emittance such that it can be injected and stored in a high precision compact storage magnet~\cite{bib:Iinuma16, bib:Abe18},
where the time dependence of the positrons from the muon decay is measured for the measurement of the anomalous spin precession~\cite{bib:Aoyagi20, bib:Kishishita20}.
\par
\begin{figure}[!h]
\centering\includegraphics[width=0.95\textwidth]{fig/e34.eps}
\caption{Overview of the J-PARC E34 experiment. }
\label{fig:e34}
\end{figure}%
\par
\subsection{Overview of muon linac}
Because muons have a finite lifetime of approximately \rev{2.2}~$\mu$s, they need to be accelerated faster to avoid decay losses to obtain the necessary experimental statistics.
From this perspective, a linac is one of the best options for muon acceleration.
\par
Owing to its intermediate mass between that of protons and electrons,
the change in velocity upon acceleration is slower than that of electrons, as shown in Fig.~\ref{fig:velocity}.
For this reason, both proton and electron linac technologies will be used to accelerate the muon to near the speed of light.
\par
\begin{figure}[!h]
\centering\includegraphics[width=0.95\textwidth]{fig/211001_beta.eps}
\caption{Particle velocity as a function of kinetic energy for electron (black line), muon (red hatched line), and proton (blue dot-dash line)}
\label{fig:velocity}
\end{figure}%
Accelerators should be designed appropriately to reduce costs.
To reduce the expenses, a spare radio-frequency quadrupole (RFQ) of the J-PARC linac~\cite{bib:Kondo13} will be used as
a first stage acceleration.
It operates with a resonant frequency of 324~MHz.
The L-band high-power klystron developed for the KEKB linac upgrade~\cite{bib:Kubosaki11} will be used as the RF power supply for the acceleration cavity in the high-energy section to achieve a further cost reduction.
\par
A schematic of the muon linac is shown in Fig.~\ref{fig:mulinac}.
The RFQ bunches and accelerates the muons to 0.3~MeV after the initial electrostatic acceleration~\cite{bib:Kondo15}.
After the RFQ, an inter-digital H-mode drift tube linac (IH-DTL) is employed during the particle velocity $\beta=0.08$ to $0.28$ (4~MeV)~\cite{bib:Otani16}.
After the muon is accelerated to $\beta=0.28$, a disk-and-washer (DAW) type coupled cavity linac (CCL) with an operational frequency of 1296~MHz is employed~\cite{bib:Otani19daw}.
Because the $\beta$ variation is modest in the high-$\beta$ region, to realize a sufficiently
short distance, the design emphasis has been shifted to achieving a high accelerating gradient.
A disk-loaded structure (DLS) traveling-wave linac is used when $\beta$ is greater than $0.7$ (42~MeV)~\cite{bib:Kondo17}.
The details of each acceleration cavity are described below.
\begin{figure}[!h]
\centering\includegraphics[width=0.98\textwidth]{fig/mulinac_config.eps}
\caption{Configuration of the muon linac. \rev{In this figure, $W$ is kinetic energy and $\beta$ is the ratio of the muon velocity to the speed of light}.}
\label{fig:mulinac}
\end{figure}%
\subsection{Initial electrostatic acceleration}
In front of the RFQ, there is an electrostatic lens called a Soa lens~\cite{bib:Can86}, which accelerates and extracts the thermal muons.
\par
Figure~\ref{fig:soa} shows a schematic view of the Soa lens.
The Soa lens consists of two mesh electrodes and three cylindrical electrodes.
The first mesh electrode (target mesh) covers the downstream surface of the silica aerogel target.
The laser ionization region is between the target mesh and second mesh electrode (S1).
The voltage applied to the target mesh and S1 is set to 5.7 and 5.6~keV, respectively, corresponding to the input energy of the RFQ.
The dimensions of the electrodes are designed to cover the primary surface muons (rms size of 31 mm and 14 mm in the horizontal and vertical directions, respectively~\cite{bib:Otani18_surfmu})
and to provide a sufficient extraction efficiency for ultra slow muons.
The voltage applied to other electrodes (S2, S3 and S4) is determined using the simulation described below
such that the phase space of the beam matched with the design acceptance of the RFQ and a high transmission efficiency is obtained.
\par
\begin{figure}[!h]
\centering\includegraphics[width=4in]{fig/SOA.eps}
\caption{Schematic of the SOA lens. \rev{Lengths in the figure are expressed in millimeters.}}
\label{fig:soa}
\end{figure}%
Input muon distributions are estimated using a simulation based on measurements.
The simulation for a surface muon beamline (MLF H-line) is constructed using the g4beamline~\cite{bib:g4bl, bib:Kawamura18}.
The absolute number of surface muons is normalized by measurements from another beamline (MLF D-line~\cite{bib:Stra10}) that uses the same muon production target in J-PARC MLF.
The stopping distribution inside the silica aerogel is estimated using GEANT4~\cite{bib:g4}.
The simulation for muonium diffusion is developed based on a three-dimensional random walk.
The simulation parameters of the thermal temperature and the diffusion constant were determined from our measurements at \rev{Canada's particle accelerator centre (TRIUMF)}~\cite{bib:Beer14}.
The muon trajectories are simulated using the GEANT4 simulation where the electrostaic field of the Soa lens calculated OPERA~\cite{bib:opera} is implemented.
\par
Figure~\ref{fig:soa_xyps} shows the transverse phase space distributions at the entrance of the RFQ.
The difference between the horizontal and vertical directions is due to the difference in the primary surface muon distribution at the entrance of the silica aerogel.
Owing to the spatial distribution of the muoniums in the laser ionization region, the muons are distributed over time with a full width of approximately 10~ns.
The transmission efficiency in the Soa lens is estimated to be 72\%, including a 17\% decay loss.
\rev{Because the structure of the mesh electrodes (target mesh and S1) is not implemented in the GEANT4 simulation,
the transmission efficiency of the mesh electrode was estimated to be 78\% based on the products of the aperture ratio of the two meshes.
In total, the efficiency is 56\% in the initial electrostatic acceleration. }
\par
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/211001_rfqin.eps}
\caption{Phase space distributions at the entrance of the RFQ: (a) Horizontal divergence angle x'
vs. x, (b) the vertical divergence angle y' vs. y, (c) y vs. x, and (d) $\Delta W$ vs. time}
\label{fig:soa_xyps}
\end{figure}%
\subsection{RFQ}
After the initial acceleration by the Soa lens, the RFQ accelerates the muons to 340~keV.
In addition to the acceleration, the RFQ bunches the muons at a frequency of 324 MHz.
\par
\rev{
The principles of the RFQ were first invented in 1969~\cite{bib:Kap70_1, bib:Kap70_2}
and \revr{they proved} at Los Alamos National Laboratory (LANL) in 1980~\cite{bib:Sto81}.
The RFQ consists of a four electrodes excited with quadrupole-mode.
The electrode is modulated longitudinally, generating the axial electric field.
By changing modulation pattern gradually so that the synchronous phase is changed from $-90$ degree to higher, the beam bunching accomplished.
Because the RFQ supply velocity-independent electric focusing, it has great advantage in a low-velocity part compared with conventional linacs
that used velocity-dependent magnetic focusing.
}
\par
Table~\ref{tbl:rfqpqr} lists the design parameters of the RFQ for the muon acceleration along with that for negative hydrogen ion (H$^-$).
To accelerate muons using the spare of the J-PARC RFQ, which was originally developed for H$^-$ acceleration,
the intervane voltage must be reduced to the mass ratio to match the particle velocity.
As a result, the required power is reduced to the square of the mass ratio.
The input and output energies are also scaled to the design velocity;
the input and output $\beta$ are 0.01 and 0.08, respectively.
\begin{table}[h!tbp]
\centering
\caption{Parameters of the RFQ~II for the H$^{-}$ and muon acceleration.}
\label{}
\begin{tabular}{lll}
\hline
& H$^-$ & muon \\
\hline
Frequency (MHz) & \multicolumn{2}{c}{324} \\
Number of cells & \multicolumn{2}{c}{295} \\
Length (m) & \multicolumn{2}{c}{3.17} \\
Intervane voltage (kV) & 82.9 & 9.3 \\
Power (kW) & 330 & 4.2 \\
Injection energy (keV) & 50 & 5.6 \\
Extraction energy (MeV) & 3 & 0.34 \\
\hline
\end{tabular}
\label{tbl:rfqpqr}
\end{table}
\par
In order to confirm that the RFQ can accelerate muons without any problems,
particle simulations were performed using PARMTEQM~\cite{bib:parmteqm}.
Figure~\ref{fig:rfqsim} shows the phase space distributions at the RFQ exit, and
Table \ref{tbl:rfqsim} summarizes the input and output beam parameters.
The horizontal and vertical normalized rms emittances at the RFQ exit are 0.30 and 0.17 $\pi$~mm~mrad, respectively.
The simulated transmission is 94.7\%.
Since the transit time through one cell of the RFQ is half the resonant frequency, the transit time through the entire
RFQ can be calculated as $\frac{1}{324 \times 10^6} \times \frac{1}{2} \times 295 = 455$ ns.
Ignoring relativistic corrections, the muon lifetime is 2.2~$\mu$s, and the survival rate of muons passing through the RFQ is $\exp(-0.455/2.2) = 0.813$.
Therefore, the total transmission is 77.0\%.
\par
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/220402_tdr.eps}
\caption{Phase space distributions at the RFQ exit: (a) Horizontal divergence angle x'
vs. x, (b) vertical divergence angle y' vs. y, (c) y vs. x, and (d) $\Delta W$ vs. \revc{rf phase ($=$ time)}}
\label{fig:rfqsim}
\end{figure}%
\begin{table}[h!tbp]
\centering
\caption{Input and output beam parameters of the \revc{beam} dynamics simulation of the RFQ.
\revc{In this table, $\alpha$ and $\beta$ are Twiss parameters. } }
\begin{tabular}{lll}
\hline
& Input & Output \\
\hline
$\alpha_x$ & 0.32 & -1.51 \\
$\beta_x$ [mm/mrad] & 0.041 & 0.21 \\
$\varepsilon_x$ [$\pi$~mm~mrad, rms, normalized ] & 0.376 & 0.297 \\
$\alpha_y$ & 0.092 & 0.606 \\
$\beta_y$ [mm/mrad] & 0.080 & 0.076 \\
$\varepsilon_y$ [$\pi$~mm~mrad, rms, normalized ] & 0.106 & 0.167 \\
$\alpha_z$ & - & 0.17 \\
$\beta_z$[deg/MeV] & - & 1360 \\
$\varepsilon_z$[$\pi$~MeV~deg, rms, normalized ] & - & 0.0381 \\
Time width & 10~ns (full width) & - \\
Energy spread & 0.00989~keV (rms) & - \\
\hline
Transmission & \multicolumn{2}{c}{ 94.7\% } \\
Transient time & \multicolumn{2}{c}{ 455 ns } \\
Survival rate &\multicolumn{2}{c}{ 81.3\% } \\
Transmission total & \multicolumn{2}{c}{ 77.0\%} \\
\hline
\end{tabular}
\label{tbl:rfqsim}
\end{table}
The operation test of the RFQ has already been conducted~\cite{bib:Otani15_pasj}.
The RFQ was powered on by a solid state amplifier at up to 6 kW and a 25 Hz repetition.
Vacuuming is done with an ion pump and reaches $10^{-6}$ Pa.
Figure~\ref{fig:rfq_offline} shows the forward, reflection, and pick-up power in the RFQ when 5~kW power is applied.
There is a small reflection because the coupling was tuned to be overcoupled originally for a high current H$^-$ beam.
A few hours of operation at 5~kW was successfully achieved without any problems, such as an RF failure from a spark.
The background associated with the RF operation was measured using a micro-channel plate detector connected downstream of the RFQ.
The detector count rates were consistent and negligibly small with and without the RF operation, within a statistical uncertainty range of 0.1~Hz.
\par
In conclusion, we are ready for a muon acceleration using the RFQ.
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/RFQ_for_back.eps}
\caption{Forward, reflection wave, and pick-up power in RFQ with nominal power of 5 kW }
\label{fig:rfq_offline}
\end{figure}%
\subsection{IH-DTL}
After the initial acceleration and bunching by the RFQ, the IH-DTL accelerates the muons to 4.3~MeV.
\par
The IH-DTL has alternating drift tubes up and down through the stem allowing the TE11 mode to be used for acceleration.
\rev{The IH-DTL was first proposed in Japan in 1949~\cite{bib:Mor49}.
\revr{Although} there were several efforts for this inventions~\cite{bib:Ble56, bib:Zei62, bib:Pot69}, an IH-DTL was realized after a quarter century as the heavy ion post-accelerator at the Munich tandem accelerator~\cite{bib:Nol79}. }
Compared to the conventional Arvaretz DTL, a higher acceleration efficiency is achieved, particularly within the range of $\beta=0.1$--$0.2$~\cite{bib:Rat05}.
In recent IH-DTLs, the alternating phase focusing (APF) method~\cite{bib:Good53, bib:Min99}, which was successfully implemented in the Hadron therapy machine~\cite{bib:Iwata06}, enables simultaneous focusing in the horizontal and vertical directions using only electric fields, resulting in a higher acceleration efficiency.
Owing to the small focusing strength in the APF method, there is a limit to its use in high current machines,
but it can be applied in the muon linac, which has a small current.
\par
The IH-DTL with the APF method for accelerating muons was designed using several types of simulation software~\cite{bib:Otani16}.
The arrangement of the drift tubes and acceleration gaps, i.e., the so-called longitudinal beam dynamics design, should be determined according to the evolution of the beam velocity,
which is determined by the energy gain at each gap.
Unlike conventional accelerating cavities where transverse focusing is performed conducted using additional elements such as quadrupoles,
an APF cavity requires simultaneous non-independent longitudinal and transverse beam dynamics designs.
This beam dynamics design was performed based on linacsapf~\cite{bib:jameson14}, where the
beam dynamics is calculated using the so-called {\it drift-kick-drift} approximation.
To achieve less emittance growth and a better transmission, a nonlinear optimization is applied to the synchronous phase array.
The cavity was designed using CST-MW Studio~\cite{bib:cst} based on the beam dynamics design.
In order to mitigate the peak electric field and obtain a higher efficiency, the dimensions of the drift tube were changed from \cite{bib:Otani16}.
Table~\ref{tbl:ihdesign} summarizes the basic parameters of the IH-DTL cavity.
\par
\begin{table}[h!tbp]
\centering
\caption{Parameters of the IH-DTL cavity. Because the frequency is tuned to 324~MHz with conductive tuners, the frequency shown here is slightly smaller than 324~MHz.
In this table, \rev{$E_{surf.}$ is the surface electric field and} $E_{kp}$ is the Kilpatrick limit at 324~MHz (17.8~MV/m).}
\begin{tabular}{lll}
\hline
Parameters & Values \\
\hline
Frequency (MHz) & 323.3 \\
Number of cells & 16 \\
Length (m) & 1.32 \\
Synchronous phase [deg.] & $-44$--$+48$ \\
Unloaded $Q$ & $1.04\times10^{4}$ \\
Power (kW) & 322 \\
Peak $E_{surf.}$ (MV/m) & 35.3 ($\sim1.99 E_{kp}$) \\
\hline
\end{tabular}
\label{tbl:ihdesign}
\end{table}
The input distribution was given from the upstream RFQ simulation.
The input was used for the beam dynamics simulation after simulation of the beam transport line to match the beam to the IH-DTL.
The beam dynamics simulation is performed using GPT~\cite{bib:gpt}.
Figure~\ref{fig:ihsim} shows the phase space distributions at the IH-DTL exit, and
Table \ref{tbl:ihsim} summarizes the output beam parameters.
The vertical emittance growth is larger than the horizontal one,
which is due to the vertical meandering track caused by the vertical electric field and horizontal magnetic fields of the first and last cells, which is inevitable in IH-DTL structures.
The transmission is \rev{99.97}\% and it is sufficiently large.
The beam transit time \rev{$t_{\mathrm{trans.}}$} is 25~ns and the muon survival rate is calculated to be $\exp(t_{\mathrm{tran.}}/\tau_{\mu}\overline{\gamma})=98.9$\%, where the average Lorentz factor during acceleration is labeled $\overline{\gamma}$ \rev{and $\tau_{\mu}$ is the muon lifetime (2.2~$\mu$s)}.
\par
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/220402_ih_tdr.eps}
\caption{Phase space distributions at the IH-DTL exit: (a) Horizontal divergence angle x'
vs. x, (b) vertical divergence angle y' vs. y, (c) y vs. x, and (d) $\Delta W$ vs. \revc{rf phase ($=$ time)}}
\label{fig:ihsim}
\end{figure}%
\begin{table}[h!tbp]
\centering
\caption{Output beam parameters of the IH-DTL. \revc{In this table, $\alpha$ and $\beta$ are Twiss parameters. } }
\begin{tabular}{ll}
\hline
& Output \\
\hline
$\alpha_x$ & -4.3 \\
$\beta_x$ [mm/mrad] & 0.38 \\
$\varepsilon_x$ [$\pi$~mm~mrad, rms, normalized ] & 0.316 \\
$\alpha_y$ & -3.4 \\
$\beta_y$ [mm/mrad] & 0.33 \\
$\varepsilon_y$ [$\pi$~mm~mrad, rms, normalized ] & 0.190 \\
$\alpha_z$ & -1.0 \\
$\beta_z$[deg/MeV] & 174 \\
$\varepsilon_z$[$\pi$~MeV~deg, rms, normalized ] & 0.0274 \\
\hline
Transmission & 99.97\% \\
Transient time & 25 ns \\
Survival rate & 98.9\% \\
Transmission total & 98.9\% \\
\hline
\end{tabular}
\label{tbl:ihsim}
\end{table}
An error study of the beam dynamics was conducted by assuming several possible cases~\cite{bib:Otani16pasj}.
The error in the on-axis electric field owing to a fabrication error is estimated to be approximately 2\% using the CST-MW Studio
assuming a general fabrication error of 100~$\mu$m.
The emittance growth due to the fabrication errors is estimated to be 10\%, which still satisfies the requirement.
It also shows that the error field can be controlled by some movable conductive tuners installed on the cavity side wall.
\par
A prototype of the IH-DTL was fabricated (Fig.~\ref{fig:ihproto}) to confirm the design and evaluate the performance~\cite{bib:Nakazawa19}.
The prototype was a three-piece design with two semi-cylindrical shells attached to the center frame where the drift tubes are mounted.
It has the first five drift tubes of the actual design, and the total length is approximately 0.5 m.
The resonant frequency and unloaded quality factor ($Q_0$) were measured and are consistent with the \rev{CST-MW Studio} calculation at 0.1\% and 11\%, respectively.
The on-axis electric field distribution was measured using the bead-pull method~\cite{bib:bpm} and agreed with the \rev{CST-MW Studio} calculation within 3\% uncertainty.
The emittance growth owing to the field error is estimated to be 3\%, which is negligible.
A high power coupler was fabricated and a high power test will soon be conducted.
Based on the experience with the prototype, we completed a detailed design of the actual IH-DTL and will soon start production.
\begin{figure}[!h]
\centering\includegraphics[width=0.95\textwidth]{fig/ihproto.eps}
\caption{Prototype of the IH-DTL: (Left) overall structure, (Right) center plate, where the five drift tubes are implemented. }
\label{fig:ihproto}
\end{figure}%
\subsection{DAW-CCL}
After the IH-DTL, the DAW-CCL accelerates the muons to 40~MeV.
\par
A DAW-CCL is a cylindrical cavity with conductive washers placed across a disk-shaped disc.
\rev{A DAW-CCL was first proposed in early 1970s~\cite{bib:and72, bib:Mur72} and full scale cavity was demonstrated in late 1970s~\cite{bib:and76}.
The first DAW-CCL was operated in a proton and H$^-$ linac at the Moscow meson factory~\cite{bib:Esi88}.
In Japan, DAW-CCLs were developed in KEK and Kyoto for electron~\cite{bib:Ina86, bib:Iwa94}. }
DAW-CCLs have advantages over other CCLs, such as a higher shunt impedance and higher coupling between the accelerating and coupling cells.
\par
The DAW-CCL cavity is designed using several types of softwares~\cite{bib:Otani19daw}.
As a first step, the two-dimensional dimensions are optimized using \rev{Poisson Superfish}~\cite{bib:sf} for a higher acceleration efficiency
and lower peak-to-average value ($E_{\mathrm{max}}/E_{0}$) in satisfying the confluence condition.
After the two-dimensional calculation, the three dimensional calculations including the washer support are conducted using CST-MW Studio.
The bi-periodic L-support~\cite{bib:ao00},
in which a washer is fixed by two supports and a pair of supports are located azimuthally 90 degrees apart from the adjacent supports,
is adopted among several support structures because perturbation to the acceleration mode can be minimized by adjusting the support structure.
Finally, the dispersion curve is investigated to check whether an unfavorable mode exists around the operation frequency.
\par
Figure~\ref{fig:dawmodel} shows the three-dimensional model (left) and the dispersion curve (right) of the designed cavity for $\beta=0.3$.
TM02 and TM01 are accelerating and coupling mode, respectively.
Because of the bi-periodic structure, some stop bands appear in $\pi/2$.
Although the TM11 mode is near the operational frequency, the cavity is tuned during the optimization process such that the operational frequencies sit within the stop band at $\pi/2$.
Although the dipole mode passband TE11 crosses the line where the phase velocity matches the speed of the muons, it is not considered to be a problem because the muon beam current is negligible and the transverse kick owing to this mode is estimated to be much smaller than the requirement.
Table~\ref{tbl:dawcavity} summarizes the cavity parameters for $\beta=0.3, 0.4, 0.5,$ and $0.6$.
\par
\begin{figure}[!htb]
\centering
\includegraphics*[width=0.95\textwidth]{fig/DAW_result.eps}
\caption{(Left) Three-dimensional model of the DAW-CCL.
The electric field of the acceleration mode in the two-cell model is shown by the colored arrows.
(Right) Dispersion curve with optimized cavity in $\beta=0.3$ calculated using CST-MW Studio.
\rev{The dotted line represents target frequency (1.3~GHz) and the dashed line corresponds to the phase velocity of the muons. }
\revr{The sup mode is associated with the L-support.}
}
\label{fig:dawmodel}
\end{figure}%
\begin{table}[!htb]
\centering
\caption{Parameters of the DAW-CCL cavity for $\beta=0.6, 0.5, 0.4,$ and $0.3$. For the simplicity, the resonant frequency is designed to be 1.3~GHz.
\rev{In this Table, L is the cell length,
$\mathrm{f_a}$ is the resonant frequency of the accelerating mode,
$\mathrm{f_c}$ is the resonant frequency of the coupling mode,
ZTT is the effective shunt impedance per unit length,
$E_{\mathrm{max}}/E_{0}$ is the peak to average axial electric-field, and
$\lambda$ is the resonant wavelength. }}
\begin{tabular}{lrrrr}
\hline
Parameters & \multicolumn{4}{c} {Values} \\
\hline
$\beta$ & 0.6 & 0.5 & 0.4 & 0.3 \\
$\mathrm{L}$ & \multicolumn{4}{c} {$\beta\lambda/4$} \\
$\mathrm{f_{a}[GHz]}$ & 1.300 & 1.300 & 1.299 & 1.301 \\
$\mathrm{f_{c}[GHz]}$ & 1.299 & 1.301 & 1.302 & 1.301 \\
$\mathrm{ZTT[M\Omega/m]}$ & 57.8 & 46.3 & 33.8 & 18.0 \\
Transit time factor & 0.84 & 0.85 & 0.84 & 0.81 \\
$E_{\mathrm{max}}/E_{0}$ & 4.4 & 4.8 & 5.1 & 5.0 \\
$\mathrm{Q_{0}}$ & $2.91\times10^4$ & $2.41\times10^4$ & $1.91\times10^4$ & $1.42\times10^4$ \\
Synchronous phase [deg.] & \multicolumn{4}{c}{ -30 } \\
\hline
\end{tabular}
\label{tbl:dawcavity}
\end{table}%
The beam dynamics is designed using PARMILA~\cite{bib:parmila} and TRACE3D~\cite{bib:t3d} based on the designed cavity performance.
For ease of fabrication, a constant cell length \revc{($L=\beta_{s}\lambda/2$)} in a tank with multiple cells is designed.
The design velocity $\beta_{s}$ and number of cells for each tank are determined based on the design of the beam dynamics.
The inter-tank spacing is set to 4.5$\beta\lambda$ considering the feasibility of a magnet installation.
The average acceleration field is determined to be 5.6~MV/m according to the Kilpatrick limit~\cite{bib:Kil57} and $E_{\mathrm{max}}/E_{0}$;
the bravery factor is set to 0.9 in maximum.
The number of cells in a tank is determined by the limitation of the quadrupole strength owing to transverse instabilities.
There may be instabilities or resonances when the zero current phase advance ($\sigma_{0}$) is greater than 90 degrees~\cite{bib:Reiser94}.
Although the muon beam intensity is expected to be much smaller than that in the region of such instabilities, the number of cells is chosen to match this criterion.
The maximum $\sigma_{0}$ is 83~degrees at the first tank when the number of cells is ten, which determines the number of cells for all tanks.
Table~\ref{tbl:dawtanks} shows the basic parameters of each DAW tank.
Because the shunt impedance of the DAW cell increases as a function of $\beta$,
the power required for the tank decreases with a higher energy.
The total power required is 4.5~MW.
The {\it phase slippage} is the greatest in the first tank, ranging from $-14$ to $+10$ degrees.
\begin{table}[h!tbp]
\centering
\caption{Parameters of each DAW tank.}
\begin{tabular}{lrrrrr}
\hline
tank & $N_{\mathrm{cells}}$ & $\beta_s$ & length [m] & Energy [MeV] & Power [MW] \\
\hline
1 & 10 & 0.29 & 0.34 & 5.6 & 0.39 \\
2 &$\uparrow$& 0.33 & 0.38 & 7.1 & 0.35 \\
3 &$\uparrow$ & 0.37 & 0.42 & 8.8 & 0.33 \\
4 &$\uparrow$ & 0.40 & 0.46 & 10.7 & 0.31 \\
5 &$\uparrow$& 0.43 & 0.50 & 12.7 & 0.30 \\
6 &$\uparrow$& 0.47 & 0.54 & 14.9 & 0.29 \\
7 &$\uparrow$& 0.50 & 0.57 & 17.2 & 0.29 \\
8 &$\uparrow$& 0.52 & 0.61 & 19.7 & 0.29 \\
9 &$\uparrow$& 0.55 & 0.64 & 22.3 & 0.28 \\
10 &$\uparrow$& 0.58 & 0.67 & 25.0 & 0.28 \\
11 &$\uparrow$& 0.60 & 0.69 & 27.9 & 0.28 \\
12 &$\uparrow$& 0.62 & 0.72 & 30.8 & 0.28 \\
13 &$\uparrow$& 0.64 & 0.74 & 33.8 & 0.28 \\
14 &$\uparrow$& 0.66 & 0.77 & 37.0 & 0.28 \\
15 &$\uparrow$& 0.68 & 0.79 & 40.2 & 0.28 \\
\hline
\end{tabular}
\label{tbl:dawtanks}
\end{table}
\par
The input distribution was given from the upstream IH-DTL simulation and used for the beam dynamics simulation after the calculation for the beam transport line to match the beam to the DAW-CCL.
Figure~\ref{fig:dawsim} shows the phase space distributions at the DAW exit, and
Table \ref{tbl:dawsim} summarizes the output beam parameters.
The total length is 16.3~m with 15 modules, which corresponds to 138$\beta\lambda$, and the beam transit time \rev{$t_{\mathrm{trans.}}$} is calculated as $\frac{138\beta\lambda}{\beta c}=106.0$~ns.
The survival probability is calculated as $\exp(t_{\mathrm{tran.}}/\tau_{\mu}\overline{\gamma})=96.1$\%, where $\overline{\gamma}=(\gamma_{\mathrm{in}}+\gamma_{\mathrm{out}})/2=1.212$.
The emittance growth is estimated to be less than a few percent, and the output emittance is 0.32~$\pi$ and 0.21~$\pi$~ mm~mrad
for the horizontal and vertical directions, respectively.
The effect of errors on the emittance owing to a misalignment of the quadrupole magnet, power errors in the cavity, and other factors is evaluated to be less than 5\%~\cite{bib:takeuchi19}.
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/220402_tdr_daw.eps}
\caption{Phase space distributions at the DAW-CCL exit: (a) Horizontal divergence angle x'
vs. x, (b) vertical divergence angle y' vs. y, (c) y vs. x, and (d) $\Delta W$ vs. \revc{rf phase ($=$ time)}}
\label{fig:dawsim}
\end{figure}%
\begin{table}[h!tbp]
\centering
\caption{Output beam parameters of the DAW-CCL. \revc{In this table, $\alpha$ and $\beta$ are Twiss parameters. } }
\begin{tabular}{ll}
\hline
& Output \\
\hline
$\alpha_x$ & 4.1 \\
$\beta_x$ [mm/mrad] & 0.23 \\
$\varepsilon_x$ [$\pi$~mm~mrad, rms, normalized ] & 0.332 \\
$\alpha_y$ & -6.9 \\
$\beta_y$ [mm/mrad] & 0.36 \\
$\varepsilon_y$ [$\pi$~mm~mrad, rms, normalized ] & 0.201 \\
$\alpha_z$ & -0.002 \\
$\beta_z$[deg/MeV] & 24 \\
$\varepsilon_z$[$\pi$~MeV~deg, rms, normalized ] & 0.108 \\
\hline
Transmission & 99.80\% \\
Transient time & 106 ns \\
Survival rate & 96.1\% \\
Transmission total & 95.9\% \\
\hline
\end{tabular}
\label{tbl:dawsim}
\end{table}
\par
A cold model of the first DAW cells was fabricated to confirm the design.
\rev{The cold model is made of aluminum. }
Figure~\ref{fig:dawproto} shows a mechanical drawing (left) and photograph of the assembly (right).
\rev{The resonant frequency was measured and consistent with the calculation using CST-MW Studio at 0.4\%. }
The on-axis electric field distribution was measured and the variation in the fields per cell was observed, which is considered to be due to the assembly process of the cavity.
Based on the measurement results in the cold model, the actual DAW-CCL has been designed and the first tanks will be soon fabricated.
\begin{figure}[!h]
\centering\includegraphics[width=0.95\textwidth]{fig/dawproto.eps}
\caption{Prototype of the DAW-CCL: (Left) mechanical three-dimensional drawing, (Right) assembly of the prototype.
\rev{The end plate with the antenna ports is attached for the low power measurements.}}
\label{fig:dawproto}
\end{figure}%
\subsection{DLS}
Finally, muons are accelerated to 212~MeV using the DLS.
\par
\rev{DLS is classified as a radio-frequency linac for electron.
After the first invention of the radio-frequency linac by Wider\"{o}e, Sloan and Lawrence~\cite{bib:Wid28, bib:Law31}, Hansen studied an electromagnetic field with radio-frequency resonator for accelerating electron~\cite{bib:Hansen38}.
The acceleration of electron using a traveling wave accelerator was demonstrated in late 1940s~\cite{bib:Gin48}. }
Recent electron accelerators using room-temperature cavities have been based on the results of the Mark III linac~\cite{bib:Cho55} and SLAC.
Unlike linear accelerators for electrons, which quickly reach the speed of light, muon linacs, which slowly approach the speed of light, require a gradual change in the length of the cell.
\par
The geometrical parameters of the DLS cell are designed using \rev{Poisson Superfish}.
\rev{Poisson Superfish} generates the standing-wave-mode electric fields of each cell with open-open and short-short boundary conditions and
the electric field of the traveling wave is represented by superposing these two fields with a phase difference of $\pi/2$.
The L-band structure was adopted to make the acceptance sufficiently large for the input muon beam and
the conventional $2\pi/3$ acceleration mode is adopted.
The synchronous phase is set to $-10$ degrees to ensure sufficient longitudinal acceptance, and
the average acceleration field ($E_0$) is assumed to be 20~MV/m.
Figure~\ref{fig:dlsparam} shows the parameters of the DLS cells.
In this study, a constant impedance design was adopted for simplicity.
The calculated fields are implemented in the beam dynamics simulation conducted using the GPT.
Figure~\ref{fig:dlssim} shows the phase space distributions at the DLS exit, and
Table \ref{tbl:dlssim} summarizes the output beam parameters.
Almost no emittance growth is observed.
The transmission through the DLS section is \rev{99.9}\%, and the loss owing to the muon decay is estimated to be 1\%.
\begin{figure}[!h]
\centering\includegraphics[width=4in]{fig/dlspar.eps}
\caption{Calculated cell parameters of the DLS section. $D$ is the cell length calculated by $\beta\lambda/3$ \rev{and $w$ is the kinetic energy of the muons.}}
\label{fig:dlsparam}
\end{figure}%
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/220402_tdr_dls.eps}
\caption{Phase space distributions at the DLS exit: (a) Horizontal divergence angle x'
vs. x, (b) vertical divergence angle y' vs. y, (c) y vs. x, and (d) $\Delta W$ vs. \revc{rf phase ($=$ time)}}
\label{fig:dlssim}
\end{figure}%
\begin{table}[h!tbp]
\centering
\caption{Output beam parameters of the DLS. \revc{In this table, $\alpha$ and $\beta$ are Twiss parameters. } }
\begin{tabular}{ll}
\hline
& Output \\
\hline
$\alpha_x$ & -1.9 \\
$\beta_x$ [mm/mrad] & 0.86 \\
$\varepsilon_x$ [$\pi$~mm~mrad, rms, normalized ] & 0.331 \\
$\alpha_y$ & -3.3 \\
$\beta_y$ [mm/mrad] & 1.5 \\
$\varepsilon_y$ [$\pi$~mm~mrad, rms, normalized ] & 0.211 \\
$\alpha_z$ & 0.07 \\
$\beta_z$[deg/MeV] & 150 \\
$\varepsilon_z$[$\pi$~MeV~deg, rms, normalized ] & 1.94 \\
\hline
Transmission & 99.9\% \\
Transient time & 30 ns \\
Survival rate & 99\% \\
Transmission total & 99\% \\
\hline
\end{tabular}
\label{tbl:dlssim}
\end{table}
\subsection{Summary of the muon linac design}
Figure~\ref{fig:emitlinac} shows the emittance evolution in the entire muon linac.
There is no significant growth of the beam emittance and the output emittance is comparable to the \rev{required total emittance of 1.5$\pi$~mm~mrad}.
Table~\ref{tbl:linacsum} summarizes the transmission, decay loss, and emittance at each section.
The intensity of the low emittance muon beam at the linac exit is estimated to be \rev{$1.2\times10^4$ muons per pulse with a repetition of 25-Hz}.
\rev{Table~\ref{tbl:allsum} shows the breakdown of the estimated transmission efficiency for all the experimental components and}
the statistical uncertainty of the \ensuremath{a_{\mu}}~measurement is estimated to be 450~ppb for $2.2\times10^7$ s of data taking~\cite{bib:Abe20}.
The statistical precision is comparable to that of the BNL experiment, and we are able to examine the \ensuremath{a_{\mu}}~ anomaly with a completely different scheme.
\par
\begin{figure}[!h]
\centering\includegraphics[width=0.95\textwidth]{fig/211001_emi.eps}
\caption{Emittance evolution from the RFQ entrance to the linac exit. }
\label{fig:emitlinac}
\end{figure}%
\begin{table}[h!tbp]
\centering
\caption{Summary of the particle simulations through the muon linac: Transmission and decay loss in each section, and the emittance at the exit of each section.
\rev{The transmission in SOA includes that of the mesh electrodes. }
}
\begin{tabular}{lccccc}
\hline
& Soa & RFQ & IH-DTL & DAW-CCL & DLS \\
\hline
Transmission \rev{[\%]} & \rev{56} & 95 & 100 & 100 & 100 \\
Decay loss \rev{[\%]} & 17 & 19 & 1 & 4 & 1\\
$\varepsilon_x$ [$\pi$~mm~mrad, rms, normalized]&0.38 & 0.30& 0.32 & 0.32 & 0.33\\
$\varepsilon_y$ [$\pi$~mm~mrad, rms, normalized]&0.11 & 0.17& 0.20 & 0.21 & 0.21\\
\hline
\end{tabular}
\label{tbl:linacsum}
\end{table}
\begin{table}[h!tbp]
\centering
\caption{\rev{Breakdown of the estimated transmission efficiency. For the \ensuremath{a_{\mu}}\ measurement, $e^{+}$ energy window (12\%), detector acceptance (100\%), and reconstruction efficiency (12\%) should be considered. Details are described in~\cite{bib:Abe20}. }}
\begin{tabular}{lcccccc}
\hline
& H-line & \begin{tabular}{c}muonium \\emission\end{tabular} & \begin{tabular}{c}laser \\ionization \end{tabular}& acceleration & injection & kicker \\
\hline
Transmission [\%] & 16 & 0.34 & 73 & 40 & 84 & 93\\
\hline
\end{tabular}
\label{tbl:allsum}
\end{table}
Because the systematic uncertainty is estimated to be less than 70~ppb and the measurement is statistically limited,
further improvement of the beam intensity directly impacts the sensitivity of the physics.
\rev{The goal in the next phase in the J-PARC \revr{E34} experiment is to measure \ensuremath{a_{\mu}}~with an accuracy of 100~ppb, comparable to the goal of the FNAL experiment.
As shown in Table~\ref{tbl:allsum}, primary improvement should be in the muonium emission efficiency.
The next one is the H-line and the acceleration efficiency.
According to the results of previous muonium measurements~\cite{bib:Bak13, bib:Beer14, bib:Bea20},
the arrangement of the silica aerogel target is being investigated to enhance the overlap between the effective area of the target and the laser ionization region.
}
In the muon linac, the most effective way for the improvement is a reduction of the decay loss in the low-$\beta$ section.
For this purpose, several efforts have been devoted: a new RFQ dedicated to the muons~\cite{bib:Kondo15}, an L-band RFQ combined with CC-DTL~\cite{bib:kondo20, bib:kondo18linac},
and a new scheme of acceleration for the low-energy muons~\cite{bib:otani20}.
\rev{}
\par
Further, development is carried out for spin tracking~\cite{bib:yasuda20}.
Because a spin precession in the muon storage ring starts with the initial spin state determined by the muon linac exit,
the initial state should be understood to examine the potential systematic uncertainties of the physics measurement.
For this purpose, a spin tracking simulation is being developed.
In addition, a spin-rotator based on the Wien-filter type is designed to enable a spin-flipping analysis~\cite{bib:yasuda20jparc}, which is
usually used in many types of spin experiments.
\section{Demonstration of muon acceleration}\label{sec:muacc}
As described in Section~\ref{sec:e34}, a low emittance muon beam is
realized through thermal muon production and acceleration using the linac dedicated to the muons.
The production of thermal muons has been developed for over a quarter of a century~\cite{bib:Mills86, bib:Chu88, bib:Nagamine95} and has matured sufficiently and can be used in experiments.
Muon acceleration, however, was an unproven technology.
Therefore, a muon acceleration should have been demonstrated prior to the construction of the actual linac.
In addition, beam monitors are needed to diagnose the muon beam with an unprecedented beam size.
\par
To demonstrate muon acceleration prior to the construction of the experimental
setup that included the thermal muon source system, a faster and simpler system for slow muons was needed.
The scheme of muon cooling using a simple metal degrader and
acceleration using an RFQ proposed at Los Alamos National Laboratory~\cite{bib:Miyadera07}
are suitable for this purpose.
We basically followed this method, but to further reduce the emittance and separate accelerated muons from background
high energy muons that pass through the RFQ without being accelerated, we decided to use negative muonium ions
($\mu^{+} e^{-} e^{-}$; $\mathrm{Mu^{-}}$) produced through an electron capture process~\cite{bib:Kua89}.
Prior to the demonstration of muon acceleration by $\mathrm{Mu^{-}}$, an experiment dedicated to a $\mathrm{Mu^{-}}$~ measurement was carried out
to evaluate the expected amount of accelerating $\mathrm{Mu^{-}}$~ signal.
\par
This chapter describes the $\mathrm{Mu^{-}}$~ measurement followed by a demonstration of muon acceleration along with a description of the development of the beam monitor.
\subsection{Development of $\mathrm{Mu^{-}}$~source}
As mentioned above, the $\mathrm{Mu^{-}}$~production process can be used to cool a muon beam down to 1~keV
by simply irradiating positive muons ($\mu^+$'s) onto a thin metal film such as aluminum.
However, since the first observation during the 1980s~\cite{bib:Kua89, bib:Har86}, there have been no measurements or no proof that a significant intensity
can be obtained for an acceleration test.
Therefore, we decided to conduct the $\mathrm{Mu^{-}}$~measurement prior to the muon acceleration test.
\par
In addition to the estimation of the accelerated muon signal,
the method for identifying the background caused by positrons from muon decay (decay-positron) was also important.
This is because the conversion efficiency from the primary $\mu^+$ to $\mathrm{Mu^{-}}$~ is expected to be less than $10^{-4}$, which results in a large amount of decay-positrons derived from muon decays at the target.
Therefore, the decay-positron background and low energy muon (LE-$\mu$) measurements were carried out~\cite{bib:Otani19mcp}
using a micro-channel plate (MCP) detector (Hamamatsu photonics, F1217-01~\cite{bib:hamamatsu}) in advance of the $\mathrm{Mu^{-}}$~ measurements.
In the decay-positron measurement, the decay-positrons from a muon beam target were identified by a series of triple scintillation detectors installed in front of the MCP detector.
In the LE-$\mu$ measurement, the 7-keV $\mu^+$s are measured using the same setup as the $\mathrm{Mu^{-}}$~measurement described below.
Figure~\ref{fig:phbgmu} shows the pulse height distribution
of the decay-positron (red triangle) and LE-$\mu$ (blue box).
The difference between the decay-positrons and LE-$\mu$ in the pulse height distributions was
evident, which can be explained by the single- and multi- channel amplification model described in Section 4 of ~\cite{bib:Otani19mcp}.
The results show that event selection using the pulse height can reject the decay-positron more than 75\% while maintaining the muon efficiency at 90\%.
\par
\begin{figure}[!h]
\centering\includegraphics[width=0.85\textwidth]{fig/phbgmu_rev.eps}
\caption{Pulse height distribution for decay-positrons (red triangle) and LE-$\mu$ (blue box).
Red dashed line and blue dotted line show the pulse height distribution with multi- and single-channel excitation models, respectively, as discussed in~\cite{bib:Otani19mcp}.
\rev{This figure is cited from~\cite{bib:Otani19mcp}. }}
\label{fig:phbgmu}
\end{figure}%
The $\mathrm{Mu^{-}}$~ measurements were conducted in a series of three experiments at the J-PARC MLF
(2015A0324~\cite{bib:Otani19mcp}, 2016A0067~\cite{bib:rkita21}, and 2018B0007~\cite{bib:Otani19mum}).
Figure~\ref{fig:mumsetup} shows a typical setup for the experiments.
The $\mu^+$'s were injected into a $\mathrm{Mu^{-}}$~production target after passing through a \rev{steel} window.
The $\mathrm{Mu^{-}}$~generated in the target was accelerated at up to 20~keV by the SOA electrostatic lens.
Then, the $\mathrm{Mu^{-}}$~was transported to the detector location through a series of electrostatic quadrupole (EQ1-4), an electrostatic deflector (ED), and a bending magnet (BM).
The energy acceptance of the beamline is estimated to be 1.4\% by the GEANT4 simulation.
The MCP detector was employed to measure the time of flight (TOF) from the $\mathrm{Mu^{-}}$~ production target.
The electrical signal from the MCP detector was amplified using a fast-filter amplifier (ORTEC 579) and
digitized using CAEN V1720.
The waveform within an interval of 10~$\mu$sec around each 25-Hz beam pulse was recorded for analysis.
A pulse higher than the noise level was regraded as a signal pulse.
The leading edge of the signal pulse was defined as the signal timing.
The pulse height is defined by the maximum height within the signal window of 40~ns.
The $\mu^+$ arrival time at the $\mathrm{Mu^{-}}$\ production target was measured with a set of scintillating counters located at the side of the $\mathrm{Mu^{-}}$\ production target.
\par
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/mumsetup.eps}
\caption{Schematic of the typical experimental setup for the $\mathrm{Mu^{-}}$~ measurements.
\revr{This figure is cited from~\cite{bib:Otani19mum}.}}
\label{fig:mumsetup}
\end{figure}%
In the 2015A0324 experiment, the LE-$\mu$ measurement was conducted to
commission the experimental setup and compare the MCP signal to the decay-positron signal as described above.
The aluminum (Al) thin foil is adopted as the target and LE-$\mu$ is extracted instead of $\mathrm{Mu^{-}}$~by flipping the beamline polarity and transported to the MCP detector.
The LE-$\mu$ signal was used to tune the applied voltage of the quadrupole and other experimental setups.
In addition to the tuning using the LE-$\mu$, negative hydrogen ions generated by ultraviolet light~\cite{bib:nakazawa19} were used
for offline commissioning of the experimental setup.
\par
The $\mathrm{Mu^{-}}$ s were observed during the 2016A0067 experiment, and high-statistics $\mathrm{Mu^{-}}$~data were obtained in the 2018B0007 experiment using three types of production targets:
an Al foil, a C12A7 electride foil~\cite{bib:hosono03, bib:hosono17}, and a \rev{steel} foil.
The electrical signals of the MCP detector were recorded and analyzed in a way similar to the 2015A0324 experiment.
Figure~\ref{fig:tq} (A) shows a pulse height versus the TOF of the observed signal.
The background events are considered to be decay-positrons from muons stopped in the experimental apparatus,
based on the results of the analysis of time constants.
The decay-positron events can be eliminated efficiently by the pulse height selection, as described above.
After pulse height selection, the TOF distribution is obtained as shown in Fig.~\ref{fig:tq} (B).
The two peaks at approximately $-300$~ns and $300$~ns are due to prompt positrons carried through the surface muon beamline with the same momentum as $\mu^+$'s.
Because the prompt positron is faster than $\mu^+$, the prompt positron arrives earlier than $\mu^+$.
The blue and red curves show the fitting results assuming the remaining decay-positron background (blue) and the $\mathrm{Mu^{-}}$\ signals (red), respectively.
The background is consistent with the exponential decay curve with the muon decay constant ($\tau_{\mu}=2.2$~$\mu$sec).
The $\mathrm{Mu^{-}}$~peak width is consistent with that of the primary $\mu^+$ beam.
The time interval of the $\mathrm{Mu^{-}}$~peaks is consistent with that of the primary proton beam pulses.
The $\mathrm{Mu^{-}}$~TOF is consistent within a few percentage points with the expectation estimated through the GEANT4 simulation where the initial energy of $\mathrm{Mu^{-}}$\ is assumed to be 0.2~keV.
\par
\begin{figure}[!h]
\centering\includegraphics[width=5in]{fig/mumresult.eps}
\caption{(A) Pulse height vs. TOF for observed signal. (B) TOF distribution after pulse height selection.
The solid red line and blue hatched line shows the fitting results assuming the $\mathrm{Mu^{-}}$\ event and the decay-positron background.
\revr{This figure is cited from~\cite{bib:Otani19mum}. }
}
\label{fig:tq}
\end{figure}%
Based on the measured $\mathrm{Mu^{-}}$~intensity during the 2016A0067 experiment,
The expected intensity of the accelerated $\mathrm{Mu^{-}}$~signal was estimated to be $6.8\pm1.2$ (stat.)$^{+0.4}_{-0.0}$ (syst.)/s~\cite{bib:rkita21}, \rev{where (stat.) represents the statistical error assigned by the measured $\mathrm{Mu^{-}}$~intensity in the experiment and (syst.) represents the systematic error. }
The effect of a misalignment of the $\mathrm{Mu^{-}}$~transport beamline is taken into account as a systematic error.
From these results, it can be concluded that several hundred events of accelerated $\mathrm{Mu^{-}}$~signals can be expected with a beam time of a few days,
and that the background of the decay-positrons can be sufficiently suppressed by a pulse height discrimination, such that a sufficient amount of the accelerated $\mathrm{Mu^{-}}$'s can be observed with a sufficient signal-to-noise ratio.
\subsection{First muon acceleration}
A demonstration of the muon acceleration was conducted in J-PARC MLF over a 6-day period starting on October 24, 2017 (2017A0263)~\cite{bib:Bae18}.
\par
Figure~\ref{fig:muaccsetup} shows a schematic drawing of the experimental setup.
The $\mu^+$'s were incident on the Al foil target used in the $\mathrm{Mu^{-}}$~measurements.
The $\mu^+$'s were decelerated through the target, and some of the
$\mu^+$'s become Mu$^-$'s at the
downstream surface of the Al target.
Using the same Soa lens as used in the $\mathrm{Mu^{-}}$ measurements, the generated $\mathrm{Mu^{-}}$'s
were accelerated to 5.6~keV and focused on the entrance of the RFQ.
In this experiment, the prototype RFQ~\cite{bib:kondo06} was used.
The length of this RFQ is equivalent to two-thirds of the RFQ in the muon linac, and
this RFQ was originally designed using KEKRFQ~\cite{bib:Ueno90} to accelerate negative hydrogen ions at up to 810~keV.
To use this RFQ for a muon acceleration, the intervane voltage should
be normalized to the muon mass, and the input velocity $\beta$ should
be the same as that of H$^-$.
Table~\ref{tbl:xrfqpqr} summarizes the parameters of the prototype RFQ.
The $\mathrm{Mu^{-}}$'s are accelerated to 89~keV and then transported to the MCP detector through the diagnostic beamline
consisting of two quadrupole magnets (QM1 and QM2) and a bending magnet (BM).
Because the expected number of $\mathrm{Mu^{-}}$~ signals is only a few hundred, and it is impossible to obtain the correct field setting using the $\mathrm{Mu^{-}}$~ signals themselves,
the diagnostic beamline was commissioned using negative hydrogen ions generated by ultraviolet light~\cite{bib:nakazawa19}, and the field setting of the magnets was verified prior to the experiment.
The electrical signals of the MCP detector are recorded in a similar way as the experiments of the $\mathrm{Mu^{-}}$~measurements.
\par
\begin{figure*}[!hbt]
\centering
\includegraphics[width = 0.95\textwidth]{fig/muaccsetup.eps}
\caption{Schematic drawing of the experimental setup for the demonstration of muon acceleration.
\rev{In this figure, TMP is turbo molecular pump. }}
\label{fig:muaccsetup}
\end{figure*}
\begin{table}[h!tbp]
\centering
\caption{Parameters of the prototype RFQ for the H$^{-}$ and muon acceleration.}
\label{}
\begin{tabular}{lll}
\hline
& H$^-$ & muon \\
\hline
Frequency (MHz) & \multicolumn{2}{c}{324} \\
Number of cells & \multicolumn{2}{c}{297} \\
Length (m) & \multicolumn{2}{c}{1.97} \\
Intervane voltage (kV) & 81 & 9.1 \\
Power (kW) & 180 & 2.3 \\
Injection energy (keV) & 50 & 5.6 \\
Extraction energy (keV) & 810 & 89 \\
\hline
\end{tabular}
\label{tbl:xrfqpqr}
\end{table}
At the start of the experiment, the polarity of the beamline was set to transport positively charged particles.
In this positive-charge configuration, muons that passed through the target and were decelerated to 89 keV, which corresponds to the acceleration energy, were measured.
The observed TOF is consistent with calculations based on the distance between the target and the MCP detector and the velocity of the 89-keV muon.
This confirms the beam diagnostic system.
\par
After measurement of 89-keV $\mu^+$'s, the polarities of the magnets were flipped to a negative-charge configuration.
Figure~\ref{fig:tof} shows the TOF spectrum with and without
the RFQ operation after the pulse-height cut was applied.
With the RFQ operation, a clear peak was observed at $830\pm11$~ns.
The time required to reach the RFQ entrance from the target while being accelerated by the Soa lens was estimated as 307~ns using the GEANT4 simulation.
The number of cells of the prototype RFQ is 297, and thus it takes
$\frac{297}{2\times324\times10^6} = 458$~ns to fully accelerate the
particles through the 324~MHz field.
The length of the diagnostic beamline is 0.91~m, and thus the transit time of the 89-keV Mu$^-$ is 72~ns.
The total flight time of the accelerated Mu$^-$ from the target to the MCP detector was calculated as
$307 + 458 + 72 = 837$~ns, which is consistent with the measurement.
The hatched histogram in Fig.~\ref{fig:tof} represents the simulated TOF spectrum of the accelerated Mu$^-$.
The number of simulation events was normalized to the number of incident muons of this data set.
The 46~ns rms width of the TOF spectrum is consistent with that
from the timing distribution of the primary $\mu^+$ at the Al target.
\par
From these experimental results, it is concluded that the observed
TOF peak is due to the Mu$^-$'s accelerated by the RFQ to 89~keV.
The event rate was estimated as $(5\pm1) \times 10^{-4}$/s by subtracting the decay-positron events
estimated from the timing region outside the signal range.
This is consistent with the expectation based on the $\mathrm{Mu^{-}}$~ measurements.
The intensity of the accelerated $\mathrm{Mu^{-}}$~ is limited by the low conversion efficiency of $\mu^+$ to $\mathrm{Mu^{-}}$.
\par
\begin{figure}[hbt]
\centering
\includegraphics[width=0.9\textwidth]{fig/muacc_rev.eps}
\caption{TOF spectra of the negative-charge configuration with RF on and off.
The clear peak of the RF on the spectrum at 830~ns corresponds to
the accelerated Mu$^-$'s. The error bars are statistical.
A simulated TOF spectrum of the accelerated Mu$^-$'s is also plotted.
\rev{This figure is cited from~\cite{bib:Bae18}. } }
\label{fig:tof}
\end{figure}
\subsection{Development of beam monitors}\label{sec:monitor}
After the first demonstration of the muon acceleration described above,
the beam monitors for the accelerated muons were tested using the accelerated $\mathrm{Mu^{-}}$'s.
To diagnose the beam in both the transverse and longitudinal directions,
a beam profile monitor (BPM) and a bunch width monitor (BWM) are necessary.
\par
The BPM has been developed based on a MCP assembly (Hamamatsu F2225-21P) and a charge-coupled device (CCD) camera.
The avalanche electrons generated by the MCP are injected into the phosphor screen, from which the light output is focused onto the CCD camera.
The spatial resolution was estimated as 0.3~mm using an ultraviolet light and surface muons~\cite{bib:kim18}.
During the 2016B0214 experiment, the detector in the setup for the $\mathrm{Mu^{-}}$~measurement shown in Fig.~\ref{fig:mumsetup} was replaced with the BPM
and the performance against LE-$\mu$ was demonstrated~\cite{bib:rkita17}.
During the 2017B0006 experiment, the detector in the setup shown in Fig.~\ref{fig:muaccsetup} is replaced with the BPM
and the beam profiles of the accelerated $\mathrm{Mu^{-}}$'s were measured.
The Kolmogorov–Smirnov test was performed and the measurement and expectation were consistent within a statistical error (\rev{P-value is 21\% and 38\%} for the horizontal and vertical direction, respectively)~\cite{bib:Otani18muaccbpm}.
\par
The BWM employed a MCP assembly (Hamamatsu F1217-11G) containing two stages of chevron-type MCPs.
The signal-processing electronics of constant-fraction discriminators~\cite{bib:Les75},
which were adopted from the technology of the other experiment~\cite{bib:Inami14}, measured the timing of the MCP signal.
It detects single muon with a high temporal resolution, accumulates muon events, and reconstructs the beam bunch
by taking the difference from the RF reference time of the muon linac.
A test bench was developed to evaluate the temporal resolution of the BWM using a picosecond pulse laser (Hamamatsu PLP10-040)~\cite{bib:yotsuzuka19}.
Recent results have evaluated the resolution of the BWM to be 40 ps, which corresponds to 1\% of the acceleration frequency of 324 MHz and satisfies the requirement.
During the \rev{2018A0222} experiment, the detector in the setup shown in Fig.~\ref{fig:muaccsetup} is replaced with the BWM to measure the bunch width of the accelerated $\mathrm{Mu^{-}}$'s.
In addition, a buncher cavity developed as a prototype for the muon linac~\cite{bib:Otani19buncher} was installed between QM2 and BM to focus the beam in the longitudinal direction.
The bunch width was successfully measured to be $0.54\pm0.11$~ns, which is consistent with the simulation~\cite{bib:sue20}.
\par
From these measurements, the beam monitors are ready to diagnose the accelerated muon beam, particular for low energy part.
Because these monitors are based on an MCP and the detection efficiency of an MCP is expected to decrease as the muon energy increases,
to conduct sufficient beam commissioning, it is necessary to develop a beam monitor for the high-energy part.
\section{Summary and outlook}\label{sec:sum}
\rev{
The design of the muon linac for the J-PARC E34 experiment and the first muon acceleration were presented.
The experiment for demonstrating the acceleration of the thermal muons is scheduled in 2022
at a new muon experiment area (S2) in J-PARC MLF.
The experiment aims to study the laser ionization of the muonium via the 1S-2S excitation with 244-nm laser~\cite{bib:Ced21} and
to accelerate the thermal muons using the RFQ.
The muon beam emittance before and after the acceleration will be measured using
the beam monitors described in \revr{Section}~\ref{sec:monitor}.
After the experiment at S2, the muon acceleration using the RFQ and IH-DTL at the H-line
will be conducted.
}
\par
We are entering a new era in which the accelerated muons are available.
One can naturally imagine an imaging technique with accelerated muons, allowing for a better resolution with less time than cosmic-ray muons.
It is natural to think of a collider using an accelerated muon beam.
A type of collider in which muons collide with other particles has been actively discussed~\cite{bib:Men20, bib:Kin21}.
If a cooling method of negative muons with an efficiency comparable to that of the muonium laser ionization is realized,
a muon collider will be promised.
\par
The J-PARC E34 experiment will soon be realized as a flagship to new horizon for a better understanding of nature.
\section*{Acknowledgment}
We express our appreciation to the many manufacturing companies
involved in this project, especially to Toshiba Co., who fabricated
the \rev{prototype} RFQ, \rev{Toyama Co.,Ltd., who fabricated the actual RFQ}, and
\rev{TIME Co., Ltd.}, who fabricated the
buncher cavity and the prototype IH-DTL.
This work is supported by JSPS KAKENHI (Grant Numbers
, 2580016
, 15H0366
, 15H0574
, 16H0398
, 16J0778
, 18H0370
, 18J2212
, 19J2176
, 20J2144
, 20H0562
, 21K1863
, 21H0508
, 22H0014
)
, JST FOREST Program (Grant Number JPMJFR212O)
and the natural science grant of the Mitsubishi Foundation.
This work is also supported by
the Korean National Research Foundation grants
NRF-2015H1A2A1030275,
NRF-2015K2A2A4000092,
and NRF-2017R1A2B3007018;
the Russian Foundation for Basic Research
grant RFBR 17-52-50064;
and the Russian Science Foundation grant
RNF 17-12-01036.
This experiment at the Materials and Life Science Experimental Facility
of the J-PARC was performed under user programs (Proposal No. 2015A0324, 2016A0067, 2016B0214, 2017A0263, 2017B0006, 2018A0222, and 2018B0007).
\let\doi\relax
|
\section{INTRODUCTION}
Scanning tunneling microscopy (STM) and atomic force microscopy (AFM) are powerful analytical techniques that probe surface phenomena down to the atomic scale for a wide range of materials and experimental conditions\cite{Wiesendanger}. As such, STM and AFM experiments are often performed at cryogenic temperatures to shut down all thermally activated processes—a condition that minimizes thermal drift, providing stability to the probe-sample junction; molecular and atomic migrations, providing stability both to the probe apex and to the sample surface; and thermal broadening, enhancing both spatial and energy resolutions. In STM, the low-temperature conditions are critical for tunneling spectroscopy measurements, making it possible to explore the physical and chemical phenomena with high energy resolution; in non-contact AFM, the thermal frequency noise that fluctuates the frequency shifts are substantially reduced. Owing to these advantages, many atomic-scale phenomena have been investigated using various scanning probe techniques at low temperatures\cite{Wiesendanger, Morita}.
In the past decades, the severe shortage of helium\cite{Nuttall} has made it increasingly difficult to conduct low-temperature scanning probe microscopy (SPM) experiments. While the use of low helium boil-off cryostats\cite{Okamoto, Trofimov, Unisoku} can alleviate the problem, the ultimate solution lies in the integration of a cryogen-free refrigerator into the SPM system itself. Such integration would not only eliminate the dependency on scarce helium, but it would also enable experiments to be carried out uninterrupted without the need for cryogen refills. To this end, several SPM systems that have integrated cryogen-free refrigerators using elaborate mechanical vibration isolations have been reported\cite{denHaan, Hackley, Zhang, Pabbi, Meng, Chaudhary} and made commercially available\cite{Omicron, RHK}. However, realizing an instrument with both a base temperature close to 4.2 ~K, and high-resolution measurement capabilities similar to liquid-helium-based SPM systems has remained elusive.
In this paper, we present a cryogen-free ultrahigh vacuum (UHV) SPM system capable of operating at the base temperature of near 5 K using an integrated pulse-tube refrigerator (PTR). We achieve this by combining a home-made interface between the PTR cold head and the SPM head with a new automatic gas-handling system. The home-made interface is designed to suppress resonant coupling of the mechanical vibrations from the PTR to the SPM. The gas-handling system automatically regulates the pressure of helium heat-exchange gas to cool the cryostat down to and keep the cryostat at the base temperature. We discuss the effects of vibration isolation using systematic vibration noise measurements. We demonstrate the performance of the cryogen-free SPM by measuring the superconducting gap of a Pb surface, by performing STM experiments on a Au(111) surface, and by performing non-contact AFM experiments on a NaCl(100) surface. We show that both the base temperature and the data quality from our cryogen-free SPM system are comparable to conventional liquid-helium-based SPM systems.
\section{SYSTEM OVERVIEW}
\begin{figure*}
\includegraphics[width=16cm]{Fig1.jpg}
\caption{\label{SystemOverview}Schematic of the entire cryogen-free low-temperature SPM system. The letters A–D denote the locations where vibration noise spectra are measured.}
\end{figure*}
Figure~\ref{SystemOverview} shows the schematic of the cryogen-free low-temperature SPM, located at the first floor of the factory in UNISOKU Co., Ltd. amongst several other SPM systems. The SPM system in the figure is not isolated by sound-proof walls, and is composed of three UHV chambers: a load-lock chamber for tip and sample exchange, a tip and sample preparation chamber, and a SPM observation chamber. The three chambers are mounted on a passive vibration isolation table (Stable 66-1209-8829-CE, Kurashiki Kako, Japan). The preparation and observation chambers are pumped by two ion pumps (240 L/s and 125 L/s), each equipped with a titanium sublimation pump. The base pressure of the observation chamber is $2\times10^{-8}$~Pa at room temperature and $6\times10^{-9}$~Pa when the SPM head is cooled down.
Next, we discuss the coupling of the PTR to the SPM observation chamber. The PTR (RP-082B2S, Sumitomo Heavy Industries, Japan) consists of a water-cooled compressor (F-70LP, Sumitomo Heavy Industries), a valve unit, and a cold head attached to two cooling stages (referred to as 1st and 2nd cooling stages in Fig. 2). The cooling power of the PTR is 35~W at 45~K in the 1st cooling stage and 0.9~W at 4.2~K in the 2nd cooling stage. The power line frequency of 60 Hz is used in this study. The compressor is encased in a home-made acoustic isolation box and is located $\sim 7$~m away from the SPM system for sound isolation. The length of the flexible tubes interconnecting the compressor and the valve unit is 20~m. Ideally, the compressor should be located in a separate room for optimum sound isolation. The valve unit is mounted on a purposefully rigid and heavy supporting frame, and connected to the PTR cold head via two stainless-steel tubes and one flexible tube. The two stainless-steel tubes are fixed together to reduce vibration transmission from the valve unit to the cold head. The valve unit, the cold head, and the tubes are encased in an acoustic isolation box.
\begin{figure}
\includegraphics[width=8cm]{Fig2.jpg}
\caption{\label{Interface} Schematic of the interface between the PTR and the SPM head. The area enclosed by the cold head, the PTFE bellows, and the cryostat is filled with helium that serves as heat-exchange gas between the cooling stages and the bottom of the cryostat. The observation chamber is kept in UHV condition. The letters B–D denote the locations where vibration noise spectra are measured.}
\end{figure}
The cold head is attached at the top of the observation chamber using polytetrafluoroethylene (PTFE) bellows to suppress vibration transmissions to the chamber (Fig.~\ref{Interface}). The top of the observation chamber is mounted on its own supporting frame separated from the vibration isolation table (Fig.~\ref{SystemOverview}). To further suppress the vibration transmission from the cold head, stainless-steel bellows are placed between the top and bottom part of the observation chamber (Fig.~\ref{Interface}).
The space enclosed by the cold head, the PTFE bellows, and the cryostat (area indicated by the arrow labeled "He heat-exchange gas" in Fig.~\ref{Interface}) houses the 1st and 2nd cooling stages of the PTR, and is filled with helium gas for heat exchange between the cooling stages and the cryostat. The pressure of helium heat-exchange gas during cooling is automatically controlled by the gas-handling system which will be described in detail later. We note that this design is suitable for the baking of the observation chamber because components that are not heat-resistant (i.e., the cold head, the two cooling stages, and the PTFE bellows and O-rings) can be easily removed.
Next, we discuss the setup of the SPM head. The SPM head is surrounded by two layers of thermal radiation shields (Fig.~\ref{Interface}). The outer aluminum (purity $\geq 99.7$~\%) shield is attached to the cryostat near the 1st cooling stage, and the inner oxygen-free copper shield is directly attached to the bottom of the cryostat. The outside surface of the inner shield is gold-plated. The outer radiation shield has four windows and the inner radiation shield has three through-holes and a window aligned with the UHV chamber flanges near the SPM head. Two pairs of outer window and inner through-hole are aligned for the observation of the tip apex and sample surface. A pair of outer window and inner window is aligned for
transferring the tip and sample to and from the SPM head. Another pair of outer window and inner through-hole is aligned for locking the SPM head during the tip/sample transfer. Each window can be closed by its own shutter.
\begin{figure}
\includegraphics[width=8.5cm]{Fig3.jpg}
\caption{\label{SPMhead}(a) Three-dimensional illustration of the SPM unit. An eddy current damper is used between the SPM stage and the 5~K stage for vibration isolation. The copper heat links between SPM and 5~K stages and electrical wirings are not shown for clarity. (b) Cross-section of the SPM head. A tip holder for STM is shown in both figures. The tip and sample are aligned horizontally in the x direction.}
\end{figure}
The SPM head is mounted on two cold stages. Figure 2 shows both the 5~K and the 60~K stages that are connected to the inner and to the outer radiation shield via copper heat links, respectively. The copper heat links not only cool down the cold stages, but they also dampen vibration noise from the cold head. The copper heat links connecting the outer shield and the 60~K stage are composed of twenty-four copper cables each braded from 3,200 copper wires (diameter $\phi = 0.05 $~mm, and length $L = 55$~mm for each wire). The total cross-sectional area of the heat links is $\sim 150~\rm{mm}^2$. The copper heat links connecting the inner shield and the 5~K stage are composed of twenty-four copper cables each braded from 1,600 copper wires ($\phi = 0.05$~mm, $L = 45$~mm for each wire). The total cross-sectional area of the heat links is $\sim 75~\rm{mm}^2$.
The 5 K and 60 K stages are supported by a carbon and an alumina fiber-reinforced plastic (FRP) cylinder, respectively (Fig.~\ref{Interface}). We choose these materials for their low thermal conductivities and large Young’s moduli, which promote the efficient cooling and enhance both the thermal and the vibration isolations of the SPM head. The alumina FRP has been used as supporting rods in the ultra-low-vibration cryocooler system for the gravitational wave detector\cite{Ikushima}. The carbon FRP is known to exhibit sufficiently low thermal conductivity at low temperatures\cite{Takeno} with larger Young’s modulus than the alumina FRP.
The SPM head is further vibrationally isolated by an eddy current damper using four stainless-steel springs and eight samarium-cobalt (SmCo) magnets as shown in Fig.~\ref{SPMhead}(a). Sixteen stranded copper cables consisting of $\sim$~380 copper wires with a diameter $\phi = 0.05$~mm and the length $L = 40$~mm are used as heat links between the 5~K stage and the SPM stage (total cross-sectional area is $\sim 12~\rm{mm}^2$).
The SPM head is built according to the UNISOKU USM1300 system, with some component modifications to increase rigidity. The sample XY stage and the prism in the Z stage for tip coarse approach, both moved by stick and slip motion of shear piezo stacks, are made of sapphire, also for rigidity [Fig.~\ref{SPMhead}(b)]. To enhance the SPM head resonance frequencies even when a heavy tip holder ($\sim 1.8$~g) is mounted on top of the piezo tube for scanning, we use the double piezo tube design, where both the inner and outer tubes are used for XY scanning, and the outer tube is used for the feedback Z motion [Fig.~\ref{SPMhead}(b)]. The configuration of the electrodes in the piezo tubes is shown in Appendix A. The maximum scan range is $1.5~\rm{\mu m} \times 1.5~\rm{\mu m}$ at low temperatures. The lowest resonance frequencies at the Z stage and at the XY stage are measured to be $\sim 4$~kHz and $\sim 3.8$~kHz, respectively. The tip and sample are aligned horizontally in the x direction, as defined in Fig.~\ref{SystemOverview}.
We use home-made vacuum-insulated cables as reported by Mykkänen $\it{et~al}$.\cite{Mykkanen} ($L = 17$~cm) to carry the tunneling-current and the AFM signals at the 5~K stage. The vacuum-insulated cable consists of a stainless tube ($\phi = 2.8$~mm, thickness of 0.18~mm) and a CuNi wire ($\phi = 0.1$~mm) and is terminated with SMPM connectors. The total length of the coaxial cable between the tip and the input of the preamplifier located outside the observation chamber is $\sim 50$~cm.
\section{HELIUM GAS-HANDLING SYSTEM}
\begin{figure*}
\includegraphics[width=16cm]{Fig4.jpg}
\caption{\label{GasHandlingSystem} (a) Schematic diagram of the automatic gas-handling system. (b) Photograph of the control panel. (c) Screenshot of the touch panel.}
\end{figure*}
The regulation of the helium heat-exchange gas pressure in the cryostat is important to both the cooling down to and the keeping of a stable base temperature. To this end, we have developed an automatic bimodal helium gas-handling system based on the programmable logic controller [Figs.~\ref{GasHandlingSystem}(a) and (b)]. First, to cool down the cryostat from room temperature to the base temperature of $\sim 5$~K, we need to add the adequate amount of helium heat-exchange gas in the cryostat. The program to cool down from room temperature to the base temperature is referred to as “cool-down mode” hereafter. Second, after cooling down, the helium heat-exchange gas needs to be occasionally refilled to keep the base temperature because the SPM head warms up during tip/sample transfers, expanding and releasing helium gas through the cryostat relief valve. This second program mode is referred to as “maintenance mode” hereafter. The gas-handling system developed in this study enables us to cool the cryostat down to the base temperature and keep the base temperature stable just by pressing a start button on the touch-screen control panel [Fig.~\ref{GasHandlingSystem}(c)].
\begin{figure}
\includegraphics[width=8.5cm]{Fig5.jpg}
\caption{\label{CoolingMode}(a) Flowchart describing the cool-down mode from room temperature to 5~K. (b) Flowchart describing the maintenance mode to keep the base temperature of $\sim 5$~K at the 2nd cooling stage of the PTR.}
\end{figure}
Figure~\ref{CoolingMode}(a) describes each process executed during the cool-down mode. The program starts by flushing and filling all plumbing lines and the cryostat with helium gas. After the pressure in the cryostat ($P_{\rm{cryo}}$) reaches atmospheric pressure, the PTR is turned on. In this study, we describe pressure in relative terms, defining the latter condition of 1~atm of helium as $P_{\rm{cryo}}=0~\rm{kPa}$.
As $P_{\rm{cryo}}$ drops with temperature, helium gas must be added to the cryostat to prevent the cooling rate from slowing down. The threshold pressure that triggers the helium-gas addition is set to the empirically chosen 10~kPa lower than atmospheric pressure ($P_{\rm{cryo}} < -10~\rm{kPa}$). When the condition $P_{\rm{cryo}} < -10~\rm{kPa}$ is met, the program opens the solenoid valves in Fig.~\ref{GasHandlingSystem}(a) and fills the cryostat with helium gas up to $P_{\rm{cryo}}=0~\rm{kPa}$. This process repeats until the temperature at the 2nd cooling stage ($T_{\rm{2nd}}$) drops down below 5 K. The end of the cool-down mode is reached
when both conditions $T_{\rm{2nd}} < 5~\rm{K}$ and $P_{\rm{cryo}} \leq 0~\rm{kPa}$ are met.
After completion of the cool-down mode, the program switches to the maintenance mode to keep $T_{\rm{2nd}} \leq 5~\rm{K}$. Figure~\ref{CoolingMode}(b) describes how the program evaluates when to start/stop the addition of helium gas into the cryostat. At $T_{\rm{2nd}} \leq 5~\rm{K}$, helium gas is added when $P_{\rm{cryo}}$ is both lower than the vapor pressure of helium at $T_{\rm{2nd}}$ [$P_{\rm{cryo}} < P_{\rm{vapor}}(T_{\rm{2nd}})$], and lower than the threshold pressure of $-10$~kPa ($P_{\rm{cryo}} < -10~\rm{kPa}$). At $T_{\rm{2nd}} > 5~\rm{K}$, only the latter condition ($P_{\rm{cryo}} < -10~\rm{kPa}$) is used to decide whether helium gas needs to be added or not because the vapor pressure of helium is calculated only at $T_{\rm{2nd}} \leq 5~\rm{K}$.
Upon starting the helium-gas addition, the program evaluates when to stop the addition. For this purpose, we set three following conditions: $P_{\rm{cryo}} > +8~\rm{kPa}$, $T_{\rm{2nd}} > 30~\rm{K}$, and $P_{\rm{cryo}} > P_{\rm{vapor}}(T_{\rm{2nd}})$ [Fig.~\ref{CoolingMode}(b)]. When any one of these three conditions is met, the solenoid valves immediately close to stop the helium-gas addition. The first two conditions are designed to be met by either the $P_{\rm{cryo}}$ and $T_{\rm{2nd}}$ surges during the sample/tip transfers. The threshold values of 8~kPa and 30~K are empirically chosen. The third condition $P_{\rm{cryo}} > P_{\rm{vapor}}(T_{\rm{2nd}})$ is designed to be met during a long-term SPM measurement when $T_{\rm{2nd}} \leq 5~\rm{K}$ occurs. Upon closing the solenoid valves to stop adding helium gas, the program waits for one minute before repeating the entire procedure.
\section{COOLING PERFORMANCE}
\begin{figure}
\includegraphics[width=8.5cm]{Fig6.jpg}
\caption{\label{CoolingTest}(a) (Top) Typical cool-down temperature curves measured at the 2nd cooling stage, 60~K stage, 5~K stage and SPM head shown in Fig.~\ref{Interface}. (Bottom) Corresponding cryostat pressure controlled by the gas-handling system. The atmospheric pressure corresponds to 0~kPa. (b) Histogram of the SPM head temperature controlled to stabilize at 6~K for three days.}
\end{figure}
Figure~\ref{CoolingTest}(a) shows the typical temperature curves at the 2nd cooling stage, 60~K stage, 5~K stage and SPM head (top panel), as well as the corresponding $P_{\rm{cryo}}$ curve (bottom panel) during the cool-down and the maintenance modes. It takes $\sim 9$~hours to cool the 2nd cooling stage from room temperature down to $\sim 5$~K. While the temperatures at the 60~K and the 5~K stages are sensitive to the variations in $P_{\rm{cryo}}$, the temperature at the SPM head is not. After twelve hours, the base temperatures are reached at the 2nd cooling stage, 60~K stage, 5~K stage and SPM head at 3.9~K, 63~K, 5.2~K and 5.5~K, respectively. The small temperature difference between the 5~K stage and the SPM head (0.3~K) indicates good thermal contact by the heat links. After reaching the base temperature, the helium-gas addition to the cryostat becomes a rare event that only occurs during tip/sample transfers.
The base temperature fluctuates over a 24-hr period depending on room temperature, typically in the range of 100~mK at the SPM head. We minimize this temperature fluctuation using a feedback-controlled heater attached to the SPM stage. As shown in Fig.~\ref{CoolingTest}(b), the temperature fluctuation at 6~K for three days is reduced to within $\pm1~\rm{mK}$, which is stable for long-term SPM measurements.
\section{VIBRATION MEASUREMENTS}
The interface structure connecting the PTR and the SPM head in Fig.~\ref{Interface} indicates two routes that can transmit vibrations from the cold head to the SPM head. The first route transmits through the PTFE bellows, the cryostat, the inner and outer shields, and the heat links between the radiation shields and the cold stages. For this route, both the PTFE bellows and the heat links are the key components that suppress vibration transmission. The second route transmits through the PTFE bellows, the stainless-steel bellows, the vibration isolation table, and the two FRP cylinders. For this route, since the SPM head is placed on the vibration isolation table via the 5 K and 60 K stages, suppressing vibration transmission using the vibration isolation table is critical.
\begin{figure}
\includegraphics[width=8.5cm]{Fig7.jpg}
\caption{\label{VibrationNoise} (a-c) Vibration spectra at valve unit (A), cold head (B), frame (C) and vibration isolation table (D) in the x, y and z directions, respectively. Refer to Fig.~\ref{SystemOverview} for the depictions of letters A–D. All measurements are acquired both the PTR and the passive vibration isolation table in operation. The scale in y axis is the same in all figures. }
\end{figure}
To analyze the vibration transmissions, we measure the vibration velocity spectra along the x, y and z direction on the valve unit (A), the cold head (B), the supporting frame (C) and the vibration isolation table (D) during PTR operation (see Fig.~\ref{SystemOverview} for locations of A–D). The vibration spectra are measured using a commercial accelerometer (NP-7310, Onosokki, Japan), its amplifier (PS-1300, Onosokki, Japan), and a Nanonis SPM controller (SPECS GmbH, Germany).
\begin{figure}
\includegraphics[width=8.5cm]{Fig8.jpg}
\caption{\label{VibrationNoise2} (a-c) Effects of the PTR on velocity spectra at the passive vibration isolation table (D in Fig.~\ref{SystemOverview}) in the x, y and z directions, respectively. For comparison, the velocity spectra at the passive vibration isolation table in a conventional SPM system with a liquid helium cryostat (UNISOKU USM1300), placed next to the PTR-based SPM, are plotted. The scale in y axis is the same in all figures.}
\end{figure}
We first compare the vibration spectra acquired at the valve unit and the cold head (light-solid and dotted lines in Fig.~\ref{VibrationNoise}, respectively). The noise spectra at the valve unit exhibit similar series of peaks in all three directions. The peaks at 2~Hz observed in all three directions correspond to the fundamental frequency of the PTR. The series of peaks at higher frequencies are its higher harmonics. Since the cold head and the valve unit are mechanically connected along the x direction (Fig.~\ref{SystemOverview}), the peak at 2~Hz in the x direction is most striking. In addition, several broad peak structures at 5, 20, and between 30 and 50~Hz are observed in all three directions.
At the cold head, a series of peaks between 30 and 90~Hz and the broad peaks between 30 and 50~Hz are effectively reduced in all directions. This is because the two stainless tubes connecting the valve unit and the cold head are rigidly fixed together. In contrast, several peaks are enhanced or newly appear at the cold head, including the peak at 2~Hz and the broad peak near 20~Hz in the y and z directions, and a broad peak structure near 10~Hz in the x and y directions. We also find that compared to the x and y directions, the vibration spectrum in the z direction at the cold head exhibits a qualitatively different feature, as represented by the broad peak near 20~Hz [Fig.~\ref{VibrationNoise}(c)]. We attribute these peaks to characteristic vibration modes of the cold head.
Next, we discuss the effects of the PTFE bellows by comparing the vibration spectra acquired at the supporting frame with those at the cold head (dashed and dotted lines in Fig.~\ref{VibrationNoise}, respectively). The PTFE bellows are clearly effective at suppressing the noise peak at 2~Hz in all directions; in particular, the vibration noise spectra at the frame show no peak at 2~Hz in the y and z directions [Fig.~\ref{VibrationNoise}(b, c)]. Similarly, other peaks at 5 Hz in the x and y directions, above 40~Hz in all directions, and the broad peak at 20~Hz in the z direction are all substantially reduced at the supporting frame.
Next, we compare the vibration spectra acquired at the vibration isolation table with those at the frame (dark-solid and dashed lines in Fig.~\ref{VibrationNoise}, respectively). While many peaks remain visible, the vibration noise levels above 4~Hz are substantially suppressed in all directions at the vibration isolation table. This suppression is mainly due to the stainless-steel bellows and the heavy vibration isolation table (Fig.~\ref{Interface}). The heat links and the FRP cylinders might also contribute to the noise suppression. Below 4~Hz, however, the noise levels are not improved; the noise peak at 2~Hz at the vibration isolation table in the x direction remains unchanged compared to that at the frame. A broad peak structure near 3~Hz appears in all directions. We attribute the enhancement of vibration near 3~Hz to the resonance vibration of the passive vibration isolation damper rather than the vibration transmission from the PTR.
We examine this effect by comparing noise spectra at the vibration isolation table with the PTR turned on and off (solid and dotted lines in Fig.~\ref{VibrationNoise2}, respectively), and find that the noise peaks at 2-3~Hz in all directions (except the sharp peak at 2~Hz in the x direction) remain almost unchanged before and after the PTR is turned on. This observation indicates that those noise peaks originate from the passive vibration isolation damper rather than from the PTR. A series of peaks between 10 and 70~Hz are possibly caused by noise from the PTR suggesting insufficient isolation by the acoustic isolation box.
We also compare the vibration noise spectra at the vibration isolation table with those of a liquid-helium-based SPM system (USM1300, UNISOKU) placed next to the cryogen-free SPM system (solid and dashed lines in Fig. 8, respectively). The noise floor of the cryogen-free SPM system is comparable to that of the liquid-helium-based SPM system and even lower above 50~Hz in all directions. We conclude that while many peaks remain observable (for instance, at 2~Hz in the x direction and a series of peaks between 10 and 70~ Hz in all directions), the vibrational-noise levels at the vibration isolation table in the cryogen-free SPM system are effectively suppressed for SPM measurements.
\section{TUNNELING-CURRENT NOISE MEASUREMENTS}
\begin{figure}
\includegraphics[width=8.5cm]{Fig9.jpg}
\caption{\label{CurrentNoise} The tunneling-current noise spectrum measured on a Au(111) surface at $T = 7.0$~K with an open feedback loop (tunneling conditions: sample bias voltage, $V = -100$~mV; tunneling current, $I_{\rm t} = 100$~pA). The sample temperature was controlled to stabilize at 7.0~K. The inset shows the tunneling current as a function of the tip-sample distance. The light-solid (dotted) curve shows the tunneling-current noise spectrum when the tip is withdrawn from the sample surface with the PTR turned on (off).}
\end{figure}
To demonstrate the effects of vibration isolation on practical SPM measurements, we measure the tunneling-current noise spectra using a commercial current preamplifier ($10^9$~V/A, DLPCA-200, Femto Messtechnik GmbH, Germany), and a Nanonis SPM controller.
Figure~\ref{CurrentNoise} compares the current noise spectra acquired with the tip withdrawn out of tunneling range from the sample surface with the PTR turned on and off (light-solid and dotted lines, respectively). The absence of any peak at 2~Hz during PTR operation indicates effective vibration isolation from the PTR. Although several peaks (e.g., at 30~Hz and 180~Hz) are observed, the peak heights are measurably low (below $10~\rm{fA}/\sqrt{\rm {Hz}}$). The noise floor remains low up to 1 kHz owing to the small capacitance afforded by the short length (50~cm) of the cable. This latter feature is advantageous for tunneling spectroscopy measurements using a lock-in amplifier.
We also show in Fig.~\ref{CurrentNoise} the tunneling-current noise spectrum acquired with a mechanically polished PtIr tip in tunneling range with an opened feedback loop over a Au(111)/mica surface prepared by repetitive Ar sputtering and annealing. The inset in Fig.~\ref{CurrentNoise} confirms the exponential decay of the tunneling current as a function of the tip-sample distance, ensuring an ideal tunneling condition.
In the tunneling condition, a broad peak appears near 2~Hz, originating from the vibrations from both the PTR and the vibration isolation table. Several peak structures near 10, 15 and 30~Hz are attributed to the vibration isolation table as identified in Fig.~\ref{VibrationNoise2}. Although these noise peaks below 40~Hz are observed, the noise floor is measurably low for higher frequencies between 40~ Hz and 1~kHz.
From the above discussions, we believe that a different relative configuration between the PTR and the SPM system may improve the noise level in the tunneling condition. In the current configuration shown in Fig.~\ref{SystemOverview}, the tip and sample are aligned in parallel with the cold head and valve unit in the x direction so that the tip-sample distance is most severely affected by the vibration noise from the PTR. Therefore, aligning the PTR in the y direction in Fig.~\ref{SystemOverview} while keeping the tip–sample direction unchanged could reduce vibrational coupling in the x direction. Another possibility would be to use an active vibration isolation table as reported in Ref.~\cite{Zhang}. Since the noise level on the active vibration isolation table in Ref.~\cite{Zhang} is substantially lower than in this study, its approach could further reduce noise peaks in the tunneling current.
\section{PERFORMANCE DEMONSTRATIONS}
\begin{figure}
\includegraphics[width=7.5cm]{Fig10.jpg}
\caption{\label{SCgap} A typical differential-conductance spectrum of Pb at the base temperature of 5.0~K, measured using the standard lock-in technique with an AC modulation amplitude $V_{\rm mod}$ of 212~$\mu V_{\rm rms}$ at 1 kHz (tunneling conditions: $I_{\rm t} = 500~{\rm pA}$, $V = +10~{\rm mV}$. The solid line shows the data fit using the Dynes function. Both thermal broadening and lock-in broadening are included.}
\end{figure}
First, we demonstrate the measurement of a Pb superconducting gap spectrum at $T = 5.0$~K and the evaluation of its effective electron temperature (Fig.~\ref{SCgap}). Mechanically polished PtIr tips are used for all STM measurements in this study. We use a polycrystalline Pb plate (purity: 99.9\%) purchased from Nilaco Co., Ltd. To avoid RF noise propagating to the SPM head, we use low-pass filters for thermometry (500~kHz cutoff), a heater (2~MHz cutoff), XY and Z stage lines (100~kHz cutoff). In the scanning line, we add a capacitor of $\sim 10$~nF between the ground at the feedthrough of the chamber. All lines unused for the measurement are grounded.
The superconducting gap spectrum is fitted with the Dynes function (see Appendix B about the detail), yielding the effective temperature of 5.7~K with the superconducting gap $\Delta = 1.29~{\rm meV}$ and a quasiparticle-lifetime broadening factor $\Gamma = 0.48~{\rm meV}$.
\begin{figure}
\includegraphics[width=8.5cm]{Fig11.jpg}
\caption{\label{STM} (a) Constant-current STM image of Au(111) on mica at $T = 5.6~{\rm K}$ (tunneling conditions: $I_{\rm t} = 1~{\rm nA}, V = +5~{\rm mV}$). (b) Non-contact AFM image of NaCl(100) at $T = 6.0~{\rm K}$ with a frequency shift $\Delta f = -13.2~{\rm Hz}$ at a resonance frequency $f_{0} = 29.29~{\rm kHz}$. $V = +1.7~{\rm V}$. The line profiles are shown under the corresponding images. The red line in each image indicates the location of the line profile. Both images are processed only by line subtraction.}
\end{figure}
Next, we demonstrate the constant-current STM imaging of a Au(111) surface, resolving its hexagonal atomic lattice along with the herringbone structure [Fig.~\ref{STM}(a)]. The atomic corrugation on a metallic surface such as Au(111) is considerably smaller compared to semiconducting and non-metallic surfaces so that a mechanically stable SPM is generally required to obtain atomic resolution. The line profile in Fig.~\ref{STM}(a) shows a well-defined atomic corrugation of $\sim 10$~pm without noticeable noise.
\begin{figure}
\includegraphics[width=8.5cm]{Fig12.jpg}
\caption{\label{QPI} (a) Constant-current STM image of Au(111) on mica at $T = 7.0~{\rm K}$ (tunneling conditions: $I_{\rm t} = 1~{\rm nA}, V = -600~{\rm mV}$). (b) d$I$/d$V$ image at $V = -10~{\rm mV}$ acquired at the same location as in (a). $I_{\rm t} = 1~{\rm nA}$. $V_{\rm mod} = 7~{\rm mV_{rms}}$ at 1 kHz. (c) Fourier-transformed image of (b). (d) Energy dispersion of the QPI signal in the direction indicated in (c). The solid line serves as an eye guide. }
\end{figure}
Next, we demonstrate the non-contact AFM measurements on a NaCl(100) surface using the qPlus sensor\cite{Giessibl} [Fig.~\ref{STM}(b)]. The NaCl(100) single crystal (Goodfellow) is glued onto a silicon substrate and cleaved in air. We anneal the sample in UHV at about $100~^\circ$C for 15~min by a direct-current heating of the silicon substrate to avoid charging the surface. We use an electrochemically etched tungsten tip glued onto the free prong of the qPlus sensor. We mechanically excite the sensor using the inner piezo tube and the resonance frequency of the sensor is 29.29~kHz with a Q-factor of 60,000 at 6.0~K. The signal is amplified using a commercial charge amplifier (HQA-15M-10T, FEMTO Messtechnik) placed outside the observation chamber. For imaging, the tip-sample distance is controlled with a feedback loop to maintain a constant frequency shift $\Delta f = -13.2~{\rm Hz}$ while the tip is driven at a constant oscillation amplitude of 200~pm. The line profile in Fig.~\ref{STM}(b) shows a typical corrugation of approximately 10~pm, comparable to the spatial resolution of known low-temperature AFM systems cooled by liquid helium\cite{Bettac}.
Finally, we demonstrate the quasiparticle interference (QPI) imaging of a Au(111) surface. Figures~\ref{QPI}(a) and (b) show the topographic and the differential-conductance (d$I$/d$V$) images acquired simultaneously on Au(111), respectively. The interference pattern caused by electrons scattering at surface defects are clearly observed in Fig.~\ref{QPI}(b). The d$I$/d$V$ image in Fig.~\ref{QPI}(b) is acquired using a lock-in amplifier at $\sim13$~min imaging time. We use Fourier analysis to identify QPI signals appearing at scattering wave $q$ vectors. The Fourier-transformed d$I$/d$V$ image in Fig.~\ref{QPI}(c) exhibits a ring-like feature around $q = 0$, reflecting the isotropic backscattering process within the ring-like constant energy contour of the parabolic surface state centered at the $\Gamma$ point\cite{Petersen}. By varying the bias voltage, the QPI signal shows a parabolic energy dispersion [Fig.~\ref{QPI}(d)], as expected for the two-dimensional surface state. The dispersion is in good agreement with previously reported data\cite{Hoesch, Tesch}. We note that thirteen d$I$/d$V$ images are used to compile the data shown in Fig.~\ref{QPI}(d) which requires a minimum of 3 hours uninterrupted acquisition time. Without the need for cryogen refills, the current SPM system can, in principle, provide uninterrupted measurement time indefinitely.
\section{CONCLUSION}
This work reports the design of a cryogen-free, PTR-based UHV SPM system capable of operating at the base temperature of $\sim 5$~K. We achieve this by combining a home-made interface design between the PTR cold head and the SPM head with an automatic gas-handling system: while the interface design enhances both vibration isolation and cooling capability, the gas-handling system facilitates both the cool down to and the maintenance of the base temperature. Our report provides both the detailed information on the vibration isolation structures, and the corresponding vibration noise measurements. Our SPM data show comparable quality with those acquired by conventional cryogen-cooled instruments. Our work demonstrates the possibility for achieving delicate low-temperature measurements without cryogenic liquids.
\begin{acknowledgments}
We are grateful to Yo Miyazaki, Tetsuya Maeda, Nozomi Nishiyama, Yutaka Miyatake (UNISOKU) and Tokkyokiki Corporation for their technical assistance. We would also like to thank Tadashi Machida (RIKEN) for helpful discussion. This study was partly supported by the Small and Medium Enterprise Agency (Grant No.~2527213203).
\end{acknowledgments}
\section*{AUTHOR DECLARATIONS}
\subsection*{Conflict of Interest}
The authors have no conflicts to disclose.
\section*{DATA AVAILABILITY}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
|
\section{Introduction}
\subsection{Background}
In recent ten years since Kaul \emph{et al.} proposed a framework to quantify the timeliness of information in 2012 \cite{Zerowait2}, one of the most popular ideas in timely update system design has been how to keep information as fresh as possible and ensure timely information delivery. For timely update systems such as vehicle networks where the vehicle's velocity and location are disseminated to ensure safe transportation \cite{vehicle}, environmental sensor networks where the updates of a time-varying phenomenon are collected for large-scale monitoring \cite{sensor}, and wireless communication networks where adaptive scheduling algorithms are adopted based on the time-varying channel state information \cite{channel}, achieving timely delivery can freshen the monitor’s awareness of the sources and thus assist correct and efficient decision making.
This has aroused new interest in analyzing \emph{Age of Information (AoI)} performance metrics. AoI has been broadly used to capture the \emph{freshness} of a monitor's knowledge of an entity or process. Different from conventional performance metrics such as delay and throughput, AoI comprehensively measures the effects of update rate, latency, and system utilization. Initial works on this issue were mainly based on queue analysis, which originated from single-source single-server queues \cite{packetmagnament1,packetmagnament2,Zerowait,Zerowait2}, and subsequently developed to multiple-sources single-server queues \cite{ISITMSDS,ISITMSDS2,ISITMSDS3,ISITMSDS4} and wireless queuing networks \cite{ISITNW1,ISITNW2,ISITNW3,ISITNW4,ISITNW5,ISITNW6}. These works are based on an ideal assumption that the status update is transmitted through a perfect channel without packet errors and losses. In practice, however, packet errors and losses are inevitable due to ubiquitous noises, signal interference, and channel fading. As the incorrectly decoded message does not bring about \emph{fresh} awareness, the packet errors and losses will result in \emph{staleness} of information, leading to uncontrollable residual errors, system instability, and wrong decisions. Therefore, it is imperative to analyze the AoI over unreliable channels.
\subsection{Related Works}
Some recent works have noticed the above limitation and have extended the AoI analyses to the physical (PHY) layer. One pioneering work concerning this issue was accomplished by Chen, \emph{et al.} in 2016 \cite{Channelcoding3}, in which the update is delivered over an erasure channel and the Peak Age of Information (PAoI) is studied. This work has aroused extensive research interest in understanding the effect of system reliability on AoI. From then on, including but not limited to the follow-up works that also analyzed AoI over the erasure channel \cite{Channelcoding4,Channelcoding5,Blockoptim}, various transmission protocols, ranging from conventional protocols like non-ARQ, classical ARQ and truncated ARQ protocols, to state-of-the-art protocols such as HARQ with Chase Combing (HARQ-CC) and HARQ with Incremental Redundancy (HARQ-IR) protocols, have been investigated under different types of noisy channels \cite{Age_LDPC,Channelcoding2,Channelcoding1,FBL_AoI,ShortAoI_2}.
We notice that the above AoI analyses focus on the transmission delay, and neglect other types of system delay such as coding delay, propagation delay, decoding delay and feedback delay. An exception work is \cite{Channelcoding_delay1}, which considers non-trivial propagation delay and studies the AoI of HARQ-IR with a fixed number of retransmitted packets $m=2$ under Satellite-IoT Systems, but also assumes negligible coding delay and decoding delay. Nevertheless, in practical communication systems, especially the short-packet communication, the coding delay and decoding delay are also nontrivial compared to the transmission delay, resulting in the \emph{staleness} of information by nature. Thus, we focus on a more realistic (or general) scenario where different types of delay elements naturally exist and the number of retransmitted packets is not fixed to $m=2$. In this regard, we would like to provide a basic framework to comprehensively study the trade-off among coding complexity, decoding complexity, code length, number of retransmitted packets and error probability from the AoI perspective.
Up to this point, we have only introduced AoI research based on conventional \emph{reactive} HARQ (also known as stop-and-wait HARQ), which allows for retransmissions only upon the reception of a Negative ACKnowledgment (NACK). As such, the retransmission process is not truly automatic. In the Release-16 5G NR specifications by the $3^{rd}$ Generation Partnership Project (3GPP), a new HARQ protocol named \emph{proactive} HARQ is designated for the Up-Link Grant-Free communication to enable the potential for meeting the stringent requirements for URLLC \cite{Proactive2}. Some recent works have shown the superiority of \emph{proactive} HARQ in terms of latency and throughput compared to \emph{reactive} HARQ \cite{GFHARQ1,GFHARQ2,GFHARQ3}. Inspiringly, these available studies also witness the potential for \emph{proactive} HARQ to be applied in the freshness-critical status update system. To this end, we would like to theoretically analyze the AoI performance of \emph{reactive} HARQ and \emph{proactive} HARQ to investigate whether \emph{proactive} HARQ will facilitate timeliness of information in the freshness-critical status update system.
\begin{table*}[t]
\normalsize
\centering
\renewcommand\arraystretch{1.5}
\caption{Construsting The Novelty Of Our Work To The Literature}
\setlength{\tabcolsep}{0.5mm}{
\begin{tabular}{c|c|c|c|c|c|c}
\hline
Contributions & {\bfseries This Work} & \cite{Zerowait2,vehicle,sensor,channel,packetmagnament1,packetmagnament2,Zerowait,ISITMSDS,ISITMSDS2,ISITMSDS3,ISITMSDS4,ISITNW1,ISITNW2,ISITNW3,ISITNW4,ISITNW5,ISITNW6} & \cite{Channelcoding3,Channelcoding4,Channelcoding5,Age_LDPC,Channelcoding2} & \cite{FBL_AoI,ShortAoI_2} & \cite{Blockoptim,Channelcoding1} & \cite{Channelcoding_delay1} \\ \hline \hline
Age of Information (AoI) & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ \\ \hline
Finite Block-Length Regime & $\checkmark$ & ~ & ~ & $\checkmark$ & ~ & ~ \\ \hline
Reactive Hybrid-ARQ & $\checkmark$ & ~ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ \\ \hline
Flexible Number of Retransmissions& $\checkmark$ & ~ & ~ & ~ & $\checkmark$ & ~ \\ \hline
\tabincell{c}{Effect of Delay Elements \\Other Than Transmission Delay} & $\checkmark$ & ~ & ~ & ~ & ~ & $\checkmark$ \\ \hline
Proactive Hybrid-ARQ & $\checkmark$ & ~ & ~ & ~ & ~ & ~ \\ \hline
\end{tabular}}
\end{table*}
\subsection{Contributions}
The research on the HARQ-based timely status update system is still in the ascendant, and some open issues remain to be addressed. First, there have been a lot of works providing explicit average age results under different types of protocols and systems. Examples include the average AoI expressions under fixed-length non-ARQ protocols, truncated-ARQ, classical ARQ, and the explicit results of some advanced ARQ-based techniques like HARQ-CC and HARQ-IR. However, there has not been a unified expression that can unify the aforementioned expressions in a single closed-form formula. By providing such a unified result, the comparative insights and the intrinsic relationships among different protocols will be further investigated. Second, the existing literature only considers certain types of delay in the status update system and assumes others to be negligible. However, as the delay exists by nature and plays as a critical part in affecting the \emph{freshness} of information, to comprehensively consider the coding delay (or processing delay), propagation delay, transmission delay, decoding delay and feedback delay in the status update system and provide a unified closed-form result will provide a systematic understanding in analyzing the age of a realistic freshness-critical status update system. Third, the age performance has been extensively studied over erasure channels. However, little research considers the short-packet AoI analysis over the AWGN channel. Finally, the majority of existing works mainly focus on AoI analysis of conventional \emph{reactive} HARQ. Some recent works analyzing the performance of \emph{proactive} HARQ are based on some conventional performance metrics, such as throughput and latency. Thus, to analyze the AoI of \emph{proactive} HARQ will fill this research gap and may further facilitate new system advancement of a status update system.
Motivated by the above, this work achieves several key contributions and we summarize them as follows:
\begin{itemize}
\item We derive unified closed-form average AoI and average Peak AoI expressions for \emph{reactive} HARQ, wherein: $i$) different kinds of delay elements (i.e., coding delay, transmission delay, propagation delay, decoding delay, and feedback delay) are comprehensively considered; $ii$) the number of repeated packets is not fixed as $m=2$, but is relaxed to a variable value; $iii$) different types of protocols are unified to a single expression.
\item We investigate the AoI explicit expressions and comparative insights for \emph{proactive} HARQ, which is the first work analyzing \emph{proactive} HARQ from the AoI perspective. Theoretical and numerical comparisons are given to show the superiority of \emph{proactive} HARQ in enabling timely information delivery.
\item We also try to further optimize the AoI for both \emph{reactive} HARQ and \emph{proactive} HARQ. By formulating an AoI minimization problem in the FBL regime, we solve out the age-optimal block assignment strategy for \emph{reactive} HARQ and \emph{proactive} HARQ, respectively. The results show that the optimal strategy for \emph{proactive} HARQ turns out to be the finest grained symbol-by-symbol transmission, while that for \emph{reactive} HARQ is highly dependent on the propagation delay and SNR.
\end{itemize}
\subsection{Organization}
The rest of this paper is organized as follows. In Section II, we briefly introduce the considered system model. The generalized closed-form expressions of average AoI and average Peak AoI for reactive HARQ and proactive HARQ are provided in Section III, where the effect of different delay elements is added into the analysis. In Section IV, we design an optimization problem to reduce the average AoI of reactive HARQ and proactive HARQ, respectively. Numerical results and discussions are given in Section V, followed by conclusions in Section VI.
\section{System Model}
We consider an end-to-end (E2E) code-based timely status update system. The update generator (source) is monitoring a time-varying phenomenon $\mathcal{F}\left(t\right)\in\left\{0,\cdots,2^k-1\right\}$, where the time $t$ is divided into some time slots in units of channel use such that $t \in \mathbb{N}$\footnote{Here we consider the symbol-level AoI analysis. Some recent works focusing on PHY-layer AoI analysis also discretize the time into time slots to analyze symbol-level AoI \cite{Channelcoding4,Blockoptim,Channelcoding2,Age_LDPC}.}. We assume that the monitored phenomenon is modeled as a sequence of independent and uniformly distributed symbols. In such a case, the size of the generated observation is $k$ information bits. The monitored data is transmitted through a noisy channel to a central location. We use the notation $\mathbb{N}$ for non-negative integers and the notation $\mathbb{Z}^+$ for positive integers. Also, we define the notation $\left[m\right]$ as $\left[m\right]\triangleq\left\{1,2,\cdots,m\right\}$ for any positive integer $m \in \mathbb{Z}^+$.
\begin{figure}[t]
\centering
\includegraphics[angle=0,width=0.8\textwidth]{./figures/systemmodel.eps
\caption{A general HARQ-based real-time status update system.}
\label{systemmodel}
\end{figure}
\subsection{Channel Model}
We consider an E2E communication setup leveraging a power-limited AWGN model:
\begin{equation}
Y=\sqrt{P}X+Z,
\end{equation}
where $P$ is the average transmit power, $X$ is the unit-variance coded symbol and $Z \sim \mathcal{N}\left(0,1\right)$ is the independent and identically distributed (i.i.d) AWGN.
\begin{figure}[t]
\centering
\subfigtopskip=2pt
\subfigbottomskip=2pt
\subfigcapskip=-5pt
\subfigure[Reactive HARQ]{
\label{reactive}
\includegraphics[angle=0,width=0.48\textwidth]{./figures/RE_HARQ.eps}}
\subfigure[Proactive HARQ]{
\label{proactive}
\includegraphics[angle=0,width=0.48\textwidth]{./figures/PR_HARQ.eps}}
\caption{Instantaneous age evolutions of reactive HARQ and proactive HARQ. Here the maximum retransmissions (the number of sub-codeblocks) is set as $m=4$. The length of the yellow block represents the coding delay $\tau_{\rm c}$, the length of the green block denotes the decoding delay $\tau_{\rm d}$, the length of the blue block represents the transmission delay $\ell_i$, the length of the oblique arrow projective on the timeline describes the propagation delay $\tau_{\rm p}$ (or feedback delay $\tau_{\rm f}$).}
\label{protocols}
\end{figure}
\begin{remark}
We notice that the AoI analysis over erasure channels has been extensively studied in the existing literature. However, the performance over AWGN channels is still not clear so far. As such, we consider the AWGN channel in this paper to reveal the AoI performance over the AWGN channel.
\end{remark}
\subsection{Hybrid ARQ}
The overall system model is shown in Fig. \ref{systemmodel}. The considered system is in close-up fashion with perfect HARQ feedback\footnote{In this article, we assume that the feedback is error-free. The research with erroneous feedback can be extended following this work.}. At the transmitter end, the update generator (source) generates a $k$-bit short-packet update $u$ and encodes it to a \emph{parent codeword} with length $\sum\nolimits_{i = 1}^m {{\ell _i}}$ channel uses, which is then divided into $m$ sub-codeblocks with length $\ell_i, i\in\left[m\right]$ and stored in a buffer waiting to be transmitted. The coding process above will take up $\tau_{\rm c}$ channel uses, and we call it as coding delay. Next, the stored sub-codeblocks are transmitted over a noisy channel sub-codeblock by sub-codeblock, with each transmission taking a transmission delay $\ell_i, i\in\left[m\right]$ channel uses. The transmitted sub-blocks will take $\tau_{\rm p}$ channel uses to arrive at the receiver end. At the receiver end, we assume that the decoding process is conducted once receiving any complete sub-codeblock. The decoding delay in each transmission round is assumed to be the same and is denoted by $\tau_{\rm d}$. If the update is decoded correctly such that $\hat{u}=u$, an ACKnowledgment (ACK) will be fed back to the transmitter; otherwise, a NACK will be sent back. The feedback, similar to the forwarding information propagation, generally takes time and results in delay by nature, and we denote the delay as $\tau_{\rm f} $ channel uses.
The \emph{generate-at-will} model is adopted in the considered E2E status update system. That is, when the transmitter receives an ACK, the process of sensing and sampling will be performed, and a new update will be generated. In such a case, we mainly focus on two types of HARQ schemes: \emph{reactive} HARQ and \emph{proactive} HARQ. The detailed processes are shown in Fig. \ref{reactive} and Fig. \ref{proactive}, respectively.
\subsubsection{Reactive HARQ}
Reactive HARQ is also know as stop-and-wait HARQ. In Fig. \ref{reactive}, we demonstrate a detailed stop-and-wait retransmission process of \emph{reactive} HARQ, wherein the maximum number of sub-codeblocks (or maximum retransmissions) is set as $m=4$. The so-called \emph{reactive} scheme implies that the transmitter allows for retransmissions only upon the reception of a NACK. As such, the transmitter should always wait for a feedback to decide whether to generate a new update or retransmit the old update's sub-codeblocks. The waiting time, however, is referred ro as the HARQ round trip time (RTT) and will result in additional latency. Therefore, the \emph{reactive} HARQ scheme allows for only a limited number of retransmissions in the URLLC application scenarios and thereby enables great potential to be further advanced \cite{GFHARQ3}.
\subsubsection{Proactive HARQ}
The \emph{proactive} HARQ scheme with maximum sub-codeblocks $m=4$ is shown in Fig. \ref{proactive}. As its name indicates, the retransmission process is completely spontaneous and proactive, which is interrupted only when an ACK is received. The core idea of \emph{proactive} HARQ is to eliminate the need for waiting for a feedback and implement consecutive retransmitting. By \emph{proactive} retransmitting, the latency introduced by waiting for a feedback is reduced, and thus the issue of long HARQ round trip time (RTT) is resolved.
\subsection{Performance Metric}
We focus on AoI analysis and optimization in this paper. Here we simply review the definition of instantaneous AoI as in Definition \ref{def1}. For more intuitive and visualized results, Fig. \ref{protocols} also gives the instantaneous age evolutions of reactive HARQ and proactive HARQ respectively.
\begin{definition}\label{def1}
\textbf{(AoI)} { Denote $t_i^{\rm G}$ as the generation time instant of the $i^{\rm th}$ status update packet that can be correctly decoded, and denote $t_i^{\rm S}$ as the time instant at which this packet is correctly decoded. At a time instant $\tau$, the index of the most recently generated update can be given by $N(\tau)=\min\left\{i|t_i^{\rm S}>\tau \right\}$ and the time stamp is $U(\tau)=t_{N(\tau)}^{\rm G}$. Then, the instantaneous AoI is defined as $\Delta\left(t\right) \triangleq t-U\left(t\right)$.}
\end{definition}
\subsubsection{Average AoI}
Also known as time-average AoI, Average AoI is a statistical metric that measures the long-term average age of a status update system. In our considered discrete symbol-level system where the time is divided into some time slots in units of channel use, the average AoI is defined as follows.
\begin{definition}
\textbf{(Average AoI)} The average AoI of a real-time status update system is defined as:
\begin{equation}
\bar{\Delta} \triangleq \mathop {\lim }\limits_{N \to \infty } \frac{1}{N}\sum\nolimits_{t = 1}^N {\Delta \left( t \right)}.
\end{equation}
\end{definition}
\subsubsection{Average Peak AoI }
We also provide explicit expressions for the average Peak AoI in this paper. The peak age indicates the maximum value of age in each renewal process. In our considered system, the average Peak AoI is defined as follows.
\begin{definition}\label{peakaoidef}
\textbf{(Average Peak AoI)} The average Peak AoI of a real-time status update system is
\begin{equation}
\bar{\Delta}^{\rm P}\triangleq\lim\limits_{N \to \infty }\frac{1}{N}\sum_{j=1}^{N}\Delta\left(t^S_j-1\right).
\end{equation}
\end{definition}
\section{Analytical Results}\label{section3}
In this section, we study the symbol-level AoI of reactive HARQ and proactive HARQ. We first give the closed-form expressions for the AoI in Proposition \ref{reactive AoI} and Proposition \ref{proactive AoI}, and then conduct a theoretical AoI comparison between the two considered transmission protocols in Corollary \ref{coro1}. The AoI expressions, given in (\ref{eqreactive}) and (\ref{eqproactive}), are functions of the block assignment vector $\bf n$ and its dependent error probability vector $\bf e$, where the element $n_i$ in vector $\bf n$ denotes the number of cumulative transmitted symbols up to the $i^{\rm th}$ transmission round with $n_i= \sum\nolimits_{j = 1}^i {{\ell _j}}$, and the element $\epsilon_i$ in vector $\bf e$ denotes the probability that the $i^{\rm th}$ re-transmitted message remains incorrectly decoded.
By flexible choices of the vector $\bf n$ and the vector $\bf e$, we also demonstrate that our derived expression for reactive transmission protocol also unifies the available AoI analyses in the existing literature. Moreover, by using the result of the achievable rate of finite-length codes, we can obtain the AoI closed-form expression under the finite block-length (FBL) regime.
\subsection{Reactive Scheme}
\subsubsection{Average AoI}
Denote the generation time of the $j^{\rm th}$ collected message as $t_j$, and denote the $j^{\rm th}$ collected message as $M_j\triangleq\mathcal{F}\left(t_j\right)$. The $j^{\rm th}$ collected message $M_j$ is encoded to a \emph{parent code} $\mathcal{C}\left(M_j\right)$ with size $n_m=\sum\nolimits_{a\in[m]}\ell_a$ being stored in a transmission buffer. Then, the transmission is evoked round by round until all the symbols stored in the buffer is transmitted or an ACK is received. In the $i^{\rm th}$ round of transmission, the transmitter will transmit $\ell_i$ symbols, and the decoder will leverage the cumulatively received $n_i={\sum\nolimits_{a\in[i]}}\ell_a$ symbols to decode the message.
In such a case, we introduce $\varsigma _{j,i} \in \left\{0,1\right\}$ to denote the feedback signal in the $i^{\rm th}$ transmission round when transmitting message $M_j$: If $\varsigma_{j,i}=1$ or $i=m$, the transmitter will no longer transmit message $M_j$, instead, it will collect new update $M_{j+1}$ for transmission; If $\varsigma_{j,i}=0$ and $i<m$, the source will transmit additional $\ell_{i+1}$ encoded symbols from $\mathcal{C}\left(M_j\right)$. As such, the probability that the $i^{\rm th}$ transmission is not correctly decoded is given as $\epsilon_i=1-\mathbb{E}\varsigma_{j,i}$.
Let $Q_j\triangleq\inf\left\{i>Q_{j-1}:\varsigma_{i,m}=1\right\}, Q_0=0$ be the cumulative number of generated packets until the $j^{\rm th}$ decoding success. Let $R_j\triangleq Q_j-Q_{j-1}-1$ represent the decoding failures between two successful decoding and $V_j\triangleq\inf\left\{i\in [m]:\varsigma_{Q_j,i}=1 \right\}$ denote the round in which the $Q_j^{\rm th}$ packet gets decoded, we have that
\begin{lemma}\label{lemma1}
The random sequences $R_j$ and $V_j$ are independent, and they are i.i.d with distributions
\begin{equation}\label{distribution}
\begin{aligned}
&\mathbb{P}\left(R_j=a\right)=\left(1-\epsilon_m\right)\epsilon^a_m, \quad a\in\mathbb{N},\\
&\mathbb{P}\left(V_j=i\right)=\frac{\epsilon_{i-1}-\epsilon_i}{1-\epsilon_m}, \quad i\in[m].
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
Please refer to Appendix \ref{appendixa}.
\end{proof}
\begin{lemma}\label{lemma2}
The first and second moments for $R_j$ are
\begin{equation}\label{momentsRj}
\begin{aligned}
\mathbb{E}R_j&=\frac{\epsilon_m}{1-\epsilon_m},\\ \mathbb{E}R_j^2&=\frac{\epsilon_m^2+\epsilon_m}{\left(1-\epsilon_m\right)^2}.
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
By utilizing the distributions in (\ref{distribution}) and the definitions that $\mathbb{E}R_j=\sum_{a\in \mathbb{N}}a\cdot\mathbb{P}\left(R_j=a\right)$ and $\mathbb{E}R_j^2=\sum_{a\in \mathbb{N}}a^2\cdot\mathbb{P}\left(R_j=a\right)$, we obtain the results in (\ref{momentsRj}).
\end{proof}
We then let $\tau^{\rm Reac}_i$ represent the elapsed time form generating an update to receiving its $i^{\rm th}$ feedback signal for reactive HARQ. For reactive HARQ, the transmitter should wait for an integral RTT to receive a feedback. As such, we have $\tau^{\rm Reac}_{i}=n_i+\tau_{\rm c}+i\left(\tau_{\rm d}+\tau_{\rm f}+\tau_{\rm p}\right)$. Denote $\mathcal{T}=\tau_{\rm d}+\tau_{\rm f}+\tau_{\rm p}$, the first and second moments for $\tau^{\rm Reac}_{V_j}$ is given as follows.
\begin{lemma}\label{lemma3}
The first and second moments for $\tau^{\rm Reac}_{V_j}$ are
\begin{equation}\label{momentstsouR}
\begin{aligned}
\mathbb{E}\tau^{\rm Reac}_{V_j}&=n_1+\tau_{\rm c}+\mathcal{T}+\sum_{i\in [m-1]}\frac{\left(n_{i+1}-n_i+\mathcal{T}\right)\epsilon_i}{1-\epsilon_m},\\
\mathbb{E}\left(\tau^{\rm Reac}_{V_j}\right)^2&=\left(n_1+\tau_{\rm c}+\mathcal{T}\right)^2+\sum_{i\in [m-1]}\frac{\left(n_{i+1}-n_i+\mathcal{T}\right)\left(n_{i+1}+n_i+2\tau_{\rm c}+\left(2i+1\right)\mathcal{T}\right)\epsilon_i}{1-\epsilon_m}.
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
By utilizing the distributions in (\ref{distribution}) and the definitions that $\mathbb{E}\tau^{\rm Reac}_{V_j}=\sum_{i\in [m]}\tau^{\rm Reac}_i\cdot\mathbb{P}\left({V_j}=i\right)$ and $\mathbb{E}\left(\tau^{\rm Reac}_{V_j}\right)^2=\sum_{i\in [m]}\left(\tau^{\rm Reac}_i\right)^2\cdot\mathbb{P}\left({V_j}=i\right)$, we obtain the results in (\ref{momentstsouR}).
\end{proof}
With these notations, we can recursively write the time-instant of the $j^{\rm th}$ successful decoding $t_j^S$ as follows.
\begin{equation}
t_j^S=t_{j-1}^S+\tau^{\rm Reac}_{m}R_j+\tau^{\rm Reac}_{V_j}+\tau_{\rm f}.
\end{equation}
Therefore, the interval between the ${\left(j-1\right)}^{\rm th}$ and the $j^{\rm th}$ successful decoding for reactive HARQ is given as follows.
\begin{equation}\label{ReacTj}
\begin{aligned}
T^{\rm Reac}_j&=t_j^S-t_{j-1}^S=\tau^{\rm Reac}_{m}R_j+\tau^{\rm Reac}_{V_j}+\tau_{\rm f}.
\end{aligned}
\end{equation}
Since both the random sequences $R_j$ and $V_j$ are i.i.d. and have finite first and second moments, it turns out that the sequence $T^{\rm Reac}_j$ is also i.i.d. with finite
finite first and second moments. We give the first and second moments of $T^{\rm Reac}_j$ as follows.
\begin{lemma}
The first and second moments for $T^{\rm Reac}_j$ are
\begin{equation}\label{momentsTj}
\begin{aligned}
&\mathbb{E}T^{\rm Reac}_j=\frac{n_1+\tau_{\rm c}+\mathcal{T}+\sum_{i\in [m-1]}{\left(n_{i+1}-n_i+\mathcal{T}\right)\epsilon_i}}{1-\epsilon_m},\\
&\mathbb{E}\left(T^{\rm Reac}_j\right)^2=\frac{\left(n_m+\tau_{\rm c}+m\mathcal{T}\right)^2\left(1+\epsilon_m\right)}{\left(1-\epsilon_m\right)^2}\\
&-\sum_{i\in [m-1]}\frac{1-\epsilon_i}{1-\epsilon_m}\left(n_{i+1}-n_i+\mathcal{T}\right)\left[n_i+n_{i+1}+2\tau_{\rm c}+\left(2i+1\right)\mathcal{T}+\frac{2\left(n_m+\tau_{\rm c}+m\mathcal{T}\right)\epsilon_m}{1-\epsilon_m}\right].
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
From (\ref{ReacTj}), we obtain
\begin{equation}\label{momentsTjreac}
\begin{aligned}
&\mathbb{E}T^{\rm Reac}_j=\tau^{\rm Reac}_{m}\mathbb{E}R_j+\mathbb{E}\tau^{\rm Reac}_{V_j}+\tau_{\rm f},\\
&\mathbb{E}\left(T^{\rm Reac}_j\right)^2=\mathbb{E}\left(\tau^{\rm Reac}_{m}R_j+\tau^{\rm Reac}_{V_j}+\tau_{\rm f}\right)^2.\\
\end{aligned}
\end{equation}
Note that the random variables $R_j$ and ${V_j}$ are independent with each other and the first and second moments of $R_j$ and $\tau^{\rm Reac}_{V_j}$ have been given in Lemma \ref{lemma2} and Lemma \ref{lemma3}. Substitute them into (\ref{momentsTjreac}), we can obtain the results in (\ref{momentsTj}).
\end{proof}
As Definition \ref{def1} indicates, the generation time instant $U\left(\tau\right)$ is given by $U\left(\tau\right)=t^G_{N\left(\tau\right)}$. Because $N\left(t^S_{j-1}\right)=\min\left\{i|t_i^{\rm S}>t^S_{j-1} \right\}=j$, we obtain that
\begin{equation}\nonumber
U\left(t^S_{j-1}\right)=t^G_{j}=t^S_{j-1}+\tau_{\rm f}-\tau^{\rm Reac}_{V_{j-1}}.
\end{equation}
Thus, for any time slots $t$ in the $j^{\rm th}$ renewal interval $\mathcal{I}_j\triangleq\left\{t^S_{j-1},\cdots,t^S_{j}-1\right\}$, we have
\begin{equation}
U\left(t\right)=U\left(t^S_{j-1}\right)=t^S_{j-1}+\tau_{\rm f}-\tau^{\rm Reac}_{V_{j-1}}, \text{ for } t \in \mathcal{I}_j.
\end{equation}
As such, the instantaneous age is given as
\begin{equation}\label{deltreact}
\begin{aligned}
\Delta\left(t\right)&=t-U\left(t\right)\\
&=t-t^S_{j-1}-\tau_{\rm f}+\tau^{\rm Reac}_{V_{j-1}}, \text{ for } t \in \mathcal{I}_j .
\end{aligned}
\end{equation}
\begin{lemma}\label{lemma5}
For the considered reactive HARQ model, the average AoI can be calculated by
\begin{equation} \label{ReacAoI}
\bar{\Delta}_{\rm Reactive}=\frac{\mathbb{E}\sum_{t\in \mathcal{I}_j}\Delta{\left(t\right)}}{\mathbb{E}T^{\rm Reac}_j}=\frac{\mathbb{E}T^{\rm Proac}_j}{2\mathbb{E}\left(T^{\rm Reac}_j\right)^2}+\mathbb{E}\tau^{\rm Reac}_{V_j}-\tau_{\rm f}-\frac{1}{2}.
\end{equation}
\end{lemma}
\begin{proof}
Please refer to Appendix \ref{appendixb}.
\end{proof}
Then, by adopting the available first and second moments of $T^{\rm Reac}_j$ and $\tau^{\rm Reac}_{V_j}$ given in (\ref{momentstsouR}) and (\ref{momentsTjreac}), we have the average AoI for reactive HARQ as follows.
\begin{proposition} \label{reactive AoI}
{ \textbf{(The Generalized Closed-form Average AoI Expression for Reactive Scheme)} For reactive HARQ with maximum retransmissions $m$, block assignment vector ${\bf n}=\left(n_1,n_2,\cdots,n_m\right)$ and error probability vector ${\bf e}=\left(\epsilon_1,\epsilon_2,\cdots,\epsilon_m\right)$, the average AoI can be calculated by}
\begin{equation} \label{eqreactive}
\begin{aligned}
{\bar{\Delta} _{\rm Reactive}} = - \frac{1}{2} - {\tau _{\rm f}} + \frac{{{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i} + \mathcal{T}} \right){\epsilon_i}} }}{{1 - {\epsilon_m}}} + \\
\frac{{{{\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T}} \right)}^2} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i} + \mathcal{T}} \right)\left( 2{{\tau _{\rm c}} + {n_{i + 1}} + {n_i} + \left( {2i + 1} \right)\mathcal{T}} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i} + \mathcal{T}} \right){\epsilon_i}} } \right)}},
\end{aligned}
\end{equation}
where $\mathcal{T}=\tau_{\rm f}+\tau_{\rm p}+\tau_{\rm d}$ with $\tau _{\rm c}$, $\tau _{\rm p}$, $\tau _{\rm d}$ and $\tau _{\rm f}$ denoting the coding delay, propagation delay, decoding delay and feedback delay, respectively.
\end{proposition}
\subsubsection{Case Study: A Unified Result}
With Proposition \ref{reactive AoI} in hand, we can conduct some case studies by flexibly considering the choices of the block assignment vector $\bf n$ and the error probability vector $\bf e$. By this means, we theoretically show that the closed-form AoI expressions given in this paper is a unified result.\footnote{The average AoI expressions in Case 1, Case 2 and Case 3 are corresponding to Proposition 1, Proposition 3 and Proposition 2 of \cite{Age_LDPC}, respectively. The average AoI expression in Case 4 is a variant of the result in \cite{Blockoptim}.}. Though the given examples are not exhaustive in this paper, we can observe from these case studies that the unified expression given in (\ref{eqreactive}) enables potential for exploring the intrinsic relationship and comparative insights among different types of transmission protocols.
\begin{case}
{\textbf{(Average AoI for Fixed-rate Codes without ARQ)} We show that the available average AoI expression for fixed-rate codes in \cite{Age_LDPC} is a specific case of our unified result in (\ref{eqreactive}). For fixed-rate codes without ARQ, the maximum retransmissions turns to $m=1$. Substitute $m=1$ into (\ref{eqreactive}) and remove the effect of delay elements such that $\tau_{\rm c}=\tau_{\rm p}=\tau_{\rm d}=\tau_{\rm f}=0$, we can obtain the average AoI as the Proposition 1 in \cite{Age_LDPC}: }
\begin{equation}
\nonumber
{\bar{\Delta} _{\rm Non-ARQ}} = - \frac{1}{2} + \frac{{{n_1}}}{{1 - {\epsilon_1}}} + \frac{{{n_1}}}{2},
\end{equation}
where $n_1$ is the code length and $\epsilon_1$ is the error probability of the fixed-rate codes.
\end{case}
\begin{case}
{ \textbf{(Average AoI for Truncated ARQ (TARQ))} We demonstrate that the average AoI expression for TARQ is also a specific case of our unified result in (\ref{eqreactive}). For truncated ARQ, the transmitter retransmits the same packet till the allowable maximum retransmissions $m$ is reached or this packet is successfully received. Since the retransmitted packet is the same as the first packet, the cumulative transmitted message length is $n_i=in_1$ and the corresponding error probability is $\epsilon_i={\epsilon_1}^i$. Then, by substituting them back into (\ref{eqreactive}) and similarly remove the effect of delay elements such that $\tau_{\rm c}=\tau_{\rm p}=\tau_{\rm d}=\tau_{\rm f}=0$, we can obtain the average AoI as the Proposition 3 in \cite{Age_LDPC}: }
\begin{equation}
\nonumber
{\bar{\Delta}_{\rm TARQ}} = - \frac{1}{2} + {n_1}\left( {\frac{2}{{1 - {\epsilon_1}}} - \frac{1}{2} - \frac{{m\epsilon_1^m}}{{1 - \epsilon_1^m}}} \right).
\end{equation}
\end{case}
\begin{case}
{ \textbf{(Average AoI for Classical ARQ)} We also find that the average AoI expression for TARQ is a specific case of our unified result in (\ref{eqreactive}). For classical ARQ, the transmitter re-transmits the same packet till the packet is successfully received, while the maximum retransmissions is not limited. The classical ARQ is a special case of TARQ where $m \to \infty$. Then, by calculating the limit $\mathop {\lim }\limits_{m \to \infty } {\Delta _{\rm TARQ}}$, we can obtain the average AoI as the Proposition 2 in \cite{Age_LDPC}: }
\begin{equation} \nonumber
{\bar{\Delta} _{\rm Classical - ARQ}} = - \frac{1}{2} + {n_1}\left( {\frac{2}{{1 - {\epsilon_1}}} - \frac{1}{2}} \right).
\end{equation}
\end{case}
\begin{case}
{ \textbf{(Average AoI for HARQ-IR)} By removing the effect of delay elements such that $\tau_{\rm c}=\tau_{\rm p}=\tau_{\rm d}=\tau_{\rm f}=0$, we find that the result in (\ref{eqreactive}}) is transformed to a variant of that in \cite{Blockoptim}, given as:
\begin{equation} \nonumber
\bar{\Delta} _{\rm HARQ-IR} = - \frac{1}{2} + \frac{{{n_1} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon _i}} }}{{1 - {\epsilon _m}}} + \frac{{n_1^2 + \sum\limits_{i = 1}^{m - 1} {\left( {n_{i + 1}^2 - n_i^2} \right){\epsilon _i}} }}{{2\left( {{n_1} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon _i}} } \right)}}.
\end{equation}
\end{case}
\subsubsection{Average Peak AoI}
Definition \ref{peakaoidef} has indicated that $\bar{\Delta}^{\rm P}\triangleq\lim\limits_{N \to \infty }\frac{1}{N}\sum_{j=1}^{N}\Delta\left(t^S_j-1\right)$. Note that for reactive scheme, the terms in the summation can be obtained from (\ref{deltreact}), given as $\Delta\left(t^S_j-1\right)=T^{\rm Reac}_j-1-\tau_{\rm f}+\tau^{\rm Reac}_{V_{j-1}}$. Then, from the law of large numbers, we can obtain the following almost sure equality
\begin{equation}\nonumber
\bar{\Delta}^{\rm P}_{\rm Reactive}=-1-\tau_{\rm f}+\mathbb{E}T^{\rm Reac}_j+\mathbb{E}\tau^{\rm Reac}_{V_{j-1}}.
\end{equation}
Then, by applying the available first moments of $T^{\rm Reac}_j$ and $\tau^{\rm Reac}_{V_j}$ given in (\ref{momentstsouR}) and (\ref{momentsTjreac}), we obtain the average Peak AoI for reactive HARQ in the following proposition.
\begin{proposition} \label{theorem6}
{ \textbf{(The Generalized Closed-form Average Peak AoI Expression for Reactive HARQ)} For reactive HARQ with maximum retransmissions $m$, block assignment vector ${\bf n}=\left(n_1,n_2,\cdots,n_m\right)$ and error probability vector ${\bf e}=\left(\epsilon_1,\epsilon_2,\cdots,\epsilon_m\right)$, the average Peak AoI can be calculated by}
\begin{equation}
\nonumber
{\bar{\Delta}^{\rm P} _{\rm Reactive}} = -1 - {\tau _{\rm f}} - \frac{{\left( {{\tau _{\rm c}} + {n_m} + m\mathcal{T}} \right){\epsilon_m}}}{{1 - {\epsilon_m}}} +
\frac{2}{{1 - {\epsilon_m}}}\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i} + \mathcal{T}} \right){\epsilon_i}} } \right).
\end{equation}
\end{proposition}
\subsection{Proactive Scheme}
\subsubsection{Average AoI}
Let $\tau^{\rm Proac}_i$ represent the elapsed time form generating an update to receiving its $i^{\rm th}$ feedback signal for proactive HARQ, we can observe from Fig. \ref{proactive} that $\tau^{\rm Proac}_{i}=n_i+\tau_{\rm c}+\mathcal{T}$. As such, the first and second moments for $\tau^{\rm Proac}_{V_j}$ is correspondingly given as follows.
\begin{lemma}\label{proaclemma1}
The first and second moments for $\tau^{\rm Proac}_{V_j}$ are
\begin{equation}\label{momentstouPr}
\begin{aligned}
\mathbb{E}\tau^{\rm Proac}_{V_j}&=n_1+\tau_{\rm c}+\mathcal{T}+\sum_{i\in [m-1]}\frac{\left(n_{i+1}-n_i\right)\epsilon_i}{1-\epsilon_m},\\
\mathbb{E}\left(\tau^{\rm Proac}_{V_j}\right)^2&=\left(n_1+\tau_{\rm c}+\mathcal{T}\right)^2+\sum_{i\in [m-1]}\frac{\left(n_{i+1}-n_i\right)\left(n_{i+1}+n_i+2\tau_{\rm c}+2\mathcal{T}\right)\epsilon_i}{1-\epsilon_m}.
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
By utilizing the distributions in (\ref{distribution}) and the definitions that $\mathbb{E}\tau^{\rm Proac}_{V_j}=\sum_{i\in [m]}\tau^{\rm Proac}_i\cdot\mathbb{P}\left({V_j}=i\right)$ and $\mathbb{E}\left(\tau^{\rm Proac}_{V_j}\right)^2=\sum_{i\in [m]}\left(\tau^{\rm Proac}_i\right)^2\cdot\mathbb{P}\left({V_j}=i\right)$, we obtain the results in (\ref{momentstouPr}).
\end{proof}
Denote the interval between the ${\left(j-1\right)}^{\rm th}$ and the $j^{\rm th}$ successful decoding for proactive HARQ as $T^{\rm Proac}_j$, we can similarly derive that
\begin{equation}\label{Tjproac}
T^{\rm Proac}_j=\tau^{\rm Proac}_{m}R_j+\tau^{\rm Proac}_{V_j}+\tau_{\rm f}.
\end{equation}
Then we have the first and second moments of $T^{\rm Proac}_j$ as follows.
\begin{lemma}\label{proaclemma2}
The first and second moments for $T^{\rm Proac}_j$ are
\begin{equation}\label{momentsTjProac}
\begin{aligned}
&\mathbb{E}T^{\rm Proac}_j=\frac{n_1+\tau_{\rm c}+\mathcal{T}+\sum_{i\in [m-1]}{\left(n_{i+1}-n_i\right)\epsilon_i}}{1-\epsilon_m},\\
&\mathbb{E}\left(T^{\rm Proac}_j\right)^2=\frac{\left(n_m+\tau_{\rm c}+\mathcal{T}\right)^2\left(1+\epsilon_m\right)}{\left(1-\epsilon_m\right)^2}\\
&-\sum_{i\in [m-1]}\frac{1-\epsilon_i}{1-\epsilon_m}\left(n_{i+1}-n_i\right)\left[n_i+n_{i+1}+2\tau_{\rm c}+2\mathcal{T}+\frac{2\left(n_m+\tau_{\rm c}+\mathcal{T}\right)\epsilon_m}{1-\epsilon_m}\right].
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
From (\ref{Tjproac}), we obtain that
\begin{equation}\label{momentsTjproac}
\begin{aligned}
&\mathbb{E}T^{\rm Proac}_j=\tau^{\rm Proac}_{m}\mathbb{E}R_j+\mathbb{E}\tau^{\rm Proac}_{V_j}+\tau_{\rm f},\\
&\mathbb{E}\left(T^{\rm Proac}_j\right)^2=\mathbb{E}\left(\tau^{\rm Proac}_{m}R_j+\tau^{\rm Proac}_{V_j}+\tau_{\rm f}\right)^2.\\
\end{aligned}
\end{equation}
Substitute (\ref{distribution}) and (\ref{momentsRj}) into (\ref{momentsTjproac}), we obtain the results in (\ref{momentsTjProac}).
\end{proof}
For the considered proactive HARQ model, the average AoI can be similarly expressed as
\begin{equation} \label{ProacAoI}
\bar{\Delta}_{\rm Proactive}=\frac{\mathbb{E}\sum_{t\in \mathcal{I}_j}\Delta{\left(t\right)}}{\mathbb{E}T^{\rm Proac}_j}=\frac{\mathbb{E}T^{\rm Proac}_j}{2\mathbb{E}\left(T^{\rm Proac}_j\right)^2}+\mathbb{E}\tau^{\rm Proac}_{V_j}-\tau_{\rm f}-\frac{1}{2}.
\end{equation}
Hence, applying the first and second moments given in Lemma \ref{proaclemma1} and Lemma \ref{proaclemma2} leads to the explicit expression in the following proposition.
\begin{proposition} \label{proactive AoI}
{\textbf{(The Generalized Closed-form Average AoI Expression for Proactive HARQ)} For proactive HARQ with maximum retransmissions $m$, block assignment vector ${\bf n}=\left(n_1,n_2,\cdots,n_m\right)$ and error probability vector ${\bf e}=\left(\epsilon_1,\epsilon_2,\cdots,\epsilon_m\right)$, the average AoI can be calculated by}
\begin{equation} \label{eqproactive}
\begin{aligned}
{\bar{\Delta} _{\rm Proactive}} = - \frac{1}{2} - {\tau _{\rm f}} + \frac{{{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} }}{{1 - {\epsilon_m}}} + \\
\frac{{{{\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T}} \right)}^2} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2{\tau _{\rm c}} + 2\mathcal{T} + {n_{i + 1}} + {n_i}} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right)}}.
\end{aligned}
\end{equation}
\end{proposition}
\subsubsection{Average Peak AoI}
As the definition indicates, we have $\bar{\Delta}^{\rm P}\triangleq\lim\limits_{N \to \infty }\frac{1}{N}\sum_{j=1}^{N}\Delta\left(t^S_j-1\right)$. For proactive HARQ, $\Delta\left(t^S_j-1\right)$ is given by $\Delta\left(t^S_j-1\right)=T^{\rm Proac}_j-1-\tau_{\rm f}+\tau^{\rm Proac}_{V_{j-1}}$. Then, from the law of large numbers, we get the following equality as
\begin{equation}\nonumber
\bar{\Delta}^{\rm P}_{\rm Proactive}=-1-\tau_{\rm f}+\mathbb{E}T^{\rm Proac}_j+\mathbb{E}\tau^{\rm Proac}_{V_{j-1}}.
\end{equation}
Finally, by applying the available first of $T^{\rm Proac}_j$ and $\tau^{\rm Proac}_{V_j}$ given in (\ref{momentstouPr}) and (\ref{momentsTjproac}), we obtain the average Peak AoI for reactive HARQ in the following proposition.
\begin{proposition} \label{theorem7}
{ \textbf{(The Generalized Closed-form Average Peak AoI Expression for Proactive HARQ)} For proactive HARQ with maximum retransmissions $m$, block assignment vector ${\bf n}=\left(n_1,n_2,\cdots,n_m\right)$ and error probability vector ${\bf e}=\left(\epsilon_1,\epsilon_2,\cdots,\epsilon_m\right)$, the average Peak AoI can be calculated by}
\begin{equation}
\nonumber
{\bar{\Delta}^{\rm P} _{\rm Proactive}} = -1 - {\tau _{\rm f}} - \frac{{\left( {{\tau _{\rm c}} + {n_m} + \mathcal{T}} \right){\epsilon_m}}}{{1 - {\epsilon_m}}} +
\frac{2}{{1 - {\epsilon_m}}}\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right).
\end{equation}
\end{proposition}
\subsubsection{Case Study: Rateless Codes }
For rateless codes, the encoder can generate as many symbols as possible to achieve error-free transmission. As such, rateless codes can be regarded as a type of proactive HARQ with infinite code-length setup. By leveraging the obtained results regarding to proactive HARQ, the average AoI and average Peak AoI of rateless code are give in the following Propositions.
\begin{proposition} \label{theorem5}
{ \textbf{(The Generalized Closed-form Average AoI Expression for Rateless Codes)} For rateless codes transmitted over a noisy channel with non-trivial coding delay $\tau_{\rm c}$, propagation delay $\tau_{\rm p}$, decoding delay $\tau_{\rm d}$ and feedback delay $\tau_{\rm f}$, the average AoI can be calculated by}
\begin{equation} \label{rateless}
\begin{aligned}
{\bar{\Delta} _{\rm Rateless}} = - \frac{1}{2} + \mathcal{T} +{n_1} + + \sum\limits_{i = 1}^{\infty} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} \\
+\frac{{{{\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T}} \right)}^2} + \sum\limits_{i = 1}^{\infty} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2{\tau _{\rm c}} + 2\mathcal{T} + {n_{i + 1}} + {n_i}} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{\infty} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right)}}.
\end{aligned}
\end{equation}
\end{proposition}
\begin{proof}
Rateless codes is a special type of proactive HARQ where $m\to \infty$. Since $\lim\limits_{m \to \infty}\epsilon_m=0$, we can obtain the average AoI of rateless codes as in (\ref{rateless}) by calculating the limit $\mathop {\lim }\limits_{m \to \infty } {\bar{\Delta} _{\rm Proactive}}$.
\end{proof}
\begin{proposition} \label{retelesspeakaoi}
{ \textbf{(The Generalized Closed-form Average Peak AoI Expression for Rateless Codes)} For rateless codes transmitted over the channel with non-trivial coding delay $\tau_{\rm c}$, propagation delay $\tau_{\rm p}$, decoding delay $\tau_{\rm d}$ and feedback delay $\tau_{\rm f}$, the average AoI can be calculated by}
\begin{equation} \label{ratelesspeak}
{\bar{\Delta}^{\rm P} _{\rm Rateless}} = -1- {\tau _{\rm f}} +
{2}\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{\infty} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right).
\end{equation}
\end{proposition}
\begin{proof}
Rateless codes is a special case of proactive HARQ where $m\to \infty$. By calculating the limit $\mathop {\lim }\limits_{m \to \infty } {\bar{\Delta} _{\rm Proactive}}$, we can obtain the average AoI of rateless codes as in (\ref{ratelesspeak}).
\end{proof}
\begin{remark}
Note that there are infinite series in (\ref{rateless}) and (\ref{ratelesspeak}), which are
\begin{equation}\label{inseries}
\begin{aligned}
&\sum\limits_{i = 1}^{\infty} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}},\\
&\sum\limits_{i = 1}^{\infty} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2{\tau _{\rm c}} + 2\mathcal{T} + {n_{i + 1}} + {n_i}} \right){\epsilon_i}}.
\end{aligned}
\end{equation}
In this regard, the sufficient condition that Proposition \ref{theorem5} and Proposition \ref{retelesspeakaoi} exist is that the infinite series in (\ref{inseries}) converge to some finite values. In the following Lemma \ref{lemma8}, we would like to discuss this issue.
\end{remark}
\begin{lemma}\label{lemma8}
The infinite series in (\ref{inseries}) are always bounded (less than or equal to some finite number).
\end{lemma}
\begin{proof}
Please refer to Appendix \ref{appendixc}.
\end{proof}
\subsection{Reactive HARQ vs. Proactive HARQ}
\begin{corollary} \label{coro1}
{ \textbf{(Reactive HARQ vs. Proactive HARQ)} The average age performance of reactive HARQ would not exceed that of proactive HARQ under the same block assignment vector $\bf n$ and the same error probability vector $\bf e$. The necessary and sufficient condition for their equivalence is $m=1$ or $\mathcal{T}=0$. }
\end{corollary}
\begin{proof}
Please refer to Appendix \ref{appendixd}.
\end{proof}
Corollary \ref{coro1} demonstrates that $\bar{\Delta}_{\rm Reactive}\ge\bar{\Delta}_{\rm Proactive}$, where the equivalence happens only if $i$) $m=1$, in such a case, both the reactive HARQ and the proactive HARQ turns to a open-loop fashion non-ARQ system, and the system does not send any incremental redundancy; $ii$) $\mathcal{T}=0$, this condition infers to an ideal assumption where the propagation delay, decoding delay and feedback delay are negligible. In this regard, the RTT issue of reactive HARQ does not exist any more, and thus the considered reactive scheme is the same as the proactive one.
\subsection{Average Age in the FBL Regime}
With the above closed-form results, we observe that the average AoI of a HARQ-based system can be directly evaluated by determining the error probability vector $\textbf{e}$ and the block assignment vector $\textbf{n}$. The error probability $\epsilon_i$ is affected by three factors, which are, $i$) the channel condition; $ii$) the coding and decoding technique; $iii$) the message length $k$ and the code length $n_i$. As such, the framework given in this paper is general-purpose, enabling potential AoI research under different coding schemes and channel conditions.
For instance, the given generalized expressions allow us to adopt the FBL results in \cite{FBL} to evaluate the AoI of the considered HARQ protocols. Over the power-limited AWGN channel with SNR $\gamma$, the error probability $\epsilon_i$ can be approximated by the Theorem 54 of \cite{FBL} as:\footnote{Here we focus on FBL analysis as a case study. Note that $\epsilon_i$ can also be characterized by other specific error-correcting techniques. }
\begin{equation} \label{Rateerr}
{\epsilon_i} \approx Q\left( {\frac{C\left( { \gamma } \right) - k/{n_i} -\frac{1}{2}{\log_2{n_i}}/{n_i} } {{\sqrt {V\left( \gamma \right)/{n_i}} }}} \right).
\end{equation}
where $C$ is the channel capacity with $C=\log_2\left(1+\gamma\right)$, $V$ is the channel dispersion with $V=(1-\frac{1}{\left(1+\gamma\right)^2})\log_2^2e$, and $Q\left(\cdot\right)$ denotes the $Q$ function with
\begin{equation} \nonumber
Q\left(x\right)=\int_{x}^{\infty}\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{t^2}{2}\right)dt.
\end{equation}
Substitute (\ref{Rateerr}) into (\ref{eqreactive}) and (\ref{eqproactive}), we obtain the average AoI closed-form expressions in the FBL regime.
\section{Age-optimal Block Assignment}
In addition to the error probability $\bf e$, which has been determined by the available finite-length results in (\ref{Rateerr}), the other factor that can significantly affect the age performance is the block assignment vector $\bf n$. The block assignment vector $\bf n$ is an important system parameter regarding the design of the transmission strategy. To this end, this section provides design guidelines for the system parameters selection for an status update system to improve the average information timeliness. By this means, we would like to answer how many retransmissions should be and what lengths they are in an age-optimal system.
\subsection{Problem Formulation}
\begin{algorithm}[h]
\label{Algorithm 1}
\caption{The algorithm for solving Problem \ref{p1}.}
\LinesNumbered
\KwIn{The signal-to-noise ratio $\gamma$; The message length $k$; The lower bound of the range of block length $n_{\min}$; The upper bound of the range of block length $n_{\max}$; The system delay $\tau_{\rm c}$, $\tau_{\rm p}$, $\tau_{\rm d}$ and $\tau_{\rm f}$; }
\KwOut{The optimal block assignment vector ${\bf n}_{\rm optimal}$; The minimum average age $\bar{\Delta}_{\min}$;}
Initialization: $\bar{\Delta}_{\min} = \infty$; $\mathcal{S}=\left\{0,1\right\}^{n_{\max}-n_{\min}+1}$\;
\For{${\bf p}$ in $\mathcal{S}$ }
{
Map vector $\bf p$ to the block assignment vector $\bf n$\;
According to the obtained $\bf n$, calculate the average age $\bar{\Delta}$ by using (\ref{eqreactive}) or (\ref{eqproactive}) \;
\If{$\bar{\Delta} < \bar{\Delta}_{\min}$}
{
Update $\bar{\Delta}_{\min}=\bar{\Delta}$\;
Update ${\bf n}_{\rm optimal}={\bf n}$\;
}
}
\Return ${\bf n}_{\rm optimal}$ and $\bar{\Delta}_{\min}$
\end{algorithm}
We establish an average AoI minimization problem here to further explore the age-optimal transmission mechanism in the FBL regime with non-trivial delay:
{\rm 1) Objective function: To minimize the average age $\bar{\Delta}$. }
{\rm 2) Decision variable: The block assignment vector ${\bf n}=\left(n_1, n_2, \cdots, n_m\right)$.
\begin{problem} Age-optimal block assignment for reactive HARQ (or proactive HARQ)\label{p1}
\begin{equation}\nonumber
\begin{aligned}
\min_{\bf n} \quad &{\bar{\Delta} _{\rm Reactive}} \quad {\rm or} \quad {\bar{\Delta} _{\rm Proactive}}\\
\mbox{s.t.} \quad
&c_1:{{n_{\min }} \le {n_1} < {n_2} < \ldots < {n_m} \le {n_{\max }}},\\
&c_2:{1 \le m \le n_{\max}-n_{\min}+1},\\
&c_3:{\epsilon_i} = Q\left( {\frac{C\left( { \gamma } \right) - k/{n_i} -\frac{1}{2}{\log_2n_i}/{n_i} } {{\sqrt {V\left( \gamma \right)/{n_i}} }}} \right),\\
&c_4:{m,{n_i} \in {\mathbb{Z}^ + },i = 1, \ldots ,m}.
\end{aligned}
\end{equation}
\end{problem}
Note that the decision variable $\bf n$ is a variable-length vector with infinite solution space, we introduce $n_{\min}$ and $n_{\max}$ as constraints of the solution space, which denotes the lower bound and the upper bound of the range of block length, respectively.
\subsection{Solutions and Discussions}
Problem \ref{p1} is a nonlinear integer problem. To solve the optimal solution of Problem \ref{p1}, an auxiliary vector $ {\bf p} \in \mathcal{S} \triangleq \left\{0,1\right\}^{n_{\max}-n_{\min}+1}$ can be introduced here.
\begin{lemma} \label{mappingnp}
There exists an one-to-one mapping between vectors $\bf n$ and $\bf p$.
\end{lemma}
\begin{proof}
Please refer to appendix \ref{appendixe}, where we construct a specific one-to-one mapping function between.
\end{proof}
Lemma \ref{mappingnp} illustrates that the introduced auxiliary vector $\bf p$ can be regarded as an index of the solution space of Problem \ref{p1}, which can help us traverse the entire solution space efficiently and find the optimal solution. The detailed algorithm process is provided in Algorithm \ref{Algorithm 1}.
\begin{figure}[t]
\centering
\includegraphics[angle=0,width=0.9\textwidth]{./figures/blockassignmentv2.eps
\caption{Age-optimal block assignment of reactive HARQ and proactive HARQ. Here the length of message is $k=100$, the minimum code length $n_{\min}=100$, the maximum code length $n_{\max}=120$, the coding delay is $\tau _{\rm c}=20$, the decoding delay is $\tau _{\rm d}=30$.}
\label{Age-optimal block assignment}
\end{figure}
Fig. \ref{Age-optimal block assignment} gives some detailed examples of the solved optimal block assignment vector ${\bf n}_{\rm optimal}$ under different protocols, SNRs, and propagation delays. For example, under SNR$=0.7$ dB and $\tau_{\rm p}=0$, the optimal block assignment vector for reactive HARQ is ${\bf n}_{\rm optimal}=\left(105,120\right)$; under SNR$=1.9$ dB and $\tau_{\rm p}=20$, the optimal block assignment vector for reactive HARQ is ${\bf n}_{\rm optimal}=\left(100,112,120\right)$.
Fig. \ref{Age-optimal block assignment}. leads to the following conclusions:
\begin{itemize}
\item For proactive HARQ, the finest grained symbol-by-symbol strategy always minimizes the average AoI.
\item For reactive HARQ, the age-optimal block assignment varies among different SNRs and propagation delays. As the propagation delay increases, the number of retransmissions will monotonically decrease and finally converge to $m=1$. In such a case, the transmission scheme turns an open-loop fashion without any retransmission. This indicates that there exists a threshold of the propagation delay, only within which retransmission is beneficial to AoI.
\item From a perspective of channel coding, we can see that the trade-off between reliability and effectiveness can be well evaluated by the new metric, AoI. It is well known that a longer code length can improve the reliability while sacrificing the effectiveness; however, what is not fully explored is that an appropriate choice of code length can minimize the AoI.
\end{itemize}
\subsection{A Heuristic Algorithm for Reactive HARQ}
For proactive HARQ, it has been empirically shown that the finest grained strategy minimizes the average AoI; however, for reactive HARQ, the age-optimal strategies vary along with channel conditions and propagation delay. Thus, to repeatedly determine the age-optimal scheme requires amounts of calculations. It is also pertinent to note that the implementation of Algorithm \ref{Algorithm 1} aims to exhaustively search the whole solution space to find an age-optimal block assignment strategy. As such, the complexity of such an Algorithm is exponentially increasing with the range of code length $n_{\max}-n_{\min}+1$. For a broader range of code length, here we heuristically provide a sub-optimal algorithm to circumvent the high-complexity issue. Specifically, the heuristic sub-optimal algorithm is given in Algorithm \ref{Algorithm 2}.
\begin{algorithm}
\label{Algorithm 2}
\caption{The sub-optimal algorithm for solving Problem \ref{p1}}
\LinesNumbered
\KwIn{The signal-to-noise ratio (SNR); The message length $k$; The lower bound of the range of block length $n_{\min}$; The upper bound of the range of block length $n_{\max}$; The system delay $\tau_{\rm c}$, $\tau_{\rm p}$, $\tau_{\rm d}$ and $\tau_{\rm f}$;}
\KwOut{The optimal block assignment vector ${\bf n}_{\rm optimal}$; The minimum average age $\bar{\Delta}_{\min}$;}
Initialization: $\bar{\Delta}_{\min}=\infty$\;
\For{$m \leftarrow$ $1$ to $n_{\max}-n_{\min}+1$ }
{
$\bar{\Delta}_{\min,m}=\infty$\;
Construct the sub set $\mathcal{S}_m$\;
\For{$\bf p$ in $\mathcal{S}_m$}
{
Map vector $\bf p$ to the block assignment vector $\bf n$\;
According to the obtained $\bf n$, calculate the average age $\bar{\Delta}$ by using (\ref{eqreactive}) or (\ref{eqproactive}) \;
\If{$\bar{\Delta} < \bar{\Delta}_{\min,m}$}
{
Update $\bar{\Delta}_{\min,m}=\bar{\Delta}$\;
Update ${\bf n}_{{\rm optimal},m}={\bf n}$\;
}
}
\If{$\bar{\Delta}_{\min,m} > \bar{\Delta}_{\min,m-1}$}
{
Update $\bar{\Delta}_{\min}=\bar{\Delta}_{\min,m-1}$\;
Update ${\bf n}_{{\rm optimal}}={\bf n}_{{\rm optimal},m-1}$\;
break\;
}
\Else{Update $\bar{\Delta}_{\min}=\bar{\Delta}_{\min,m}$\;
Update ${\bf n}_{{\rm optimal}}={\bf n}_{{\rm optimal},m}$\;}
}
\Return ${\bf n}_{\rm optimal}$ and $\bar{\Delta}_{\min}$
\end{algorithm}
The design of such an algorithm is based on the empirical observation that the age-optimal $m$ is always monotonically decreasing with propagation delay (see Fig. \ref{Age-optimal block assignment}). Meanwhile, the age-optimal $m$ for reactive HARQ tends to remain small since large $m$ will lead to multiple RTT and thus result in \emph{staleness} of information. Attributed to the above factors, heuristically, we denote the minimal age under a fixed $m$ as $\bar{\Delta}_{\min,m}$ and recursively search the solution space $\mathcal{J}_m\triangleq\left\{\left(n_1, \cdots, n_i\right): i=m, n_{\min}<n_1< n_2 \cdots < n_i<n_{\max} \right\}$ with increasing value of $m$. The on-the-fly searching process will terminate only if $\bar{\Delta}_{\min,m}>\bar{\Delta}_{\min,m-1}$. By this means, the algorithm outputs $\bar{\Delta}_{\min,m-1}$ as the optimal age. In such a case, this algorithm eliminates the need for searching the sub-space $\cup_{i=m+1}^{n_{\max}-n_{\min}+1}\mathcal{S}_{i} \subseteq \mathcal{S}$, thereby bypassing the calculations required for searching for the whole solution space.
Note that an auxiliary set $ \mathcal{S}_m \triangleq \left\{{\bf p} \in \left\{0,1\right\}^{n_{\max}-n_{\min}+1}:\left\|{\bf p}\right\|_1=m\right\}$ is also introduced in Algorithm \ref{Algorithm 2} to assist high-efficiency searching, where $\left\|\cdot\right\|_1$ represents the implementation of $\ell_1$-norm.
\begin{lemma}\label{mappinghnp}
There exists an one-to-one mapping between ${\bf n}\in \mathcal{J}_m$ and ${\bf p} \in \mathcal{S}_m$.
\end{lemma}
\begin{proof}
The one-to-one mapping function between ${\bf n}\in \mathcal{J}_m$ and ${\bf p} \in \mathcal{S}_m$ is the same as that of Lemma \ref{mappingnp}, which has been discussed in Appendix \ref{appendixe} in detail.
\end{proof}
\section{Numerical Results}
\subsection{The Closed-Form Results}
\begin{figure*}[htbp]
\centering
\includegraphics[angle=0,width=0.98\textwidth]{./figures/Simu.eps
\caption{Instantaneous age evolution and the statistic characterizations among reactive HARQ, proactive HARQ and rateless codes. Here the encoding delay is $\tau_{\rm c}=2$, the decoding delay is $\tau_{\rm d}=3$, the propagation delay is $\tau_{\rm p}=5$ and the feedback delay is $\tau_{\rm f}=6$. }
\label{Simulation}
\end{figure*}
In addition to the case studies given in Section III. B, we also carry out Monte Carlo simulations to verify our closed-form expressions. For the simulation setup, we leverage an i.i.d uniformly distributed random sequence $\mathcal{X}_j\sim \mathcal{U}\left(0,1\right)$ to generate the feedback signal sequence $\varsigma_{j,i}, i\in [m]$ when transmitting message $M_j$. Specifically, the feedback signal sequence is generated by
\begin{equation}\nonumber
\varsigma_{j,i}=sign\left(\mathcal{X}_j-\epsilon_i\right) \text{, for } i\in[m],
\end{equation}
where $\epsilon_i$ is obtained by (\ref{Rateerr}) and $sign\left(\cdot\right)$ is defined as
\begin{equation}\nonumber
sign\left( x \right) = \left\{
{\begin{array}{*{20}{c}}
1\text{ , for } x\ge 0\\
0\text{ , for } x<0
\end{array}} \right..
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[angle=0,width=0.9\textwidth]{./figures/ReactivevsProactive.eps}
\caption{Reactive HARQ vs proactive HARQ. Here the message length is $k=100$, the encoding delay is $\tau_{\rm c}=20$, the decoding delay is $\tau_{\rm c}=30$, the propagation delay range is $\tau_{\rm p}=0:200$ and the feedback delay is calculated by $\tau_{\rm f}=\tau_{\rm p}+1$. For fairness, here we consider the finest grained block assignment vectors for both reactive HARQ and proactive HARQ, with $n_1=100, n_i=n_{i-1}+1$.}
\label{revspro}
\end{figure}
Then, with the feedback signal sequence $\varsigma_{j,i}$ in hand, the transmission-decoding model is almost sure, and we can recursively obtain the instantaneous age evolution as shown in Fig. \ref{Simulation}. For reactive HARQ and proactive HARQ, we set $k=100$, $m=11$ and $\mathbf{n}=100:110$. For rateless codes, we find that a sufficiently large value $m$ will directly lead to an almost convergent AoI. Thus, we set $m=10000$ for the simulation setup of rateless codes.
Fig. 4(a) demonstrates the instantaneous age evolution for reactive HARQ, proactive HARQ, and rateless codes, respectively. Intuitively, we can observe that the age of reactive HARQ tends to exhibit a number of large sawtooth waveforms, while that of proactive HARQ and rateless codes cut off the large sawtooth waveforms and keep at a relatively low level.
\begin{figure}[t]
\centering
\includegraphics[angle=0,width=0.9\textwidth]{./figures/Ageoptimal.eps
\caption{Average age comparisons among the finest grained reactive HARQ, the optimal reactive HARQ and the finest grained proactive HARQ. Here the message length is $k=100$, the encoding delay is $\tau_{\rm c}=20$, the decoding delay is $\tau_{\rm c}=200$, the propagation delay is $\tau_{\rm p}=50$ and the feedback delay is $\tau_{\rm f}=51$. }
\label{blockassignment}
\end{figure}
Fig. 4(b) and Fig. 4(c) depict the average AoI and average peak AoI comparisons between the simulation results and the analytical closed-form results, wherein the discrete orange points are obtained through Monte Carlo simulations, while the blue curves are plotted by utilizing the available closed-form results given in Section III. It can be seen that the simulation results fit well with the analytical results, verifying that our provided closed-form expressions enable exact and efficient AoI evaluations.
\subsection{Reactive HARQ vs. Proactive HARQ}
Fig. \ref{revspro} demonstrates the average AoI comparison between reactive HARQ and proactive HARQ from a multi-dimensional perspective. The comparisons are conducted among different settings of $m$, $\tau_{\rm p}$ and SNR. It is shown that the proactive HARQ surface remains below the reactive HARQ surface. Also, they intersects with each other at $m=1$. These numerical results are consistent with Corollary \ref{coro1}. In addition, Fig. \ref{revspro} also illustrates the impact that $\tau_{\rm p}$ and $m$ exert on average AoI. On the one hand, the average AoI is monotonically increasing with respect to the propagation delay $\tau_{\rm p}$. On the other hand, the impact of $m$ on average AoI could be complex: $i$) for proactive HARQ, retransmitting redundancy remains beneficial for AoI performance metric; $ii$) for reactive HARQ, however, retransmitting redundancy naturally brings about RTT and thus results in \emph{staleness} of information when the SNR is high enough to achieve reliable communication; in contrast, if the channel condition is poor, retransmitting redundancy is essential for reliable delivery, and thus may even compensate the AoI losses due to RTT.
\subsection{The Age-Optimal Block Assignment}
Fig. \ref{blockassignment} shows a comparison among the finest grained reactive HARQ, the optimal reactive HARQ, and the finest grained proactive HARQ. It is shown that the optimized reactive HARQ approaches proactive HARQ in average AoI performance. Notice that this gain lies in an adaptive block assignment strategy which requires accurate channel status information. In this regard, we find that adopting proactive HARQ for freshness-critical status update systems would be a robust and timeliness-efficient approach.
\section{Conclusion and Future Work }
In this paper, we have comprehensively considered different types of nontrivial system delay and derived unified closed-form average AoI and average Peak AoI expressions for both reactive HARQ and proactive HARQ. The unifying characteristic of our result has been shown by several case studies, wherein some existing PHY-layer AoI expressions in the literature are shown to be only some specific cases of our result. With these closed-form results in hand, we have theoretically proven that under the same communication conditions, the proactive scheme always outperforms the reactive scheme in terms of the average AoI. Also, a block assignment design framework at the PHY layer has been provided to further achieve timely delivery in a status update system. The simulation and analytical results demonstrate that the age-optimal block assignment strategy of reactive schemes is sensitive to both channel conditions and propagation delays, while that of proactive scheme exhibits both strategy robustness and age superiority. In this regard, we witness the potential for the proactive HARQ to be applied in the freshness-critical system.
The research in this paper also leaves some open challenges and issues for future research. First, it will be an interesting work to carry out AoI analyses and comparisons for some specific state-of-the-art channel coding techniques, such as polar codes, LDPC codes, Turbo codes, and rateless Raptor codes, etc. As such, from the AoI perspective, the trade-off among coding complexity, decoding complexity, codelength, the number of retransmitted packets, and the error probability can be explored. Second, since this work is based on an ideal assumption of perfect feedback, the analysis considering lossy feedback can be further conducted. Third, notwithstanding the AoI superiority of proactive HARQ compared to reactive HARQ, proactive HARQ may consume more energy due to the consecutive retransmissions. To this end, to further investigate the trade-off of proactive HARQ between timeliness and energy efficiency would be an interesting topic.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\appendices
\section{Proof of Lemma \ref{lemma1}} \label{appendixa}
The event $\left\{R_j=a\right\}$ is equivalent to
\begin{equation}\nonumber
\left\{R_j=a\right\}=\left\{\varsigma_{Q_{j},m}=1\right\}\bigcap_{i\in [a]}\left\{\varsigma_{i+Q_{j-1},m}=0\right\}.
\end{equation}
Note that the AWGN is i.i.d, we have that
\begin{equation}\nonumber
\begin{aligned}
\mathbb{P}\left(R_j=a\right)&=\mathbb{P}\left(\varsigma_{Q_{j},m}=1\right)\cdot\prod_{i\in [a]}\mathbb{P}\left(\varsigma_{i+Q_{j-1},m}=0\right)\\
&=\left(1-\epsilon_m\right)\epsilon^a_m.
\end{aligned}
\end{equation}
Also, the probability of event $\left\{V_j=i\right\}$ can be expressed as
\begin{equation}\label{Vj}
\mathbb{P}\left(V_j=i\right)=\frac{\mathbb{P}\left(\left\{\varsigma_{Q_j,m}=1, \varsigma_{Q_j,i}=1 \right\}\bigcap_{r\in [i-1]}\left\{ \varsigma_{Q_j,r}=0\right\}\right)}{\mathbb{P}\left(\varsigma_{Q_j,m}=1\right)}.
\end{equation}
As the variable $\varsigma_{Q_j,i}$ follows the monotonic property such that
\begin{equation}\nonumber
\begin{aligned}
&\left\{\varsigma_{Q_j,i}=1\right\}\subseteq\left\{\varsigma_{Q_j,m}=1\right\}, \text{\rm for } i \le m,\\
&\left\{\varsigma_{Q_j,i}=0\right\}\subseteq\left\{\varsigma_{Q_j,r}=0\right\}, \text{\rm for } r \le i.
\end{aligned}
\end{equation}
The event in (\ref{Vj}) can be simplified as
\begin{equation}\label{event}
\begin{aligned}
&\left\{\varsigma_{Q_j,m}=1, \varsigma_{Q_j,i}=1 \right\}\bigcap_{r\in [i-1]}\left\{ \varsigma_{Q_j,r}=0\right\}\\
&=\left\{\varsigma_{Q_j,i}=1,\varsigma_{Q_j,i-1}=0\right\}\\
&=\left\{\varsigma_{Q_j,i}=1\right\} /\left\{\varsigma_{Q_j,i-1}=1\right\}.
\end{aligned}
\end{equation}
Substituting (\ref{event}) into (\ref{Vj}} results in the probability as
\begin{equation}\nonumber
\mathbb{P}\left(V_j=i\right)=\frac{\epsilon_{i-1}-\epsilon_i}{1-\epsilon_m}.
\end{equation}
\section{Proof of Lemma \ref{lemma5}}\label{appendixb}
Define $G^{\rm Reac}_j=\sum_{t\in \mathcal{I}_j}\bar{\Delta}\left(t\right)$, we have that
\begin{equation}\nonumber
\begin{aligned}
&G^{\rm Reac}_j=\sum_{t=t^S_{j-1}}^{t^S_{j}-1}\left(t-t^S_{j-1}-\tau_{\rm f}+\tau^{\rm Reac}_{V_{j-1}}\right)\\
&\overset{x=t-t^S_{j-1}}{=}\sum_{x=0}^{T^{\rm Reac}_j-1}\left(x-\tau_{\rm f}+\tau^{\rm Reac}_{V_{j-1}}\right)\\
&=\frac{T^{\rm Reac}_j\left(T^{\rm Reac}_j-1\right)}{2}-\tau_{\rm f}T^{\rm Reac}_j+\tau^{\rm Reac}_{V_{j-1}}T^{\rm Reac}_j.
\end{aligned}
\end{equation}
Thus, we have the first moment of $S_j$ as
\begin{equation}\label{Sjmoments}
\mathbb{E}G^{\rm Reac}_j=\frac{\mathbb{E}\left(T^{\rm Reac}_j\right)^2}{2}+\mathbb{E}T^{\rm Reac}_j\left(\mathbb{E}\tau^{\rm Reac}_{V_{j-1}}-\tau_{\rm f}-\frac{1}{2}\right).
\end{equation}
With (\ref{Sjmoments}), we can obtain the average age as
\begin{equation}\nonumber
\bar{\Delta}_{\rm Reactive}=\frac{\mathbb{E}G^{\rm Reac}_j}{\mathbb{E}T^{\rm Reac}_j}=\frac{\mathbb{E}T^{\rm Proac}_j}{2\mathbb{E}\left(T^{\rm Reac}_j\right)^2}+\mathbb{E}\tau^{\rm Reac}_{V_j}-\tau_{\rm f}-\frac{1}{2}.
\end{equation}
\section{Proof of Lemma \ref{lemma8}}\label{appendixc}
Recall that $\epsilon_i$ a monotonically decreasing infinite sequence with $\epsilon_1>\epsilon_2>\epsilon_3>\cdots$, we can prove Lemma 8 by adopting the Dirichlet’s test. With Dirichlet’s test, this proof is equivalent to proving that the partial sums
\begin{equation}\label{partialsum}
\begin{aligned}
&\sum\limits_{i = 1}^{N} {\left( {{n_{i + 1}} - {n_i}} \right)},\\
&\sum\limits_{i = 1}^{N} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2{\tau _{\rm c}} + 2\mathcal{T} + {n_{i + 1}} + {n_i}} \right)}.
\end{aligned}
\end{equation}
are bounded. Evidently, we have the solutions to (\ref{partialsum}) as
\begin{equation}
\begin{aligned}
&\sum\limits_{i = 1}^{N} {\left( {{n_{i + 1}} - {n_i}} \right)}=n_{N+1}-n_1,\\
&\sum\limits_{i = 1}^{N} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2{\tau _{\rm c}} + 2\mathcal{T} + {n_{i + 1}} + {n_i}} \right)}=\left(n_{N+1}+\tau_{\rm c}+\mathcal{T}\right)^2-\left(n_1+\tau_{\rm c}+\mathcal{T}\right)^2.
\end{aligned}
\end{equation}
Thus, we have that the infinite series in (\ref{inseries}) are bounded.
\section{Proof of Corollary \ref{coro1}}\label{appendixd}
Subtract ${\bar{\Delta} _{\rm Proactive}}$ from ${\bar{\Delta} _{\rm Reactive}}$, we have
\begin{equation} \label{HARQvs}
\begin{aligned}
&{\bar{\Delta} _{\rm Reactive}} - {\bar{\Delta} _{\rm Proactive}} = \frac{{\mathcal{T}\sum\limits_{i = 1}^{m - 1} {{\epsilon_i}} }}{{1 - {\epsilon_m}}} + \\
&\frac{{{{\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T}} \right)}^2} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i} + \mathcal{T}} \right)\left( {2{\tau _{\rm c}} + {n_{i + 1}} + {n_i} + \left( {2i + 1} \right)\mathcal{T}} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i} + \mathcal{T}} \right){\epsilon_i}} } \right)}}\\
&- \frac{{{{\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T}} \right)}^2} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2{\tau _{\rm c}} + 2\mathcal{T} + {n_{i + 1}} + {n_i}} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right)}}\\
&\overset{(a)}{\ge} \frac{{\mathcal{T}\sum\limits_{i = 1}^{m - 1} {{\epsilon_i}} }}{{1 - {\epsilon_m}}} + \frac{{\mathcal{T}\sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right)\left( {2i - 1} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right)}} + \\
&\frac{{\mathcal{T}\sum\limits_{i = 1}^{m - 1} {\left( {2{\tau _{\rm c}} + {n_{i + 1}} + {n_i} + \left( {2i + 1} \right)\mathcal{T}} \right){\epsilon_i}} }}{{2\left( {{\tau _{\rm c}} + {n_1} + \mathcal{T} + \sum\limits_{i = 1}^{m - 1} {\left( {{n_{i + 1}} - {n_i}} \right){\epsilon_i}} } \right)}}\overset{(b)}{\ge} 0.
\end{aligned}
\end{equation}
\subsubsection{Proof of Sufficiency}
If $\mathcal{T}=0$, the equal signs at both $(a)$ and $(b)$ in (\ref{HARQvs}) are established ; if $m=1$, all those sums $\sum\nolimits_{i=1}^{m-1} {x_i}\equiv 0 $, and thus we have ${\bar{\Delta} _{\rm Reactive}} - {\bar{\Delta} _{\rm Proactive}}=0$.
Therefore, we prove the sufficiency that $ m=1$ or ${\mathcal{T}=0\Rightarrow \bar{\Delta} _{\rm Reactive}} = {\bar{\Delta} _{\rm Proactive}}$.
\subsubsection{Proof of Necessity}
Assume $m>1$ and $\mathcal{T}\ne0$, we can easily obtain from (\ref{HARQvs}) that ${\bar{\Delta} _{\rm Reactive}} - {\bar{\Delta} _{\rm Proactive}}>0$. This is equivalent to the necessity that ${\bar{\Delta} _{\rm Reactive}} = {\bar{\Delta} _{\rm Proactive}} \Rightarrow m=1$ or $ \mathcal{T}=0$.
\section{Constructing the Mapping in Lemma \ref{mappingnp}}\label{appendixe}
A constructing mapping from $\bf p$ to $\bf n$ is shown below:
Step 1: Find the indexes of all the zero-value positions of vector ${\bf p} = \left(p_1,p_2,\cdots,p_{n_{\max}-n_{\min}+1}\right)$, and store them in an empty set $\mathcal{A}$. For example, if $p_i=0$, then $i$ is stored into $\mathcal{A}$.
Step 2: Sort the elements in set $\mathcal{A}$ in the ascending order and denote the ordered elements as a vector $\left(a_1, a_2, \dots, a_{\left| {\mathcal{A}} \right|}\right)$, with $a_1 < a_2 < \cdots < a_{\left| {\mathcal{A}} \right|}$.
Step 3: The vector $\bf n$ can be obtained by
\begin{equation}
n_i=a_i+n_{\min}-1, i=1,2,\cdots,\left|\mathcal{A}\right|.
\end{equation}
The above process is reversible, and the mapping from $\bf n$ to $\bf p$ is elaborated below:
Step 1: Construct the vector ${\bf a} = \left(a_1,a_2,\cdots,a_{m}\right)$ by
\begin{equation}
a_i=n_i-n_{\min}+1, i=1,2,\cdots,m.
\end{equation}
Step 2: Initialize ${\bf p} = {\bf 1}^{n_{\max}-n_{\min}+1}$ and let $p_{a_i}=0, i = 1, 2, \dots, m$, the vector $\bf p$ is obtained.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:Intro}
Energetic partons created at early stage of heavy-ion collisions (HIC) may suffer energy loss due to interacting with the quark-gluon plasma (QGP), a novel state of matter with deconfined quarks and gluons under extreme high temperature and energy density. This
phenomenon is referred as jet quenching~\cite{Wang:1991xy,Gyulassy:2003mc,Qin:2015srf}, which could provide powerful tools to study the creation and properties of the QCD medium. In the last decade the investigations of jet quenching have been extended from the leading hadron productions suppression~\cite{Khachatryan:2016odn, Acharya:2018qsh, Aad:2015wga, Burke:2013yra,Chen:2010te,Chen:2011vt,Liu:2015vna,Dai:2015dxa,Dai:2017piq,Dai:2017tuy,Ma:2018swx,Xie:2019oxg,Zhang:2022fau} to medium modifications of a wealth of reconstructed jet observables, such as inclusive jets productions, di-jets asymmetry, correlations of gauge bosons and jets , as well as heavy flavor jets~\cite{Vitev:2008rz,Vitev:2009rd,Aad:2010bu,Chatrchyan:2011sx,Chatrchyan:2012gt,Aad:2014bxa,Chatrchyan:2012gw,Chatrchyan:2013kwa, Aad:2014wha,Sirunyan:2017jic, CasalderreySolana:2010eh, Sirunyan:2018ncy,Young:2011qx,He:2011pd,ColemanSmith:2012vr,Neufeld:2010fj,Zapp:2012ak, Dai:2012am, Ma:2013pha, Senzel:2013dta, Casalderrey-Solana:2014bpa,Milhano:2015mng,Chang:2016gjp,Majumder:2014gda, Chen:2016cof, Chien:2016led, Apolinario:2017qay,Connors:2017ptx,Zhang:2018urd,Dai:2018mhw,Luo:2018pto,Chang:2019sae,Wang:2019xey,Chen:2019gqo,Chen:2020kex,Wang:2020qwe,Yan:2020zrz,Wang:2020ukj,Zhang:2021sua}.
A full reconstructed jet is a collimated spray of hadrons created in $e^+e^-$ collisions, p+p reactions as well as nucleus-nucleus collisions with large momentum transfer, and the existence of QCD medium
should naturally modify the yields and the internal structures of full jets, and thus the medium modifications of jet observables could be used to tomography of QGP formed in HIC.
The nuclear modifications of jet substructure have received a growing attention in the heavy-ion community.
One interesting jet substructure is jet $p_{T}$ dispersion($p_{T}D$), which characterizes the fragmentation of a jet~\cite{Giele:1997hd,Acharya:2018uvf,KunnawalkamElayavalli:2017hxo,Agafonova:2019tqe,Wan:2018zpq}. The nuclear modification of jet $p_{T}D$ distribution may help us understand deeper the specific features of jet-medium interaction, and put new insight on how jet substructure is modified. Recently, ALICE Collaboration has measured $p_{T}D$ distributions for small-radius ($R=0.2$) jets in heavy-ion collisions~\cite{Acharya:2018uvf}, which further facilitates the studies of $p_{T}D$ in distributions HIC, since the theoretical calculations could be confronted with the data directly to infer some crucial information of jet propagation in the QCD medium.
In this paper, we present our study on the normalized $p_{T}D$ distributions for inclusive jets with jet radius $R=0.2$ both in p+p and Pb+Pb collisions at $\sqrt{s}=2.76$~TeV. We employ POWHEG+PYTHIA~\cite{Alioli:2010xa,Alioli:2010qp,Buckley:2016bhy}, a Monte Carlo model matching NLO matrix elements with parton shower (PS), including hadronization process to obtain the solid baseline results of jet $p_{T}D$ in p+p collisions, which are then served as input to simulate parton energy loss within the higher-twist approach~\cite{Guo:2000nz,Zhang:2003yn,Zhang:2003wk,Majumder:2009ge} to compute the $p_{T}D$ distribution in heavy-ion collisions. Our model calculations of $p_{T}D$ distribution for inclusive jets could provide satisfactory descriptions of ALICE data both in p+p and Pb+Pb collisions, where we observe a shift of $p_{T}D$ distribution toward higher values in Pb+Pb collisions relative to that in p+p. We further make a comprehensive understanding of the distinct feature between quark and gluon initiated jets, and the nuclear modification ratio $p_{T}D$ distribution. We find $p_{T}D$ can be analytical expressed as the standard deviation and the multiplicity of jet constitutes. After jet quenching, more jet constituents lied further from the mean value of $p_{T}$.
The remainder of this paper is organized as follows. In Sec.~\ref{sec:framework} we will introduce the framework used to calculate the normalized $p_{T}D$ distributions in both p+p and Pb+Pb collisions. Our numerical results and detailed discussions of the medium modifications of $p_{T}D$ distributions for groomed and ungroomed jets are presented in Section~\ref{sec:results}. In Sec.~\ref{sec:summary} we will give a summary.
\section{Analysis framework}
\label{sec:framework}
We study a jet substructure observable, the $p_{T}D$, which characterizes the second moment of the constituent $p_{T}$ distribution inside a jet~\cite{Giele:1997hd,Acharya:2018uvf}, and is defined
as:
\begin{eqnarray}
p_{T}D=\frac{\sqrt{\sum_{i}p^{2}_{T,i}}}{p_{T,{\rm jet}}}
\label{eq:g}
\end{eqnarray}
Where $p_{T,i}$ represents the transverse momentum of $i$th jet ingredient inside the jet with transverse momentum $p_{T,{\rm jet}}$. $p_{T}D$ is connected to
how hard or soft the jet fragmentation is, and whether $p_{T,i}$ distribution is uniform or not. For example, in the extreme case of very few constituents carrying a lion share of the jet momentum, $p_{T}D$ will be close to unity; while in the case of jets containing a large number
of constituents with soft momentum, $p_{T}D$ may approach to zero. It is noted that jet dispersion is one of
a class of jet substructure observables, the generalized jet angularities~\cite{Larkoski:2014pca,ALICE:2021njq}, which are defined as $\lambda^{\kappa}_{\beta}=\Sigma_{i}z^{\kappa}_{i}\theta^{\beta}_{i}$, where $z_{i}=p_{T,i}/p_{T,jet}$ is the momentum fraction of jet constituents, $\theta_{i}=\Delta{R_{i}}/R$, $\Delta{R_{i}}$ is the opening angle from constituent to jet axis, $\kappa$ and $\beta$ are free parameters. One can see that $(p_{T}D)^2$ is equal to $\lambda^{\kappa}_{\beta}(\kappa=2, \beta=0)$.
In this work, we use a Monte Carlo model POWHEG+PYTHIA, which performs next-to-leading order (NLO) matrix elements matched with parton showering~\cite{Alioli:2010xa,Alioli:2010qp,Buckley:2016bhy},
to generate jet productions in p+p collisions. In our simulation the POWHEG BOX code is utilized~\cite{Alioli:2010xa,Alioli:2010qp}, which provides a computer framework for performing NLO calculations in parton shower Monte Carlo programs in accordance with the POWHEG method~\cite{Frixione:2007vw}. Previous studies have shown that, POWHEG BOX Monte Carlo program matched with parton showering could give nice description of productions and correlations for a variety of processes in p+p collisions, such as di-jet, gauge boson tagged jets, heavy flavour jets etc.~\cite{powheg-box}. We generate the NLO matrix elements for QCD dijet events with POWHEG BOX, and then matched with PYTHIA6 to perform parton showering and hadronization~\cite{Sjostrand:2006za}. After that,
Fastjet package~\cite{Cacciari:2008gp} is employed to reconstruct final state hadrons into full jets.
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{disper-pp.pdf} \\
\caption{Normalized $p_{T}D$ distribution of inclusive jets in p+p collisions at
$\sqrt{s}=7$~TeV from POWHEG+PYTHIA calculation as compared with ALICE data~\cite{Acharya:2018uvf}.}
\label{g-ALICE}
\end{figure}
In order to compare with the available experimental data,
we selected events according to the same kinematic
cuts as adopted by the experimental measurements. In ALICE Collaboration data, jets are reconstructed using anti-$k_{T}$
algorithm with radius parameter $R=0.2$ from charged hadrons which are required to have $p_{T}>0.15$~GeV.
TThose reconstructed jets are accepted in transverse momentum range of $40 \ {\rm GeV} <p_{T, \rm jet}<60 $~GeV and rapidity range of $\left|\eta_{\rm jet}\right|<0.7$.
Our numerical results of normalized distributions of $p_{T}D$ in p+p collisions at
$\sqrt{s}=7$~TeV and their comparison with ALICE data are shown in Fig.~\ref{g-ALICE}. We can observe that, POWHEG+PYTHIA calculations show well agreements with experimental measurements in p+p collisions in the overall $p_{T}D$ region, which will be served as input for the subsequent study of nuclear modification in HIC. The $p_{T}D$ distributions of quark- and gluon-initiated jets are also plotted in Fig.~\ref{g-ALICE}, respectively. We find that at the same jet $p_T$, the peak of $p_{T}D$ distribution for gluon jets is located at smaller region relative to that for quark jets. It implies as compared to quark jets, gluon jets favor harder radiation, on average. To make a comprehensive understanding of the distinct feature between quark and gluon jets, we derived the $p_{T}D$ with standard deviation(labeled as $\sigma$) and the multiplicity of jet constitutes(labeled as $n$) in the following.
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{delta2.pdf} \\
\vspace{-1cm}
\caption{Top panel: normalized $\sigma^2$ distribution of quark- and gluon-initiated jets in p+p collisions at
$\sqrt{s}=7$~TeV from POWHEG+PYTHIA calculation; Bottom panel: ratio of normalized $\sigma^2$ distribution of gluon and quark jets.}
\label{delta2-qg}
\end{figure}
The standard deviation could describe the average degree of a dataset. It tells us, on average, how far each value lies away from the mean value.
A high standard deviation signifies that values are generally far from the mean value, while a low standard deviation means that values are clustered near to the mean value.
Aa a result of that, in our study, the standard deviation of transverse momentum of jet constituents can be written as:
\begin{eqnarray}
\sigma=\frac{\sqrt{\sum_{i}(p_{T,i}-\langle p_{T,i} \rangle)^2}}{p_{T,{\rm jet}}}
\label{eq:g2}
\end{eqnarray}
Then $\sigma^2$ can be expressed as:
\begin{eqnarray}
\sigma^2 &=& \frac{\sum_{i}(p_{T,i}-\langle p_{T,i} \rangle)^2}{(n \cdot \langle p_{T,i} \rangle)^2} \nonumber \\
&=& \frac{\sum_{i}(p^{2}_{T,i}-2p_{T,i}\langle p_{T,i} \rangle + \langle p_{T,i} \rangle^2)}{(n \cdot \langle p_{T,i} \rangle)^2} \nonumber \\
\label{eq:g3}
&=& (p_{T}D)^2-1/n
\label{eq:sigma}
\end{eqnarray}
Conversely, we have $$(p_{T}D)^2=\sigma^2+1/n \, .$$ Shown in Fig.~\ref{delta2-qg} are the normalized $\sigma^2$ distributions (top) of quark and gluon jets in p+p collisions at $\sqrt{s}=7$~TeV, as well as their gluon/quark ratios (bottom). We observed more gluon jets distributed in lower $p_{T}D$ region compared to quark jets.
It is because gluon jets contain more fragment ingredients. Thus, at the same energy, the value of standard deviation for gluon jets is smaller than that for quark jets.
In heavy-ion collisions, fast partons produced from hard scattering will interact with medium partons and lose their energy. In our calculations, the initial jet shower partons are generated by POWHEG+PYTHIA,
then they are arranged to have initial positions which are sampled from Glauber model~\cite{Alver:2008aq}.
We assume all the partons move in QGP as the same of classical particles. The probability for gluon
radiation happens in QGP during each time step $\Delta t$ can be expressed as~\cite{He:2015pra,Cao:2016gvr,Cao:2017hhk,Wang:2019xey}:
\begin{eqnarray}
P_{rad}(t,\Delta t)=1-e^{-\left\langle N(t,\Delta t)\right\rangle} \, .
\label{eq:g4}
\end{eqnarray}
Here $\left\langle N(t,\Delta t)\right\rangle$ is the averaged number of emitted gluons, which can be integrated
from the medium induced radiated gluon spectrum within Higher-Twist(HT) method~\cite{Guo:2000nz,Zhang:2003yn,Zhang:2003wk,Majumder:2009ge}:
\begin{eqnarray}
\frac{dN}{dxdk^{2}_{\perp}dt}=\frac{2\alpha_{s}C_sP(x)\hat{q}}{\pi k^{4}_{\perp}}\sin^2(\frac{t-t_i}{2\tau_f})(\frac{k^2_{\perp}}{k^2_{\perp}+x^2M^2})^4
\label{eq:g5}
\end{eqnarray}
Here $\alpha_{s}$ denotes the strong coupling constant, $x$ is the energy fraction of the radiated gluon, $M$ is the mass of parent parton, and $k_\perp$ is the $p_{T}$ of the radiated gluon.
A a lower $p_{T}$ cut-off with $x_{min}=\mu_{D}/E$ of the emitted gluon is applied in our calculations, and $\mu_{D}$ is the Debye screening mass.
$P(x)$ is the parton splitting function in vacuum, $C_s$ is the Casimir factor for gluons ($C_A$) and quarks ($C_F$). The formation time of the radiated gluons is $\tau_f=2Ex(1-x)/(k^2_\perp+x^2M^2)$.
$\hat{q}$ is jet transport parameter, which is proportional to the local parton distribution density in the QCD medium,
The jet transport parameter $\hat{q}$ is proportional to the local parton density distribution in the QCD medium, and related to the space and time evolution of the medium relative to its initial value $\hat{q}_0$ in the central region when QGP formed, which controls the magnitude of energy loss due to jet-medium interaction.
The number of emitted gluons is sampled from a Poisson distribution during each time step.
To calculate the collisional energy loss of these showered partons\cite{Dai:2018mhw,Wang:2019xey}, a Hard Thermal Loop (HTL) formula has been adopted in this work:~\cite{Neufeld:2010xi}
$\frac{dE^{coll}}{dt}=\frac{\alpha_{s}C_{s}\mu^{2}_{D}}{2}ln\frac{\sqrt{ET}}{\mu_{D}}$.
The space time evolution of bulk medium is given by
the smooth iEBE-VISHNU hydrodynamical code~\cite{Shen:2014vra}.
Jet partons stop their propagation in QGP medium when local temperature fall below ($T_c=165$~MeV).
After all the partons escape from QGP, PYQUEN method is used to carry out the hadronization process~\cite{Lokhtin:2000wm,Lokhtin:2005px}. In the model, the radiated gluons are rearranged in the same string as their parent partons,
and these partons could fragment into hadrons
by standard PYTHIA hadronization procedure.
\label{sec:results}
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{disper-aapp.pdf} \\
\caption{Normalized $p_{T}D$ distributions of inclusive jets in p+p and Pb+Pb collisions at $\sqrt{s}=2.76$~TeV as compared with ALICE data~\cite{Acharya:2018uvf}.}
\label{g-AA}
\end{figure}
\section{Results and discussion}
\subsection{$p_{T}D$ distributions of ungroomed jets in Pb+Pb collisions}
We now could calculate
the jet number normalized $p_{T}D$ distributions in Pb+Pb collisions at $\sqrt{s}=2.76$~TeV.
We use the same jet selection criterion as we did in p+p collisions in Sec.~\ref{sec:framework}.
Our numerical results of jet number normalized $p_{T}D$ distributions for inclusive jets in p+p and Pb+Pb collisions
are shown in Fig.~\ref{g-AA}, which are confronted with the existing
experimental data in Pb+Pb by ALICE Collaboration~\cite{Acharya:2018uvf}. We find our theoretical calculations could provide quite decent descriptions of experimental measurements.
Relative to that in p+p collisions, the observed normalized $p_{T}D$ distribution in Pb+Pb is shifted towards higher values.
That is to say, $p_{T}D$ distribution of inclusive jets in Pb+Pb collisions is shifted towards quark jets after jet quenching.
It indicate that a jet in Pb+Pb collisions may has more softer constituents than that in p+p.
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{R-disper.pdf} \\
\caption{Nuclear modification ratio of $p_{T}D$ distribution of inclusive jets as well as quark and gluon jets. The ratio from ALICE Collaboration is preformed by Pb+Pb measurements scaled by MC simulation in p+p~\cite{Acharya:2018uvf}.
}
\label{ratio-jets}
\end{figure}
To investigate the deviation of jet $p_{T}D$ distributions in HIC to those in p+p in a more straightforward way,
it is essential and helpful to define
the nuclear modification ratio($R_{AA}^{\rm p_{T}D}$) of $p_{T}D$ distributions as:
\begin{eqnarray}
R_{AA}^{\rm p_{T}D}={\dfrac{1}{N_{AA}} \dfrac{dN_{AA}}{dp_{T}D}}/{\dfrac{1}{N_{pp}} \dfrac{dN_{pp}}{dp_{T}D}} \,\, .
\label{eq:ratio}
\end{eqnarray}
Shown in Fig.~\ref{ratio-jets} are $R_{AA}^{\rm p_{T}D}$ of normalized $p_{T}D$ distributions for inclusive jets, as compared with ALICE ratio preformed form Pb+Pb measurements scaled by MC simulation in p+p~\cite{Acharya:2018uvf}. Our calculated results could provide nice description of ALICE ratio in the overall $p_{T}D$ region. The $R_{AA}^{\rm p_{T}D}$ of quark jets and gluon jets as the components of inclusive jets are also plotted in Fig.~\ref{ratio-jets}. One can observe there is a suppression of $p_{T}D$ distribution for both quark jets and gluon jets at low $p_{T}D$ region,
while an enhancement at high $p_{T}D$ region. The nuclear corrections of gluon jets $p_{T}D$ distribution is much stronger than that for quark jets. The curve of $R_{AA}^{\rm p_{T}D}$ for inclusive jets goes between the curves of $R_{AA}^{\rm p_{T}D}$ for quark jets and gluon jets, since inclusive jets are the combinations of quark jets and gluon jets.
To further understand the nuclear modification mechanism of $p_{T}D$ distribution for inclusive jets, we start with the modifications of the relative fraction of quark and gluon jets due to jet quenching.
In our calculations, as most of other jet quenching models, gluons may lose more energy than quarks in QGP with their larger color charge. Therefore, generally we should see an enhancement of contribution fraction of quark jets in A+A collisions relative to that in p+p. Qualitatively such enhancement will naturally lead $p_{T}D$ distributions for inclusive jets to larger value region, since quark jets peak at larger value of $p_{T}D$ than gluon jets.
This is illustrated in Fig.~\ref{ratio-jets}, where the curve labelled ‘inclusive jet (q/g)’ represents our numerical result of $R_{AA}^{\rm p_{T}D}$ by only considering the effect of quark/gluon jets fraction alterations due to jet quenching while assuming there are no medium modifications for $p_{T}D$ distributions of pure quark and gluon jets in heavy-ion collisions.
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{delta2aa.pdf} \\
\vspace{-1cm}
\caption{Top: normalized $\sigma^2$ distribution of inclusive jets in Pb+Pb and p+p collisions at
$\sqrt{s}=2.76$~TeV; Bottom: ratio of normalized $\sigma^2$ distribution in Pb+Pb and p+p collisions.}
\label{delta2aa}
\end{figure}
\begin{figure}[!htb]
\centering
\vspace{0.in}
\includegraphics[width=9.5cm,height=9.6cm]{z-ratio.pdf} \\
\vspace{0.in}
\caption{Nuclear modification ratio of z distributions for inclusive jets at $\sqrt{s}=2.76$~TeV as compared with ATLAS data~\cite{Aad:2014wha}.}
\label{zratio}
\end{figure}
To explore why for normalized jet $p_{T}D$ distributions $R_{AA}^{\rm p_{T}D} < 1$ at small $p_{T}D$ region, and $R_{AA}^{\rm p_{T}D} > 1$ at large $p_{T}D$ region, as shown in Fig.~\ref{ratio-jets}. Eq.(\ref{eq:sigma}) indicate the nuclear modification of $p_{T}D$ distributions has a very strong correlation with the nuclear correction of standard deviation of $p_{T,i}$.
Thus we turn to investigate the standard deviation of $p_{T,i}$ for jets in p+p and Pb+Pb collisions. Presented in Fig.~\ref{delta2aa} are the normalized distribution of standard deviation of $p_{T,i}$ in p+p and Pb+Pb collisions at $\sqrt{s}=2.76$~TeV. We can observed the distributions of jet standard deviation is also shifted to higher value region in heavy-ion collisions compared with that in p+p. Which indicates after jet quenching, the value of $p_{T,i}$ lies further from the mean value in heavy-ion collisions relative to p+p. It is noted that the changes of mean multiplicities of Pb+Pb relative to p+p is rather small, and
the estimated mean values of jet constituents number ($\bar{n}$) in p+p and Pb+Pb collisions are $\bar{n}_{pp}=6.72$ and $\bar{n}_{PbPb}=6.54$ respectively.
To see the point more clearly, we also plotted the nuclear modification ratio of momentum fraction for jet constituents($z=p_{T,i}/p_{T, \rm jet}$) in Fig.~\ref{zratio}. One can see, our model calculations could provide nice description of ATLAS data for jets with $p_T>100$~GeV~\cite{Aad:2014wha}. For jets with $40<p_T<60$~GeV, the nuclear modification ratios of charged-particle transverse momentum distributions in Pb+Pb collisions to those measured in p+p exhibit an enhancement in fragment yield in central collisions for $0.02<z<0.05$,
a reduction in fragment yields for $0.05<z<0.3$,
and an enhancement in the fragment yield for $0.3<z<1$. Which means the number of jet constituents with lower and higher value of $p_{T}$ are enhanced, more constituents lied further from the mean value. Therefore, the jet standard deviation is shifted to higher region in heavy-ion collisions.
\subsection{$p_{T}D$ distributions of groomed jets in Pb+Pb collisions}
In this section, we will study $p_{T}D$ distributions of groomed jets in Pb+Pb collisions. Jet grooming techniques have seen a particularly great of interest from both experimental and theoretical side~\cite{Larkoski:2014wba,Dasgupta:2013ihk}. It is designed to remove soft wide-angle radiation from the jet,
allowing for a more direct comparison between experimental data and purely perturbative QCD calculations, since hadronization and underlying event contributions are significantly reduced during grooming procedure.
A full jet constructed using radius R via the anti-$k_{T}$ algorithm is first re-clustered using the Cambridge-Aachen (C/A) algorithm~\cite{Dokshitzer:1997in,Wobisch:1998wt} until two hard branches are found to
to satisfy the following condition:
\begin{eqnarray}
\frac{min(p_{T1},p_{T2})}{p_{T1}+p_{T2}}\equiv{z_{g}}>z_{cut}\left (\frac{\Delta R}{R} \right)^{\beta}
\label{eq:sd}
\end{eqnarray}
where $\left (\frac{\Delta R}{R} \right)$ is an additional parameter of the relative angular distance between the two sub-jets, $z_{cut}$ and $\beta$ are free
parameters which can be used to control how strict the soft drop condition is. For the heavy-ion studies conducted so far, $z_{cut}$ has been set to $0.1$ and $\beta$
has been set to zero.
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{groomed.pdf} \\
\vspace{-1cm}
\caption{Top: nuclear modification factor of $p_{T}D$ distribution for groomed and ungroomed jets; Bottom: ratio of $p_{T}D$ distribution for groomed and ungroomed jets in p+p and Pb+Pb collisions.}
\label{delta2aag}
\end{figure}
In the top panel of Fig.~\ref{delta2aag}, we plot the nuclear modification ratio of $p_{T}D$ distributions for groomed and ungroomed jets respectively. One can observe that the nuclear modification pattern
of $p_{T}D$ distributions for groomed jets are similar with that for ungroomed jets, and
the $p_{T}D$ distribution for groomed jets
is also shifted to higher $p_{T}D$ region. Besides, compared with ungroomed jets,
the nuclear modification of groomed jets becomes weaker. It implies grooming procedure could not only remove soft radiation from the jet in QCD vacuum, but also reduce soft radiation
in QCD medium. In the bottom panel of Fig.~\ref{delta2aag} we present the ratios of $p_{T}D$ distributions for the groomed jet
to that for
the ungroomed both in p+p and Pb+Pb collisions. It is shown that these ratios are below to unity at small $p_{T}D$, whereas
larger than one at large $p_{T}D$. To understand deeper the alteration of $p_{T}D$ distributions originate from jet grooming procedure, in the following we may investigate the difference of standard deviation of jet constituents $p_{T,i}$ and jet constituents number between groomed and ungroomed jets in Pb+Pb collisions.
\begin{figure}[!htb]
\centering
\includegraphics[width=9.5cm,height=9.6cm]{sigma-PbPb.pdf} \\
\vspace{-1cm}
\caption{Top: normalized $\sigma^2$ distribution of groomed and ungroomed jets in Pb+Pb collisions at
$\sqrt{s}=2.76$~TeV; Bottom: ratio of normalized $\sigma^2$ distribution for groomed and ungroomed jets in Pb+Pb collisions.}
\label{delta2aa2}
\end{figure}
Firstly, in Fig.~\ref{delta2aa2} we plotted the standard deviation ($\sigma^{2}$) distributions of jet constituents $p_{T,i}$ for groomed and ungroomed jets in Pb+Pb collisions. The distributions of jet $\sigma^{2}$ are shifted to lower region for groomed jets compared with that for ungroomed jets, which is in contrast to the alteration of $p_{T}D$ distributions. It indicates that after jet soft-drop procedure, the values of $p_{T,i}$ in groomed jets lies closer to the mean value relative to those in ungroomed jets. That is because during the grooming, some particles with low $p_{T,i}$ are dropped from the jet constituents.
\begin{figure}[!htb]
\includegraphics[width=9.5cm,height=9.6cm]{number-PbPb.pdf} \\
\vspace{-1cm}
\caption{Top: normalized (1/n) distribution of groomed and ungroomed jets in Pb+Pb collisions at
$\sqrt{s}=2.76$~TeV; Bottom: ratio of normalized (1/n) distribution for groomed and ungroomed jets in Pb+Pb collisions.}
\label{nn2aa}
\end{figure}
Secondly, the number of jets constituents are modified during soft drop grooming process, which will be contribute to the correction of $p_{T}D$ distributions. Presented in Fig.~\ref{nn2aa} are the number of jets constituents distributions of groomed and ungroomed jets in p+p and Pb+Pb collisions. One cam see the value of $1/n$ is enhanced after grooming process. As we shown in Eq.~3, the value of $p_{T}D$ is equivalent to the standard deviation added by $1/n$. Therefore, even though the grooming process will lead to lower value of standard deviation for jet constituents transverse momenta, it will enhance the value of $1/n$ meanwhile. Though these two effects offset with each other, the correction of $1/n$ are more pronounced, which results in the increase of the ratio of $p_{T}D$ distributions for groomed jets to that for ungroomed jets with $p_{T}D$ (as shown in the bottom panel
of Fig.~\ref{delta2aag}), a trend similar to the increase of ratio of normalized (1/n) distribution for groomed to ungroomed jets with $1/n$ (as demonstrated in the bottom panel of Fig.~\ref{nn2aa}).
\section{Summary}
\label{sec:summary}
In this paper, by using a NLO$+$PS event generator POWHEG$+$PYTHIA for p+p baseline and HT parton energy loss approach for jet quenching, we have studied
the nuclear modifications of $p_{T}D$ distributions for inclusive jets with small radius $R=0.2$ in Pb+Pb at $\sqrt{s}=2.76$~TeV.
Our simulated results of inclusive jets could provide decent description of ALICE measurements. The $p_{T}D$ distributions for inclusive jets
are shifted toward higher $p_{T}D$ region in Pb+Pb collisions compared to those in p+p, and similar trends have also been found for quark and gluon jets.
We further find two elements could contribute to the nuclear modifications of $p_{T}D$ distributions: more uneven $p_T$ of jet constituents, and the enhanced fraction of quark-initiated jets after jet-medium interaction in HIC.
The observed nuclear modifications of $p_{T}D$ distributions for gluon jets are stronger than that for quark jets in HIC
since gluons may lose more energy than quarks in our model.
Additionally, we also investigate the medium modifications of $p_{T}D$ distributions for groomed jets in Pb+Pb collisions. We observe weaker nuclear modifications of $p_{T}D$ distributions for groomed jets
compare to that for ungroomed jets.
{\bf Acknowledgments:} The authors would like to thank P Ru, S L Zhang, S Wang and Q Zhang for helpful discussions.This research was supported in
part by Guangdong Major Project of Basic and Applied Basic Research No. 2020B0301030008, by Natural Science Foundation of China (NSFC) under Project
Nos. 11935007 and and 12035007. S.-Y. Chen is supported by the MOE Key Laboratory of Quark and Lepton Physics (CCNU) under Project
No. QLPL2020P01.
\vspace*{-.6cm}
|
\section{Introduction}
An agile approach to projects has become extremely popular both within \citep{dingsoyr20121213}, and outside software development \citep{pajares2017project}. In contrast to traditional plan-driven development (a waterfall process), this paradigm has been given the name \emph{agile}, which literally means to be able to move quickly, easily and be flexible \citep{williams}.
Feasible solutions to customer needs should be developed in close customer collaboration, rather than being subject to the uncertainties of large up-front design and planning \citep{isleanagile}. With the explosion of software in the car, this complexity has forced automotive companies to, not just focus on customer value, but to also to rely on teams, and teams of teams, to develop modern cars with around 150 Million lines of code. This implies that the social aspect of group dynamics need to be understood in order to scale the development projects. The car projects also differ in that the agile approaches are used across both software, hybrid, and pure hardware teams. This study, though, focuses on the software teams in this quite different context since the $\approx 800$ teams essentially build one single product together.
Research on automotive software engineering has mainly focused on testing and building software models \citep{haghighatkhah2017automotive}. We have not been able to find many studies on organizational psychology in general except studies on ``human factors'' associated with individuals (e.g., \citet{maro2017challenges}). We have not been able to find any studies on social psychology in general, or team maturity in particular, in automotive software development.
Some studies have been conducted on self-organization in general agile software development (i.e., not automotive) teams in a broader software engineering context, also referred to as team autonomy. \citet{moe2008} concluded that one of the barriers to agile team autonomy was a lack of system support for teams and reduced external autonomy.
More recently, \citet{hodgson2013controlling} investigated how teams adopt agile practices over time. They conclude that some teams do a ``big bang'' while other implement agile practices one-by-one. They saw a trend that team members without much previous agile knowledge prefer the latter while experienced agile practitioners prefer the former. Volvo Cars has gone through an agile transformation in the last five years and there is a strong push towards providing teams with the mandate to lead themselves to a larger extent, i.e., a ``big bang'' on a very large scale impacting close to 800 teams.
\citet{hoda2011supporting} showed that senior management have a key role in enabling team agility. That there is a challenge in fitting the agile approach into traditional management in a software engineering context was also shown by \citet{hodgson2013controlling}. In more general leadership research, there are traits that a good leader must have, but on top of that, good leaders in all fields adapt their leadership style dynamically to the situation \citep{2009tei}. Leadership could also be seen as a function of group action \citep{von1986leadership} and not a role, which makes it possible to share or change over time. This means that there is no best leadership style since many different styles are needed depending on the context, the group, and the people involved. In the agile space, e.g., servant leadership is advocated as the foundation for leading, but the definition of what that is remains vague \citep{parris2013systematic}. Leadership is a difficult construct to research since it happens in complex systems and is influenced by a vast number of factors.
\citet{grenjss2} found initial indications confirming that the definition of agile teams overlaps with what is meant by a mature group in social psychology \citep{wheelan}. Agile teams are dependent on also being mature in terms of collaborative behaviur plus having navigated through the initial stages of group development. Mature teams are also known to collaborate better with other teams \citep{wheelandev}, which is of interest when scaling the agile ways of working at Volvo Cars. We define team maturity in this study as the degree to which a team has navigated through the group development stages according to \citet{wheelandev}. In Wheelan's Integrated Model of Group Development (or IMGD), small groups (or teams) start in the Dependency and Inclusion stage where team members are tentative and polite in their communication and need to build an initial level of psychological safety. Team members do not have a common mental model of the group goal nor do they know the real competencies of the other team members. When higher levels of psychological safety start to emerge, the team transitions into the second phase, Counter-Dependency and Fight. This is a conflict stage but is absolutely necessary to determine who can contribute to what and create a shared mental model of the team's purpose. After the second stage, the team can organize work better and set productive group norms. With higher levels of trust and a continuous improvements of work organization, the team can reach stage four, which is the Work and Productivity stage \citep{wheelandev}.
Leadership, self-organization, and agility can thus be mapped onto the team development stages \citep{grenjss2}. As the development cycle moves along from a psychological perspective, all members of the team (including the managers or leaders) must change their behavior \citep{wheelandev}. Team maturity, as measured by the Group Development Questionnaire (GDQ), has been shown to be correlated to measurements of productivity also in software engineering \citep{al2018connections}. In this study, we used a recently developed short version of the GDQ comprising 13 items instead of 60 \citep{gren2020gdqs}, which makes it practically more useful considering the time availability of teams.
If the results on the connection between team maturity and team productivity from social psychology \citep{wheelandev} apply to automotive software development, aspects of productivity should also be connected to measurements of team maturity in that context. One quite objective proxy for productivity for these Volvo Cars teams was their issue management, i.e., how long the teams take to fix issues reported back from their deliveries. This metric is external to the team, which makes it in many ways more appropriate than e.g., flow of backlog items or self-assessed productivity.
\section{Method}
We distributed the short version of the GDQ, (the GDQS \citep{gren2020gdqs}) to 14 teams engaged in software-only development at Volvo Cars. The teams also logged the number of issues they have with the delivered software in the same month. Issues could be for example bugs or other types of change requests. The number of issues can accumulate since many of them remain unaddressed so it may not a good measurement of the current team performance. We, therefore, instead looked at the average age of issues in the same month as the teams filled out the GDQS. The average is sensitive to outliers but Volvo Cars chose that metric as a KPI since all issues need to be resolved. 82 team members answered the GDQS from all the 14 teams. Since this is an average of responses per team and the teams are usually from 3 to 12 individuals, we estimate this response rate to be around 80\%.
We describe the data and its patterns or lack thereof and avoid using $p$ values for null hypothesis significance testing in accordance with \citet{mcshane2019abandon}. We instead use an approach based on Bayesian Data Analysis (BDA), i.e., we will plot the re-sampled (using Markov-Chain Monte Carlo) distribution of all our parameters and be explicit about our assumptions (see supplementary material). For a description and tutorial applied to software engineering see \citet{furia2019bayesian}.
Note all the shared data in this paper, plots and supplementary material were transformed using arbitrary linear transformations. This was done so as to not reveal the absolute values of issues and team maturity, since that was deemed as sensitive information by the company. The average ages of issues where downloaded from the Volvo Cars instance of Jira and the team maturity data was collected through the company's own survey tool.
In the second step, we interviewed the Release Train Engineer (RTE), which is a role that has an overview of this whole set of teams. The interviewee was asked to describe the situation around the data collection and also critically assess the connection between team maturity and the teams' capacity to resolve reported issues.
Based on previous studies of team maturity and aspects of productivity (see Section~1), we hypothesize that the first two stages (Stage 1 and 2) should be positively connected to the average age of issues, i.e., the more collaborative aspects to solve in these stages, the longer it would take to solve reported issues. We would also hypothesize that the latter stages (Stage 3 and 4) would have the opposite relationship to reported issues, since higher values on these two scales imply more maturity (e.g., higher levels of trust etc.).
\section{Results}
\subsection{Quantitative results}
The result shown in Fig.~\ref{corrplots} indicates that the direction of the lines are in the hypothesized direction. However, Fig.~\ref{mcmc} shows that only the distributions of stages 1 and 2 are distinct from zero with close to a 95\% credible interval.
\begin{figure}
\subfloat{\includegraphics[width = 1.75in]{corrstage1final.pdf}}
\subfloat{\includegraphics[width = 1.75in]{corrstage2final.pdf}}\\
\subfloat{\includegraphics[width = 1.75in]{corrstage3final.pdf}}
\subfloat{\includegraphics[width = 1.75in]{corrstage4final.pdf}}
\caption{Scatter plots with each with one of the developmental stages of team maturity as independent variable (the x axis) and the average age of issues as dependent variable (the y axis).}
\label{corrplots}
\end{figure}
\begin{figure}
\subfloat[Stage 1]{\includegraphics[width = 1.75in]{stage1trans.pdf}}
\subfloat[Stage 2]{\includegraphics[width = 1.75in]{stage2trans.pdf}}\\
\subfloat[Stage 3]{\includegraphics[width = 1.75in]{stage3trans.pdf}}
\subfloat[Stage 4]{\includegraphics[width = 1.75in]{stage4trans.pdf}}
\caption{Re-sample distributions of four separate models, each with one of the developmental stages of team maturity as independent variable and the average age of issues as dependent variable.}
\label{mcmc}
\end{figure}
As mentioned, we do not want to use classical NHST inference, which means that we see this quantitative result as indicative, but further studies are needed. We would like to, instead, highlight the importance of the qualitative data below.
In summary, when the teams are newly formed (or less mature) or in the conflict stage of creating shared mental models of the team goals and figuring out who can really contribute with what in order to fulfill these goals, the average age of issues is higher for those teams.
\subsection{Qualitative validation}
Below is a summary of what the RTE commented upon in the interview.
During the time of the data collection, the teams were focusing on removing issues that were never to be dealt with. The team developed their software in-house and wanted to remove issues they had inherited from suppliers. The teams were working on two different platforms, one old and one new. Different teams were working on the different platforms. The teams working on the old platforms were described as having lower motivation for their work compared to the teams working on the new platform. The teams that worked on the old platform would have preferred to work on the new platform instead. The teams working on the new platform used to even be located at a different municipality.
The performance and collaboration of the teams were described as very different. Regarding their reported issues, some teams could solve those internally while other teams were highly dependent on suppliers. For some teams, solving issues was a complex task with many steps.
The teams working on the old platform had closed a high number of issues, other teams received reports of defects that were incorrectly perceived as being errors.
The teams working on the old platform were assessed as having lower team maturity compared to the ones working on the new platform. They closed a lot of issues that they realized would never be dealt with. Fixing issues from problem reports was a part of all the teams' daily work.
As a form of validation, the team with the highest values on the y axis in Fig.~\ref{corrplots}, was the team the interviewee identified as the least mature team (denoted as a triangle).
\section{Discussion}
The results of this study suggest that the automotive software development having transformed into agile ways of working, might be dependent on team maturity in order to deliver value in the shape of fixing issues in their delivery. Since the team as a unit of analysis lies at the heart of an agile organization, research on automotive agile software development should also use theories and models from social psychology in order to understand the mechanisms at play.
On the necessary journey towards being more responsive to change, Volvo Cars has been on an agile transformation for the last five years. Some parts of the company has gotten farther than others at the time of this research, but the vehicle software and electronics department were the forerunner in the transformation.
Now that the automotive industry has software as a key to gaining a competitive advantage, research on software engineering adapted to the automotive industry is needed and conducted to understand this somewhat specific context. The psychological aspects of automotive software engineering has, though, not been much in focus and this study is a tiny step in beginning to fill that gap. Research so far has focused on testing and building software models \citep{haghighatkhah2017automotive} and the human factors investigated have been associated with individual psychology (e.g., \citet{maro2017challenges}). Since the aspects of agile team autonomy now also surfaces in the automotive software development, we need to investigate this context further also with regards to self-organization. \citet{moe2008} concluded, among other things, that one of the barriers of agile team autonomy was a lack of system for team support. This is an aspect identified in the agile transformation at Volvo Cars as well and the participating teams (and many others) were trained for at least three hours in Wheelan's Integrated Model of Group Development. They were also given access to a toolbox connected to the different stages after having filled out the GDQS, which hopefully helped them maturing further as a team, but this remains to be investigated.
\subsection{Practical implications}
We have assumed that the team maturity would be connected to aspects of productivity in the automotive software development at Volvo Cars, but this current study is the first one showing data on this actually being the case. If the results of this study would hold for automotive software engineering in general, it has a multitude of implications for developing such software. Many of the challenges go hand-in-hand with the agile transformation, but also provides way of getting teams to mature faster. One of these implications is to focus even more on training leaders and managers (in addition to training teams, of course) in adapting their leadership style dynamically to the team maturity level \citep{2009tei}. A team will not self-organize on day one because the team members are unlikely to have met before. Also there are the leadership challenges of designing teams well and assuring the emergence of good work principles and norms \citep{magdalucasxp}, instead of micromanaging technical details. Agile leadership has been described as ``servant leadership,'' which remains vague \citep{parris2013systematic}. But a key aspect of thinking about servant leadership is that it challenges many managers' mental model of leadership being based on hierarchical power, however, agile leadership need to be a servant in a way that serves the teams with what they actually need right now. That could also be to confront behavior that does not build psychological safety, as an example \citep{agileleadicse}.
This current study is in line with the result by \citet{al2018connections} that team maturity as measured by the GDQ is connected to productivity aspect of software development, but extends it into the automotive software development domain. A very practical implication for the automotive industry is to measure team maturity in a short validated survey, such as the GDQS \citep{gren2020gdqs}, train both managers, agile leaders, and team members in team development, by for example using the Integrated Model of Group Development \citep{wheelan} in order to increase the speed of fixing issues in the teams' deliveries. We infer this causality by logic and the qualitative data, since it does not make sense to hypothesize that fixing issues fast would lead to the team maturing, however, the mediating role of team maturity between team building interventions and the speed of fixing issues, of course needs much more research.
\subsection{Threats to validity}
We use relevant parts of the checklist created by \citet{gren2018standards} for behavioral software engineering research.
\paragraph{Reliability}
Both the internal consistency and stability have been checked multiple times in the development of the GDQ \citep{wheelan} and GDQS \citep{gren2020gdqs}. The lack of a formal definition of what becomes an issue in the problem reports at Volvo Cars is a threat to the reliability of us actually measuring the average age of issues the teams have. Further studies should explicitly ask for such a definition, or help define it, if it does not exist. We do not know how extreme values might have affected this average that Volvo Cars chose to use as a KPI. Perhaps the median age of issues might have been more reliable.
\paragraph{Construct validity}
The constructs under investigation were the four team development scales and the average age of issues. The GDQS is a well-validated scale for measuring team maturity and we only selected teams in a part of Volvo Cars that we know reported and maintained their problem reports continuously. The validity of the used constructs are, therefore, assessed as high.
\paragraph{Internal validity}
We tried to argue for a theoretical reason for causality between team maturity and average age of issues, but we did not measure any other confounding factors. Potential confounds could be type of work, team composition, conformance to processes, individual competencies, etc., and in a social complex adaptive system there could be many more.
\paragraph{External validity}
The small sample of 14 teams and a collection of data from only one part of Volvo Cars are other large threats to this study. We do not know if these teams and the way they work is representative to any teams outside this context. A random sample was not possible in this case since it was hard to find a set of teams that were in a position to continuously report on, and receive feedback from their deliveries. This is an ongoing transformation at Volvo Cars, i.e., to become more data-driven. However, it does seem plausible that they could be applicable to other similar safety-critical software development environments.
\section{Conclusion and Future Work}
We conclude that the average ages of issues in automotive software development teams were associated with teams having higher values on the first two stages of team development. This could mean that mature teams fix issues faster. However, this is just an initial study with a small sample and future studied should sample randomly and also use other productivity measurements.
Another part of the automotive software development that could be studied in the future is extending the work of \citet{safetylenberg} and further look at how teams deal with safety-critical systems.
\section*{Supplementary Material}
The Supplementary Material is available on Zenodo\footnote{\url{https://www.doi.org/10.5281/zenodo.6400809}}. It includes an anonymized raw data CSV file and all the R scripts used.
\section*{Acknowledgments}
Both authors would like to thank the Swedish Research Council funding this work under the auspices of the 2022 Tage Erlander research professorship. Both authors thank Volvo for the provision of the data that made this work possible. This project is an associated project under the Software Center, Gothenburg, Sweden.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\label{sec:intro}
This paper is concerned with solving the following unconstrained optimization
problem:
\begin{equation}
\label{eq:pb}
\min_{\vct{x} \in \mathbb{R}^n} f(\vct{x}),
\end{equation}
where $f: \mathbb{R}^n \rightarrow \mathbb{R}$ is a continuously differentiable function. We
suppose that the derivative of $f$ is unavailable for algorithmic purposes,
thereby precluding the use of standard optimization techniques. This setup
is that of derivative-free optimization, which has a now
well-established class of algorithms~\cite{CAudet_WHare_2017,
ARConn_KScheinberg_LNVicente_2009b,JLarson_MMenickelly_SMWild_2019}. Two
main paradigms can be identified in derivative-free optimization:
model-based algorithms, that construct a model of the objective function,
which is used to guide the optimization process into selecting a new point
to evaluate, and direct-search methods, that proceed by exploring the space
along suitably chosen directions.
A recent trend in nonlinear optimization, and in derivative-free optimization
in particular, is to compare optimization methods based on their
complexity guarantees. In the context of derivative-free optimization, we are
interested in the number of function evaluations required by these methods
to reach a point $\vct{x}$ such that $\|\nabla f(\vct{x})\| \le \epsilon$, where
$\epsilon \in (0,1)$ is a given tolerance. Recent results in worst-case
complexity analyzes have established that this number is of order
$n^2 \epsilon^{-2}$
for standard deterministic derivative-free frameworks such as trust
region~\cite{RGarmanjani_DJudice_LNVicente_2016} and direct search with
sufficient decrease~\cite{LNVicente_2013}, as well as $n^2 \epsilon^{-3/2}$
for cubic regularization based on finite-difference
estimates~\cite{CCartis_NIMGould_PhLToint_2012b}. In the case of direct
search, it was shown that the factor $n^2$ could not be improved
upon using deterministic algorithms~\cite{MDodangeh_LNVicente_ZZhang_2016}.
Despite this negative result, a number of recent algorithmic proposals
based on randomized techniques have been shown to reduce the dependency in
the dimension from $n^2$ to $n$. A direct-search method based on random
directions was endowed with a complexity bound of order $n \epsilon^{-2}$
in a probabilistic sense, provided the directions were chosen uniformly on
the unit sphere~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}. In particular,
choosing two opposite directions was identified as the best choice from a
theoretical viewpoint (in that it maximizes the probability of having a
good direction among two uniformly drawn), that also yielded the best
results in numerical experiments~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015,
SGratton_CWRoyer_LNVicente_ZZhang_2019}. Similar results were obtained through
a different analysis for zeroth-order methods that constructed a gradient
estimate, typically based on Gaussian
sampling~\cite{EBergou_EGorbunov_PRichtarik_2020,YuNesterov_2011,
YuNesterov_VSpokoiny_2017}. Very recently, Kozak et
al.~\cite{DKozak_SBecker_ADoostan_LTenorio_2021} also proposed to use
directional derivatives sampled within random orthogonal subspaces to
approximate the gradient. In a related, subsequent work, Kozak et
al.~\cite{DKozak_CMolinari_LRosasco_LTenorio_SVilla_2021} described a variant
of this approach using finite difference estimates based on orthogonal
directions drawn from random subspaces. In all the aforementioned approaches,
a linear dependence in the problem dimension could be identified,
representing an improvement over the deterministic setting.
The situation was different in the model-based framework. Although
randomized frameworks were proposed based on the same reasoning than
for direct-search methods, the complexity analysis did not suggest
any possible improvement in terms of the problem
dimension~\cite{ASBandeira_KScheinberg_LNVicente_2014,
SGratton_CWRoyer_LNVicente_ZZhang_2018}. However, recent approaches
focusing on constructing models in random, low-dimensional subspaces
did manage to achieve such a theoretical
improvement~\cite{CCartis_LRoberts_2021}
by leveraging random embeddings using for instance Gaussian or random
orthogonal matrices. Such techniques appeared as promising to develop
scalable derivative-free optimization techniques, which was done by
Cartis et al.~\cite{CCartis_TFerguson_LRoberts_2020,CCartis_LRoberts_2021}. Their approach
builds on a general framework for derivative-based optimization in
random subspaces~\cite{CCartis_JFowkes_ZShao_2020,CCartis_JFowkes_ZShao_2022,
ZShao_2022}, and adapts it to a derivative-free, model-based setting. However,
the connection between the subspace arguments in the model-based literature
and the use of random one-dimensional subspaces in direct search was
not investigated.
In this paper, we revisit the analysis of direct-search methods based on
probabilistic properties in order to improve our understanding of their
behavior. To this end, we propose a framework that relies on search directions
chosen within subspaces of the variable space. We define probabilistic
properties on these subspaces and directions, by combining elements
from the direct-search literature and that of model-based methods based on
random subspaces. Such properties are then combined to yield complexity
bounds for our framework, in a way that departs from existing
analyses~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}. In particular,
our reasoning allows for unbounded directions, and thus encompasses
popular practical choices such as Gaussian directions. Using these
bounds, we introduce a suite of new methods that all match the best known
bound in terms of dependencies on $n$, and find a new interpretation of
direct search based on opposite, random directions. As a result, our framework
for direct search together with the model-based
framework~\cite{CCartis_LRoberts_2021} provides a coherent understanding of
the benefits of randomization for scalable derivative-free optimization of
nonconvex functions, resolving the disconnect between randomized direct-search
and randomized model-based techniques that was previously
observed~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015,
SGratton_CWRoyer_LNVicente_ZZhang_2018}. Our experiments confirm that a
randomized subspace approach can be quite beneficial in a direct-search
setting.
The rest of this document is organized as follows. In
Section~\ref{sec:ds}, we recall the main features of direct-search
methods, and the associated complexity guarantees. We then describe a new
paradigm to generate polling directions based on subspaces in
Section~\ref{sec:sds}, for which we establish probabilistic
complexity guarantees.
Numerical experiments for our proposed techniques are given in
Section~\ref{sec:num}.
\paragraph{Software}
All the algorithms discussed here are available in an open-source Python
package available on
Github.\footnote{\url{https://github.com/lindonroberts/directsearch}}
\paragraph{Notation} In what follows, $\|\cdot\|$ will denote the Euclidean
norm for vectors or the operator 2-norm for matrices. We use $\log(\cdot)$ to
denote the reciprocal of the exponential function and $\log_{a}(\cdot)$ to
denote the base-$a$ logarithmic function. The vectors will be denoted
by bold lowercase letters (e.g. $\vct{x}$) while the matrices will be denoted by
bold uppercase letters (e.g. $\mtx{S}$). Sets will be denoted by cursive uppercase
letters (e.g. $\mathcal{D}$). The letters $m,n,r$ will always denote integers
greater than or equal to $1$. Finally, $\mtx{I}_r$ will denote the identity matrix
in $\mathbb{R}^{r \times r}$.
\section{Direct-search framework and complexity results}
\label{sec:ds}
In this section, we recall the key components of a direct-search method based
on sufficient decrease. We focus on the properties that guarantee decrease
in the objective function and, as a result, convergence to approximate
stationary points. Section~\ref{subsec:ds:det} recalls the fundamentals
of direct search based on deterministic properties, while
Section~\ref{subsec:ds:pba} extends the analysis to the case of
probabilistically descent directions, following the reasoning of Gratton et
al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}.
\subsection{Direct search based on deterministic descent}
\label{subsec:ds:det}
Algorithm~\ref{algo:ds} presents a simplified direct-search framework based
on sufficient decrease, for which complexity results can easily be
established. At every iteration, the algorithm chooses a set of polling
directions of fixed cardinality, and evaluates the objective at the
corresponding points. If one of these trial points satisfies the decrease
condition~\eqref{eq:ds:suffdec}, this point becomes the new iterate and the
stepsize parameter is (possibly) increased. Otherwise, the current point does
not change, and the stepsize is decreased.
\begin{algorithm}[h!]
\SetAlgoLined
\DontPrintSemicolon
\BlankLine
\textbf{Inputs}: $\vct{x}_0 \in \mathbb{R}^n$, $\alpha_{\max} > 0$,
$\alpha_0 \in (0,\alpha_{\max}]$, $c>0$, $0 < \gamma_{\rm dec} < 1 < \gamma_{\rm inc}$,
$m \in \mathbb{N}$.\;
\For{$k=0,1,...$}{
Compute a polling set $\mathcal{D}_k \subset \mathbb{R}^n$ of $m$ vectors.\;
If there exists $\vct{d}_k \in \mathcal{D}_k$ such that
\begin{equation}
\label{eq:ds:suffdec}
f(\vct{x}_k+\alpha_k \vct{d}_k) <
f(\vct{x}_k) - \frac{c}{2} \alpha_k^2 \|\vct{d}_k\|^2,
\end{equation}
set $\vct{x}_{k+1}:=\vct{x}_k+\alpha_k \vct{d}_k$ and
$\alpha_{k+1}:=\min\{\gamma_{\rm inc} \alpha_k,\alpha_{\max}\}$.\;
Otherwise, set $\vct{x}_{k+1}:=\vct{x}_k$ and $\alpha_{k+1}:=\gamma_{\rm dec} \alpha_k$.\;
}
\caption{Direct-search framework based on sufficient decrease.\label{algo:ds}}
\end{algorithm}
The choice of the polling sets is crucial for obtaining theoretical guarantees on the
behavior of Algorithm~\ref{algo:ds}. The standard requirements rely on the following
concept of cosine measure, which we define for an arbitrary dimension $r$ for later
use in the paper.
\begin{definition}
\label{de:cosmeas}
Given a set of vectors $\mathcal{D} \subset \mathbb{R}^r$ and a nonzero vector
$\vct{v} \in \mathbb{R}^r$, the \emph{cosine measure of $\mathcal{D}$ at $\vct{v}$} is
defined by
\begin{equation}
\label{eq:cosmeasvec}
\cm{\mathcal{D},\vct{v}}:= \max_{\vct{d} \in \mathcal{D}} \frac{\vct{d}^\mathrm{T} \vct{v}}{\|\vct{d}\|\|\vct{v}\|}.
\end{equation}
The \emph{cosine measure of $\mathcal{D}$} is then given by
\begin{equation}
\label{eq:cosmeas}
\cm{\mathcal{D}}:=\min_{\substack{\vct{v} \in \mathbb{R}^{r} \\ \|\vct{v}\|\neq 0}}
\cm{\mathcal{D},\vct{v}}.
\end{equation}
\end{definition}
Any set $\mathcal{D}$ such that $\cm{\mathcal{D}}>0$ is called a positive spanning set (PSS) for
$\mathbb{R}^r$, as its
elements span $\mathbb{R}^r$ by nonnegative linear combinations~\cite{CAudet_WHare_2017,ARConn_KScheinberg_LNVicente_2009b}.
Using PSSs leads to complexity results for Algorithm~\ref{algo:ds} under
the following standard assumptions on the objective function.
\begin{assumption}
\label{as:flow}
There exists $f_{\rm low} \in \mathbb{R}$ such that $f(\vct{x}) \ge f_{\rm low}$ for all $\vct{x} \in \mathbb{R}^n$.
\end{assumption}
\begin{assumption}
\label{as:fC11}
The function $f$ is continuously differentiable, and its derivative is
$L$-Lipschitz continuous with $L>0$.
\end{assumption}
The analysis of Algorithm~\ref{algo:ds} is based on two key arguments. First, one
can use the sufficient decrease property~\eqref{eq:ds:suffdec} to guarantee that the
step size converges to zero: this is the purpose of the following lemma.
\begin{lemma}
\label{le:ds:sumalpha}
Let Assumption~\ref{as:flow} hold, and consider the step size sequence
$\{\alpha_k\}$ produced by Algorithm~\ref{algo:ds}. Suppose that the
directions in $\{\mathcal{D}_k\}_k$ are uniformly bounded in norm. Then, there exists
$\beta>0$ that does not depend on $\{\mathcal{D}_k\}$ such that
\begin{equation}
\label{eq:ds:sumalpha}
\sum_{k=0}^{\infty} \alpha_k^2 \le \beta < \infty.
\end{equation}
As a result, $\lim_{k \rightarrow \infty} \alpha_k = 0$.
\end{lemma}
We emphasize that the result of Lemma~\ref{le:ds:sumalpha} requires the
directions to be bounded in norm, which is easily achieved by restricting
the polling sets to the unit sphere. Note, however, that the cosine measures of
these sets is not relevant to the result of Lemma~\ref{le:ds:sumalpha}.
The second main ingredient of our complexity analysis concerns unsuccessful
iterations. Provided the polling set is of sufficiently good quality, we can
relate the stepsize to the norm of the gradient at the current iterate.
\begin{lemma}
\label{le:ds:unsucc}
Consider the $k$-th iteration of Algorithm~\ref{algo:ds} under the assumption
that\\
$\cm{\mathcal{D}_k,-\nabla f(\vct{x}_k)} \ge \kappa \in (0,1)$. If iteration
$k$ is unsuccessful, then one has
\begin{equation}
\label{eq:ds:unsucc}
\alpha_k \ge \frac{2}{L+c}\kappa \|\nabla f(\vct{x}_k)\|.
\end{equation}
\end{lemma}
In practice, the gradient is unknown, thus the assumption
$\cm{\mathcal{D}_k,-\nabla f(\vct{x}_k)} \ge \kappa$ is replaced by
$\cm{\mathcal{D}_k} \ge \kappa$, which is equivalent to assuming that $\mathcal{D}_k$ is a
PSS in $\mathbb{R}^n$.
The updating process on $\{\alpha_k\}_k$ together with the result of
Lemma~\ref{le:ds:unsucc} guarantees convergence of Algorithm~\ref{algo:ds}
provided the cosine measure sequence $\{\cm{\mathcal{D}_k}\}_k$ is uniformly
bounded below by $\kappa \in (0,1)$.
Under this assumption, it is known~\cite{LNVicente_2013} that
Algorithm~\ref{algo:ds} reaches an iterate $\vct{x}_k$
such that $\|\nabla f(\vct{x}_k)\| \le \epsilon$ using at most
\begin{equation}
\label{eq:wccdsdeterm}
\mathcal{O}\left( m\,\kappa^{-2}\,\epsilon^{-2}\right)
\end{equation}
function evaluations.
A classical choice in direct search consists in selecting $\mathcal{D}_k$ as the
set of coordinate vectors and their negatives in $\mathbb{R}^n$. In that case, one
has $m=2n$, $\kappa=\tfrac{1}{\sqrt{n}}$, and the bound becomes
\begin{equation}
\label{eq:wccdplusdeterm}
\mathcal{O}\left( n^2\,\epsilon^{-2}\right).
\end{equation}
In a deterministic setting, the dependency in $n^2$ cannot be improved while
using positive spanning sets without additional
information~\cite{MDodangeh_LNVicente_ZZhang_2016}.
\subsection{Direct search based on probabilistic descent}
\label{subsec:ds:pba}
As highlighted in the previous section, the use of positive spanning sets is
instrumental for convergence of classical direct-search methods. However, this
property also incurs a dependency of $n^2$ in the complexity bounds, due to
the need to cover an $n$-dimensional space. Gratton et
al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015} recently established that
randomly generated direction sets \emph{that do not form a PSS} can still
provide a good approximation of a particular vector in that space, and that
this is sufficient to produce good directions. The following property was
thus introduced.
\begin{definition}
\label{de:ds:pbadesc}
Given $p \in (0,1]$ and $\kappa \in (0,1)$, the polling set sequence
$\{\mathcal{D}_k\}$ used in Algorithm~\ref{algo:ds} is called
$(p,\kappa)$-descent if
\begin{equation}
\label{eq:ds:pbadesc}
\left\{
\begin{array}{lll}
\P{ \cm{\mathcal{D}_0,-\nabla f(\vct{x}_0)} \ge \kappa} &\ge &p \\
& & \\
\P{ \cm{\mathcal{D}_k,-\nabla f(\vct{x}_k)}
\ge \kappa \middle| \mathcal{F}_{k-1} } &\ge &p, \qquad \forall k \ge 1,
\end{array}
\right.
\end{equation}
where $\mathcal{F}_{k-1}$ is the $\sigma$-algebra generated by the random sets
$\mathcal{D}_0,\dots,\mathcal{D}_{k-1}$.
\end{definition}
The $(p,\kappa)$-descent property is enough to establish probabilistic
complexity results for Algorithm~\ref{algo:ds}. Indeed, the result of
Lemma~\ref{le:ds:sumalpha} holds for every realization of the method provided
the directions are bounded, while that of Lemma~\ref{le:ds:unsucc} now
depends on the occurrence of the random event
$\{\cm{\mathcal{D}_k,-\nabla f(\vct{x}_k)} \ge \kappa\}$.
Using martingale-type arguments, it is then possible to show that a method
employing a $(p,\kappa)$-descent sequence converges almost surely to a point
with zero gradient. In addition, high probability complexity guarantees hold,
in that the method reaches an iterate satisfying
$\|\nabla f(\vct{x}_k)\|\le \epsilon$ using at most
\[
\mathcal{O}(m \kappa^{-2}\epsilon^{-2})
\]
function evaluations \emph{with probability at least
$1-\mathcal{O}(-\exp(C\epsilon^{-2}))$}.
Assuming $\mathcal{D}_k$ is randomly generated using $m \ge 2$ directions uniformly
distributed on the unit sphere, one obtains an high-probability evaluation
complexity bound in
\begin{equation}
\label{eq:wccdspd}
\mathcal{O}(n\epsilon^{-2}).
\end{equation}
Using $m=2$ emerged as a good practical alternative, with the use of two
opposite directions allowing to maximize the probability of having a descent
set~\cite[Appendix B]{SGratton_CWRoyer_LNVicente_ZZhang_2015}.
Other proposals based on random directions~\cite{YuNesterov_VSpokoiny_2017,
EBergou_EGorbunov_PRichtarik_2020} relied on evaluations along
random Gaussian opposite directions, leading to a same improvement in the
complexity bound (note that the probabilistic descent analysis does not
apply to Gaussian direction since those are not deterministically bounded
in norm). For such approaches, using evaluations along a one-dimensional
subspace stood out as an efficient and theoretically sound strategy, but the
role of the subspace was not further investigated. Taking the subspace nature
of those directions into account is the key goal of this paper, and the subject
of the next section.
\section{Probabilistic descent in reduced spaces}
\label{sec:sds}
Building in the framework of Algorithm~\ref{algo:ds}, we propose a method that
operates in a reduced space by selecting both a subspace of $\mathbb{R}^n$ and a set
of polling directions within that subspace. This two-step process allows to
identify deterministic and probabilistic conditions under which the subspace
(resp. the directions) are of suitable quality.
\subsection{Algorithm and suitable properties}
\label{subsec:sds:algo}
Algorithm~\ref{algo:sds} details our proposed method. At every iteration, we
first generate directions in $\mathbb{R}^r$ with $r\le n$. We then combine those
directions with a matrix $\mtx{P}_k \in \mathbb{R}^{r \times n}$ to obtain a polling set
in $\mathbb{R}^n$ \emph{that belongs to an $r$-dimensional subspace of $\mathbb{R}^n$}. This
property is a key feature of our method, as it allows our method to operate
in subspaces of lower dimension than that of the original problem: at
iteration $k$, our polling set is $\{\mtx{P}_k^\mathrm{T} \vct{d} | \vct{d} \in \mathcal{D}_k\}$.
\begin{algorithm}[h!]
\SetAlgoLined
\DontPrintSemicolon
\BlankLine
\textbf{Inputs}: $\vct{x}_0 \in \mathbb{R}^n$, $\alpha_{\max} > 0$,
$\alpha_0 \in (0,\alpha_{\max}]$, $c>$, $0 < \gamma_{\rm dec} < 1 < \gamma_{\rm inc}$;
$m \in \mathbb{N}$, $r \le n$.\;
\For{$k=0,1,...$}{
Compute a matrix $\mtx{P}_k \in \mathbb{R}^{r \times n}$.\;
Compute a set $\mathcal{D}_k \subset \mathbb{R}^r$ of $m$ vectors.\;
If there exists $\vct{d}_k \in \mathcal{D}_k$ such that
\begin{equation}
\label{eq:ssuffdec}
f(\vct{x}_k+\alpha_k \mtx{P}^\mathrm{T}_k \vct{d}_k) < f(\vct{x}_k)
- \frac{c}{2} \alpha_k^2 \|\mtx{P}_k^\mathrm{T} \vct{d}_k\|^2,
\end{equation}
set $\vct{x}_{k+1}:=\vct{x}_k+\alpha_k \mtx{P}^T_k \vct{d}_k$ and
$\alpha_{k+1}:=\min\{\gamma_{\rm inc}\alpha_k,\alpha_{\max}\}$.\;
Otherwise, set $\vct{x}_{k+1}:=\vct{x}_k$ and
$\alpha_{k+1}:=\gamma_{\rm dec} \alpha_k$.\;
}
\caption{Direct-search method in reduced spaces.\label{algo:sds}}
\end{algorithm}
To assess the quality of the polling sets, we define separate properties for
the matrix $\mtx{P}_k$ and the set $\mathcal{D}_k$, starting with the former. The matrix
$\mtx{P}_k$ produces an $r$-dimensional subspace of $\mathbb{R}^n$ in which polling
directions will be generated. When $r<n$, the use of $\mtx{P}_k$ will prevent the
polling set from providing good approximations to all of $\mathbb{R}^n$. Nevertheless,
for optimization purposes, we are merely interested in approximating the
negative gradient (and its norm). Consequently, we require the matrix $\mtx{P}_k$
to capture a significant portion of gradient information. In addition,
defining the polling set through application of $\mtx{P}_k^\mathrm{T}$ should alter the
directions in the reduced subspace in a controlled way, which we express
through bounds on the singular values of the matrix. These considerations lead
to the following definition, motivated by a similar concept in the model-based
setting~\cite{CCartis_LRoberts_2021,CCartis_JFowkes_ZShao_2020,
CCartis_JFowkes_ZShao_2022,ZShao_2022}.
\begin{definition}
\label{de:sds:Pk}
Let $\eta$, $\sigma$ and $P_{\max}$ be positive quantities.
For any realization of Algorithm~\ref{algo:sds} and any $k \in \mathbb{N}$, the
matrix $\mtx{P}_k$ is called \emph{$(\eta,\sigma,P_{\max})$-well aligned} for $f$
at $\vct{x}_k$ provided
\begin{equation}
\label{eq:sds:Pk:grad}
\|\mtx{P}_k \nabla f(\vct{x}_k)\| \ge \eta \|\nabla f(\vct{x}_k)\|,
\end{equation}
\begin{equation}
\label{eq:sds:Pk:pmax}
\|\mtx{P}_k\| \le P_{\max},
\end{equation}
\begin{equation}
\label{eq:sds:Pk:PTd}
\sigma_{\min}(\mtx{P}_k) \ge \sigma,
\end{equation}
where $\sigma_{\min}(\cdot)$ denotes the minimum nonzero
singular value of the matrix $\mtx{P}_k$.
\end{definition}
Conditions~\eqref{eq:sds:Pk:grad}--\eqref{eq:sds:Pk:PTd} are satisfied with
$\eta=\sigma=P_{\max}=1$ when $r=n$ and $\mtx{P}_k$ is the identity matrix, but may
not hold when $r< n$, or when $\mtx{P}_k$ is a random matrix. For this reason,
we introduce a probabilistic counterpart to Definition~\ref{de:sds:Pk}.
\begin{definition}
\label{de:sds:pbaPk}
The sequence $\{\mtx{P}_k\}_k$ generated by Algorithm~\ref{algo:sds} is called
\emph{$(\eta,\sigma,P_{\max},q)$-well-aligned} for $q \in (0,1]$ if
\begin{equation}
\label{eq:sds:pbaPk}
\begin{array}{rll}
\P{\mtx{P}_0\ \mbox{is ($\eta,\sigma,P_{\max}$)-well aligned}}
&\ge &q \\
\forall k \ge 1, \qquad
\P{ \mtx{P}_k\ \mbox{is ($\eta,\sigma,P_{\max}$)-well aligned} \
\middle|\ \mathcal{F}_{k-1}}
&\ge &q,
\end{array}
\end{equation}
where $\mathcal{F}_{k-1}$ is the $\sigma$-algebra generated by
$\mtx{D}_0,\mtx{P}_0,\dots,\mtx{D}_{k-1},\mtx{P}_{k-1}$.
\end{definition}
Our requirement on $\mathcal{D}_k$ is given below, and is similar to that used in
Section~\ref{subsec:ds:det}.
\begin{definition}
\label{de:sds:Dk}
Let $\kappa \in (0,1]$ and $D_{\max}>0$.
For any realization of Algorithm~\ref{algo:sds} and any index $k \in \mathbb{N}$,
the set $\mathcal{D}_k$ is called $(\kappa,D_{\max})$-descent for $f$ and $\mtx{P}_k$ at
$\vct{x}_k$ if
\begin{equation}
\label{eq:sds:Dkcm}
\cm{\mathcal{D}_k,-\mtx{P}_k \nabla f(x_k)}
= \max_{\vct{d} \in \mathcal{D}_k}
\frac{-\vct{d}^\mathrm{T} \mtx{P}_k \nabla f(\vct{x}_k)}{\|\vct{d}\|\|\mtx{P}_k \nabla f(\vct{x}_k)\|} \ge \kappa
\end{equation}
and
\begin{equation}
\label{eq:sds:Dkdmax}
\forall \vct{d} \in \mathcal{D}_k, \quad D_{\max}^{-1} \le \|\vct{d}\| \le D_{\max}.
\end{equation}
\end{definition}
Examples of sets satisfying Definition~\ref{de:sds:Dk} are positive spanning
sets in $\mathbb{R}^r$ with unitary elements. As for the properties of $\mtx{P}_k$, we provide a
probabilistic counterpart of Definition~\ref{de:sds:Dk} below.
\begin{definition}
\label{de:sds:pbaDk}
The sequence $\{\mtx{D}_k\}_k$ generated by Algorithm~\ref{algo:sds} is called
\emph{$(\kappa,D_{\max},p)$-descent} for $p \in (0,1]$ if
\begin{equation}
\label{eq:sds:pbaDk}
\begin{array}{rll}
\P{\mathcal{D}_0\ \mbox{is ($\kappa,D_{\max}$)-descent}
\middle|\ \mathcal{F}_{-1/2}}
&\ge &p \\
\forall k \ge 1, \qquad
\P{\mathcal{D}_k\ \mbox{is ($\kappa,D_{\max}$)-descent} \
\middle|\ \mathcal{F}_{k-1/2}}
&\ge &p,
\end{array}
\end{equation}
where $\mathcal{F}_{k-1}$ is the $\sigma$-algebra generated by
$\mtx{P}_0,\mathcal{D}_0,\dots,\mtx{P}_{k-1},\mtx{D}_{k-1},\mtx{P}_k$ and $\mathcal{F}_{-1/2}$ is
the $\sigma$-algebra generated by $\mtx{P}_0$.
\end{definition}
Definition~\ref{de:sds:pbaDk} departs from Definition~\ref{de:ds:pbadesc} in
that it allows for unbounded directions, as long as such directions occur with
a small probability. This enables for instance the use of Gaussian vectors, a
distribution that was not immediately covered by the analysis of Gratton et
al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}.
In Section~\ref{subsec:sds:ps} we give several possible choices for $\mtx{P}_k$
and $\mtx{D}_k$ that satisfy these requirements.
\subsection{Complexity analysis}
\label{subsec:sds:wcc}
In this section, we leverage the probabilistic properties defined above to
derive complexity results for Algorithm~\ref{algo:sds}. Our analysis follows
a reasoning previously developed for derivative-free methods based on
probabilistic properties~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015,
SGratton_CWRoyer_LNVicente_ZZhang_2018}, but involves two properties of this
form at every iteration, respectively related to $\mtx{P}_k$ and $\mathcal{D}_k$.
Although these properties can be handled jointly, the analysis still departs
from existing ones as we allow for directions that are unbounded in norm. At
the same time, we point out that our approach still relies on exact function
values, and therefore does not require Lyapunov functions
similar to those used in stochastic
optimization~\cite{CPaquette_KScheinberg_2020}.
For the rest of this section, let $\mathcal{S}$ denote the index set of
\emph{successful iterations} (i.e.~the $k$ for which $\vct{x}_{k+1} \neq \vct{x}_k$)
and $\mathcal{U}$ denote the index set of \emph{unsuccessful iterations} (for which
$\vct{x}_{k+1}=\vct{x}_k$).
\begin{lemma}
\label{le:sds:succit}
Let Assumption~\ref{as:fC11} hold, and consider the $k$-th iteration of
a realization of Algorithm~\ref{algo:sds}. Suppose further that
$\mathcal{D}_k$ is $(\kappa,D_{\max})$-descent and that $\mtx{P}_k$ is
$(\eta,\sigma,P_{\max})$-well aligned. Finally, suppose that
\begin{equation}
\label{eq:sds:alphabnd}
\alpha_k < \bar{\alpha}\|\nabla f(\vct{x}_k)\|, \quad \mbox{where}
\quad \bar{\alpha}:= \frac{2 \kappa \eta}{(L+c)P_{\max}^2 D_{\max}^3}.
\end{equation}
Then, the $k$-th iteration is successful.
\end{lemma}
\begin{proof}
To find a contradiction, suppose that iteration $k$ is unsuccessful.
Then, by Assumption~\ref{as:fC11}, for all $\vct{d}\in\mathcal{D}_k$ we have
\begin{align*}
-\frac{c}{2}\alpha_k^2 \|\mtx{P}_k^T \vct{d}\|^2
&\le f(\vct{x}_k+\alpha_k \mtx{P}_k^T \vct{d}) - f(\vct{x}_k), \\
&\le \alpha_k \vct{d}^T \mtx{P}_k \nabla f(\vct{x}_k)
+ \frac{L}{2}\alpha_k^2 \|\mtx{P}_k^T \vct{d}\|^2.
\end{align*}
Since $\mathcal{D}_k$ is a $(\kappa,D_{\max})$-descent set, there exists
$\vct{d}_k\in\mathcal{D}_k$ such that
\begin{align}
\frac{\vct{d}_k^T (-\mtx{P}_k \nabla f(\vct{x}_k))}{\|\vct{d}_k\| \: \|\mtx{P}_k \nabla f(\vct{x}_k)\|}
= \cm{\mathcal{D}_k,-\mtx{P}_k \nabla f(\vct{x}_k)} \geq \kappa.
\end{align}
Therefore, we obtain
\begin{eqnarray*}
-\frac{c}{2}\alpha_k^2 \|\mtx{P}_k^T \vct{d}_k\|^2
&\le &-\kappa \alpha_k \|\vct{d}_k\| \|\mtx{P}_k \nabla f(\vct{x}_k)\|
+ \frac{L}{2}\alpha_k^2\|\mtx{P}_k^\mathrm{T} \vct{d}_k\|^2 \\
\kappa \alpha_k \|\vct{d}_k\| \: \|\mtx{P}_k \nabla f(\vct{x}_k)\|
&\le & \frac{L+c}{2} \alpha_k^2\|\mtx{P}_k^\mathrm{T} \vct{d}_k\|^2 \\
\kappa \|\vct{d}_k\| \: \|\mtx{P}_k \nabla f(\vct{x}_k)\|
&\le & \frac{L+c}{2} \alpha_k\|\mtx{P}_k^\mathrm{T} \vct{d}_k\|^2 \\
\end{eqnarray*}
Using now the property~\eqref{eq:sds:Pk:grad} on $\mtx{P}_k$ together with
the bound~\eqref{eq:sds:Dkdmax} on $\|\vct{d}_k\|$ leads to
\begin{equation*}
\kappa \|\vct{d}_k\| \: \|\mtx{P}_k \nabla f(\vct{x}_k)\|
\ge \kappa \eta D_{\max}^{-1} \|\nabla f(\vct{x}_k)\|
\end{equation*}
as well as
\begin{equation*}
\|\mtx{P}_k^\mathrm{T} \vct{d}_k\|^2 \le \|\mtx{P}_k^\mathrm{T}\|^2 \|\vct{d}_k\|^2
\le P_{\max}^2 D_{\max}^2.
\end{equation*}
Putting everything together, we arrive at
\begin{equation*}
\kappa \eta D_{\max}^{-1} \|\nabla f(\vct{x}_k)\|
\le \frac{L+c}{2} P_{\max}^2 D_{\max}^2 \alpha_k \quad
\Leftrightarrow \quad \alpha_k \ge \bar{\alpha}\|\nabla f(\vct{x}_k)\|,
\end{equation*}
and this contradicts~\eqref{eq:sds:alphabnd}.
\end{proof}
We now introduce the following indicator variables:
\begin{subequations}
\label{eq:sds:indic}
\begin{align}
\label{eq:sds:indiczk}
Z_k &:= \mathbf{1}\left( \mathcal{D}_k\ \mbox{($\kappa,D_{\max}$)-descent and }
\ \mtx{P}_k\ \mbox{($\eta,\sigma,P_{\max}$)-well aligned}
\right), \\
\label{eq:sds:indicvk}
V_k(\alpha) &:= \mathbf{1}\left( \alpha_k < \alpha \right)
\quad \forall \alpha>0, \\
\label{eq:sds:indicwk}
W_k &:= \mathbf{1}\left( k \in \mathcal{S}\right),
\end{align}
\end{subequations}
whose realizations will be denoted by $z_k,v_k(\alpha),w_k$, respectively. Our
goal is to bound the sum of $z_k$ by a quantity involving the gradient
norm. To this end, we study careful combinations of the indicator variables
above, for which we can provide bounds that are independent of the iteration:
this technique has proven useful in establishing complexity guarantees for
derivative-free optimization schemes with probabilistic
components~\cite{CCartis_LRoberts_2021,CCartis_KScheinberg_2018,
CPaquette_KScheinberg_2020,SGratton_CWRoyer_LNVicente_ZZhang_2015,
SGratton_CWRoyer_LNVicente_ZZhang_2018}.
Our first result bounds the sum of squared stepsizes for a certain
subset of successful iterations. Unlike in analyses of direct search
based solely on descent
properties~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}, where such a
bound can be obtained for all iterations, here it only holds for those
successful iterations that use both well-aligned subspace matrices and
descent directions.
\begin{lemma}
\label{le:sds:sumzalpha}
Let Assumption~\ref{as:flow} hold. For any realization of
Algorithm~\ref{algo:sds},
\begin{equation}
\label{eq:sds:sumzalpha}
\sum_{k=0}^{\infty} z_k w_k \alpha_k^2 \le
\beta:=\frac{2 D_{\max}^2(f(\vct{x}_0)-f_{\rm low})}{c \sigma^2}.
\end{equation}
\end{lemma}
\begin{proof}
It suffices to consider iterations for which $z_k w_k=1$, i.e.
successful iterations for which $\mathcal{D}_k$ is $(\kappa,D_{\max})$-descent
and $\mtx{P}_k$ is $(\eta,\sigma,P_{\max})$-well aligned. For such an iteration,
there exists $\vct{d}_k \in \mathcal{D}_k$ such that
$\vct{x}_{k+1}=\vct{x}_k+\alpha_k \mtx{P}_k^\mathrm{T} \vct{d}_k$, and
\begin{eqnarray*}
f(\vct{x}_k) - f(\vct{x}_{k+1})
\ge \frac{c}{2}\alpha_k^2 \|\mtx{P}_k^\mathrm{T} \vct{d}_k\|^2
\ge \frac{c}{2}\alpha_k^2 \sigma^2 \|\vct{d}_k\|^2
\ge \frac{c}{2}\alpha_k^2 \sigma^2 D_{\max}^{-2}
= \frac{c}{2}\sigma^2 D_{\max}^{-2} \alpha_k^2.
\end{eqnarray*}
On the other hand, Assumption~\ref{as:flow} guarantees that
\begin{equation*}
f(\vct{x}_0) - f_{\rm low}
\ge \sum_{k=0}^{\infty} f(\vct{x}_k)-f(\vct{x}_{k+1})
\ge \sum_{k=0}^{\infty} z_k w_k\,(f(\vct{x}_k) - f(\vct{x}_{k+1})).
\end{equation*}
Thus, we obtain
\[
\sum_{k=0}^{\infty} z_k w_k \alpha_k^2
\le \frac{2 D_{\max}^2(f(\vct{x}_0)-f_{\rm low})}{c \sigma^2},
\]
proving the desired result.
\end{proof}
The next two results are obtained by carefully examining the behavior
of the stepsize sequence. This is another notable departure from
the analysis of direct search based on probabilistic
descent~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}, that is
due to our two probabilistic properties. We note that similar
results have been derived in the context of randomized model-based
methods~\cite{CCartis_LRoberts_2021}.
\begin{lemma}
\label{le:sds:bndwv}
Let Assumption~\ref{as:flow} hold. Consider a realization of
Algorithm~\ref{algo:sds} and an index $k$ such that
$\min_{0 \le \ell \le k-1} \|\nabla f(\vct{x}_{\ell})\| > 0$.
Then,
\begin{equation}
\label{eq:sds:bndwv}
\sum_{j=0}^{k-1} v_j\left(\overline{\alpha}_k\right) w_j
\; \le \;
\mu \sum_{j=0}^{k-1}
v_j\left(\frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)
(1-w_j)
\end{equation}
with $\mu:=\log_{\gamma_{\rm inc}}(\gamma_{\rm dec}^{-1})$.
\begin{equation}
\label{eq:sds:kalpha}
\overline{\alpha}_k:=\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\min\left\{
\gamma_{\rm inc}^{-1} \alpha_0,\bar{\alpha}
\min_{0 \le \ell \le k-1} \|\nabla f(\vct{x}_{\ell})\| \right\},
\end{equation}
and $\bar{\alpha}$ is defined as in Lemma~\ref{le:sds:succit}.
\end{lemma}
\begin{proof}
If $v_j\left(\overline{\alpha}_k\right) w_j=0$ for every $j \leq k-1$, the
bound clearly holds. For the rest of the proof, we thus
assume that there exist at least one index $j \in \{0,\dots,k-1\}$
such that $v_j\left(\overline{\alpha}_k\right) w_j=1$. Since
$\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}} > 1$, we also have
$v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)w_j=1$,
thus there also exists at least one index satisfying this property.
Consider a sequence of iterates $j_1,\dots,j_2$ such that
$v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right) =1$ for
every $j=j_1,\dots,j_2$, with
$v_{j_2}\left(\overline{\alpha}_k\right)w_{j_2}=1$ and either $j_1=0$ or
$v_{j_1-1}\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)=0$
(note that such a sequence necessarily exists by assumption).
Using the updating rules on the stepsize together with
the bound
$\overline{\alpha}_k < \gamma_{\rm inc}^{-1}\alpha_0 \leq \gamma_{\rm inc}^{-1}\alpha_{\max}$,
we obtain for any $j=j_1,\dots,j_2$ that
\begin{equation}
\label{eq:sds:alphaupvw}
\left\{
\begin{array}{lll}
\alpha_{j+1}
&= \min\{\gamma_{\rm inc} \alpha_j,\alpha_{\max}\}
= \gamma_{\rm inc} \alpha_j,
&\mbox{if\ } v_j(\overline{\alpha}_k) w_j=1, \\
\alpha_{j+1}
&\ge \alpha_j,
&\mbox{if\ } (1-v_j(\overline{\alpha}_k))
v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right) w_j=1, \\
\alpha_{j+1}
&= \gamma_{\rm dec} \alpha_j,
&\mbox{if\ } v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)(1-w_j)=1.
\end{array}
\right.
\end{equation}
Note that the rules~\eqref{eq:sds:alphaupvw} cover all possible
cases. Applying rule~\eqref{eq:sds:alphaupvw} iteratively gives:
\begin{eqnarray*}
\alpha_{j_2+1}
&\ge &\gamma_{\rm inc}^{\sum_{j=j_1}^{j_2} v_j(\overline{\alpha}_k) w_j}
\gamma_{\rm dec}^{\sum_{j=j_1}^{j_2}
v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)(1-w_j)} \alpha_{j_1}.
\end{eqnarray*}
Moreover, since
$v_{j_2}\left(\overline{\alpha}_k\right) w_{j_2}=1$, we have
\[
\alpha_{j_2+1} = \gamma_{\rm inc} \alpha_{j_2} \le \gamma_{\rm inc} \overline{\alpha}_k,
\]
leading to
\begin{equation}
\label{eq:vwboundsubseq}
\frac{\gamma_{\rm inc}\overline{\alpha}_k}{\alpha_{j_1}}
\ge \gamma_{\rm inc}^{\sum_{j=j_1}^{j_2} v_j(\overline{\alpha}_k) w_j}
\gamma_{\rm dec}^{\sum_{j=j_1}^{j_2}
v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)(1-w_j)}.
\end{equation}
To bound the ratio $\tfrac{\gamma_{\rm inc}\overline{\alpha}_k}{\alpha_{j_1}}$, we consider
two cases. If $j_1=0$, then $\gamma_{\rm dec}\overline{\alpha}_k \le \alpha_0$ by definition of
$\overline{\alpha}_k$, and thus $\gamma_{\rm dec}\frac{\overline{\alpha}_k}{\alpha_{j_1}} \le 1$. Otherwise,
if $j_1>0$, we have by definition of $j_1$ that
$v_{j_1-1}\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)=0$, that is,
$\alpha_{j_1-1} \ge \frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k$. Per the updating rules
on the step size, this implies
\[
\alpha_{j_1} \ge \gamma_{\rm dec} \alpha_{j_1-1} \ge \gamma_{\rm inc} \overline{\alpha}_k,
\]
hence $\gamma_{\rm inc}\frac{\overline{\alpha}_k}{\alpha_{j_1}} \le 1$ also holds in this case.
Plugging this bound into~\eqref{eq:sds:1v1wboundsubseq} yields
\[
1 \ge \gamma_{\rm inc}^{\sum_{j=j_1}^{j_2} v_j(\overline{\alpha}_k) w_j}
\gamma_{\rm dec}^{\sum_{j=j_1}^{j_2}
v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)(1-w_j)}.
\]
Taking logarithms, we obtain that
\begin{equation*}
0 \ge \log(\gamma_{\rm inc})\sum_{j=j_1}^{j_2} v_j(\overline{\alpha}_k) w_j
+ \log(\gamma_{\rm dec})\sum_{j=j_1}^{j_2}
v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)(1-w_j),
\end{equation*}
which after rearranging becomes
\begin{equation*}
\sum_{j=j_1}^{j_2} v_j(\overline{\alpha}_k) w_j
\le \log_{\gamma_{\rm inc}}(\gamma_{\rm dec}^{-1})
\sum_{j=j_1}^{j_2} v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)
(1-w_j)
= \mu \sum_{j=j_1}^{j_2} v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)
(1-w_j).
\end{equation*}
To conclude, we simply observe that
\[
\left\{ j\ \middle|\ v_j(\overline{\alpha}_k) w_j=1 \right\}
\; \subset \;
\left\{ j\ \middle|\ v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)=1 \right\},
\]
on equivalently, that every $j\in\{0,\ldots,k-1\}$ for which $v_j(\overline{\alpha}_k) w_j=1$
is in such a subsequence $j_1,\ldots,j_2$. As a result, we obtain
\[
\sum_{j=0}^{k-1} v_j(\overline{\alpha}_k) w_j
\; \le \;
\mu \sum_{j=0}^{k-1} v_j\left(\tfrac{\gamma_{\rm inc}}{\gamma_{\rm dec}}\overline{\alpha}_k\right)
(1-w_j).
\]
\end{proof}
\begin{lemma}
\label{le:sds:bnd1w1v}
Let Assumption~\ref{as:flow} hold. Consider a realization of
Algorithm~\ref{algo:sds} and an index $k$ such that
$\min_{0 \le \ell \le k-1} \|\nabla f(\vct{x}_{\ell})\| >0$. Then,
\begin{equation}
\label{eq:sds:bnd1w1v}
\sum_{j=0}^{k-1}
\left(1-v_j\left(\overline{\alpha_k}\right)\right)(1-w_j)
\; \le \;
\frac{1}{\mu}\sum_{j=0}^{k-1} \left(1-v_j\left(
\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha_k}\right)\right)
w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right),
\end{equation}
where $\mu$ and $\overline{\alpha}_k$ are defined as in
Lemma~\ref{le:sds:bndwv}.
\end{lemma}
\begin{proof}
The proof follows the template of that of Lemma~\ref{le:sds:bndwv}.
The bound trivially holds if
$\left(1-v_j\left(\overline{\alpha_k}\right)\right)(1-w_{j}) =0$
for every $j \leq k-1$. Therefore, we suppose that there exists at least one
index $j \in \{0,\dots,k-1\}$ such that
$(1-v_j(\overline{\alpha}_k))(1-w_j)=1$. Since $\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}} < 1$, we
also have
$\left(1-v_j\left(
\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)(1-w_j)=1$, thus
there also exists at least one index satisfying this property.
Consider now a sequence of iterates $j_1,\dots,j_2$ such that
$\left(1-v_j\left(
\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)=1$
for every $j=j_1,\dots,j_2$, with
$\left(1-v_{j_2}\left(\overline{\alpha}_k\right)\right)(1-w_{j_2})=1$ and
either $j_1=0$ or $v_{j_1-1}\left(
\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)=1$: such a sequence
necessarily exists by assumption.
From the updating rules on the stepsize, we
obtain the following possible cases:
\begin{equation}
\label{eq:sds:alphaup1v1w}
\left\{
\begin{array}{lll}
\alpha_{j+1}
&= \gamma_{\rm dec} \alpha_j ,
&\mbox{if\ } (1-v_j(\overline{\alpha}_k)) (1-w_j)=1, \\
\alpha_{j+1}
&\le \alpha_j,
&\mbox{if\ } v_j(\overline{\alpha}_k)
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)
(1-w_j)=1, \\
\alpha_{j+1}
&\le \gamma_{\rm inc} \alpha_j,
&\mbox{if\ }
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)
w_j=1,
\end{array}
\right.
\end{equation}
for any $j=j_1,\dots,j_2$. By applying rule~\eqref{eq:sds:alphaup1v1w}
iteratively, we thus obtain
\[
\alpha_{j_2+1} \le
\gamma_{\rm dec}^{\sum_{j=j_1}^{j_2} (1-v_j(\overline{\alpha}_k)) (1-w_j)}
\gamma_{\rm inc}^{\sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j}
\alpha_{j_1}.
\]
In addition, using that $\left(1-v_{j_2}\left(\overline{\alpha}_k\right)\right)(1-w_{j_2})=1$,
we also have
\[
\alpha_{j_2+1} = \gamma_{\rm dec} \alpha_{j_2}
\ge \gamma_{\rm dec} \overline{\alpha}_k,
\]
thus
\begin{equation}
\label{eq:sds:1v1wboundsubseq}
\gamma_{\rm dec}\frac{\overline{\alpha}_k}{\alpha_{j_1}}
\le
\gamma_{\rm dec}^{\sum_{j=j_1}^{j_2} (1-v_j(\overline{\alpha}_k)) (1-w_j)}
\gamma_{\rm inc}^{\sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j}.
\end{equation}
Taking the logarithm, we get
\begin{equation*}
\log\left(\gamma_{\rm dec}\frac{\overline{\alpha}_k}{\alpha_{j_1}}\right) \le
-\log(\gamma_{\rm dec}^{-1}) \sum_{j=j_1}^{j_2} (1-v_j(\overline{\alpha}_k)) (1-w_j)
+ \log(\gamma_{\rm inc}) \sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j,
\end{equation*}
which after re-arranging gives
\begin{eqnarray*}
\sum_{j=j_1}^{j_2} (1-v_j(\overline{\alpha}_k)) (1-w_j)
&\le
&\log_{\gamma_{\rm dec}^{-1}}(\gamma_{\rm inc})\sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
- \log_{\gamma_{\rm dec}^{-1}}\left(\gamma_{\rm dec}\frac{\overline{\alpha}_k}{\alpha_{j_1}}\right) \\
&= &\frac{1}{\mu}\sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_{j_1}}{\gamma_{\rm dec}\overline{\alpha}_k}\right)
\end{eqnarray*}
by definition of $\mu$.
To bound the last term, we consider two cases.
If $j_1=0$, then $\frac{\alpha_{j_1}}{\gamma_{\rm dec}\overline{\alpha}_k}
= \frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k}$. Otherwise,
if $j_1>0$, we have by definition of $j_1$ that
$v_{j_1-1}\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)=1$, that is,
$\alpha_{j_1-1} < \frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k$. Per the updating rules
on the step size, this implies
\[
\alpha_{j_1} \le \gamma_{\rm inc} \alpha_{j_1-1} < \gamma_{\rm dec} \overline{\alpha}_k,
\]
hence $\frac{\alpha_{j_1}}{\gamma_{\rm dec}\overline{\alpha}_k} \le 1$ and thus
$\log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_{j_1}}{\gamma_{\rm dec}\overline{\alpha}_k}\right) < 0$.
Overall, we thus obtain that
\[
\sum_{j=j_1}^{j_2} (1-v_j(\overline{\alpha}_k)) (1-w_j)
\le
\left\{
\begin{array}{ll}
\frac{1}{\mu}\sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right)
&\mbox{if\ $j_1=0$} \\
\frac{1}{\mu}\sum_{j=j_1}^{j_2}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
&\mbox{otherwise.}
\end{array}
\right.
\]
Finally, we note that
\[
\left\{ j\ \middle|\ (1-v_j(\overline{\alpha}_k)) (1-w_j)=1 \right\}
\; \subset \;
\left\{ j\ \middle|\ 1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)=1 \right\},
\]
on equivalently, that every $j\in\{0,\ldots,k-1\}$ for which $(1-v_j(\overline{\alpha}_k)) (1-w_j)=1$
is in such a subsequence $j_1,\ldots,j_2$.
This allows us to conclude
\[
\sum_{j=0}^{k-1} (1-v_j(\overline{\alpha}_k)) (1-w_j)
\le \frac{1}{\mu}\sum_{j=0}^{k-1}
\left(1-v_j\left(\tfrac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right).
\]
\end{proof}
The results of Lemmas~\ref{le:sds:bndwv} and~\ref{le:sds:bnd1w1v} are
sufficient to obtain a bound on the number of iterations for which the
directions are generated from both a descent set and a well-aligned subspace
matrix.
\begin{proposition}
\label{pr:sds:bndz}
For any realization of Algorithm~\ref{algo:sds} and any positive integer
$k$,
\begin{equation}
\label{eq:sds:sumindic}
\sum_{j=0}^{k-1} z_{j} \le
\frac{(1-p_0)C}
{\min\{\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
+ (1-p_0)\log_{\gamma_{\rm dec}^{-1}}\left(\frac{\gamma_{\rm inc}\alpha_0}{\gamma_{\rm dec}^2
\min\{\gamma_{\rm inc}^{-1}\alpha_0,
\bar{\alpha} \tilde{g}_k\}} \right)
+ p_0 k,
\end{equation}
where $\tilde{g}_k = \min_{0 \le j \le k-1} \|\nabla f(\vct{x}_j)\|$,
\begin{equation}
\label{eq:sds:constant}
C :=
\frac{(\mu\gamma_{\rm dec}^2+\gamma_{\rm inc}^2)\gamma_{\rm inc}^2 \beta}{ \mu\gamma_{\rm dec}^4},
\end{equation}
and
\begin{equation}
\label{eq:sds:p0}
p_0 := \max\left\{
\frac{\ln(\gamma_{\rm dec})}{\ln(\gamma_{\rm inc}^{-1}\gamma_{\rm dec})},
\frac{\ln(\gamma_{\rm inc})}{\ln(\gamma_{\rm inc}\gamma_{\rm dec}^{-1})} \right\} =
\max\left\{\frac{1}{1+\mu},\frac{\mu}{1+\mu}\right\}.
\end{equation}
\end{proposition}
\begin{proof}
For any $j=0,\dots,k-1$, we have
\begin{equation}
\label{eq:sds:zdecomp}
z_j = z_j v_j(\overline{\alpha}_k) w_j + z_j (1-v_j(\overline{\alpha}_k))(1-w_j)
+ z_j (1-v_j(\overline{\alpha}_k))w_j.
\end{equation}
Indeed, the bound clearly holds when $z_j=0$. Assuming $z_j=1$,
it also holds when $v_j(\overline{\alpha}_k) w_j=1$ or $(1-v_j(\overline{\alpha}_k))(1-w_j)=1$, while
Lemma~\ref{le:sds:succit} implies that we cannot have $z_j v_j(\overline{\alpha}_k) (1-w_j)=1$. This last property will be instrumental in the proof
of our result.
Summing~\eqref{eq:sds:zdecomp} over all $j=0,\dots,k-1$, we obtain:
\begin{equation}
\label{eq:sds:sumzdecomp}
\sum_{j=0}^{k-1} z_j
= \sum_{j=0}^{k-1} z_j (1-v_j(\overline{\alpha}_k))w_j
+ \sum_{j=0}^{k-1} z_j v_j(\overline{\alpha}_k) w_j
+ \sum_{j=0}^{k-1} z_j (1-v_j(\overline{\alpha}_k)) (1-w_j).
\end{equation}
We will provide separate bounds on the three sums on the right-hand side
of~\eqref{eq:sds:sumzdecomp}.
Consider first an index such that $z_j\,(1-v_j(\overline{\alpha}_k))\,w_j=1$. By
definition of $v_j(\overline{\alpha}_k)$, we obtain from
$1-v_j(\overline{\alpha}_k)=1$ that
\[
1
\le \left(\frac{\alpha_j}{\overline{\alpha}_k}\right)^2
= \frac{\gamma_{\rm inc}^2\alpha_j^2}{\gamma_{\rm dec}^2\min\{\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 \min_{0 \le \ell \le k-1}\|\nabla f(\vct{x}_{\ell})\|^2\}}.
\]
Thus,
\begin{eqnarray*}
z_j\,(1-v_j(\overline{\alpha}_k))\,w_j
\le \frac{z_j w_j \gamma_{\rm inc}^2\alpha_j^2}{\gamma_{\rm dec}^2\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}} \\
\end{eqnarray*}
and summing over all indices up to $k-1$ gives
\begin{equation}
\label{eq:sds:sumz1vw}
\sum_{j=0}^{k-1} z_j\,(1-v_j(\overline{\alpha}_k))\,w_j
\le \frac{\gamma_{\rm inc}^2}{\gamma_{\rm dec}^2\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
\sum_{j=0}^{k-1} z_j w_j \alpha_j^2
\le \frac{\gamma_{\rm inc}^2 \beta}{\gamma_{\rm dec}^2\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}},
\end{equation}
where the last inequality follows from Lemma~\ref{le:sds:sumzalpha}.
We now bound the second term on the right-hand side
of~\eqref{eq:sds:sumzdecomp}. Thanks to Lemma~\ref{le:sds:bndwv}, we
have
\begin{eqnarray*}
\sum_{j=0}^{k-1} z_j v_j(\overline{\alpha}_k) w_j
&\le &\sum_{j=0}^{k-1} v_j(\overline{\alpha}_k) w_j \\
&\le
&\mu \sum_{j=0}^{k-1} v_j\left(\frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}
\overline{\alpha}_k\right) (1-w_j) \\
&=
&\mu\left[
\sum_{j=0}^{k-1} z_j v_j\left(\frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}
\overline{\alpha}_k\right) (1-w_j)
+
\sum_{j=0}^{k-1} (1-z_j) v_j\left(\frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}
\overline{\alpha}_k\right) (1-w_j) \right] \\
&\le
&\mu\left[
\sum_{j=0}^{k-1} z_j v_j\left(\frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}
\overline{\alpha}_k\right) (1-w_j)
+
\sum_{j=0}^{k-1} (1-z_j) (1-w_j) \right] \\
\end{eqnarray*}
Now, for any $j=0,\dots,k-1$, we have
\[
z_j v_j\left(\frac{\gamma_{\rm inc}}{\gamma_{\rm dec}}
\overline{\alpha}_k\right) (1-w_j)
\le z_j v_j(\bar{\alpha} \tilde{g}_k) (1-w_j) = 0,
\]
hence the first sum in the above expression is always zero.
Thus,
\begin{equation}
\label{eq:sds:sumzvw}
\sum_{j=0}^{k-1} z_j v_j(\overline{\alpha}_k) w_j
\le
\mu \sum_{j=0}^{k-1} (1-z_j)(1-w_j).
\end{equation}
We finally consider the last sum in the right-hand side
of~\eqref{eq:sds:sumzdecomp}. Applying Lemma~\ref{le:sds:bnd1w1v},
we get
\begin{eqnarray*}
\sum_{j=0}^{k-1} z_j (1-v_j(\overline{\alpha}_k)) (1-w_j)
&\le
&\sum_{j=0}^{k-1} (1-v_j(\overline{\alpha}_k)) (1-w_j) \\
&\le
&\frac{1}{\mu}\sum_{j=0}^{k-1} \left(1-v_j\left(
\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)
w_j + \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right)\\
&=
&\frac{1}{\mu}\sum_{j=0}^{k-1} z_j \left(1-v_j\left(
\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j \\
&
&+\frac{1}{\mu}\sum_{j=0}^{k-1} (1-z_j) \left(1-v_j\left(
\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right)\\
&\le
&\frac{1}{\mu}\sum_{j=0}^{k-1} z_j \left(1-v_j\left(
\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
\\
&
&+\frac{1}{\mu}\sum_{j=0}^{k-1} (1-z_j) w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right).
\end{eqnarray*}
By the same reasoning used to obtain~\eqref{eq:sds:sumz1vw}, we can
bound the first sum in the last expression as follows:
\[
\sum_{j=0}^{k-1} z_j \left(1-v_j\left(
\frac{\gamma_{\rm dec}}{\gamma_{\rm inc}}\overline{\alpha}_k\right)\right)w_j
\le
\frac{\gamma_{\rm inc}^4 \beta}{\gamma_{\rm dec}^4\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 \|\tilde{\vct{g}}_k\|^2\}}.
\]
Therefore, the following bound holds:
\begin{equation}
\label{eq:sds:sumz1v1w}
\sum_{j=0}^{k-1} z_j (1-v_j(\overline{\alpha}_k)) (1-w_j)
\le \frac{\gamma_{\rm inc}^4 \beta}{\mu\gamma_{\rm dec}^4\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 \|\tilde{\vct{g}}_k\|^2\}}
+ \frac{1}{\mu}\sum_{j=0}^{k-1} (1-z_j) w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right).
\end{equation}
To conclude the proof, we combine~\eqref{eq:sds:sumzdecomp},
\eqref{eq:sds:sumz1vw}, \eqref{eq:sds:sumzvw} and~\eqref{eq:sds:sumz1v1w}
as follows:
\begin{eqnarray*}
\sum_{j=0}^{k-1} z_j
&\le
&\sum_{j=0}^{k-1} z_j (1-v_j(\overline{\alpha}_k))w_j
+ \sum_{j=0}^{k-1} z_j v_j(\overline{\alpha}_k) w_j
+ \sum_{j=0}^{k-1} z_j (1-v_j(\overline{\alpha}_k)) (1-w_j) \\
&\le
&\frac{\gamma_{\rm inc}^2 \beta}{\gamma_{\rm dec}^2\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
+\mu \sum_{j=0}^{k-1} (1-z_j)(1-w_j)
+\frac{\gamma_{\rm inc}^4 \beta}{\mu\gamma_{\rm dec}^4\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}} \\
&
&+ \frac{1}{\mu}\sum_{j=0}^{k-1} (1-z_j) w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right) \\
&=
&\frac{(\mu\gamma_{\rm dec}^2+\gamma_{\rm inc}^2)\gamma_{\rm inc}^2 \beta}{ \mu\gamma_{\rm dec}^4\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
+ \mu \sum_{j=0}^{k-1} (1-z_j)(1-w_j) \\
&
&+ \frac{1}{\mu}\sum_{j=0}^{k-1} (1-z_j) w_j
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right) \\
&\le
&\frac{(\mu\gamma_{\rm dec}^2+\gamma_{\rm inc}^2)\gamma_{\rm inc}^2 \beta}{ \mu\gamma_{\rm dec}^4\min\{
\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
+ \max\left\{\mu,\frac{1}{\mu}\right\} \sum_{j=0}^{k-1}(1-z_j)
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right) \\
&=
&\frac{C}{\min\{\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
+ \max\left\{\mu,\frac{1}{\mu}\right\} \sum_{j=0}^{k-1}(1-z_j)
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right),
\end{eqnarray*}
where the last line simply uses the formula~\eqref{eq:sds:constant} for
$C$. Re-arranging the terms gives
\[
\sum_{j=0}^{k-1} z_j
\le
\frac{\max\left\{\mu,\tfrac{1}{\mu}\right\}}
{1+\max\left\{\mu,\tfrac{1}{\mu}\right\}} k
+\frac{1}{1+\max\left\{\mu,\tfrac{1}{\mu}\right\}}
\left[\frac{C}{\min\{\gamma_{\rm inc}^{-2}\alpha_0^2,
\bar{\alpha}^2 [\tilde{g}_k]^2\}}
+ \log_{\gamma_{\rm dec}^{-1}}\left(\frac{\alpha_0}{\gamma_{\rm dec}\overline{\alpha}_k} \right)
\right].
\]
By using the formula~\eqref{eq:sds:kalpha} for $\overline{\alpha}_k$ and observing that
\begin{eqnarray*}
\frac{\max\left\{\mu,\tfrac{1}{\mu}\right\}}
{1+\max\left\{\mu,\tfrac{1}{\mu}\right\}}
&= &\max\left\{\frac{\mu}{1+\mu},\frac{1}{1+\mu}\right\}
= p_0
\end{eqnarray*}
we arrive at the desired result.
\end{proof}
Note that Proposition~\ref{pr:sds:bndz} yields an alternate definition of
$p_0$ than that identified by Gratton et
al~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2019}, which was
$\tfrac{\log(\gamma_{\rm dec})}{\log(\gamma_{\rm inc}^{-1}\gamma_{\rm dec})}$. However, we point out
that both terms in~\eqref{eq:sds:p0} are equal whenever
$\gamma_{\rm inc} = \gamma_{\rm dec}^{-1}$, corresponding to $p_0=\tfrac{1}{2}$: this
particular setting has been widely used in analyzing derivative-free
methods based on probabilistic properties~\cite{CCartis_KScheinberg_2018}.
In addition to the result of Proposition~\ref{pr:sds:bndz}, we also
have the following concentration inequality on the sum of the $z_k$ variables.
\begin{lemma}
\label{le:sds:conineq}
Consider the sequences $\{\mtx{D}_k\}_k$ and $\{\mtx{P}_k\}_k$ generated by
Algorithm~\ref{algo:sds}, and suppose that the sequences are
$(\kappa,D_{\max},p)$-descent and $(\eta,\sigma,P_{\max},q)$-well-aligned,
respectively.
Let $\pi_k(\lambda):=\P{\sum_{j=0}^{k-1} Z_j \le \lambda k}$
for any $\lambda \in (0,pq)$. Then,
\begin{equation}
\label{eq:sds!conineq}
\pi_k(\lambda) \le \exp\left[-\frac{(pq-\lambda)^2}{2 pq}k\right].
\end{equation}
\end{lemma}
The reasoning behind this result is the same than Lemma 4.5 of Gratton
et al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}, except that the
variable $Z_k$ involves two random events. To connect it with our
probabilistic properties of interest, one must use
\begin{eqnarray*}
\P{Z_k=1 | Z_0,\dots,Z_{k-1}}
&=
&\P{\mathcal{D}_k\ \mbox{is ($\kappa,D_{\max}$)-descent}\ \&\
\mathcal{D}_k\ \mbox{is ($\kappa,D_{\max}$)-descent}|\mathcal{F}_{k-1}} \\
&=
&\P{\mathcal{D}_k\ \mbox{is ($\kappa,D_{\max}$)-descent} |\mathcal{F}_{k-1/2}} \\
&
&\times \P{\mathcal{D}_k\ \mbox{is ($\kappa,D_{\max}$)-descent} |\mathcal{F}_{k-1}},
\end{eqnarray*}
which holds because the events on $\mathcal{D}_k$ and $\mtx{P}_k$ are conditionally
independent.
Combining the results of Proposition~\ref{pr:sds:bndz} and
Lemma~\ref{le:sds:conineq}
lead to our main high-probability complexity result. The proof of this
result follows from that of Theorem 4.6 in Gratton et
al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}, and merely differs
by the presence of the additional logarithmic term in~\eqref{eq:sds:sumindic}.
\begin{theorem}
\label{th:highprobawcc}
Suppose that the sequences $\{\mtx{D}_k\}_k$ and $\{\mtx{P}_k\}_k$ generated by
Algorithm~\ref{algo:sds} are $(\kappa,D_{\max},p)$-descent and
$(\eta,\sigma,P_{\max},q)$-well-aligned, respectively. Suppose further that
$pq>p_0$, where $p_0$ is defined as in Proposition~\ref{pr:sds:bndz}, and
consider an index $k$ and a tolerance $\epsilon>0$ such that
\begin{equation}
\label{eq:largekwcc}
k \ge \frac{2}{pq-p_0}\left[
\frac{(1-p_0)C}{\bar{\alpha}^2}\epsilon^{-2}
+(1-p_0)\log_{\gamma_{\rm dec}}\left(\tfrac{\gamma_{\rm dec}^2 \bar{\alpha}}
{\gamma_{\rm inc}\alpha_0}\epsilon\right)
\right]
\end{equation}
with $C$ defined as in Proposition~\ref{pr:sds:bndz}, and
\begin{equation}
\label{eq:smallepswcc}
\epsilon \le \min\left\{1,
\frac{\gamma_{\rm inc}^{-1}\alpha_0}
{\bar{\alpha}} \right\}.
\end{equation}
Then,
\begin{equation}
\label{eq:highprobawcc}
\P{\tilde{G}_k \le \epsilon} \ge 1 - \exp\left[
-\frac{(pq-p_0)^2}{8pq} k \right],
\end{equation}
where $\tilde{G}_k$ is the random variable associated with $\tilde{g}_k$.
\end{theorem}
When $r=n$ and $\mtx{P}_k=\mtx{I}_n$, our result is of the same order than
that of Gratton et al~\cite[Theorem 4.6]{SGratton_CWRoyer_LNVicente_ZZhang_2015}.
The additional logarithmic term in~\eqref{eq:largekwcc} can be
viewed as an additional cost incurred by the generalization of
the reasoning to handle unbounded directions and random subspaces.
The result of Theorem~\ref{th:highprobawcc} can be declined in several
ways, depending on the quantity of interest (gradient norm, number of
iterations, number of iterations) and the type of result that is sought (high
probability, fixed probability guarantee, in expectation). We
provide below two results that are of particular interest to
our approach. The first one is a high-probability complexity bound
on the number of function evaluations required to reach an
approximate stationary point.
\begin{corollary}
\label{co:sds:wccevalsproba}
Under the assumptions of Theorem~\ref{th:highprobawcc}, let
$N_{\epsilon}$ be the number of function evaluations required by
Algorithm~\ref{algo:sds} to reach a point $\vct{x}_k$ such that
$\|\nabla f(\vct{x}_k)\| \le \epsilon$, where $\epsilon>0$
satisfies~\eqref{eq:smallepswcc}. Then,
\begin{equation}
\label{eq:sds:wccevalsproba}
\P{N_{\epsilon} \le \left\lceil
\frac{2m}{pq-p_0}\phi(\epsilon)
\right\rceil} \ge
1 - \exp\left[
-\frac{(pq-p_0)}{4pq}\phi(\epsilon) \right]
\end{equation}
where $\phi(\epsilon)= \frac{(1-p_0)C} {\bar{\alpha}^2}\epsilon^{-2}
+(1-p_0)\log_{\gamma_{\rm dec}}\left(\tfrac{\gamma_{\rm dec}^2 \bar{\alpha}}
{\gamma_{\rm inc}\alpha_0}\epsilon\right)$.
\end{corollary}
The proof of Corollary~\ref{co:sds:wccevalsproba} follows from
that of Theorem 4.8 in Gratton et
al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}. Its result shows that
Algorithm~\ref{algo:sds} has a high-probability complexity bound in
\begin{align}
\frac{2m}{pq-p_0}\phi(\epsilon) = \mathcal{O}\left(m \frac{1}{pq-p_0}
\eta^{-2} \sigma^{-2} P_{\max}^4 D_{\max}^8 \kappa^{-2} \epsilon^{-2}\right) \label{eq_asymptotic_eval_bound}
\end{align}
The second corollary of our main result illustrates how our analysis leads to
bounds in expectation: this result can be obtained following the argument
developed for derivative-free algorithms based on probabilistic
properties~\cite[Theorem 2.14]{SGratton_CWRoyer_LNVicente_ZZhang_2018}.
\begin{corollary}
\label{co:sds:wccevalsexp}
Under the assumptions of Theorem~\ref{th:highprobawcc},
\begin{equation}
\label{eq:sds:wccevalsexp}
\E{N_{\epsilon}} \le \frac{2 m}{pq-p_0} \phi(\epsilon)
+ \frac{1}{1-\exp\left(-\frac{(pq-p_0)^2}{8pq}\right)},
\end{equation}
where $N_{\epsilon}$ and $\phi(\epsilon)$ are defined as in
Corollary~\ref{co:sds:wccevalsproba}.
\end{corollary}
As $\phi(\epsilon) = \mathcal{O}(\epsilon^{-2})$ for $\epsilon<1$, we obtain a
complexity bound that matches that of other probabilistic techniques in terms
of dependencies on $\epsilon$~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2018}.
In addition, the dependencies on $m$, $\kappa$ and $\epsilon$ match that
obtained for direct search based on deterministic
descent~\eqref{eq:wccdsdeterm}. In the next two sections, we will establish
evaluation complexity bounds for several choices of subspaces and polling
sets.
\subsection{Examples of direction generation techniques}
\label{subsec:sds:ps}
The above analysis of Algorithm~\ref{algo:sds} requires that our poll
directions $\{\mtx{D}_k\}_k$ are $(\kappa,D_{\max},p)$-descent and that our
subspace matrices $\{\mtx{P}_k\}_k$ are $(\eta,\sigma,P_{\max},q)$-well-aligned.
We now discuss several approaches to satisfying these requirements, both
deterministic and probabilistic.
We begin by noting that our framework encompasses direct search based on
deterministic and probabilistic descent. Indeed, if we take $r=n$ and
$\mtx{P}_k=\mtx{I}_n$ for every $k$, $\mtx{P}_k$ is $(\eta,\sigma,P_{\max},q)$-well-aligned with
$\eta=\sigma=P_{\max}=q=1$, and Algorithm~\ref{algo:sds} reduces to
Algorithm~\ref{algo:ds}. We then recover classical, deterministic direct search by choosing
all $\mtx{D}_k$s to be the same PSS: the sequence $\{\mtx{D}_k\}$ is then
$(\kappa,D_{\max},p)$-descent with $p=1$, and $m \ge n+1$. Table 1 shows the
values of $m$, $\kappa$ and $D_{\max}$ for three popular approaches: the
coordinate vectors and their negatives, a set of $n+1$ vectors with uniform
angles~\cite[Chapter 2.1]{ARConn_KScheinberg_LNVicente_2009b}, or the
coordinate vectors and the vector with all entries $-1$.\footnote{We thank
Warren Hare and Gabriel Jarry-Bolduc for checking the value of $\kappa$ in
that latter case~\cite{WHare_GJarryBolduc_2020}.}
For probabilistic descent, we form $\mtx{D}_k$ by generating $m$ independent vectors
uniformly distributed on the unit sphere: tin that case, $\mtx{D}_k$ is
$(\kappa,D_{\max},p)$-descent with $D_{\max}=1$, $\kappa=\tau/\sqrt{n}$ and
$p= 1-\left(\frac{1}{2}+\frac{\tau}{\sqrt{2\pi}}\right)^m$ for any
$\tau \in [0,\sqrt{n}]$~\cite[Appendix B]{SGratton_CWRoyer_LNVicente_ZZhang_2015}.
Choosing $\tau=1$, we can make $p$ sufficiently large (so that it exceeds $p_0$)
by taking $m=\bigO(1)$. These choices are summarized in
Table~\ref{tab_pss_methods}.
\begin{table}[tb]
\centering
{\small
\begin{tabular}{c|ccccc}
\hline Method & $m$ & $\kappa$ & $D_{\max}$ & Success prob.~$p$ \\ \hline
$[{\bf I}_r,-{\bf I}_r]$ & $2r$ & $r^{-1/2}$ & 1 & 1 \\
Uniform angle PSS & $r+1$ & $r^{-1}$ & 1 & 1 \\
$[{\bf I}_r,-\vct{e}_r]$ & $r+1$ & $(r^2+2(r-1)\sqrt{r})^{-1/2}$ & $\sqrt{r}$ & 1 \\
Random unit vectors & $\bigO(1)$ & $r^{-1/2}$ & 1
& $1-(1/2 + 1/\sqrt{2\pi})^m$. \\
\hline
\end{tabular}
}
\caption{Summary of methods for generating a direction set $\mtx{D}_k$ in
$\mathbb{R}^r$.}
\label{tab_pss_methods}
\end{table}
We now explain how to generate $\mtx{P}_k$ matrices that are probabilistically
well-aligned (the previous paragraph gives an example of a deterministically
well-aligned matrix).
Consider first the requirements \eqref{eq:sds:Pk:grad}
and~\eqref{eq:sds:Pk:pmax}. Three possible approaches can be used:
\begin{itemize}
\item If $\mtx{P}_k$ has entries which are i.i.d.~$\mathcal{N}(0,1/r)$ and
$r = \Omega((1-\eta)^{-2}|\log (1-q)|)$, then $\mtx{P}_k$
satisfies \eqref{eq:sds:Pk:grad} and \eqref{eq:sds:Pk:pmax}~\cite[Theorem 2.13]{
SBoucheron_GLugosi_PMassart_2013} with
$P_{\max} = \Theta(\sqrt{n/r})$ with high probability~\cite[Corollary 3.11]{ASBandeira_RvHansel_2016}.
\item If $\mtx{P}_k$ is an $s$-hashing matrix\footnote{i.e.~every column of
$\mtx{P}_k$ has exactly $s$ nonzero entries at randomly selected locations,
each taking value $\pm 1/\sqrt{s}$ with independent probability $1/2$.}
with $s=\Theta((1-\eta)^{-1}|\log (1-q)|)$ and
$r = \Omega((1-\eta)^{-2}|\log (1-q)|)$, then $\mtx{P}_k$
satisfies \eqref{eq:sds:Pk:grad} and \eqref{eq:sds:Pk:pmax} for
$P_{\max}=\sqrt{n}$, where the value of $P_{\max}$ comes from
$\|\mtx{P}_k\|_2 \leq \|\mtx{P}_k\|_F = \sqrt{n}$~\cite{DKane_JNelson_2014}.
\item If $\mtx{P}_k = \sqrt{n/r} {\bf I}_{r\times n}\mtx{Q}^T $ where ${\bf I}_{r\times n}$
denotes the first $r$ rows of the $n\times n$ identity matrix $\mtx{I}_n$, and
$\mtx{Q}\in\mathbb{R}^{n\times n}$ is the orthogonal factor in the QR decomposition
$\mtx{Z}=\mtx{Q}\mtx{R} \in \mathbb{R}^{n\times n}$ of a matrix $\mtx{Z}$ with i.i.d. standard
normal entries such that the diagonal entries of $\mtx{R}$ are positive, then
$\|\mtx{P}_k\|_2 \leq \sqrt{n/r}$, so $\mtx{P}_k$ satisfies~\eqref{eq:sds:Pk:pmax}.
Moreover, for a given $r$ and $\eta$, $\mtx{P}_k$ satisfies~\eqref{eq:sds:Pk:grad}
with probability $q=1-I_{\eta r/d}(r/2, (n-r)/2)$, where $I_{p}(\alpha,\beta)$ is the
regularized incomplete Beta function
\cite[Lemma 1]{DKozak_SBecker_ADoostan_LTenorio_2021}.
Although this does not give a closed form for a suitable choice of $r$,
numerical evidence suggests that $r$ may be chosen independently of
$n$~\cite[Figure 1]{DKozak_SBecker_ADoostan_LTenorio_2021}.
\end{itemize}
We now focus on condition \eqref{eq:sds:Pk:PTd}, that corresponds
to bounding the smallest singular value of $\mtx{P}_k$ away from zero. For
Gaussian matrices with entries in $\mathcal{N}(0,1/r)$ as above, the
estimate $\sigma_{\min}(\mtx{P}_k^T) \sim \sqrt{n/r}-1$ holds with high
probability~\cite[Theorem 1.1]{MRudelson_RVershynin_2009}. In the case of
orthogonal subsampling (the third case above), we automatically have
$\sigma_{\min}(\mtx{P}_k^T)=\sqrt{n/r}$. We are unaware of theoretical
results showing that $\sigma_{\min}(\mtx{P}_k) = \Theta(\sqrt{n/r})$ for hashing
matrices, we present numerical evidence that this estimate holds when $n$
is sufficiently large compared to $r$ in Appendix~\ref{app_hashing_sing_val}.
\begin{table}[tb]
\centering
{\small
\begin{tabular}{c|ccccc}
\hline Method & $r$ & Success prob.~$q$ & $P_{\max}$ & $\sigma$ & Comments \\ \hline
Identity & $n$ & 1 & 1 & 1 & --- \\
Gaussian & $\bigO(1)$ & $1-e^{-\frac{(1-\eta)^2}{Ar}}$ & $\Theta(\sqrt{n/r})$ & $\Theta(\sqrt{n/r})$ & $P_{\max}$, $\sigma$ valid with high prob. \\
$s$-Hashing & $\bigO(1)$ & $1-e^{-\frac{(1-\eta)^2}{Ar}}$ & $\sqrt{n}$ & $\Theta(\sqrt{n/r})$ & $s=\Theta(\frac{|\log(1-q)|}{1-\eta})$, $\sigma$ from App.~\ref{app_hashing_sing_val} \\
Orthogonal & $\bigO(1)$ & $1-I_{\eta r/d}(\frac{r}{2},\frac{n-r}{2})$ & $\sqrt{n/r}$ & $\sqrt{n/r}$ & $r$ estimated from \cite[Figure 1]{DKozak_SBecker_ADoostan_LTenorio_2021}. \\
\hline
\end{tabular}
}
\caption{Summary of methods for generating $\mtx{P}_k$. The value $A$ is used to
represent a universal constant.}
\label{tab_subspace_methods}
\end{table}
Table~\ref{tab_subspace_methods} summarizes our results for the three random choices of
$\mtx{P}_k$ described above: for such choices, since $r$ can be chosen as small compared to
the ambient dimension $n$, we may choose $\mtx{D}_k$ to be a standard PSS in $\mathbb{R}^r$, such as
$[{\bf I}_r\ -{\bf I}_r]$. As result, our poll step tests the points
$\vct{x}_k \pm \alpha_k \vct{p}_{k,i}$, where $\vct{p}_{k,i}$ is the $i$-th row of $\mtx{P}_k$. In that
case, we obtain directions that are very similar to those used in Gratton et
al.~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015}. In fact, using $r=1$ and orthogonal
$\mtx{P}_k$ corresponds to using $\{\pm\sqrt{n}\ \vct{v}\}$, where $\vct{v}\in\mathbb{R}^n$ is a randomly
drawn unit vector. Up to a scaling constant, this recovers the method described in
Gratton et al.~\cite[p.~1535]{SGratton_CWRoyer_LNVicente_ZZhang_2015}. Our framework
is however more general: for instance, using a Gaussian matrix $\mtx{P}_k$ with $r=1$ leads
to the directions $\{\pm\vct{v}\}$, where $\vct{v}\in\mathbb{R}^n$ is a vector with i.i.d. standard
Gaussian components.
\begin{remark}
Compared to traditional direct search, the use of random subspaces defined by
$\mtx{P}_k$ increases the per-iteration linear algebra cost of Algorithm~\ref{algo:sds}:
however, this cost depends on the ensemble from which $\mtx{P}_k$ is generated.
For instance, a Gaussian $\mtx{P}_k$ would have a per-iteration cost of $\bigO(nr)$ to
generate the matrix, and $\bigO(mnr)$ to construct each $\mtx{P}_k^T \vct{d}$, noting that
$m=\bigO(r)$ for most methods considered above. By contrast, an orthogonal $\mtx{P}_k$
requires $\bigO(nr^2)$ to generate $\mtx{P}_k$ via QR factorization, and again
$\bigO(mnr)$ to construct $\mtx{P}_k^T \vct{d}$.
Finally, using hashing to generate $\mtx{P}_k$ would take approximately
$\bigO(ns)$ work, depending on the specific method used to select the nonzero
locations, and only $\bigO(m\,\operatorname{nnz}(\mtx{P}_k))=\bigO(mns )$ to
construct $\mtx{P}_k^T \vct{d}$ using sparse multiplication. Note that this cost is lower
when $\mtx{D}_k$ is sparse, e.g. $\mtx{D}_k = [{\bf I}_r\ -{\bf I}_r]$.
\end{remark}
\subsection{Evaluation complexity results}
\label{sec:sds:evalwcc}
As illustrated in Section~\ref{subsec:sds:ps}, the framework given by
Algorithm~\ref{algo:sds} may be implemented in various ways. We now compare
several instances of this method using the evaluation complexity
bound~\eqref{eq_asymptotic_eval_bound} derived in Section~\ref{subsec:sds:wcc}.
We pay particular attention to the dependency on $n$ that arises as we
instantiate the method, since our reduced space approach aims at reducing this
dependency.
Table~\ref{tab_algo_complexities} highlights the dependency on $n$ for several
variants of the method. Note that for classical direct search ($\mtx{P}_k={\bf I}_n$,
$\mtx{D}_k$ deterministic), there is a substantial difference in complexity
depending on the choice of $\mtx{D}_k$. On the contrary, using a randomized
subspace choice $\mtx{P}_k$ always leads to an evaluation complexity in
$\bigO(n\epsilon^{-2})$, independently of the choice of $\mtx{D}_k$.
Our result recovers the $\bigO(n\epsilon^{-2})$ complexity from direct search
based on probabilistic descent~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2015},
and shows how the same result may be achieved with substantially greater
flexibility. It also gives further insight into the choice $\mtx{P}_k={\bf I}_n$ and
$\mtx{D}_k=\{\pm\vct{v}\}$ where $\vct{v}$ is a random unit vector, briefly discussed in
Gratton et al.~\cite[p.~1535]{SGratton_CWRoyer_LNVicente_ZZhang_2015}. As
explained above, this choice may be viewed as taking $\mtx{P}_k$ to be orthogonal
with $r=1$ and $\mtx{D}_k=[{\bf I}_n,-{\bf I}_n]$.
\begin{table}[tb]
\centering
{\small
\begin{tabular}{c|cccc}
\hline
$\mtx{D}_k$ choice $\backslash$ $\mtx{P}_k$ choice
& Identity & Gaussian & $s$-Hashing & Orthogonal \\
\hline
$[{\bf I},-{\bf I}]$ & $n^2$ & $n$ & $n$ & $n$ \\
Uniform angle PSS & $n^3$ & $n$ & $n$ & $n$ \\
$[{\bf I},-\vct{e}]$ & $n^7$ & $n$ & $n$ & $n$ \\
Random unit vectors & $n$ & $n$ & $n$ & $n$ \\
\hline
\end{tabular}
}
\caption{Summary of evaluation complexity dependency on $n$ for
different choices of $\mtx{P}_k$ and $\mtx{D}_k$ in Algorithm~\ref{algo:sds}.}
\label{tab_algo_complexities}
\end{table}
We conclude this section by commenting on the connections between our
results and that obtained for model-based techniques. Unlike for direct search,
in the model-based setting, using models that are of probabilistically
good quality (or probabilistically fully
linear~\cite{ARConn_KScheinberg_LNVicente_2009b}) is not known to improve the
dependence on the dimension~\cite{SGratton_CWRoyer_LNVicente_ZZhang_2018}.
However, such an improvement can be obtained by using deterministically fully
linear models in randomly drawn subspaces~\cite{CCartis_LRoberts_2021}. This
result, together with the complexity analysis of this section, suggests that
using random subspace exploration is a more general approach method for
producing the complexity improvement observed in both direct-search and
model-based techniques.
\section{Numerical experiments}
\label{sec:num}
In this section, we investigate the practical performance of
Algorithm~\ref{algo:sds} by comparing the following direct-search methods:
\begin{itemize}
\item Algorithm~\ref{algo:ds} with deterministic descent, setting
$\mathcal{D}_k$ as either the columns of $[{\bf I}_n\ -{\bf I}_n]$ or
$\mathcal{D}_k=[{\bf I}_n\ -\vct{e}_n]$, where $\vct{e}_n\in\mathbb{R}^n$ is the vector of ones;
\item Algorithm~\ref{algo:ds} with probabilistic descent, using
$\mathcal{D}_k = \{\pm \vct{v}\}$ where $\vct{v}\in\mathbb{R}^n$ is drawn from the
uniform distribution on the unit sphere $\|\vct{v}\|=1$;
\item The Stochastic Three-Points (STP) algorithm from Bergou et
al.\cite{EBergou_EGorbunov_PRichtarik_2020}, that relies on opposite
directions uniform on the sphere, with step size schedule
$\alpha_k = \alpha_0 / (k+1)$;
\item Algorithm~\ref{algo:sds} with $\mtx{P}_k$ taken either to be a
Gaussian matrix with $r\in\{1,2,3,4,5\}$, a hashing matrix, or a
subsampled orthogonal matrix with $r\in\{1,5\}$. The poll directions
were chosen to be columns of $[{\bf I}_r\ -{\bf I}_r]$.
\end{itemize}
All methods used $\alpha_0=1$ and terminated if $\alpha_k < 10^{-6}$.
The direct search methods used $\gamma_{\rm inc}=2$, $\gamma_{\rm dec}=0.5$, $\alpha_{\max}=1000$, and used opportunistic polling\footnote{i.e.~Do not check any other poll directions as soon as a $\vct{d}_k$ satisfying \eqref{eq_new_sufficient_decrease} is found.} with the sufficient decrease condition
\begin{align}
f(\vct{x}_k + \alpha_k \mtx{P}_k^T \vct{d}_k) < f(\vct{x}_k) - \min\left(10^{-5}, 10^{-5} \alpha_k^2 \|\mtx{P}_k^T \vct{d}_k\|^2\right), \label{eq_new_sufficient_decrease}
\end{align}
where $\mtx{P}_k=\mtx{I}$ for standard and probabilistic direct search.
All methods used in our experiments have been implemented within a
Python package that has been made publicly
available.\footnote{\url{https://github.com/lindonroberts/directsearch}}
\subsection{Robust Regression}
\label{sec_robust_regression}
Our first set of experiments considers a robust linear regression problem with
Gaussian data and a significant fraction of outliers, described by Carmon et
al.~\cite{YCarmon_JDuchi_OHinder_ASidford_2017}. Specifically, we perform
linear regression using a smoothed biweight loss function:
\begin{align}
f(\vct{x}) = \frac{1}{m}\sum_{i=1}^{m} \phi(\vct{a}_i^T\vct{x}-b_i),
\qquad \text{where} \qquad \phi(\theta) := \frac{\theta^2}{1+\theta^2}. \label{eq_nonconvex_regression}
\end{align}
The vectors $\vct{a}_i\sim N(0,{\bf I}_n)$ are i.i.d.~Gaussian vectors and the values
$b_i$ are the coordinates of $\vct{b}=\mtx{A}\vct{z} + 3\vct{u}_1+\vct{u}_2$, where $\mtx{A}$ has
columns $\vct{a}_i$, $\vct{z}\sim N(0,4{\bf I}_m)$, $\vct{u}_1\sim N(0,{\bf I}_m)$ and the entries of
$\vct{u}_2$ are i.i.d.~Bernoulli distributed with $p=0.3$. We choose $n=100$,
$m=200$, start all solvers from the origin and run each solver for a maximum
of $50(n+1)$ objective evaluations. For this initial study, we only compare
direct search with $2n$ poll directions, probabilistic direct search, STP and
Algorithm~\ref{algo:sds} with Gaussian $\mtx{P}_k$ and $r=1$. All randomized
methods are run 10 times for a given problem.
\begin{figure}[tbh]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./nonconvex_regression_d100_seed3_alt_rho_avg.pdf}
\caption{First problem instance}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./nonconvex_regression_d100_seed5_alt_rho_avg.pdf}
\caption{Second problem instance}
\end{subfigure}
\caption{Objective value versus number of objective evaluations (in units of $n+1$) for the nonconvex regression problem \eqref{eq_nonconvex_regression}. For randomized methods, we show the average value achieved over 10 instances.}
\label{fig_nonconvex_regression}
\end{figure}
In Figure~\ref{fig_nonconvex_regression}, we plot the average objective
value reached (over the 10 runs per solver) versus the number of
objective evaluations normalized in terms of simplex gradients (i.e.~units
of $n+1$ evaluations). We show these results for two instances of
problem~\eqref{eq_nonconvex_regression}, with different choices of $\vct{a}_i$ and
$\vct{b}$. All three randomized variants based on Algorithm~\ref{algo:sds}
substantially outperform the traditional (deterministic) direct search method.
Moreover, using randomized subspaces (effectively using Gaussian poll
directions) leads to better results than using unit vectors as poll
directions, as done in probabilistic direct search and STP.
\subsection{CUTEst Collections}
\label{sec_cutest}
Our second set of experiments is based on two collections of problems from
the CUTEst optimization test
set~\cite{NIMGould_DOrban_PhLToint_2015}:
\begin{itemize}
\item The \textit{CFMR} collection \cite[Section 7]{CCartis_JFiala_BMarteau_LRoberts_2019} of 90 medium-scale problems ($25 \leq n \leq 120$, with $n\approx 100$ for 67 problems, approximately 75\%), ignoring bound constraints if any;
\item The \textit{CR-large} collection \cite[Appendix B]{CCartis_LRoberts_2021} of 28 large-scale problems ($1000 \leq n \leq 5000$).
\end{itemize}
All solvers were run for at most $200(n+1)$ objective evaluations for the
CFMR collection, and at most $10(n+1)$ evaluations for the CR-large collection.
As above, all randomized methods were run 10 times on every problem.
In this case, for each problem $P$ and solver instance $\mathcal{S}$, we measure
the number of objective evaluations $N_{P,\mathcal{S}}$ required to achieve
\begin{align}
f(\vct{x}_k) \leq f^* + \tau(f(\vct{x}_0) - f^*),
\end{align}
where $\tau\in(0,1)$ measures the required accuracy and $f^*$ is the true
minimum objective value for each problem.
If this accuracy level is never achieved for a particular solver instance
(within the desired budget or before termination), we take the convention
$N_{P,\mathcal{S}}=\infty$. We then compare solver performance using performance
profiles \cite{EDolan_JMore_2002}, which, for a given solver $\mathcal{S}$, plots
the fraction of test problems for which $N_{P,\mathcal{S}}$ is within some factor of
$\min_{\mathcal{S}} N_{P,\mathcal{S}}$ (averaged over all solver instances).
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./r_comparison_gsn_cutest100_budget200_perf1.pdf}
\caption{Medium-scale (CFMR), $\tau=10^{-1}$}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./r_comparison_gsn_cutest100_budget200_perf3.pdf}
\caption{Medium-scale (CFMR), $\tau=10^{-3}$}
\end{subfigure}
\\
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./r_comparison_gsn_cutest1000_budget10_perf1.pdf}
\caption{Large-scale (CR-large), $\tau=10^{-1}$}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./r_comparison_gsn_cutest1000_budget10_perf3.pdf}
\caption{Large-scale (CR-large), $\tau=10^{-3}$}
\end{subfigure}
\caption{Comparison of $r$ values for Gaussian sketching}
\label{fig_r_comparison}
\end{figure}
Figure~\ref{fig_r_comparison} compares Algorithm~\ref{algo:sds} with Gaussian
$\mtx{P}_k$ for different values of $r$, for both CFMR and CR-large test sets and
accuracy levels $\tau\in\{10^{-1}, 10^{-3}\}$. In the case of the medium-scale
CFMR problems, we see that smaller $r$ values give the best performance, but
all values are able to solve essentially the same number of problems.
In the case of large-scale CR-large problems, again smaller $r$ values give
the best performance, but we find that larger $r$ values are less robust in
terms of the total number of problems solved (within the given evaluation
budget).
Motivated by these results, in Figure~\ref{fig_solver_comparison} we compare
Algorithm~\ref{algo:sds} (with Gaussian and hashing $\mtx{P}_k$ and $r=1$) with
the other direct-search variants.
For both problem sets and accuracy levels, the best-performing methods are
probabilistic direct search (Algorithm~\ref{algo:ds} with uniformly random
directions) and the two Algorithm~\ref{algo:sds} variants, which for larger
performance ratios strongly outperform deterministic direct search and STP.
Comparing Algorithm~\ref{algo:sds} with probabilistic direct search, we see
broadly a very similar level of performance, reflecting the similarity of the
two methods (both randomized with at most two objective evaluations per
iteration at $\vct{x}_k\pm\alpha_k \vct{v}$ for a random vector $\vct{v}$).
However, for low-accuracy solutions to the larger test problems CR-large, we
see that allowing non-unit poll directions in Algorithm~\ref{algo:sds} gives
a notable performance improvement over probabilistic direct search.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./paper_comparison_cutest100_budget200_perf1.pdf}
\caption{Medium-scale (CFMR), $\tau=10^{-1}$}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./paper_comparison_cutest100_budget200_perf3.pdf}
\caption{Medium-scale (CFMR), $\tau=10^{-3}$}
\end{subfigure}
\\
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./paper_comparison_cutest1000_budget10_perf1.pdf}
\caption{Large-scale (CR-large), $\tau=10^{-1}$}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{./paper_comparison_cutest1000_budget10_perf3.pdf}
\caption{Large-scale (CR-large), $\tau=10^{-3}$}
\end{subfigure}
\caption{Comparison of different solvers}
\label{fig_solver_comparison}
\end{figure}
Overall, we find that Algorithm~\ref{algo:sds} performs similarly well to
probabilistic direct search in many instances, although there appears to be a
reasonable fraction of problems (in CUTEst and the nonconvex regression
problem) for which our new approach gives superior performance.
\section{Conclusion}
\label{sec:conc}
We have proposed a general direct-search framework equipped with complexity
guarantees that allows for polling in low-dimensional subspaces. We identified
properties of both the subspaces and the polling directions that enable the
derivation of complexity guarantees, which we expressed in a probabilistic
form. Both our theoretical analysis and our numerical experiments suggest that
using one-dimensional subspaces is an overall efficient approach, that allows
for applying direct-search methods to larger dimensional problems than usually
considered in the literature. Interestingly, the use of random embeddings
does not remove all dependencies on the original dimension of the problem,
but comes with the computational advantage of drawing only two directions
per iteration.
Our approach is based on characterizing properties of the polling directions
at the iteration level, while information from previous iterations is
typically used in practice to improve the polling set. On the other hand, removing our
conditional independence yields a number of mathematical challenges, that
would have to be overcome in order to obtain a complexity result. Similarly,
extending this method to handle noisy function values would require to
combine our analysis with existing frameworks for stochastic optimization.
Both avenues of research represent natural perspectives for future work.
\bibliographystyle{plain}
|
\section{Introduction}
\subsection{Background and the statement of the main result}
Let $(Y,\lambda)$ be a closed contact three manifold. $\lambda$ determines a vector field $X_{\lambda}$ which satisfies $i_{X_{\lambda}}\lambda=1$, $d\lambda(X_{\lambda},\,\,)=0$ and it is called the Reeb vector field on $(Y,\lambda)$. A smooth map $\gamma : \mathbb{R}/T\mathbb{Z} \to Y$ is called a Reeb orbit with periodic $T$ if $\dot{\gamma} = X_{\lambda}(\gamma)$ and simple if $\gamma$ is a embedding map. In this paper, two Reeb orbits are considered equivalent if they differ by reparametrization. If its return map $d\phi^{T}|_{\mathrm{Ker}\lambda=\xi} :\xi_{\gamma(0)} \to \xi_{\gamma(0)}$ has no eigenvalue 1, we call it a non-degenerate Reeb orbit and we call a contact manifold $(Y,\lambda)$ non-degenerate if all Reeb orbits are non-degenerate.
The three-dimensional Weinstein conjecture was shown by C. H. Taubes by using Seiberg Witten Floer (co)homology \cite{T1}. On the other hand, there is an usuful tool to research the properties of contact three manifolds, called Embedded contact homology(ECH), which was introduced by M. Hutchings in several papers (for example, it is briefly explained in \cite{H2}).
In \cite{CH}, D. Cristofaro-Gardiner and M. Hucthings showed that every contact closed three manifold $(Y,\lambda)$ has at least two simple periodic orbits by using the volume property with respect to ECH spectrums.
It will be summarized as follows.
\begin{them}
A contact (possibly degenerate) closed three manifold $(Y,\lambda)$ has at least two simple periodic orbit.
\end{them}
Non-degenerate periodic orbits are classified into three types according to their eigenvalues of their return maps. A periodic orbit is called a negative hyperbolic if $d\phi^{T}|_{\xi}$ has eigenvalues $h,h^{-1} < 0$, a positive hyperbolic if $d\phi^{T}|_{\xi}$ has eigenvalues $h,h^{-1} > 0$ and an elliptic if $d\phi^{T}|_{\xi}$ has eigenvalues $e^{\pm i2\pi\theta}$ for some $\theta \in \mathbb{R}\backslash \mathbb{Q}$.
As a refinement of the Weinstein conjecture, we can raise questions; Does $(Y,\lambda)$ have infinity many Reeb orbits? or if the number of Reeb orbits is finite, how many are there?
In association with this question, the following holds.
\begin{them}[\cite{HT3}]
Let $(Y,\lambda)$ be a closed contact non-degenerate three manifold. Assume that there exists exactly two simple Reeb orbit, then both of them are elliptic and $Y$ is a lens space.
\end{them}
\begin{them}[\cite{HCP}]\label{a}
Let $(Y,\lambda)$ be a non-degenerate contact three manifold. Let $\mathrm{Ker}\lambda=\xi$.
Then
\item[1.] if $c_{1}(\xi)$ is torsion, there exists infinity many periodic orbits, or there exists exactly two elliptic simple periodic orbits and $Y$ is diffeomorphic to a lens space.
\item[2.] if $c_{1}(\xi)$ is not torsion, there exists at least four periodic orbit.
\end{them}
In the above theorem, the ways of the proof are completely different if $c_{1}(\xi)$ is torsion or not.
If $c_{1}(\xi)$ is torsion, some indexes defined in ECH are well-controlled and this property plays an important role in the proof. But if $c_{1}(\xi)$ is not torsion, such a property does not hold. In \cite{HCP}, they proved this by splitting ECH into two parts $\mathrm{ECH}_{\mathrm{even}}$ and $\mathrm{ECH}_{\mathrm{odd}}$. In particular, if $b_{1}(Y)>0$, both elements called $U$-sequence does not vanish and hence by combining with the volume property of ECH spectrums, the proof was completed.
On the other hand, the above $\mathrm{ECH}_{\mathrm{odd}}$ is the part which detects the existence of a positive hyperbolic orbit and thus they also obtained the following theorem.
\begin{them}[\cite{HCP}]\label{0}
If $b_{1}(Y)>0$, there exists at least one positive hyperbolic orbit.
\end{them}
In general, they asked the following question.
\begin{Que}[\cite{HCP}]\label{ques}
Let $Y$ be a closed connected three-manifold which is not $S^3$ or a lens space, and let $\lambda$ be a nondegenerate contact form on $Y$. Does $\lambda$ have a positive hyperbolic simple Reeb orbit?
\end{Que}
This is a kind of natural refining the Weinstein conjecture in the sense of looking for Reeb orbits of particular types.
On contrary to the case $b_{1}(Y)>0$, if $b_{1}(Y)=0$, then $\mathrm{ECH}_{\mathrm{odd}}$ may vanish. So to answer this question, we have to use something different way.
The next theorem is the main theorem in this paper.
\begin{them}\label{main}
Let $(Y,\lambda)$ be a nondegenerate contact three manifold with $b_{1}(Y)=0$. Suppose that $(Y,\lambda)$ has infinity many simple periodic orbits (that is, $(Y,\lambda)$ is not a lens space with exactly two simple Reeb orbits) and has at least one elliptic orbit. Then, there exists at least one simple positive hyperbolic orbit.
\end{them}
We note that the existence of infinity many simple periodic orbits is a generic condition (see \cite{Ir}).
Properties of vector fields and maps with elliptic orbits have been classically studied for a long time related to classical mechanics and area preserving maps on surfaces and it has been known that such objects have many rich properties under some generic conditions (for example, see \cite{Ze}, \cite{Ne}, \cite{AMa} and their references).
In addition to this, the existence of an elliptic orbit also have been studied. For example, the conjecture that the Reeb flow of every convex hypersurface in $\mathbb{R}^{2n}$ carries an elliptic periodic orbit is in general open and has been researched under some additional hypothesises (for example, see \cite{AM1}, \cite{AM2}, \cite{DDE}, \cite{HrS}, \cite{Lo} and their references). Here, it is worth to note that we can apply Theorem \ref{main} to such already known cases.
Under Theorem \ref{main}, to complete the answer of Question \ref{ques}, it is sufficient to consider the next question.
\begin{Que}
Let $(Y,\lambda)$ be a non-degenerate contact three manifold with $b_{1}(Y) =0$. If all periodic orbits are hyperbolic, does there exist at least one simple positive
hyperbolic orbit?
\end{Que}
For example, many nondegenerate contact hyperbolic three manifolds whose all periodic orbits are hyperbolic were constructed in \cite{FHa} as contact Anosov flows. In associated with this, the next question occurs.
\begin{Que}
What kind of three manifold admits a contact form whose all periodic orbits are hyperbolic?
\end{Que}
In the proof of this paper, we will exclude a case such that exactly one simple orbit is elliptic and all the others are negative hyperbolic orbits. So the next question also occurs.
\begin{Que}
Are there any contact three manifolds such that exactly one simple orbit is elliptic and all the others are hyperbolic?
\end{Que}
Here, we introduce some results which will be proved in the forthcoming papers \cite{S1} and \cite{S2} by using methods and results in this paper.
\begin{enumerate}
\item[\bf 1.] Let $L(p,q)$ $(p\neq \pm 1)$ be a lens space and $\lambda$ be a non-degenerate contact form on it with infinity many simple periodic orbits. If $p$ is odd, then there is a simple positive hyperbolic orbit (\cite{S1}).
\item[\bf 2.] Let $(S^{3},\lambda)$ be a non-degenerate contact three sphere with infinity many simple orbits. If the Conley Zehnder index of a minimal periodic orbit with respect to the trivialization induced by a bounding disc is more than or equal to three, then there is a simple positive hyperbolic orbit. In particular, generic compact strictly convex energy surface in $\mathbb{R}^4$ carries a simple positive hyperbolic orbit and more generally, every non-degenerate dynamically convex contact three sphere with infinity many simple orbits has a positive hyperbolic orbit (\cite{S2}).
\end{enumerate}
The above notion of dynamically convex was introduced in \cite{HWZ}.
The behaviors of ECH are completely different whether elliptic orbits exist or not. In particular, the way in the present paper does not work if there is no elliptic orbit. But as results mentioned above, the author expects that the observations in this paper also furthermore help to study these kinds of question.
\subsection{Idea of the proof of the main theorem and the structure of this paper.}
First of all, in Section 2, we will review facts of Embedded contact homology that we will need in this paper.
To prove the main theorem, we divide the problem into two cases. One case is that $(Y,\lambda)$ has at least two simple elliptic orbits and the other is that $(Y,\lambda)$ has only one simple elliptic orbit. In Section 3, we will prove the former case by contradiction. Suppose that there is no simple positive hyperbolic orbit. Then the boundary operator $\partial$ used to define ECH always vanishes because of the property of ECH index (\ref{mod2}). Under the assumption, from $\partial=0$ we will introduce some notations and use them to cause a contradiction with the volume property (Theorem \ref{volume}).
In Section 4 and beyond, we will prove the main theorem under the case that $(Y,\lambda)$ has only one simple elliptic orbit. This is the most difficult part in this paper. The strategy of this part is inspired by the method in \cite{HCP} for proving the existence of infinity many Reeb orbits if $c_{1}(\mathrm{Ker}(\lambda)=\xi)$ is torsion (Theorem \ref{a}). Let us recall that way. Suppose that $c_{1}(\xi)$ is torsion and there are only finitely Reeb orbits. Then the increasing rate of the difference of ECH index $I$ and $J_{0}$ index are at most linear. By combining with volume property (Theorem \ref{volume}), we have either $J_{0}\leq 1$ holomorphic curve counted by the $U$-map or sufficiently large consecutive $J_{0}=2$ holomorphic curves counted by the $U$-map. Moreover we can pick up such curves so that their energies are sufficiently small. If there is a $J_{0}\leq 1$ holomorphic curve, such a curve becomes a global section of the Reeb vector field and this implies there are infinity many Reeb orbits. If there are such sufficiently large consecutive $J_{0}=2$ holomorphic curves, then at least one curve in them becomes a global section of the Reeb vector field and in the same as the other case, this implies there are infinity many Reeb orbits. As a result, we can conclude that $(Y,\lambda)$ has infinity many Reeb orbits. With this in mind, we now explain the strategy of the proof of the rest of the main theorem.
Suppose that there is no positive hyperbolic orbit. Note that $\partial=0$. At first, in the first half of Section 4, we will show Proposition \ref{asympdense} which asserts that the density of elements which does not have good properties in a sense tend to 0 at the limit. In this proof, we will also introduce some notations and technically use them combining with the volume property. In the latter half of this section, we will show Proposition \ref{main index 2} under Lemma \ref{mainlemma} by combining Proposition \ref{asympdense}. Proposition \ref{main index 2} asserts that there are sufficiently large consecutive $J_{0}=2$ holomorphic curves counted by the $U$-map between two elements whose behaviors are good in a sense especially whose energies are sufficiently small. Here the term “good” means that the $J$-holomorphic curves counted by the $U$-map are well-controlled for futher arguments. As the former part in this section, the volume property will also play an important role.
Next, in Section 5 we will prove Lemma \ref{mainlemma}. Lemma \ref{mainlemma} asserts that there is no $J_{0}\leq 1$ holomorphic curve counted by the $U$-map between two elements whose behaviors are good. We will prove this by contradiction. Suppose that such a curve exists. To derive a contradiction, we will consider the compactification of the moduli spaces of the holomorphic curves counted by the $U$-map. Here we note that the $J$-holomorphic curves counted by the $U$-map through a fixed generic point $z\in Y$ and their topological types are controlled by $J_{0}$ index. By moving this $z$ and considering the properties of ECH index, we have some certain splitting curves and also approximate relations of actions in some orbits. This contradicts some properties of ECH index, especially the partition conditions of ends (Definition \ref{part}) and the properties of $S_{\pm\theta}$.
In Section 6 and Section 7, we will state and prove Proposition \ref{nagai}. Proposition \ref{nagai} asserts that such $J_{0}=2$ holomorphic curves obtained in Proposition \ref{main index 2} can be classified into six types. In particular, each type has some approximate relations of actions in some orbits. In order to determine approximate relations, we will also observe the splitting behaviors of the $J$-holomorphic curves with small energy. More specifically, in this step, we will list all possibilities of their splitting ways and solve dozens of simultaneous approximate equations to determine the approximate relations.
In Section 8, we will derive a contradiction from Proposition \ref{nagai} and Proposition \ref{main index 2}. More precisely, the approximate relations of actions in some orbits obtained in Proposition \ref{nagai} restrict the consecutiveness of $J_{0}=2$ holomorphic curves counted by the $U$-map. This contradicts the result obtained in Proposition \ref{main index 2} that there are sufficiently large consecutive $J_{0}=2$ holomorphic curves counted by the $U$-map. As a result, we will complete the proof of the main theorem.
\subsection*{Acknowledgement}
The author would like to thank his advisor Professor Kaoru Ono
for his discussion and checking the preliminary version of this paper in detail, and Suguru Ishikawa for a series of discussion. He also thanks Masayuki Asaoka and Kei Irie for some comments.
This work was supported by JSPS KAKENHI Grant Number JP21J20300.
\section{Preliminaries}
\subsection{The definitions and properties of Embedded contact homology}
Let $(Y,\lambda)$ be a non-degenerate contact three manifold. For $\Gamma \in H_{1}(Y;\mathbb{Z})$, Embedded contact homology $\mathrm{ECH}(Y,\lambda,\Gamma)$ is defined. At first, we define the chain complex $(\mathrm{ECC}(Y,\lambda,\Gamma),\partial)$. In this paper, we consider ECH over $\mathbb{Z}/2\mathbb{Z}=\mathbb{F}$.
\begin{dfn} [{\cite[Definition 1.1]{H1}}]\label{qdef}
An orbit set $\alpha=\{(\alpha_{i},m_{i})\}$ is a finite pair of distinct simple periodic orbit $\alpha_{i}$ with positive integer $m_{i}$.
If $m_{i}=1$ whenever $\alpha_{i}$ is hyperboric orbit, then $\alpha=\{(\alpha_{i},m_{i})\}$ is called an admissible orbit set.
\end{dfn}
Set $[\alpha]=\sum m_{i}[\alpha_{i}] \in H_{1}(Y)$. For two orbit sets $\alpha=\{(\alpha_{i},m_{i})\}$ and $\beta=\{(\beta_{j},n_{j})\}$ with $[\alpha]=[\beta]$, we define $H_{2}(Y,\alpha,\beta)$ to be the set of relative homology classes of
2-chains $Z$ in $Y$ with $\partial Z =\sum_{i}m_{i} \alpha_{i}-\sum_{j}m_{j}\beta_{j}$ . This is an affine space over $H_{2}(Y)$. From now on. we fix a trivialization of contact plane $\xi$ over each simple orbit and write it by $\tau$.
\begin{dfn}[{\cite[{\S}8.2]{H1}}]
Let $\alpha_{1}$, $\beta_{1}$, $\alpha_{2}$ and $\beta_{2}$ be orbit sets with $[\alpha_{1}]=[\beta_{1}]$ and $[\alpha_{2}]=[\beta_{2}]$. For a trivialization $\tau$, we can define
\begin{equation}
Q_{\tau}:H_{2}(Y;\alpha_{1},\beta_{1}) \times H_{2}(Y;\alpha_{2},\beta_{2}) \to \mathbb{Z}
\end{equation}
This is well-defined. Moreover if $Z_{1}\in H_{2}(Y;\alpha_{1},\beta_{1})$ and $Z_{2} \in H_{2}(Y;\alpha_{2},\beta_{2})$, then
\begin{equation}
Q_{\tau}(Z_{1}+Z_{2},Z_{1}+Z_{2})=Q_{\tau}(Z_{1},Z_{1})+2Q_{\tau}(Z_{1},Z_{2})+Q_{\tau}(Z_{2},Z_{2}).
\end{equation}
See {\cite[{\S}8.2]{H1}} for more detail definitions.
\end{dfn}
\begin{dfn}[{\cite[Definition 1.5]{H1}}]
For $Z\in H_{2}(Y,\alpha,\beta)$, we define
\begin{equation}
I(\alpha,\beta,Z):=c_{1}(\xi|_{Z},\tau)+Q_{\tau}(Z)+\sum_{i}\sum_{k=1}^{m_{i}}\mu_{\tau}(\alpha_{i}^{k})-\sum_{j}\sum_{k=1}^{n_{j}}\mu_{\tau}(\beta_{j}^{k}).
\end{equation}
We call $I(\alpha,\beta,Z)$ an ECH index. Here, $\mu_{\tau}$ is the Conely Zhender index with respect to $\tau$ and $c_{1}(\xi|_{Z},\tau)$ is a reative Chern number and $Q_{\tau}(Z)=Q_{\tau}(Z,Z)$. Moreover this is independent of $\tau$ (see \cite{H1} for more details).
\end{dfn}
\begin{prp}[{\cite[Proposition 1.6]{H1}}]
The ECH index $I$ has the following properties.
\item[1.] For orbit sets $\alpha, \beta, \gamma$ with $[\alpha]=[\beta]=[\gamma]=\Gamma\in H_{1}(Y)$ and $Z\in H_{2}(Y,\alpha,\beta)$, $Z'\in H_{2}(Y,\beta,\gamma)$,
\begin{equation}\label{adtiv}
I(\alpha,\beta,Z)+I(\beta,\gamma,Z')=I(\alpha,\gamma,Z+Z').
\end{equation}
\item[2.] For $Z, Z'\in H_{2}(Y,\alpha,\beta)$,
\begin{equation}\label{homimi}
I(\alpha,\beta,Z)-I(\alpha,\beta,Z')=<c_{1}(\xi)+2\mathrm{PD}(\Gamma),Z-Z'>.
\end{equation}
\item[3.] If $\alpha$ and $\beta$ are admissible orbit sets,
\begin{equation}\label{mod2}
I(\alpha,\beta,Z)=\epsilon(\alpha)-\epsilon(\beta) \,\,\,\mathrm{mod}\,\,2.
\end{equation}
Here, $\epsilon(\alpha)$, $\epsilon(\beta)$ are the numbers of positive hyperbolic orbits in $\alpha$, $\beta$ respectively.
\end{prp}
For $\Gamma \in H_{1}(Y)$, we define $\mathrm{ECC}(Y,\lambda,\Gamma)$ as freely generated module over $\mathbb{Z}/2$ by admissible orbit sets $\alpha$ such that $[\alpha]=\Gamma$. That is
\begin{equation}
\mathrm{ECC}(Y,\lambda,\Gamma):= \bigoplus_{\alpha:\mathrm{admissibe\,\,orbit\,\,set\,\,with\,\,}{[\alpha]=\Gamma}}\mathbb{Z}_{2}\langle \alpha \rangle.
\end{equation}
To define the differential $\partial:\mathrm{ECC}(Y,\lambda,\Gamma)\to \mathrm{ECC}(Y,\lambda,\Gamma) $, we pick a generic almost complex structure $J$ on $\mathbb{R}\times Y$ which satisfies
\begin{enumerate}
\item $\mathbb{R}$-invariant
\item $J(\frac{d}{ds})=X_{\lambda}$
\item $J\xi=\xi$
\end{enumerate}
We consider $J$-holomorphic curves $u:(\Sigma,j)\to (\mathbb{R}\times Y,J)$ where
the domain $(\Sigma, j)$ is a punctured compact Riemann surface. Here the domain $\Sigma$ is
not necessarily connected. Let $\gamma$ be a (not necessarily simple) Reeb orbit. If a puncture
of $u$ is asymptotic to $\mathbb{R}\times \gamma$ as $s\to \infty$, we call it a positive end of $u$ at $\gamma$ and if a puncture of $u$ is asymptotic to $\mathbb{R}\times \gamma$ as $s\to -\infty$, we call it a negative end of $u$ at $\gamma$ ( see \cite{H1} for more details ).
Let $u:(\Sigma,j)\to (\mathbb{R}\times Y,J)$ and $u':(\Sigma',j')\to (\mathbb{R}\times Y,J)$ be two $J$-holomorphic curves. If there is a biholomorphic map $\phi:(\Sigma,j)\to (\Sigma',j')$ with $u'\circ \phi= u$, we regard $u$ and $u'$ as equivalent.
Let $\alpha=\{(\alpha_{i},m_{i})\}$ and $\beta=\{(\beta_{i},n_{i})\}$ be orbit sets. Let $\mathcal{M}^{J}(\alpha,\beta)$ denote the set of $J$-holomorphic curves with positive ends
at covers of $\alpha_{i}$ with total covering multiplicity $m_{i}$, negative ends at covers of $\beta_{j}$
with total covering multiplicity $n_{j}$, and no other punctures. Moreover, in $\mathcal{M}^{J}(\alpha,\beta)$, we consider two
$J$-holomorphic curves to be equivalent if they represent the same current in $\mathbb{R}\times Y$.
For $u \in \mathcal{M}^{J}(\alpha,\beta)$, we naturally have $[u]\in H_{2}(Y;\alpha,\beta)$ and set $I(u)=I(\alpha,\beta,[u])$. Moreover we define
\begin{equation}
\mathcal{M}_{k}^{J}(\alpha,\beta):=\{\,u\in \mathcal{M}^{J}(\alpha,\beta)\,|\,I(u)=k\,\,\}
\end{equation}
Under this notations, we define $\partial_{J}:\mathrm{ECC}(Y,\lambda,\Gamma)\to \mathrm{ECC}(Y,\lambda,\Gamma)$ as follows.
For admissible orbit set $\alpha$ with $[\alpha]=\Gamma$, we define
\begin{equation}
\partial_{J} \langle \alpha \rangle=\sum_{\beta:\mathrm{admissible\,\,orbit\,\,set\,\,with\,\,}[\beta]=\Gamma} \# (\mathcal{M}_{1}^{J}(\alpha,\beta)/\mathbb{R})\cdot \langle \beta \rangle.
\end{equation}
Note that the above counting is well-defined and $\partial_{J} \circ \partial_{J}$. We can see the reason of the former in Proposition \ref{ind} and the later was proved in \cite{HT1} and \cite{HT2}. Moreover, the homology defined by $\partial_{J}$ does not depend on $J$ (see Theorem \ref{test}, or see \cite{T1}).
Before we get to the next subsection, recall (Fredholm) index.
For $u\in \mathcal{M}^{J}(\alpha,\beta)$, the its (Fredholm) index is defined by
\begin{equation}
\mathrm{ind}(u):=-\chi(u)+2c_{1}(\xi|_{[u]},\tau)+\sum_{k}\mu_{\tau}(\gamma_{k}^{+})-\sum_{l}\mu_{\tau}(\gamma_{l}^{-}).
\end{equation}
Here $\{\gamma_{k}^{+}\}$ is the set consisting of (not necessarilly simple) all positive ends of $u$ and $\{\gamma_{l}^{-}\}$ is that one of all negative ends. Note that for generic $J$, if $u$ is connected and somewhere injective, then the moduli space of $J$-holomorphic
curves near $u$ is a manifold of dimension $\mathrm{ind}(u)$ (see \cite[Definition 1.3]{HT1}).
\subsection{$J$-holomorphic curves with small ECH index and partition conditions of multiplicities}
For $\theta\in \mathbb{R}\backslash \mathbb{Q}$, we define $S_{\theta}$ to be the set of positive integers $q$ such that $\frac{\lceil q\theta \rceil}{q}< \frac{\lceil q'\theta \rceil}{q'}$ for all $q'\in \{1,\,\,2,...,\,\,q-1\}$ and write $S_{\theta}=\{q_{0}=1,\,\,q_{1},\,\,q_{2},\,\,q_{3},...\}$ in increasing order. Also $S_{-\theta}=\{p_{0}=1,\,\,p_{1},\,\,p_{2},\,\,p_{3},...\}$.
\begin{prp}[{\cite[Proof of Lemma 3.3]{HT3}}, and {\cite[Proof of Remark 4.4]{H1}}]\label{s}
For $\theta \in \mathbb{R}\backslash \mathbb{Q}$,
\item[1.] $q_{i+1}-q_{i}$ (resp. $p_{i+1}-p_{i}$) are nondecreasing with respect to $i$ and some elements of $S_{-\theta}$ (resp. $S_{\theta}$).
\item[2.] $S_{\theta}\cap{S_{-\theta}}=\{1\}$,
\item[3.] $q_{i+1}-q_{i}\to \infty$ (resp. $p_{i+1}-p_{i}\to \infty$) as $i\to \infty$.
\end{prp}
\begin{dfn}[{\cite[Definition 7.1]{HT1}}, or {\cite[{\S}4]{H1}}]
For non negative integer $M$, we inductively define the incoming partition $P_{\theta}^{\mathrm{in}}(M)$ as follows.
For $M=0$, $P_{\theta}^{\mathrm{in}}(0)=\emptyset$ and for $M>0$,
\begin{equation}
P_{\theta}^{\mathrm{in}}(M):=P_{\theta}^{\mathrm{in}}(M-a)\cup{(a)}
\end{equation}
where $a:=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$.
Define outgoing partition
\begin{equation}
P_{\theta}^{\mathrm{out}}(M):= P_{-\theta}^{\mathrm{in}}(M).
\end{equation}
\end{dfn}
\begin{dfn}[{\cite[Definition 7.11]{HT1}}, or cf.{\cite[{\S}4]{H1}}]\label{part}
For a simple Reeb orbit $\gamma$ and positive integer $M$, define two partitions of $M$, the incoming partition $P_{\gamma}^{\mathrm{in}}(M)$ and the outgoing partition $P_{\gamma}^{\mathrm{out}}(M)$ as follows.
\item[1.] If $\gamma$ is positive hyperbolic, then
\begin{equation}
P_{\gamma}^{\mathrm{in}}(M)=P_{\gamma}^{\mathrm{out}}:=(1,...,\,1)
\end{equation}
\item[2.] If $\gamma$ is negative hyperbolic, then
\begin{equation}
P_{\gamma}^{\mathrm{in}}(M)=P_{\gamma}^{\mathrm{out}}(M):=
\begin{cases}
(2,...,\,2) & \mathrm{if}\,\,M\,\,\mathrm{is\,\,even}, \\
(2,...,\,2,\,1) & \mathrm{if}\,\,M\,\,\mathrm{is\,\,odd},
\end{cases}
\end{equation}
\item[3.] If $\gamma$ is elliptic, then
\begin{equation}
P_{\gamma}^{\mathrm{in}}(M):=P_{\theta}^{\mathrm{in}}(M),\,\,\,\,P_{\gamma}^{\mathrm{out}}(M):=P_{\theta}^{\mathrm{out}}(M).
\end{equation}
where $\theta$ is the rotation number of $\gamma$ up to $\mathbb{Z}$.
The standard ordering convention for $P_{\gamma}^{\mathrm{in}}(M)$ or $P_{\gamma}^{\mathrm{out}}(M)$ is to list the entries
in``nonincreasing'' order.
\end{dfn}
Let $\alpha=\{(\alpha_{i},m_{i})\}$ and $\beta=\{(\beta_{i},n_{i})\}$. For $u\in \mathcal{M}^{J}(\alpha,\beta)$, it can be uniquely
written as $u=u_{0}\cup{u_{1}}$ where $u_{0}$ are unions of all components which maps to $\mathbb{R}$-invariant cylinders in $u$ and $u_{1}$ is the rest of $u$.
For $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha,\beta)$, let $P_{\alpha_{i}}^{+}$ denote the set consisting of the multiplicities of the positive ends of $u_{1}$ at covers of $\alpha_{i}$. Define $P_{\beta_{j}}^{-}$ analogously for the negative end.
\begin{dfn}[{\cite[Definition 7.13]{HT1}}]\label{adm}
$u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha,\beta)$ is admissible if
\item[1.] $u_{1}$ is embedded and does not intersect $u_{0}$
\item[2.] For each simple Reeb orbit $\alpha_{i}$ in $\alpha$ (resp. $\beta_{j}$ in $\beta$), under the standerd ordering convention, $P_{\alpha_{i}}^{+}$(resp. $P_{\beta_{j}}^{-}$) is an initial segment of $P_{\alpha_{i}}^{\mathrm{out}}(m_{i})$(resp. $P_{\beta_{j}}^{\mathrm{in}}(n_{j})$).
\end{dfn}
\begin{prp}[{\cite[Proposition 7.15]{HT1}}]\label{ind}
Suppose that $J$ is generic and $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha,\beta)$. Then
\item[1.] $I(u)\geq 0$
\item[2.] If $I(u)=0$, then $u_{1}=\emptyset$
\item[3.] If $I(u)=1$, then $u$ is admissible and $\mathrm{ind}(u_{1})=1$.
\item[4.] If $I(u)=2$ and $\alpha$ and $\beta$ are admissible, then u is admissible and $\mathrm{ind}(u_{1})=2$.
\end{prp}
\subsection{The gradings and the $U$-map}
By (\ref{homo}) and (\ref{adi}), we can see that if $c_{1}(\xi)+2\mathrm{PD(\Gamma)}$ is divisible by $d$, admissible orbits sets have relative $\mathbb{Z}/d$-grading. So we can decompose $\mathrm{ECC}(Y,\lambda,\Gamma)$ as direct sum by $\mathbb{Z}/d$-grading.
\begin{equation}\label{direc}
\mathrm{ECC}(Y,\lambda,\Gamma)= \bigoplus_{*:\,\,\mathbb{Z}/d\,\,\mathrm{grading}}\mathrm{ECC}_{*}(Y,\lambda,\Gamma).
\end{equation}
Note that if $c_{1}(\xi)+2\mathrm{PD}(\Gamma)$ is torsion, there exists the relative $\mathbb{Z}$-grading.
The same as (\ref{direc}), we can see that
\begin{equation}
\mathrm{ECH}(Y,\lambda,\Gamma):= \bigoplus_{*:\,\,\mathbb{Z}\,\,\mathrm{grading}}\mathrm{ECH}_{*}(Y,\lambda,\Gamma).
\end{equation}
Let $Y$ be connected.
Then there is degree$-2$ map $U$.
\begin{equation}\label{Umap}
U:\mathrm{ECH}_{*}(Y,\lambda,\Gamma) \to \mathrm{ECH}_{*-2}(Y,\lambda,\Gamma).
\end{equation}
To define this, choose a base point $z\in Y$ which is not on the image of any Reeb orbit and let $J$ be generic.
Then define a map
\begin{equation}
U_{J,z}:\mathrm{ECC}_{*}(Y,\lambda,\Gamma) \to \mathrm{ECC}_{*-2}(Y,\lambda,\Gamma).
\end{equation}
by
\begin{equation}
U_{J,z} \langle \alpha \rangle=\sum_{\beta:\mathrm{admissible\,\,orbit\,\,set\,\,with\,\,}[\beta]=\Gamma} \# \{\,u\in \mathcal{M}_{2}^{J}(\alpha,\beta)/\mathbb{R})\,|\,(0,z)\in u\,\}\cdot \langle \beta \rangle.
\end{equation}
The above map $U_{J,z}$ is a chain map, and we define the $U$ map
as the induced map on homology. Under the assumption, this map is independent on $z$ (for a generic $J$). See \cite[{\S}2.5]{HT3} for more details. Moreover, in the same reason as $\partial$, $U_{J,z}$ does not depend on $J$ (see Theorem \ref{test}, and see \cite{T1}).
In this paper, we choose a suitable generic $J$ as necessary (Specifically, we choose a generic $J$ so that $U_{J,z}$ is well-defined for some countable sequences $z$ appearing in the future discussions).
The next isomorphism is important.
\begin{them}[\cite{T1}]\label{test}
For each $\Gamma\in H_{1}(Y)$, there is an isomorphism
\begin{equation}
\mathrm{ECH}_{*}(Y,\lambda,\Gamma) \cong \reallywidecheck{HM}_{*}(-Y,\mathfrak{s}(\xi)+2\mathrm{PD}(\Gamma))
\end{equation}
of relatively $\mathbb{Z}/d\mathbb{Z}$-graded abelian groups. Here $d$ is the divisibility of $\mathfrak{s}(\xi)+2\mathrm{PD}(\Gamma)$ in $H_{1}(Y)$ mod torsion and $\mathfrak{s}(\xi)$ is the
spin-c structure associated to the oriented 2–plane field as in \cite{KM}.
Moerover, the above isomorphism interchanges
the map $U$ in (\ref{Umap}) with the map
\begin{equation}
U_{\dag}: \reallywidecheck{HM}_{*}(-Y,\mathfrak{s}(\xi)+2\mathrm{PD}(\Gamma)) \longrightarrow \reallywidecheck{HM}_{*-2}(-Y,\mathfrak{s}(\xi)+2\mathrm{PD}(\Gamma))
\end{equation}
defined in \cite{KM}.
\end{them}
Here $\reallywidecheck{HM}_{*}(-Y,\mathfrak{s}(\xi)+2\mathrm{PD}(\Gamma))$ is a version of Seiberg-Witten Floer homology with
$\mathbb{Z}/2\mathbb{Z}$ coefficients defined by Kronheimer-Mrowka \cite{KM}.
\subsection{$J_{0}$ index and topological complexity of $J$-holomorphic curve}
In this subsection, we recall the $J_{0}$ index.
\begin{dfn}[{\cite[{\S}3.3]{HT3}}]
Let $\alpha=\{(\alpha_{i},m_{i})\}$ and $\beta=\{(\beta_{j},n_{j})\}$ be orbit sets with $[\alpha]=[\beta]$.
For $Z\in H_{2}(Y,\alpha,\beta)$, we define
\begin{equation}
J_{0}(\alpha,\beta,Z):=-c_{1}(\xi|_{Z},\tau)+Q_{\tau}(Z)+\sum_{i}\sum_{k=1}^{m_{i}-1}\mu_{\tau}(\alpha_{i}^{k})-\sum_{j}\sum_{k=1}^{n_{j}-1}\mu_{\tau}(\beta_{j}^{k}).
\end{equation}
\end{dfn}
\begin{prp}[{\cite[{\S}3.3]{HT3}} {\cite[{\S}2.6]{CHR}}]
The index $J_{0}$ has the following properties.
\item[1.] For orbit sets $\alpha, \beta, \gamma$ with $[\alpha]=[\beta]=[\gamma]=\Gamma\in H_{1}(Y)$ and $Z\in H_{2}(Y,\alpha,\beta)$, $Z'\in H_{2}(Y,\beta,\gamma)$,
\begin{equation}\label{adi}
J_{0}(\alpha,\beta,Z)+J_{0}(\beta,\gamma,Z')=J_{0}(\alpha,\gamma,Z+Z').
\end{equation}
\item[2.] For $Z, Z'\in H_{2}(Y,\alpha,\beta)$,
\begin{equation}\label{homo}
J_{0}(\alpha,\beta,Z)-J_{0}(\alpha,\beta,Z')=<-c_{1}(\xi)+2\mathrm{PD}(\Gamma),Z-Z'>.
\end{equation}
\end{prp}
\begin{dfn}[{\cite[just befor Proposition 5.8]{H3}}]
Let $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha,\beta)$. Suppose that $u_{1}$ is somewhere injective. Let $n_{i}^{+}$
be the number of positive ends of $u_{1}$ which are asymptotic to $\alpha_{i}$, plus 1 if $u_{0}$ includes the trivial cylinder $\mathbb{R}\times \alpha_{i}$ with some multiplicity. Likewise, let $n_{j}^{-}$ be the number of negative ends of $u_{1}$ which are asymptotic to $\beta_{j}$, plus 1 if $u_{0}$ includes the trivial cylinder $\mathbb{R}\times \beta_{j}$ with some multiplicity.
\end{dfn}
Write $J_{0}(u)=J_{0}(\alpha,\beta,[u])$.
\begin{prp}[{\cite[Lemma 3.5]{HT3}} {\cite[Proposition 5.8]{H3}}]
Let $\alpha=\{(\alpha_{i},m_{i})\}$ and $\beta=\{(\beta_{j},n_{j})\}$ be admissible orbit sets, and let $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha,\beta)$. Then
\begin{equation}
-\chi(u_{1})+\sum_{i}(n_{i}^{+}-1)+\sum_{j}(n_{j}^{-}-1)\leq J_{0}(u)
\end{equation}
If $u$ is counted by the ECH differential or the $U$-map, then the above equality holds. Note that $J_{0}(u)\geq -1$ in any case.
\end{prp}
\subsection{ECH spectrum and the volume property}
The action of an orbit set $\alpha=\{(\alpha_{i},m_{i})\}$ is defined by
\begin{equation}
A(\alpha)=\sum m_{i}A(\alpha_{i})=\sum m_{i}\int_{\alpha_{i}}\lambda.
\end{equation}
For any $L>0$, $\mathrm{ECC}^{L}(Y,\lambda,\Gamma)$ denotes the subspace of $\mathrm{ECC}(Y,\lambda,\Gamma)$ which is generated by admissible orbit sets whose actions are less than $L$. In the same way, $(\mathrm{ECC}^{L}(Y,\lambda,\Gamma),\partial)$ becomes a chain complex and the homology group $\mathrm{ECH}^{L}(Y,\lambda,\Gamma)$ is obtained. Here, we use the fact that if two admissible orbit sets $\alpha=\{(\alpha_{i},m_{i})\}$ and $\beta=\{(\beta_{i},n_{i})\}$ have $A(\alpha)\leq A(\beta)$, then the coefficient of $\beta$ in $\partial \alpha$ is $0$ because of the positivity of $J$ holomorphic curves over $d\lambda$ and the fact that $A(\alpha)-A(\beta)$ is equivalent to the integral value of $d\lambda$ over $J$-holomorphic punctured curves which is asymptotic to $\alpha$ at $+\infty$, $\beta$ at $-\infty$.
By construction, there exists a natural homomorphism $i_{L}:\mathrm{ECH}^{L}(Y,\lambda,\Gamma) \to \mathrm{ECH}(Y,\lambda,\Gamma)$.
\begin{dfn}[{\cite[Definition 4.1]{H2}}]
Let $Y$ be a closed oriented three manifold with a nondegenerate contact form $\lambda$ and $\Gamma \in H_{1}(Y,\mathbb{Z})$. If $0\neq \sigma \in \mathrm{ECH}(Y,\lambda,\Gamma)$, define
\begin{equation}\label{spect}
c_{\sigma}(Y,\lambda)=\inf\{L>0 |\, \sigma \in \mathrm{Im}(i_{L}:\mathrm{ECH}^{L}(Y,\lambda,\Gamma) \to \mathrm{ECH}(Y,\lambda,\Gamma))\, \}
\end{equation}
\end{dfn}
Note that we can define $c_{\sigma}$ if $\lambda$ is degenerate but in this paper, this is not necessary.(for more details, see \cite{H2})
In the case that $c_{1}(\xi)+2\mathrm{PD}(\Gamma)$ is torsion, the above spectrum recover the volume of $\mathrm{Vol}(Y,\lambda)=\int_{Y}\lambda \wedge d\lambda$.
\begin{them}[{\cite[Theorem 1.3]{CHR}}]\label{volume}
Let $Y$ be a closed connected three-manifold with
a contact form $\lambda$, let $\Gamma \in H_{1}(Y)$ with $c_{1}(\xi)+2\mathrm{PD}(\gamma)$ torsion, and let $I$ be any
refinement of the relative $\mathbb{Z}$-grading on $\mathrm{ECH}(Y,\lambda,\Gamma)$ to an absolute $\mathbb{Z}$-grading.
Then for any sequence of nonzero homogeneous classes $\{\sigma_{k} \}_{1\geq k}$ in $\mathrm{ECH}(Y,\lambda,\Gamma)$
with $\lim_{k\to \infty}I(\sigma)=+\infty$, we have
\begin{equation}\label{vol}
\lim_{k \to \infty} \frac{c_{\sigma_{k}}(Y,\lambda)^{2}}{I(\sigma_{k})}=\mathrm{Vol}(Y,\lambda).
\end{equation}
\end{them}
\section{The case that the number of simple elliptic orbits is at least two.}
Suppose that $b_{1}(Y)=0$. In this situation, for any orbit sets $\alpha$ and $\beta$ with $[\alpha]=[\beta]$, $H_{2}(Y,\alpha,\beta)$ consists of only one component since $H_{2}(Y)=0$. So we may omit the homology component from the notation of ECH index $I$ and $J_{0}$, that is, they are just written by $I(\alpha,\beta)$ and $J_{0}(\alpha,\beta)$ respectively.
Moreover, for any orbit sets $\alpha$ with $[\alpha]=0\in H_{1}(Y)$, we set $I(\alpha,[\emptyset]):=I(\alpha)$ and also $J_{0}(\alpha):=J_{0}(\alpha,[\emptyset])$. This $I(\alpha)$ defines an absolute $\mathbb{Z}$ grading in $\mathrm{ECH}(Y,\lambda,0)$. From now on, we suppose that $\mathrm{ECH}(Y,\lambda,0)$ is graded in this way.
The aim of this section is to prove the next proposition.
\begin{them}\label{el2}
Let $(Y,\lambda)$ be a connected non-degenerate closed contact three manifold with $b_{1}(Y)=0$. Assume that the number of simple elliptic orbits is at least two and the number of all simple orbit is infinity. then there exists at least one positive hyperbolic orbit.
\end{them}
We prove this by contradiction. At first, we show the following lemma.
\begin{lem}\label{isomero}
Let $(Y,\lambda)$ be a connected non-degenerate closed contact three manifold with $b_{1}(Y)=0$. Then for $\Gamma\in H_{1}(Y)$, there is some finite generated vector space $E_{\Gamma}$ such that
\begin{equation}\label{iso}
\mathrm{ECH}(Y,\lambda,\Gamma) \cong \mathbb{F}[U_{\dag}^{-1},U_{\dag}]/U_{\dag}\mathbb{F}[U_{\dag}]\bigoplus E_{\Gamma}.
\end{equation}
Morover, the above isomorphism interchanges
the map $U$ in (\ref{Umap}) with the action of the product by $U_{\dag}$ on $\mathbb{F}[U_{\dag}^{-1},U_{\dag}]/U_{\dag}\mathbb{F}[U_{\dag}]$.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{isomero}]
There are three type homologies $\reallywidecheck{HM}_{*}(-Y,\mathfrak{s})$, $\widehat{HM}_{*}(-Y,\mathfrak{s})$ and $\overline{HM}_{*}(-Y,\mathfrak{s})$ in Seiberg-Witten Floer homologies and there exist an exact sequence.
\begin{equation}
... \longrightarrow \reallywidecheck{HM}_{*}(-Y,\mathfrak{s})\longrightarrow \widehat{HM}_{*}(-Y,\mathfrak{s}) \longrightarrow \overline{HM}_{*}(-Y,\mathfrak{s}) \longrightarrow \reallywidecheck{HM}_{*-1}(-Y,\mathfrak{s}) \longrightarrow ...
\end{equation}
As \cite[Proposition 35.3.1]{KM},
\begin{equation}
\bigoplus_{*}\overline{HM}_{*}(-Y,\mathfrak{s}) \cong \mathbb{F}[U^{-1}_{\dag},U_{\dag}]
\end{equation}
By construction \cite[Definition 14.5.2, Subsection 22.1 and Subsection 22.3]{KM}, $\reallywidecheck{HM}_{*}(-Y,\mathfrak{s})$ vanishes if its grading is sufficiently low. Moreover the image of
\begin{equation}
\bigoplus_{*}\reallywidecheck{HM}_{*}(-Y,\mathfrak{s})\longrightarrow \bigoplus_{*}\widehat{HM}_{*}(-Y,\mathfrak{s})
\end{equation}
is finite rank \cite[Proposition 22.2.3]{KM}. This finishes the proof of the lemma.
\end{proof}
\textbf{From now on, in this section we suppose that there is no positive hyperbolic orbit.}
For $M>0$ and $\Gamma\in H_{1}(Y)$, we set
\begin{equation}
\Lambda(M,\Gamma):=\{\,\,\alpha\,\,|\,\,\alpha\,\,\mathrm{is\,\,admissible\,\,orbit\,\,set\,\,such\,\,that\,\,}[\alpha]=\Gamma \mathrm{\,\,and}\,\,A(\alpha)<M\,\,\}
\end{equation}
\begin{lem}\label{asympp}
For every $\Gamma\in H_{1}(Y)$,
\begin{equation}\label{volp}
\lim_{M \to \infty} \frac{M^{2}}{|\Lambda(M,\Gamma)|}=2\mathrm{Vol}(Y,\lambda).
\end{equation}
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{asympp}]
Fix $\Gamma\in H_{1}(Y)$. By $\partial=0$ and (\ref{iso}), there are admissible orbit sets $\{\alpha_{i}^{\Gamma}\}_{0 \leq i }$ and $\{\beta_{j}^{\Gamma}\}_{0\leq j \leq m_{\Gamma}}$ with $[\alpha_{i}^{\Gamma}]=[\beta_{j}^{\Gamma}]=\Gamma$ satisfy
\begin{equation}\label{importantiso}
\mathrm{ECH}(Y,\lambda,\Gamma) \cong \bigoplus_{i=0}^{\infty}\mathbb{F}\langle \alpha_{i}^{\Gamma} \rangle \bigoplus \bigoplus_{j=1}^{m_{\Gamma}}\mathbb{F}\langle \beta _{j}^{\Gamma}\rangle
\end{equation}
and $U\langle \alpha_{i}^{\Gamma} \rangle=\langle \alpha_{i-1}^{\Gamma} \rangle$ for $i\geq 1$. Moreover since $\partial=0$, each admissible orbit set $\alpha$ with $[\alpha]=\Gamma$ is equal to either $\beta_{j}^{\Gamma}$ for some $1\leq j \leq m_{\Gamma}$ or $\alpha_{k}^{\Gamma}$ for some $0\leq k$.
Note that
\begin{equation}\label{la}
A(\alpha_{l}^{\Gamma} )-A(\alpha_{k}^{\Gamma} )>0\,\,\,\,\,\mathrm{if}\,\,\,\,l>k,
\end{equation}
\begin{equation}\label{2}
I(\alpha_{l}^{\Gamma})-I( \alpha_{k}^{\Gamma})=2(l-k).
\end{equation}
Since $A(\alpha_{k}^{\Gamma} )\to \infty$ as $k\to \infty$, there is $k_{0}>0$ such that for every $k>k_{0}$ $A(\alpha_{k}^{\Gamma} )> A(\beta_{j}^{\Gamma})$ for $1\leq j \leq m_{\Gamma}$.
By (\ref{la}) and (\ref{2}), we have
$I(\alpha_{k}^{\Gamma})=2k+I(\alpha_{0}^{\Gamma})$ and for sufficiently large $k$, $|\Lambda(A(\alpha_{k}^{\Gamma}),\Gamma)|=k+m$. So for sufficiently large $k$, we have
\begin{equation}
|2|\Lambda(A(\alpha_{k}^{\Gamma}),\Gamma)|-I(\alpha_{k}^{\Gamma})|\leq 2m+I(\alpha_{0}^{\Gamma})
\end{equation}
By the above we have,
\begin{equation}
\lim_{k\to \infty}\frac{2|\Lambda(A(\alpha_{k}^{\Gamma}),\Gamma)|}{I( \alpha_{k}^{\Gamma})}=1.
\end{equation}
Since $\partial=0$ and the definition of the spectrum (\ref{spect}), $c_{\langle \alpha \rangle}(Y,\lambda)=A(\alpha)$.
So by (\ref{vol}),
\begin{equation}
\lim_{k\to \infty}\frac{c_{\langle \alpha_{k} \rangle}(Y,\lambda)^{2}}{I(\alpha_{k}^{\Gamma})}=\lim_{k\to \infty}\frac{A(\alpha_{k}^{\Gamma})^{2}}{2|\Lambda(A(\alpha_{k}^{\Gamma}),\Gamma)|}=\mathrm{Vol}(Y,\lambda).
\end{equation}
Note that for large $M>0$, there is a large $k>0$ such that $A(\alpha_{k}^{\Gamma})\geq M \geq A(\alpha_{k-1}^{\Gamma})$.
Therefore we complete the proof of Lemma \ref{asympp}.
\end{proof}
\begin{proof}[\bf Proof of Theorem \ref{el2}]
We pick up two simple elliptic orbits $\gamma_{1}$, $\gamma_{2}$. Let $s_{1}$ and $s_{2}$ denote the orders of $[\gamma_{1}]$ and $[\gamma_{2}]$ in $H_{1}(Y)$ respectively.
Since $|H_{1}(Y)|<\infty$, we can choose infinity sequence of simple orbit $\{\delta_{i}\}_{i>0}$ satisfying
\begin{enumerate}
\item their homology classes $[\delta_{i}]$ are in a same one.
\item $A(\delta_{i})<A(\delta_{j})$ if $i<j$.
\item Any $\delta_{i}$ is not equivalent to $\gamma_{1}$ and $\gamma_{2}$
\end{enumerate}
Let $r$ be the order of $[\delta_{i}]$ in $H_{1}(Y)$. Then we define a sequence of admissible orbit sets $\{\epsilon_{n}\}_{n>0}$ by $\epsilon_{n}:=\{(\delta_{r(n-1)+i},1)\}_{1\leq i \leq r}$.
By construction, $[\epsilon_{n}]=0$ and $A(\epsilon_{n})<A(\epsilon_{n+1})$.
For $t_{1},\,t_{2}\in \mathbb{Z}_{\geq 0}$, we set
\begin{equation}
\epsilon_{(t_{1},t_{2},n)}:=(\gamma_{1},t_{1}s_{1})\cup{(\gamma_{2},t_{2}s_{2})}\cup{\epsilon_{n}}
\end{equation}
Note that $\epsilon_{(t_{1},t_{2},n)}$ is admissible with $[\epsilon_{(t_{1},t_{2},n)}]=0$ and morover if $(t_{1},t_{2},n)\neq(t_{1}',t_{2}',n')$ then $\epsilon_{(t_{1},t_{2},n)}\neq \epsilon_{(t_{1}',t_{2}',n')}$.
Let $T_{n}:=A(\epsilon_{n})$ and $R_{i}=s_{i}A(\delta_{i})$ for $i=1,2$, then
\begin{equation}
A(\epsilon_{(t_{1},t_{2},n)})=t_{1}R_{1}+t_{2}R_{2}+T_{n}
\end{equation}
So for $n$,
\begin{equation}\label{eq}
\{\,\,(t_{1},t_{2})\,\,|\,\,A(\epsilon_{(t_{1},t_{2},n)})<M\,\,\}=\{\,\,(t_{1},t_{2})\in \mathbb{Z}_{\geq 0}\times\mathbb{Z}_{\geq 0}\,\,|\,\,t_{1}R_{1}+t_{2}R_{2}<M-T_{n}\,\,\}
\end{equation}
In general, for any $S_{1},\,\,S_{2},\,\,T>0$, the number of
\begin{equation}
\{\,\,(t_{1},t_{2})\in \mathbb{Z}_{\geq 0}\times\mathbb{Z}_{\geq 0}\,\,|\,\,t_{1}S_{1}+t_{2}S_{2}<T\,\,\}
\end{equation}
is $\frac{(T)^{2}}{2S_{1}S_{2}}+O(T)$ (for example, we can see the same argument in \cite{H2}).
So the right hand side of (\ref{eq}) is equal to $\frac{(M-T_{n})^{2}}{2R_{1}R_{2}}+O(M-T_{n})$.
For any $N\in \mathbb{Z}_{>0}$, we pick a sufficiently large $M>0$ satisfying $M>T_{N}$. Then
\begin{equation}
\begin{split}
|\Lambda(M,0)|>&|\{\,\,(t_{1},t_{2},n)\,\,|\,\,A(\epsilon_{(t_{1},t_{2},n)})<M\,\,,\,\,1\leq n\leq N\,\,\}|\\
=&\sum_{k=1}^{N}|\{\,\,(t_{1},t_{2})\,\,|\,\,t_{1}R_{1}+t_{2}R_{2}<M-T_{k}\,\,\}|\\
=&\sum_{k=1}^{N}\frac{(M-T_{k})^{2}}{2R_{1}R_{2}}+O(M-T_{k})\\
=&\frac{(M)^{2}N}{2R_{1}R_{2}}+O(M)
\end{split}
\end{equation}
So
\begin{equation}
\lim_{M \to \infty} \frac{M^{2}}{|\Lambda(M,0)|}<\frac{2R_{1}R_{2}}{N}.
\end{equation}
Since we choose $N$ arbitrarily, by (\ref{volp}) we can see that $\mathrm{Vol}(Y,\lambda)=0$. This is a contradiction. We complete the proof of Theorem \ref{el2}
\end{proof}
\section{The case that the number of simple elliptic orbits is exactly one.}
We use the rest of this paper to prove the next theorem.
\begin{them}\label{mainmainmain}
Let $Y$ be a closed connected three manifold with $b_{1}(Y)=0$. Then, $Y$ does not admit a non-degenerate contact form $\lambda$ such that exactly one simple orbit is elliptic orbit and all the others are negative hyperbolic.
\end{them}
We prove this by contradiction.
\textbf{From now on, we assume that $(Y,\lambda)$ is non-degenerate contact three manifold such that exactly one simple orbit is elliptic and all the others are negative hyperbolic.}
Let $\gamma$ be the simple elliptic orbit. Moreover, let $A(\gamma)=R$ and $\theta$ be the rotation number with respect to some fixed trivialization $\tau$ over $\gamma$. This means $e^{\pm 2\pi \theta}$ are eigenvalues of $d\phi^{R}|_{\xi}$ and for every $k\in \mathbb{Z}$, $\mu_{\tau}(\gamma^{k})=2\lfloor k\theta \rfloor +1$.
\subsection{Density of orbit sets with some properties}
For admissible orbit set $\alpha$, Let $E(\alpha)$, $H(\alpha)$ be the multiplicity at $\gamma$ in $\alpha$ and the number of hyperbolic orbits in $\alpha$, respectively.
Recall that for $M\in \mathbb{R}$ and $\Gamma \in H_{1}(M)$,
\begin{equation}
\Lambda(M,\Gamma)=\{\,\alpha\,|\,\alpha\,\,\mathrm{is\,\,an\,\,admissible\,\,orbit\,\,set\,\,such\,\,that\,\,}[\alpha]=\Gamma \mathrm{\,\,and}\,\,A(\alpha)<M\,\}.
\end{equation}
In addition to this, we introduce some notations as follows.
\begin{equation}
\Lambda_{(n,m)}(M,\Gamma):=\{\,\,\alpha \in \Lambda(M,\Gamma)\,|\,(E(\alpha),H(\alpha))=(n,m)\,\}
\end{equation}
\begin{equation}
\Lambda_{(n,\infty)}(M,\Gamma):=\bigcup_{m=0}^{\infty}\Lambda_{(n,m)}(M,\Gamma)
\end{equation}
\begin{equation}
\Lambda_{(\infty,m)}(M,\Gamma):=\bigcup_{n=0}^{\infty}\Lambda_{(n,m)}(M,\Gamma)
\end{equation}
\begin{equation}
\Lambda(M):=\bigcup_{\Gamma\in H_{1}(M)}\Lambda(M,\Gamma)
\end{equation}
Note that if $M\leq 0$, the above sets become empty.
\begin{prp}\label{asympdense}
For every $\Gamma \in H_{1}(Y)$,
\item[1.] For every positive integer $n$,
\begin{equation}
\lim_{M\to \infty}\frac{|\Lambda_{(n,\infty)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}=0
\end{equation}
\item[2.] For every positive integer $m$,
\begin{equation}
\lim_{M\to \infty}\frac{|\Lambda_{(\infty,m)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}=0
\end{equation}
\item [3.]
\begin{equation}
\lim_{M\to \infty}\frac{|\bigcup_{p_{i}\in S_{\theta} } \Lambda_{(p_{i},\infty)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}=0
\end{equation}
\end{prp}
Before we try to prove Proposition \ref{asympdense}, we show the next almost trivial claim but this makes the proof of Proposition \ref{asympdense} easier.
\begin{cla}\label{easier}
For every $\Gamma\in H_{1}(Y)$, we have
\begin{equation}
\lim_{M\to \infty}\frac{|\Lambda(M)|}{|\Lambda(M,\Gamma)|}=|H_{1}(Y)|
\end{equation}
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{easier}]
By the definition, we have
\begin{equation}
|\Lambda(M)|=\sum_{\Gamma\in H_{1}(Y)}|\Lambda(M,\Gamma)|
\end{equation}
And since Lemma \ref{asympp},
\begin{equation}
\lim_{M\to \infty}\frac{|\Lambda(M)|}{M^{2}}= \lim_{M\to \infty}\sum_{\Gamma\in H_{1}(Y)}\frac{|\Lambda(M,\Gamma)|}{M^2}= \frac{|H_{1}(Y)|}{2\mathrm{Vol}(Y,\lambda)}.
\end{equation}
Hence
\begin{equation}
\lim_{M\to \infty}\frac{|\Lambda(M)|}{|\Lambda(M,\Gamma)|}= \lim_{M\to \infty}\frac{|\Lambda(M)|}{M^{2}}\frac{M^{2}}{|\Lambda(M,\Gamma)|}=|H_{1}(Y)|.
\end{equation}
\end{proof}
\begin{proof}[\bf Proof of Proposition \ref{asympdense}]
Note that $|\Lambda_{(n,\infty)}(M,\Gamma)|=|\Lambda_{(n-k,\infty)}(M-kR,\Gamma-k[\gamma])|$. This is because the correspondence by adding $(\gamma,k)$ from $\Lambda_{(n-k,\infty)}(M-kR,\Gamma-k[\gamma])$ to $\Lambda_{(n,\infty)}(M,\Gamma)$ is bijective. Hence
\begin{equation}
|\Lambda(M,\Gamma)|=\sum_{n=0}^{\infty}|\Lambda_{(n,\infty)}(M,\Gamma)|=\sum_{n=0}^{\infty}|\Lambda_{(0,\infty)}(M-nR,\Gamma-n[\gamma])|.
\end{equation}
Since $\lim_{M\to\infty}\frac{|\Lambda(M-R,\Gamma-[\gamma])|}{|\Lambda(M,\Gamma)|}=\lim_{M\to\infty}\frac{(M-R)^2}{M^2}=1$, we have
\begin{equation}
\begin{split}
\lim_{M\to\infty}\frac{|\Lambda_{(0,\infty)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}
=&1-\lim_{M\to \infty}\frac{\sum_{n=0}^{\infty}|\Lambda_{(0,\infty)}(M-(n+1)R,\Gamma-(n+1)[\gamma])|}{\sum_{n=0}^{\infty}|\Lambda_{(0,\infty)}(M-nR,\Gamma-n[\gamma])|}\\
=&1-\lim_{M\to\infty}\frac{|\Lambda(M-R,\Gamma-[\gamma])|}{|\Lambda(M,\Gamma)|}\\
=&0.
\end{split}
\end{equation}
And hence for $n>0$,
\begin{equation}
\begin{split}
\lim_{M\to\infty}\frac{|\Lambda_{(n,\infty)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}
=& \lim_{M\to\infty}\frac{|\Lambda_{(0,\infty)}(M-nR,\Gamma-n[\gamma])|}{|\Lambda(M-nR,\Gamma-n[\gamma])|}\frac{|\Lambda(M-nR,\Gamma-n[\gamma])|}{|\Lambda(M,\Gamma)|}\\
=&\lim_{M\to\infty}\frac{|\Lambda_{(0,\infty)}(M-nR,\Gamma-n[\gamma])|}{|\Lambda(M-nR,\Gamma-n[\gamma])|}\frac{(M-nR)^{2}}{M^{2}}=0.
\end{split}
\end{equation}
This completes the proof of the first statement.
To prove the second statement, we change the denominator in the statement to $|\Lambda(M)|$. By Claim \ref{easier}, it is sufficient to prove this version.
We define the map
\begin{equation}
f:\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)\times \Lambda_{(\infty,m)}(\frac{M}{2},\Gamma) \to \Lambda(M)
\end{equation}
as follows.
Let $\alpha$, $\beta \in I_{(\infty,m)}(\frac{M}{2})$. If $\alpha$ and $\beta$ have no common negative hyperbolic orbit, we define $f(\alpha,\beta)=\alpha\cup{\beta}$. Otherwise, that is if $\alpha$ and $\beta$ have some common negative hyperbolic orbit, we define $f(\alpha,\beta)$ by changing all multiplicities of negative hyperbolic orbits in $\alpha\cup{\beta}$ to one. This definition is well-defined.
Let $\delta$ be an element in the image of $f$. Then the number of multiplicity at $\gamma$ in $\delta$ is at most $\frac{M}{R}$. So
by combinatorial arguments, we can find that for every $m$, there is $C_{m}$ such that $\frac{C_{m}M}{R}>|f^{-1}(\delta)|$ for any $\delta \in \Lambda(M)$. So,
\begin{equation}
|\Lambda(M)|>\frac{|\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)|^2}{\frac{C_{m}M}{R}}
\end{equation}
thus
\begin{equation}
\frac{\frac{C_{m}M}{R}}{|\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)|}>\frac{|\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)|}{|\Lambda(M)|}.
\end{equation}
Suppose that there are $\epsilon>0$ and $M_{k}\to \infty$ such that $\frac{|\Lambda_{(\infty,m)}(\frac{M_{k}}{2},\Gamma)|}{|\Lambda(M_{k})|}>\epsilon$ then,
\begin{equation}
\begin{split}
\lim_{k\to \infty}\frac{|\Lambda_{(\infty,m)}(\frac{M_{k}}{2},\Gamma)|}{|\Lambda(M_{k})|} \leq&
\lim_{k\to \infty} \frac{C_{m}M_{k}}{R|\Lambda_{(\infty,m)}(\frac{M_{k}}{2},\Gamma)|}\\
=& \lim_{k\to \infty} \frac{C_{m}M_{k}}{R|\Lambda(M_{k})|}\frac{|\Lambda(M_{k})|}{|\Lambda_{(\infty,m)}(\frac{M_{k}}{2})|}\\
\leq&\lim_{k\to \infty} \frac{C_{m}M_{k}^{2}}{\epsilon R|\Lambda(M_{k})|}\cdot \frac{1}{M_{k}}\\
=&\frac{2C_{m}\mathrm{Vol}(Y,\lambda)}{\epsilon R|H_{1}(Y)|}\lim_{k\to \infty}\frac{1}{M_{k}}=0.
\end{split}
\end{equation}
This contradicts $\frac{|\Lambda_{(\infty,m)}(\frac{M_{k}}{2},\Gamma)|}{|\Lambda(M_{k})|}>\epsilon$. Hence $\lim_{M\to \infty}\frac{|\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)|}{|\Lambda(M)|}=0$ and so
\begin{equation}
\begin{split}
\lim_{M\to \infty}\frac{|\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)|}{|\Lambda(\frac{M}{2})|}=&\lim_{M\to \infty}\frac{|\Lambda_{(\infty,m)}(\frac{M}{2})|}{|\Lambda(M)|}\frac{|\Lambda(M)|}{|\Lambda(\frac{M}{2})|}\\
=&\lim_{M\to \infty}\frac{|\Lambda_{(\infty,m)}(\frac{M}{2},\Gamma)|}{|\Lambda(M)|}\frac{M^2}{(\frac{M}{2})^2}=0
\end{split}
\end{equation}
This completes the proof of the second statement.
Finally, we prove the third statement.
Note that for $p_{i}\in S_{\theta}$, the sequence of $p_{i+1}-p_{i}$ is monotone increasing with respect to $i$ and diverges to infinity as $i\to \infty$ (Proposition \ref{s}). Let $s\in \mathbb{Z}_{>0}$ be the order of $[\gamma]$ in $H_{1}(Y)$. Then for every $N\in \mathbb{Z}_{>0}$, there is $l\in \mathbb{Z}_{>0}$ such that $p_{i+1}-p_{i}>sN$ for every $i>l$.
By the first statement,
\begin{equation}\label{to0}
\lim_{M\to \infty}\frac{|\bigcup_{p_{i}\in S_{\theta},\,\,p_{i}<p_{l} } \Lambda_{(p_{i},\infty)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}=0.
\end{equation}
Since $|\Lambda_{(n,\infty)}(M,\Gamma)|\leq |\Lambda_{(n',\infty)}(M,\Gamma-(n-n')[\gamma])|$ for $n>n'$, if $p_{i}\geq p_{l}$,
\begin{equation}
\sum_{n=p_{i}+1}^{p_{i+1}}|\Lambda_{(n,\infty)}(M,\Gamma)|\geq N|\Lambda_{(p_{i+1},\infty)}(M,\Gamma)|
\end{equation}
and then
\begin{equation}
|\Lambda(M,\Gamma)|\geq N|\bigcup_{p_{i}\in S_{\theta},\,\,p_{l}\leq p_{i} } \Lambda_{(p_{i},\infty)}(M,\Gamma)|.
\end{equation}
By combining with (\ref{to0}), we have
\begin{equation}
\frac{1}{N}\geq \lim_{M\to \infty}\frac{|\bigcup_{p_{i}\in S_{\theta} } \Lambda_{(p_{i},\infty)}(M,\Gamma)|}{|\Lambda(M,\Gamma)|}.
\end{equation}
Since we can pick up $N$ arbitrary large, we complete the proof of the third statement.
\end{proof}
\subsection{Proof of Proposition \ref{main index 2} under Lemma \ref{mainlemma}}
Recall that since Lemma \ref{isomero} there is an isomorphism
\begin{equation}\label{isom}
\mathrm{ECH}(Y,\lambda,0) \cong \bigoplus_{k=0}^{\infty}\mathbb{F}\langle \alpha_{k} \rangle \bigoplus \bigoplus_{j=1}^{m}\mathbb{F}\langle \beta_{j}\rangle
\end{equation}
with $U\langle \alpha_{k} \rangle=\langle \alpha_{k-1} \rangle$ for $k\geq 1$ (note that $U\langle \alpha_{0} \rangle$ is not necessarily $0$). Moreover all but finite admissible orbit sets are in $\{\alpha_{k}\}_{k\in \mathbb{Z}_{\geq0}}$. Note that $A(\alpha_{k})>A(\alpha_{l})$ if and only if $k>l$. Here we omit some notations from (\ref{importantiso}) and from now on, we do under this notations unless there is confusion.
The aim of this subsection is the proof of the next proposition under Lemmma \ref{mainlemma}.
\begin{prp}\label{main index 2}
For every $\epsilon>0$ and positive integer $l$, there is $k\in \mathbb{Z}_{>0}$ which satisfies the following condition.
The $l+1$ consecutive orbit sets $\alpha_{k},\,\,\alpha_{k+1},....\,\,\alpha_{k+l}$ satisfies for every $0\leq i \leq l$,
\begin{enumerate}
\item $J(\alpha_{k+i+1},\alpha_{k+i})=2$
\item $A(\alpha_{k+i+1})-A(\alpha_{k+i})<\epsilon$
\item $E(\alpha_{k+i})\notin S_{\theta}\cup{S_{-\theta}}$
\item $E(\alpha_{k+i})>p_{1},\,\,q_{1}$
\item $H(\alpha_{k+i})>4$
\item In the notation of (\ref{importantiso}), for any $\Gamma \in H_{1}(Y)$ and $1\leq j \leq m_{\Gamma}$, $A(\alpha_{k+i})>A(\gamma)+A(\beta^{\Gamma}_{j})$ .
\end{enumerate}
\end{prp}
The next lemma plays an important role in the proof of Proposition \ref{main index 2}.
\begin{lem}\label{mainlemma}
For any sufficiently small $\epsilon>0$, there is no positive integer $k$ which
satisfies
\begin{enumerate}
\item $J(\alpha_{k+1},\alpha_{k})\leq 1$
\item $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$
\item $E(\alpha_{k+1}),\,\,E(\alpha_{k})>p_{1},\,\,q_{1}$
\item $E(\alpha_{k+1}),\,\, E(\alpha_{k})\notin S_{\theta}\cup{ S_{-\theta}}$
\item $H(\alpha_{k+1}),\,\,H(\alpha_{k})>4$
\item In the notation of (\ref{importantiso}), for any $\Gamma \in H_{1}(Y)$ and $1\leq j \leq m_{\Gamma}$, $A(\alpha_{k})>A(\gamma)+A(\beta^{\Gamma}_{j})$.
\end{enumerate}
Here we note that the number of $k$ which does not satisfy the sixth condition is finite.
\end{lem}
Before proving Lemma \ref{mainlemma}, we will give the proof of Proposition \ref{main index 2} under Lemma \ref{mainlemma}. We will prove Lemma \ref{mainlemma} in the next section.
To prove Proposition \ref{main index 2}, we introduce some notations as follows.
For positive integer $k$ and $\epsilon>0$, we set
\begin{equation}
\hat{I}_{(n,m)}(k):=\{\,\,\alpha_{k'}\,\,|\,\, k' \leq k, \,\, (E(\alpha_{k'}),H(\alpha_{k'}))=(n,m)\,\,\}
\end{equation}
\begin{equation}
\hat{I}_{(n,\infty)}(k):=\bigcup_{m=0}^{\infty}\hat{I}_{(n,m)}(k)
\end{equation}
\begin{equation}
\hat{I}_{(\infty,m)}(k):=\bigcup_{n=0}^{\infty}\hat{I}_{(n,m)}(k)
\end{equation}
\begin{equation}
\hat{I}_{\geq\epsilon}(k):=\{\,\,\alpha_{k'}\,\,|\,\, k' \leq k, \,\, A(\alpha_{k'+1})-A(\alpha_{k'})\geq \epsilon\,\,\}
\end{equation}
\begin{equation}
\hat{I}_{=2,<\epsilon}(k):=\{\,\,\alpha_{k'}\,\,|\,\, k' \leq k, \,\, J(\alpha_{k'+1},\alpha_{k'})=2,\,\,\,\alpha_{k'+1},\,\,\alpha_{k'}\,\,\mathrm{satisfy\,\,2,3,4,5,6\,\, in\,\,Lemma\,\, \ref{mainlemma}}\}
\end{equation}
\begin{equation}
\hat{I}_{>2,<\epsilon}(k):=\{\,\,\alpha_{k'}\,\,|\,\, k' \leq k, \,\, J(\alpha_{k'+1},\alpha_{k'})>2,\,\,\,\alpha_{k'+1},\,\,\alpha_{k'}\,\,\mathrm{satisfy\,\,2,3,4,5,6\,\, in\,\,Lemma\,\, \ref{mainlemma}}\}
\end{equation}
\begin{proof}[\bf Proof of Proposition \ref{main index 2} under Lemma \ref{mainlemma}]
For large $n$, there is a $k_{n}\in \mathbb{Z}$ such that $\alpha_{k_{n}}=\{(\gamma,sn)\}$ where $s\in \mathbb{Z}_{>0}$ is the order of $[\gamma]$ in $H_{1}(Y)$. Here we use $[(\gamma,sn)]=0$ and (\ref{isom}).
\begin{cla}\label{asdg}
\begin{enumerate}
\item For every positive integer $n$,
$
\lim_{k\to \infty}\frac{|\hat{I}_{(n,\infty)}(k)|}{k}=0
$
\item For every positive integer $m$,
$
\lim_{k\to \infty}\frac{|\hat{I}_{(\infty,m)}(k)|}{k}=0
$
\item
$
\lim_{k\to \infty}\frac{|\bigcup_{n\in S_{\theta} } \hat{I}_{(n,\infty)}(k)|}{k}=0
$
\item $
\lim_{n\to \infty}\frac{|\hat{I}_{\geq\epsilon}(k_{n})|}{k_{n}}=0
$
\end{enumerate}
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{asdg}]
The first three statement is just restatements of Proposition \ref{asympdense}. So we have only to prove the forth statement.
Since $I(\alpha_{k_{n}})=2k_{n}+I(\alpha_{0})$, $\lim_{n \to \infty}\frac{(snR)^2}{2k_{n}}=\mathrm{Vol}(Y,\lambda)$ and so $n<C\sqrt{k_{n}}$ for some $C>0$. So by the definition, for some $C>0$,
\begin{equation}
\epsilon |\hat{I}_{\geq\epsilon}(k_{n})|<A(\alpha_{k_{n}})-A(\alpha_{0})=snR-A(\alpha_{0})<C\sqrt{k_{n}}
\end{equation}
Thus
\begin{equation}
0= \lim_{n \to \infty}\frac{C\sqrt{k_{n}}}{\epsilon k_{n}}\geq \lim_{n\to \infty}\frac{|\hat{I}_{\geq\epsilon}(k_{n})|}{k_{n}}.
\end{equation}
This finishes the proof of Claim \ref{asdg}.
\end{proof}
By definition,
\begin{equation}
I(\alpha_{k_{n}})-J_{0}(\alpha_{k_{n}})=2nc_{s\gamma}+2\lfloor n\theta \rfloor +1
\end{equation}
where $c_{s\gamma}=c_{1}(\xi|_{*},\tau)$ with $\{*\}=H_{2}(Y;(\gamma,s),\emptyset)$ and $J_{0}(\alpha_{k_{n}})=J_{0}(\alpha_{k_{n}},\emptyset)$. So there is $C>0$ such that $|I(\alpha_{k_{n}})-J_{0}(\alpha_{k_{n}})|<Cn$ and then
\begin{equation}
J_{0}(\alpha_{k_{n}})=\sum_{i=0}^{k_{n}-1}J_{0}(\alpha_{i+1},\alpha_{i})<2k_{n}+C\sqrt{k_{n}}.
\end{equation}
for some $C>0$. By considering $J_{0}(\alpha_{i+1},\alpha_{i})\geq -1$ and Lemma \ref{mainlemma}, for some $C>0$ we have
\begin{equation}
\begin{split}
2|\hat{I}_{=2,<\epsilon}(k_{n})|+3|\hat{I}_{>2,<\epsilon}(k_{n})|\leq & 2k_{n}+C\sqrt{k_{n}}+|\bigcup_{p_{i}\in S_{\theta}}\hat{I}_{(p_{i},\infty)}(k_{n}) |\\
&+|\bigcup_{q_{i}\in S_{-\theta}}\hat{I}_{(q_{i},\infty)}(k_{n}) |
+\sum_{i=0}^{\mathrm{max}(p_{1},q_{1})}|\hat{I}_{(i,\infty)}(k_{n})|\\
&+\sum_{i=0}^{4}|\hat{I}_{(\infty,i)}(k_{n})|+|\hat{I}_{\geq\epsilon}(k_{n})|
\end{split}
\end{equation}
Since the right hand side over $k_{n}$ converges to 2 as $n\to \infty$, we have
\begin{equation}\label{ikkaime}
\lim_{n\to \infty}\frac{ 2|\hat{I}_{=2,<\epsilon}(k_{n})|+3|\hat{I}_{>2,<\epsilon}(k_{n})|}{k_{n}}\leq 2.
\end{equation}
On the other hand, we have
\begin{equation}
\begin{split}
k_{n}\geq |\hat{I}_{=2,<\epsilon}(k_{n})|+|\hat{I}_{>2,<\epsilon}(k_{n})|\geq & k_{n}-C\sqrt{k_{n}}-|\bigcup_{p_{i}\in S_{\theta}}\hat{I}_{(p_{i},\infty)}(k_{n}) |\\
&-|\bigcup_{q_{i}\in S_{-\theta}}\hat{I}_{(q_{i},\infty)}(k_{n}) |
-\sum_{i=0}^{\mathrm{max}(p_{1},q_{1})}|\hat{I}_{(i,\infty)}(k_{n})|\\
&-\sum_{i=0}^{4}|\hat{I}_{(\infty,i)}(k_{n})|-|\hat{I}_{\geq\epsilon}(k_{n})|.
\end{split}
\end{equation}
And so we have
\begin{equation}\label{nikaime}
\lim_{n\to \infty}\frac{ |\hat{I}_{=2,<\epsilon}(k_{n})|+|\hat{I}_{>2,<\epsilon}(k_{n})|}{k_{n}}=1.
\end{equation}
From (\ref{ikkaime}) and (\ref{nikaime}), we have
\begin{equation}
\lim_{n\to \infty}\frac{ |\hat{I}_{=2,<\epsilon}(k_{n})|}{k_{n}}=1.
\end{equation}
This implies that for every positive integer $l$, if $n$ is sufficiently large, we can pick up $l+1$ consecutive orbit sets $\alpha_{k}$, $\alpha_{k+1}$,...$\alpha_{k+l}$ which satisfy 2, 3, 4, 5 in Lemma \ref{mainlemma} and $J(\alpha_{k+i+1},\alpha_{k+i})=2$ for $0\leq i \leq l-1$. We complete the proof of Proposition \ref{main index 2}.
\end{proof}
\section{Proof of Lemma \ref{mainlemma}}
For our purpose, we choose $\epsilon>0$ such that $\epsilon < \frac{1}{10^{5}}\mathrm{min}\{A(\alpha)\,|\alpha\,\,\mathrm{is\,\,a\,\,Reeb\,\,orbit}\}$ and make it smaller as needed.
Since $U\langle \alpha_{k+1} \rangle= \langle \alpha_{k} \rangle$, there is at least one $J$-holomorphic curve whose ECH index is equal to 2, and we write $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha_{k+1},\alpha_{k})$. Note that $u_{1}$ is through a fixed generic point $z$.
To prove Lemma \ref{mainlemma}, we prepare some notations as follows. Let $g$, $k$ and $l$ be the genus of $u_{1}$, the number of punctures of $u_{1}$ and $\sum_{i}(n_{i}^{+}-1)+\sum_{j}(n_{j}^{-}-1)$ respectively. In this notation, $J_{0}(\alpha_{k+1},\alpha_{k})=-2+2g+k+l$. Note that $k$ is definitely positive and $u_{1}$ has at least one positive end because of the maximum principle. In the proof of Lemma \ref{mainlemma}, we have only to consider the cases $J_{0}=-1$, $0$, $1$. To make the proof easier to understand, we make a list of their topological types as follows.
\begin{itemize}
\item[\bf Case] $J_{0}=-1$
In this case, $(g,k,l)=(0,1,0)$ may appear as $J$-holomorphic curves counted by $U$-map.
\item $(g,k,l)=(0,1,0)$
\begin{tikzpicture}
\draw (7,-2.5)--(12,-2.5);
\draw (7,-3.5)--(12,-3.5);
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (9,-4.5) to [out=0,in=180] (12,-4);
\draw (9,-5.5) to [out=0,in=180] (12,-5);
\draw (9,-4.5) to [out=180,in=90] (8,-5);
\draw (8,-5) to [out=270,in=180] (9,-5.5);
\draw[densely dotted, thick] (7,-2)--(7,-6) ;
\draw[densely dotted, thick] (12,-2)--(12,-6) ;
\draw (7,-6.3) node{$-\infty$};
\draw (12,-6.3) node{$+\infty$};
\draw (10,-5) node{$u_{1}$};
\end{tikzpicture}
\item[\bf Case] $J_{0}=0$
In this case, $(g,k,l)=(0,1,1)$, $(0,2,0)$ may appear as $J$- holomorphic curves counted by the $U$-map.
\item $(g,k,l)=(0,1,1)$
This case has only one type as follows.
\begin{tikzpicture}
\draw (7,-2.5)--(12,-2.5);
\draw (7,-3.5)--(12,-3.5);
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4)--(12,-4);
\draw (7,-5)--(12,-5);
\draw (12,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (9,-4.5) to [out=0,in=180] (12,-4);
\draw (9,-5.5) to [out=0,in=180] (12,-5);
\draw (9,-4.5) to [out=180,in=90] (8,-5);
\draw (8,-5) to [out=270,in=180] (9,-5.5);
\draw[densely dotted, thick] (7,-2)--(7,-6) ;
\draw[densely dotted, thick] (12,-2)--(12,-6) ;
\draw (7,-6.3) node{$-\infty$};
\draw (12,-6.3) node{$+\infty$};
\draw (9.8,-5.2) node{$u_{1}$};
\end{tikzpicture}
\item $(g,k,l)=(0,2,0)$
This case has two types as follows.
\begin{tikzpicture}
\draw (7,-2.5)--(12,-2.5);
\draw (7,-3.5)--(12,-3.5);
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4.5) to [out=0,in=180] (12,-4);
\draw (7,-5.5) to [out=0,in=180] (12,-5);
\draw[densely dotted, thick] (7,-2)--(7,-6) ;
\draw[densely dotted, thick] (12,-2)--(12,-6) ;
\draw (7,-6.3) node{$-\infty$};
\draw (12,-6.3) node{$+\infty$};
\draw (10,-4.8) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (19,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-1.8)--(14,-1.8) ;
\draw (19,-2.8)--(14,-2.8) ;
\draw (14,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (14,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (15.5,-5) to [out=90,in=180] (19,-3.5);
\draw (15.5,-5) to [out=270,in=180] (19,-6.8);
\draw (19,-5.8) to [out=180,in=270] (17,-5);
\draw (17,-5) to [out=90,in=180] (19,-4.5);
\draw[densely dotted, thick] (14,-1.6)--(14,-7);
\draw[densely dotted, thick] (19,-1.6)--(19,-7) ;
\draw (14,-7.5) node{$-\infty$};
\draw (19,-7.5) node{$+\infty$};
\draw (16.5,-4.9) node{$u_{1}$};
\end{tikzpicture}
\item[\bf Case] $J_{0}=1$
In this case, $(g,k,l)=(0,3,0)$, $(0,2,1)$, $(1,1,0)$ may appear as $J$-holomorphic curves counted by the $U$-map. Note that we can see from the definitions of $g$, $k$, $l$ and geometric observation that the case $(g,k,l)=(0,1,2)$ satisfies the equation $J_{0}=-2+2g+k+l=1$ but this can not appear as a $J$-holomorphic curve.
\item $(g,k,l)=(0,3,0)$
This case has three types as follows.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) to [out=0,in=180] (11,-3.5);
\draw (6,-5) to [out=0,in=180] (11,-6.8);
\draw (11,-5.8) to [out=180,in=270] (9,-5);
\draw (9,-5) to [out=90,in=180] (11,-4.5);
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (8,-4.7) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (18,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (18,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (18,-1.8)--(13,-1.8) ;
\draw (18,-2.8)--(13,-2.8) ;
\draw (13,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (18,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (18,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (18,-5) to [out=180,in=0] (13,-6.8);
\draw (13,-5.8) to [out=0,in=270] (15,-4.8);
\draw (15,-4.8) to [out=90,in=0] (13,-4.3);
\draw (18,-4) to [out=180,in=0] (13,-3.3);
\draw (13,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (13,-1.6)--(13,-7);
\draw[densely dotted, thick] (18,-1.6)--(18,-7) ;
\draw (13,-7.5) node{$-\infty$};
\draw (18,-7.5) node{$+\infty$};
\draw (16,-4.7) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-6.8) to [out=180,in=270] (8,-5);
\draw (8,-5) to [out=90,in=180] (11,-3);
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (9,-4.9) node{$u_{1}$};
\draw (11,-4) to [out=180,in=90] (10.5,-4.2);
\draw (10.5,-4.2) to [out=270,in=180] (11,-4.4);
\draw (11,-5.4) to [out=180,in=90] (10.5,-5.6);
\draw (10.5,-5.6) to [out=270,in=180] (11,-5.8);
\end{tikzpicture}
\item $(g,k,l)=(0,2,1)$
This case has three types as follows.
\begin{tikzpicture}
\draw (2,-2.5)--(7,-2.5);
\draw (2,-3.5)--(7,-3.5);
\draw (2,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-4)--(7,-4);
\draw (2,-5)--(7,-5);
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) to [out=0,in=180] (7,-4);
\draw (2,-7) to [out=0,in=180] (7,-5);
\draw[densely dotted, thick] (2,-2.2)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.2)--(7,-7.3) ;
\draw (4.5,-5.5) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (2,-2.5)--(7,-2.5);
\draw (2,-3.5)--(7,-3.5);
\draw (2,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-4)--(7,-4);
\draw (2,-5)--(7,-5);
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-4) to [out=0,in=180] (7,-6);
\draw (2,-5) to [out=0,in=180] (7,-7);
\draw[densely dotted, thick] (2,-2.2)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.2)--(7,-7.3) ;
\draw (4.5,-5.5) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (19,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-1.8)--(14,-1.8) ;
\draw (19,-2.8)--(14,-2.8) ;
\draw (14,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (14,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (15.5,-5) to [out=90,in=180] (19,-3.5);
\draw (15.5,-5) to [out=270,in=180] (19,-6.8);
\draw (19,-5.8) to [out=180,in=270] (17,-5);
\draw (17,-5) to [out=90,in=180] (19,-4.5);
\draw[densely dotted, thick] (14,-1.6)--(14,-7);
\draw[densely dotted, thick] (19,-1.6)--(19,-7) ;
\draw (14,-7.5) node{$-\infty$};
\draw (19,-7.5) node{$+\infty$};
\draw (16.5,-4.9) node{$u_{1}$};
\draw (14,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (14,-4.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-3.5)--(14,-3.5) ;
\draw (19,-4.5)--(14,-4.5) ;
\end{tikzpicture}
\item $(g,k,l)=(1,1,0)$
This case has only one type as follows.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) to [out=180,in=0] (8.5,-3.5);
\draw (11,-5.4) to [out=180,in=0] (8.5,-6.3);
\draw (8.5,-3.5) to [out=180,in=90] (7,-4.7);
\draw (7,-4.7) to [out=270,in=180] (8.5,-6.3);
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (8,-4.8) to [out=30,in=150] (9.5,-4.8);
\draw (8.2,-4.7) to [out=350,in=200] (9.3,-4.7);
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (8.4,-5.5) node{$u_{1}$};
\end{tikzpicture}
\end{itemize}
\begin{proof}[\bf Proof of Lemma \ref{mainlemma}]
Since $J_{0}\geq -1$, we prove Lemma \ref{mainlemma} by dividing into three cases $J_{0}(\alpha_{k+1},\alpha_{k})=-1,\,\,0,\,\,1$.
\item[\bf Case1.] $J_{0}(\alpha_{k+1},\alpha_{k})=-1$
In this case, only $(g,k,l)=(0,1,0)$ satisfies this equation. The integral value of this $J$-holomorphic curve over $d\lambda$ is equal to $A(\alpha_{k+1})-A(\alpha_{k})$ by Stoke's theorem and so moreover equal to some action of a Reeb orbit. This contradicts $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon < \frac{1}{10^5}\mathrm{max}\{A(\alpha)\,\,|\alpha\,\,\mathrm{is\,\,a\,\,Reeb\,\,orbit}\,\,\}$.
\item[\bf Case2.] $J_{0}(\alpha_{k+1},\alpha_{k})=0$
In this case, $(g,k,l)=(0,2,0)$, $(0,1,1)$. For the same reason as Case1, we have only to consider the case $(g,k,l)=(0,2,0)$ and $u_{1}$ has both positive and negative ends and their two orbits are different each other. Since $E(\alpha_{k+1}),\,\, E(\alpha_{k})\notin S_{\theta}\cup{S_{-\theta}}$, $l=0$ and the partition conditions of the ends of admissible curves (Definition \ref{part}, Definition \ref{adm} and Proposition \ref{ind}), $u_{1}$ has no end asymptotic to $\gamma$. Moreover since $E(\alpha_{k+1}),\,\, E(\alpha_{k})>p_{1},\,\,q_{1}>1$, $u_{0}$ contains some covering of $\mathbb{R}\times \gamma$ and so we have $E(\alpha_{k+1})=E(\alpha_{k})$. Let $\delta_{1}$ and $\delta_{2}$ be the Reeb orbits where the positive and negative end of $u_{1}$ are asymptotic respectively. We set $E(\alpha_{k+1})=E(\alpha_{k})=M$. Then we can describe $\alpha_{k+1}$, $\alpha_{k}$ as $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{1},1)}}$ and $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{2},1)}}$ respectively where $\hat{\alpha}$ is an admissible orbit set consisting of some negative hyperbolic orbits which do not contain $\delta_{1}$, $\delta_{2}$.
By the above argument, we can see that for any generic $z\in Y$, non trivial parts of $J$-holomorhic curves counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$ through $z$ are in $\mathcal{M}^{J}(\delta_{1},\delta_{2})$ and of genus zero.
Now, we consider the behaviors of such $J$-holomorphic curves as $z\to \gamma$. By compactness argument (in this situation, for example see \cite[$\S 9$]{H1}), its some subsequence have a limiting $J$-holomorphic curve $u_{1}^{\infty}$ up to $\mathbb{R}$-action which may be splitting into two floors.
Suppose that $u_{1}^{\infty}\in \mathcal{M}^{J}(\delta_{1},\delta_{2})$, then by construction, it intersects with $\mathbb{R}\times \gamma$ but this contradicts the admissibility of curves of ECH index 2. So we may assume that the limiting curve has two floors. By construction, both have ends asymptotic to $\gamma$ and same multiplicity.
we set $u_{1}^{\infty}=(u_{-}^{\infty},u_{+}^{\infty})$ where $u_{\pm}^{\infty}$ are top and bottom curves up to $\mathbb{R}$-action respectively (see the below figure). The additivity and non negativeness of ECH index, we have $I(u_{-}^{\infty}\cup{u_{0}})=I(u_{+}^{\infty}\cup{u_{0}})=1$ and thus the multiplicity of positive or negative ends of $u_{\pm}^{\infty}$ are one since $S_{\theta}\cap{S_{-\theta}}=\{1\}$. Then we have $u_{-}^{\infty}\cup{u_{0}}\in \mathcal{M}^{J}(\hat{\alpha}\cup{(\gamma,M+1)},\alpha_{k})$ and $u_{+}^{\infty}\cup{u_{0}}\in \mathcal{M}^{J}(\alpha_{k+1},\hat{\alpha}\cup{(\gamma,M+1)})$.
By definition assumption, $\hat{\alpha}\cup{(\gamma,M+1)}$ is admissible and have no positive hyperbolic orbit. So by (\ref{mod2}), $I(u_{-}^{\infty}\cup{u_{0}})=I(u_{+}^{\infty}\cup{u_{0}})=0\,\,\mathrm{mod}\,2$. This is a contradiction.
Here, we introduce another way to derive a contradiction from $u_{-}^{\infty}\cup{u_{0}}\in \mathcal{M}^{J}(\hat{\alpha}\cup{(\gamma,M+1)},\alpha_{k})$, $u_{+}^{\infty}\cup{u_{0}}\in \mathcal{M}^{J}(\alpha_{k+1},\hat{\alpha}\cup{(\gamma,M+1)})$ and $I(u_{-}^{\infty}\cup{u_{0}})=I(u_{+}^{\infty}\cup{u_{0}})=1$. From the partition condition of admissible $J$-holomorphic curve at $\gamma$, we have $1=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M+1\}})=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M+1\}})$. But by the assumption $M=E(\alpha_{k+1})=E(\alpha_{k})>p_{1},\,\,q_{1}$, we have $\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M+1\}})$, $\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M+1\}}) >p_{1},\,\,q_{1}>1$. This is a contradiction.
\begin{tikzpicture}\label{fig;1}
\draw (2,-4)--(7,-4);
\draw (2,-5)--(7,-5);
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) to [out=0,in=180] (7,-4);
\draw (2,-7) to [out=0,in=180] (7,-5);
\draw[densely dotted, thick] (2,-3.7)--(2,-7.3);
\draw[densely dotted, thick] (7,-3.7)--(7,-7.3) ;
\draw (7,-1.3) node{$u_{1}$};
\draw (4.5,-2.5) node{$\alpha_{k}$};
\draw (9.5,-2.5) node{$\alpha_{k+1}$};
\draw (10.7,0.5) node{$(\gamma,M)$};
\draw (10.5,-1) node{$(\delta_{1},1)$};
\draw (3.5,-1.5) node{$(\delta_{2},1)$};
\draw (4.5,-5.6) node{$u_{-}^{\infty}$};
\draw (7,-7.5) node{$\hat{\alpha}\cup{(\gamma,M+1)}$};
\draw (9.7,-5.5) node{$u_{+}^{\infty}$};
\draw (2,-7.5) node{$\alpha_{k}$};
\draw (12,-7.5) node{$\alpha_{k+1}$};
\draw (13.2,-4.5) node{$(\gamma,M)$};
\draw (13,-6) node{$(\delta_{1},1)$};
\draw (1.2,-6.5) node{$(\delta_{2},1)$};
\draw (4.5,1)--(9.5,1);
\draw (4.5,0)--(9.5,0);
\draw (4.5,1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (4.5,0)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (4.5,-2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-1) to [out=0,in=180] (9.5,-0.5);
\draw (4.5,-2) to [out=0,in=180] (9.5,-1.5);
\draw[densely dotted, thick] (4.5,1.3)--(4.5,-2.3);
\draw[densely dotted, thick] (9.5,1.3)--(9.5,-2.3) ;
\draw (9.5,1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (9.5,0) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (9.5,-0.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (9.5,-1.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-3)node{$\Downarrow$};
\draw (7,-4)--(12,-4);
\draw (7,-5)--(12,-5);
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-5.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-6.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) to [out=0,in=180] (12,-5.5);
\draw (7,-5) to [out=0,in=180] (12,-6.5);
\draw[densely dotted, thick] (12,-3.7)--(12,-7.3) ;
\end{tikzpicture}
\item[\bf Case3.] $J_{0}(\alpha_{k+1},\alpha_{k})=1$
In this case, $(g,k,l)=(0,3,0)$, $(0,2,1)$, $(1,1,0)$.
We can exclude the case $(g,k,l)=(1,1,0)$ in the same way as Case1. If $(g,k,l)=(0,3,0)$, we have $E(\alpha_{k+1})=E(\alpha_{k})$ just like the way explained in Case2. On the other hand, If $(g,k,l)=(0,2,1)$, the image of $u_{0}$ contains $\mathbb{R}\times \gamma$ and also one positive end or negative end of $u_{1}$ is asymptotic to $\gamma$ and thus $E(\alpha_{k+1})\neq E(\alpha_{k})$. This implies that the pair $\alpha_{k+1}$, $\alpha_{k}$ which $(g,k,l)=(0,3,0)$ or $(g,k,l)=(0,2,1)$ types $J$-holomorphic curves by the $U$-map can occur are mutually exclusive.
\item[\bf 1.] If $(g,k,l)=(0,3,0)$.
Let $u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha_{k+1},\alpha_{k})$ be a $J$-holomorphic curve counted the by $U$-map. Since $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$, $u_{1}$ has either two positive ends and one negative end or one positive end and two negative ends. Without loss of generality, we may assume that $u_{1}$ has two positive ends asymptotic to $\delta_{1}$, $\delta_{2}$ and one negative end asymptotic to $\delta_{3}$. Note that $\delta_{1}$, $\delta_{2}$ and $\delta_{3}$ are mutually different because of the smallness of $A(\alpha_{k+1})-A(\alpha_{k})$.
In this notation. we have $u_{1}\in \mathcal{M}^{J}((\delta_{1},1)\cup{(\delta_{2},1)},(\delta_{3},1))$ for any non-trivial parts of $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle = \langle \alpha_{k} \rangle$ and also we write $\alpha_{k+1}=\hat{\alpha}\cup{(\delta_{2},1)\cup{(\delta_{1},1)\cup{(\gamma,M)}}}$, $\alpha_{k}=\hat{\alpha}\cup{(\delta_{3},1)\cup{(\gamma,M)}}$.
As $z\to \gamma$, we have three possibilities of splitting of $u_{1}$(see the below figure). In any case, this is a contradiction in the same reason as Case2.
\begin{tikzpicture}
\draw (13.5,-1.0) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,-2.0) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,-1)--(8.5,-1) ;
\draw (13.5,-2)--(8.5,-2) ;
\draw (8.5,-1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (8.5,-2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,-6.5) node{$\alpha_{k+1}$};
\draw (8.5,-6.5) node{$\alpha_{k}$};
\draw (14.5,-1.5) node{$(\gamma,M)$};
\draw (14.5,-3.3) node{$(\delta_{1},1)$};
\draw (14.5,-5.5) node{$(\delta_{2},1)$};
\draw (7.5,-3.8) node{$(\delta_{3},1)$};
\draw (16,-13.5) node{$\alpha_{k+1}$};
\draw (6,-13.5) node{$\alpha_{k}$};
\draw (11,-13.5) node{$\hat{\alpha}\cup{(\gamma,M+1)\cup{(\delta_{1},1)}}$};
\draw (11.2,-14) node{$(\mathrm{resp.}\,\,\hat{\alpha}\cup{(\gamma,M+1)\cup{(\delta_{2},1)}})$};
\draw (17,-8.5) node{$(\gamma,M)$};
\draw (17,-10.3) node{$(\delta_{1},1)$};
\draw (17,-12.5) node{$(\delta_{2},1)$};
\draw (17,-15.5) node{$(\gamma,M)$};
\draw (17,-17.3) node{$(\delta_{1},1)$};
\draw (17,-19.5) node{$(\delta_{2},1)$};
\draw (17.5,-10.8) node{$(\mathrm{resp.}\,\,(\delta_{2},1))$};
\draw (17.5,-13) node{$(\mathrm{resp.}\,\,(\delta_{1},1))$};
\draw (16,-20.5) node{$\alpha_{k+1}$};
\draw (6,-20.5) node{$\alpha_{k}$};
\draw (11,-20.5) node{$\hat{\alpha}\cup{(\gamma,M+1)}$};
\draw (5,-10.8) node{$(\delta_{3},1)$};
\draw (5,-17.8) node{$(\delta_{3},1)$};
\draw (8.5,-3.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (8.5,-4.2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,-2.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,-3.7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,-6) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (8.5,-3.2) to [out=0,in=180] (13.5,-2.7);
\draw (8.5,-4.2) to [out=0,in=180] (13.5,-6);
\draw (13.5,-5) to [out=180,in=270] (11.5,-4.2);
\draw (11.5,-4.2) to [out=90,in=180] (13.5,-3.7);
\draw[densely dotted, thick] (8.5,-0.7)--(8.5,-6.2);
\draw[densely dotted, thick] (13.5,-0.7)--(13.5,-6.2) ;
\draw (10.5,-3.9) node{$u_{1}$};
\draw (11,-7)node{$\Downarrow$};
\draw (11,-8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-9)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-8)--(6,-8) ;
\draw (11,-9)--(6,-9) ;
\draw (6,-8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-9)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-10.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-11.2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-12) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-13)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-10.2) to [out=0,in=180] (11,-8);
\draw (6,-11.2) to [out=0,in=180] (11,-13);
\draw (11,-9) to [out=180,in=90] (9,-10.5);
\draw (9,-10.5) to [out=270,in=180] (11,-12);
\draw (16,-9.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-10.7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-8) to [out=0,in=180] (16,-9.7);
\draw (11,-9) to [out=0,in=180] (16,-10.7);
\draw[densely dotted, thick] (6,-7.7)--(6,-13.2);
\draw[densely dotted, thick] (11,-7.7)--(11,-13.2) ;
\draw[densely dotted, thick] (16,-7.7)--(16,-13.2) ;
\draw (16,-12) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-13) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-12)--(11,-12) ;
\draw (16,-13)--(11,-13) ;
\draw (16,-8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-9) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-8)--(11,-8) ;
\draw (16,-9)--(11,-9) ;
\draw (16,-15) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-16) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-15)--(6,-15) ;
\draw (16,-16)--(6,-16) ;
\draw (11,-15) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-16)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-15) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-16)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-17.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-18.2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-16.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-17.7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-15) to [out=180,in=0] (6,-17.2);
\draw (11,-16) to [out=180,in=0] (6,-18.2);
\draw (16,-19) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-20) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-15) to [out=0,in=180] (16,-16.7);
\draw (11,-16) to [out=0,in=180] (16,-20);
\draw (16,-19) to [out=180,in=270] (14,-17.3);
\draw (14,-17.3) to [out=90,in=180] (16,-17.7);
\draw[densely dotted, thick] (6,-14.7)--(6,-20.2) ;
\draw[densely dotted, thick] (11,-14.7)--(11,-20.2);
\draw[densely dotted, thick] (16,-14.7)--(16,-20.2) ;
\end{tikzpicture}
\item[\bf 2.]If $(g,k,l)=(0,2,1)$.
From now on, we consider $(g,k,l)=(0,2,1)$. This case is more complicated but the way in this case also play an important role in Section 6 and beyond.
Since $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$, $u_{1}$ has both positive and negative end. Moreover by definition, $u_{0}$ contains some covering of $\mathbb{R}\times \gamma$ and either positive or negative end of $u_{1}$ asymptotic to $\gamma$. Because symmetry allows the same argument, here we consider only the case that the positive end of $u_{1}$ is asymptotic to $\gamma$ (see the below figure). Let $E(\alpha_{k+1})=M$. Then, By the admissibility of $u$, the multiplicity of positive end of $u_{1}$ at $\gamma$ is $p_{i}:=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$ and so $E(\alpha_{k})=M-p_{i}$. Let $\delta$ be the orbit where the negative end of $u_{1}$ is asymptotic . By the discussion so far, we can see that for any generic point $z\in Y$, any non trivial parts of $J$-holomorphic curves through $z$ counted by the $U$-map are in $\mathcal{M}^{J}((\gamma,p_{i}),(\delta,1))$.
We set $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$ and then $\alpha_{k}=\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,M-p_{i})}}$ where $\hat{\alpha}$ only contains negative hyperbolic orbits.
\begin{tikzpicture}
\draw (2,-2.5)--(7,-2.5);
\draw (2,-3.5)--(7,-3.5);
\draw (2,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-3.8) node{...};
\draw (8,-3) node{$\hat{\alpha}$};
\draw (2,-4)--(7,-4);
\draw (2,-5)--(7,-5);
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) to [out=0,in=180] (7,-4);
\draw (2,-7) to [out=0,in=180] (7,-5);
\draw[densely dotted, thick] (2,-2.2)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.2)--(7,-7.3) ;
\draw (8,-4.5) node{$(\gamma,M)$};
\draw (0.8,-4.5) node{$(\gamma,M-p_{i})$};
\draw (1,-6.5) node{$(\delta,1)$};
\draw (4.5,-5.5) node{$u_{1}$};
\draw (7,-7.5) node{$\alpha_{k+1}$};
\draw (2,-7.5) node{$\alpha_{k}$};
\end{tikzpicture}
\begin{cla}\label{mi1} In the above notation,
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,M-1)},\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i}-1)})=2
\end{equation}
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{mi1}]
Suppose that this claim is false. Since $A(\alpha_{k+1})-A(\alpha_{k})=A(\hat{\alpha}\cup{(\gamma,M-1)})-A(\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i}-1)})>0$ and (\ref{isom}), we have
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,M-1)},\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i}-1)})>2.
\end{equation}
By considering the sixth condition in Lemma \ref{mainlemma}, there is an admissible orbit set $\zeta$ with $[\zeta]=[\hat{\alpha}\cup{(\gamma,M-1)}]=[\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i}-1)}]$ such that $U\langle \hat{\alpha}\cup{(\gamma,M-1)} \rangle=\langle \zeta \rangle$ and $A(\hat{\alpha}\cup{(\gamma,M-1)})>A(\zeta)>A(\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i}-1)})$. This implies that $A(\alpha_{k+1})>A(\zeta\cup{(\gamma,1)})>A(\alpha_{k})$. This contradicts (\ref{importantiso}).
\end{proof}
\begin{rem}
In essence, Claim \ref{mi1} and Claim \ref{index2to4} come from only their topological conditions and the properties of ECH index (in particular the equation (\ref{espe}); For example see \cite[Proof of Proposition 7.1]{H1}). But to understand the proof only with the facts in this paper as much as possible, we prove them by using some special conditions.
\end{rem}
\begin{cla}\label{index2to4}
In the above notation, for any $p_{i}\leq N<p_{i+1}$
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,N)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i})}})=2.
\end{equation}
Moreover
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,p_{i+1})},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})=4.
\end{equation}
And for any $p_{i}<N\leq p_{i+1}$,
\begin{equation}
J_{0}(\hat{\alpha}\cup{(\gamma,N)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i})}})=1,
\end{equation}
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{index2to4}]
Let $\{Z\}=H_{2}(Y;\hat{\alpha}\cup{(\gamma,p_{i}),\hat{\alpha}\cup{(\delta_{1},1)}})$ and $\{Z_{\gamma}\}=H_{2}(Y;\gamma,\gamma)$ respectively.
Then by the definition, we have
\begin{equation}
\begin{split}
2=I(\alpha_{k+1},\alpha_{k})=&I(\hat{\alpha}\cup{(\gamma,M)},\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i})})\\
=&c_{1}(\xi|_{Z},\tau)+Q_{\tau}(Z,Z)+2(M-p_{i})Q_{\tau}(Z,Z_{\gamma})\\
&+\sum_{k=M-p_{i}+1}^{M}(2\lfloor k\theta \rfloor+1)-\mu_{\tau}(\delta).
\end{split}
\end{equation}
Here, we use the property $Q_{\tau}(Z_{1}+Z_{2},Z_{1}+Z_{2})=Q_{\tau}(Z_{1},Z_{1})+2Q_{\tau}(Z_{1},Z_{2})+Q_{\tau}(Z_{2},Z_{2})$ in Definition \ref{qdef} and $Q_{\tau}(Z_{\gamma})=0$ and $Q_{\tau}(Z,mZ_{\gamma})=mQ_{\tau}(Z,Z_{\gamma})$. These properties easily follows from the definition of $Q_{\tau}$ (see \cite[Lemma 8.5]{H1}).
And also
\begin{equation}
\begin{split}
2=I(\hat{\alpha}\cup{(\gamma,M-1)},&\hat{\alpha}\cup{(\delta,1)}\cup{(\gamma,M-p_{i}-1)})\\
=&c_{1}(\xi|_{Z},\tau)+Q_{\tau}(Z,Z)+2(M-p_{i}-1)Q_{\tau}(Z,Z_{\gamma})\\
&+\sum_{k=M-p_{i}}^{M-1}(2\lfloor k\theta \rfloor+1)-\mu_{\tau}(\delta).
\end{split}
\end{equation}
By taking the difference from the above equations, we have
\begin{equation}
2Q_{\tau}(Z,Z_{\gamma})+2(\lfloor M\theta \rfloor - \lfloor (M-p_{i})\theta \rfloor)=0
\end{equation}
Note that for any $p_{i} \leq N<p_{i+1}$, $\lfloor N\theta \rfloor - \lfloor (N-p_{i})\theta \rfloor=\lfloor p_{i}\theta \rfloor$ and moreover $\lfloor p_{i+1}\theta \rfloor - \lfloor (p_{i+1}-p_{i})\theta \rfloor=\lfloor p_{i}\theta \rfloor+1$. These facts are special cases of \cite[Lemma 4.5]{H1}. Hence
\begin{equation}\label{espe}
2Q_{\tau}(Z,Z_{\gamma})+2\lfloor p_{i}\theta \rfloor =0
\end{equation}
On the other hand, in the same way, for any $p_{i}< N\leq p_{i+1}$ we have
\begin{equation}
\begin{split}
I(\hat{\alpha}\cup{(\gamma,N)},&\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i})}})\\
&-I(\hat{\alpha}\cup{(\gamma,N-1)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i}-1)}})
\\
=&2Q_{\tau}(Z,Z_{\gamma})+2(\lfloor N\theta \rfloor - \lfloor (N-p_{i})\theta \rfloor)\\
=&2(\lfloor N\theta \rfloor - \lfloor (N-p_{i})\theta \rfloor-\lfloor p_{i}\theta \rfloor).
\end{split}
\end{equation}
This implies that for any $p_{i}\leq N<p_{i+1}$, $I(\hat{\alpha}\cup{(\gamma,N)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i})}})$ are equal to each other and hence 2, moreover we have $I(\hat{\alpha}\cup{(\gamma,p_{i+1})},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})=4$.
In the same way, we have
\begin{equation}
\begin{split}
J_{0}(\hat{\alpha}\cup{(\gamma,N)},&\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i})}})\\
&-J_{0}(\hat{\alpha}\cup{(\gamma,N-1)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i}-1)}})
\\
=&2Q_{\tau}(Z,Z_{\gamma})+2(\lfloor (N-1)\theta \rfloor - \lfloor (N-p_{i}-1)\theta \rfloor)\\
=&2(\lfloor(N-1)\theta \rfloor - \lfloor (N-p_{i}-1)\theta \rfloor-\lfloor p_{i}\theta \rfloor).
\end{split}
\end{equation}
This implies that for any $p_{i}< N \leq p_{i+1}$, $J_{0}(\hat{\alpha}\cup{(\gamma,N)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,N-p_{i})}})=1$.
We complete the proof of Claim \ref{index2to4}.
\end{proof}
Since $ I(\hat{\alpha}\cup{(\gamma,p_{i+1})},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})=4$, there is an unique admissible orbit set $\zeta$ such that $ I(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta)=2= I(\zeta,\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})$. Note that $U\langle \hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$, $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}} \rangle$.
\begin{cla}\label{el0}
The above $\zeta$ satisfies $E(\zeta)=0$.
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{el0}]
Suppose that $E(\zeta)>0$. Since $A(\hat{\alpha}\cup{(\gamma,p_{i+1})})>A(\zeta)>A(\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})$, we also have $A(\hat{\alpha}\cup{(\gamma,p_{i+1}-1)})>A(\zeta-(\gamma,1))>A(\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i}-1)}})$. Since (\ref{isom}), this indicates $I(\hat{\alpha}\cup{(\gamma,p_{i+1}-1)},\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i}-1)}})>2$. This contradicts Claim \ref{index2to4}. Therefore we complete the proof of Claim \ref{el0}.
\end{proof}
Since $J_{0}\geq -1$ and additivity of $J_{0}$, we have $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}))$=$(2,-1)$, $(1,0)$, $(0,1)$ or $(-1,2)$.
If $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}))$=$(2,-1)$ or $(-1,2)$, we can derive a contradiction in the same way as Case1 since $A(\hat{\alpha}\cup{(\gamma,p_{i+1})})-A(\zeta)<\epsilon$ and $A(\zeta)-A(\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})<\epsilon$. This is because $A(\hat{\alpha}\cup{(\gamma,p_{i+1})})-A(\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}})$ $=A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$. Thus we have only to consider the case $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}))$=$(1,0)$ or $(0,1)$. Here we note that the assumption $H(\alpha_{k})$, $H(\alpha_{k+1})>4$ in Proposition \ref{main index 2} implies that $\hat{\alpha}$ contains more than four negative hyperbolic orbits. In these cases, we derive contradictions by using the splitting behaviors of $J$-holomorphic curves as follows.
\item[\bf $\rm(\hspace{.18em}i\hspace{.18em})$.] $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}))$=$(0,1)$
From $A(\hat{\alpha}\cup{(\gamma,p_{i+1})})-A(\zeta)<\epsilon$ and $E(\zeta)=0$ and the partition condition of end, we have that there is a negative hyperbolic orbit $\delta'$ with $\delta'\notin \hat{\alpha}$ such that $\zeta=\hat{\alpha}\cup{(\delta',1)}$. Moreover the nontrivial parts of any $J$-holomorphic curve counted by $U\langle \hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ are of genus 0 and in $\mathcal{M}^{J}((\gamma,p_{i+1}),(\delta',1))$.
\ref{picture1}
\begin{tikzpicture}\label{picture1}
\draw (7,-2.5)--(12,-2.5);
\draw (7,-3.5)--(12,-3.5);
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (12,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4.5) to [out=0,in=180] (12,-4);
\draw (7,-5.5) to [out=0,in=180] (12,-5);
\draw[densely dotted, thick] (7,-2)--(7,-6) ;
\draw[densely dotted, thick] (12,-2)--(12,-6) ;
\draw (7,-6.5) node{$\zeta=\hat{\alpha}\cup{(\delta',1)}$};
\draw (12,-6.5) node{$\hat{\alpha}\cup{(\gamma,p_{i+1})}$};
\draw (13.1,-4.5) node{$(\gamma,p_{i+1})$};
\draw (6,-5) node{$(\delta',1)$};
\draw (13.2,-3) node{$(\eta,1)\in \hat{\alpha}$};
\end{tikzpicture}
Let us consider the behaviors of such curves as $z\to \eta$. In the same way as Case2, the limiting curve of such sequence splits and each of them has end at $\eta$. Furthermore both ECH indexes are one. Its multiplicities are two because of the admissibility of curves of ECH index 1. This implies that $|2A(\eta)-p_{i+1}R|<\epsilon$.
Consider back to the $J$-homolorphic curves of $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$. Its nontrivial parts are in $\mathcal{M}^{J}((\gamma,p_{i}),(\delta,{1}))$. By considering the behaviors of this curves as $z \to \eta$, we have $|2A(\eta)-p_{i}R|<\epsilon$. By combining with $|2A(\eta)-p_{i+1}R|<\epsilon$, we have $(p_{i+1}-p_{i})R<2\epsilon$ and so $p_{i+1}=p_{i}$. This is a contradiction.
\item[\bf $\rm(\hspace{.08em}ii\hspace{.08em})$.]If $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}))$=$(1,0)$
Consider the $J$-holomorphic curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}) \rangle$. In the same way as the above argument, we can find that there is an hyperbolic orbit $\delta'$ such that $\zeta=\hat{\alpha}\cup{(\delta,1)}\cup{(\delta',1)}$ and that nontrivial parts of any $J$-holomorphic curve counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\delta,1)\cup{(\gamma,p_{i+1}-p_{i})}}) \rangle$ are of genus 0 and in $\mathcal{M}^{J}((\delta',1),(\gamma,p_{i+1}-p_{i}))$. And also we have $|2A(\eta)-p_{i}R|<\epsilon$ and $|2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$. Note that since $E(\alpha_{k+1}),\,\, E(\alpha_{k})>p_{1},\,\, q_{1}$, we have $p_{i}>1$.
The next claim is obvious but often used later on. So we state here.
\begin{cla}\label{fre}
Suppose that $q_{i}\in S_{\theta}$ (resp. $p_{i}\in S_{-\theta}$). If $q_{i}>1$ (resp. $p_{i}>1$), then $ q_{i}\neq q_{i+1}-q_{i}$ (resp. $p_{i}\neq p_{i+1}-p_{i}$).
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{fre}]
Suppose that $q_{i}\in S_{\theta}$, then by Proposition \ref{s}, $q_{i+1}-q_{i}\in S_{-\theta}$ and since $S_{\theta}\cap{S_{-\theta}}=\{1\}$, if $q_{i}>1$, we have $q_{i}\neq q_{i+1}-q_{i}$. We can do the same in the case $p_{i}$.
\end{proof}
From $|2A(\eta)-p_{i}R|<\epsilon$ and $|2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$, we have $|p_{i}R-(p_{i+1}-p_{i})R|<2\epsilon$. This implies that $p_{i}=p_{i+1}-p_{i}$ but this contradicts Claim \ref{fre}.
\begin{tikzpicture}
\draw (3.1,-3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (2,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (-3,-2.5)--(2,-2.5);
\draw (-3,-3.5)--(2,-3.5);
\draw (-3,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (-3,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (-3,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (-3,-7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (-3,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (-3,-5.4)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (-3,-4.4) to [out=0,in=180] (2,-4);
\draw (-3,-5.4) to [out=0,in=180] (2,-5);
\draw (-3,-6)--(2,-6);
\draw (-3,-7)--(2,-7);
\draw[densely dotted, thick] (-3,-2)--(-3,-7.5);
\
\draw[densely dotted, thick] (2,-2)--(2,-7.5) ;
\draw (2,-8) node{$\zeta=\hat{\alpha}\cup{(\delta,1)\cup{(\delta',1)}}$};
\draw (-3,-8) node{$\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta,1)}$};
\draw (-4,-6.5) node{$(\delta,1)$};
\draw (-4.4,-4.8) node{$(\gamma,p_{i+1}-p_{i})$};
\draw (2.9,-4.5) node{$(\delta',1)$};
\end{tikzpicture}
Combining the above arguments, we complete the proof of Lemma \ref{mainlemma}.
\end{proof}
\section{The properties of certain $J_{0}=2$ curves}
To derive a contradiction from Proposition \ref{main index 2}, at first we state Proposition \ref{nagai} and use Section 6 and Section 7 to prove Proposition \ref{nagai}.
\textbf{Notation}
For any $a,\,b\in \mathbb{R}$, we write $a\approx b$ if $|a-b|<\frac{1}{100}\mathrm{min}\{A(\alpha)\,|\alpha\,\,\mathrm{is\,\,a\,\,Reeb\,\,orbit}\}$. In this notation, for $n,\,m\in \mathbb{Z}$ and $\tau>\frac{1}{100}$, if $n\tau R\approx m \tau R$, then $n=m$.
\begin{prp}\label{nagai}
Let $\epsilon<\frac{1}{10^5}\mathrm{min}\{A(\alpha)\,\,|\alpha\,\,\mathrm{is\,\,a\,\,Reeb\,\,orbit}\,\,\}$ and $k$ sufficiently large. Suppose that two admissible orbit sets $\alpha_{k+1}$ and $\alpha_{k}$ with $U\langle \alpha_{k+1} \rangle =\langle \alpha_{k} \rangle$ satisfy the following conditions.
\begin{enumerate}
\item $J(\alpha_{k+1},\alpha_{k})=2$
\item $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$
\item $E(\alpha_{k+1}),\,E(\alpha_{k})\notin S_{\theta}\cup{ S_{-\theta}}$
\item $E(\alpha_{k+1}),\,E(\alpha_{k})>p_{1},\,\,q_{1}$
\item $H(\alpha_{k+1}),\,H(\alpha_{k})>4$.
\end{enumerate}
Let $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha_{k+1},\alpha_{k})$ be any $J$-holomorphic curve counted by the $U$-map.
Then one of the following conditions holds.
\begin{description}
\item[(a).] Let $E(\alpha_{k+1})=M$ and $p_{i}:=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{1},1)}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{2},1)}}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1)\cup{(\gamma,p_{i})},(\delta_{2},1))$.
Moreover, $A(\delta_{1})\approx (p_{i+1}-p_{i})R$, $A(\delta_{2})\approx p_{i+1}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}p_{i+1}R$ or $A(\eta)\approx \frac{1}{2}(p_{i+1}-p_{i})R$.
\item[(a').]Let $E(\alpha_{k})=M$ and $q_{i}:=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{2},1)}}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1),(\delta_{2},1)\cup{(\gamma,q_{i})})$.
Moreover, $A(\delta_{1})\approx q_{i+1}R$, $A(\delta_{2})\approx (q_{i+1}-q_{i})R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}q_{i+1}R$ or $A(\eta)\approx \frac{1}{2}(q_{i+1}-q_{i})R$.
\item[(b).]Let $E(\alpha_{k+1})=M$ and $p_{i}:=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then $\frac{3}{2}p_{i}=p_{i+1}$ and there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$ and $u_{1}\in \mathcal{M}^{J}((\gamma,p_{i}),(\delta_{1},1)\cup{(\delta_{2},1)})$.
Moreover, $A(\delta_{1})\approx \frac{1}{2}p_{i}R$, $A(\delta_{2})\approx \frac{1}{2}p_{i}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}p_{i}R$ or $A(\eta)\approx \frac{1}{4}p_{i}R$.
\item[(b').]Let $E(\alpha_{k})=M$ and $q_{i}:=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then $\frac{3}{2}q_{i}=q_{i+1}$ and there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1)\cup{(\delta_{2},1)},(\gamma,q_{i}))$.
Moreover, $A(\delta_{1})\approx \frac{1}{2}q_{i}R$, $A(\delta_{2})\approx \frac{1}{2}q_{i}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}q_{i}R$ or $A(\eta)\approx \frac{1}{4}q_{i}R$.
\item[(c).]Let $E(\alpha_{k+1})=M$ and $p_{i}:=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then $\frac{4}{3}p_{i}=p_{i+1}$ and there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$ and $u_{1}\in \mathcal{M}^{J}((\gamma,p_{i}),(\delta_{1},1)\cup{(\delta_{2},1)})$.
Moreover, $A(\delta_{1})\approx \frac{2}{3}p_{i}R$, $A(\delta_{2})\approx \frac{1}{3}p_{i}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}p_{i}R$ or $A(\eta)\approx \frac{1}{6}p_{i}R$.
\item[(c').]Let $E(\alpha_{k})=M$ and $q_{i}:=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then $\frac{4}{3}q_{i}=q_{i+1}$ and there are and two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1)\cup{(\delta_{2},1)},(\gamma,p
q_{i}))$. .
Moreover, $A(\delta_{1})\approx \frac{2}{3}q_{i}R$, $A(\delta_{2})\approx \frac{1}{3}q_{i}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}q_{i}R$ or $A(\eta)\approx \frac{1}{6}q_{i}R$.
\end{description}
Note that \textbf{(a)} and \textbf{(a')}, \textbf{(b)} and \textbf{(b')}, \textbf{(c)} and \textbf{(c')} are symmetrical
respectively and, \textbf{(a}), \textbf{(a')}, \textbf{(b)}, \textbf{(b')}, \textbf{(c)} and \textbf{(c')} are mutually exclusive.
\end{prp}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (14.2,-3) node{$A(\eta)\approx \frac{1}{2}p_{i+1}R$ or $ \frac{1}{2}(p_{i+1}-p_{i})R$};
\draw (8.5,-3.2) node{...};
\draw (6,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) node{$(\delta_{1},1)$ } ;
\draw (13.5,-4.7) node{ $A(\delta_{1})\approx (p_{i+1}-p_{i})R$ } ;
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-6.3) node{$(\gamma,M)$} ;
\draw (4.8,-6.3) node{$(\gamma,M-p_{i})$} ;
\draw (6,-4) to [out=0,in=180] (11,-3.5);
\draw (6,-5) to [out=0,in=180] (11,-6.8);
\draw (11,-5.8) to [out=180,in=270] (9,-5);
\draw (9,-5) to [out=90,in=180] (11,-4.5);
\draw (4.5,-1.5) node{\Large \textbf{(a)}} ;
\draw (11,-5.8)--(6,-5.8) ;
\draw (11,-6.8)--(6,-6.8) ;
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-4.5) node{$(\delta_{2},1)$} ;
\draw (4,-5.2) node{$A(\delta_{2})\approx p_{i+1}R$} ;
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{2},1)}}$};
\draw (11,-7.5) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{1},1)}}$};
\draw (8,-4.7) node{$u_{1}$};
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (14.2,-3) node{$A(\eta)\approx \frac{1}{2}q_{i+1}R$ or $ \frac{1}{2}(q_{i+1}-q_{i})R$};
\draw (8.5,-3.2) node{...};
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4.4) node{$(\delta_{1},1)$ } ;
\draw (12.7,-5) node{ $A(\delta_{1})\approx q_{i+1}R$ } ;
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.5,-6.3) node{$(\gamma,M-q_{i})$} ;
\draw (5,-6.3) node{$(\gamma,M)$} ;
\draw (11,-5) to [out=180,in=0] (6,-6.8);
\draw (6,-5.8) to [out=0,in=270] (8,-4.8);
\draw (8,-4.8) to [out=90,in=0] (6,-4.3);
\draw (11,-4) to [out=180,in=0] (6,-3.3);
\draw (11,-5.8)--(6,-5.8) ;
\draw (11,-6.8)--(6,-6.8) ;
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-4) node{$(\delta_{2},1)$} ;
\draw (4,-4.7) node{$A(\delta_{2})\approx (q_{i+1}-q_{i})R$} ;
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{2},1)}}$};
\draw (11,-7.5) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}}$};
\draw (9,-4.7) node{$u_{1}$};
\draw (4.5,-1.5) node{\Large \textbf{(a')}} ;
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (14,-3) node{$A(\eta)\approx \frac{1}{2}p_{i}R$ or $ \frac{1}{4}p_{i}R$};
\draw (8.5,-3.2) node{...};
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5)--(11,-4.5);
\draw (6,-5.5)--(11,-5.5);
\draw (12,-5) node{$(\gamma,M)$ } ;
\draw (4.8,-5) node{$(\gamma,M-p_{i})$} ;
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (8,-5);
\draw (8,-5) to [out=90,in=0] (6,-4.3);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-4) node{$(\delta_{2},1)$} ;
\draw (4,-3.3) node{$A(\delta_{2})\approx \frac{1}{2} p_{i}R$} ;
\draw (4.8,-6.7) node{$(\delta_{2},1)$} ;
\draw (4,-6) node{$A(\delta_{1})\approx \frac{1}{2} p_{i}R$} ;
\draw[densely dotted, thick] (6,-1.5)--(6,-7.3);
\draw[densely dotted, thick] (11,-1.5)--(11,-7.3) ;
\draw (6,-7.5) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$};
\draw (11,-7.5) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$};
\draw (8.5,-4.3) node{$u_{1}$};
\draw (4.5,-1.5) node{\Large \textbf{(b)}} ;
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (14,-2.8) node{$A(\eta)\approx \frac{1}{2}q_{i}R$ or $ \frac{1}{4}q_{i}R$};
\draw (4.5,-1.5) node{\Large \textbf{(b')}} ;
\draw (8.5,-3.2) node{...};
\draw (6,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) node{$(\delta_{1},1)$ } ;
\draw (13.3,-4.5) node{ $A(\delta_{1})\approx \frac{1}{2} q_{i}R$ } ;
\draw (11,-6.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-7.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-5.3) node{$(\gamma,M-q_{i})$} ;
\draw (12,-6.8) node{$(\delta_{2},1)$} ;
\draw (13,-6.3) node{$A(\delta_{2})\approx \frac{1}{2}q_{i}R$} ;
\draw (6,-4.8) to [out=0,in=180] (11,-3.5);
\draw (6,-5.8) to [out=0,in=180] (11,-7.3);
\draw (11,-6.3) to [out=180,in=270] (9,-5.5);
\draw (9,-5.5) to [out=90,in=180] (11,-4.5);
\draw (11,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8)--(11,-4.8);
\draw (6,-5.8)--(11,-5.8);
\draw (4.8,-5.3) node{$(\gamma,M)$} ;
\draw[densely dotted, thick] (6,-1.6)--(6,-7.5);
\draw[densely dotted, thick] (11,-1.6)--(11,-7.5) ;
\draw (6,-7.8) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)}$};
\draw (11,-7.8) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}}$};
\draw (8.4,-4.5) node{$u_{1}$};
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-1.5) node{\Large \textbf{(c)}} ;
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (14,-3) node{$A(\eta)\approx \frac{1}{2}p_{i}R$ or $ \frac{1}{6}p_{i}R$};
\draw (8.5,-3.2) node{...};
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5)--(11,-4.5);
\draw (6,-5.5)--(11,-5.5);
\draw (12,-5) node{$(\gamma,M)$ } ;
\draw (4.8,-5) node{$(\gamma,M-p_{i})$} ;
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (8,-5);
\draw (8,-5) to [out=90,in=0] (6,-4.3);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-4) node{$(\delta_{2},1)$} ;
\draw (4,-3.3) node{$A(\delta_{2})\approx \frac{1}{3} p_{i}R$} ;
\draw (4.8,-6.7) node{$(\delta_{2},1)$} ;
\draw (4,-6) node{$A(\delta_{1})\approx \frac{2}{3} p_{i}R$} ;
\draw[densely dotted, thick] (6,-1.8)--(6,-7.3);
\draw[densely dotted, thick] (11,-1.8)--(11,-7.3) ;
\draw (6,-7.7) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$};
\draw (11,-7.7) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$};
\draw (8.5,-4.3) node{$u_{1}$};
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (11,-2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-2)--(6,-2) ;
\draw (11,-3)--(6,-3) ;
\draw (6,-2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (14,-2.8) node{$A(\eta)\approx \frac{1}{2}q_{i}R$ or $ \frac{1}{6}q_{i}R$};
\draw (4.5,-1.5) node{\Large \textbf{(c')}} ;
\draw (8.5,-3.2) node{...};
\draw (6,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) node{$(\delta_{1},1)$ } ;
\draw (13.3,-4.5) node{ $A(\delta_{1})\approx \frac{1}{3} q_{i}R$ } ;
\draw (11,-6.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-7.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-5.3) node{$(\gamma,M-q_{i})$} ;
\draw (12,-6.8) node{$(\delta_{2},1)$} ;
\draw (13,-6.3) node{$A(\delta_{2})\approx \frac{2}{3}q_{i}R$} ;
\draw (6,-4.8) to [out=0,in=180] (11,-3.5);
\draw (6,-5.8) to [out=0,in=180] (11,-7.3);
\draw (11,-6.3) to [out=180,in=270] (9,-5.5);
\draw (9,-5.5) to [out=90,in=180] (11,-4.5);
\draw (11,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8)--(11,-4.8);
\draw (6,-5.8)--(11,-5.8);
\draw (4.8,-5.3) node{$(\gamma,M)$} ;
\draw[densely dotted, thick] (6,-1.9)--(6,-7.5);
\draw[densely dotted, thick] (11,-1.9)--(11,-7.5) ;
\draw (6,-7.8) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)}$};
\draw (11,-7.8) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}}$};
\draw (8.4,-4.5) node{$u_{1}$};
\end{tikzpicture}
\vspace{1cm}
\subsection{Restriction of topological types of the $J$-holomorphic curves}
If $J_{0}(\alpha_{k+1},\alpha_{k})=-2+2g+k+l=2$, each topological type of $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$ is $(g,k,l)=(0,4,0)$, $(0,3,1)$, $(0,2,2)$, $(1,1,1)$ or $(1,2,0)$.
But in Proposition \ref{nagai}, only the type $(g,k,l)=(0,3,1)$ appears. So at first, we exclude the others.
As is the same with the case $J_{0}\leq 1$, we make a list of topological types of $J_{0}=2$ as follows.
\begin{itemize}
\item $(g,k,l)=(0,4,0)$
This case has four types as follows.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (7,-5.9);
\draw (7,-5.9) to [out=90,in=0] (6,-5.7);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-4.7) to [out=0,in=270] (7,-4.5);
\draw (7,-4.5) to [out=90,in=0] (6,-4.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1.8)--(6,-7.3);
\draw[densely dotted, thick] (11,-1.8)--(11,-7.3) ;
\draw (6,-7.7) node{$-\infty$};
\draw (11,-7.7) node{$+\infty$};
\draw (8.5,-5.3) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-2)--(6,-2) ;
\draw (11,-3)--(6,-3) ;
\draw (6,-2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-6.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-7.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8) to [out=0,in=180] (11,-3.5);
\draw (6,-5.8) to [out=0,in=180] (11,-7.3);
\draw (11,-6.3) to [out=180,in=270] (10,-6);
\draw (10,-6) to [out=90,in=180] (11,-5.8);
\draw (11,-4.5) to [out=180,in=90] (10,-4.7);
\draw (10,-4.7) to [out=270,in=180] (11,-4.8);
\draw (11,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-2)--(6,-7.5);
\draw[densely dotted, thick] (11,-2)--(11,-7.5) ;
\draw (6,-7.8) node{$-\infty$};
\draw (11,-7.8) node{$+\infty$};
\draw (8.4,-5.5) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) to [out=0,in=180] (11,-3.5);
\draw (6,-6.8) to [out=10,in=170] (11,-6.8);
\draw (11,-5.8) to [out=180,in=270] (9,-5);
\draw (9,-5) to [out=90,in=180] (11,-4.5);
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (8,-5.7) node{$u_{1}$};
\draw (6,-5.8) to [out=0,in=270] (7,-5.5);
\draw (7,-5.5) to [out=90,in=0] (6,-5);
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-8) to [out=180,in=270] (8,-5.5);
\draw (8,-5.5) to [out=90,in=180] (11,-3);
\draw (11,-4) to [out=180,in=90] (10.5,-4.2);
\draw (10.5,-4.2) to [out=270,in=180] (11,-4.4);
\draw (11,-5.4) to [out=180,in=90] (10.5,-5.6);
\draw (10.5,-5.6) to [out=270,in=180] (11,-5.8);
\draw (11,-7) to [out=180,in=270] (10.5,-6.9);
\draw (10.5,-6.9) to [out=90,in=180] (11,-6.8);
\draw[densely dotted, thick] (6,-1.6)--(6,-8.1);
\draw[densely dotted, thick] (11,-1.6)--(11,-8.1) ;
\draw (6,-8.4) node{$-\infty$};
\draw (11,-8.4) node{$+\infty$};
\draw (9,-4.9) node{$u_{1}$};
\end{tikzpicture}
\item $(g,k,l)=(0,3,1)$
This case has eight types as follows.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) to [out=0,in=180] (11,-3.5);
\draw (6,-5) to [out=0,in=180] (11,-6.8);
\draw (11,-5.8) to [out=180,in=270] (9,-5);
\draw (9,-5) to [out=90,in=180] (11,-4.5);
\draw (11,-5.8)--(6,-5.8) ;
\draw (11,-6.8)--(6,-6.8) ;
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (8,-4.7) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5) to [out=180,in=0] (6,-6.8);
\draw (6,-5.8) to [out=0,in=270] (8,-4.8);
\draw (8,-4.8) to [out=90,in=0] (6,-4.3);
\draw (11,-4) to [out=180,in=0] (6,-3.3);
\draw (11,-5.8)--(6,-5.8) ;
\draw (11,-6.8)--(6,-6.8) ;
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (9,-4.7) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5)--(11,-4.5);
\draw (6,-5.5)--(11,-5.5);
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (8,-5);
\draw (8,-5) to [out=90,in=0] (6,-4.3);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1.8)--(6,-7.3);
\draw[densely dotted, thick] (11,-1.8)--(11,-7.3) ;
\draw (6,-7.7) node{$-\infty$};
\draw (11,-7.7) node{$+\infty$};
\draw (8.5,-4.3) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-2)--(6,-2) ;
\draw (11,-3)--(6,-3) ;
\draw (6,-2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-6.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-7.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8) to [out=0,in=180] (11,-3.5);
\draw (6,-5.8) to [out=0,in=180] (11,-7.3);
\draw (11,-6.3) to [out=180,in=270] (9,-5.5);
\draw (9,-5.5) to [out=90,in=180] (11,-4.5);
\draw (11,-4.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.8)--(11,-4.8);
\draw (6,-5.8)--(11,-5.8);
\draw[densely dotted, thick] (6,-1.9)--(6,-7.5);
\draw[densely dotted, thick] (11,-1.9)--(11,-7.5) ;
\draw (6,-7.8) node{$-\infty$};
\draw (11,-7.8) node{$+\infty$};
\draw (8.4,-4.5) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (2,-3)--(7,-3);
\draw (2,-4)--(7,-4);
\draw (2,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-4)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-5.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-6.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-6) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-5.5) to [out=270,in=180] (7,-5);
\draw (2,-6.5) to [out=0,in=180] (7,-6);
\draw (5,-4.8) to [out=0,in=180] (7,-5);
\draw (2,-5.5) to [out=0,in=180] (5,-4.8);
\draw (4.5,-5.5) to [out=90,in=150] (5.5,-5.5);
\draw [dashed] (5.5,-5.5) to [out=330,in=180] (7,-6);
\draw[densely dotted, thick] (2,-2.7)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.7)--(7,-7.3) ;
\draw (3.8,-5.7) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (2,-3)--(7,-3);
\draw (2,-4)--(7,-4);
\draw (2,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-4)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-5.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-6.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-6) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6.5) to [out=330,in=200] (7,-6);
\draw (4.8,-4.8) to [out=0,in=200] (7,-5);
\draw (2,-5.5) to [out=0,in=180] (4.8,-4.8);
\draw (2,-5.5) to [out=330,in=270] (4.5,-5.8);
\draw (4.5,-5.8) to [out=90,in=40] (3,-6);
\draw[dashed] (3,-6) to [out=220,in=30] (2,-6.5);
\draw[densely dotted, thick] (2,-2.7)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.7)--(7,-7.3) ;
\draw (5.8,-5.7) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (2,-3)--(7,-3);
\draw (2,-4)--(7,-4);
\draw (2,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-4)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-6.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-7.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-6) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-5.5) to [out=270,in=180] (7,-5);
\draw (5,-6.7) to [out=90,in=180] (7,-6);
\draw (5,-6.7) to [out=270,in=180] (7,-6.8);
\draw (3,-6.5) to [out=270,in=180] (7,-7.8);
\draw (5,-4.8) to [out=0,in=180] (7,-5);
\draw (3,-6.5) to [out=90,in=180] (5,-4.8);
\draw (4.5,-5.5) to [out=90,in=150] (5.5,-5.5);
\draw [dashed] (5.5,-5.5) to [out=330,in=180] (7,-6);
\draw[densely dotted, thick] (2,-2.7)--(2,-8.1);
\draw[densely dotted, thick] (7,-2.7)--(7,-8.1) ;
\draw (4.2,-6.5) node{$u_{1}$};
\draw (7,-8.3) node{$+\infty$};
\draw (2,-8.3) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8)--(6,-5.8) ;
\draw (11,-6.8)--(6,-6.8) ;
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-6.8) to [out=180,in=270] (8,-5);
\draw (8,-5) to [out=90,in=180] (11,-3);
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (6,-7.5) node{$-\infty$};
\draw (11,-7.5) node{$+\infty$};
\draw (9,-4.9) node{$u_{1}$};
\draw (11,-4) to [out=180,in=90] (10.5,-4.2);
\draw (10.5,-4.2) to [out=270,in=180] (11,-4.4);
\draw (11,-5.4) to [out=180,in=90] (10.5,-5.6);
\draw (10.5,-5.6) to [out=270,in=180] (11,-5.8);
\end{tikzpicture}
\item $(g,k,l)=(0,2,2)$
This case has four types as follows. Note that under the assumption that there is only one simple elliptic orbit, the first and the fourth cases can not occur.
\begin{tikzpicture}
\draw (2,-2.5)--(7,-2.5);
\draw (2,-3.5)--(7,-3.5);
\draw (2,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-4)--(7,-4);
\draw (2,-5)--(7,-5);
\draw (2,-7)--(7,-7);
\draw (2,-6)--(7,-6);
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-6) to [out=0,in=180] (7,-4);
\draw (2,-7) to [out=0,in=180] (7,-5);
\draw[densely dotted, thick] (2,-2.2)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.2)--(7,-7.3) ;
\draw (4.5,-5.5) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (2,-2.5)--(7,-2.5);
\draw (2,-3.5)--(7,-3.5);
\draw (2,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-3.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-2.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-3.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-4)--(7,-4);
\draw (2,-5)--(7,-5);
\draw (2,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-5.5) to [out=0,in=180] (7,-4);
\draw (4.5,-6.5) to [out=0,in=180] (7,-5);
\draw (4.5,-5.5) to [out=180,in=0] (2,-4);
\draw (4.5,-6.5) to [out=180,in=0] (2,-5);
\draw[densely dotted, thick] (2,-2.2)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.2)--(7,-7.3) ;
\draw (4.5,-6) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (2,-3)--(7,-3);
\draw (2,-4)--(7,-4);
\draw (2,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-4)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (2,-5)--(7,-5);
\draw (2,-6)--(7,-6);
\draw (2,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (2,-6)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (7,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (7,-6) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.5,-5.5) to [out=270,in=180] (7,-5);
\draw (3.5,-5.5) to [out=270,in=180] (7,-6);
\draw (3.5,-5.5) to [out=90,in=180] (7,-5);
\draw (4.5,-5.5) to [out=90,in=150] (5.5,-5.5);
\draw [dashed] (5.5,-5.5) to [out=330,in=180] (7,-6);
\draw[densely dotted, thick] (2,-2.7)--(2,-7.3);
\draw[densely dotted, thick] (7,-2.7)--(7,-7.3) ;
\draw (4,-5.5) node{$u_{1}$};
\draw (7,-7.5) node{$+\infty$};
\draw (2,-7.5) node{$-\infty$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (19,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-1.8)--(14,-1.8) ;
\draw (19,-2.8)--(14,-2.8) ;
\draw (14,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (14,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (19,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (15.5,-5) to [out=90,in=180] (19,-3.5);
\draw (15.5,-5) to [out=270,in=180] (19,-6.8);
\draw (19,-5.8) to [out=180,in=270] (17,-5);
\draw (17,-5) to [out=90,in=180] (19,-4.5);
\draw[densely dotted, thick] (14,-1.6)--(14,-7);
\draw[densely dotted, thick] (19,-1.6)--(19,-7) ;
\draw (14,-7.5) node{$-\infty$};
\draw (19,-7.5) node{$+\infty$};
\draw (16.5,-4.9) node{$u_{1}$};
\draw (19,-3.5)--(14,-3.5) ;
\draw (19,-4.5)--(14,-4.5) ;
\draw (14,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (14,-4.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (19,-5.8)--(14,-5.8) ;
\draw (19,-6.8)--(14,-6.8) ;
\draw (14,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (14,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\end{tikzpicture}
\item $(g,k,l)=(1,1,1)$
This case has only one type as follows.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.4) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.4) to [out=180,in=0] (8.5,-3.5);
\draw (11,-5.4) to [out=180,in=0] (8.5,-6.3);
\draw (8.5,-3.5) to [out=180,in=90] (7,-4.7);
\draw (7,-4.7) to [out=270,in=180] (8.5,-6.3);
\draw (11,-4.4)--(6,-4.4) ;
\draw (11,-5.4)--(6,-5.4) ;
\draw (6,-4.4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.4)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (8,-4.8) to [out=30,in=150] (9.5,-4.8);
\draw (8.2,-4.7) to [out=350,in=200] (9.3,-4.7);
\draw (6,-7.3) node{$-\infty$};
\draw (11,-7.3) node{$+\infty$};
\draw (8.4,-5.8) node{$u_{1}$};
\end{tikzpicture}
\item $(g,k,l)=(1,2,0)$
This case has two types as follows.
\begin{tikzpicture}
\draw (13.5,6.2)--(8.5,6.2) ;
\draw (13.5,5.2)--(8.5,5.2) ;
\draw (8.5,6.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (8.5,5.2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,6.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,5.2) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (8.5,3.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (8.5,2.7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,3.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,2.2) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (8.5,3.7) to [out=0,in=180] (11,4.7);
\draw (11,4.7) to [out=0,in=180] (13.5,3.2);
\draw (8.5,2.7) to [out=0,in=180] (11,1.2);
\draw (11,1.2) to [out=0,in=180] (13.5,2.2);
\draw (10,3) to [out=30,in=150] (12,3);
\draw (10.3,3.1) to [out=-30,in=210] (11.7,3.1);
\draw[densely dotted, thick] (8.5,6.5)--(8.5,1);
\draw[densely dotted, thick] (13.5,6.5)--(13.5,1);
\draw (8.5,0.7) node{$-\infty$};
\draw (13.5,0.7) node{$+\infty$};
\draw (11,2) node{$u_{1}$};
\end{tikzpicture}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.8) to [out=180,in=0] (8.5,-3.5);
\draw (11,-6.3) to [out=180,in=0] (8.5,-6.3);
\draw (8.5,-3.5) to [out=180,in=90] (7,-4.7);
\draw (7,-4.7) to [out=270,in=180] (8.5,-6.3);
\draw (11,-4.8) to [out=180,in=90] (10,-5);
\draw (10,-5) to [out=270,in=180] (11,-5.3);
\draw[densely dotted, thick] (6,-1.6)--(6,-7);
\draw[densely dotted, thick] (11,-1.6)--(11,-7) ;
\draw (7.7,-4.8) to [out=30,in=150] (9.2,-4.8);
\draw (7.9,-4.7) to [out=350,in=200] (9,-4.7);
\draw (6,-7.3) node{$-\infty$};
\draw (11,-7.3) node{$+\infty$};
\draw (8.4,-5.8) node{$u_{1}$};
\end{tikzpicture}
\end{itemize}
\begin{lem}\label{toptyp}
Suppose that $\alpha_{k+1}$ and $\alpha_{k}$ satisfy the assumptions 1, 2, 3, 4 and 5 in Proposition \ref{nagai}.
Let $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha_{k+1},\alpha_{k})$ be
any $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$. Then $u$ is $(g,k,l)=(0,3,1)$.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{toptyp}]
In the cases $(g,k,l)=(0,2,2)$, $(1,1,1)$, we can see from the topological types that $A(\alpha_{k+1})-A(\alpha_{k})$ have to be larger than some action of orbit. But this contradicts $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$.
From now on, we consider $(g,k,l)=(0,4,0)$, $(1,2,0)$. As a matter of fact, we can easily exclude these cases in almost the same way as Lemma \ref{mainlemma}. But to make sure, we explain how to do in detail.
Since $l=0$ and $E(\alpha_{k+1})$, $E(\alpha_{k})\notin S_{-\theta}\cup{S_{\theta}}$, $u_{1}$ has no end asymptotic to $\gamma$ and since $E(\alpha_{k+1})$, $E(\alpha_{k})>p_{1}$, $q_{1}>0$, $u_{0}$ has some covering of $\mathbb{R}\times \gamma$. Let $z_{i}\to \gamma$. Then, we obtain a sequence of $J$-holomorphic curves $u_{1}^{i}$ which are through $z_{i}$ and either $(g,k)=(0,4)$ or $(1,2)$. Note that their topological types and orbit where their ends are asymptotic may change in the sequence.
At first, suppose that the sequence contains infinity many $J$-holomorphic curves whose topological types are $(g,k)=(0,4)$. By the compactness argument, there is an $J$- holomoprhic curve $u_{1}^{\infty}$ which may be splitting into some floors. By its topological type and properties of ECH and Fredholm indexes, we can find that the number of floors are at most two. If $u_{1}^{\infty}$ does not split, $\mathbb{R}\times \gamma \cap{u_{1}^{\infty}}\neq \emptyset$ and so $u_{0}\cap{u_{1}^{\infty}}\neq \emptyset$. This contradict $I(u_{0}\cup{u_{1}^{\infty}})=2$ and its admissibility. So we may assume that $u_{1}^{\infty}$ has two floors and write $u_{1}^{\infty}=(u_{-}^{\infty},u_{+}^{\infty})$ up to $\mathbb{
R}$-action. By the additivity of ECH index and Fredholm index, we have $I(u_{0}\cup{u_{-}^{\infty}})=I(u_{0}\cup{u_{+}^{\infty}})=1$ and each non trivial part of $u_{\pm}^{\infty}$ are connected. Moreover, by the assumption of their topological type, one of the non trivial parts of $u_{\pm}^{\infty}$ is of genus 0 with one end of it asymptotic to $\gamma$. This indicates that the middle orbit set, that is, the orbit set consisting of orbits where positive ends of $u_{0}\cup{u_{-}^{\infty}}$ are asymptotic is admissible. This contradicts (\ref{mod2}) and $I(u_{0}\cup{u_{-}^{\infty}})=I(u_{0}\cup{u_{+}^{\infty}})=1$.
Next, suppose that the sequence contains infinity many $J$-holomorphic curves whose topological types are $(g,k)=(1,2)$. In the same way, we have a splitting curve. Let $u_{\pm}^{\infty}$ be the curves in top and bottom floors of the splitting curve respectively. Then $I(u_{0}\cup{u_{-}^{\infty}})=I(u_{0}\cup{u_{+}^{\infty}})=1$.
By geometric observation, $u_{+}^{\infty}$ has one or two negative ends and moreover at least one of them is elliptic. Also $u_{-}^{\infty}$ has one or two positive ends and at least one of them is elliptic. If all negative ends of $u_{-}^{\infty}$ are elliptic, this contradicts (\ref{mod2}). Also if all positive ends of $u_{+}^{\infty}$ are elliptic, this contradicts (\ref{mod2}).
This means that the number of negative ends of $u_{+}^{\infty}$ and that one of positive ends of $u_{-}^{\infty}$ are both two and only one of them is asymptotic to the elliptic orbit respectively. Moreover, their multiplicities are the same (see the below figure).
Let $E(\alpha_{k+1})=E(\alpha_{k})=M$. Then the total multiplicity of the middle orbit set is $M+1$ and by the partition condition, we have $1=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M+1\}})=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M+1\}})$. But this is a contradiction because of the assumption $M>q_{1},\,\,p_{1}$.
Combining the above argument, we complete the proof of Lemma \ref{toptyp}.
\begin{tikzpicture}
\draw (16,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-1.8)--(11,-1.8) ;
\draw (16,-2.8)--(11,-2.8) ;
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (17.2,-2.3) node{$(\gamma,M)$};
\draw (16,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-5.5) to [out=180,in=0] (11,-7.1);
\draw (11,-6.1) to [out=0,in=270] (13,-4.5);
\draw (13,-4.5) to [out=90,in=0] (11,-2.8);
\draw (16,-4.5) to [out=180,in=0] (11,-1.8);
\draw (11,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (11,-1.5)--(11,-7.3);
\draw[densely dotted, thick] (16,-1.5)--(16,-7.3) ;
\draw (16,-7.7) node{$\alpha_{k+1}$};
\draw (11,0)node{$\Downarrow$};
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.8) to [out=0,in=180] (11,-1.8);
\draw (6,-4.8) to [out=0,in=180] (11,-7.1);
\draw (11,-6.1) to [out=180,in=270] (9,-5);
\draw (9,-5) to [out=90,in=180] (11,-2.8);
\draw (6,-6.1)[dashed] arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-6.1)[dashed] arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw [dashed] (6,-6.1)--(16,-6.1);
\draw [dashed] (6,-7.1)--(16,-7.1) ;
\draw[densely dotted, thick] (6,-1.5)--(6,-7.5);
\draw (6,-8) node{$\alpha_{k}$};
\draw (13.5,6.2)--(8.5,6.2) ;
\draw (13.5,5.2)--(8.5,5.2) ;
\draw (8.5,6.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (8.5,5.2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,6.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,5.2) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (8.5,3.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (8.5,2.7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (13.5,3.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (13.5,2.2) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (8.5,3.7) to [out=0,in=180] (11,4.7);
\draw (11,4.7) to [out=0,in=180] (13.5,3.2);
\draw (8.5,2.7) to [out=0,in=180] (11,1.2);
\draw (11,1.2) to [out=0,in=180] (13.5,2.2);
\draw (10,3) to [out=30,in=150] (12,3);
\draw (10.3,3.1) to [out=-30,in=210] (11.7,3.1);
\draw[densely dotted, thick] (8.5,6.5)--(8.5,0.5);
\draw[densely dotted, thick] (13.5,6.5)--(13.5,0.5);
\draw (8.5,0) node{$\alpha_{k}$};
\draw (13.5,0) node{$\alpha_{k+1}$};
\end{tikzpicture}
\end{proof}
More precisely, we have the next lemma.
\begin{lem}\label{top}
Suppose that $\alpha_{k+1}$ and $\alpha_{k}$ satisfy the assumptions 1, 2, 3, 4 and 5 in Proposition \ref{nagai}.
Let $u=u_{0}\cup{u_{1}}\in \mathcal{M}^{J}(\alpha_{k+1},\alpha_{k})$ be
any $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$. Then $\alpha_{k+1}$, $\alpha_{k}$ and $u$ hold one of the following conditions.
\item[\bf (A).] Let $E(\alpha_{k+1})=M$ and $p_{i}:=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{1},1)}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{2},1)}}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1)\cup{(\gamma,p_{i})},(\delta_{2},1))$.
\item[\bf (A').]Let $E(\alpha_{k})=M$ and $q_{i}:=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Then there are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{2},1)}}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1),(\delta_{2},1)\cup{(\gamma,q_{i})})$.
\item[\bf (B).]Let $E(\alpha_{k+1})=M$ and $p_{i}:=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. There are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$ and $u_{1}\in \mathcal{M}^{J}((\gamma,p_{i}),(\delta_{1},1)\cup{(\delta_{2},1)})$.
\item[\bf (B').]Let $E(\alpha_{k})=M$ and $q_{i}:=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. There are two negative hyperbolic orbits $\delta_{1}$, $\delta_{2}$ and an admissible orbit set $\hat{\alpha}$ consisting of negative hyperbolic orbits such that $\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M-q_{i})\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}}$, $\alpha_{k}=\hat{\alpha}\cup{(\gamma,M)}$ and $u_{1}\in \mathcal{M}^{J}((\delta_{1},1)\cup{(\delta_{2},1)},(\gamma,q_{i}))$.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{top}]
Since $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$, $u_{1}$ has at least one negative end. Moreover, at least one end of $u_{1}$ have to be asymptotic to some negative hyperbolic orbit because the fact causes a contradiction that if all ends are asymptotic to $\gamma$, the value $A(\alpha_{k+1})-A(\alpha_{k})$ have to be lager than or equal to $A(\gamma)$. From the assumptions $E(\alpha_{k+1})$, $E(\alpha_{k})>p_{1}$, $q_{1}>0$ and the partition conditions of the ends, we have Lemma \ref{top}.
\end{proof}
\subsection{Restriction of $J_{0}$ combinations}
In the previous subsection, we decided the topological type of $J$-holomorphic curves counted by $U\langle \alpha_{k+1} \rangle =\langle \alpha_{k} \rangle$. To prove Proposition \ref{nagai}, we have to decide the approximate relations in the actions of the orbits in $\alpha_{k+1}$ and $\alpha_{k}$.
From now on, because symmetry allows the same argument, , we only consider the cases (A) and (B).
The next claim is almost the same as Claim \ref{index2to4}.
\begin{cla}\label{index22}
In the case of (A) in Lemma \ref{top},
\item[\bf (A).]
For any $p_{i}\leq N<p_{i+1}$ ,
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,N)\cup{(\delta_{1},1)}},\hat{\alpha}\cup{(\gamma,N-p_{i})}\cup{(\delta_{2},1)})=2.
\end{equation}
Moreover
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,p_{i+1})}\cup{(\delta_{1},1)},\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)})=4.
\end{equation}
And for any $p_{i}<N\leq p_{i+1}$,
\begin{equation}
J_{0}(\hat{\alpha}\cup{(\gamma,N)}\cup{(\delta_{1},1)},\hat{\alpha}\cup{(\gamma,N-p_{i})}\cup{(\delta_{2},1)})=1.
\end{equation}
In the case of (B) in Lemma \ref{top},
\item[\bf (B).]
For any $p_{i}\leq N<p_{i+1}$ ,
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,N)},\hat{\alpha}\cup{(\gamma,N-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)})=2.
\end{equation}
Moreover
\begin{equation}
I(\hat{\alpha}\cup{(\gamma,p_{i+1})},\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)})=4.
\end{equation}
And for any $p_{i}<N\leq p_{i+1}$,
\begin{equation}
J_{0}(\hat{\alpha}\cup{(\gamma,N)},\hat{\alpha}\cup{(\gamma,N-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)})=1.
\end{equation}
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{index22}]
We can prove this in the same way as Claim \ref{mi1} and Claim \ref{index2to4}.
\end{proof}
\begin{cla}\label{basis}
In the case of (A) in Lemma \ref{top},
\item[\bf (A).]
There is an admissible orbit set $\zeta$ with $U\langle \hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle= \langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)} \rangle$. Moreover, $E(\zeta)=0$.
In the case of (B) in Lemma \ref{top},
\item[\bf (B).]
There is an admissible orbit set $\zeta$ with $U\langle \hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle= \langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{{(\delta_{1},1)}}\cup{(\delta_{2},1)} \rangle$. Moreover, $E(\zeta)=0$.
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{basis}]
In the same way as Claim \ref{el0} and just before that.
\end{proof}
Since $J_{0}\geq-1$, there are five possibilities,
$(J_{0}(\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}))(\mathrm{resp.}\,\,(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}))) =(3,-1)$, $(2,0)$, $(1,1)$, $(0,2)$ or $(-1,3)$. But except for $(1,1)$, the behaviors of $J$-holomorphic curves counted by the $U$-map cause contradictions.
Now, we use the rest of Section
6 to prove the next lemma.
\begin{lem}\label{exc}
Under the notation in Claim \ref{basis}, $J$-holomorphic curves counted by the $U$-map cause contradictions except for $(J_{0}(\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}))=(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}))=(1,1)$.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{exc}]
At first, we can easily exclude the cases $(J_{0}(\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}))=(3,-1)$, $(-1,3)$ because of their smallness of the difference of their actions.
Next, we will consider the cases (A), (B) respectively.
\item[\bf Case] (A).
At first, we consider the splitting behaviors of $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$ as $z \to \eta$ for some fixed $\eta\in \hat{\alpha}$. Then there are three possibilities of splitting of holomorphic curves and also we have three possibilities of approximate relations as follows.
\item[$(\mathfrak{a}_{1})$.] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{a}_{2})$.] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{a}_{3})$.] $|p_{i}R-2A(\eta)|<\epsilon$.
Moreover, we always have
\item[$(\clubsuit)$]$A(\alpha_{k+1})-A(\alpha_{k})=|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (8.5,-3.2) node{...};
\draw (6,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-4) node{$(\delta_{1},1)$ } ;
\draw (11,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12,-6.3) node{$(\gamma,M)$} ;
\draw (4.8,-6.3) node{$(\gamma,M-p_{i})$} ;
\draw (6,-4) to [out=0,in=180] (11,-3.5);
\draw (6,-5) to [out=0,in=180] (11,-6.8);
\draw (11,-5.8) to [out=180,in=270] (9,-5);
\draw (9,-5) to [out=90,in=180] (11,-4.5);
\draw (11,-5.8)--(6,-5.8) ;
\draw (11,-6.8)--(6,-6.8) ;
\draw (6,-5.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-4.5) node{$(\delta_{2},1)$} ;
\draw[densely dotted, thick] (6,-1.5)--(6,-7);
\draw[densely dotted, thick] (11,-1.5)--(11,-7) ;
\draw (6,-7.5) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{2},1)}}$};
\draw (11,-7.5) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)\cup{(\delta_{1},1)}}$};
\draw (8,-4.7) node{$u_{1}$};
\end{tikzpicture}
\item [$\rm(\hspace{.18em}i\hspace{.18em})$.] If $(J_{0}(\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}))=(2,0)$
Note that genus of each $J$-holomorphic curve counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ are 0. Moreover, each curve has both negative end covering at $\gamma$ with multiplicity $p_{i+1}-p_{i}$ and positive end covering at negative hyperbolic orbit $(\delta',1)$ which is not equivalent to $\delta_{2}$ because $A(\zeta)-A(\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)})<\epsilon$ and so in $\mathcal{M}^{J}((\delta',1),(\gamma,p_{i+1}-p_{i}))$.
Then from the splitting behavior as $z \to \delta_{2}$, we have $|2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$ and also as $z\to \eta\in \hat{\alpha}$, we have $|2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$. These two relations indicate that $|A(\eta)-A(\delta_{2})|<\epsilon$. Since $|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $A(\delta_{2})>p_{i}R$, $ A(\delta_{1})$ and hence $A(\eta)>p_{i}R$, $ A(\delta_{1})$. These relations contradict $(\mathfrak{a}_{1})$, $(\mathfrak{a}_{2})$ and $(\mathfrak{a}_{3})$ in any case. Therefore, this case can not occur.
\begin{tikzpicture}
\draw (6,-2.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-3.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-2.5)--(6,-7.5);
\draw (1,-2.8)--(6,-2.8) ;
\draw (1,-3.8)--(6,-3.8) ;
\draw (1,-4.3)--(6,-4.3) ;
\draw (1,-5.3)--(6,-5.3) ;
\draw (1,-2.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-3.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-4.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-5.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-6.1) to [out=0,in=180] (3,-5.5);
\draw (1,-7.1) to [out=0,in=180] (3,-6.5);
\draw (3,-5.5) to [out=0,in=180] (6,-6.1);
\draw (3,-6.5) to [out=0,in=180] (6,-7.1);
\draw[densely dotted, thick] (1,-2.5)--(1,-7.5) ;
\draw (1,-8) node{$\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}$};
\draw (6,-8) node{$\zeta=\hat{\alpha}\cup{(\delta_{2},1)}\cup{(\delta',1)}$};
\draw (-0.5,-6.7) node{$(\gamma,p_{i+1}-p_{i})$};
\draw (7,-4.9) node{$(\delta_{2},1)$};
\draw (7.3,-3.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (7,-6.6) node{$(\delta',1)$};
\end{tikzpicture}
\item[$\rm(\hspace{.08em}ii\hspace{.08em})$.] If $(J_{0}(\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}))=(0,2)$
In the same way as above, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ are of genus 0 and have both negative end covering at $\gamma$ with multiplicity $p_{i+1}$ and positive end covering one negative hyperbolic orbit $(\delta',1)$ which is not equivalent to $\delta_{1}$ and so in $\mathcal{M}^{J}((\gamma,p_{i+1}),(\delta',1))$.
Then from the splitting behavior as $z \to \delta_{1}$, we have $|2A(\delta_{1})-p_{i+1}R|<\epsilon$ and also as $z\to \eta$, we have $|2A(\eta)-p_{i+1}R|<\epsilon$. Since $(\clubsuit)$.$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $|(\frac{1}{2}p_{i+1}+p_{i})R-A(\delta_{2})|<\frac{3}{2}\epsilon$.
These relations contradict $(\mathfrak{a}_{1})$, $(\mathfrak{a}_{2})$ and $(\mathfrak{a}_{3})$ in any case. Here, we use Claim \ref{fre} implicitly.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-6)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(16,-1.8) ;
\draw (11,-2.8)--(16,-2.8) ;
\draw (11,-3.5)--(16,-3.5) ;
\draw (11,-4.5)--(16,-4.5) ;
\draw (16,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-5.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-6.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (11,-1.5)--(11,-7) ;
\draw[densely dotted, thick] (16,-1.5)--(16,-7) ;
\draw (16,-7.5) node{$\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})\cup{(\delta_{1},1)}}$};
\draw (11,-7.5) node{$\zeta=\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\delta',1)}$};
\draw (11,-5) to [out=0,in=180] (13.5,-5.5);
\draw (11,-6) to [out=0,in=180] (13.5,-6.5);
\draw (13.5,-5.5) to [out=0,in=180] (16,-5.3);
\draw (13.5,-6.5) to [out=0,in=180] (16,-6.3);
\draw (17,-4) node{$(\delta_{1},1)$};
\draw (17,-5.8) node{$(\gamma,p_{i+1})$};
\draw (10,-5.6) node{$(\delta',1)$};
\draw (17.3,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\end{tikzpicture}
By the arguments so far, we can see that only the case $(J_{0}(\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}))=(1,1)$ may occur.
\item[\bf Case (B).]
At first, we consider the splitting behaviors of $J$-holomorphic curves counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$ as $z \to \eta$ for some $\eta\in \hat{\alpha}$. Then there are three possibilities of splitting of holomorphic curve and also we have three possibilities of approximate relations as follows.
\item[($\mathfrak{b}_{1})$.] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{b}_{2})$.] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{b}_{3})$.] $|p_{i}R-2A(\eta)|<\epsilon$.
Moreover, we always have
\item[$(\spadesuit)$.]$A(\alpha_{k+1})-A(\alpha_{k})=|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (12.2,-2.3) node{$(\eta,1)\in \hat{\alpha}$};
\draw (8.5,-3.2) node{...};
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5)--(11,-4.5);
\draw (6,-5.5)--(11,-5.5);
\draw (12,-5) node{$(\gamma,M)$ } ;
\draw (4.8,-5) node{$(\gamma,M-p_{i})$} ;
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (8,-5);
\draw (8,-5) to [out=90,in=0] (6,-4.3);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-4) node{$(\delta_{2},1)$} ;
\draw (4.8,-6.7) node{$(\delta_{1},1)$} ;
\draw[densely dotted, thick] (6,-1.5)--(6,-7.3);
\draw[densely dotted, thick] (11,-1.5)--(11,-7.3) ;
\draw (6,-7.7) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,M-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$};
\draw (11,-7.7) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,M)}$};
\draw (8.5,-4.3) node{$u_{1}$};
\end{tikzpicture}
\item[$\rm(\hspace{.18em}i\hspace{.18em})$.] If $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}))=(0,2)$
In the same as before, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ are of genus 0 and in $\mathcal{M}^{J}((\gamma,p_{i+1}),(\delta',1))$ for some negative hyperbolic orbit $\delta'$ with $\zeta = \hat{\alpha}\cup{(\delta',1)}$.
Then from the splitting behavior as $z \to \eta \in \hat{\alpha}$, we have $|2A(\eta)-p_{i+1}R|<\epsilon$. Since ($\spadesuit$). $|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$, we have $p_{i}R>A(\delta_{1})$, $A(\delta_{2})$. So we can easily see that the relations contradict $(\mathfrak{b}_{1})$, $(\mathfrak{b}_{2})$ and $(\mathfrak{b}_{3})$ in any case. Therefore this case can not occur.
\begin{tikzpicture}
\draw (11,-1.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.3)--(16,-1.3) ;
\draw (11,-2.3)--(16,-2.3) ;
\draw (16,-1.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-2.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-4.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-4.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-3.8) to [out=0,in=180] (16,-3.3);
\draw (11,-4.8) to [out=0,in=180] (16,-4.3);
\draw[densely dotted, thick] (11,-1)--(11,-5.5) ;
\draw[densely dotted, thick] (16,-1)--(16,-5.5) ;
\draw (11,-6) node{$\zeta=\hat{\alpha}\cup{(\delta',1)}$};
\draw (16,-6) node{$\hat{\alpha}\cup{(\gamma,p_{i+1})}$};
\draw (17,-3.8) node{$(\gamma,p_{i+1})$};
\draw (17.3,-1.8) node{$(\eta,1)\in \hat{\alpha} $};
\end{tikzpicture}
\item[$\rm(\hspace{.08em}ii\hspace{.08em})$.] If $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}))=(2,0)$
In the same as before, $J$-holomorphic curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}) \rangle$ are of genus 0 and in $\mathcal{M}^{J}((\delta',1),(\gamma,p_{i+1}-p_{i}))$ for some negative hyperbolic orbit $\delta'$ with $\zeta = \hat{\alpha}\cup{(\delta',1)}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}$.
Then from the splitting behavior as $z \to \delta_{1}$, $\delta_{2}$, we have $|2A(\delta_{1})-(p_{i+1}-p_{i})R|<\epsilon$ and $|2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$. From these relations, we also have $|(A(\delta_{1})+A(\delta_{2}))-(p_{i+1}-p_{i})R|<\epsilon$. By combining with ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$, we have $|p_{i}R-(p_{i+1}-p_{i})R|<\epsilon$ and so $p_{i}=p_{i+1}-p_{i}$. This contradicts Claim \ref{fre}.
\begin{tikzpicture}
\draw (1,-1.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-2.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-1.3)--(6,-1.3) ;
\draw (1,-2.3)--(6,-2.3) ;
\draw (6,-1.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-2.6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-3.6) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-2.6)--(6,-2.6) ;
\draw (1,-3.6)--(6,-3.6) ;
\draw (1,-6.1)--(6,-6.1) ;
\draw (1,-7.1)--(6,-7.1) ;
\draw (1,-4.5) to [out=0,in=180] (3,-4.1);
\draw (3,-4.1) to [out=0,in=180] (6,-4.5);
\draw (1,-5.5) to [out=0,in=180] (3,-5.1);
\draw (3,-5.1) to [out=0,in=180] (6,-5.5);
\draw (1,-2.6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-3.6)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (1,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (1,-7.1) [dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw[densely dotted, thick] (6,-1)--(6,-7.5);
\draw (0.7,-8) node{$\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}$};
\draw (6.3,-8) node{$\zeta =\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}\cup{(\delta',1)}$};
\draw (0,-3.2) node{$(\delta_{1},1)$};
\draw (0,-6.7) node{$(\delta_{2},1)$};
\draw (-0.5,-5) node{$(\gamma,p_{i+1}-p_{i})$};
\draw (7,-5) node{$(\delta',1)$};
\draw (7.2,-1.8) node{$(\eta,1) \in \hat{\alpha}$};
\draw[densely dotted, thick] (1,-1)--(1,-7.5) ;
Summarizing the above arguments, we complete the proof of Lemma \ref{exc}.
\end{tikzpicture}
By the arguments so far, we can see that only the case $(J_{0}(\hat{\alpha}\cup{(\gamma,p_{i+1})},\zeta),J_{0}(\zeta,\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}))=(1,1)$ may occur.
\end{proof}
\section{Calculations of the approximate values of the actions of the orbits}
In this section, we compute the approximate values of the actions of the orbits and complete the proof of Proposition \ref{nagai} under the result obtained so far.
In the same way as before, the splitting behaviors of $J$-holomorphic curves counted by the $U$-map play an important rule.
In this section, we consider $(A)$ and $(B)$ in Lemma \ref{top} respectively.
\subsection{Type (A)}
Since $E(\zeta)=0$, $p_{i+1}\in S_{-\theta}$ and $p_{i+1}-p_{i}\in S_{\theta}$, the topological types of $J$-holomoprhic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ are both $(g,k,l)=(0,3,0)$.
By considering which ends of $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ correspond to ones of $J$-holomorphic curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ respectively, we can see that there are three pairs of $\zeta$, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ as follows.
\begin{enumerate}
\item [\bf Type $(A_{1})$]
There is a negative hyperbolic orbit $\delta'$ with $\zeta=\hat{\alpha}\cup{
(\delta',1)}$. Moreover any $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ are in $\mathcal{M}^{J}((\delta_{1},1)\cup{(\gamma,p_{i+1})},(\delta',1))$ and $\mathcal{M}^{J}((\delta',1),(\gamma,p_{i+1}-p_{i})\cup{(\delta_{2},1)})$ respectively.
\item [\bf Type $(A_{2})$]
There is a negative hyperbolic orbit $\delta'$ with $\zeta=\hat{\alpha}\cup{
(\delta',1)}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}$. Moreover any $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ are in $\mathcal{M}^{J}((\gamma,p_{i+1}),(\delta',1)\cup{(\delta_{2},1)})$ and $\mathcal{M}^{J}((\delta',1)\cup{(\delta_{1})},(\gamma,p_{i+1}-p_{i}))$ respectively.
\item[\bf Type $(A_{3})$]
There is a negative hyperbolic orbit $\delta'$ with $\delta'\in \hat{\alpha}$ such that any $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ are in $\mathcal{M}^{J}((\gamma,p_{i+1})\cup{(\delta',1)},(\delta_{2},1))$ and in $\mathcal{M}^{J}((\delta_{1}),(\gamma,p_{i+1}-p_{i})\cup{(\delta',1)})$ respectively.
\end{enumerate}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (8,-5);
\draw (8,-5) to [out=90,in=0] (6,-4.3);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-3.9) node{$(\delta_{2},1)$} ;
\draw (4.6,-6.7) node{$(\gamma,p_{i+1}-p_{i})$} ;
\draw[densely dotted, thick] (6,-1.5)--(6,-7.5);
\draw (6,-7.7) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})\cup{(\delta_{2},1)}}$};
\draw (16,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-1.8)--(11,-1.8) ;
\draw (16,-2.8)--(11,-2.8) ;
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (17,-2.3) node{$(\eta,1)$};
\draw (16,-3.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-4.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-6.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-7.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (17.2,-4) node{$(\delta_{1}.1)$} ;
\draw (17,-6.8) node{$(\gamma,p_{i+1})$} ;
\draw (11,-4.5) to [out=0,in=180] (16,-3.5);
\draw (11,-5.5) to [out=0,in=180] (16,-7.3);
\draw (16,-6.3) to [out=180,in=270] (14,-5.3);
\draw (14,-5.3) to [out=90,in=180] (16,-4.5);
\draw[densely dotted, thick] (11,-1.5)--(11,-7.5);
\draw[densely dotted, thick] (16,-1.5)--(16,-7.5) ;
\draw (11,-8.2) node{\Large Type $(A_{1})$} ;
\draw (16,-7.7) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,p_{i+1})\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}}$};
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (16,-0.3)--(6,-0.3) ;
\draw (16,-1.3)--(6,-1.3) ;
\draw (6,-0.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-1.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-0.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-1.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-0.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-1.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.7)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-2.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-3.2)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-2.2)--(11,-2.2) ;
\draw (16,-3.2)--(11,-3.2) ;
\draw (16,-2.2) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-3.2) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.7) to [out=180,in=0] (6,-5);
\draw (11,-2.2) to [out=180,in=0] (6,-4);
\draw (11,-4.7) to [out=180,in=270] (9,-4);
\draw (11,-3.2) to [out=180,in=90] (9,-4);
\draw (4.6,-4.5) node{$(\gamma,p_{i+1}-p_{i})$} ;
\draw (17,-2.7) node{$(\delta_{1},1)$} ;
\draw (4.9,-7.8) node{$(\delta_{2},1)$} ;
\draw (17.2,-6.5) node{$(\gamma,p_{i+1})$} ;
\draw (17,-0.8) node{$(\eta,1)$} ;
\draw[densely dotted, thick] (6,0)--(6,-8.5);
\draw[densely dotted, thick] (11,0)--(11,-8.5) ;
\draw[densely dotted, thick] (16,0)--(16,-8.5);
\draw (7.2,-8.9) node{$\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})\cup{(\delta_{2},1)}}$};
\draw (16,-8.9) node{$\hat{\alpha}\cup{(\gamma,p_{i+1})\cup{(\delta_{1},1)}}$};
\draw (16,-6) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.7) to [out=0,in=180] (16,-6);
\draw (11,-5.7) to [out=0,in=90] (13,-6.4);
\draw (13,-6.4) to [out=270,in=0] (11,-7.3);
\draw (16,-7) to [out=180,in=0] (11,-8.3);
\draw (11,-7.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-8.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-7.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-8.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-7.3)--(6,-7.3) ;
\draw (11,-8.3)--(6,-8.3) ;
\draw (11,-9.5) node{\Large Type $(A_{2})$} ;
\end{tikzpicture}
\vspace{5mm}
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-1.8)--(6,-1.8) ;
\draw (11,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5.5) to [out=180,in=0] (6,-7.1);
\draw (6,-6.1) to [out=0,in=270] (8,-5);
\draw (8,-5) to [out=90,in=0] (6,-4.3);
\draw (11,-4.5) to [out=180,in=0] (6,-3.3);
\draw (6,-6.1) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-7.1)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.8,-3.9) node{$(\delta',1)$} ;
\draw (4.6,-6.7) node{$(\gamma,p_{i+1}-p_{i})$} ;
\draw (4.8,-2.3) node{$(\delta_{2},1)$} ;
\draw[densely dotted, thick] (6,-0.5)--(6,-7.5);
\draw (6,-7.7) node{$\alpha_{k}=\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})\cup{(\delta_{2},1)}}$};
\draw (16,-4.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-5.5) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-4.5)--(11,-4.5) ;
\draw (16,-5.5)--(11,-5.5) ;
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (17,-1.3) node{$(\gamma,p_{i+1})$};
\draw (16,-0.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-1.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-4.3) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16.9,-3.8) node{$(\delta',1)$} ;
\draw (17,-5) node{$(\delta_{1},1)$} ;
\draw (11,-1.8) to [out=0,in=180] (16,-0.8);
\draw (11,-2.8) to [out=0,in=180] (16,-4.3);
\draw (16,-3.3) to [out=180,in=270] (14,-2.6);
\draw (14,-2.6) to [out=90,in=180] (16,-1.8);
\draw[densely dotted, thick] (11,-0.5)--(11,-7.5);
\draw[densely dotted, thick] (16,-0.5)--(16,-7.5) ;
\draw (16,-7.7) node{$\alpha_{k+1}=\hat{\alpha}\cup{(\gamma,p_{i+1})\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}}$};
\draw (11,-8.2) node{\Large Type $(A_{3})$} ;
\end{tikzpicture}
To prove Proposition \ref{nagai}, at first, we calculate the approximate actions of $\delta_{1}$, $\delta_{2}$ and $\eta \in \hat{\alpha}$ from $(A_{1})$ by using the splitting behavior of $J$-holomorphic curves. Next, we will show that any splitting behaviors of $(A_{2})$ and $(A_{3})$ cause contradictions.
\subsubsection{Type $(A_{1})$}
\begin{lem}\label{typa1}
Suppose that the pair of $\zeta$, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ is Type $(A_{1})$. Then, we have $A(\delta_{1})\approx (p_{i+1}-p_{i})R$, $A(\delta_{2})\approx p_{i+1}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}p_{i+1}R$ or $A(\eta)\approx \frac{1}{2}(p_{i+1}-p_{i})R$.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{typa1}]
We will consider the behavior of $J$-holomorphic curves counted by $U$-map as $z \to \eta$. Then each $J$-holomorphic curve counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ have three possibilities of splitting and we have three estimates respectively as follows.
From $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$, we have
\begin{enumerate}
\item[($\mathfrak{c}_{1}$).] $|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$
\item [($\mathfrak{c}_{2}$).] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item[($\mathfrak{c}_{3}$).]$|(p_{i+1}-p_{i})R+A(\delta_{2})-2A(\eta)|<\epsilon$
\end{enumerate}
From $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$, we have
\begin{enumerate}
\item[($\mathfrak{d}_{1}$).] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item [($\mathfrak{d}_{2}$).] $|p_{i+1}R-2A(\eta)|<\epsilon$
\item[($\mathfrak{d}_{3}$).]$|p_{i+1}R+A(\delta_{1})-2A(\eta)|<\epsilon$
\end{enumerate}
Here recall that we have from $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$,
\begin{enumerate}
\item[($\mathfrak{a}_{1}$).] $|p_{i}R-2A(\eta)|<\epsilon$
\item [($\mathfrak{a}_{2}$).] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[($\mathfrak{a}_{3}$).]$|A(\delta_{2})-2A(\eta)|<\epsilon$
\end{enumerate}
and moreover we always have
$(\clubsuit)$.$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$.
At first, we prove the next lemma.
\begin{lem}\label{restA1}
Any above pair (($\mathfrak{c}_{i_{1}}$),($\mathfrak{d}_{i_{2}}$),($\mathfrak{a}_{i_{3}}$)) causes a contradiction except for the following two cases.
$(i_{1},i_{2},i_{3})=(1,1,2),(2,2,3)$.
\end{lem}
In the next claim, we exclude pairs (($\mathfrak{c}_{i_{1}}$),($\mathfrak{d}_{i_{2}}$)) such that we can derive contradictions by only them.
\begin{cla}\label{firstA1}
Any pair (($\mathfrak{c}_{i_{1}}$),($\mathfrak{d}_{i_{2}}$)) causes a contradiction except for the following cases
$(i_{1},i_{2})=(1,1),(2,2),(3,3)$.
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{firstA1}]
We have to derive a contradiction from $(i_{1},i_{2})=(1,2)$, $(1,3)$, $(2,1)$, $(2,3)$, $(3,1)$, $(3,2)$ respectively.
\item Case $(i_{1},i_{2})=(1,2)$
By cancelling the terms $p_{i+1}R$ and $2A(\eta)$ from ($\mathfrak{c}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$ and ($\mathfrak{d}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$, we have $|p_{i}R|<2\epsilon$. This is a contradiction since $\epsilon$ is sufficiently small.
\item Case $(i_{1},i_{2})=(1,3)$
By cancelling the terms $p_{i+1}R$ and $2A(\eta)$ from
($\mathfrak{c}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$ and ($\mathfrak{d}_{3}$).$|p_{i+1}R+A(\delta_{1})-2A(\eta)|<\epsilon$, we have $|A(\delta_{1})+p_{i}R|<2\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2})=(2,1)$
From
($\mathfrak{c}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{d}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have $|p_{i}R|<3\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2})=(2,3)$
In the same way, from
($\mathfrak{c}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{d}_{3}$).$|p_{i+1}R+A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have $|(p_{i+1}-p_{i})R|<3\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2})=(3,1)$
By cancelling the terms $A(\delta_{1})$, $A(\delta_{2})$ and $2A(\eta)$ from
($\mathfrak{c}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{d}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have $|p_{i+1}R|<3\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2})=(3,2)$
By cancelling the terms $p_{i+1}R$ and $2A(\eta)$ from
($\mathfrak{c}_{3}$).$|p_{i+1}R+A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\mathfrak{d}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$, we have$|A(\delta_{1})|<2\epsilon$. This is a contradiction.
By the above arguments, we complete the proof of Claim \ref{firstA1}.
\end{proof}
\begin{proof}[\bf Proof of Lemma \ref{restA1}]
By Claim \ref{firstA1}, we can see that the rest cases which we have to exclude are
$(i_{1},i_{2},i_{3})=(1,1,1)$, $(1,1,3)$, $(2,2,1)$, $(2,2,2)$, $(3,3,1)$, $(3,3,2)$, $(3,3,3)$.
\item Case $(i_{1},i_{2},i_{3})=(1,1,1)$
By cancelling the term $2A(\eta)$ from
($\mathfrak{c}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$ and ($\mathfrak{a}_{1}$).$|p_{i}R-2A(\eta)|<\epsilon$, we have $|(p_{i+1}-2p_{i})R|<\epsilon$ and thus $p_{i}=p_{i+1}-p_{i}$. This contradicts Claim \ref{fre}.
\item Case $(i_{1},i_{2},i_{3})=(1,1,3)$, (resp. $(2,2,2)$)
By cancelling the term $2A(\eta)$ from
($\mathfrak{d}_{1}$).(resp. ($\mathfrak{a}_{2}$).)$|A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\mathfrak{a}_{3}$).(resp. ($\mathfrak{c}_{2}$).)$|A(\delta_{2})-2A(\eta)|<\epsilon$, we have $|A(\delta_{1})-A(\delta_{2})|<2\epsilon$. This contradicts ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})=(2,2,1)$
By cancelling the terms $2A(\eta)$, $A(\delta_{2})$ and $p_{i}R$ from
($\mathfrak{c}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{a}_{1}$).$|p_{i}R-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have $|A(\delta_{1})|<3\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2},i_{3})=(3,3,1)$
By cancelling the terms $2A(\eta)$ from ($\mathfrak{d}_{3}$).$|p_{i+1}R+A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\mathfrak{a}_{1}$).$|p_{i}R-2A(\eta)|<\epsilon$, we have $|(p_{i+1}-p_{i})R+A(\delta_{1})|<2\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2},i_{3})=(3,3,2)$
($\mathfrak{d}_{3}$).$|p_{i+1}R+A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\mathfrak{a}_{2}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ indicate $|p_{i+1}R|<2\epsilon$ and so $p_{i+1}=0$. This is a contradiction.
\item Case $(i_{1},i_{2},i_{3})=(3,3,3)$
In the same way,
($\mathfrak{a}_{3}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{c}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{2})-2A(\eta)|<\epsilon$ indicate $p_{i+1}=p_{i}$. This is a contradiction.
Combining the above argument, we complete the proof of Lemma \ref{restA1}.
\end{proof}
By the discussions so far, the rest cases are $(i_{1},i_{2},i_{3})=(1,1,2)$ and $(2,2,3)$.
\begin{enumerate}
\item[ Case] $(i_{1},i_{2},i_{3})=(1,1,2)$.
From ($\mathfrak{c}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, ($\mathfrak{d}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have
\begin{equation}
A(\delta_{1})\approx (p_{i+1}-p_{i})R,\,\,A(\delta_{2})\approx p_{i+1}R,\,\,A(\eta)\approx \frac{1}{2} (p_{i+1}-p_{i})R.
\end{equation}
\item [Case] $(i_{1},i_{2},i_{3})=(2,2,3)$.
From
($\mathfrak{c}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{d}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|p_{i}R+A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have
\begin{equation}
A(\delta_{1})\approx (p_{i+1}-p_{i})R,\,\,A(\delta_{2})\approx p_{i+1}R,\,\,A(\eta)\approx \frac{1}{2}p_{i+1}R.
\end{equation}
\end{enumerate}
Combining the arguments, we complete the proof of Lemma \ref{typa1}.
\end{proof}
\subsubsection{Type $(A_{2})$}
\begin{lem}\label{typa2}
Suppose that the pair of $\zeta$, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ is Type $(A_{2})$. Then,
the $J$-holomorphic curves counted by the $U$-map cause a contradiction.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{typa2}]
Consider the behaviors of $J$-holomorphic curves counted by $U$-map as $z\to \eta$. Then we obtain three possibilities from $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$.
\begin{enumerate}
\item[($\mathfrak{e}_{1}$).] $|p_{i+1}R-2A(\eta)|<\epsilon$
\item [($\mathfrak{e}_{2}$).] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item[($\mathfrak{e}_{3}$).]$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$
By the same way, we obtain three possibilities from $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$.
\item[($\mathfrak{f}_{1}$).] $|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$
\item [($\mathfrak{f}_{2}$).] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[($\mathfrak{f}_{3}$).]$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$
Also as $z\to \delta_{2}$. then by the splitting behaviors of the curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$, we have
\item[($\mathfrak{g}_{1}$).] $|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$
\item [($\mathfrak{g}_{2}$).] $|A(\delta_{1})-2A(\delta_{2})|<\epsilon$
\item[($\mathfrak{g}_{3}$).]$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$
Also as $z\to \delta_{1}$. then From $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$, we have
\item[($\mathfrak{h}_{1}$).] $|p_{i+1}R-2A(\delta_{1})|<\epsilon$
\item [($\mathfrak{h}_{2}$).]$|A(\delta_{2})-2A(\delta_{1})|<\epsilon$
\item[($\mathfrak{h}_{3}$).]$|A(\delta_{2})+2A(\delta_{1})-p_{i+1}R|<\epsilon$
Recall that we always have ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, since $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$.
We exclude the pairs (($\mathfrak{e}_{i_{1}}$),($\mathfrak{f}_{i_{2}}$),($\mathfrak{g}_{i_{3}}$),($\mathfrak{h}_{i_{4}}$)) in the same way as Type $(A_{1})$.
\end{enumerate}
At first, we prove the next lemma.
\begin{lem}\label{afo}
Any above pairs (($\mathfrak{e}_{i_{1}}$),($\mathfrak{f}_{i_{2}}$),($\mathfrak{g}_{i_{3}}$),($\mathfrak{h}_{i_{4}}$)) causes a contradiction except for the following cases.
$(i_{1},i_{2},i_{3},i_{4})$=$(3,3,1,2)$, $(3,3,1,3)$, $(3,3,3,2)$.
\end{lem}
To prove Lemma \ref{afo}, first of all, we will exclude pairs (($\mathfrak{e}_{i_{1}}$),($\mathfrak{f}_{i_{2}}$)) such that we can derive contradictions by only them.
\begin{cla}\label{abu}
The following each pair (($\mathfrak{e}_{i_{1}}$),($\mathfrak{f}_{i_{2}}$)) causes a contradiction.
$(i_{1},i_{2})$=$(1,1)$, $(3,1)$, $(2,2)$, $(1,3)$.
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{abu}]
\item Case $(i_{1},i_{2})$=$(1,1)$
This is trivial.
\item Case $(i_{1},i_{2})$=$(3,1)$
By cancelling the term $2A(\eta)$ from ($\mathfrak{e}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$ and ($\mathfrak{f}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, we have $|p_{i}R-A(\delta_{2})|<3\epsilon$. But this contradicts ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2})$=$(2,2)$
By cancelling the term $2A(\eta)$ from ($\mathfrak{e}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{f}_{2}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$, we have $|A(\delta_{1})-A(\delta_{2})|<2\epsilon$. This obviously contradict ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2})$=$(1,3)$
($\mathfrak{e}_{1}$).$|p_{i+1}R-2A(\eta)|<\epsilon$ and ($\mathfrak{f}_{3}$).$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$ imply that $|A(\delta_{1})+p_{i}R|<2\epsilon$. This is a contradiction.
\end{proof}
\begin{cla}\label{adddv}
Any pairs (($\mathfrak{e}_{i_{1}}$),($\mathfrak{f}_{i_{2}}$),($\mathfrak{g}_{i_{3}}$)) causes a contradiction except for the following cases.
$(i_{1},i_{2},i_{3})$=$(3,3,1)$, $(3,3,3)$.
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{adddv}]
At first, note that ($\mathfrak{g}_{2}$).$|A(\delta_{1})-2A(\delta_{2})|<\epsilon$ obviously contradicts ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$. So we have only to consider the cases $i_{3}=1$, $2$.
Hence by Claim \ref{abu}, we can find that it is sufficient to exclude the cases $(i_{1},i_{2},i_{3})=(2,1,1)$, $(2,1,3)$, $(1,2,1)$, $(1,2,3)$, $(3,2,1)$, $(3,2,3)$, $(2,3,1)$, $(2,3,3)$.
\item Case $(i_{1},i_{2},i_{3})$=$(2,1,1)$
By cancelling the term $2A(\eta)$ from ($\mathfrak{e}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{f}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, we have $|A(\delta_{2})-(p_{i+1}-p_{i})R|<2\epsilon$. This contradicts ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(2,1,3)$
In the same way as above, by cancelling the term $2A(\eta)$ from ($\mathfrak{e}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{f}_{1}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, we have $|A(\delta_{2})-(p_{i+1}-p_{i})R|<2\epsilon$. This contradicts ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(1,2,1)$
By cancelling the terms $2A(\eta)$ and $A(\delta_{1})$ from
($\mathfrak{e}_{1}$).$|p_{i+1}R-2A(\eta)|<\epsilon$, ($\mathfrak{f}_{2}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $|A(\delta_{2})-(p_{i+1}+p_{i})R|<3\epsilon$. This contradicts ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(1,2,3)$
In the same way as above, by cancelling the terms $2A(\eta)$ and $A(\delta_{1})$ from
($\mathfrak{e}_{1}$).$|p_{i+1}R-2A(\eta)|<\epsilon$, ($\mathfrak{f}_{2}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $|A(\delta_{2})-(p_{i+1}+p_{i})R|<3\epsilon$. This contradicts ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(3,2,1)$
By cancelling the terms $2A(\eta)$ and $A(\delta_{1})$ from ($\mathfrak{e}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, ($\mathfrak{f}$).$|A(\delta_{i})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $|2A(\delta_{2})-(p_{i+1}+p_{i})R|<3\epsilon$. This contradicts ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(3,2,3)$
In the same way as above, by cancelling the terms $2A(\eta)$ and $A(\delta_{1})$ from ($\mathfrak{e}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, ($\mathfrak{f}$).$|A(\delta_{i})-2A(\eta)|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $|2A(\delta_{2})-(p_{i+1}+p_{i})R|<3\epsilon$. This contradicts ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(2,3,1)$
($\mathfrak{e}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{f}_{3}$).$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$ and ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$ imply that $|A(\delta_{1})-A(\delta_{2})|<3\epsilon$. This contradicts ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(2,3,3)$
($\mathfrak{e}_{2}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{f}_{3}$).$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$ imply $|A(\delta_{1})+A(\delta_{2})-(p_{i+1}-p_{i})R|<2\epsilon$. This obviously contradicts ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$.
\end{proof}
\begin{proof}[\bf Proof of Lemma \ref{afo}]
By the above arguments, we can find that the remaining cases are $(i_{1},i_{2},i_{3},i_{4})=(3,3,1,1)$, $(3,3,1,2)$, $(3,3,1,3)$, $(3,3,3,1)$, $(3,3,3,2)$, $(3,3,3,3)$.
To prove the lemma, we will exclude the cases $(i_{1},i_{2},i_{3},i_{4})=(3,3,1,1)$, $(3,3,1,3)$ $(3,3,3,1)$, $(3,3,3,3)$ as follows.
\item Case $(i_{1},i_{2},i_{3},i_{4})=(3,3,1,1)$
By cancelling the term $p_{i+1}R$ from ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$ and ($\mathfrak{h}_{1}$).$|p_{i+1}R-2A(\delta_{1})|<\epsilon$, we have $|A(\delta_{1})-\frac{1}{2}p_{i}R-A(\delta_{2})|<\epsilon$. This contradicts ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$
\item Case $(i_{1},i_{2},i_{3},i_{4})=(3,3,3,1)$
By cancelling the term $A(\delta_{1})$ from ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$ and ($\mathfrak{h}_{1}$). $|p_{i+1}R-2A(\delta_{1})|<\epsilon$, we have $|4A(\delta_{2})-(p_{i+1}-2p_{i})R|<3\epsilon$. Also from ($\mathfrak{h}_{1}$). $|p_{i+1}R-2A(\delta_{1})|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have $|2A(\delta_{2})-(p_{i+1}+2p_{i})R|<3\epsilon$.
$|4A(\delta_{2})-(p_{i+1}-2p_{i})R|<3\epsilon$ and $|2A(\delta_{2})-(p_{i+1}+2p_{i})R|<3\epsilon$ imply $|(p_{i+1}+6p_{i})R|<9\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2},i_{3},i_{4})=(3,3,3,3)$
By cancelling the term $p_{i+1}R$ from ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$ and ($\mathfrak{h}_{3}$).$|A(\delta_{2})+2A(\delta_{1})-p_{i+1}R|<\epsilon$, we have $|A(\delta_{2})-A(\delta_{1})+p_{i}R|<2\epsilon$. This contradicts ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$.
Combining the above consequences, we finish the proof of Lemma \ref{afo}.
\end{proof}
By Lemma \ref{afo}, we still have the following pairs.
$(i_{1},i_{2},i_{3},i_{4})$=$(3,3,1,2)$, $(3,3,1,3)$, $(3,3,3,2)$.
From these pairs. we can decide the approximate relations as follows.
\item[Case] $(i_{1},i_{2},i_{3},i_{4})$=$(3,3,1,2)$
From ($\mathfrak{e}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, ($\mathfrak{f}_{3}$).$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$, ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$, ($\mathfrak{h}_{2}$).$|A(\delta_{2})-2A(\delta_{1})|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have
\begin{equation}\label{i}
A(\delta_{1})\approx p_{i}R,\,\,A(\delta_{2})\approx2p_{i}R,\,\,A(\eta)\approx\frac{3}{2}p_{i}R
\end{equation}
\item [ Case] $(i_{1},i_{2},i_{3},i_{4})$=$(3,3,1,3)$
From ($\mathfrak{e}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, ($\mathfrak{f}_{3}$).$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$, ($\mathfrak{g}_{1}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$, ($\mathfrak{h}_{3}$).$|A(\delta_{2})+2A(\delta_{1})-p_{i+1}R|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have
\begin{equation}\label{ii}
A(\delta_{1})\approx 2p_{i}R,\,\,A(\delta_{2})\approx 3p_{i}R,\,\,A(\eta)\approx 2p_{i}R
\end{equation}.
\item [\bf Case] $(i_{1},i_{2},i_{3},i_{4})$=$(3,3,3,2)$
From ($\mathfrak{e}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, ($\mathfrak{f}_{3}$).$|A(\delta_{1})+2A(\eta)-(p_{i+1}-p_{i})R|<\epsilon$, ($\mathfrak{g}_{3}$).$|A(\delta_{1})+2A(\delta_{2})-(p_{i+1}-p_{i})R|<\epsilon$, ($\mathfrak{h}_{2}$).$|A(\delta_{2})-2A(\delta_{1})|<\epsilon$ and ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$, we have
\begin{equation}\label{iii}
A(\delta_{1})\approx p_{i}R,\,\,A(\delta_{2})\approx2p_{i}R,\,\,A(\eta)\approx2p_{i}R
\end{equation}
Recall that from the splitting behaviors of $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$ as $z \to \eta$, we have three possibilities of splitting of holomorphic curve and also three possibilities of approximate relations as follows.
\item[($\mathfrak{a}_{1}$).] $|p_{i}R-2A(\eta)|<\epsilon$
\item [($\mathfrak{a}_{2}$).] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[($\mathfrak{a}_{3}$).]$|A(\delta_{2})-2A(\eta)|<\epsilon$
But it is easy to check that (\ref{i}), (\ref{ii}), (\ref{iii}) can not hold any these relation. We complete the proof of Lemma \ref{typa2}.
\end{proof}
\subsubsection{Type $(A_{3})$}
\begin{lem}\label{typa3}
Suppose that the pair of $\zeta$, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$ is Type $(A_{3})$. Then,
the $J$-holomorphic curves counted by the $U$-map cause a contradiction.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{typa3}]
Consider the behaviors of $J$-holomorphic curves counted by $U$-map as $z\to \delta_{2}$.
Then we obtain three possibilities of the approximate relations in the actions from the splitting behaviors of the curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{2},1)}) \rangle$.
\begin{enumerate}
\item[($\mathfrak{i}_{1}$).] $|A(\delta_{1})-(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$
\item[($\mathfrak{i}_{2}$).] $|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$
\item[($\mathfrak{i}_{3}$).] $|A(\delta_{1})-2A(\delta_{2})|<\epsilon$.
\end{enumerate}
But Every ($\mathfrak{i}_{1}$), ($\mathfrak{i}_{2}$) and ($\mathfrak{i}_{3}$) causes a contradiction as follows
\begin{enumerate}
\item Case ($\mathfrak{i}_{1}$)
Recall ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$. This obviously contradicts $|A(\delta_{1})-(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$. In fact, by cancelling the term $A(\delta_{1})$ from them, we have $|p_{i+1}R+A(\delta_{2})|<2$. This is a contradiction.
\item Case ($\mathfrak{i}_{2}$)
Since any $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\delta_{1},1)}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ $\delta'\in \hat{\alpha}$ are in $\mathcal{M}^{J}((\gamma,p_{i+1})\cup{(\delta',1)},(\delta_{2},1))$, we have $|p_{i+1}R+A(\delta')-A(\delta_{2})|<\epsilon$.
By cancelling the term $A(\delta_{2})$ from the above inequality and ($\mathfrak{i}_{2}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$, we have $|\frac{1}{2}(p_{i+1}+p_{i})R+A(\delta')|<\frac{3}{2}\epsilon$. This is a contradiction.
\item Case ($\mathfrak{i}_{3}$)
Recall ($\clubsuit$).$|A(\delta_{1})+p_{i}R-A(\delta_{2})|<\epsilon$. This obviously contradicts ($\mathfrak{i}_{3}$).$|A(\delta_{1})-2A(\delta_{2})|<\epsilon$.
\end{enumerate}
Combining the above arguments, we can see that Type $(A_{3})$ can not occur.
\end{proof}
\subsection{Type (B)}
In the same way as Type (A), the topological types of $J$-holomoprhic curves counted by $U\langle\hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}) \rangle$ are both $(g,k,l)=(0,3,0)$.
And also we can see that there is a negative hyperbolic orbit $\delta'$ such that $\zeta=\hat{\alpha}\cup{
(\delta',1)}\cup{(\delta_{2},1)}$ and, any $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}) \rangle$ are in $\mathcal{M}^{J}((\gamma,p_{i+1}),(\delta',1)\cup{(\delta_{2},1)})$ and $\mathcal{M}^{J}((\delta',1),(\gamma,p_{i+1}-p_{i})\cup{(\delta_{1},1)})$ respectively.
\begin{tikzpicture}
\draw (11,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-1.8)--(6,-1.8) ;
\draw (16,-2.8)--(6,-2.8) ;
\draw (6,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-2.8)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (16,-1.8) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-2.8) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-3.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-4.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-5) to [out=180,in=0] (6,-6.5);
\draw (6,-5.5) to [out=0,in=270] (8,-4.7);
\draw (8,-4.7) to [out=90,in=0] (6,-4.3);
\draw (11,-4) to [out=180,in=0] (6,-3.3);
\draw (6,-5.5) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-6.5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (4.6,-6) node{$(\gamma,p_{i+1}-p_{i})$} ;
\draw (4.9,-3.9) node{$(\delta_{1},1)$} ;
\draw (4.9,-7.8) node{$(\delta_{2},1)$} ;
\draw (17.2,-5.2) node{$(\gamma,p_{i+1})$} ;
\draw (17.1,-2.3) node{$(\eta,1) \in \hat{\alpha}$} ;
\draw[densely dotted, thick] (6,-1.5)--(6,-8.5);
\draw[densely dotted, thick] (11,-1.5)--(11,-8.5) ;
\draw[densely dotted, thick] (16,-1.5)--(16,-8.5);
\draw (7.2,-8.9) node{$\hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})\cup{(\delta_{1},1)\cup{(\delta_{2},1)}}}$};
\draw (16,-8.9) node{$\hat{\alpha}\cup{(\gamma,p_{i+1})}$};
\draw (12,-8.9) node{$\hat{\alpha}\cup{(\delta',1)}\cup{(\delta_{2},1)}$};
\draw (11.5,-5.5) node{$(\delta',1)$};
\draw (16,-4.7) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (16,-5.7) arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-4) to [out=0,in=180] (16,-4.7);
\draw (11,-5) to [out=0,in=90] (13,-6.1);
\draw (13,-6.1) to [out=270,in=0] (11,-7.3);
\draw (16,-5.7) to [out=180,in=0] (11,-8.3);
\draw (11,-7.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-8.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (6,-7.3) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (6,-8.3)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-7.3)--(6,-7.3) ;
\draw (11,-8.3)--(6,-8.3) ;
\draw (11,-4) arc [start angle=90,end angle=270,x radius=0.25,y radius=0.5];
\draw (11,-5)[dashed] arc [start angle=270,end angle=450,x radius=0.25,y radius=0.5];
\draw (11,-9.8) node{\Large Type $(B)$} ;
\end{tikzpicture}
\begin{lem}\label{typb}
Suppose that the pair of $\zeta$, $J$-holomorphic curves counted by $U\langle\hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$ and $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}) \rangle$ are above. Then, one of the following holds.
\item[($\Delta_{1}$).] $\frac{3}{2}p_{i}=p_{i+1}$. Moreover
$A(\delta_{1})\approx \frac{1}{2}p_{i}R$, $A(\delta_{2})\approx \frac{1}{2}p_{i}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}p_{i}R$ or $A(\eta)\approx \frac{1}{4}p_{i}R$.
\item[($\Delta_{2}$).]$\frac{4}{3}p_{i}=p_{i+1}$. Moreover, $A(\delta_{1})\approx \frac{2}{3}p_{i}R$, $A(\delta_{2})\approx \frac{1}{3}p_{i}R$ and for each $\eta \in \hat{\alpha}$, either $A(\eta)\approx \frac{1}{2}p_{i}R$ or $A(\eta)\approx \frac{1}{6}p_{i}R$.
\end{lem}
\begin{proof}[\bf Proof of Proposition \ref{nagai}]
Since ($\Delta_{1}$) and ($\Delta_{2}$) are correspond to ($b$) and ($c$) respectively, combine with the result of Type (A), we complete the proof of Proposition \ref{nagai}.
\end{proof}
\begin{proof}[\bf Proof of Lemma \ref{typb}]
Let $z\to \eta$. Then we obtain three possibilities from the splitting behaviors of the curves counted by $U\langle\hat{\alpha}\cup{(\gamma,p_{i+1})} \rangle=\langle \zeta \rangle$.
\item[($\mathfrak{j}_{1}$).] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item [($\mathfrak{j}_{2}$).] $|p_{i+1}R-2A(\eta)|<\epsilon$
\item[($\mathfrak{j}_{3}$).]$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$
By the same way, we have three possibilities of the approximate actions in the orbits from the splitting behaviors of the curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}) \rangle$.
\item[($\mathfrak{k}_{1}$).] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item [($\mathfrak{k}_{2}$).] $|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$
\item[($\mathfrak{k}_{3}$).]$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$
Also let $z\to \delta_{2}$. then by the splitting behaviors of the curves counted by $U\langle \zeta \rangle=\langle \hat{\alpha}\cup{(\gamma,p_{i+1}-p_{i})}\cup{(\delta_{1},1)}\cup{(\delta_{2},1)}) \rangle$, we have
\item[($\mathfrak{l}_{1}$).] $|A(\delta_{1})-2A(\delta_{2})|<\epsilon$
\item [($\mathfrak{l}_{2}$).] $|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$
\item[($\mathfrak{l}_{3}$).]$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\delta_{2})|<\epsilon$
Recall that we always have ($\spadesuit$).$|A(\delta_{1})+A(\delta_{2})-p_{i}R|<\epsilon$, since $A(\alpha_{k+1})-A(\alpha_{k})<\epsilon$.
As is the same with so far, Some pair (($\mathfrak{j}_{i_{1}}$),($\mathfrak{k}_{i_{2}}$),($\mathfrak{l}_{i_{3}}$)) cause contradictions as follows.
\begin{lem}\label{typB}
Any above pairs (($\mathfrak{j}_{i_{1}}$),($\mathfrak{k}_{i_{2}}$),($\mathfrak{l}_{i_{3}}$)) causes a contradiction except for the following cases.
$(i_{1},i_{2},i_{3})=(1,1,3),(1,2,1),(1,2,3),(3,3,1),(3,3,3)$.
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{typB}]
At first, we will exclude pairs (($\mathfrak{j}_{i_{1}}$),($\mathfrak{k}_{i_{2}}$)) such that we can derive a contradiction by only them.
\begin{cla}\label{clB}
Each of the following pairs (($\mathfrak{j}_{i_{1}}$),($\mathfrak{k}_{i_{2}}$)) causes a contradiction.
$(i_{1},i_{2})$=$(2,1)$, $(2,2)$, $(2,3)$, $(3,1)$, $(3,2)$.
\end{cla}
\begin{proof}[\bf Proof of Claim \ref{clB}]
\item Case $(i_{1},i_{2})$=$(2,1)$
By cancelling the term $2A(\eta)$ from
($\mathfrak{j}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$ and ($\mathfrak{k}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$, we have $|p_{i+1}R-A(\delta_{1})|<2\epsilon$. This contradicts ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$.
\item Case $(i_{1},i_{2})$=$(2,2)$
By cancelling the term $2A(\eta)$ from
($\mathfrak{j}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$ and ($\mathfrak{k}_{2}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, we have $|p_{i}R|<2\epsilon$. This is a contradiction.
\item Case $(i_{1},i_{2})$=$(2,3)$
By cancelling the term $2A(\eta)$ from
($\mathfrak{j}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$ and ($\mathfrak{k}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$, we have $|p_{i}R-A(\delta_{1})|<2\epsilon$. This contradicts ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$.
\item Case $(i_{1},i_{2})$=$(3,1)$
By cancelling the term $2A(\eta)$ from
($\mathfrak{j}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$ and ($\mathfrak{k}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$, we have $|p_{i+1}R-(A(\delta_{1})+A(\delta_{2}))|<2\epsilon$. But this contradicts ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$.
\item Case $(i_{1},i_{2})$=$(3,2)$
By cancelling the term $2A(\eta)$ from
($\mathfrak{j}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$ and ($\mathfrak{k}_{2}$).$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, we have $|p_{i}R-A(\delta_{2})|<2\epsilon$. But this contradicts ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$.
\end{proof}
By the arguments so far, we have only to show that
each of the following pairs (($\mathfrak{j}_{i_{1}}$),($\mathfrak{k}_{i_{2}}$),($\mathfrak{l}_{i_{3}}$)) causes a contradiction.
$(i_{1},i_{2},i_{3})$=$(1,1,1)$, $(1,1,2)$, $(1,3,1)$, $(1,3,2)$, $(1,3,3)$, $(1,2,2)$, $(3,3,2)$.
\item Case $(i_{1},i_{2},i_{3})$=$(1,1,1)$
Obviously, ($\mathfrak{j}_{1}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{k}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ contradict
($\mathfrak{l}_{1}$).$|A(\delta_{1})-2A(\delta_{2})|<\epsilon$
\item Case $(i_{1},i_{2},i_{3})$=$(1,1,2)$
By cancelling the terms $A(\delta_{1})$, $A(\delta_{2})$ and $2A(\eta)$ from ($\mathfrak{j}_{1}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{k}_{1}$).$|A(\delta_{1})-2A(\eta)|<\epsilon$ ($\mathfrak{l}_{2}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$, and ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$, we have $|p_{i}R-(p_{i+1}-p_{i})R|<4\epsilon$. Hence $p_{i}=p_{i+1}-p_{i}$. This is a contradiction.
\item Case $(i_{1},i_{2},i_{3})$=$(1,3,1)$
By cancelling the terms $A(\delta_{1})$ and $2A(\eta)$ from ($\mathfrak{j}_{1}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{k}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$ and ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$, we have $|p_{i+1}R-2A(\delta_{2})|<3\epsilon$ and hence $|(p_{i}-\frac{1}{2}p_{i+1})R-A(\delta_{1})|<\frac{5}{2}\epsilon$.
By combining with ($\mathfrak{l}_{1}$).$|A(\delta_{1})-2A(\delta_{2})|<\epsilon$, we have $|(\frac{3}{2}p_{i+1}-p_{i})R|<\frac{13}{2}\epsilon$ This is a contradiction.
\item Case $(i_{1},i_{2},i_{3})$=$(1,3,2)$
In the same way as above, from ($\mathfrak{j}_{1}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{k}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$, and ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$, we have $|p_{i+1}R-2A(\delta_{2})|<3\epsilon$. This obviously contradicts ($\mathfrak{l}_{2}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(1,3,3)$
In the same way as above, from ($\mathfrak{j}_{1}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$, ($\mathfrak{k}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$, and ($\spadesuit$).$|p_{i}R-(A(\delta_{1})+A(\delta_{2}))|<\epsilon$, we have $|p_{i+1}R-2A(\delta_{2})|<3\epsilon$ and $|(p_{i}-\frac{1}{2}p_{i+1})R-A(\delta_{1})|<\frac{5}{2}\epsilon$.
By combining with ($\mathfrak{l}_{3}$).$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\delta_{2})|<\epsilon$, we have $|\frac{1}{2}p_{i+1}R|<\frac{13}{2}\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(1,2,2)$
By cancelling the term $2A(\eta)$ from ($\mathfrak{j}_{1}$).$|A(\delta_{2})-2A(\eta)|<\epsilon$ and ($\mathfrak{j}_{2}$).$|p_{i+1}R-2A(\eta)|<\epsilon$, we have $|p_{i+1}R-A(\delta_{2})|<2\epsilon$. This obviously contradicts ($\mathfrak{l}_{2}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$.
\item Case $(i_{1},i_{2},i_{3})$=$(3,3,2)$
From ($\mathfrak{j}_{3}$).$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, ($\mathfrak{l}_{2}$).$|(p_{i+1}-p_{i})R-2A(\delta_{2})|<\epsilon$ and ($\spadesuit$).$|p_{i}R-A(\delta_{1})-A(\delta_{2})|<\epsilon$, we have
\begin{equation}\label{atla}
(\frac{3}{2}p_{i}-\frac{1}{2}p_{i+1})R\approx A(\delta_{1}),\,\,\frac{1}{2}(p_{i+1}-p_{i})R\approx A(\delta_{2}),\,\,\frac{1}{4}(p_{i+1}+p_{i})R\approx A(\eta).
\end{equation}
Recall that when we consider the splitting behaviors of $J$-holomorphic curve counted by $U\langle \alpha_{k+1} \rangle=\langle \alpha_{k} \rangle$ as $z \to \eta$ for some $\eta\in \hat{\alpha}$. Then there are three possibilities of splitting of holomorphic curve and also we have three possibilities of the approximate relation in the actions as follows.
\item[($\mathfrak{b}_{1})$.] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{b}_{2})$.] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{b}_{3})$.] $|p_{i}R-2A(\eta)|<\epsilon$.
But obviously, (\ref{atla}) contradicts the above relations in any case.
Combining the arguments, we complete the proof of Lemma \ref{typB}.
\end{proof}
To complete the proof of Lemma \ref{typb}, it is sufficient to compute the approximate relations from the rest cases $(i_{1},i_{2},i_{3})=(1,1,3)$, $(1,2,1)$, $(1,2,3)$, $(3,3,1)$, $(3,3,3)$.
\item Case $(i_{1},i_{2},i_{3})=(1,1,3)$
From $(\mathfrak{j}_{1})$.$|A(\delta_{2})-2A(\eta)|<\epsilon$, $(\mathfrak{k}_{1})$.$|A(\delta_{1})-2A(\eta)|<\epsilon$,
$(\mathfrak{l}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\delta_{2})|<\epsilon$ and ($\spadesuit$).$|A(\delta_{1})+A(\delta_{2})-p_{i}R|<\epsilon$, we have
\begin{equation}
A(\delta_{1})\approx \frac{1}{2}p_{i}R,\,\, A(\delta_{2}) \approx \frac{1}{2}p_{i}R,\,\,p_{i+1}= \frac{3}{2}p_{i},\,\,A(\eta)\approx \frac{1}{4}p_{i}.
\end{equation}
\item Case $(i_{1},i_{2},i_{3})=(1,2,1)$
From $(\mathfrak{j}_{1})$.$|A(\delta_{2})-2A(\eta)|<\epsilon$, $(\mathfrak{k}_{2})$.$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, $(\mathfrak{l}_{1})$.$|A(\delta_{1})-2A(\delta_{2})|<\epsilon$ and ($\spadesuit$).$|A(\delta_{1})+A(\delta_{2})-p_{i}R|<\epsilon$. we have
\begin{equation}
A(\delta_{1})\approx \frac{2}{3}p_{i}R,\,\,A(\delta_{2})\approx \frac{1}{3}p_{i}R,\,\,p_{i+1}=\frac{4}{3}p_{i},\,\,A(\eta)\approx \frac{1}{6}p_{i}R.
\end{equation}
\item Case $(i_{1},i_{2},i_{3})=(1,2,3)$
From $(\mathfrak{j}_{1})$.$|A(\delta_{2})-2A(\eta)|<\epsilon$, $(\mathfrak{k}_{2})$.$|(p_{i+1}-p_{i})R-2A(\eta)|<\epsilon$, $(\mathfrak{l}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\delta_{2})|<\epsilon$ and ($\spadesuit$).$|A(\delta_{1})+A(\delta_{2})-p_{i}R|<\epsilon$, we have
\begin{equation}
A(\delta_{1})\approx\frac{1}{2}p_{i}R,\,\, A(\delta_{2}) \approx \frac{1}{2}p_{i}R,\,\,p_{i+1}= \frac{3}{2}p_{i},\,\,A(\eta)\approx \frac{1}{4}p_{i}.
\end{equation}
\item Case $(i_{1},i_{2},i_{3})=(3,3,1)$
Recall that by considering the splitting behavior of $J$-holomorphic curves counted by $U$-map from $\langle \alpha_{k+1} \rangle$ to $\langle \alpha_{k} \rangle$ as $z\to \eta$. Then we have three possibilties,
\item[$(\mathfrak{b}_{1})$.] $|A(\delta_{1})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{b}_{2})$.] $|A(\delta_{2})-2A(\eta)|<\epsilon$
\item[$(\mathfrak{b}_{3})$.] $|p_{i}R-2A(\eta)|<\epsilon$
The first inequality contradicts $(\mathfrak{k}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$ and also the second one does $(\mathfrak{k}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$ and $|A(\delta_{1})-2A(\delta_{2})|<\epsilon$. So only third one may be possible.
By combining with $(\mathfrak{j}_{3})$.$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, $(\mathfrak{k}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$, $(\mathfrak{l}_{1})$.$|A(\delta_{1})-2A(\delta_{2})|<\epsilon$ and ($\spadesuit$).$|A(\delta_{1})+A(\delta_{2})-p_{i}R|<\epsilon$, we have
\begin{equation}\label{}
A(\delta_{1})\approx \frac{2}{3}p_{i}R,\,\,A(\delta_{2})\approx \frac{1}{3}p_{i}R,\,\,p_{i+1}=\frac{4}{3}p_{i},\,\,A(\eta)\approx \frac{1}{2}p_{i}R.
\end{equation}
\item Case $(i_{1},i_{2},i_{3})=(3,3,3)$
From $(\mathfrak{j}_{3})$.$|p_{i+1}R-(2A(\eta)+A(\delta_{2}))|<\epsilon$, $(\mathfrak{k}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\eta)|<\epsilon$, $(\mathfrak{l}_{3})$.$|(p_{i+1}-p_{i})R+A(\delta_{1})-2A(\delta_{2})|<\epsilon$ and ($\spadesuit$).$|A(\delta_{1})+A(\delta_{2})-p_{i}R|<\epsilon$ we have
\begin{equation}
A(\delta_{1})\approx\frac{1}{2}p_{i}R,\,\, A(\delta_{2}) \approx \frac{1}{2}p_{i}R,\,\,p_{i+1}= \frac{3}{2}p_{i},\,\,A(\eta)\approx \frac{1}{2}p_{i}.
\end{equation}
Combining the arguments, we complete the proof of Lemma \ref{typb}.
\end{proof}
\section{Proof of Theorem \ref{mainmainmain}}
Suppose that $\alpha_{k}$, $\alpha_{k+1}$ satisfy the assumptions 1, 2, 3, 4 and 5 in Proposition \ref{nagai}, then the pair $(\alpha_{k},\alpha_{k+1})$ is one of the types of \textbf{(a)}, \textbf{(a')}, \textbf{(b)}, \textbf{(b')}, \textbf{(c)}, \textbf{(c')}. Moreover, we can see that any actions of negative hyperbolic orbits in $\alpha_{k+1}$, $\alpha_{k}$ are in $(\frac{1}{12}R\mathbb{Z})_{\epsilon'}=\{\,x \in \mathbb{R}\,|\,\mathrm{dist}(x,\frac{1}{12}R\mathbb{Z})<\epsilon'\,\,\}$ where $\epsilon'=\frac{1}{100}\mathrm{max}\{A(\alpha)\,|\alpha\,\,\mathrm{is\,\,a\,\,Reeb\,\,orbit}\,\}$. Therefore, we can define a map
\begin{equation}
f:\{\,\eta\in \alpha_{k}\,|\,\eta\,\,\mathrm{is\,\,negative\,\,hyperbolic}\,\}\to \frac{1}{12}R\mathbb{Z}
\end{equation}
Here, $f(\eta)\in \frac{1}{12}R\mathbb{Z}$ is an image by the natural projection of $A(\eta)$ from $(\frac{1}{12}R\mathbb{Z})_{\epsilon'}$ to $\frac{1}{12}R\mathbb{Z}$. Note that $\frac{1}{12}R\mathbb{Z}$ is discrete.
Let $\alpha_{k+1}$, $\alpha_{k}$ and $\alpha_{k-1}$ be such 3 consecutive admissible orbit sets. then each type of the pairs $(\alpha_{k},\alpha_{k+1})$ and $(\alpha_{k-1},\alpha_{k})$ decides the action relations in $\alpha_{k}$ respectively. But sometimes, they may be incompatible. In particular, we can show that such fact induces contradictions with Proposition \ref{main index 2}.
At first, we show the next lemma.
\begin{lem}\label{las}
Suppose that $\alpha_{k-1}$, $\alpha_{k}$ and $\alpha_{k+1}$ satisfy the assumptions 1, 2, 3, 4 and 5 in Proposition \ref{nagai}. Then the following pairs of types of $(\alpha_{k},\alpha_{k+1})$ and $(\alpha_{k-1},\alpha_{k})$ classified in Proposition \ref{nagai} cause contradictions.
(type of ($\alpha_{k-1},\alpha_{k}$), type of ($\alpha_{k},\alpha_{k+1}$))\\
=(\textbf{a'}, \textbf{a}), (\textbf{b'}, \textbf{a'}), (\textbf{b'}, \textbf{a}), (\textbf{b'}, \textbf{b}), (\textbf{b'}, \textbf{b'}), (\textbf{b'}, \textbf{c}), (\textbf{b'}, \textbf{c'}), (\textbf{c'}, \textbf{a}), (\textbf{c'}, \textbf{a'}),
(\textbf{c'}, \textbf{b'}), (\textbf{c'}, \textbf{c}), (\textbf{c'}, \textbf{c'}), (\textbf{a'}, \textbf{b}), (\textbf{b}, \textbf{b}), (\textbf{c'}, \textbf{b}), (\textbf{c}, \textbf{b}), (\textbf{a}, \textbf{c}),
(\textbf{b}, \textbf{c}), (\textbf{c}, \textbf{c}), (\textbf{a}, \textbf{a}), (\textbf{a'}, \textbf{a'}), (\textbf{a}, \textbf{b}), (\textbf{a'}, \textbf{c}), (\textbf{a}, \textbf{a'}).
\end{lem}
\begin{proof}[\bf Proof of Lemma \ref{las}]
We derive a contradiction respectively.
\item [\bf Case] (\textbf{b'}, \textbf{a})(or by symmetry, (\textbf{a'}, \textbf{b}))
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{b'}), the number of the inverse image of the largest value of $f:\{\,\eta\in \alpha_{k}\,|\,\eta\,\,\mathrm{is\,\,negative\,\,hyperbolic}\,\}\to \frac{1}{12}R\mathbb{Z}$ is at least two. But the assumption that the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{a}) indicates that the number of the inverse image of the largest value of $f:\{\,\eta\in \alpha_{k}\,|\,\eta\,\,\mathrm{is\,\,negative\,\,hyperbolic}\,\}\to \frac{1}{12}R\mathbb{Z}$ is one. This is a contradiction.
\item [\bf Case] (\textbf{b'}, \textbf{a'})(or by symmetry, (\textbf{a}, \textbf{b}))
Let $E(\alpha_{k-1})=M$, then $E(\alpha_{k})=M-q_{i}$ where $q_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Let $q_{i'}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M-q_{i}\}})$.
Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{a'}), the image of $f$ on $\alpha_{k}$ is in $\{\,\frac{1}{2}q_{i'+1}R,\,\,\frac{1}{2}(q_{i'+1}-q_{i'})R,\,\,(q_{i'+1}-q_{i'})R\,\}$.
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{a'}), we have $q_{i+1}=\frac{3}{2}q_{i}$.
Hence
\begin{equation}
q_{i'} \leq M-q_{i}\leq q_{i+1}-q_{i}= \frac{1}{2}q_{i}<q_{i}.
\end{equation}
Therefore $q_{i'+1}\leq q_{i}$.
If $q_{i'+1}<q_{i}$, the actions of each orbit in $\hat{\alpha}$ are less than $\frac{1}{2} q_{i}R$. The number of the positive ends of (\textbf{b'}) whose image of $f$ are largest is at least two but that one of negative ends of (\textbf{a'}) whose image of $f$ are largest is only $\delta_{2}$. This is a contradiction.
If $q_{i'+1}=q_{i}$, by considering the correspondence of the approximate action values and Claim \ref{fre}, we can see $\frac{1}{4}q_{i}R = (q_{i'+1}-q_{i'})R$ and so $\frac{1}{4}q_{i}=q_{i'+1}-q_{i'}=q_{i}-q_{i'}$ hence $\frac{3}{4}q_{i}=q_{i'}$. This contradict $q_{i'}\leq \frac{1}{2}q_{i}$.
\item [\bf Case] (\textbf{b'}, \textbf{b})
Let $E(\alpha_{k-1})=M$ $E(\alpha_{k+1})=M'$, then $E(\alpha_{k})=M-q_{i}=M'-p_{i'}$ where $q_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$ and $p_{i'}=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M'\}})$.
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{b'}), the largest value of $f$ on $\alpha_{k}$ is $\frac{1}{2}q_{i}R$ and Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{b}), the largest value of $f$ on $\alpha_{k}$ is $\frac{1}{2} p_{i'}R$. Therefore we have $q_{i}=p_{i'}$. This contradicts $S_{\theta}\cap{S_{-\theta}}=\{1\}$.
\item [\bf Case] (\textbf{b'}, \textbf{b'}) (or by symmetry, (\textbf{b}, \textbf{b}))
Let $E(\alpha_{k-1})=M$, then $E(\alpha_{k})=M-q_{i}<\frac{1}{2}q_{i}$. This implies that the largest approximate value decided by the positive ends of former (\textbf{b'}) is larger than that one decided by the negative ends of the later (\textbf{b'}) and so this is a contradiction.
\item [\bf Case] (\textbf{b'}, \textbf{c}) (or by symmetry, (\textbf{c'}, \textbf{b}))
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{b'}), the number of the image of $f$ on $\alpha_{k}$ is at most 2. On the other hand, since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{c}), the number of the image of $f$ on $\alpha_{k}$ is at least three. This is a contradiction.
\item [\bf Case] (\textbf{b'}, \textbf{c'}) (or by symmetry, (\textbf{c}, \textbf{b}))
Let $E(\alpha_{k-1})=M$. Then $E(\alpha_{k})=M-q_{i}<\frac{1}{2}q_{i}$. This implies that the largest approximate value decided by the action of the positive ends of (\textbf{b'}) is larger than that one decided by the action of the negative ends of (\textbf{c'}) and so this is a contradiction.
\item [\bf Case] (\textbf{c'}, \textbf{a}) (or by symmetry, (\textbf{a'}, \textbf{c}))
Let $E(\alpha_{k-1})=M$ $E(\alpha_{k+1})=M'$, then $E(\alpha_{k})=M-q_{i}=M'-p_{i'}$ where $q_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$ and $p_{i'}=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M'\}})$.
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{c'}), the largest value of $f$ on $\alpha_{k}$ is $\frac{2}{3}q_{i}R$ and Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{a}), the largest value of $f$ on $\alpha_{k}$ is $p_{i'+1}R$. Therefore we have $\frac{2}{3}q_{i}=p_{i'+1}$.
By considering the correspondence of the actions of the part of trivial cylinders, we have two possibilities, $\frac{1}{2}q_{i}=\frac{1}{2}(p_{i'+1}-p_{i'})$ or $\frac{1}{6}q_{i}=\frac{1}{2}(p_{i'+1}-p_{i'})$.
If $\frac{1}{2}q_{i}=\frac{1}{2}(p_{i'+1}-p_{i'})$, this contradicts $\frac{2}{3}q_{i}=p_{i'+1}$.
If $\frac{1}{6}q_{i}=\frac{1}{2}(p_{i'+1}-p_{i'})$, we have $\frac{1}{2}p_{i'+1}=(p_{i'+1}-p_{i'})$ since $\frac{2}{3}q_{i}=p_{i'+1}$. This contradicts Claim \ref{fre}.
Combining the above arguments, we can see that this case causes a contradiction.
\item [\bf Case] (\textbf{c'}, \textbf{a'}) (or by symmetry, (\textbf{a}, \textbf{c}))
Let $E(\alpha_{k-1})=M$. Then $E(\alpha_{k})=M-q_{i}$ $E(\alpha_{k+1})=M-q_{i}-q_{i'}$ where $q_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$ and $q_{i'}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M-q_{i}\}})$.
Moreover, since $\frac{4}{3}q_{i+1}=q_{i}$,
\begin{equation}\label{addd}
q_{i'}<E(\alpha_{k})=M-q_{i}<\frac{1}{3}q_{i}<q_{i}
\end{equation}
and so
\begin{equation}\label{asda}
q_{i'+1}\leq q_{i}.
\end{equation}
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{c'}), the largest value of $f$ on $\alpha_{k}$ is $\frac{2}{3}q_{i}R$. On the other hand, Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{a'}), the largest value of $f$ on $\alpha_{k}$ is $(q_{i'+1}-q_{i'})R$ or $\frac{1}{2}q_{i'+1}$. This indicates that $\frac{2}{3}q_{i}=q_{i'+1}-q_{i'}$ or $\frac{2}{3}q_{i}=\frac{1}{2}q_{i'+1}$ but the later case contradicts (\ref{asda}). So it is sufficient to consider the former case.
Suppose that $\frac{2}{3}q_{i}=q_{i'+1}-q_{i'}$.
By considering the correspondence of the actions of the part of trivial cylinder, we have two possibilities, $\frac{1}{2}q_{i}=\frac{1}{2}q_{i'+1}$ or $\frac{1}{6}q_{i}=\frac{1}{2}q_{i'+1}$.
If $\frac{1}{2}q_{i}=\frac{1}{2}q_{i'+1}$, we can see $\frac{1}{3}q_{i}=q_{i'}$ by $\frac{2}{3}q_{i}=q_{i'+1}-q_{i'}$. But this contradict (\ref{addd}) $q_{i'}<\frac{1}{3}q_{i}$.
If $\frac{1}{6}q_{i}=\frac{1}{2}q_{i'+1}$, we can see $\frac{1}{3}q_{i}+q_{i'}=0$ by $\frac{2}{3}q_{i}=q_{i'+1}-q_{i'}$. This is a contradiction.
Combining the above argument, we can see that this case cause a contradiction.
\item [\bf Case] (\textbf{c'}, \textbf{b'}) (or by symmetry, (\textbf{b}, \textbf{c}))
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{c'}), the number of the image of $f$ on $\alpha_{k}$ is at least three. On the other hand, Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{b'}), the number of the image of $f$ on $\alpha_{k}$ is at most two. This is a contradiction.
\item [\bf Case] (\textbf{c'}, \textbf{c})
Let $E(\alpha_{k-1})=M$ and $E(\alpha_{k+1})=M'$. Let $q_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$ and $p_{i'}=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M'\}})$.
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{c'}), the largest value of $f$ on $\alpha_{k}$ is $\frac{2}{3}q_{i}R$. On the other hand, Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{c}), the largest value of $f$ on $\alpha_{k}$ is $\frac{2}{3}p_{i'}R$. Therefore $q_{i}=p_{i'}$. This contradicts $S_{\theta}\cap{S_{-\theta}}=\{1\}$.
\item [\bf Case] (\textbf{c'}, \textbf{c'}) (or by symmetry, (\textbf{c}, \textbf{c}))
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{c'}), the number of the image of $f$ on $\alpha_{k}$ on $\alpha_{k}$ is at least three. On the other hand, Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{c'}), the number of the image of $f$ on $\alpha_{k}$ is at most two. This is a contradiction.
\item [\bf Case] (\textbf{a}, \textbf{a}) (or by symmetry, (\textbf{a'}, \textbf{a'}))
Let $E(\alpha_{k+1})=M$. Then $E(\alpha_{k})=M-p_{i}$ where $p_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$. Set $p_{i'}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M-p_{i}\}})$.
By construction, $p_{i}\leq p_{i'}$ and so $p_{i+1}\leq p_{i'+1}$.
Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{a}), the largest value of $f$ on $\alpha_{k}$ is $p_{i+1}R$. On the other hand, since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{a}), the image of $f$ on $\alpha_{k}$ is in $\{\,\frac{1}{2}p_{i'+1}R,\,\,\frac{1}{2}(p_{i'+1}-p_{i'})R,\,\,(p_{i'+1}-p_{i'})R\,\}$. This is obviously a contradiction.
\item [\bf Case] (\textbf{a}, \textbf{a'})
Let $E(\alpha_{k})=M$, $p_{i}=\mathrm{max}(S_{\theta}\cap{\{1,\,\,2,...,\,\,M\}})$, $q_{i'}=\mathrm{max}(S_{-\theta}\cap{\{1,\,\,2,...,\,\,M\}})$.
By considering the correspondence of the actions of the part of trivial cylinder and Claim \ref{fre}, we have two possibilities, $\frac{1}{2}p_{i+1}=\frac{1}{2}(q_{i'+1}-q_{i'})$ or $\frac{1}{2}q_{i'+1}=\frac{1}{2}(p_{i+1}-p_{i})$. Here we use $S_{\theta}\cap{S_{-\theta}}=\{1\}$ implicitly.
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{a}), the image of $f$ on $\alpha_{k}$ is in $\{\,\frac{1}{2}p_{i+1}R,\,\,\frac{1}{2}(p_{i+1}-p_{i})R,\,\,(p_{i+1}-p_{i})R\,\}$.
Suppose that $\frac{1}{2}p_{i+1}=\frac{1}{2}(q_{i'+1}-q_{i'})$. Since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{a'}), the image of $f$ on $\alpha_{k}$ contains $(q_{i'+1}-q_{i'})R$. But since $\frac{1}{2}p_{i+1}=\frac{1}{2}(q_{i'+1}-q_{i'})$, then $(q_{i'+1}-q_{i'})R$ is larger than any $\{\,\frac{1}{2}p_{i+1}R,\,\,\frac{1}{2}(p_{i+1}-p_{i})R,\,\,(p_{i+1}-p_{i})R\,\}$. This is a contradiction.
In the case of $\frac{1}{2}q_{i'+1}=\frac{1}{2}(p_{i+1}-p_{i})$, we can also derive a contradiction in the same way.
\item [\bf Case] (\textbf{a'}, \textbf{a})
Since the type of\,\,($\alpha_{k-1},\alpha_{k}$) is (\textbf{a'}),
the largest value in the image of $f$ on $\alpha_{k}$ is $ q_{i+1}R$ and also since the type of\,\,($\alpha_{k},\alpha_{k+1}$) is (\textbf{a}), we have that the largest value in the image of $f$ on $\alpha_{k}$ is $ p_{i+1}R$. This indicates $p_{i+1}=q_{i+1}$ but this contradicts $S_{\theta}\cap{S_{-\theta}}=\{1\}$.
Combining the discussions so far, we complete the proof of Lemma.
\end{proof}
\begin{proof}[\bf Proof of Theorem \ref{mainmainmain}]
By Lemma \ref{las}, we have the rest possibilities of pairs, (\textbf{b}, \textbf{a}), (\textbf{a}, \textbf{b'}), (\textbf{b}, \textbf{b'}), (\textbf{c}, \textbf{b'}), (\textbf{a'}, \textbf{b'}), (\textbf{c}, \textbf{a}), (\textbf{a}, \textbf{c'}), (\textbf{b}, \textbf{c'}), (\textbf{c}, \textbf{c'}), (\textbf{a'}, \textbf{c'}), (\textbf{b}, \textbf{a'}), (\textbf{c}, \textbf{a'}).
It is easy to check that we can not connect the above pairs more than two. This contradicts Proposition \ref{main index 2} and we complete the proof of Theorem \ref{mainmainmain}.
\end{proof}
|
\section{Introduction}
\label{chap:introduction}
The proposed India-based Neutrino Observatory (INO) at Bodi hill
($9^\circ 58'N, 77^\circ 16' E$) in the Southern part of India, with
at least 1\,km rock overburden in all directions will be
the largest experimental facility of basic science
in India to carry out front-ranking experiments in the field of
particle and astroparticle physics.
The Iron Calorimeter (ICAL) detector is the proposed underground neutrino
physics experiment
in the INO cavern, which aims to observe the neutrino oscillation
pattern at least over one full period.
Main goals of this experiment
are the precise measurements of neutrino oscillation parameters
including the sign of the 2-3 mass-squared difference,
$\Delta m_{32}^2 (= m^2_3 - m^2_2)$ through matter effects,
and, last but not the least, the search for any
non-standard effect beyond neutrino oscillations \cite{ino_pramana}.
The detector size is
48\,m\,$\times$\,16\,m\,$\times$\,14.5\,m made of 151 layers of
5.6\,cm iron plates interleaved with the Resistive Plate Chamber (RPC)
for the detection of the muon trajectory. There will be $\sim$ 30000 RPCs
of size $\sim$2\,m\,$\times$\,2\,m, made of glass gap with pickup
panels and corresponding electronics to collect and store the induced signals which are produced due to the avalanche of electrons along
the trajectory of the charged particles.
Initially, a small number of this large RPC \cite{santanico} were built in the laboratory in
collaboration with the industry. After establishing the whole procedure, jobs are
given to different industries, e.g., production of the gas gap, pickup panel, electronics,
power supply etc., are given in different companies. After assuring the quality control (QC)
of each component separately, all are brought back to the laboratory for final assembly and
test the performance of the RPC. Though there is a stringent condition on the QC at
the vendor's place, the performance of the finally produced RPC is slightly worse than the earlier
produced RPC in the laboratory. One example is the uniformity of the gas gap and consequently
the variation of gain/noise rate of RPC as a function of the position in the RPC. To have
a higher gain of the RPC in the whole surface area, the noise rate, as well as the probability of
streamer formation, is also increased in the high gain area of the detector, which eventually deteriorates the
position and timing performance of the RPC detector. The optimal choice of the applied HV is
obtained by considering the efficiency and resolutions. The outline of the paper is the
following. Section \ref{chap_exptsetup}
describes the experimental setup. Data analysis techniques are described in
Section \ref{chap_analysis}. Section \ref{chap_conventional} shows the overall efficiency of a poor
quality RPC as a function of HV and the V-I curve of the
detector. Section~\ref{chap_multiplicity} discusses the response of
the RPC in different regions of the detector area. The overall
efficiency is calculated based on the optimum selection criteria explained
in Section~\ref{chap_efficiency}. The section
\ref{chap_resolution} shows the position and time resolution as a function of HV and finally
we conclude our findings in section \ref{chap_conclusion}.
\section{Experimental Setup}
\label{chap_exptsetup}
RPCs were introduced in the 1980s in the particle physics experiments as
a replacement for spark counters. Due to its rugged structure and
its large area coverage with uncompromised efficiency and time
resolution made it a significant element in the High Energy
Physics as Trigger detector and Time-of-Flight detector in various experiments.
The operation of the RPCs are based on the detection of the gas
ionization produced by the charged particle when it passes through the active area of the
gaseous chamber and subsequently the avalanche due to the strong electric field
in that region. The RPCs used as trigger detectors have a
gas gap in the order of a few mm. If the RPCs are used in the Time of Flight
(ToF), the gap in the order of a few 100\,$\mu$m used to get the better
time resolution (few 10 of ps). The gas mixture along with an operating
voltage decides the working mode of the detector whether it is
an avalanche or streamer. In avalanche mode, the electron
multiplication occurs due to the drifting and collision of the
electrons with gas molecules. The avalanche became the precursor of
the new process called streamer if the electron multiplication
reaches the extreme value \cite{raether,meek}.
In the Streamer phase, the kinetic energy of the electron is low and
the recombination of the electron-ion results in the photons and this
create delayed secondary avalanches with respect to the first
one. If the number of photons are large enough and the applied
electric potential is strong enough will produce a large number of
the secondary avalanches until the local density and electron-ion
distributions are such to create a plasma between two electrodes (electron and ion
distribution connects the two electrodes). In this process, the
extremely high current such as 100 time larger than the avalanche is
flown through the gas gap, until the electrons and ions are collected
by the electrode.
The RPCs used in this study are made by placing two thin glass plates of 3\,mm thickness, 2\,mm apart from
each other. The gap between the two glass plates is maintained to be 2\,mm using
poly-carbonate "button" spacers. The sides are sealed with side spacers to form a closed chamber.
There are gas inlet and outlet nozzles at the corners of the chamber through which the mixed
gas is flown. The outer surfaces of the glass chamber are coated with a thin film of graphite
paint for the application of external high voltage. The resistivity of the coat is
chosen by optimising the application of uniform
electric field in the gap as well as localising the induced signal in a small area.
A thin layer (50\,$\mu$m) of insulated mylar sheet is kept in between the chamber and
pickup panels on both sides. The pickup panels are made of the copper strips with a
width of 2.8\,cm and an inter-strip gap of 0.2\,cm pasted on the plastic honeycomb
material, where the other side is pasted with a thin aluminium layer for grounding.
The RPCs are operational in avalanche mode with a non-flammable gas mixture of
$\mathrm{C_2 H_2 F_4}$ (95.2\%), iso-$\mathrm{C_4H_{10}}$ (4.5\%) and $\mathrm{SF_6}$ (0.3\%), which are
continuously flown through the gas gap with a rate of 6SCCM per RPC.
To work on the RPCs in avalanche mode, the major component of
the gas mixture is the electronegative freon gas ($\mathrm{C_2 H_2 F_4}$) with enough primary
ionization production along with that low pathlength for the
electron capture. Due to the electronegative property of the gas, it
controls the electron multiplication within the Meek limit \cite{raether,meek}. The
iso-$\mathrm{C_4 H_{10}}$ is the polyatomic gas, which has high absorption
probability for the UV photons produced in the electron-ion
recombination. The absorbed energy is dissipated as the
rotational-vibrational energy levels. The third gas $\mathrm{SF_6}$ is used
with minimal volume fraction but it plays a dominant role in operating the
RPCs in the avalanche mode by quenching the electrons.
Few RPCs with a large non-uniformity in terms of efficiency and noise are studied to find the
optimum operating HV. There are many sources of the non-uniformity, e.g., uniformity
of the thickness of the glasses \& supporting spacers to maintain the gas gap, resistivity
in the graphite coating, flatness of the pickup panels, difference in the gain of
amplifiers as well as the threshold of the discriminators, the characteristic impedance of the
pickup strips due to variation of width as well as depth of the supporting material etc.
RPCs are tested using the existing 85\,ton miniICAL experimental
setup \cite{miniical_iichep}, which is a miniature version of ICAL and a schematic view
is shown in figure \ref{fig:miniical_view}. In absence of easily available test beam
facilities with muon beams, the miniICAL is used to trigger cosmic muon trajectory. Though
the miniICAL performance is being tested with the 1.4\,T magnetic field, this study used the
data collected without any magnetic field. Also, there are 20 RPC (two RPC in a gap
with 10 RPC in a vertical column) in the system, but this study is based on 10 RPC of
a column \footnote{The numbering of RPCs starts from the bottom and the top layer is numbered as 9.}.
The characteristic impedance of the strip is around $\sim$50\,$\Omega$. The
readout strips are placed orthogonally on either side of the RPC to measure the position
of the traversed charged particle (X-coordinate from the bottom strip and Y-coordinate from
the top strip plane). The induced signals due to the avalanche multiplication
inside the RPC are amplified and discriminated by an 8-channel NINO ASIC chip \cite{nino}.
The LVDS output of NINO is fed to the FPGA based RPC DAQ system. Finally the electronic
outputs
of an RPC are 128 logic units, to identify the induced signal in the strips and sixteen time
measurements. The 128 logic unit is defined as hit\footnote{Induced signal in a strip larger
than 100\,fC.} and used to localise the position of the traversing charged particle.
For the time measurement, every 8$^{th}$ strip are connected together to have a single output
of 8 strips.
This study is performed using cosmic muon events, which are triggered by the coincidence of signals in four fixed RPCs in
that column with the time window of 100\,ns. The logical ``AND'' is done for X- and Y-plane independently to form the trigger signals from X- and Y-plane. The final trigger is created by logical ``OR'' of the trigger signals from X- and Y-plane.
\begin{figure}
\center
\includegraphics[width=0.75\textwidth]{mini_ical_orgpic.png}
\caption{A view of 85\,ton miniICAL detector with 10 layers of RPC detector.}
\label{fig:miniical_view}
\end{figure}
\section{Data Analysis}
\label{chap_analysis}
One of the main disadvantages of this setup
is the uncertainty of the position and time measurement of muon trajectories in the test RPC.
In test beam setup, muon position is measured with the tracking device of position resolution
about $\sim$mm or less and time is measured very precisely from the beam delivery
system \cite{cmsrpc}. In this setup, those are measured in a crude way.
The typical noise rate of a strip is $\sim$100\,Hz at 10\,kV. The four-fold coincidence logic reduces the
events triggered by the random coincidence of noise. But, muon is contaminated by the
correlated electronic noise as well as cosmic showers mainly due to interactions of remaining
hadrons of cosmic ray particles on the roof of the building and the iron in miniICAL.
Most of the analysis framework is done independently using the information in X-strips and Y-strips
of different layers independently and then combined together to localise the position of the muon
in two dimensional space in an RPC.
The muon position in the RPC is
simply the mean of the strip positions in the RPC. For the timing measurement, strips with
an earlier timestamp amongst different strips are used as the measured time for that layer.
To remove the background events, each layer is considered only when there are at most three
hits in a plane and all three hits should be in consecutive strips. This criterion is used also
to reduce the small fraction of streamer pulse. Position resolution of an RPC is drastically
deteriorated with the multiplicity 4 or more \cite{pethu1} and will be discussed more in
section \ref{chap_multiplicity}.
The position alignment of an RPC with respect to the remaining system is done through an
iterative procedure, where the muon trajectory is fitted with
a simple straight-line equation using the information of all other detectors.
The test layer is aligned by comparing the measured position in that layer with respect to
the extrapolated position from the fit using other layers. This entire process is repeated iteratively where residual distributions of
positions are corrected and updated for all layers.
The details of the alignment technique are given in \cite{sumanta1} and the same iterative
procedure is also used for the correction of time offset, but for the timing in a layer
there are sixty-four offset corrections (for each and every individual strip) per plane.
After all these criteria, straight lines for the muon are fitted in both X-Z plane and Y-Z
plane independently. Each fit must contain at least six layers and $\chi^2/ndf$ is less than
two are considered for further study. Applying these conditions simultaneously on both X-Z
and Y-Z plan select almost 100\% pure muon signal in the event.
Using the fit parameters, muon trajectories are
extrapolated to the test RPC. The position resolution
of the RPC's are $\sim$8\,mm \cite{pethu1}. Thus the error on the extrapolation on top of the
miniICAL is also about $\sim$6-7\,mm, but this is minimum in the middle layers, which is
about $\sim$3\,mm. Similarly, the uncertainty of the time measurement in each layer is
$\sim$1\,ns \cite{apoorva} and error on the extrapolated time at the top of miniICAL is
0.8-0.9\,ns, whereas at the centre of the stack it is $\sim$0.4\,ns. These extrapolated uncertainties
on position and timing are larger than the typical values achieved in test beam
facilities \cite{hcaltestbeam}. Measured position and time resolutions are convoluted with the
error due to extrapolation, but for the comparison of the performance as a function of applied HV,
the effect will be the same for all HV. The choice of optimum HV will not affect much with
these extrapolated errors
\section {Efficiency Plateau of the RPC detectors}
\label{chap_conventional}
During the preliminary testing of the RPC under the study, the first
parameter is to measure the Voltage-Current characteristics of the RPC.
The measured RPC count rate per strip and current at different HV is
shown in Figs.~\ref{fig:effvshv}(a) and (b) respectively.
In general, the choice of the operating high voltage (HV) of the RPC detector is determined
by the overall detection efficiency of the RPCs for the charged particle
as a function of applied HV. The operating point (workingpoint) of the RPCs are decided by the ``knee'' voltage; the voltage at
which the overall efficiency of the RPCs reached 95$\%$ of the
maximum efficiency\cite{cmsrpc1,argoybj1}.
The efficiency of the RPC discussed in this section is estimated for the
pixel\footnote{Area of 3\,cm$\times$3\,cm in the RPC detector to match the strip pitch.} when
muon passes through the pixel. The size of the pixel is 3\,cm, whereas the extrapolation
error $\sim$8\,mm, to incorporate the uncertainty of extrapolation, any hit in the RPC strips
with the 3\,cm from the extrapolated position is used for the efficiency measurement.
Even with the disadvantage due to error in track extrapolation, there is also an advantage over
test beam or conventional muon telescope.
In general, the detection efficiency of the selected strips is estimated using the
cosmic muon telescope made of scintillators as trigger detectors and RPCs as
detectors under test. In that method, the performance of only a selected
region of the RPC is studied. The same is true for test beam activities also. But to study
the detailed (pixel-wise) performance of the RPC detectors over the whole area, the RPC
detectors kept in miniICAL or any other RPC stack is certainly a better option. It was
mentioned earlier that the uncertainty of the extrapolation is minimum in the middle of
the stack, but due to other considerations, the HV scan measurement is performed for the
RPC, which is kept in the 8th layer of the miniICAL. Muons events are collected with the
in-situ trigger from the time coincidence of layers 2, 3, 4 and 6.
In the different runs, the HV and all other conditions of all RPCs are kept the same except
the HV of layer 8. The muon trajectory is fitted with the hit points of all layers
excluding layer-8 and interpolated to layer-8 and matched with the measured hit position
in that layer.
The overall efficiency as a function of High Voltage for the RPC under observation is given
Fig.~\ref{fig:effvshv}(c). From the overall
efficiency measurement it is observed that the
the efficiency of the RPC is saturated at around 10.2\,kV of the applied
high voltage. The efficiency as a function of the HV is fitted using
the sigmoidal function is given in Eqn~\ref{eqn:sigmoid}\footnote{The observed behaviour of the X- and Y-plane is similar. Thus the results from one of the planes is described in further sections.},
\begin{equation}
\centering
\epsilon = \frac{\epsilon_{max}}{1+exp[-(HV-HV_{0.5})/\beta]}
\label{eqn:sigmoid}
\end{equation}
where $\epsilon_{max}$ is the maximum efficiency as HV is
$\infty$. $HV_{0.5}$ is the $HV$ value for which the efficiency is
half of the maximum efficiency ($\epsilon_{max}$).
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{VI_Noise_Eff_Plat.pdf
\caption{(a), (b) and (c) are average count rate per strip, the V-I curve and the measured efficiency vs high voltage for the RPC under observation.}
\label{fig:effvshv}
\end{figure}
\section{Pixel-wise efficiency and Strip Multiplicity for different HV}
\label{chap_multiplicity}
Comparing the knee of the efficiency plots with other results, e.g., Ref.~\cite{cmsrpc}, the large
difference is observed between the applied HV for the saturation in the efficiency and the 50\% efficiency
of the RPC. This behaviour might be due to the variation of gain in different regions of the RPC,
where certain regions have larger gain and reach the maximum efficiency earlier than the regions with low gas gain.
The figure \ref{fig:effhitmult9.4kv}(a) shows
the pixel-wise combined efficiency of X\&Y-side for the whole RPC at 9.4\,kV. The three
bands parallel to the Y-axis are due to the malfunction of electronics in the three strips. There
are regions where efficiency is almost zero and some regions are more than 90\% efficient.
The efficiency will increase with applied HV, but to have saturated efficiency of these
low gain regions, one needs a substantial increase of applied HV.
Depending on the efficiency in the RPC at 9.4\,KV, the whole surface area of the RPC
is divided in 10 zones,
with efficiency, 2-10\%, 10-20\% and so on.
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Trig_Eff_Mult_2zone_9p4kv.pdf
\caption{(a), (b) and (c) are the pixel-wise efficiencies and hit multiplicity distribution for efficiency zone 20-30\,$\%$ and 50-60\,$\%$ of the RPC at applied voltage of 9.4\,kV.}
\label{fig:effhitmult9.4kv}
\end{figure}
Along with the overall efficiency of the RPCs, the other parameter which tells about the
behaviour of the RPC detectors is the cluster size or hit multiplicity distribution.
The hit multiplicity distribution for the regions with efficiency of 20-30\,$\%$ and 50-60\,$\%$ at the applied voltage of 9.4\,kV are
shown in Figs.\ref{fig:effhitmult9.4kv}(b) and (c). In general for a stable operational
RPC, the strip
multiplicity produced by avalanche due to the passage of the muon through RPC detector
is upto three strips. The hit multiplicity beyond three and until $\sim$20 is mostly due
to the hadronic shower and steamer development inside the gas gap. The events with strip
multiplicity of more than $\sim$20 is due to electronic noise that occurred due to EMI from various
sources around the experiment. But, there is no sharp boundary between the streamer and
electronic noise. From the operational point of view, hit multiplicity increases with the applied HV.
\begin{figure}
\center
\includegraphics[width=0.45\linewidth]{Trig_Eff_Zone_wise_Plat.pdf
\includegraphics[width=0.45\linewidth]{Mult_Zone_wise.pdf}
\caption{(a) and (b) are the measured efficiency and average cluster size vs high voltage for different zones of the test RPC.}
\label{fig:effvshvrange}
\end{figure}
The average efficiencies
in one of the RPC planes at different HV
for different zones are shown in the Fig.~\ref{fig:effvshvrange}(a). It is clearer from the
figure that the different zone reaches the saturated efficiency at different HV. The inset plot
of Fig.~\ref{fig:effvshvrange}(a) shows the decrease of efficiency at very high voltage, which is
due to a visible dead time of the detector caused by the large noise rate.
The average hit multiplicities (cluster size) for
different HV at various efficiency zones are shown in Fig.~\ref{fig:effvshvrange}(b). Similar to the efficiencies, the multiplicities are also a
combined function of the zone and applied HV. It was mentioned earlier that one of the sources
of high multiplicity is the streamer formation and it causes the deterioration of position resolution.
Earlier studies \cite{pethu1} showed that the position resolution drastically deteriorates
with the multiplicity of four or more. The operation of this RPC at 10.1\,kV will have a
large region with an average multiplicity of four or more and can not be used for any physics study,
e.g., identification of trajectories and the accurate measurement of momentum
due to poor position resolution. The position resolution for different hit multiplicities from one to five and multiplicities greater than four hits for the observation RPC at 10.1\,kV for efficiency region of 60-70\,$\%$ are given in Figs.~\ref{fig:pos_reso_mult12345} (a) to (f). The observed position resolution for applied HV vs efficiency zones for different multiplicities (from one to five and greater than four) are given in Figs.~\ref{fig:pos_reso_mult12345_2d} (a) to (f). Some regions are empty due to the omission of resolutions with a very low number of events.
At the low HV, events with larger multiplicities, e.g., three or more are dominated by noise and similarly at larger HV, events with
low multiplicity are also dominated by noise associated with inefficiency of the detector. The
multiplicity increases with HV for the whole surface area, but it is always less in the low-efficiency
zone in comparison with higher efficiency zones at each high voltage. The lowest resolutions for each
multiplicity, which is nearly diagonal points in these plots, are the regions of high purity signal. With high voltage there are relatively more events with a large avalanche pulse,
thus resolution with multiplicity four is reduced
with HV. Though the resolutions for multiplicity more than four are always larger than the events with multiplicity 1 to 3, except
for multiplicity one with very high voltage. The larger resolution is mainly due to the contamination of
streamer pulses. Similarly, for five-hit multiplicity, it is similar but happening at an even higher
voltage. At larger HV, e.g., 10.6\,V, events with multiplicity one is dominated by the region of
dead strips and consequently observed large resolution.
Overall, the observed position resolution for the events with four hits or more shows a substantial
deterioration in comparison with the events with multiplicities upto three.
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Pos_Reso_1d_mult12345.pdf
\caption{The measured position resolution of the observation layer for different multiplicity from (a) one to (e) five and (f) greater than four at the 60-70\,$\%$ efficiency region.}
\label{fig:pos_reso_mult12345}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Layer_Reso_2D.pdf
\caption{The measured position resolution of the observation layer for HV vs efficiency region for different multiplicity from (a) one to (e) five and (f) greater than four.}
\label{fig:pos_reso_mult12345_2d}
\end{figure}
\section{Efficiency using Selected Events}
\label{chap_efficiency}
The efficiency of the RPC which was discussed in the section \ref{chap_conventional} is estimated without
any selection criteria on the multiplicities, the event can have any number of hits in the RPC
under the study. It was discussed in the previous section that the avalanche can induce
upto $\sim$ three strip hits and beyond that is not recommended to use for any physics
study due to worse localisation of signal. Thus, the definition of the
efficiency is modified with the criteria of accepting events with at most three consecutive hit
in this test RPC. The average efficiencies of the selected events defined as selected
efficiency for the whole detector as a function of applied voltage are shown in Fig.~\ref{fig:seleffhv}
\begin{figure}
\center
\includegraphics[width=0.66\linewidth]{Sel_Eff_Plat_RPC155.pdf
\caption{The measured efficiency from the selected events vs high voltage for the RPC under observation.}
\label{fig:seleffhv}
\end{figure}
The average efficiency has a peak at $\sim$9.8-9.9\,kV for the RPC under observation
suggests that the optimum operating voltage of the RPC under the study
should be 9.8-9.9\,kV.
Thus from the efficiency point of view, the optimum operating voltage of this RPC is less than
10.1\,kV which was obtained earlier with the simple efficiency measurement. Beyond 10.1\,kV, the
efficiency of the selected events are decreasing because of the rejection of events with large
multiplicities, mainly due to the formation of streamer pulses and
it is more for high gain regions.
\section{Position and Time resolution of the RPCs}
\label{chap_resolution}
The efficiency is not the only ultimate test for the optimisation of the applied HV. The position
resolution drastically deteriorated for multiplicity four or more, but also these are worse
for multiplicity three in comparison with multiplicity one or two. Even after rejecting the events with hit multiplicities four or more, the physics parameters can be affected bit by the hit multiplicity three.
Until now the main parameter considered only the efficiency (with and without multiplicity
criteria) to choose the operating voltage of the RPC. But to look at more detailed performance
the physics parameters of the RPCs such as position and time resolution of the RPC
are very important. The position residues, the difference between the observed and extrapolated
positions are distributed and the resulting distribution follows the gaussian shape. The
observed distribution is fitted with the Gaussian function the fitted $\sigma_{pos}$ is considered
as the position resolution of the RPC. Similar to the position resolution, the time resolution is also
estimated by fitting the distribution of the difference between the observed and extrapolated
time of the RPC layer. The distribution of position and time residues at 9.8\,kV are shown in
Figs.~\ref{fig:postimereso} (a) to (f) for different efficiency
regions such as 20-30\,$\%$, 50-60\,$\%$ and 80-90\,$\%$.
\begin{figure}
\center
\includegraphics[width=1.1\linewidth]{Pos_Time_Residue_9p8.pdf
\caption{(a), (b) and (c) are the position residue distribution at
9.8\,kV for efficiency zones of 20-30\,$\%$, 50-60\,$\%$ and
80-90\,$\%$ respectively. (d), (e) and (f) are the time residue
distribution at 9.8\,kV for efficiency zones of 20-30\,$\%$,
50-60\,$\%$ and 80-90\,$\%$ respectively.}
\label{fig:postimereso}
\end{figure}
The position and time offsets calculated from the muon data recorded
with different HV.
The mean of the time residues (can
be called avalanche delay with respect to the trigger signal) is expected to
show the correlation with HV (which is calculated with the time corrections calculated using the data from 9.8\,kV). As the HV increases the signal height
increases at a faster speed, consequently the threshold crossing time for the avalanche signal is earliar with respect to the trigger signal. From Fig.~\ref{fig:postimereshv}(a), the avalanche
delay vs HV is fitted using the straight-line upto 10.2\,kV until
there the linearity exists. The obtained slope for one of the planes is $-$7.51\,ns/kV.
The change in the slope beyond 10.2\,kV is mainly due to the effect of a large noise rate.
The fitted $\sigma_{pos}$ and $\sigma_{time}$ as a function of the
applied voltage for the RPC under study is shown in
Figs.~\ref{fig:postimereshv} (b) and (c).
The variation in the position resolution with different applied voltage shows improving behaviour
upto HV$_{pos}^{min}$=9.8-9.9\,kV. But beyond the HV$_{pos}^{min}$, the position resolution is
deteriorating due to the increased contribution of the multiplicities which
results in the poor localisation of the signal. The time resolution of the RPC under the study
improves as the applied voltage increases upto HV$_{time}^{min}$=10.4\,kV. The larger applied voltage
above HV$_{time}^{min}$ shows an increase in the time resolution.
During each avalanche, even for the noise pulse, the effective high voltage at the avalanche
position (few mm$^2$) is reduced below the threshold voltage to form an avalanche signal,
which causes the inefficiency as seen in the inset plot of Fig.~\ref{fig:effvshvrange}(a) and
gradually the high voltage is built up in that region. During the recovery time, the time taken
by the system to regain its stable voltage configuration, the effective HV is low and consequently
the gain is also low.
This explains the change in slope in Figs.~\ref{fig:postimereshv} (a) and increase of resolution in Figs.~\ref{fig:postimereshv} (c).
This argument is supported by the time residues distribution in the larger HV samples. Figs~\ref{fig:timeskew} (a), (b) and (c) are the time residue distribution for the efficiency zone of 50-60\,$\%$ for 10.2\,kV, 10.4\,kV and 10.6\,kV respectively. The time residue distribution for larger HV has an increasing number of events in the right side bump (which is small signals that cross the discriminator threshold later than the muon trigger time) and has increasing skewness at a larger voltage in comparison of skewness in Fig.~\ref{fig:postimereso}(d) to Fig.~\ref{fig:postimereso}(f).
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Avalanche_Delay_RPC155.pdf
\caption{(a) is the mean of the time residues
as a function of applied HV. (b) and (c) are the
position and time resolution as a function of the applied voltage.}
\label{fig:postimereshv}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Time_Reso_10p2_10p4_10p6.pdf
\caption{(a), (b) and (c) are the time residue distribution for the efficiency zone of 50-60\,$\%$ for 10.2, 10.4 and 10.6\,kV.}
\label{fig:timeskew}
\end{figure}
The results of selected efficiency, as well as the position resolution, suggests that
the optimum operating point of this RPC is 9.8-9.9\,kV, not the 10.1-10.2\,kV, which was obtained using the
standard efficiency plots.
Another RPC, of non-uniform gain, is also studied in the same
manner. The default efficiency and selected efficiency are shown in
Figs.~\ref{fig:newrpcomb}(a) and (b). Fig.~\ref{fig:newpostimerms}(a) is the mean of the time
residue as a function of HV is fitted using a straight line and the obtained slope for one of the planes is -6.21\,ns/kV. The
position resolution and time resolution at different operating voltages are shown in
Figs.~\ref{fig:newpostimerms}(b) and (c) respectively. That also shows that the optimum operating HV of the RPC for the best
physics performance is obtained at $\sim$9.8\,kV, not 10.1-10.2\,kV, which is obtained from the plateau of default efficiency.
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Eff_Sel_Eff_L7_Paper.pdf
\caption{(a) and (b) are the default efficiency and selection
efficiency as a function of applied HV for another RPC.}
\label{fig:newrpcomb}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.99\linewidth]{Avalanche_Delay_RPCL7.pdf
\caption{(a) is the mean of the time residues
as a function of applied HV. (b) and (c) are the position and time resolution as a function of the applied voltage.}
\label{fig:newpostimerms}
\end{figure}
\section{Conclusion}
\label{chap_conclusion}
Two RPCs were tested at different operating HV to choose the optimum operating HV for the
best physics performance of the system. Due to the non-uniform response of these RPCs over
the whole surface area, different regions show the saturation of efficiencies at different
HV. Different zones can be considered as different individual RPCs, which should have
different optimum operating HV. Though the plateau of efficiency apparently shows
a larger operating voltage in terms of uniform gain, the optimum operating HV for the best
performance of the detector is lower than that. This paper shows examples of two RPCs with
large non-uniformities in gain. In general, these kinds of chambers is not installed in an
experiment. But, even one installs a good RPC in beginning, during the operational phase,
characteristics of the RPC might change, e.g. change in resistivity in the graphite coating,
bulging due to button popup etc., and then certainly one needs to check the optimum
operating HV for the best physics out of it, not simple efficiency plateau. Even with a small
variation of gain in an RPC, one should choose the operating HV by optimising the selected
efficiency, resolutions but not only overall efficiency.
\acknowledgments
We would also like to thank all members of the INO collaboration for their valuable inputs.
|
\section{INTRODUCTION}
Spintronics, which uses the spin degrees of freedom of electrons for information transmission, storage and processing, has attracted extensive attention because of its unique advantages of low power consumption, fast data processing speed and high integration density~\cite{1S. A. Wolf-Science-2001}. Two-dimensional (2D) ferromagnets have potential applications in nanoscale spintronic devices. A large number of 2D materials~\cite{2K. S. Novoselov-Science-2004,3K. S. Novoselov-Proceedings-2005,4C. Jin-Phys. Rev. Lett.-2009,5M. Chhowalla-Nature Chemistry-2013,6K. F. Mak-Phys. Rev. Lett.-2010,7A. Splendiani-Nano Lett.-2010} have been discovered in recent years. However, the lack of intrinsic ferromagnetism heavily restricts their application in spintronic devices. The ferromagnetic (FM) ordering of 2D materials can be obtained by doping~\cite{8B. Wang-Nanoscale Horiz.-2018,9N. Miao-J. Am. Chem. Soc.-2017,10B. Li-Nature Communications-2017}, external strain~\cite{11B. Xu-Applied Physics Letters-2020,12Y. Ma-ACS Nano-2012}, external electric field~\cite{13Z. Fei-Nature Materials-2018} and defect engineering~\cite{14Z. Zhang-ACS Nano-2013,15J. Guan-ChemPhysChem-2013,16Y. Tong-Advanced Materials-2017}. Where external strain is an effective approach to adjust the electronic structures and magnetic properties of the low-dimensional materials. Many theoretical reports have indicated the tunable electronic structures and magnetic characteristics in strained monolayers~\cite{Y. Ma-ACS Nano-2012,S. Bertolazzi-ACS Nano-2011,X. Chen-Phys. Lett. A-2015,Y. Ma-Nanoscale-2011,L. Kou-ACS Nano-2011}. In the experiment, the application of adjustable biaxial strain to 2D materials has made remarkable progress~\cite{F. Ding-Nano Lett.-2010}. Additionally, under a biaxial tensile strain of approximately 13\% in MnPSe$_3$~\cite{17Q. Pei-Frontiers of Physics-2018}, it occurs a magnetic phase from antiferromagnetic (AFM) to FM state, and this transition is also achieved by electron and hole doping induce~\cite{18X. Li-J. Am. Chem. Soc.-2014}. The tunable magnetic properties of 2D ferromagnets have attracted great interest. On the other hand, the discovery of 2D magnetic materials with high spin polarization, large magnetic anisotropy energy (MAE), and high Curie temperature (\emph{T}$_c$) would promote the development of spintronic devices, and also would provide new opportunities in low-power-consumption spintronics and quantum computing, among many other applications~\cite{Z. Fei-Nat. Mater-2018,Y. Deng-Nature-2018,Y. Deng-Nature-2018}. Half-metallic materials are conducting in one spin channel but insulating in another optional channel which exhibiting 100\% spin polarization~\cite{19X. Li-J. Am. Chem. Soc.-2014,20M. Frik-New Journal of Physics-2007}. These properties of half-metallic materials are highly desired in many spin-based devices. Therefore, the 2D half-metals are ideal materials for spintronic nanoscale devices~\cite{21C. Felser-Angewandte Chemie International Edition-2007}.
In fact, the electronic structures largely determine the physical properties of materials, due to the destruction of structural symmetry of low dimensional materials, it can be significantly modulated. In recent years, 2D Janus materials have received extensive attention in the research field. The Janus monolayers due to out-of-plane asymmetry exhibit extraordinary physical characteristics, such as the piezoelectric polarization~\cite{22C. Zhang-Nano Lett.-2019} and preferred catalytic performance~\cite{23D. Er-Nano Lett.-2018,24X. Ma-J. Mater. Chem. A-2018}. The 2D Cr$_2$XS$_3$ (X = Br, I) are room-temperature magnetism Janus semiconductors by substituted one layer of halogon atoms with sulfur atoms to break symmetry of CrX$_3$ (X = Br, I) monolayers~\cite{25D. Wu-J. Phys. Chem. Lett.-2021}. Janus transition metal dichalcogenides MXY (M = Mo, W; X, Y = S, Se, Te; X $\neq$ Y)~\cite{26L. Dong-ACS Nano-2017} and Janus MoSSe monolayer show intrinsic dipole and piezoelectric effects~\cite{27A.-Y. Lu-Nature Nanotechnology-2017}. Multi-Functional Janus vanadium dichalcogenides VXX' (X/X' = S, Se, Te) also exhibit excellent physical properties, and their Curie temperature can be enhanced by the built-in electric field effect~\cite{S. Ji-Chin. Phys. Lett-2020}. Other Janus materials also show excellent properties, such as FeXY (X, Y = Cl, Br, and I, X $\neq$ Y)~\cite{28R. Li-Nature Nanotechnology-2017}, V$_2$X$_3$Y$_3$(X, Y = Cl, Br and I; X $\neq$ Y)~\cite{29Y. Ren-Phys. Rev. B-2020},M$_2$SeTe(M = Ga, In)~\cite{30Y. Guo-Appl. Phys. Lett.-2017}, and the single-sided hydrogenated graphene~\cite{31J. Zhou-Nano Lett.-2009}. Overall, 2D Janus materials greatly promote the development of spintronic devices and expect to have potential applications in electronic and electromechanical devices.
In this work, based on first-principles, we systematically investigated the electronic and magnetic properties of Janus V$_2$AsP monolayer, demonstrating highly mechanical and dynamic stability. The Janus V$_2$AsP monolayer with large half-metallic band gap and spin band gap is an FM half metallic material. Moreover, we applied a biaxial strain to Janus V$_2$AsP monolayer. It is accompanied by the transition from an FM to AFM. Furthermore, the predicted \emph{T}$_c$ of V$_2$AsP monolayer is 83 K by Monte Carlo simulations. And the \emph{T}$_c$ can be enhanced with the increase of compressive strain.
\section{Methods}
The calculations of this work were performed by using the Vienna ab initio simulation software package (VASP)~\cite{32G. Kresse-Phys. Rev. B-1993,33G. Kresse-Phys. Rev. B-1993} based on the spin polarization density function theory (DFT)~\cite{34W. Kohn-Phys. Rev.-1965,35P. Hohenberg-Phys. Rev.-1964}. The exchange correlation potential was calculated under the Perdew-Burke-Ernzerhof (PBE) function of Generalized Gradient Approximation (GGA)~\cite{36J. P. Perdew-Phys. Rev. Lett.-1996}. The plane-wave cutoff energy was set to be 500 eV. The tolerance criterion of energy and force were set to be 10$^{-6}$ and 0.01 eV/{\AA}, respectively. A $9\times9\times1$ Monkhorst-Pack special k-point mesh was used in Brillouin zone~\cite{37H. J. Monkhorst-Phys. Rev. B-1976}. A vacuum slab of 20 {\AA} was added along the z axis to avoid the interactions between the adjacent monolayers. The strong field Coulomb interaction was considered by using the GGA+U method, onsite Coulomb interaction parameter was set to be 3 eV based on the relevant previous reports~\cite{38J. He-J. Mater. Chem. C-2016}. Phonon dispersion spectrum was analyzed based on density functional perturbation theory (DFPT) as implemented in the Phonopy code~\cite{39A. Togo-Phys. Rev. B-2008}.
\begin{figure}[t!hp]
\centerline{\includegraphics[width=0.85\textwidth]{fig1.pdf}}
\caption{(a) Top and (b)(c) side views of Janus V$_2$AsP monolayer. (d)-(h) One Ferromagnetic (FM) and four antiferromagnetic (AFM) configurations.
\label{fig:stru}}
\end{figure}
\begin{figure}[tbp!]
\centerline{\includegraphics[width=0.80\textwidth]{fig2.pdf}}
\caption{Phonon spectrum of the Janus V$_2$AsP monolayer
\label{fig:64-E-m}}
\end{figure}
\begin{figure}[tbp!]
\centerline{\includegraphics[width=0.95\textwidth]{fig3.pdf}}
\caption{(a)(b) The band structures, (c)total density of states (TDOS) and projected density of states (PDOS) for Janus V$_2$AsP monolayer; (d)charge density difference isosurface plot and (e)spin density isosurface plots for ground state spin arrangements
\label{fig:64-E-k}}
\end{figure}
\section{ RESULTS and DISCUSSION}
As shown in Fig.~\ref{fig:stru}(a)-(c), the Janus V$_2$AsP monolayer consists of one V layer sandwiched between the As layer and P layer with lattice parameters a = b = 4.43 {\AA}, which is isostructural to the MnX (X=P, As) monolayer~\cite{B. Wang-Nanoscale-2019}. The bond lengths of V-As, V-P, and V-V are 2.54 {\AA}, 2.40 {\AA}, and 3.13 {\AA}. The $2\times2\times1$ supercell with one FM and four AFM configurations shown in Fig.~\ref{fig:stru}(d)-(h) are considered to ascertain the magnetic ground state of Janus V$_2$AsP monolayer. We calculated the energy difference ($\Delta$E = E$_F$$_M$ $-$ E$_A$$_F$$_M$) relative to FM configurations are 33.32, 21.18, 45.98, and 76.94 meV per V atom for AFM1, AFM2, AFM3, and AFM4 configurations, our results suggest that the magnetic ground state of Janus V$_2$AsP monolayer is FM. Besides, the non-magnetic (NM) state can be neglected due to the tremendous energy difference between NM state and magnetic states.
Next, the elastic stiffness tensors C$_{11}$ (C$_{11}$ = C$_{22}$) and C$_{12}$ were calculated as 28.95 and 8.90 N/m. It indicates that the Janus V$_2$AsP monolayer is mechanically stable according to the Born stability criterion (C$_{11}{>}$0, C$_{22}{>}$0 and C$_{11}$-C$_{12}{>}$0)~\cite{40Z.-j-Phys. Rev. B-2007}. The Young's modulus (\emph{Y}$_{2D}$) and Poission's ratio were evaluated to be 26.21 N/m and 0.31, respectively. The value of Young's modulus is lower than that of MoS$_2$~\cite{41S. Bertolazzi-ACS Nano-2011}, indicating that V$_2$AsP monolayer is softer and can sustain a large strain. To verify the dynamic stability of V$_2$AsP monolayer, phonon spectrum was calculated. The phonon dispersion relation has no imaginary frequency (Fig.~\ref{fig:64-E-m}), indicating that Janus V$_2$AsP monolayer is dynamically stable.
The band structures and density of states for V$_2$AsP monolayer are shown in Fig.~\ref{fig:64-E-k}(a)(b). We find that the spin-up channels cross the Fermi energy level, while the spin-down channels have band gap. Therefore, the Janus V$_2$AsP monolayer is a half-metallic material which satisfies 100\% spin polarization. Half-metallic materials are semiconductors in one spin channel and in another optional channel exhibit metallic character, which have potential applications for spintronic devices~\cite{42C. Felser-Angewandte-2007}. The half-metallic band gap and spin gap with PBE+U functional are 0.38 eV and 1.34 eV. The PBE+U method always underestimates the band gap, then we used the Heyd-Scuseria-Ernzerh (HSE06) hybrid functional method to obtain more accurate electronic structures(in Fig.S1, supplementary material). Note that the half-metallic gap (0.61 eV) and the band gap of spin-down channel (1.93 eV) are larger than that of PBE+U method. Fig.~\ref{fig:64-E-k}(c), the density of states show that the spin polarization at the Fermi level mainly derives from the V atoms, the valence band maximum (VBM) of the spin-down channel is mainly contributed by V, As and P atoms, while the contribution of conduction band minimum (CBM) is V atoms. Fig.~\ref{fig:64-E-k}(d) shows the differential charge density which is the difference between the charge density at the bonding point and the atomic charge density at the corresponding point of Janus V$_2$AsP monolayer. The yellow and blue regions represent the net charge gain and loss, respectively. It is obvious that As and P atoms gain electrons, while the V atoms lose electrons. The V-As and V-P bonds show more ionic. As illustrated in Fig.~\ref{fig:64-E-k}(e), the local magnetic moments are mainly the contribution of V atom which is consistent with previous analysis of magnetic moment.
To estimate the \emph{T}$_c$ of Janus V$_2$AsP monolayer, we performed Monte Carlo simulations of the Heisenberg model. The Hamiltonian:
\begin{eqnarray}
{H = - \sum\limits_{ < ij > } {{J_{1}}{S_i}{S_j}}- \sum\limits_{ < ik > } {{J_{2}}{S_i}{S_k}}- A{{S_i^Z}{S_i^Z}}},
\end{eqnarray}
where J$_1$ and J$_2$ are the exchange parameters between the nearest and the next-nearest, S is the spin vector of V atom, S$_i^Z$ is the spin along z direction, and A is the parameter of magnetic anisotropy. To extract the exchange parameters, we utilize:
\begin{eqnarray}
{E(FM) = {E_0} - 2{J_1}{S^2}- 2{J_2}{S^2}- A{S^2}},
\end{eqnarray}
\begin{eqnarray}
{E(AFM1) = {E_0} + 2{J_1}{S^2}- 2{J_2}{S^2}- A{S^2}},
\end{eqnarray}
\begin{eqnarray}
{E(AFM3) = {E_0} + 2{J_2}{S^2}- A{S^2}},
\end{eqnarray}
Where E$_0$ is the energy without spin polarization, and the exchange parameters J$_1$ and J$_2$ of V$_2$AsP monolayer are calculated to be 8.3 and 7.3 meV per V atom. The result is shown in Fig.S2, supplementary material; indicating the \emph{T}$_c$ of Janus V$_2$AsP monolayer is 83 K. It is significantly that the \emph{T}$_c$ is higher than CrI$_3$ monolayer (45 K)~\cite{43B. Huang-Nature-2017}, Cr$_2$Ge$_2$Te$_6$ bilayer (30 K)~\cite{44C. Gong-Nature-2017}, CrSiTe$_3$ (35.7 K) and CrGeTe$_3$ (57.2 K)~\cite{45C. Gong-Nature-2017}. The determination of easy axis is able to study the magnetic coupling effect. In Fig.~\ref{fig:64-E-b}(a), we plot $\Delta$E = E$_a${$_b$ $-$ E$_{\theta}$ as a function of $\theta$, the E$_a${$_b$} and E$_{\theta}$ represent the energy difference between the spin directions of parallel and $\theta$ angled to the ab plane. We can find that the V$_2$AsP monolayer prefers an in-plane array and tend to decrease. Magnetic anisotropy is the key to the establishment of long-range 2D ferromagnetism. MAE mainly comes from the influence of spin-orbit coupling (SOC), and it can effectively counteract the influence of thermal fluctuation. MAE is defined as the energy difference between in plane and out of plane magnetization directions, the MAE of Janus V$_2$AsP monolayer is calculated as 177.39 $\mu$eV, which easy axis is in-plane spin arrays. The observed large MAE indicates that V$_2$AsP monolayer has potential for application in magnetic storage devices.
Besides, we analyzed the effect of biaxial strain on the properties of V$_2$AsP monolayer. The strain $\varepsilon$ is defined as (a $-$ a$_0$)$/$a$_0$, where a$_0$ and a are the lattice constant for the unstrained and strained system, respectively. As shown in Fig.~\ref{fig:64-E-b}(b), the red line represents the energy difference ($\Delta$E) between FM and AFM configurations, illustrating that the $\Delta$E increases with the decrease in compressive strain and the increase in tensile strain. when the $\Delta$E becomes greater than zero, the phase transition from FM to AFM state occurs at 4.9\% tensile strain. Meanwhile, the black line shows the change of Curie temperature under biaxial strain. With the decrease of $\varepsilon$, the Curie temperature tends to increase. We estimated that a compressive strain of -5\% can make the \emph{T}$_c$ reach to 130 K.
\begin{figure}[tbp!]
\centerline{\includegraphics[width=0.95\textwidth]{fig4.pdf}}
\caption{(a) Energy difference $\Delta$E between E$_a${$_b$} and E$_{\theta}$ per unit cell as a function of the $\theta$. (b) Energy difference between FM and AFM phases (red line) and the Curie temperature (black line) under the biaxial strain for V$_2$AsP monolayer
\label{fig:64-E-b}}
\end{figure}
\section{CONCLUSION}
In summary, we have investigated the stability, electronic and magnetic properties of Janus V$_2$AsP monolayer by using first-principles calculations. The phonon spectrum calculations indicate that it is dynamically stable. The Janus V$_2$AsP monolayer with 100\% spin polarization is an FM half-metal. The half-metallic band gap is 0.38 eV and the spin gap for the semiconducting channel is 1.34 eV. Moreover, the MAE with an in-plane easy magnetization direction is 177.39 $\mu$eV. Monte Carlo simulations based on the Heisenberg model evaluate the \emph{T}$_c$ of 83 K for the V$_2$AsP monolayer, which is higher than recently reported \emph{T}$_c$ of CrI$_3$ (45 K) and Cr$_2$Ge$_2$Te$_6$ (30 K). A tensile strain of 4.9\% is applied to Janus V$_2$AsP monolayer, then a phase transition occurs from FM to AFM state. The \emph{T}$_c$ of Janus V$_2$AsP monolayer can be elevated to 130 K by a compressive strain of -5\%. The Janus V$_2$AsP monolayer might promote the development of spintronic nanoscale devices.
\begin{acknowledgments}
This work was supported by the Natural Science Foundation of Hebei Province (No. A2019203507 and B2021203030). The authors thank the High Performance Computing Center of Yanshan University.
\end{acknowledgments}
\section{References}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.